1
linux/arch/x86/events/intel/ds.c

2548 lines
72 KiB
C
Raw Permalink Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 07:07:57 -07:00
// SPDX-License-Identifier: GPL-2.0
#include <linux/bitops.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/sched/clock.h>
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
#include <asm/cpu_entry_area.h>
#include <asm/debugreg.h>
#include <asm/perf_event.h>
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
#include <asm/tlbflush.h>
#include <asm/insn.h>
#include <asm/io.h>
#include <asm/timer.h>
#include "../perf_event.h"
/* Waste a full page so it can be mapped into the cpu_entry_area */
DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
/* The size of a BTS record in bytes: */
#define BTS_RECORD_SIZE 24
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
#define PEBS_FIXUP_SIZE PAGE_SIZE
/*
* pebs_record_32 for p4 and core not supported
struct pebs_record_32 {
u32 flags, ip;
u32 ax, bc, cx, dx;
u32 si, di, bp, sp;
};
*/
union intel_x86_pebs_dse {
u64 val;
struct {
unsigned int ld_dse:4;
unsigned int ld_stlb_miss:1;
unsigned int ld_locked:1;
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
unsigned int ld_data_blk:1;
unsigned int ld_addr_blk:1;
unsigned int ld_reserved:24;
};
struct {
unsigned int st_l1d_hit:1;
unsigned int st_reserved1:3;
unsigned int st_stlb_miss:1;
unsigned int st_locked:1;
unsigned int st_reserved2:26;
};
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
struct {
unsigned int st_lat_dse:4;
unsigned int st_lat_stlb_miss:1;
unsigned int st_lat_locked:1;
unsigned int ld_reserved3:26;
};
struct {
unsigned int mtl_dse:5;
unsigned int mtl_locked:1;
unsigned int mtl_stlb_miss:1;
unsigned int mtl_fwd_blk:1;
unsigned int ld_reserved4:24;
};
struct {
unsigned int lnc_dse:8;
unsigned int ld_reserved5:2;
unsigned int lnc_stlb_miss:1;
unsigned int lnc_locked:1;
unsigned int lnc_data_blk:1;
unsigned int lnc_addr_blk:1;
unsigned int ld_reserved6:18;
};
};
/*
* Map PEBS Load Latency Data Source encodings to generic
* memory data source information
*/
#define P(a, b) PERF_MEM_S(a, b)
#define OP_LH (P(OP, LOAD) | P(LVL, HIT))
perf/x86: Fix data source decoding for Skylake Skylake changed the encoding of the PEBS data source field. Some combinations are not available anymore, but some new cases e.g. for L4 cache hit are added. Fix up the conversion table for Skylake, similar as had been done for Nehalem. On Skylake server the encoding for L4 actually means persistent memory. Handle this case too. To properly describe it in the abstracted perf format I had to add some new fields. Since a hit can have only one level add a new field that is an enumeration, not a bit field to describe the level. It can describe any level. Some numbers are also used to describe PMEM and LFB. Also add a new generic remote flag that can be combined with the generic level to signify a remote cache. And there is an extension field for the snoop indication to handle the Forward state. I didn't add a generic flag for hops because it's not needed for Skylake. I changed the existing encodings for older CPUs to also fill in the new level and remote fields. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: http://lkml.kernel.org/r/20170816222156.19953-3-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-16 15:21:54 -07:00
#define LEVEL(x) P(LVLNUM, x)
#define REM P(REMOTE, REMOTE)
#define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
/* Version for Sandy Bridge and later */
static u64 pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
perf/x86: Fix data source decoding for Skylake Skylake changed the encoding of the PEBS data source field. Some combinations are not available anymore, but some new cases e.g. for L4 cache hit are added. Fix up the conversion table for Skylake, similar as had been done for Nehalem. On Skylake server the encoding for L4 actually means persistent memory. Handle this case too. To properly describe it in the abstracted perf format I had to add some new fields. Since a hit can have only one level add a new field that is an enumeration, not a bit field to describe the level. It can describe any level. Some numbers are also used to describe PMEM and LFB. Also add a new generic remote flag that can be combined with the generic level to signify a remote cache. And there is an extension field for the snoop indication to handle the Forward state. I didn't add a generic flag for hops because it's not needed for Skylake. I changed the existing encodings for older CPUs to also fill in the new level and remote fields. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: http://lkml.kernel.org/r/20170816222156.19953-3-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-16 15:21:54 -07:00
P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
OP_LH | P(LVL, L1) | LEVEL(L1) | P(SNOOP, NONE), /* 0x01: L1 local */
OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
OP_LH | P(LVL, L2) | LEVEL(L2) | P(SNOOP, NONE), /* 0x03: L2 hit */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, NONE), /* 0x04: L3 hit */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, MISS), /* 0x05: L3 hit, snoop miss */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT), /* 0x06: L3 hit, snoop hit */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM), /* 0x07: L3 hit, snoop hitm */
OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x08: L3 miss snoop hit */
OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, HIT), /* 0x0a: L3 miss, shared */
OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x0b: L3 miss, shared */
OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | SNOOP_NONE_MISS, /* 0x0c: L3 miss, excl */
OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
OP_LH | P(LVL, IO) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
};
/* Patch up minor differences in the bits */
void __init intel_pmu_pebs_data_source_nhm(void)
{
perf/x86: Fix data source decoding for Skylake Skylake changed the encoding of the PEBS data source field. Some combinations are not available anymore, but some new cases e.g. for L4 cache hit are added. Fix up the conversion table for Skylake, similar as had been done for Nehalem. On Skylake server the encoding for L4 actually means persistent memory. Handle this case too. To properly describe it in the abstracted perf format I had to add some new fields. Since a hit can have only one level add a new field that is an enumeration, not a bit field to describe the level. It can describe any level. Some numbers are also used to describe PMEM and LFB. Also add a new generic remote flag that can be combined with the generic level to signify a remote cache. And there is an extension field for the snoop indication to handle the Forward state. I didn't add a generic flag for hops because it's not needed for Skylake. I changed the existing encodings for older CPUs to also fill in the new level and remote fields. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: http://lkml.kernel.org/r/20170816222156.19953-3-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-16 15:21:54 -07:00
pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
}
static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source)
perf/x86: Fix data source decoding for Skylake Skylake changed the encoding of the PEBS data source field. Some combinations are not available anymore, but some new cases e.g. for L4 cache hit are added. Fix up the conversion table for Skylake, similar as had been done for Nehalem. On Skylake server the encoding for L4 actually means persistent memory. Handle this case too. To properly describe it in the abstracted perf format I had to add some new fields. Since a hit can have only one level add a new field that is an enumeration, not a bit field to describe the level. It can describe any level. Some numbers are also used to describe PMEM and LFB. Also add a new generic remote flag that can be combined with the generic level to signify a remote cache. And there is an extension field for the snoop indication to handle the Forward state. I didn't add a generic flag for hops because it's not needed for Skylake. I changed the existing encodings for older CPUs to also fill in the new level and remote fields. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: http://lkml.kernel.org/r/20170816222156.19953-3-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-16 15:21:54 -07:00
{
u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
}
void __init intel_pmu_pebs_data_source_skl(bool pmem)
{
__intel_pmu_pebs_data_source_skl(pmem, pebs_data_source);
}
2022-08-31 07:27:02 -07:00
static void __init __intel_pmu_pebs_data_source_grt(u64 *data_source)
{
data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
}
2022-08-31 07:27:02 -07:00
void __init intel_pmu_pebs_data_source_grt(void)
{
__intel_pmu_pebs_data_source_grt(pebs_data_source);
}
void __init intel_pmu_pebs_data_source_adl(void)
{
u64 *data_source;
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
__intel_pmu_pebs_data_source_skl(false, data_source);
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
2022-08-31 07:27:02 -07:00
__intel_pmu_pebs_data_source_grt(data_source);
}
static void __init __intel_pmu_pebs_data_source_cmt(u64 *data_source)
{
data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
data_source[0x0a] = OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, NONE);
data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
data_source[0x0c] = OP_LH | LEVEL(RAM) | REM | P(SNOOPX, FWD);
data_source[0x0d] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, HITM);
}
void __init intel_pmu_pebs_data_source_mtl(void)
{
u64 *data_source;
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
__intel_pmu_pebs_data_source_skl(false, data_source);
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
__intel_pmu_pebs_data_source_cmt(data_source);
}
void __init intel_pmu_pebs_data_source_cmt(void)
{
__intel_pmu_pebs_data_source_cmt(pebs_data_source);
}
/* Version for Lion Cove and later */
static u64 lnc_pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA), /* 0x00: ukn L3 */
OP_LH | P(LVL, L1) | LEVEL(L1) | P(SNOOP, NONE), /* 0x01: L1 hit */
OP_LH | P(LVL, L1) | LEVEL(L1) | P(SNOOP, NONE), /* 0x02: L1 hit */
OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x03: LFB/L1 Miss Handling Buffer hit */
0, /* 0x04: Reserved */
OP_LH | P(LVL, L2) | LEVEL(L2) | P(SNOOP, NONE), /* 0x05: L2 Hit */
OP_LH | LEVEL(L2_MHB) | P(SNOOP, NONE), /* 0x06: L2 Miss Handling Buffer Hit */
0, /* 0x07: Reserved */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, NONE), /* 0x08: L3 Hit */
0, /* 0x09: Reserved */
0, /* 0x0a: Reserved */
0, /* 0x0b: Reserved */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD), /* 0x0c: L3 Hit Snoop Fwd */
OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM), /* 0x0d: L3 Hit Snoop HitM */
0, /* 0x0e: Reserved */
P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM), /* 0x0f: L3 Miss Snoop HitM */
OP_LH | LEVEL(MSC) | P(SNOOP, NONE), /* 0x10: Memory-side Cache Hit */
OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, NONE), /* 0x11: Local Memory Hit */
};
void __init intel_pmu_pebs_data_source_lnl(void)
{
u64 *data_source;
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
memcpy(data_source, lnc_pebs_data_source, sizeof(lnc_pebs_data_source));
data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
__intel_pmu_pebs_data_source_cmt(data_source);
}
static u64 precise_store_data(u64 status)
{
union intel_x86_pebs_dse dse;
u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
dse.val = status;
/*
* bit 4: TLB access
* 1 = stored missed 2nd level TLB
*
* so it either hit the walker or the OS
* otherwise hit 2nd level TLB
*/
if (dse.st_stlb_miss)
val |= P(TLB, MISS);
else
val |= P(TLB, HIT);
/*
* bit 0: hit L1 data cache
* if not set, then all we know is that
* it missed L1D
*/
if (dse.st_l1d_hit)
val |= P(LVL, HIT);
else
val |= P(LVL, MISS);
/*
* bit 5: Locked prefix
*/
if (dse.st_locked)
val |= P(LOCK, LOCKED);
return val;
}
static u64 precise_datala_hsw(struct perf_event *event, u64 status)
{
union perf_mem_data_src dse;
dse.val = PERF_MEM_NA;
if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
dse.mem_op = PERF_MEM_OP_STORE;
else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
dse.mem_op = PERF_MEM_OP_LOAD;
/*
* L1 info only valid for following events:
*
* MEM_UOPS_RETIRED.STLB_MISS_STORES
* MEM_UOPS_RETIRED.LOCK_STORES
* MEM_UOPS_RETIRED.SPLIT_STORES
* MEM_UOPS_RETIRED.ALL_STORES
*/
if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
if (status & 1)
dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
else
dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
}
return dse.val;
}
static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock)
{
/*
* TLB access
* 0 = did not miss 2nd level TLB
* 1 = missed 2nd level TLB
*/
if (tlb)
*val |= P(TLB, MISS) | P(TLB, L2);
else
*val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
/* locked prefix */
if (lock)
*val |= P(LOCK, LOCKED);
}
/* Retrieve the latency data for e-core of ADL */
static u64 __grt_latency_data(struct perf_event *event, u64 status,
u8 dse, bool tlb, bool lock, bool blk)
{
u64 val;
perf/x86/intel: Clean up the hybrid CPU type handling code There is a fairly long list of grievances about the current code. The main beefs: 1. hybrid_big_small assumes that the *HARDWARE* (CPUID) provided core types are a bitmap. They are not. If Intel happened to make a core type of 0xff, hilarity would ensue. 2. adl_get_hybrid_cpu_type() utterly inscrutable. There are precisely zero comments and zero changelog about what it is attempting to do. According to Kan, the adl_get_hybrid_cpu_type() is there because some Alder Lake (ADL) CPUs can do some silly things. Some ADL models are *supposed* to be hybrid CPUs with big and little cores, but there are some SKUs that only have big cores. CPUID(0x1a) on those CPUs does not say that the CPUs are big cores. It apparently just returns 0x0. It confuses perf because it expects to see either 0x40 (Core) or 0x20 (Atom). The perf workaround for this is to watch for a CPU core saying it is type 0x0. If that happens on an Alder Lake, it calls x86_pmu.get_hybrid_cpu_type() and just assumes that the core is a Core (0x40) CPU. To fix up the mess, separate out the CPU types and the 'pmu' types. This allows 'hybrid_pmu_type' bitmaps without worrying that some future CPU type will set multiple bits. Since the types are now separate, add a function to glue them back together again. Actual comment on the situation in the glue function (find_hybrid_pmu_for_cpu()). Also, give ->get_hybrid_cpu_type() a real return type and make it clear that it is overriding the *CPU* type, not the PMU type. Rename cpu_type to pmu_type in the struct x86_hybrid_pmu to reflect the change. Originally-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lore.kernel.org/r/20230829125806.3016082-6-kan.liang@linux.intel.com
2023-08-29 05:58:05 -07:00
WARN_ON_ONCE(hybrid_pmu(event->pmu)->pmu_type == hybrid_big);
dse &= PERF_PEBS_DATA_SOURCE_GRT_MASK;
val = hybrid_var(event->pmu, pebs_data_source)[dse];
pebs_set_tlb_lock(&val, tlb, lock);
if (blk)
val |= P(BLK, DATA);
else
val |= P(BLK, NA);
return val;
}
u64 grt_latency_data(struct perf_event *event, u64 status)
{
union intel_x86_pebs_dse dse;
dse.val = status;
return __grt_latency_data(event, status, dse.ld_dse,
dse.ld_locked, dse.ld_stlb_miss,
dse.ld_data_blk);
}
/* Retrieve the latency data for e-core of MTL */
u64 cmt_latency_data(struct perf_event *event, u64 status)
{
union intel_x86_pebs_dse dse;
dse.val = status;
return __grt_latency_data(event, status, dse.mtl_dse,
dse.mtl_stlb_miss, dse.mtl_locked,
dse.mtl_fwd_blk);
}
static u64 lnc_latency_data(struct perf_event *event, u64 status)
{
union intel_x86_pebs_dse dse;
union perf_mem_data_src src;
u64 val;
dse.val = status;
/* LNC core latency data */
val = hybrid_var(event->pmu, pebs_data_source)[status & PERF_PEBS_DATA_SOURCE_MASK];
if (!val)
val = P(OP, LOAD) | LEVEL(NA) | P(SNOOP, NA);
if (dse.lnc_stlb_miss)
val |= P(TLB, MISS) | P(TLB, L2);
else
val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
if (dse.lnc_locked)
val |= P(LOCK, LOCKED);
if (dse.lnc_data_blk)
val |= P(BLK, DATA);
if (dse.lnc_addr_blk)
val |= P(BLK, ADDR);
if (!dse.lnc_data_blk && !dse.lnc_addr_blk)
val |= P(BLK, NA);
src.val = val;
if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
src.mem_op = P(OP, STORE);
return src.val;
}
u64 lnl_latency_data(struct perf_event *event, u64 status)
{
struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
if (pmu->pmu_type == hybrid_small)
return cmt_latency_data(event, status);
return lnc_latency_data(event, status);
}
static u64 load_latency_data(struct perf_event *event, u64 status)
{
union intel_x86_pebs_dse dse;
u64 val;
dse.val = status;
/*
* use the mapping table for bit 0-3
*/
val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
/*
* Nehalem models do not support TLB, Lock infos
*/
if (x86_pmu.pebs_no_tlb) {
val |= P(TLB, NA) | P(LOCK, NA);
return val;
}
pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked);
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
/*
* Ice Lake and earlier models do not support block infos.
*/
if (!x86_pmu.pebs_block) {
val |= P(BLK, NA);
return val;
}
/*
* bit 6: load was blocked since its data could not be forwarded
* from a preceding store
*/
if (dse.ld_data_blk)
val |= P(BLK, DATA);
/*
* bit 7: load was blocked due to potential address conflict with
* a preceding store
*/
if (dse.ld_addr_blk)
val |= P(BLK, ADDR);
if (!dse.ld_data_blk && !dse.ld_addr_blk)
val |= P(BLK, NA);
return val;
}
static u64 store_latency_data(struct perf_event *event, u64 status)
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
{
union intel_x86_pebs_dse dse;
union perf_mem_data_src src;
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
u64 val;
dse.val = status;
/*
* use the mapping table for bit 0-3
*/
val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse];
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked);
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
val |= P(BLK, NA);
/*
* the pebs_data_source table is only for loads
* so override the mem_op to say STORE instead
*/
src.val = val;
src.mem_op = P(OP,STORE);
return src.val;
}
struct pebs_record_core {
u64 flags, ip;
u64 ax, bx, cx, dx;
u64 si, di, bp, sp;
u64 r8, r9, r10, r11;
u64 r12, r13, r14, r15;
};
struct pebs_record_nhm {
u64 flags, ip;
u64 ax, bx, cx, dx;
u64 si, di, bp, sp;
u64 r8, r9, r10, r11;
u64 r12, r13, r14, r15;
u64 status, dla, dse, lat;
};
/*
* Same as pebs_record_nhm, with two additional fields.
*/
struct pebs_record_hsw {
u64 flags, ip;
u64 ax, bx, cx, dx;
u64 si, di, bp, sp;
u64 r8, r9, r10, r11;
u64 r12, r13, r14, r15;
u64 status, dla, dse, lat;
u64 real_ip, tsx_tuning;
};
union hsw_tsx_tuning {
struct {
u32 cycles_last_block : 32,
hle_abort : 1,
rtm_abort : 1,
instruction_abort : 1,
non_instruction_abort : 1,
retry : 1,
data_conflict : 1,
capacity_writes : 1,
capacity_reads : 1;
};
u64 value;
};
#define PEBS_HSW_TSX_FLAGS 0xff00000000ULL
/* Same as HSW, plus TSC */
struct pebs_record_skl {
u64 flags, ip;
u64 ax, bx, cx, dx;
u64 si, di, bp, sp;
u64 r8, r9, r10, r11;
u64 r12, r13, r14, r15;
u64 status, dla, dse, lat;
u64 real_ip, tsx_tuning;
u64 tsc;
};
void init_debug_store_on_cpu(int cpu)
{
struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
if (!ds)
return;
wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
(u32)((u64)(unsigned long)ds),
(u32)((u64)(unsigned long)ds >> 32));
}
void fini_debug_store_on_cpu(int cpu)
{
if (!per_cpu(cpu_hw_events, cpu).ds)
return;
wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
}
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
static DEFINE_PER_CPU(void *, insn_buffer);
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
{
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
unsigned long start = (unsigned long)cea;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
phys_addr_t pa;
size_t msz = 0;
pa = virt_to_phys(addr);
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
preempt_disable();
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
cea_set_pte(cea, pa, prot);
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
/*
* This is a cross-CPU update of the cpu_entry_area, we must shoot down
* all TLB entries for it.
*/
flush_tlb_kernel_range(start, start + size);
preempt_enable();
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
}
static void ds_clear_cea(void *cea, size_t size)
{
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
unsigned long start = (unsigned long)cea;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
size_t msz = 0;
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
preempt_disable();
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
cea_set_pte(cea, 0, PAGE_NONE);
x86/events/intel/ds: Use the proper cache flush method for mapping ds buffers Thomas reported the following warning: BUG: using smp_processor_id() in preemptible [00000000] code: ovsdb-server/4498 caller is native_flush_tlb_single+0x57/0xc0 native_flush_tlb_single+0x57/0xc0 __set_pte_vaddr+0x2d/0x40 set_pte_vaddr+0x2f/0x40 cea_set_pte+0x30/0x40 ds_update_cea.constprop.4+0x4d/0x70 reserve_ds_buffers+0x159/0x410 x86_reserve_hardware+0x150/0x160 x86_pmu_event_init+0x3e/0x1f0 perf_try_init_event+0x69/0x80 perf_event_alloc+0x652/0x740 SyS_perf_event_open+0x3f6/0xd60 do_syscall_64+0x5c/0x190 set_pte_vaddr is used to map the ds buffers into the cpu entry area, but there are two problems with that: 1) The resulting flush is not supposed to be called in preemptible context 2) The cpu entry area is supposed to be per CPU, but the debug store buffers are mapped for all CPUs so these mappings need to be flushed globally. Add the necessary preemption protection across the mapping code and flush TLBs globally. Fixes: c1961a4631da ("x86/events/intel/ds: Map debug buffers in cpu_entry_area") Reported-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org Link: https://lkml.kernel.org/r/20180104170712.GB3040@hirez.programming.kicks-ass.net
2018-01-04 10:07:12 -07:00
flush_tlb_kernel_range(start, start + size);
preempt_enable();
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
}
static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
{
unsigned int order = get_order(size);
int node = cpu_to_node(cpu);
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
struct page *page;
page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
return page ? page_address(page) : NULL;
}
static void dsfree_pages(const void *buffer, size_t size)
{
if (buffer)
free_pages((unsigned long)buffer, get_order(size));
}
static int alloc_pebs_buffer(int cpu)
{
struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
struct debug_store *ds = hwev->ds;
size_t bsiz = x86_pmu.pebs_buffer_size;
int max, node = cpu_to_node(cpu);
void *buffer, *insn_buff, *cea;
if (!x86_pmu.pebs)
return 0;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
if (unlikely(!buffer))
return -ENOMEM;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
/*
* HSW+ already provides us the eventing ip; no need to allocate this
* buffer then.
*/
if (x86_pmu.intel_cap.pebs_format < 2) {
insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
if (!insn_buff) {
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
dsfree_pages(buffer, bsiz);
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
return -ENOMEM;
}
per_cpu(insn_buffer, cpu) = insn_buff;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
}
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
hwev->ds_pebs_vaddr = buffer;
/* Update the cpu entry area mapping */
cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
ds->pebs_buffer_base = (unsigned long) cea;
ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
ds->pebs_index = ds->pebs_buffer_base;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
return 0;
}
static void release_pebs_buffer(int cpu)
{
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
void *cea;
if (!x86_pmu.pebs)
return;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
kfree(per_cpu(insn_buffer, cpu));
per_cpu(insn_buffer, cpu) = NULL;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
/* Clear the fixmap */
cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
hwev->ds_pebs_vaddr = NULL;
}
static int alloc_bts_buffer(int cpu)
{
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
struct debug_store *ds = hwev->ds;
void *buffer, *cea;
int max;
if (!x86_pmu.bts)
return 0;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
if (unlikely(!buffer)) {
WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
return -ENOMEM;
}
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
hwev->ds_bts_vaddr = buffer;
/* Update the fixmap */
cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
ds->bts_buffer_base = (unsigned long) cea;
ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
ds->bts_index = ds->bts_buffer_base;
max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
ds->bts_absolute_maximum = ds->bts_buffer_base +
max * BTS_RECORD_SIZE;
ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
(max / 16) * BTS_RECORD_SIZE;
return 0;
}
static void release_bts_buffer(int cpu)
{
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
void *cea;
if (!x86_pmu.bts)
return;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
/* Clear the fixmap */
cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
ds_clear_cea(cea, BTS_BUFFER_SIZE);
dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
hwev->ds_bts_vaddr = NULL;
}
static int alloc_ds_buffer(int cpu)
{
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
x86/events/intel/ds: Map debug buffers in cpu_entry_area The BTS and PEBS buffers both have their virtual addresses programmed into the hardware. This means that any access to them is performed via the page tables. The times that the hardware accesses these are entirely dependent on how the performance monitoring hardware events are set up. In other words, there is no way for the kernel to tell when the hardware might access these buffers. To avoid perf crashes, place 'debug_store' allocate pages and map them into the cpu_entry_area. The PEBS fixup buffer does not need this treatment. [ tglx: Got rid of the kaiser_add_mapping() complication ] Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 07:07:50 -07:00
memset(ds, 0, sizeof(*ds));
per_cpu(cpu_hw_events, cpu).ds = ds;
return 0;
}
static void release_ds_buffer(int cpu)
{
per_cpu(cpu_hw_events, cpu).ds = NULL;
}
void release_ds_buffers(void)
{
int cpu;
if (!x86_pmu.bts && !x86_pmu.pebs)
return;
for_each_possible_cpu(cpu)
release_ds_buffer(cpu);
for_each_possible_cpu(cpu) {
/*
* Again, ignore errors from offline CPUs, they will no longer
* observe cpu_hw_events.ds and not program the DS_AREA when
* they come up.
*/
fini_debug_store_on_cpu(cpu);
}
for_each_possible_cpu(cpu) {
release_pebs_buffer(cpu);
release_bts_buffer(cpu);
}
}
void reserve_ds_buffers(void)
{
int bts_err = 0, pebs_err = 0;
int cpu;
x86_pmu.bts_active = 0;
x86_pmu.pebs_active = 0;
if (!x86_pmu.bts && !x86_pmu.pebs)
return;
if (!x86_pmu.bts)
bts_err = 1;
if (!x86_pmu.pebs)
pebs_err = 1;
for_each_possible_cpu(cpu) {
if (alloc_ds_buffer(cpu)) {
bts_err = 1;
pebs_err = 1;
}
if (!bts_err && alloc_bts_buffer(cpu))
bts_err = 1;
if (!pebs_err && alloc_pebs_buffer(cpu))
pebs_err = 1;
if (bts_err && pebs_err)
break;
}
if (bts_err) {
for_each_possible_cpu(cpu)
release_bts_buffer(cpu);
}
if (pebs_err) {
for_each_possible_cpu(cpu)
release_pebs_buffer(cpu);
}
if (bts_err && pebs_err) {
for_each_possible_cpu(cpu)
release_ds_buffer(cpu);
} else {
if (x86_pmu.bts && !bts_err)
x86_pmu.bts_active = 1;
if (x86_pmu.pebs && !pebs_err)
x86_pmu.pebs_active = 1;
for_each_possible_cpu(cpu) {
/*
* Ignores wrmsr_on_cpu() errors for offline CPUs they
* will get this call through intel_pmu_cpu_starting().
*/
init_debug_store_on_cpu(cpu);
}
}
}
/*
* BTS
*/
struct event_constraint bts_constraint =
EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
void intel_pmu_enable_bts(u64 config)
{
unsigned long debugctlmsr;
debugctlmsr = get_debugctlmsr();
debugctlmsr |= DEBUGCTLMSR_TR;
debugctlmsr |= DEBUGCTLMSR_BTS;
if (config & ARCH_PERFMON_EVENTSEL_INT)
debugctlmsr |= DEBUGCTLMSR_BTINT;
if (!(config & ARCH_PERFMON_EVENTSEL_OS))
debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
if (!(config & ARCH_PERFMON_EVENTSEL_USR))
debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
update_debugctlmsr(debugctlmsr);
}
void intel_pmu_disable_bts(void)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
unsigned long debugctlmsr;
if (!cpuc->ds)
return;
debugctlmsr = get_debugctlmsr();
debugctlmsr &=
~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
update_debugctlmsr(debugctlmsr);
}
int intel_pmu_drain_bts_buffer(void)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct debug_store *ds = cpuc->ds;
struct bts_record {
u64 from;
u64 to;
u64 flags;
};
struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
struct bts_record *at, *base, *top;
struct perf_output_handle handle;
struct perf_event_header header;
struct perf_sample_data data;
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
unsigned long skip = 0;
struct pt_regs regs;
if (!event)
return 0;
if (!x86_pmu.bts_active)
return 0;
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
top = (struct bts_record *)(unsigned long)ds->bts_index;
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
if (top <= base)
return 0;
memset(&regs, 0, sizeof(regs));
ds->bts_index = ds->bts_buffer_base;
perf_sample_data_init(&data, 0, event->hw.last_period);
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
/*
* BTS leaks kernel addresses in branches across the cpl boundary,
* such as traps or system calls, so unless the user is asking for
* kernel tracing (and right now it's not possible), we'd need to
* filter them out. But first we need to count how many of those we
* have in the current batch. This is an extra O(n) pass, however,
* it's much faster than the other one especially considering that
* n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
* alloc_bts_buffer()).
*/
for (at = base; at < top; at++) {
/*
* Note that right now *this* BTS code only works if
* attr::exclude_kernel is set, but let's keep this extra
* check here in case that changes.
*/
if (event->attr.exclude_kernel &&
(kernel_ip(at->from) || kernel_ip(at->to)))
skip++;
}
/*
* Prepare a generic sample, i.e. fill in the invariant fields.
* We will overwrite the from and to address before we output
* the sample.
*/
rcu_read_lock();
perf_prepare_sample(&data, event, &regs);
perf_prepare_header(&header, &data, event, &regs);
if (perf_output_begin(&handle, &data, event,
header.size * (top - base - skip)))
goto unlock;
perf/x86/intel/ds: Work around BTS leaking kernel addresses BTS leaks kernel addresses even in userspace-only mode due to imprecise IP sampling, so sometimes syscall entry points or page fault handler addresses end up in a userspace trace. Since this driver uses a relatively small buffer for BTS records and it has to iterate through them anyway, it can also take on the additional job of filtering out the records that contain kernel addresses when kernel space tracing is not enabled. This patch changes the bts code to skip the offending records from perf output. In order to request the exact amount of space on the ring buffer, we need to do an extra pass through the records to know how many there are of the valid ones, but considering the small size of the buffer, this extra pass adds very little overhead to the nmi handler. This way we won't end up with awkward IP samples with zero IPs in the perf stream. Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1441030168-6853-2-git-send-email-alexander.shishkin@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-08-31 07:09:27 -07:00
for (at = base; at < top; at++) {
/* Filter out any records that contain kernel addresses. */
if (event->attr.exclude_kernel &&
(kernel_ip(at->from) || kernel_ip(at->to)))
continue;
data.ip = at->from;
data.addr = at->to;
perf_output_sample(&handle, &header, &data, event);
}
perf_output_end(&handle);
/* There's new data available. */
event->hw.interrupts++;
event->pending_kill = POLL_IN;
unlock:
rcu_read_unlock();
return 1;
}
static inline void intel_pmu_drain_pebs_buffer(void)
{
struct perf_sample_data data;
x86_pmu.drain_pebs(NULL, &data);
}
/*
* PEBS
*/
struct event_constraint intel_core2_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_atom_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_slm_pebs_event_constraints[] = {
/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_glm_pebs_event_constraints[] = {
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_grt_pebs_event_constraints[] = {
/* Allow all events as PEBS with no flags */
INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0x3),
INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_nehalem_pebs_event_constraints[] = {
INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_westmere_pebs_event_constraints[] = {
INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_snb_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_ivb_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
perf/x86: Use INST_RETIRED.PREC_DIST for cycles: ppp Add a new 'three-p' precise level, that uses INST_RETIRED.PREC_DIST as base. The basic mechanism of abusing the inverse cmask to get all cycles works the same as before. PREC_DIST is available on Sandy Bridge or later. It had some problems on Sandy Bridge, so we only use it on IvyBridge and later. I tested it on Broadwell and Skylake. PREC_DIST has special support for avoiding shadow effects, which can give better results compare to UOPS_RETIRED. The drawback is that PREC_DIST can only schedule on counter 1, but that is ok for cycle sampling, as there is normally no need to do multiple cycle sampling runs in parallel. It is still possible to run perf top in parallel, as that doesn't use precise mode. Also of course the multiplexing can still allow parallel operation. :pp stays with the previous event. Example: Sample a loop with 10 sqrt with old cycles:pp 0.14 │10: sqrtps %xmm1,%xmm0 <-------------- 9.13 │ sqrtps %xmm1,%xmm0 11.58 │ sqrtps %xmm1,%xmm0 11.51 │ sqrtps %xmm1,%xmm0 6.27 │ sqrtps %xmm1,%xmm0 10.38 │ sqrtps %xmm1,%xmm0 12.20 │ sqrtps %xmm1,%xmm0 12.74 │ sqrtps %xmm1,%xmm0 5.40 │ sqrtps %xmm1,%xmm0 10.14 │ sqrtps %xmm1,%xmm0 10.51 │ ↑ jmp 10 We expect all 10 sqrt to get roughly the sample number of samples. But you can see that the instruction directly after the JMP is systematically underestimated in the result, due to sampling shadow effects. With the new PREC_DIST based sampling this problem is gone and all instructions show up roughly evenly: 9.51 │10: sqrtps %xmm1,%xmm0 11.74 │ sqrtps %xmm1,%xmm0 11.84 │ sqrtps %xmm1,%xmm0 6.05 │ sqrtps %xmm1,%xmm0 10.46 │ sqrtps %xmm1,%xmm0 12.25 │ sqrtps %xmm1,%xmm0 12.18 │ sqrtps %xmm1,%xmm0 5.26 │ sqrtps %xmm1,%xmm0 10.13 │ sqrtps %xmm1,%xmm0 10.43 │ sqrtps %xmm1,%xmm0 0.16 │ ↑ jmp 10 Even with PREC_DIST there is still sampling skid and the result is not completely even, but systematic shadow effects are significantly reduced. The improvements are mainly expected to make a difference in high IPC code. With low IPC it should be similar. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1448929689-13771-2-git-send-email-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-12-04 04:50:52 -07:00
/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_hsw_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
perf/x86: Use INST_RETIRED.PREC_DIST for cycles: ppp Add a new 'three-p' precise level, that uses INST_RETIRED.PREC_DIST as base. The basic mechanism of abusing the inverse cmask to get all cycles works the same as before. PREC_DIST is available on Sandy Bridge or later. It had some problems on Sandy Bridge, so we only use it on IvyBridge and later. I tested it on Broadwell and Skylake. PREC_DIST has special support for avoiding shadow effects, which can give better results compare to UOPS_RETIRED. The drawback is that PREC_DIST can only schedule on counter 1, but that is ok for cycle sampling, as there is normally no need to do multiple cycle sampling runs in parallel. It is still possible to run perf top in parallel, as that doesn't use precise mode. Also of course the multiplexing can still allow parallel operation. :pp stays with the previous event. Example: Sample a loop with 10 sqrt with old cycles:pp 0.14 │10: sqrtps %xmm1,%xmm0 <-------------- 9.13 │ sqrtps %xmm1,%xmm0 11.58 │ sqrtps %xmm1,%xmm0 11.51 │ sqrtps %xmm1,%xmm0 6.27 │ sqrtps %xmm1,%xmm0 10.38 │ sqrtps %xmm1,%xmm0 12.20 │ sqrtps %xmm1,%xmm0 12.74 │ sqrtps %xmm1,%xmm0 5.40 │ sqrtps %xmm1,%xmm0 10.14 │ sqrtps %xmm1,%xmm0 10.51 │ ↑ jmp 10 We expect all 10 sqrt to get roughly the sample number of samples. But you can see that the instruction directly after the JMP is systematically underestimated in the result, due to sampling shadow effects. With the new PREC_DIST based sampling this problem is gone and all instructions show up roughly evenly: 9.51 │10: sqrtps %xmm1,%xmm0 11.74 │ sqrtps %xmm1,%xmm0 11.84 │ sqrtps %xmm1,%xmm0 6.05 │ sqrtps %xmm1,%xmm0 10.46 │ sqrtps %xmm1,%xmm0 12.25 │ sqrtps %xmm1,%xmm0 12.18 │ sqrtps %xmm1,%xmm0 5.26 │ sqrtps %xmm1,%xmm0 10.13 │ sqrtps %xmm1,%xmm0 10.43 │ sqrtps %xmm1,%xmm0 0.16 │ ↑ jmp 10 Even with PREC_DIST there is still sampling skid and the result is not completely even, but systematic shadow effects are significantly reduced. The improvements are mainly expected to make a difference in high IPC code. With low IPC it should be similar. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1448929689-13771-2-git-send-email-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-12-04 04:50:52 -07:00
/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_bdw_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
EVENT_CONSTRAINT_END
};
struct event_constraint intel_skl_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */
perf/x86: Use INST_RETIRED.PREC_DIST for cycles: ppp Add a new 'three-p' precise level, that uses INST_RETIRED.PREC_DIST as base. The basic mechanism of abusing the inverse cmask to get all cycles works the same as before. PREC_DIST is available on Sandy Bridge or later. It had some problems on Sandy Bridge, so we only use it on IvyBridge and later. I tested it on Broadwell and Skylake. PREC_DIST has special support for avoiding shadow effects, which can give better results compare to UOPS_RETIRED. The drawback is that PREC_DIST can only schedule on counter 1, but that is ok for cycle sampling, as there is normally no need to do multiple cycle sampling runs in parallel. It is still possible to run perf top in parallel, as that doesn't use precise mode. Also of course the multiplexing can still allow parallel operation. :pp stays with the previous event. Example: Sample a loop with 10 sqrt with old cycles:pp 0.14 │10: sqrtps %xmm1,%xmm0 <-------------- 9.13 │ sqrtps %xmm1,%xmm0 11.58 │ sqrtps %xmm1,%xmm0 11.51 │ sqrtps %xmm1,%xmm0 6.27 │ sqrtps %xmm1,%xmm0 10.38 │ sqrtps %xmm1,%xmm0 12.20 │ sqrtps %xmm1,%xmm0 12.74 │ sqrtps %xmm1,%xmm0 5.40 │ sqrtps %xmm1,%xmm0 10.14 │ sqrtps %xmm1,%xmm0 10.51 │ ↑ jmp 10 We expect all 10 sqrt to get roughly the sample number of samples. But you can see that the instruction directly after the JMP is systematically underestimated in the result, due to sampling shadow effects. With the new PREC_DIST based sampling this problem is gone and all instructions show up roughly evenly: 9.51 │10: sqrtps %xmm1,%xmm0 11.74 │ sqrtps %xmm1,%xmm0 11.84 │ sqrtps %xmm1,%xmm0 6.05 │ sqrtps %xmm1,%xmm0 10.46 │ sqrtps %xmm1,%xmm0 12.25 │ sqrtps %xmm1,%xmm0 12.18 │ sqrtps %xmm1,%xmm0 5.26 │ sqrtps %xmm1,%xmm0 10.13 │ sqrtps %xmm1,%xmm0 10.43 │ sqrtps %xmm1,%xmm0 0.16 │ ↑ jmp 10 Even with PREC_DIST there is still sampling skid and the result is not completely even, but systematic shadow effects are significantly reduced. The improvements are mainly expected to make a difference in high IPC code. With low IPC it should be similar. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1448929689-13771-2-git-send-email-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-12-04 04:50:52 -07:00
/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
perf/x86: Use INST_RETIRED.TOTAL_CYCLES_PS for cycles:pp for Skylake I added UOPS_RETIRED.ALL by mistake to the Skylake PEBS event list for cycles:pp. But the event is not documented for Skylake, and has some issues. The recommended replacement for cycles:pp is to use INST_RETIRED.ANY+pebs as a base, similar to what CPUs before Sandy Bridge did. This new event is called INST_RETIRED.TOTAL_CYCLES_PS. The event is not really new, but has been already used by perf before Sandy Bridge for the original cycles:p Note the SDM doesn't document that event either, but it's being documented in the latest version of the event list on: https://download.01.org/perfmon/SKL This patch does: - Remove UOPS_RETIRED.ALL from the Skylake PEBS event list - Add INST_RETIRED.ANY to the Skylake PEBS event list, and an table entry to allow cmask=16,inv=1 for cycles:pp - We don't need an extra entry for the base INST_RETIRED event, because it is already covered by the catch-all PEBS table entry. - Switch Skylake to use the Core2 PEBS alias (which is INST_RETIRED.TOTAL_CYCLES_PS) Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: hpa@zytor.com Link: http://lkml.kernel.org/r/1448929689-13771-1-git-send-email-andi@firstfloor.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-12-04 04:50:32 -07:00
/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
INTEL_PLD_CONSTRAINT(0x1cd, 0xf), /* MEM_TRANS_RETIRED.* */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_L3_MISS_RETIRED.* */
perf/x86: Revamp PEBS event selection The basic idea is that it does not make sense to list all PEBS events individually. The list is very long, sometimes outdated and the hardware doesn't need it. If an event does not support PEBS it will just not count, there is no security issue. We need to only list events that something special, like supporting load or store addresses. This vastly simplifies the PEBS event selection. It also speeds up the scheduling because the scheduler doesn't have to walk as many constraints. Bugs fixed: - We do not allow setting forbidden flags with PEBS anymore (SDM 18.9.4), except for the special cycle event. This is done using a new constraint macro that also matches on the event flags. - Correct DataLA and load/store/na flags reporting on Haswell [Requires a followon patch] - We did not allow all PEBS events on Haswell: We were missing some valid subevents in d1-d2 (MEM_LOAD_UOPS_RETIRED.*, MEM_LOAD_UOPS_RETIRED_L3_HIT_RETIRED.*) This includes the changes proposed by Stephane earlier and obsoletes his patchkit (except for some changes on pre Sandy Bridge/Silvermont CPUs) I only did Sandy Bridge and Silvermont and later so far, mostly because these are the parts I could directly confirm the hardware behavior with hardware architects. Also I do not believe the older CPUs have any missing events in their PEBS list, so there's no pressing need to change them. I did not implement the flag proposed by Peter to allow setting forbidden flags. If really needed this could be implemented on to of this patch. v2: Fix broken store events on SNB/IVB (Stephane Eranian) v3: More fixes. Rename some arguments (Stephane Eranian) v4: List most Haswell events individually again to report memory operation type correctly. Add new flags to describe load/store/na for datala. Update description. Signed-off-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1407785233-32193-2-git-send-email-eranian@google.com Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Mark Davies <junk@eslaf.co.uk> Cc: Paul Mackerras <paulus@samba.org> Cc: Stephane Eranian <eranian@google.com> Cc: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-08-11 12:27:10 -07:00
/* Allow all events as PEBS with no flags */
INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
EVENT_CONSTRAINT_END
};
perf/x86/intel: Add Icelake support Add Icelake core PMU perf code, including constraint tables and the main enable code. Icelake expanded the generic counters to always 8 even with HT on, but a range of events cannot be scheduled on the extra 4 counters. Add new constraint ranges to describe this to the scheduler. The number of constraints that need to be checked is larger now than with earlier CPUs. At some point we may need a new data structure to look them up more efficiently than with linear search. So far it still seems to be acceptable however. Icelake added a new fixed counter SLOTS. Full support for it is added later in the patch series. The cache events table is identical to Skylake. Compare to PEBS instruction event on generic counter, fixed counter 0 has less skid. Force instruction:ppp always in fixed counter 0. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-9-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:05 -07:00
struct event_constraint intel_icl_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL), /* old INST_RETIRED.PREC_DIST */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL), /* SLOTS */
perf/x86/intel: Add Icelake support Add Icelake core PMU perf code, including constraint tables and the main enable code. Icelake expanded the generic counters to always 8 even with HT on, but a range of events cannot be scheduled on the extra 4 counters. Add new constraint ranges to describe this to the scheduler. The number of constraints that need to be checked is larger now than with earlier CPUs. At some point we may need a new data structure to look them up more efficiently than with linear search. So far it still seems to be acceptable however. Icelake added a new fixed counter SLOTS. Full support for it is added later in the patch series. The cache events table is identical to Skylake. Compare to PEBS instruction event on generic counter, fixed counter 0 has less skid. Force instruction:ppp always in fixed counter 0. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-9-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:05 -07:00
INTEL_PLD_CONSTRAINT(0x1cd, 0xff), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
perf/x86/intel: Add Icelake support Add Icelake core PMU perf code, including constraint tables and the main enable code. Icelake expanded the generic counters to always 8 even with HT on, but a range of events cannot be scheduled on the extra 4 counters. Add new constraint ranges to describe this to the scheduler. The number of constraints that need to be checked is larger now than with earlier CPUs. At some point we may need a new data structure to look them up more efficiently than with linear search. So far it still seems to be acceptable however. Icelake added a new fixed counter SLOTS. Full support for it is added later in the patch series. The cache events table is identical to Skylake. Compare to PEBS instruction event on generic counter, fixed counter 0 has less skid. Force instruction:ppp always in fixed counter 0. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-9-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:05 -07:00
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
/*
* Everything else is handled by PMU_FL_PEBS_ALL, because we
* need the full constraints from the main table.
*/
EVENT_CONSTRAINT_END
};
struct event_constraint intel_glc_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
/*
* Everything else is handled by PMU_FL_PEBS_ALL, because we
* need the full constraints from the main table.
*/
EVENT_CONSTRAINT_END
};
struct event_constraint intel_lnc_pebs_event_constraints[] = {
INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */
INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
INTEL_HYBRID_LDLAT_CONSTRAINT(0x1cd, 0x3ff),
INTEL_HYBRID_STLAT_CONSTRAINT(0x2cd, 0x3),
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
/*
* Everything else is handled by PMU_FL_PEBS_ALL, because we
* need the full constraints from the main table.
*/
EVENT_CONSTRAINT_END
};
struct event_constraint *intel_pebs_constraints(struct perf_event *event)
{
struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
struct event_constraint *c;
if (!event->attr.precise_ip)
return NULL;
if (pebs_constraints) {
for_each_event_constraint(c, pebs_constraints) {
if (constraint_match(c, event->hw.config)) {
event->hw.flags |= c->flags;
return c;
}
}
}
/*
* Extended PEBS support
* Makes the PEBS code search the normal constraints.
*/
if (x86_pmu.flags & PMU_FL_PEBS_ALL)
return NULL;
return &emptyconstraint;
}
/*
* We need the sched_task callback even for per-cpu events when we use
* the large interrupt threshold, such that we can provide PID and TID
* to PEBS samples.
*/
static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
{
if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
return false;
return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
}
perf: Rewrite core context handling There have been various issues and limitations with the way perf uses (task) contexts to track events. Most notable is the single hardware PMU task context, which has resulted in a number of yucky things (both proposed and merged). Notably: - HW breakpoint PMU - ARM big.little PMU / Intel ADL PMU - Intel Branch Monitoring PMU - AMD IBS PMU - S390 cpum_cf PMU - PowerPC trace_imc PMU *Current design:* Currently we have a per task and per cpu perf_event_contexts: task_struct::perf_events_ctxp[] <-> perf_event_context <-> perf_cpu_context ^ | ^ | ^ `---------------------------------' | `--> pmu ---' v ^ perf_event ------' Each task has an array of pointers to a perf_event_context. Each perf_event_context has a direct relation to a PMU and a group of events for that PMU. The task related perf_event_context's have a pointer back to that task. Each PMU has a per-cpu pointer to a per-cpu perf_cpu_context, which includes a perf_event_context, which again has a direct relation to that PMU, and a group of events for that PMU. The perf_cpu_context also tracks which task context is currently associated with that CPU and includes a few other things like the hrtimer for rotation etc. Each perf_event is then associated with its PMU and one perf_event_context. *Proposed design:* New design proposed by this patch reduce to a single task context and a single CPU context but adds some intermediate data-structures: task_struct::perf_event_ctxp -> perf_event_context <- perf_cpu_context ^ | ^ ^ `---------------------------' | | | | perf_cpu_pmu_context <--. | `----. ^ | | | | | | v v | | ,--> perf_event_pmu_context | | | | | | | v v | perf_event ---> pmu ----------------' With the new design, perf_event_context will hold all events for all pmus in the (respective pinned/flexible) rbtrees. This can be achieved by adding pmu to rbtree key: {cpu, pmu, cgroup, group_index} Each perf_event_context carries a list of perf_event_pmu_context which is used to hold per-pmu-per-context state. For example, it keeps track of currently active events for that pmu, a pmu specific task_ctx_data, a flag to tell whether rotation is required or not etc. Additionally, perf_cpu_pmu_context is used to hold per-pmu-per-cpu state like hrtimer details to drive the event rotation, a pointer to perf_event_pmu_context of currently running task and some other ancillary information. Each perf_event is associated to it's pmu, perf_event_context and perf_event_pmu_context. Further optimizations to current implementation are possible. For example, ctx_resched() can be optimized to reschedule only single pmu events. Much thanks to Ravi for picking this up and pushing it towards completion. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Co-developed-by: Ravi Bangoria <ravi.bangoria@amd.com> Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20221008062424.313-1-ravi.bangoria@amd.com
2022-10-07 23:24:24 -07:00
void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
perf/x86/intel: Add proper condition to run sched_task callbacks We have 2 functions using the same sched_task callback: - PEBS drain for free running counters - LBR save/store Both of them are called from intel_pmu_sched_task() and either of them can be unwillingly triggered when the other one is configured to run. Let's say there's PEBS drain configured in sched_task callback for the event, but in the callback itself (intel_pmu_sched_task()) we will also run the code for LBR save/restore, which we did not ask for, but the code in intel_pmu_sched_task() does not check for that. This can lead to extra cycles in some perf monitoring, like when we monitor PEBS event without LBR data. # perf record --no-timestamp -c 10000 -e cycles:p ./perf bench sched pipe -l 1000000 (We need PEBS, non freq/non timestamp event to enable the sched_task callback) The perf stat of cycles and msr:write_msr for above command before the change: ... Performance counter stats for './perf record --no-timestamp -c 10000 -e cycles:p \ ./perf bench sched pipe -l 1000000' (5 runs): 18,519,557,441 cycles:k 91,195,527 msr:write_msr 29.334476406 seconds time elapsed And after the change: ... Performance counter stats for './perf record --no-timestamp -c 10000 -e cycles:p \ ./perf bench sched pipe -l 1000000' (5 runs): 18,704,973,540 cycles:k 27,184,720 msr:write_msr 16.977875900 seconds time elapsed There's no affect on cycles:k because the sched_task happens with events switched off, however the msr:write_msr tracepoint counter together with almost 50% of time speedup show the improvement. Monitoring LBR event and having extra PEBS drain processing in sched_task callback showed just a little speedup, because the drain function does not do much extra work in case there is no PEBS data. Adding conditions to recognize the configured work that needs to be done in the x86_pmu's sched_task callback. Suggested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Jiri Olsa <jolsa@kernel.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Kan Liang <kan.liang@intel.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Jiri Olsa <jolsa@kernel.org> Link: http://lkml.kernel.org/r/20170719075247.GA27506@krava Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-19 00:52:47 -07:00
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
if (!sched_in && pebs_needs_sched_cb(cpuc))
intel_pmu_drain_pebs_buffer();
}
static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
{
struct debug_store *ds = cpuc->ds;
int max_pebs_events = intel_pmu_max_num_pebs(cpuc->pmu);
u64 threshold;
int reserved;
if (cpuc->n_pebs_via_pt)
return;
if (x86_pmu.flags & PMU_FL_PEBS_ALL)
reserved = max_pebs_events + x86_pmu_max_num_counters_fixed(cpuc->pmu);
else
reserved = max_pebs_events;
if (cpuc->n_pebs == cpuc->n_large_pebs) {
threshold = ds->pebs_absolute_maximum -
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
reserved * cpuc->pebs_record_size;
} else {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
}
ds->pebs_interrupt_threshold = threshold;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
static void adaptive_pebs_record_size_update(void)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
u64 pebs_data_cfg = cpuc->pebs_data_cfg;
int sz = sizeof(struct pebs_basic);
if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
sz += sizeof(struct pebs_meminfo);
if (pebs_data_cfg & PEBS_DATACFG_GP)
sz += sizeof(struct pebs_gprs);
if (pebs_data_cfg & PEBS_DATACFG_XMMS)
sz += sizeof(struct pebs_xmm);
if (pebs_data_cfg & PEBS_DATACFG_LBRS)
sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
cpuc->pebs_record_size = sz;
}
#define PERF_PEBS_MEMINFO_TYPE (PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC | \
perf/core: Add PERF_SAMPLE_WEIGHT_STRUCT Current PERF_SAMPLE_WEIGHT sample type is very useful to expresses the cost of an action represented by the sample. This allows the profiler to scale the samples to be more informative to the programmer. It could also help to locate a hotspot, e.g., when profiling by memory latencies, the expensive load appear higher up in the histograms. But current PERF_SAMPLE_WEIGHT sample type is solely determined by one factor. This could be a problem, if users want two or more factors to contribute to the weight. For example, Golden Cove core PMU can provide both the instruction latency and the cache Latency information as factors for the memory profiling. For current X86 platforms, although meminfo::latency is defined as a u64, only the lower 32 bits include the valid data in practice (No memory access could last than 4G cycles). The higher 32 bits can be used to store new factors. Add a new sample type, PERF_SAMPLE_WEIGHT_STRUCT, to indicate the new sample weight structure. It shares the same space as the PERF_SAMPLE_WEIGHT sample type. Users can apply either the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type to retrieve the sample weight, but they cannot apply both sample types simultaneously. Currently, only X86 and PowerPC use the PERF_SAMPLE_WEIGHT sample type. - For PowerPC, there is nothing changed for the PERF_SAMPLE_WEIGHT sample type. There is no effect for the new PERF_SAMPLE_WEIGHT_STRUCT sample type. PowerPC can re-struct the weight field similarly later. - For X86, the same value will be dumped for the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type for now. The following patches will apply the new factors for the PERF_SAMPLE_WEIGHT_STRUCT sample type. The field in the union perf_sample_weight should be shared among different architectures. A generic name is required, but it's hard to abstract a name that applies to all architectures. For example, on X86, the fields are to store all kinds of latency. While on PowerPC, it stores MMCRA[TECX/TECM], which should not be latency. So a general name prefix 'var$NUM' is used here. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-2-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:07 -07:00
PERF_SAMPLE_PHYS_ADDR | \
PERF_SAMPLE_WEIGHT_TYPE | \
PERF_SAMPLE_TRANSACTION | \
PERF_SAMPLE_DATA_PAGE_SIZE)
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
static u64 pebs_update_adaptive_cfg(struct perf_event *event)
{
struct perf_event_attr *attr = &event->attr;
u64 sample_type = attr->sample_type;
u64 pebs_data_cfg = 0;
bool gprs, tsx_weight;
if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
attr->precise_ip > 1)
return pebs_data_cfg;
if (sample_type & PERF_PEBS_MEMINFO_TYPE)
pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
/*
* We need GPRs when:
* + user requested them
* + precise_ip < 2 for the non event IP
* + For RTM TSX weight we need GPRs for the abort code.
*/
gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
(attr->sample_regs_intr & PEBS_GP_REGS);
perf/core: Add PERF_SAMPLE_WEIGHT_STRUCT Current PERF_SAMPLE_WEIGHT sample type is very useful to expresses the cost of an action represented by the sample. This allows the profiler to scale the samples to be more informative to the programmer. It could also help to locate a hotspot, e.g., when profiling by memory latencies, the expensive load appear higher up in the histograms. But current PERF_SAMPLE_WEIGHT sample type is solely determined by one factor. This could be a problem, if users want two or more factors to contribute to the weight. For example, Golden Cove core PMU can provide both the instruction latency and the cache Latency information as factors for the memory profiling. For current X86 platforms, although meminfo::latency is defined as a u64, only the lower 32 bits include the valid data in practice (No memory access could last than 4G cycles). The higher 32 bits can be used to store new factors. Add a new sample type, PERF_SAMPLE_WEIGHT_STRUCT, to indicate the new sample weight structure. It shares the same space as the PERF_SAMPLE_WEIGHT sample type. Users can apply either the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type to retrieve the sample weight, but they cannot apply both sample types simultaneously. Currently, only X86 and PowerPC use the PERF_SAMPLE_WEIGHT sample type. - For PowerPC, there is nothing changed for the PERF_SAMPLE_WEIGHT sample type. There is no effect for the new PERF_SAMPLE_WEIGHT_STRUCT sample type. PowerPC can re-struct the weight field similarly later. - For X86, the same value will be dumped for the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type for now. The following patches will apply the new factors for the PERF_SAMPLE_WEIGHT_STRUCT sample type. The field in the union perf_sample_weight should be shared among different architectures. A generic name is required, but it's hard to abstract a name that applies to all architectures. For example, on X86, the fields are to store all kinds of latency. While on PowerPC, it stores MMCRA[TECX/TECM], which should not be latency. So a general name prefix 'var$NUM' is used here. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-2-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:07 -07:00
tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
((attr->config & INTEL_ARCH_EVENT_MASK) ==
x86_pmu.rtm_abort_event);
if (gprs || (attr->precise_ip < 2) || tsx_weight)
pebs_data_cfg |= PEBS_DATACFG_GP;
if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
(attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_data_cfg |= PEBS_DATACFG_XMMS;
if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
/*
* For now always log all LBRs. Could configure this
* later.
*/
pebs_data_cfg |= PEBS_DATACFG_LBRS |
((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
}
return pebs_data_cfg;
}
static void
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
struct perf_event *event, bool add)
{
perf: Rewrite core context handling There have been various issues and limitations with the way perf uses (task) contexts to track events. Most notable is the single hardware PMU task context, which has resulted in a number of yucky things (both proposed and merged). Notably: - HW breakpoint PMU - ARM big.little PMU / Intel ADL PMU - Intel Branch Monitoring PMU - AMD IBS PMU - S390 cpum_cf PMU - PowerPC trace_imc PMU *Current design:* Currently we have a per task and per cpu perf_event_contexts: task_struct::perf_events_ctxp[] <-> perf_event_context <-> perf_cpu_context ^ | ^ | ^ `---------------------------------' | `--> pmu ---' v ^ perf_event ------' Each task has an array of pointers to a perf_event_context. Each perf_event_context has a direct relation to a PMU and a group of events for that PMU. The task related perf_event_context's have a pointer back to that task. Each PMU has a per-cpu pointer to a per-cpu perf_cpu_context, which includes a perf_event_context, which again has a direct relation to that PMU, and a group of events for that PMU. The perf_cpu_context also tracks which task context is currently associated with that CPU and includes a few other things like the hrtimer for rotation etc. Each perf_event is then associated with its PMU and one perf_event_context. *Proposed design:* New design proposed by this patch reduce to a single task context and a single CPU context but adds some intermediate data-structures: task_struct::perf_event_ctxp -> perf_event_context <- perf_cpu_context ^ | ^ ^ `---------------------------' | | | | perf_cpu_pmu_context <--. | `----. ^ | | | | | | v v | | ,--> perf_event_pmu_context | | | | | | | v v | perf_event ---> pmu ----------------' With the new design, perf_event_context will hold all events for all pmus in the (respective pinned/flexible) rbtrees. This can be achieved by adding pmu to rbtree key: {cpu, pmu, cgroup, group_index} Each perf_event_context carries a list of perf_event_pmu_context which is used to hold per-pmu-per-context state. For example, it keeps track of currently active events for that pmu, a pmu specific task_ctx_data, a flag to tell whether rotation is required or not etc. Additionally, perf_cpu_pmu_context is used to hold per-pmu-per-cpu state like hrtimer details to drive the event rotation, a pointer to perf_event_pmu_context of currently running task and some other ancillary information. Each perf_event is associated to it's pmu, perf_event_context and perf_event_pmu_context. Further optimizations to current implementation are possible. For example, ctx_resched() can be optimized to reschedule only single pmu events. Much thanks to Ravi for picking this up and pushing it towards completion. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Co-developed-by: Ravi Bangoria <ravi.bangoria@amd.com> Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20221008062424.313-1-ravi.bangoria@amd.com
2022-10-07 23:24:24 -07:00
struct pmu *pmu = event->pmu;
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
/*
perf/x86/intel/ds: Don't clear ->pebs_data_cfg for the last PEBS event The MSR_PEBS_DATA_CFG MSR register is used to configure which data groups should be generated into a PEBS record, and it's shared among all counters. If there are different configurations among counters, perf combines all the configurations. The first perf command as below requires a complete PEBS record (including memory info, GPRs, XMMs, and LBRs). The second perf command only requires a basic group. However, after the second perf command is running, the MSR_PEBS_DATA_CFG register is cleared. Only a basic group is generated in a PEBS record, which is wrong. The required information for the first perf command is missed. $ perf record --intr-regs=AX,SP,XMM0 -a -C 8 -b -W -d -c 100000003 -o /dev/null -e cpu/event=0xd0,umask=0x81/upp & $ sleep 5 $ perf record --per-thread -c 1 -e cycles:pp --no-timestamp --no-tid taskset -c 8 ./noploop 1000 The first PEBS event is a system-wide PEBS event. The second PEBS event is a per-thread event. When the thread is scheduled out, the intel_pmu_pebs_del() function is invoked to update the PEBS state. Since the system-wide event is still available, the cpuc->n_pebs is 1. The cpuc->pebs_data_cfg is cleared. The data configuration for the system-wide PEBS event is lost. The (cpuc->n_pebs == 1) check was introduced in commit: b6a32f023fcc ("perf/x86: Fix PEBS threshold initialization") At that time, it indeed didn't hurt whether the state was updated during the removal, because only the threshold is updated. The calculation of the threshold takes the last PEBS event into account. However, since commit: b752ea0c28e3 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG") we delay the threshold update, and clear the PEBS data config, which triggers the bug. The PEBS data config update scope should not be shrunk during removal. [ mingo: Improved the changelog & comments. ] Fixes: b752ea0c28e3 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG") Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Cc: stable@vger.kernel.org Link: https://lore.kernel.org/r/20240401133320.703971-1-kan.liang@linux.intel.com
2024-04-01 06:33:20 -07:00
* Make sure we get updated with the first PEBS event.
* During removal, ->pebs_data_cfg is still valid for
* the last PEBS event. Don't clear it.
*/
perf/x86/intel/ds: Don't clear ->pebs_data_cfg for the last PEBS event The MSR_PEBS_DATA_CFG MSR register is used to configure which data groups should be generated into a PEBS record, and it's shared among all counters. If there are different configurations among counters, perf combines all the configurations. The first perf command as below requires a complete PEBS record (including memory info, GPRs, XMMs, and LBRs). The second perf command only requires a basic group. However, after the second perf command is running, the MSR_PEBS_DATA_CFG register is cleared. Only a basic group is generated in a PEBS record, which is wrong. The required information for the first perf command is missed. $ perf record --intr-regs=AX,SP,XMM0 -a -C 8 -b -W -d -c 100000003 -o /dev/null -e cpu/event=0xd0,umask=0x81/upp & $ sleep 5 $ perf record --per-thread -c 1 -e cycles:pp --no-timestamp --no-tid taskset -c 8 ./noploop 1000 The first PEBS event is a system-wide PEBS event. The second PEBS event is a per-thread event. When the thread is scheduled out, the intel_pmu_pebs_del() function is invoked to update the PEBS state. Since the system-wide event is still available, the cpuc->n_pebs is 1. The cpuc->pebs_data_cfg is cleared. The data configuration for the system-wide PEBS event is lost. The (cpuc->n_pebs == 1) check was introduced in commit: b6a32f023fcc ("perf/x86: Fix PEBS threshold initialization") At that time, it indeed didn't hurt whether the state was updated during the removal, because only the threshold is updated. The calculation of the threshold takes the last PEBS event into account. However, since commit: b752ea0c28e3 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG") we delay the threshold update, and clear the PEBS data config, which triggers the bug. The PEBS data config update scope should not be shrunk during removal. [ mingo: Improved the changelog & comments. ] Fixes: b752ea0c28e3 ("perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG") Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Cc: stable@vger.kernel.org Link: https://lore.kernel.org/r/20240401133320.703971-1-kan.liang@linux.intel.com
2024-04-01 06:33:20 -07:00
if ((cpuc->n_pebs == 1) && add)
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
cpuc->pebs_data_cfg = PEBS_UPDATE_DS_SW;
if (needed_cb != pebs_needs_sched_cb(cpuc)) {
if (!needed_cb)
perf_sched_cb_inc(pmu);
else
perf_sched_cb_dec(pmu);
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
cpuc->pebs_data_cfg |= PEBS_UPDATE_DS_SW;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
/*
* The PEBS record doesn't shrink on pmu::del(). Doing so would require
* iterating all remaining PEBS events to reconstruct the config.
*/
if (x86_pmu.intel_cap.pebs_baseline && add) {
u64 pebs_data_cfg;
pebs_data_cfg = pebs_update_adaptive_cfg(event);
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
/*
* Be sure to update the thresholds when we change the record.
*/
if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
}
}
void intel_pmu_pebs_add(struct perf_event *event)
perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS interrupt threshold) PEBS always had the capability to log samples to its buffers without an interrupt. Traditionally perf has not used this but always set the PEBS threshold to one. For frequently occurring events (like cycles or branches or load/store) this in term requires using a relatively high sampling period to avoid overloading the system, by only processing PMIs. This in term increases sampling error. For the common cases we still need to use the PMI because the PEBS hardware has various limitations. The biggest one is that it can not supply a callgraph. It also requires setting a fixed period, as the hardware does not support adaptive period. Another issue is that it cannot supply a time stamp and some other options. To supply a TID it requires flushing on context switch. It can however supply the IP, the load/store address, TSX information, registers, and some other things. So we can make PEBS work for some specific cases, basically as long as you can do without a callgraph and can set the period you can use this new PEBS mode. The main benefit is the ability to support much lower sampling period (down to -c 1000) without extensive overhead. One use cases is for example to increase the resolution of the c2c tool. Another is double checking when you suspect the standard sampling has too much sampling error. Some numbers on the overhead, using cycle soak, comparing the elapsed time from "kernbench -M -H" between plain (threshold set to one) and multi (large threshold). The test command for plain: "perf record --time -e cycles:p -c $period -- kernbench -M -H" The test command for multi: "perf record --no-time -e cycles:p -c $period -- kernbench -M -H" ( The only difference of test command between multi and plain is time stamp options. Since time stamp is not supported by large PEBS threshold, it can be used as a flag to indicate if large threshold is enabled during the test. ) period plain(Sec) multi(Sec) Delta 10003 32.7 16.5 16.2 20003 30.2 16.2 14.0 40003 18.6 14.1 4.5 80003 16.8 14.6 2.2 100003 16.9 14.1 2.8 800003 15.4 15.7 -0.3 1000003 15.3 15.2 0.2 2000003 15.3 15.1 0.1 With periods below 100003, plain (threshold one) cause much more overhead. With 10003 sampling period, the Elapsed Time for multi is even 2X faster than plain. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:50 -07:00
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct hw_perf_event *hwc = &event->hw;
bool needed_cb = pebs_needs_sched_cb(cpuc);
cpuc->n_pebs++;
if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
cpuc->n_large_pebs++;
if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
cpuc->n_pebs_via_pt++;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_update_state(needed_cb, cpuc, event, true);
perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS interrupt threshold) PEBS always had the capability to log samples to its buffers without an interrupt. Traditionally perf has not used this but always set the PEBS threshold to one. For frequently occurring events (like cycles or branches or load/store) this in term requires using a relatively high sampling period to avoid overloading the system, by only processing PMIs. This in term increases sampling error. For the common cases we still need to use the PMI because the PEBS hardware has various limitations. The biggest one is that it can not supply a callgraph. It also requires setting a fixed period, as the hardware does not support adaptive period. Another issue is that it cannot supply a time stamp and some other options. To supply a TID it requires flushing on context switch. It can however supply the IP, the load/store address, TSX information, registers, and some other things. So we can make PEBS work for some specific cases, basically as long as you can do without a callgraph and can set the period you can use this new PEBS mode. The main benefit is the ability to support much lower sampling period (down to -c 1000) without extensive overhead. One use cases is for example to increase the resolution of the c2c tool. Another is double checking when you suspect the standard sampling has too much sampling error. Some numbers on the overhead, using cycle soak, comparing the elapsed time from "kernbench -M -H" between plain (threshold set to one) and multi (large threshold). The test command for plain: "perf record --time -e cycles:p -c $period -- kernbench -M -H" The test command for multi: "perf record --no-time -e cycles:p -c $period -- kernbench -M -H" ( The only difference of test command between multi and plain is time stamp options. Since time stamp is not supported by large PEBS threshold, it can be used as a flag to indicate if large threshold is enabled during the test. ) period plain(Sec) multi(Sec) Delta 10003 32.7 16.5 16.2 20003 30.2 16.2 14.0 40003 18.6 14.1 4.5 80003 16.8 14.6 2.2 100003 16.9 14.1 2.8 800003 15.4 15.7 -0.3 1000003 15.3 15.2 0.2 2000003 15.3 15.1 0.1 With periods below 100003, plain (threshold one) cause much more overhead. With 10003 sampling period, the Elapsed Time for multi is even 2X faster than plain. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:50 -07:00
}
static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
if (!is_pebs_pt(event))
return;
if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
}
static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct hw_perf_event *hwc = &event->hw;
struct debug_store *ds = cpuc->ds;
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
u64 value = ds->pebs_event_reset[hwc->idx];
u32 base = MSR_RELOAD_PMC0;
unsigned int idx = hwc->idx;
if (!is_pebs_pt(event))
return;
if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
base = MSR_RELOAD_FIXED_CTR0;
idx = hwc->idx - INTEL_PMC_IDX_FIXED;
if (x86_pmu.intel_cap.pebs_format < 5)
value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
else
value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
}
wrmsrl(base + idx, value);
}
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
static inline void intel_pmu_drain_large_pebs(struct cpu_hw_events *cpuc)
{
if (cpuc->n_pebs == cpuc->n_large_pebs &&
cpuc->n_pebs != cpuc->n_pebs_via_pt)
intel_pmu_drain_pebs_buffer();
}
void intel_pmu_pebs_enable(struct perf_event *event)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
u64 pebs_data_cfg = cpuc->pebs_data_cfg & ~PEBS_UPDATE_DS_SW;
struct hw_perf_event *hwc = &event->hw;
struct debug_store *ds = cpuc->ds;
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
unsigned int idx = hwc->idx;
hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
cpuc->pebs_enabled |= 1ULL << hwc->idx;
perf/x86/intel: Add Icelake support Add Icelake core PMU perf code, including constraint tables and the main enable code. Icelake expanded the generic counters to always 8 even with HT on, but a range of events cannot be scheduled on the extra 4 counters. Add new constraint ranges to describe this to the scheduler. The number of constraints that need to be checked is larger now than with earlier CPUs. At some point we may need a new data structure to look them up more efficiently than with linear search. So far it still seems to be acceptable however. Icelake added a new fixed counter SLOTS. Full support for it is added later in the patch series. The cache events table is identical to Skylake. Compare to PEBS instruction event on generic counter, fixed counter 0 has less skid. Force instruction:ppp always in fixed counter 0. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-9-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:05 -07:00
if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
cpuc->pebs_enabled |= 1ULL << 63;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (x86_pmu.intel_cap.pebs_baseline) {
hwc->config |= ICL_EVENTSEL_ADAPTIVE;
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
if (pebs_data_cfg != cpuc->active_pebs_data_cfg) {
/*
* drain_pebs() assumes uniform record size;
* hence we need to drain when changing said
* size.
*/
intel_pmu_drain_large_pebs(cpuc);
adaptive_pebs_record_size_update();
wrmsrl(MSR_PEBS_DATA_CFG, pebs_data_cfg);
cpuc->active_pebs_data_cfg = pebs_data_cfg;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
}
}
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
if (cpuc->pebs_data_cfg & PEBS_UPDATE_DS_SW) {
cpuc->pebs_data_cfg = pebs_data_cfg;
pebs_update_threshold(cpuc);
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (idx >= INTEL_PMC_IDX_FIXED) {
if (x86_pmu.intel_cap.pebs_format < 5)
idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
else
idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
}
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS interrupt threshold) PEBS always had the capability to log samples to its buffers without an interrupt. Traditionally perf has not used this but always set the PEBS threshold to one. For frequently occurring events (like cycles or branches or load/store) this in term requires using a relatively high sampling period to avoid overloading the system, by only processing PMIs. This in term increases sampling error. For the common cases we still need to use the PMI because the PEBS hardware has various limitations. The biggest one is that it can not supply a callgraph. It also requires setting a fixed period, as the hardware does not support adaptive period. Another issue is that it cannot supply a time stamp and some other options. To supply a TID it requires flushing on context switch. It can however supply the IP, the load/store address, TSX information, registers, and some other things. So we can make PEBS work for some specific cases, basically as long as you can do without a callgraph and can set the period you can use this new PEBS mode. The main benefit is the ability to support much lower sampling period (down to -c 1000) without extensive overhead. One use cases is for example to increase the resolution of the c2c tool. Another is double checking when you suspect the standard sampling has too much sampling error. Some numbers on the overhead, using cycle soak, comparing the elapsed time from "kernbench -M -H" between plain (threshold set to one) and multi (large threshold). The test command for plain: "perf record --time -e cycles:p -c $period -- kernbench -M -H" The test command for multi: "perf record --no-time -e cycles:p -c $period -- kernbench -M -H" ( The only difference of test command between multi and plain is time stamp options. Since time stamp is not supported by large PEBS threshold, it can be used as a flag to indicate if large threshold is enabled during the test. ) period plain(Sec) multi(Sec) Delta 10003 32.7 16.5 16.2 20003 30.2 16.2 14.0 40003 18.6 14.1 4.5 80003 16.8 14.6 2.2 100003 16.9 14.1 2.8 800003 15.4 15.7 -0.3 1000003 15.3 15.2 0.2 2000003 15.3 15.1 0.1 With periods below 100003, plain (threshold one) cause much more overhead. With 10003 sampling period, the Elapsed Time for multi is even 2X faster than plain. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:50 -07:00
/*
* Use auto-reload if possible to save a MSR write in the PMI.
* This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS interrupt threshold) PEBS always had the capability to log samples to its buffers without an interrupt. Traditionally perf has not used this but always set the PEBS threshold to one. For frequently occurring events (like cycles or branches or load/store) this in term requires using a relatively high sampling period to avoid overloading the system, by only processing PMIs. This in term increases sampling error. For the common cases we still need to use the PMI because the PEBS hardware has various limitations. The biggest one is that it can not supply a callgraph. It also requires setting a fixed period, as the hardware does not support adaptive period. Another issue is that it cannot supply a time stamp and some other options. To supply a TID it requires flushing on context switch. It can however supply the IP, the load/store address, TSX information, registers, and some other things. So we can make PEBS work for some specific cases, basically as long as you can do without a callgraph and can set the period you can use this new PEBS mode. The main benefit is the ability to support much lower sampling period (down to -c 1000) without extensive overhead. One use cases is for example to increase the resolution of the c2c tool. Another is double checking when you suspect the standard sampling has too much sampling error. Some numbers on the overhead, using cycle soak, comparing the elapsed time from "kernbench -M -H" between plain (threshold set to one) and multi (large threshold). The test command for plain: "perf record --time -e cycles:p -c $period -- kernbench -M -H" The test command for multi: "perf record --no-time -e cycles:p -c $period -- kernbench -M -H" ( The only difference of test command between multi and plain is time stamp options. Since time stamp is not supported by large PEBS threshold, it can be used as a flag to indicate if large threshold is enabled during the test. ) period plain(Sec) multi(Sec) Delta 10003 32.7 16.5 16.2 20003 30.2 16.2 14.0 40003 18.6 14.1 4.5 80003 16.8 14.6 2.2 100003 16.9 14.1 2.8 800003 15.4 15.7 -0.3 1000003 15.3 15.2 0.2 2000003 15.3 15.1 0.1 With periods below 100003, plain (threshold one) cause much more overhead. With 10003 sampling period, the Elapsed Time for multi is even 2X faster than plain. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:50 -07:00
*/
if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
ds->pebs_event_reset[idx] =
(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
perf/x86/intel: Fix debug_store reset field for freq events There's a bug in PEBs event enabling code, that prevents PEBS freq events to work properly after non freq PEBS event was run. freq events - perf_event_attr::freq set -F <freq> option of perf record PEBS events - perf_event_attr::precise_ip > 0 default for perf record Like in following example with CPU 0 busy, we expect ~10000 samples for following perf tool run: # perf record -F 10000 -C 0 sleep 1 [ perf record: Woken up 2 times to write data ] [ perf record: Captured and wrote 0.640 MB perf.data (10031 samples) ] Everything's fine, but once we run non freq PEBS event like: # perf record -c 10000 -C 0 sleep 1 [ perf record: Woken up 4 times to write data ] [ perf record: Captured and wrote 1.053 MB perf.data (20061 samples) ] the freq events start to fail like this: # perf record -F 10000 -C 0 sleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.185 MB perf.data (40 samples) ] The issue is in non freq PEBs event initialization of debug_store reset field, which value is used to auto-reload the counter value after PEBS event drain. This value is not being used for PEBS freq events, but once we run non freq event it stays in debug_store data and screws the sample_freq counting for PEBS freq events. Setting the reset field to 0 for freq events. Signed-off-by: Jiri Olsa <jolsa@kernel.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Kan Liang <kan.liang@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20170714163551.19459-1-jolsa@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-14 09:35:51 -07:00
} else {
perf/x86/intel: Fix PEBS-via-PT reload base value for Extended PEBS If we use the "PEBS-via-PT" feature on a platform that supports extended PBES, like this: perf record -c 10000 \ -e '{intel_pt/branch=0/,branch-instructions/aux-output/p}' uname we will encounter the following call trace: [ 250.906542] unchecked MSR access error: WRMSR to 0x14e1 (tried to write 0x0000000000000000) at rIP: 0xffffffff88073624 (native_write_msr+0x4/0x20) [ 250.920779] Call Trace: [ 250.923508] intel_pmu_pebs_enable+0x12c/0x190 [ 250.928359] intel_pmu_enable_event+0x346/0x390 [ 250.933300] x86_pmu_start+0x64/0x80 [ 250.937231] x86_pmu_enable+0x16a/0x2f0 [ 250.941434] perf_event_exec+0x144/0x4c0 [ 250.945731] begin_new_exec+0x650/0xbf0 [ 250.949933] load_elf_binary+0x13e/0x1700 [ 250.954321] ? lock_acquire+0xc2/0x390 [ 250.958430] ? bprm_execve+0x34f/0x8a0 [ 250.962544] ? lock_is_held_type+0xa7/0x120 [ 250.967118] ? find_held_lock+0x32/0x90 [ 250.971321] ? sched_clock_cpu+0xc/0xb0 [ 250.975527] bprm_execve+0x33d/0x8a0 [ 250.979452] do_execveat_common.isra.0+0x161/0x1d0 [ 250.984673] __x64_sys_execve+0x33/0x40 [ 250.988877] do_syscall_64+0x3d/0x80 [ 250.992806] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 250.998302] RIP: 0033:0x7fbc971d82fb [ 251.002235] Code: Unable to access opcode bytes at RIP 0x7fbc971d82d1. [ 251.009303] RSP: 002b:00007fffb8aed808 EFLAGS: 00000202 ORIG_RAX: 000000000000003b [ 251.017478] RAX: ffffffffffffffda RBX: 00007fffb8af2f00 RCX: 00007fbc971d82fb [ 251.025187] RDX: 00005574792aac50 RSI: 00007fffb8af2f00 RDI: 00007fffb8aed810 [ 251.032901] RBP: 00007fffb8aed970 R08: 0000000000000020 R09: 00007fbc9725c8b0 [ 251.040613] R10: 6d6c61632f6d6f63 R11: 0000000000000202 R12: 00005574792aac50 [ 251.048327] R13: 00007fffb8af35f0 R14: 00005574792aafdf R15: 00005574792aafe7 This is because the target reload msr address is calculated based on the wrong base msr and the target reload msr value is accessed from ds->pebs_event_reset[] with the wrong offset. According to Intel SDM Table 2-14, for extended PBES feature, the reload msr for MSR_IA32_FIXED_CTRx should be based on MSR_RELOAD_FIXED_CTRx. For fixed counters, let's fix it by overriding the reload msr address and its value, thus avoiding out-of-bounds access. Fixes: 42880f726c66("perf/x86/intel: Support PEBS output to PT") Signed-off-by: Like Xu <likexu@tencent.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20210621034710.31107-1-likexu@tencent.com
2021-06-20 20:47:10 -07:00
ds->pebs_event_reset[idx] = 0;
}
intel_pmu_pebs_via_pt_enable(event);
}
void intel_pmu_pebs_del(struct perf_event *event)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct hw_perf_event *hwc = &event->hw;
bool needed_cb = pebs_needs_sched_cb(cpuc);
cpuc->n_pebs--;
if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
cpuc->n_large_pebs--;
if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
cpuc->n_pebs_via_pt--;
perf/x86/intel: Implement batched PEBS interrupt handling (large PEBS interrupt threshold) PEBS always had the capability to log samples to its buffers without an interrupt. Traditionally perf has not used this but always set the PEBS threshold to one. For frequently occurring events (like cycles or branches or load/store) this in term requires using a relatively high sampling period to avoid overloading the system, by only processing PMIs. This in term increases sampling error. For the common cases we still need to use the PMI because the PEBS hardware has various limitations. The biggest one is that it can not supply a callgraph. It also requires setting a fixed period, as the hardware does not support adaptive period. Another issue is that it cannot supply a time stamp and some other options. To supply a TID it requires flushing on context switch. It can however supply the IP, the load/store address, TSX information, registers, and some other things. So we can make PEBS work for some specific cases, basically as long as you can do without a callgraph and can set the period you can use this new PEBS mode. The main benefit is the ability to support much lower sampling period (down to -c 1000) without extensive overhead. One use cases is for example to increase the resolution of the c2c tool. Another is double checking when you suspect the standard sampling has too much sampling error. Some numbers on the overhead, using cycle soak, comparing the elapsed time from "kernbench -M -H" between plain (threshold set to one) and multi (large threshold). The test command for plain: "perf record --time -e cycles:p -c $period -- kernbench -M -H" The test command for multi: "perf record --no-time -e cycles:p -c $period -- kernbench -M -H" ( The only difference of test command between multi and plain is time stamp options. Since time stamp is not supported by large PEBS threshold, it can be used as a flag to indicate if large threshold is enabled during the test. ) period plain(Sec) multi(Sec) Delta 10003 32.7 16.5 16.2 20003 30.2 16.2 14.0 40003 18.6 14.1 4.5 80003 16.8 14.6 2.2 100003 16.9 14.1 2.8 800003 15.4 15.7 -0.3 1000003 15.3 15.2 0.2 2000003 15.3 15.1 0.1 With periods below 100003, plain (threshold one) cause much more overhead. With 10003 sampling period, the Elapsed Time for multi is even 2X faster than plain. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-5-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:50 -07:00
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_update_state(needed_cb, cpuc, event, false);
}
void intel_pmu_pebs_disable(struct perf_event *event)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct hw_perf_event *hwc = &event->hw;
perf/x86/intel/ds: Flush PEBS DS when changing PEBS_DATA_CFG Several similar kernel warnings can be triggered, [56605.607840] CPU0 PEBS record size 0, expected 32, config 0 cpuc->record_size=208 when the below commands are running in parallel for a while on SPR. while true; do perf record --no-buildid -a --intr-regs=AX \ -e cpu/event=0xd0,umask=0x81/pp \ -c 10003 -o /dev/null ./triad; done & while true; do perf record -o /tmp/out -W -d \ -e '{ld_blocks.store_forward:period=1000000, \ MEM_TRANS_RETIRED.LOAD_LATENCY:u:precise=2:ldlat=4}' \ -c 1037 ./triad; done The triad program is just the generation of loads/stores. The warnings are triggered when an unexpected PEBS record (with a different config and size) is found. A system-wide PEBS event with the large PEBS config may be enabled during a context switch. Some PEBS records for the system-wide PEBS may be generated while the old task is sched out but the new one hasn't been sched in yet. When the new task is sched in, the cpuc->pebs_record_size may be updated for the per-task PEBS events. So the existing system-wide PEBS records have a different size from the later PEBS records. The PEBS buffer should be flushed right before the hardware is reprogrammed. The new size and threshold should be updated after the old buffer has been flushed. Reported-by: Stephane Eranian <eranian@google.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20230421184529.3320912-1-kan.liang@linux.intel.com
2023-04-21 11:45:28 -07:00
intel_pmu_drain_large_pebs(cpuc);
cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
perf/x86/intel: Add Icelake support Add Icelake core PMU perf code, including constraint tables and the main enable code. Icelake expanded the generic counters to always 8 even with HT on, but a range of events cannot be scheduled on the extra 4 counters. Add new constraint ranges to describe this to the scheduler. The number of constraints that need to be checked is larger now than with earlier CPUs. At some point we may need a new data structure to look them up more efficiently than with linear search. So far it still seems to be acceptable however. Icelake added a new fixed counter SLOTS. Full support for it is added later in the patch series. The cache events table is identical to Skylake. Compare to PEBS instruction event on generic counter, fixed counter 0 has less skid. Force instruction:ppp always in fixed counter 0. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-9-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:05 -07:00
if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
(x86_pmu.version < 5))
cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
perf/x86: Fix event/group validation Commit 43b4578071c0 ("perf/x86: Reduce stack usage of x86_schedule_events()") violated the rule that 'fake' scheduling; as used for event/group validation; should not change the event state. This went mostly un-noticed because repeated calls of x86_pmu::get_event_constraints() would give the same result. And x86_pmu::put_event_constraints() would mostly not do anything. Commit e979121b1b15 ("perf/x86/intel: Implement cross-HT corruption bug workaround") made the situation much worse by actually setting the event->hw.constraint value to NULL, so when validation and actual scheduling interact we get NULL ptr derefs. Fix it by removing the constraint pointer from the event and move it back to an array, this time in cpuc instead of on the stack. validate_group() x86_schedule_events() event->hw.constraint = c; # store <context switch> perf_task_event_sched_in() ... x86_schedule_events(); event->hw.constraint = c2; # store ... put_event_constraints(event); # assume failure to schedule intel_put_event_constraints() event->hw.constraint = NULL; <context switch end> c = event->hw.constraint; # read -> NULL if (!test_bit(hwc->idx, c->idxmsk)) # <- *BOOM* NULL deref This in particular is possible when the event in question is a cpu-wide event and group-leader, where the validate_group() tries to add an event to the group. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Hunter <ahh@google.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maria Dimakopoulou <maria.n.dimakopoulou@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 43b4578071c0 ("perf/x86: Reduce stack usage of x86_schedule_events()") Fixes: e979121b1b15 ("perf/x86/intel: Implement cross-HT corruption bug workaround") Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-21 01:57:13 -07:00
else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
cpuc->pebs_enabled &= ~(1ULL << 63);
intel_pmu_pebs_via_pt_disable(event);
if (cpuc->enabled)
wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
}
void intel_pmu_pebs_enable_all(void)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
if (cpuc->pebs_enabled)
wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
}
void intel_pmu_pebs_disable_all(void)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
if (cpuc->pebs_enabled)
__intel_pmu_pebs_disable_all();
}
static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
unsigned long from = cpuc->lbr_entries[0].from;
unsigned long old_to, to = cpuc->lbr_entries[0].to;
unsigned long ip = regs->ip;
int is_64bit = 0;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
void *kaddr;
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
int size;
/*
* We don't need to fixup if the PEBS assist is fault like
*/
if (!x86_pmu.intel_cap.pebs_trap)
return 1;
/*
* No LBR entry, no basic block, no rewinding
*/
if (!cpuc->lbr_stack.nr || !from || !to)
return 0;
/*
* Basic blocks should never cross user/kernel boundaries
*/
if (kernel_ip(ip) != kernel_ip(to))
return 0;
/*
* unsigned math, either ip is before the start (impossible) or
* the basic block is larger than 1 page (sanity)
*/
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
if ((ip - to) > PEBS_FIXUP_SIZE)
return 0;
/*
* We sampled a branch insn, rewind using the LBR stack
*/
if (ip == to) {
set_linear_ip(regs, from);
return 1;
}
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
size = ip - to;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
if (!kernel_ip(ip)) {
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
int bytes;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
u8 *buf = this_cpu_read(insn_buffer);
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
/* 'size' must fit our buffer, see above */
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
bytes = copy_from_user_nmi(buf, (void __user *)to, size);
if (bytes != 0)
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
return 0;
kaddr = buf;
} else {
kaddr = (void *)to;
}
do {
struct insn insn;
old_to = to;
#ifdef CONFIG_X86_64
is_64bit = kernel_ip(to) || any_64bit_mode(regs);
#endif
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
insn_init(&insn, kaddr, size, is_64bit);
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
/*
* Make sure there was not a problem decoding the instruction.
* This is doubly important because we have an infinite loop if
* insn.length=0.
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
*/
if (insn_get_length(&insn))
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
break;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
to += insn.length;
perf/x86: Optimize intel_pmu_pebs_fixup_ip() There's been reports of high NMI handler overhead, highlighted by such kernel messages: [ 3697.380195] perf samples too long (10009 > 10000), lowering kernel.perf_event_max_sample_rate to 13000 [ 3697.389509] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.331 msecs Don Zickus analyzed the source of the overhead and reported: > While there are a few places that are causing latencies, for now I focused on > the longest one first. It seems to be 'copy_user_from_nmi' > > intel_pmu_handle_irq -> > intel_pmu_drain_pebs_nhm -> > __intel_pmu_drain_pebs_nhm -> > __intel_pmu_pebs_event -> > intel_pmu_pebs_fixup_ip -> > copy_from_user_nmi > > In intel_pmu_pebs_fixup_ip(), if the while-loop goes over 50, the sum of > all the copy_from_user_nmi latencies seems to go over 1,000,000 cycles > (there are some cases where only 10 iterations are needed to go that high > too, but in generall over 50 or so). At this point copy_user_from_nmi > seems to account for over 90% of the nmi latency. The solution to that is to avoid having to call copy_from_user_nmi() for every instruction. Since we already limit the max basic block size, we can easily pre-allocate a piece of memory to copy the entire thing into in one go. Don reported this test result: > Your patch made a huge difference in improvement. The > copy_from_user_nmi() no longer hits the million of cycles. I still > have a batch of 100,000-300,000 cycles. My longest NMI paths used > to be dominated by copy_from_user_nmi, now it is not (I have to dig > up the new hot path). Reported-and-tested-by: Don Zickus <dzickus@redhat.com> Cc: jmario@redhat.com Cc: acme@infradead.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20131016105755.GX10651@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-15 03:14:04 -07:00
kaddr += insn.length;
x86: Remove arbitrary instruction size limit in instruction decoder The current x86 instruction decoder steps along through the instruction stream but always ensures that it never steps farther than the largest possible instruction size (MAX_INSN_SIZE). The MPX code is now going to be doing some decoding of userspace instructions. We copy those from userspace in to the kernel and they're obviously completely untrusted coming from userspace. In addition to the constraint that instructions can only be so long, we also have to be aware of how long the buffer is that came in from userspace. This _looks_ to be similar to what the perf and kprobes is doing, but it's unclear to me whether they are affected. The whole reason we need this is that it is perfectly valid to be executing an instruction within MAX_INSN_SIZE bytes of an unreadable page. We should be able to gracefully handle short reads in those cases. This adds support to the decoder to record how long the buffer being decoded is and to refuse to "validate" the instruction if we would have gone over the end of the buffer to decode it. The kprobes code probably needs to be looked at here a bit more carefully. This patch still respects the MAX_INSN_SIZE limit there but the kprobes code does look like it might be able to be a bit more strict than it currently is. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Acked-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: x86@kernel.org Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: "David S. Miller" <davem@davemloft.net> Link: http://lkml.kernel.org/r/20141114153957.E6B01535@viggo.jf.intel.com Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2014-11-14 08:39:57 -07:00
size -= insn.length;
} while (to < ip);
if (to == ip) {
set_linear_ip(regs, old_to);
return 1;
}
/*
* Even though we decoded the basic block, the instruction stream
* never matched the given IP, either the TO or the IP got corrupted.
*/
return 0;
}
static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
{
if (tsx_tuning) {
union hsw_tsx_tuning tsx = { .value = tsx_tuning };
return tsx.cycles_last_block;
}
return 0;
}
static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
{
u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
/* For RTM XABORTs also log the abort code from AX */
if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
return txn;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
static inline u64 get_pebs_status(void *n)
{
if (x86_pmu.intel_cap.pebs_format < 4)
return ((struct pebs_record_nhm *)n)->status;
return ((struct pebs_basic *)n)->applicable_counters;
}
#define PERF_X86_EVENT_PEBS_HSW_PREC \
(PERF_X86_EVENT_PEBS_ST_HSW | \
PERF_X86_EVENT_PEBS_LD_HSW | \
PERF_X86_EVENT_PEBS_NA_HSW)
static u64 get_data_src(struct perf_event *event, u64 aux)
{
u64 val = PERF_MEM_NA;
int fl = event->hw.flags;
bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
if (fl & PERF_X86_EVENT_PEBS_LDLAT)
val = load_latency_data(event, aux);
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
else if (fl & PERF_X86_EVENT_PEBS_STLAT)
val = store_latency_data(event, aux);
else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID)
val = x86_pmu.pebs_latency_data(event, aux);
else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
val = precise_datala_hsw(event, aux);
else if (fst)
val = precise_store_data(aux);
return val;
}
static void setup_pebs_time(struct perf_event *event,
struct perf_sample_data *data,
u64 tsc)
{
/* Converting to a user-defined clock is not supported yet. */
if (event->attr.use_clockid != 0)
return;
/*
* Doesn't support the conversion when the TSC is unstable.
* The TSC unstable case is a corner case and very unlikely to
* happen. If it happens, the TSC in a PEBS record will be
* dropped and fall back to perf_event_clock().
*/
if (!using_native_sched_clock() || !sched_clock_stable())
return;
data->time = native_sched_clock_from_tsc(tsc) + __sched_clock_offset;
data->sample_flags |= PERF_SAMPLE_TIME;
}
#define PERF_SAMPLE_ADDR_TYPE (PERF_SAMPLE_ADDR | \
PERF_SAMPLE_PHYS_ADDR | \
PERF_SAMPLE_DATA_PAGE_SIZE)
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
static void setup_pebs_fixed_sample_data(struct perf_event *event,
struct pt_regs *iregs, void *__pebs,
struct perf_sample_data *data,
struct pt_regs *regs)
{
/*
* We cast to the biggest pebs_record but are careful not to
* unconditionally access the 'extra' entries.
*/
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct pebs_record_skl *pebs = __pebs;
u64 sample_type;
int fll;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
if (pebs == NULL)
return;
sample_type = event->attr.sample_type;
fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
perf_sample_data_init(data, 0, event->hw.last_period);
data->period = event->hw.last_period;
/*
* Use latency for weight (only avail with PEBS-LL)
*/
if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE)) {
perf/core: Add PERF_SAMPLE_WEIGHT_STRUCT Current PERF_SAMPLE_WEIGHT sample type is very useful to expresses the cost of an action represented by the sample. This allows the profiler to scale the samples to be more informative to the programmer. It could also help to locate a hotspot, e.g., when profiling by memory latencies, the expensive load appear higher up in the histograms. But current PERF_SAMPLE_WEIGHT sample type is solely determined by one factor. This could be a problem, if users want two or more factors to contribute to the weight. For example, Golden Cove core PMU can provide both the instruction latency and the cache Latency information as factors for the memory profiling. For current X86 platforms, although meminfo::latency is defined as a u64, only the lower 32 bits include the valid data in practice (No memory access could last than 4G cycles). The higher 32 bits can be used to store new factors. Add a new sample type, PERF_SAMPLE_WEIGHT_STRUCT, to indicate the new sample weight structure. It shares the same space as the PERF_SAMPLE_WEIGHT sample type. Users can apply either the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type to retrieve the sample weight, but they cannot apply both sample types simultaneously. Currently, only X86 and PowerPC use the PERF_SAMPLE_WEIGHT sample type. - For PowerPC, there is nothing changed for the PERF_SAMPLE_WEIGHT sample type. There is no effect for the new PERF_SAMPLE_WEIGHT_STRUCT sample type. PowerPC can re-struct the weight field similarly later. - For X86, the same value will be dumped for the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type for now. The following patches will apply the new factors for the PERF_SAMPLE_WEIGHT_STRUCT sample type. The field in the union perf_sample_weight should be shared among different architectures. A generic name is required, but it's hard to abstract a name that applies to all architectures. For example, on X86, the fields are to store all kinds of latency. While on PowerPC, it stores MMCRA[TECX/TECM], which should not be latency. So a general name prefix 'var$NUM' is used here. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-2-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:07 -07:00
data->weight.full = pebs->lat;
data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
}
/*
* data.data_src encodes the data source
*/
if (sample_type & PERF_SAMPLE_DATA_SRC) {
data->data_src.val = get_data_src(event, pebs->dse);
data->sample_flags |= PERF_SAMPLE_DATA_SRC;
}
perf/x86/intel: Fix unwind errors from PEBS entries (mk-II) Vince reported the perf_fuzzer giving various unwinder warnings and Josh reported: > Deja vu. Most of these are related to perf PEBS, similar to the > following issue: > > b8000586c90b ("perf/x86/intel: Cure bogus unwind from PEBS entries") > > This is basically the ORC version of that. setup_pebs_sample_data() is > assembling a franken-pt_regs which ORC isn't happy about. RIP is > inconsistent with some of the other registers (like RSP and RBP). And where the previous unwinder only needed BP,SP ORC also requires IP. But we cannot spoof IP because then the sample will get displaced, entirely negating the point of PEBS. So cure the whole thing differently by doing the unwind early; this does however require a means to communicate we did the unwind early. We (ab)use an unused sample_type bit for this, which we set on events that fill out the data->callchain before the normal perf_prepare_sample(). Debugged-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Tested-by: Josh Poimboeuf <jpoimboe@redhat.com> Tested-by: Prashant Bhole <bhole_prashant_q7@lab.ntt.co.jp> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-05-10 06:48:41 -07:00
/*
* We must however always use iregs for the unwinder to stay sane; the
* record BP,SP,IP can point into thin air when the record is from a
* previous PMI context or an (I)RET happened between the record and
perf/x86/intel: Fix unwind errors from PEBS entries (mk-II) Vince reported the perf_fuzzer giving various unwinder warnings and Josh reported: > Deja vu. Most of these are related to perf PEBS, similar to the > following issue: > > b8000586c90b ("perf/x86/intel: Cure bogus unwind from PEBS entries") > > This is basically the ORC version of that. setup_pebs_sample_data() is > assembling a franken-pt_regs which ORC isn't happy about. RIP is > inconsistent with some of the other registers (like RSP and RBP). And where the previous unwinder only needed BP,SP ORC also requires IP. But we cannot spoof IP because then the sample will get displaced, entirely negating the point of PEBS. So cure the whole thing differently by doing the unwind early; this does however require a means to communicate we did the unwind early. We (ab)use an unused sample_type bit for this, which we set on events that fill out the data->callchain before the normal perf_prepare_sample(). Debugged-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Tested-by: Josh Poimboeuf <jpoimboe@redhat.com> Tested-by: Prashant Bhole <bhole_prashant_q7@lab.ntt.co.jp> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-05-10 06:48:41 -07:00
* PMI.
*/
if (sample_type & PERF_SAMPLE_CALLCHAIN)
perf_sample_save_callchain(data, event, iregs);
perf/x86/intel: Fix unwind errors from PEBS entries (mk-II) Vince reported the perf_fuzzer giving various unwinder warnings and Josh reported: > Deja vu. Most of these are related to perf PEBS, similar to the > following issue: > > b8000586c90b ("perf/x86/intel: Cure bogus unwind from PEBS entries") > > This is basically the ORC version of that. setup_pebs_sample_data() is > assembling a franken-pt_regs which ORC isn't happy about. RIP is > inconsistent with some of the other registers (like RSP and RBP). And where the previous unwinder only needed BP,SP ORC also requires IP. But we cannot spoof IP because then the sample will get displaced, entirely negating the point of PEBS. So cure the whole thing differently by doing the unwind early; this does however require a means to communicate we did the unwind early. We (ab)use an unused sample_type bit for this, which we set on events that fill out the data->callchain before the normal perf_prepare_sample(). Debugged-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Tested-by: Josh Poimboeuf <jpoimboe@redhat.com> Tested-by: Prashant Bhole <bhole_prashant_q7@lab.ntt.co.jp> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-05-10 06:48:41 -07:00
/*
perf/x86/intel: Cure bogus unwind from PEBS entries Vince Weaver reported that perf_fuzzer + KASAN detects that PEBS event unwinds sometimes do 'weird' things. In particular, we seemed to be ending up unwinding from random places on the NMI stack. While it was somewhat expected that the event record BP,SP would not match the interrupt BP,SP in that the interrupt is strictly later than the record event, it was overlooked that it could be on an already overwritten stack. Therefore, don't copy the recorded BP,SP over the interrupted BP,SP when we need stack unwinds. Note that its still possible the unwind doesn't full match the actual event, as its entirely possible to have done an (I)RET between record and interrupt, but on average it should still point in the general direction of where the event came from. Also, it's the best we can do, considering. The particular scenario that triggered the bogus NMI stack unwind was a PEBS event with very short period, upon enabling the event at the tail of the PMI handler (FREEZE_ON_PMI is not used), it instantly triggers a record (while still on the NMI stack) which in turn triggers the next PMI. This then causes back-to-back NMIs and we'll try and unwind the stack-frame from the last NMI, which obviously is now overwritten by our own. Analyzed-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@gmail.com> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davej@codemonkey.org.uk <davej@codemonkey.org.uk> Cc: dvyukov@google.com <dvyukov@google.com> Cc: stable@vger.kernel.org Fixes: ca037701a025 ("perf, x86: Add PEBS infrastructure") Link: http://lkml.kernel.org/r/20161117171731.GV3157@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-11-17 10:17:31 -07:00
* We use the interrupt regs as a base because the PEBS record does not
* contain a full regs set, specifically it seems to lack segment
* descriptors, which get used by things like user_mode().
*
perf/x86/intel: Cure bogus unwind from PEBS entries Vince Weaver reported that perf_fuzzer + KASAN detects that PEBS event unwinds sometimes do 'weird' things. In particular, we seemed to be ending up unwinding from random places on the NMI stack. While it was somewhat expected that the event record BP,SP would not match the interrupt BP,SP in that the interrupt is strictly later than the record event, it was overlooked that it could be on an already overwritten stack. Therefore, don't copy the recorded BP,SP over the interrupted BP,SP when we need stack unwinds. Note that its still possible the unwind doesn't full match the actual event, as its entirely possible to have done an (I)RET between record and interrupt, but on average it should still point in the general direction of where the event came from. Also, it's the best we can do, considering. The particular scenario that triggered the bogus NMI stack unwind was a PEBS event with very short period, upon enabling the event at the tail of the PMI handler (FREEZE_ON_PMI is not used), it instantly triggers a record (while still on the NMI stack) which in turn triggers the next PMI. This then causes back-to-back NMIs and we'll try and unwind the stack-frame from the last NMI, which obviously is now overwritten by our own. Analyzed-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@gmail.com> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davej@codemonkey.org.uk <davej@codemonkey.org.uk> Cc: dvyukov@google.com <dvyukov@google.com> Cc: stable@vger.kernel.org Fixes: ca037701a025 ("perf, x86: Add PEBS infrastructure") Link: http://lkml.kernel.org/r/20161117171731.GV3157@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-11-17 10:17:31 -07:00
* In the simple case fix up only the IP for PERF_SAMPLE_IP.
*/
*regs = *iregs;
/*
* Initialize regs_>flags from PEBS,
* Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
* i.e., do not rely on it being zero:
*/
regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
if (sample_type & PERF_SAMPLE_REGS_INTR) {
regs->ax = pebs->ax;
regs->bx = pebs->bx;
regs->cx = pebs->cx;
regs->dx = pebs->dx;
regs->si = pebs->si;
regs->di = pebs->di;
perf/x86/intel: Fix unwind errors from PEBS entries (mk-II) Vince reported the perf_fuzzer giving various unwinder warnings and Josh reported: > Deja vu. Most of these are related to perf PEBS, similar to the > following issue: > > b8000586c90b ("perf/x86/intel: Cure bogus unwind from PEBS entries") > > This is basically the ORC version of that. setup_pebs_sample_data() is > assembling a franken-pt_regs which ORC isn't happy about. RIP is > inconsistent with some of the other registers (like RSP and RBP). And where the previous unwinder only needed BP,SP ORC also requires IP. But we cannot spoof IP because then the sample will get displaced, entirely negating the point of PEBS. So cure the whole thing differently by doing the unwind early; this does however require a means to communicate we did the unwind early. We (ab)use an unused sample_type bit for this, which we set on events that fill out the data->callchain before the normal perf_prepare_sample(). Debugged-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Tested-by: Josh Poimboeuf <jpoimboe@redhat.com> Tested-by: Prashant Bhole <bhole_prashant_q7@lab.ntt.co.jp> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-05-10 06:48:41 -07:00
regs->bp = pebs->bp;
regs->sp = pebs->sp;
perf/x86/intel: Cure bogus unwind from PEBS entries Vince Weaver reported that perf_fuzzer + KASAN detects that PEBS event unwinds sometimes do 'weird' things. In particular, we seemed to be ending up unwinding from random places on the NMI stack. While it was somewhat expected that the event record BP,SP would not match the interrupt BP,SP in that the interrupt is strictly later than the record event, it was overlooked that it could be on an already overwritten stack. Therefore, don't copy the recorded BP,SP over the interrupted BP,SP when we need stack unwinds. Note that its still possible the unwind doesn't full match the actual event, as its entirely possible to have done an (I)RET between record and interrupt, but on average it should still point in the general direction of where the event came from. Also, it's the best we can do, considering. The particular scenario that triggered the bogus NMI stack unwind was a PEBS event with very short period, upon enabling the event at the tail of the PMI handler (FREEZE_ON_PMI is not used), it instantly triggers a record (while still on the NMI stack) which in turn triggers the next PMI. This then causes back-to-back NMIs and we'll try and unwind the stack-frame from the last NMI, which obviously is now overwritten by our own. Analyzed-by: Josh Poimboeuf <jpoimboe@redhat.com> Reported-by: Vince Weaver <vincent.weaver@maine.edu> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@gmail.com> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davej@codemonkey.org.uk <davej@codemonkey.org.uk> Cc: dvyukov@google.com <dvyukov@google.com> Cc: stable@vger.kernel.org Fixes: ca037701a025 ("perf, x86: Add PEBS infrastructure") Link: http://lkml.kernel.org/r/20161117171731.GV3157@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-11-17 10:17:31 -07:00
#ifndef CONFIG_X86_32
regs->r8 = pebs->r8;
regs->r9 = pebs->r9;
regs->r10 = pebs->r10;
regs->r11 = pebs->r11;
regs->r12 = pebs->r12;
regs->r13 = pebs->r13;
regs->r14 = pebs->r14;
regs->r15 = pebs->r15;
#endif
}
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
if (event->attr.precise_ip > 1) {
/*
* Haswell and later processors have an 'eventing IP'
* (real IP) which fixes the off-by-1 skid in hardware.
* Use it when precise_ip >= 2 :
*/
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
if (x86_pmu.intel_cap.pebs_format >= 2) {
set_linear_ip(regs, pebs->real_ip);
regs->flags |= PERF_EFLAGS_EXACT;
} else {
/* Otherwise, use PEBS off-by-1 IP: */
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
set_linear_ip(regs, pebs->ip);
/*
* With precise_ip >= 2, try to fix up the off-by-1 IP
* using the LBR. If successful, the fixup function
* corrects regs->ip and calls set_linear_ip() on regs:
*/
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
if (intel_pmu_pebs_fixup_ip(regs))
regs->flags |= PERF_EFLAGS_EXACT;
}
} else {
/*
* When precise_ip == 1, return the PEBS off-by-1 IP,
* no fixup attempted:
*/
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
set_linear_ip(regs, pebs->ip);
}
perf/x86/intel: Fix linear IP of PEBS real_ip on Haswell and later CPUs this patch fix a bug in how the pebs->real_ip is handled in the PEBS handler. real_ip only exists in Haswell and later processor. It is actually the eventing IP, i.e., where the event occurred. As opposed to the pebs->ip which is the PEBS interrupt IP which is always off by one. The problem is that the real_ip just like the IP needs to be fixed up because PEBS does not record all the machine state registers, and in particular the code segement (cs). This is why we have the set_linear_ip() function. The problem was that set_linear_ip() was only used on the pebs->ip and not the pebs->real_ip. We have profiles which ran into invalid callstacks because of this. Here is an example: ..... 0: ffffffffffffff80 recent entry, marker kernel v ..... 1: 000000000040044d <= user address in kernel space! ..... 2: fffffffffffffe00 marker enter user v ..... 3: 000000000040044d ..... 4: 00000000004004b6 oldest entry Debugging output in get_perf_callchain(): [ 857.769909] CALLCHAIN: CPU8 ip=40044d regs->cs=10 user_mode(regs)=0 The problem is that the kernel entry in 1: points to a user level address. How can that be? The reason is that with PEBS sampling the instruction that caused the event to occur and the instruction where the CPU was when the interrupt was posted may be far apart. And sometime during that time window, the privilege level may change. This happens, for instance, when the PEBS sample is taken close to a kernel entry point. Here PEBS, eventing IP (real_ip) captured a user level instruction. But by the time the PMU interrupt fired, the processor had already entered kernel space. This is why the debug output shows a user address with user_mode() false. The problem comes from PEBS not recording the code segment (cs) register. The register is used in x86_64 to determine if executing in kernel vs user space. This is okay because the kernel has a software workaround called set_linear_ip(). But the issue in setup_pebs_sample_data() is that set_linear_ip() is never called on the real_ip value when it is available (Haswell and later) and precise_ip > 1. This patch fixes this problem and eliminates the callchain discrepancy. The patch restructures the code around set_linear_ip() to minimize the number of times the IP has to be set. Signed-off-by: Stephane Eranian <eranian@google.com> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: kan.liang@intel.com Link: http://lkml.kernel.org/r/1521788507-10231-1-git-send-email-eranian@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-23 00:01:47 -07:00
if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
x86_pmu.intel_cap.pebs_format >= 1) {
data->addr = pebs->dla;
data->sample_flags |= PERF_SAMPLE_ADDR;
}
if (x86_pmu.intel_cap.pebs_format >= 2) {
/* Only set the TSX weight when no memory weight. */
if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll) {
perf/core: Add PERF_SAMPLE_WEIGHT_STRUCT Current PERF_SAMPLE_WEIGHT sample type is very useful to expresses the cost of an action represented by the sample. This allows the profiler to scale the samples to be more informative to the programmer. It could also help to locate a hotspot, e.g., when profiling by memory latencies, the expensive load appear higher up in the histograms. But current PERF_SAMPLE_WEIGHT sample type is solely determined by one factor. This could be a problem, if users want two or more factors to contribute to the weight. For example, Golden Cove core PMU can provide both the instruction latency and the cache Latency information as factors for the memory profiling. For current X86 platforms, although meminfo::latency is defined as a u64, only the lower 32 bits include the valid data in practice (No memory access could last than 4G cycles). The higher 32 bits can be used to store new factors. Add a new sample type, PERF_SAMPLE_WEIGHT_STRUCT, to indicate the new sample weight structure. It shares the same space as the PERF_SAMPLE_WEIGHT sample type. Users can apply either the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type to retrieve the sample weight, but they cannot apply both sample types simultaneously. Currently, only X86 and PowerPC use the PERF_SAMPLE_WEIGHT sample type. - For PowerPC, there is nothing changed for the PERF_SAMPLE_WEIGHT sample type. There is no effect for the new PERF_SAMPLE_WEIGHT_STRUCT sample type. PowerPC can re-struct the weight field similarly later. - For X86, the same value will be dumped for the PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample type for now. The following patches will apply the new factors for the PERF_SAMPLE_WEIGHT_STRUCT sample type. The field in the union perf_sample_weight should be shared among different architectures. A generic name is required, but it's hard to abstract a name that applies to all architectures. For example, on X86, the fields are to store all kinds of latency. While on PowerPC, it stores MMCRA[TECX/TECM], which should not be latency. So a general name prefix 'var$NUM' is used here. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-2-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:07 -07:00
data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
}
if (sample_type & PERF_SAMPLE_TRANSACTION) {
data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
pebs->ax);
data->sample_flags |= PERF_SAMPLE_TRANSACTION;
}
}
/*
* v3 supplies an accurate time stamp, so we use that
* for the time stamp.
*
* We can only do this for the default trace clock.
*/
if (x86_pmu.intel_cap.pebs_format >= 3)
setup_pebs_time(event, data, pebs->tsc);
if (has_branch_stack(event))
perf: Add branch stack counters Currently, the additional information of a branch entry is stored in a u64 space. With more and more information added, the space is running out. For example, the information of occurrences of events will be added for each branch. Two places were suggested to append the counters. https://lore.kernel.org/lkml/20230802215814.GH231007@hirez.programming.kicks-ass.net/ One place is right after the flags of each branch entry. It changes the existing struct perf_branch_entry. The later ARCH specific implementation has to be really careful to consistently pick the right struct. The other place is right after the entire struct perf_branch_stack. The disadvantage is that the pointer of the extra space has to be recorded. The common interface perf_sample_save_brstack() has to be updated. The latter is much straightforward, and should be easily understood and maintained. It is implemented in the patch. Add a new branch sample type, PERF_SAMPLE_BRANCH_COUNTERS, to indicate the event which is recorded in the branch info. The "u64 counters" may store the occurrences of several events. The information regarding the number of events/counters and the width of each counter should be exposed via sysfs as a reference for the perf tool. Define the branch_counter_nr and branch_counter_width ABI here. The support will be implemented later in the Intel-specific patch. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20231025201626.3000228-1-kan.liang@linux.intel.com
2023-10-25 13:16:19 -07:00
perf_sample_save_brstack(data, event, &cpuc->lbr_stack, NULL);
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
static void adaptive_pebs_save_regs(struct pt_regs *regs,
struct pebs_gprs *gprs)
{
regs->ax = gprs->ax;
regs->bx = gprs->bx;
regs->cx = gprs->cx;
regs->dx = gprs->dx;
regs->si = gprs->si;
regs->di = gprs->di;
regs->bp = gprs->bp;
regs->sp = gprs->sp;
#ifndef CONFIG_X86_32
regs->r8 = gprs->r8;
regs->r9 = gprs->r9;
regs->r10 = gprs->r10;
regs->r11 = gprs->r11;
regs->r12 = gprs->r12;
regs->r13 = gprs->r13;
regs->r14 = gprs->r14;
regs->r15 = gprs->r15;
#endif
}
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
#define PEBS_LATENCY_MASK 0xffff
#define PEBS_CACHE_LATENCY_OFFSET 32
#define PEBS_RETIRE_LATENCY_OFFSET 32
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
/*
* With adaptive PEBS the layout depends on what fields are configured.
*/
static void setup_pebs_adaptive_sample_data(struct perf_event *event,
struct pt_regs *iregs, void *__pebs,
struct perf_sample_data *data,
struct pt_regs *regs)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct pebs_basic *basic = __pebs;
void *next_record = basic + 1;
u64 sample_type;
u64 format_size;
struct pebs_meminfo *meminfo = NULL;
struct pebs_gprs *gprs = NULL;
struct x86_perf_regs *perf_regs;
if (basic == NULL)
return;
perf_regs = container_of(regs, struct x86_perf_regs, regs);
perf_regs->xmm_regs = NULL;
sample_type = event->attr.sample_type;
format_size = basic->format_size;
perf_sample_data_init(data, 0, event->hw.last_period);
data->period = event->hw.last_period;
setup_pebs_time(event, data, basic->tsc);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
/*
* We must however always use iregs for the unwinder to stay sane; the
* record BP,SP,IP can point into thin air when the record is from a
* previous PMI context or an (I)RET happened between the record and
* PMI.
*/
if (sample_type & PERF_SAMPLE_CALLCHAIN)
perf_sample_save_callchain(data, event, iregs);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
*regs = *iregs;
/* The ip in basic is EventingIP */
set_linear_ip(regs, basic->ip);
regs->flags = PERF_EFLAGS_EXACT;
if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) {
if (x86_pmu.flags & PMU_FL_RETIRE_LATENCY)
data->weight.var3_w = format_size >> PEBS_RETIRE_LATENCY_OFFSET & PEBS_LATENCY_MASK;
else
data->weight.var3_w = 0;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
/*
* The record for MEMINFO is in front of GP
* But PERF_SAMPLE_TRANSACTION needs gprs->ax.
* Save the pointer here but process later.
*/
if (format_size & PEBS_DATACFG_MEMINFO) {
meminfo = next_record;
next_record = meminfo + 1;
}
if (format_size & PEBS_DATACFG_GP) {
gprs = next_record;
next_record = gprs + 1;
if (event->attr.precise_ip < 2) {
set_linear_ip(regs, gprs->ip);
regs->flags &= ~PERF_EFLAGS_EXACT;
}
if (sample_type & PERF_SAMPLE_REGS_INTR)
adaptive_pebs_save_regs(regs, gprs);
}
if (format_size & PEBS_DATACFG_MEMINFO) {
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
u64 weight = meminfo->latency;
if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) {
data->weight.var2_w = weight & PEBS_LATENCY_MASK;
weight >>= PEBS_CACHE_LATENCY_OFFSET;
}
/*
* Although meminfo::latency is defined as a u64,
* only the lower 32 bits include the valid data
* in practice on Ice Lake and earlier platforms.
*/
if (sample_type & PERF_SAMPLE_WEIGHT) {
data->weight.full = weight ?:
intel_get_tsx_weight(meminfo->tsx_tuning);
} else {
data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?:
intel_get_tsx_weight(meminfo->tsx_tuning);
}
data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
perf/x86/intel: Add perf core PMU support for Sapphire Rapids Add perf core PMU support for the Intel Sapphire Rapids server, which is the successor of the Intel Ice Lake server. The enabling code is based on Ice Lake, but there are several new features introduced. The event encoding is changed and simplified, e.g., the event codes which are below 0x90 are restricted to counters 0-3. The event codes which above 0x90 are likely to have no restrictions. The event constraints, extra_regs(), and hardware cache events table are changed accordingly. A new Precise Distribution (PDist) facility is introduced, which further minimizes the skid when a precise event is programmed on the GP counter 0. Enable the Precise Distribution (PDist) facility with :ppp event. For this facility to work, the period must be initialized with a value larger than 127. Add spr_limit_period() to apply the limit for :ppp event. Two new data source fields, data block & address block, are added in the PEBS Memory Info Record for the load latency event. To enable the feature, - An auxiliary event has to be enabled together with the load latency event on Sapphire Rapids. A new flag PMU_FL_MEM_LOADS_AUX is introduced to indicate the case. A new event, mem-loads-aux, is exposed to sysfs for the user tool. Add a check in hw_config(). If the auxiliary event is not detected, return an unique error -ENODATA. - The union perf_mem_data_src is extended to support the new fields. - Ice Lake and earlier models do not support block information, but the fields may be set by HW on some machines. Add pebs_no_block to explicitly indicate the previous platforms which don't support the new block fields. Accessing the new block fields are ignored on those platforms. A new store Latency facility is introduced, which leverages the PEBS facility where it can provide additional information about sampled stores. The additional information includes the data address, memory auxiliary info (e.g. Data Source, STLB miss) and the latency of the store access. To enable the facility, the new event (0x02cd) has to be programed on the GP counter 0. A new flag PERF_X86_EVENT_PEBS_STLAT is introduced to indicate the event. The store_latency_data() is introduced to parse the memory auxiliary info. The layout of access latency field of PEBS Memory Info Record has been changed. Two latency, instruction latency (bit 15:0) and cache access latency (bit 47:32) are recorded. - The cache access latency is similar to previous memory access latency. For loads, the latency starts by the actual cache access until the data is returned by the memory subsystem. For stores, the latency starts when the demand write accesses the L1 data cache and lasts until the cacheline write is completed in the memory subsystem. The cache access latency is stored in low 32bits of the sample type PERF_SAMPLE_WEIGHT_STRUCT. - The instruction latency starts by the dispatch of the load operation for execution and lasts until completion of the instruction it belongs to. Add a new flag PMU_FL_INSTR_LATENCY to indicate the instruction latency support. The instruction latency is stored in the bit 47:32 of the sample type PERF_SAMPLE_WEIGHT_STRUCT. Extends the PERF_METRICS MSR to feature TMA method level 2 metrics. The lower half of the register is the TMA level 1 metrics (legacy). The upper half is also divided into four 8-bit fields for the new level 2 metrics. Expose all eight Topdown metrics events to user space. The full description for the SPR features can be found at Intel Architecture Instruction Set Extensions and Future Features Programming Reference, 319433-041. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1611873611-156687-5-git-send-email-kan.liang@linux.intel.com
2021-01-28 15:40:10 -07:00
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (sample_type & PERF_SAMPLE_DATA_SRC) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
data->data_src.val = get_data_src(event, meminfo->aux);
data->sample_flags |= PERF_SAMPLE_DATA_SRC;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (sample_type & PERF_SAMPLE_ADDR_TYPE) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
data->addr = meminfo->address;
data->sample_flags |= PERF_SAMPLE_ADDR;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (sample_type & PERF_SAMPLE_TRANSACTION) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
gprs ? gprs->ax : 0);
data->sample_flags |= PERF_SAMPLE_TRANSACTION;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
}
if (format_size & PEBS_DATACFG_XMMS) {
struct pebs_xmm *xmm = next_record;
next_record = xmm + 1;
perf_regs->xmm_regs = xmm->xmm;
}
if (format_size & PEBS_DATACFG_LBRS) {
struct lbr_entry *lbr = next_record;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
& 0xff) + 1;
next_record = next_record + num_lbr * sizeof(struct lbr_entry);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (has_branch_stack(event)) {
intel_pmu_store_pebs_lbrs(lbr);
perf/x86/intel: Support branch counters logging The branch counters logging (A.K.A LBR event logging) introduces a per-counter indication of precise event occurrences in LBRs. It can provide a means to attribute exposed retirement latency to combinations of events across a block of instructions. It also provides a means of attributing Timed LBR latencies to events. The feature is first introduced on SRF/GRR. It is an enhancement of the ARCH LBR. It adds new fields in the LBR_INFO MSRs to log the occurrences of events on the GP counters. The information is displayed by the order of counters. The design proposed in this patch requires that the events which are logged must be in a group with the event that has LBR. If there are more than one LBR group, the counters logging information only from the current group (overflowed) are stored for the perf tool, otherwise the perf tool cannot know which and when other groups are scheduled especially when multiplexing is triggered. The user can ensure it uses the maximum number of counters that support LBR info (4 by now) by making the group large enough. The HW only logs events by the order of counters. The order may be different from the order of enabling which the perf tool can understand. When parsing the information of each branch entry, convert the counter order to the enabled order, and store the enabled order in the extension space. Unconditionally reset LBRs for an LBR event group when it's deleted. The logged counter information is only valid for the current LBR group. If another LBR group is scheduled later, the information from the stale LBRs would be otherwise wrongly interpreted. Add a sanity check in intel_pmu_hw_config(). Disable the feature if other counter filters (inv, cmask, edge, in_tx) are set or LBR call stack mode is enabled. (For the LBR call stack mode, we cannot simply flush the LBR, since it will break the call stack. Also, there is no obvious usage with the call stack mode for now.) Only applying the PERF_SAMPLE_BRANCH_COUNTERS doesn't require any branch stack setup. Expose the maximum number of supported counters and the width of the counters into the sysfs. The perf tool can use the information to parse the logged counters in each branch. Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20231025201626.3000228-5-kan.liang@linux.intel.com
2023-10-25 13:16:23 -07:00
intel_pmu_lbr_save_brstack(data, cpuc, event);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
}
}
WARN_ONCE(next_record != __pebs + (format_size >> 48),
"PEBS record size %llu, expected %llu, config %llx\n",
format_size >> 48,
(u64)(next_record - __pebs),
basic->format_size);
}
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
static inline void *
get_next_pebs_record_by_bit(void *base, void *top, int bit)
{
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
void *at;
u64 pebs_status;
/*
* fmt0 does not have a status bitfield (does not use
* perf_record_nhm format)
*/
if (x86_pmu.intel_cap.pebs_format < 1)
return base;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
if (base == NULL)
return NULL;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
for (at = base; at < top; at += cpuc->pebs_record_size) {
unsigned long status = get_pebs_status(at);
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (test_bit(bit, (unsigned long *)&status)) {
/* PEBS v3 has accurate status bits */
if (x86_pmu.intel_cap.pebs_format >= 3)
return at;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (status == (1 << bit))
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
return at;
/* clear non-PEBS bit and re-check */
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_status = status & cpuc->pebs_enabled;
perf/x86: Fix spurious NMI with PEBS Load Latency event Spurious NMIs will be observed with the following command: while :; do perf record -bae "cpu/umask=0x01,event=0xcd,ldlat=0x80/pp" -e "cpu/umask=0x03,event=0x0/" -e "cpu/umask=0x02,event=0x0/" -e cycles,branches,cache-misses -e cache-references -- sleep 10 done The bug was introduced by commit: 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+") That commit clears the status bits for the counters used for PEBS events, by masking the whole 64 bits pebs_enabled. However, only the low 32 bits of both status and pebs_enabled are reserved for PEBS-able counters. For status bits 32-34 are fixed counter overflow bits. For pebs_enabled bits 32-34 are for PEBS Load Latency. In the test case, the PEBS Load Latency event and fixed counter event could overflow at the same time. The fixed counter overflow bit will be cleared by mistake. Once it is cleared, the fixed counter overflow never be processed, which finally trigger spurious NMI. Correct the PEBS enabled mask by ignoring the non-PEBS bits. Signed-off-by: Kan Liang <kan.liang@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Fixes: 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL status on HSW+") Link: http://lkml.kernel.org/r/1491333246-3965-1-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-04-04 12:14:06 -07:00
pebs_status &= PEBS_COUNTER_MASK;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
if (pebs_status == (1 << bit))
return at;
}
}
return NULL;
}
void intel_pmu_auto_reload_read(struct perf_event *event)
{
WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
perf_pmu_disable(event->pmu);
intel_pmu_drain_pebs_buffer();
perf_pmu_enable(event->pmu);
}
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
/*
* Special variant of intel_pmu_save_and_restart() for auto-reload.
*/
static int
intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
{
struct hw_perf_event *hwc = &event->hw;
int shift = 64 - x86_pmu.cntval_bits;
u64 period = hwc->sample_period;
u64 prev_raw_count, new_raw_count;
s64 new, old;
WARN_ON(!period);
/*
* drain_pebs() only happens when the PMU is disabled.
*/
WARN_ON(this_cpu_read(cpu_hw_events.enabled));
prev_raw_count = local64_read(&hwc->prev_count);
rdpmcl(hwc->event_base_rdpmc, new_raw_count);
local64_set(&hwc->prev_count, new_raw_count);
/*
* Since the counter increments a negative counter value and
* overflows on the sign switch, giving the interval:
*
* [-period, 0]
*
* the difference between two consecutive reads is:
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
*
* A) value2 - value1;
* when no overflows have happened in between,
*
* B) (0 - value1) + (value2 - (-period));
* when one overflow happened in between,
*
* C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
* when @n overflows happened in between.
*
* Here A) is the obvious difference, B) is the extension to the
* discrete interval, where the first term is to the top of the
* interval and the second term is from the bottom of the next
* interval and C) the extension to multiple intervals, where the
* middle term is the whole intervals covered.
*
* An equivalent of C, by reduction, is:
*
* value2 - value1 + n * period
*/
new = ((s64)(new_raw_count << shift) >> shift);
old = ((s64)(prev_raw_count << shift) >> shift);
local64_add(new - old + count * period, &event->count);
perf/x86/intel: Fix inaccurate period in context switch for auto-reload Perf doesn't take the left period into account when auto-reload is enabled with fixed period sampling mode in context switch. Here is the MSR trace of the perf command as below. (The MSR trace is simplified from a ftrace log.) #perf record -e cycles:p -c 2000000 -- ./triad_loop //The MSR trace of task schedule out //perf disable all counters, disable PEBS, disable GP counter 0, //read GP counter 0, and re-enable all counters. //The counter 0 stops at 0xfffffff82840 write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 0 write_msr: MSR_IA32_PEBS_ENABLE(3f1), value 0 write_msr: MSR_P6_EVNTSEL0(186), value 40003003c rdpmc: 0, value fffffff82840 write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value f000000ff //The MSR trace of the same task schedule in again //perf disable all counters, enable and set GP counter 0, //enable PEBS, and re-enable all counters. //0xffffffe17b80 (-2000000) is written to GP counter 0. write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 0 write_msr: MSR_IA32_PMC0(4c1), value ffffffe17b80 write_msr: MSR_P6_EVNTSEL0(186), value 40043003c write_msr: MSR_IA32_PEBS_ENABLE(3f1), value 1 write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value f000000ff When the same task schedule in again, the counter should starts from previous left. However, it starts from the fixed period -2000000 again. A special variant of intel_pmu_save_and_restart() is used for auto-reload, which doesn't update the hwc->period_left. When the monitored task schedules in again, perf doesn't know the left period. The fixed period is used, which is inaccurate. With auto-reload, the counter always has a negative counter value. So the left period is -value. Update the period_left in intel_pmu_save_and_restart_reload(). With the patch: //The MSR trace of task schedule out write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 0 write_msr: MSR_IA32_PEBS_ENABLE(3f1), value 0 write_msr: MSR_P6_EVNTSEL0(186), value 40003003c rdpmc: 0, value ffffffe25cbc write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value f000000ff //The MSR trace of the same task schedule in again write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value 0 write_msr: MSR_IA32_PMC0(4c1), value ffffffe25cbc write_msr: MSR_P6_EVNTSEL0(186), value 40043003c write_msr: MSR_IA32_PEBS_ENABLE(3f1), value 1 write_msr: MSR_CORE_PERF_GLOBAL_CTRL(38f), value f000000ff Fixes: d31fc13fdcb2 ("perf/x86/intel: Fix event update for auto-reload") Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Link: https://lkml.kernel.org/r/20200121190125.3389-1-kan.liang@linux.intel.com
2020-01-21 12:01:25 -07:00
local64_set(&hwc->period_left, -new);
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
perf_event_update_userpage(event);
return 0;
}
static __always_inline void
__intel_pmu_pebs_event(struct perf_event *event,
struct pt_regs *iregs,
struct perf_sample_data *data,
void *base, void *top,
int bit, int count,
void (*setup_sample)(struct perf_event *,
struct pt_regs *,
void *,
struct perf_sample_data *,
struct pt_regs *))
{
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
struct hw_perf_event *hwc = &event->hw;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
struct x86_perf_regs perf_regs;
struct pt_regs *regs = &perf_regs.regs;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
void *at = get_next_pebs_record_by_bit(base, top, bit);
static struct pt_regs dummy_iregs;
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
/*
* Now, auto-reload is only enabled in fixed period mode.
* The reload value is always hwc->sample_period.
* May need to change it, if auto-reload is enabled in
* freq mode later.
*/
intel_pmu_save_and_restart_reload(event, count);
} else if (!intel_pmu_save_and_restart(event))
return;
perf/x86/intel/ds: Fix x86_pmu_stop warning for large PEBS A warning as below may be triggered when sampling with large PEBS. [ 410.411250] perf: interrupt took too long (72145 > 71975), lowering kernel.perf_event_max_sample_rate to 2000 [ 410.724923] ------------[ cut here ]------------ [ 410.729822] WARNING: CPU: 0 PID: 16397 at arch/x86/events/core.c:1422 x86_pmu_stop+0x95/0xa0 [ 410.933811] x86_pmu_del+0x50/0x150 [ 410.937304] event_sched_out.isra.0+0xbc/0x210 [ 410.941751] group_sched_out.part.0+0x53/0xd0 [ 410.946111] ctx_sched_out+0x193/0x270 [ 410.949862] __perf_event_task_sched_out+0x32c/0x890 [ 410.954827] ? set_next_entity+0x98/0x2d0 [ 410.958841] __schedule+0x592/0x9c0 [ 410.962332] schedule+0x5f/0xd0 [ 410.965477] exit_to_usermode_loop+0x73/0x120 [ 410.969837] prepare_exit_to_usermode+0xcd/0xf0 [ 410.974369] ret_from_intr+0x2a/0x3a [ 410.977946] RIP: 0033:0x40123c [ 411.079661] ---[ end trace bc83adaea7bb664a ]--- In the non-overflow context, e.g., context switch, with large PEBS, perf may stop an event twice. An example is below. //max_samples_per_tick is adjusted to 2 //NMI is triggered intel_pmu_handle_irq() handle_pmi_common() drain_pebs() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() hwc->interrupts = 1 return 0 //A context switch happens right after the NMI. //In the same tick, the perf_throttled_seq is not changed. perf_event_task_sched_out() perf_pmu_sched_task() intel_pmu_drain_pebs_buffer() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() ++hwc->interrupts >= max_samples_per_tick return 1 x86_pmu_stop(); # First stop perf_event_context_sched_out() task_ctx_sched_out() ctx_sched_out() event_sched_out() x86_pmu_del() x86_pmu_stop(); # Second stop and trigger the warning Perf should only invoke the perf_event_overflow() in the overflow context. Current drain_pebs() is called from: - handle_pmi_common() -- overflow context - intel_pmu_pebs_sched_task() -- non-overflow context - intel_pmu_pebs_disable() -- non-overflow context - intel_pmu_auto_reload_read() -- possible overflow context With PERF_SAMPLE_READ + PERF_FORMAT_GROUP, the function may be invoked in the NMI handler. But, before calling the function, the PEBS buffer has already been drained. The __intel_pmu_pebs_event() will not be called in the possible overflow context. To fix the issue, an indicator is required to distinguish between the overflow context aka handle_pmi_common() and other cases. The dummy regs pointer can be used as the indicator. In the non-overflow context, perf should treat the last record the same as other PEBS records, and doesn't invoke the generic overflow handler. Fixes: 21509084f999 ("perf/x86/intel: Handle multiple records in the PEBS buffer") Reported-by: Like Xu <like.xu@linux.intel.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Like Xu <like.xu@linux.intel.com> Link: https://lkml.kernel.org/r/20200902210649.2743-1-kan.liang@linux.intel.com
2020-09-02 14:06:49 -07:00
if (!iregs)
iregs = &dummy_iregs;
while (count > 1) {
setup_sample(event, iregs, at, data, regs);
perf_event_output(event, data, regs);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
at += cpuc->pebs_record_size;
at = get_next_pebs_record_by_bit(at, top, bit);
count--;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
}
setup_sample(event, iregs, at, data, regs);
perf/x86/intel/ds: Fix x86_pmu_stop warning for large PEBS A warning as below may be triggered when sampling with large PEBS. [ 410.411250] perf: interrupt took too long (72145 > 71975), lowering kernel.perf_event_max_sample_rate to 2000 [ 410.724923] ------------[ cut here ]------------ [ 410.729822] WARNING: CPU: 0 PID: 16397 at arch/x86/events/core.c:1422 x86_pmu_stop+0x95/0xa0 [ 410.933811] x86_pmu_del+0x50/0x150 [ 410.937304] event_sched_out.isra.0+0xbc/0x210 [ 410.941751] group_sched_out.part.0+0x53/0xd0 [ 410.946111] ctx_sched_out+0x193/0x270 [ 410.949862] __perf_event_task_sched_out+0x32c/0x890 [ 410.954827] ? set_next_entity+0x98/0x2d0 [ 410.958841] __schedule+0x592/0x9c0 [ 410.962332] schedule+0x5f/0xd0 [ 410.965477] exit_to_usermode_loop+0x73/0x120 [ 410.969837] prepare_exit_to_usermode+0xcd/0xf0 [ 410.974369] ret_from_intr+0x2a/0x3a [ 410.977946] RIP: 0033:0x40123c [ 411.079661] ---[ end trace bc83adaea7bb664a ]--- In the non-overflow context, e.g., context switch, with large PEBS, perf may stop an event twice. An example is below. //max_samples_per_tick is adjusted to 2 //NMI is triggered intel_pmu_handle_irq() handle_pmi_common() drain_pebs() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() hwc->interrupts = 1 return 0 //A context switch happens right after the NMI. //In the same tick, the perf_throttled_seq is not changed. perf_event_task_sched_out() perf_pmu_sched_task() intel_pmu_drain_pebs_buffer() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() ++hwc->interrupts >= max_samples_per_tick return 1 x86_pmu_stop(); # First stop perf_event_context_sched_out() task_ctx_sched_out() ctx_sched_out() event_sched_out() x86_pmu_del() x86_pmu_stop(); # Second stop and trigger the warning Perf should only invoke the perf_event_overflow() in the overflow context. Current drain_pebs() is called from: - handle_pmi_common() -- overflow context - intel_pmu_pebs_sched_task() -- non-overflow context - intel_pmu_pebs_disable() -- non-overflow context - intel_pmu_auto_reload_read() -- possible overflow context With PERF_SAMPLE_READ + PERF_FORMAT_GROUP, the function may be invoked in the NMI handler. But, before calling the function, the PEBS buffer has already been drained. The __intel_pmu_pebs_event() will not be called in the possible overflow context. To fix the issue, an indicator is required to distinguish between the overflow context aka handle_pmi_common() and other cases. The dummy regs pointer can be used as the indicator. In the non-overflow context, perf should treat the last record the same as other PEBS records, and doesn't invoke the generic overflow handler. Fixes: 21509084f999 ("perf/x86/intel: Handle multiple records in the PEBS buffer") Reported-by: Like Xu <like.xu@linux.intel.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Like Xu <like.xu@linux.intel.com> Link: https://lkml.kernel.org/r/20200902210649.2743-1-kan.liang@linux.intel.com
2020-09-02 14:06:49 -07:00
if (iregs == &dummy_iregs) {
/*
* The PEBS records may be drained in the non-overflow context,
* e.g., large PEBS + context switch. Perf should treat the
* last record the same as other PEBS records, and doesn't
* invoke the generic overflow handler.
*/
perf_event_output(event, data, regs);
perf/x86/intel/ds: Fix x86_pmu_stop warning for large PEBS A warning as below may be triggered when sampling with large PEBS. [ 410.411250] perf: interrupt took too long (72145 > 71975), lowering kernel.perf_event_max_sample_rate to 2000 [ 410.724923] ------------[ cut here ]------------ [ 410.729822] WARNING: CPU: 0 PID: 16397 at arch/x86/events/core.c:1422 x86_pmu_stop+0x95/0xa0 [ 410.933811] x86_pmu_del+0x50/0x150 [ 410.937304] event_sched_out.isra.0+0xbc/0x210 [ 410.941751] group_sched_out.part.0+0x53/0xd0 [ 410.946111] ctx_sched_out+0x193/0x270 [ 410.949862] __perf_event_task_sched_out+0x32c/0x890 [ 410.954827] ? set_next_entity+0x98/0x2d0 [ 410.958841] __schedule+0x592/0x9c0 [ 410.962332] schedule+0x5f/0xd0 [ 410.965477] exit_to_usermode_loop+0x73/0x120 [ 410.969837] prepare_exit_to_usermode+0xcd/0xf0 [ 410.974369] ret_from_intr+0x2a/0x3a [ 410.977946] RIP: 0033:0x40123c [ 411.079661] ---[ end trace bc83adaea7bb664a ]--- In the non-overflow context, e.g., context switch, with large PEBS, perf may stop an event twice. An example is below. //max_samples_per_tick is adjusted to 2 //NMI is triggered intel_pmu_handle_irq() handle_pmi_common() drain_pebs() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() hwc->interrupts = 1 return 0 //A context switch happens right after the NMI. //In the same tick, the perf_throttled_seq is not changed. perf_event_task_sched_out() perf_pmu_sched_task() intel_pmu_drain_pebs_buffer() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() ++hwc->interrupts >= max_samples_per_tick return 1 x86_pmu_stop(); # First stop perf_event_context_sched_out() task_ctx_sched_out() ctx_sched_out() event_sched_out() x86_pmu_del() x86_pmu_stop(); # Second stop and trigger the warning Perf should only invoke the perf_event_overflow() in the overflow context. Current drain_pebs() is called from: - handle_pmi_common() -- overflow context - intel_pmu_pebs_sched_task() -- non-overflow context - intel_pmu_pebs_disable() -- non-overflow context - intel_pmu_auto_reload_read() -- possible overflow context With PERF_SAMPLE_READ + PERF_FORMAT_GROUP, the function may be invoked in the NMI handler. But, before calling the function, the PEBS buffer has already been drained. The __intel_pmu_pebs_event() will not be called in the possible overflow context. To fix the issue, an indicator is required to distinguish between the overflow context aka handle_pmi_common() and other cases. The dummy regs pointer can be used as the indicator. In the non-overflow context, perf should treat the last record the same as other PEBS records, and doesn't invoke the generic overflow handler. Fixes: 21509084f999 ("perf/x86/intel: Handle multiple records in the PEBS buffer") Reported-by: Like Xu <like.xu@linux.intel.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Like Xu <like.xu@linux.intel.com> Link: https://lkml.kernel.org/r/20200902210649.2743-1-kan.liang@linux.intel.com
2020-09-02 14:06:49 -07:00
} else {
/*
* All but the last records are processed.
* The last one is left to be able to call the overflow handler.
*/
if (perf_event_overflow(event, data, regs))
perf/x86/intel/ds: Fix x86_pmu_stop warning for large PEBS A warning as below may be triggered when sampling with large PEBS. [ 410.411250] perf: interrupt took too long (72145 > 71975), lowering kernel.perf_event_max_sample_rate to 2000 [ 410.724923] ------------[ cut here ]------------ [ 410.729822] WARNING: CPU: 0 PID: 16397 at arch/x86/events/core.c:1422 x86_pmu_stop+0x95/0xa0 [ 410.933811] x86_pmu_del+0x50/0x150 [ 410.937304] event_sched_out.isra.0+0xbc/0x210 [ 410.941751] group_sched_out.part.0+0x53/0xd0 [ 410.946111] ctx_sched_out+0x193/0x270 [ 410.949862] __perf_event_task_sched_out+0x32c/0x890 [ 410.954827] ? set_next_entity+0x98/0x2d0 [ 410.958841] __schedule+0x592/0x9c0 [ 410.962332] schedule+0x5f/0xd0 [ 410.965477] exit_to_usermode_loop+0x73/0x120 [ 410.969837] prepare_exit_to_usermode+0xcd/0xf0 [ 410.974369] ret_from_intr+0x2a/0x3a [ 410.977946] RIP: 0033:0x40123c [ 411.079661] ---[ end trace bc83adaea7bb664a ]--- In the non-overflow context, e.g., context switch, with large PEBS, perf may stop an event twice. An example is below. //max_samples_per_tick is adjusted to 2 //NMI is triggered intel_pmu_handle_irq() handle_pmi_common() drain_pebs() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() hwc->interrupts = 1 return 0 //A context switch happens right after the NMI. //In the same tick, the perf_throttled_seq is not changed. perf_event_task_sched_out() perf_pmu_sched_task() intel_pmu_drain_pebs_buffer() __intel_pmu_pebs_event() perf_event_overflow() __perf_event_account_interrupt() ++hwc->interrupts >= max_samples_per_tick return 1 x86_pmu_stop(); # First stop perf_event_context_sched_out() task_ctx_sched_out() ctx_sched_out() event_sched_out() x86_pmu_del() x86_pmu_stop(); # Second stop and trigger the warning Perf should only invoke the perf_event_overflow() in the overflow context. Current drain_pebs() is called from: - handle_pmi_common() -- overflow context - intel_pmu_pebs_sched_task() -- non-overflow context - intel_pmu_pebs_disable() -- non-overflow context - intel_pmu_auto_reload_read() -- possible overflow context With PERF_SAMPLE_READ + PERF_FORMAT_GROUP, the function may be invoked in the NMI handler. But, before calling the function, the PEBS buffer has already been drained. The __intel_pmu_pebs_event() will not be called in the possible overflow context. To fix the issue, an indicator is required to distinguish between the overflow context aka handle_pmi_common() and other cases. The dummy regs pointer can be used as the indicator. In the non-overflow context, perf should treat the last record the same as other PEBS records, and doesn't invoke the generic overflow handler. Fixes: 21509084f999 ("perf/x86/intel: Handle multiple records in the PEBS buffer") Reported-by: Like Xu <like.xu@linux.intel.com> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Like Xu <like.xu@linux.intel.com> Link: https://lkml.kernel.org/r/20200902210649.2743-1-kan.liang@linux.intel.com
2020-09-02 14:06:49 -07:00
x86_pmu_stop(event, 0);
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
}
}
static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct debug_store *ds = cpuc->ds;
struct perf_event *event = cpuc->events[0]; /* PMC0 only */
struct pebs_record_core *at, *top;
int n;
if (!x86_pmu.pebs_active)
return;
at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
/*
* Whatever else happens, drain the thing
*/
ds->pebs_index = ds->pebs_buffer_base;
if (!test_bit(0, cpuc->active_mask))
return;
WARN_ON_ONCE(!event);
if (!event->attr.precise_ip)
return;
n = top - at;
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
if (n <= 0) {
if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
intel_pmu_save_and_restart_reload(event, 0);
return;
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
}
__intel_pmu_pebs_event(event, iregs, data, at, top, 0, n,
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
setup_pebs_fixed_sample_data);
}
static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
{
struct perf_event *event;
int bit;
/*
* The drain_pebs() could be called twice in a short period
* for auto-reload event in pmu::read(). There are no
* overflows have happened in between.
* It needs to call intel_pmu_save_and_restart_reload() to
* update the event->count for this case.
*/
for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
event = cpuc->events[bit];
if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
intel_pmu_save_and_restart_reload(event, 0);
}
}
static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
{
x86: Replace __get_cpu_var uses __get_cpu_var() is used for multiple purposes in the kernel source. One of them is address calculation via the form &__get_cpu_var(x). This calculates the address for the instance of the percpu variable of the current processor based on an offset. Other use cases are for storing and retrieving data from the current processors percpu area. __get_cpu_var() can be used as an lvalue when writing data or on the right side of an assignment. __get_cpu_var() is defined as : #define __get_cpu_var(var) (*this_cpu_ptr(&(var))) __get_cpu_var() always only does an address determination. However, store and retrieve operations could use a segment prefix (or global register on other platforms) to avoid the address calculation. this_cpu_write() and this_cpu_read() can directly take an offset into a percpu area and use optimized assembly code to read and write per cpu variables. This patch converts __get_cpu_var into either an explicit address calculation using this_cpu_ptr() or into a use of this_cpu operations that use the offset. Thereby address calculations are avoided and less registers are used when code is generated. Transformations done to __get_cpu_var() 1. Determine the address of the percpu instance of the current processor. DEFINE_PER_CPU(int, y); int *x = &__get_cpu_var(y); Converts to int *x = this_cpu_ptr(&y); 2. Same as #1 but this time an array structure is involved. DEFINE_PER_CPU(int, y[20]); int *x = __get_cpu_var(y); Converts to int *x = this_cpu_ptr(y); 3. Retrieve the content of the current processors instance of a per cpu variable. DEFINE_PER_CPU(int, y); int x = __get_cpu_var(y) Converts to int x = __this_cpu_read(y); 4. Retrieve the content of a percpu struct DEFINE_PER_CPU(struct mystruct, y); struct mystruct x = __get_cpu_var(y); Converts to memcpy(&x, this_cpu_ptr(&y), sizeof(x)); 5. Assignment to a per cpu variable DEFINE_PER_CPU(int, y) __get_cpu_var(y) = x; Converts to __this_cpu_write(y, x); 6. Increment/Decrement etc of a per cpu variable DEFINE_PER_CPU(int, y); __get_cpu_var(y)++ Converts to __this_cpu_inc(y) Cc: Thomas Gleixner <tglx@linutronix.de> Cc: x86@kernel.org Acked-by: H. Peter Anvin <hpa@linux.intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-08-17 10:30:40 -07:00
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct debug_store *ds = cpuc->ds;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
struct perf_event *event;
void *base, *at, *top;
short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
int max_pebs_events = intel_pmu_max_num_pebs(NULL);
int bit, i, size;
u64 mask;
if (!x86_pmu.pebs_active)
return;
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
ds->pebs_index = ds->pebs_buffer_base;
mask = x86_pmu.pebs_events_mask;
size = max_pebs_events;
if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
mask |= x86_pmu.fixed_cntr_mask64 << INTEL_PMC_IDX_FIXED;
size = INTEL_PMC_IDX_FIXED + x86_pmu_max_num_counters_fixed(NULL);
}
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
if (unlikely(base >= top)) {
intel_pmu_pebs_event_update_no_drain(cpuc, size);
return;
perf/x86/intel: Fix event update for auto-reload There is a bug when reading event->count with large PEBS enabled. Here is an example: # ./read_count 0x71f0 0x122c0 0x1000000001c54 0x100000001257d 0x200000000bdc5 In fixed period mode, the auto-reload mechanism could be enabled for PEBS events, but the calculation of event->count does not take the auto-reload values into account. Anyone who reads event->count will get the wrong result, e.g x86_pmu_read(). This bug was introduced with the auto-reload mechanism enabled since commit: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Introduce intel_pmu_save_and_restart_reload() to calculate the event->count only for auto-reload. Since the counter increments a negative counter value and overflows on the sign switch, giving the interval: [-period, 0] the difference between two consequtive reads is: A) value2 - value1; when no overflows have happened in between, B) (0 - value1) + (value2 - (-period)); when one overflow happened in between, C) (0 - value1) + (n - 1) * (period) + (value2 - (-period)); when @n overflows happened in between. Here A) is the obvious difference, B) is the extension to the discrete interval, where the first term is to the top of the interval and the second term is from the bottom of the next interval and C) the extension to multiple intervals, where the middle term is the whole intervals covered. The equation for all cases is: value2 - value1 + n * period Previously the event->count is updated right before the sample output. But for case A, there is no PEBS record ready. It needs to be specially handled. Remove the auto-reload code from x86_perf_event_set_period() since we'll not longer call that function in this case. Based-on-code-from: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Fixes: 851559e35fd5 ("perf/x86/intel: Use the PEBS auto reload mechanism when possible") Link: http://lkml.kernel.org/r/1518474035-21006-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-12 15:20:31 -07:00
}
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
for (at = base; at < top; at += x86_pmu.pebs_record_size) {
struct pebs_record_nhm *p = at;
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
u64 pebs_status;
pebs_status = p->status & cpuc->pebs_enabled;
pebs_status &= mask;
/* PEBS v3 has more accurate status bits */
if (x86_pmu.intel_cap.pebs_format >= 3) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
counts[bit]++;
continue;
}
/*
* On some CPUs the PEBS status can be zero when PEBS is
* racing with clearing of GLOBAL_STATUS.
*
* Normally we would drop that record, but in the
* case when there is only a single active PEBS event
* we can assume it's for that event.
*/
if (!pebs_status && cpuc->pebs_enabled &&
!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
pebs_status = p->status = cpuc->pebs_enabled;
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
bit = find_first_bit((unsigned long *)&pebs_status,
max_pebs_events);
if (!(x86_pmu.pebs_events_mask & (1 << bit)))
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
continue;
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
/*
* The PEBS hardware does not deal well with the situation
* when events happen near to each other and multiple bits
* are set. But it should happen rarely.
*
* If these events include one PEBS and multiple non-PEBS
* events, it doesn't impact PEBS record. The record will
* be handled normally. (slow path)
*
* If these events include two or more PEBS events, the
* records for the events can be collapsed into a single
* one, and it's not possible to reconstruct all events
* that caused the PEBS record. It's called collision.
* If collision happened, the record will be dropped.
*/
if (pebs_status != (1ULL << bit)) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
for_each_set_bit(i, (unsigned long *)&pebs_status, size)
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
error[i]++;
continue;
}
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
counts[bit]++;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
for_each_set_bit(bit, (unsigned long *)&mask, size) {
if ((counts[bit] == 0) && (error[bit] == 0))
continue;
perf/x86/intel/pebs: Robustify PEBS buffer drain Vince Weaver and Stephane Eranian reported warnings in the PEBS code when running the perf fuzzer. Stephane wrote: > I can reproduce the problem on my HSW running the fuzzer. > > I can see why this could be happening if you are mixing PEBS and non PEBS events > in the bottom 4 counters. I suspect: > for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) { > if ((counts[bit] == 0) && (error[bit] == 0)) > continue; > > This test is not correct when you have non-PEBS events mixed with > PEBS events and they overflow at the same time. They will have > counts[i] != 0 but error[i] == 0, and thus you fall thru the loop > and hit the assert. Or it is something along those lines. The only way I can make this work is if ->status only has !PEBS events set, because if it has both set we'll take that slow path which masks out the !PEBS bits. After masking there are 3 options: - there is one bit set, and its @bit, we increment counts[bit]. - there are multiple bits set, we increment error[] for each set bit, we do not increment counts[]. - there are no bits set, we do nothing. The intent was to never increment counts[] for !PEBS events. Now if we start out with only a single !PEBS event set, we'll pass the test and increment counts[] for a !PEBS and hit the warn. Reported-by: Vince Weaver <vincent.weaver@maine.edu> Reported-by: Stephane Eranian <eranian@google.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 05:35:46 -07:00
perf/x86/intel: Handle multiple records in the PEBS buffer When the PEBS interrupt threshold is larger than one record and the machine supports multiple PEBS events, the records of these events are mixed up and we need to demultiplex them. Demuxing the records is hard because the hardware is deficient. The hardware has two issues that, when combined, create impossible scenarios to demux. The first issue is that the 'status' field of the PEBS record is a copy of the GLOBAL_STATUS MSR at PEBS assist time. To see why this is a problem let us first describe the regular PEBS cycle: A) the CTRn value reaches 0: - the corresponding bit in GLOBAL_STATUS gets set - we start arming the hardware assist < some unspecified amount of time later -- this could cover multiple events of interest > B) the hardware assist is armed, any next event will trigger it C) a matching event happens: - the hardware assist triggers and generates a PEBS record this includes a copy of GLOBAL_STATUS at this moment - if we auto-reload we (re)set CTRn - we clear the relevant bit in GLOBAL_STATUS Now consider the following chain of events: A0, B0, A1, C0 The event generated for counter 0 will include a status with counter 1 set, even though its not at all related to the record. A similar thing can happen with a !PEBS event if it just happens to overflow at the right moment. The second issue is that the hardware will only emit one record for two or more counters if the event that triggers the assist is 'close'. The 'close' can be several cycles. In some cases even the complete assist, if the event is something that doesn't need retirement. For instance, consider this chain of events: A0, B0, A1, B1, C01 Where C01 is an event that triggers both hardware assists, we will generate but a single record, but again with both counters listed in the status field. This time the record pertains to both events. Note that these two cases are different but undistinguishable with the data as generated. Therefore demuxing records with multiple PEBS bits (we can safely ignore status bits for !PEBS counters) is impossible. Furthermore we cannot emit the record to both events because that might cause a data leak -- the events might not have the same privileges -- so what this patch does is discard such events. The assumption/hope is that such discards will be rare. Here lists some possible ways you may get high discard rate. - when you count the same thing multiple times. But it is not a useful configuration. - you can be unfortunate if you measure with a userspace only PEBS event along with either a kernel or unrestricted PEBS event. Imagine the event triggering and setting the overflow flag right before entering the kernel. Then all kernel side events will end up with multiple bits set. Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com> Signed-off-by: Kan Liang <kan.liang@intel.com> [ Changelog improvements. ] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: acme@infradead.org Cc: eranian@google.com Link: http://lkml.kernel.org/r/1430940834-8964-4-git-send-email-kan.liang@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-06 12:33:49 -07:00
event = cpuc->events[bit];
if (WARN_ON_ONCE(!event))
continue;
if (WARN_ON_ONCE(!event->attr.precise_ip))
continue;
/* log dropped samples number */
perf/x86/intel: Account interrupts for PEBS errors It's possible to set up PEBS events to get only errors and not any data, like on SNB-X (model 45) and IVB-EP (model 62) via 2 perf commands running simultaneously: taskset -c 1 ./perf record -c 4 -e branches:pp -j any -C 10 This leads to a soft lock up, because the error path of the intel_pmu_drain_pebs_nhm() does not account event->hw.interrupt for error PEBS interrupts, so in case you're getting ONLY errors you don't have a way to stop the event when it's over the max_samples_per_tick limit: NMI watchdog: BUG: soft lockup - CPU#22 stuck for 22s! [perf_fuzzer:5816] ... RIP: 0010:[<ffffffff81159232>] [<ffffffff81159232>] smp_call_function_single+0xe2/0x140 ... Call Trace: ? trace_hardirqs_on_caller+0xf5/0x1b0 ? perf_cgroup_attach+0x70/0x70 perf_install_in_context+0x199/0x1b0 ? ctx_resched+0x90/0x90 SYSC_perf_event_open+0x641/0xf90 SyS_perf_event_open+0x9/0x10 do_syscall_64+0x6c/0x1f0 entry_SYSCALL64_slow_path+0x25/0x25 Add perf_event_account_interrupt() which does the interrupt and frequency checks and call it from intel_pmu_drain_pebs_nhm()'s error path. We keep the pending_kill and pending_wakeup logic only in the __perf_event_overflow() path, because they make sense only if there's any data to deliver. Signed-off-by: Jiri Olsa <jolsa@kernel.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vince@deater.net> Cc: Vince Weaver <vincent.weaver@maine.edu> Link: http://lkml.kernel.org/r/1482931866-6018-2-git-send-email-jolsa@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-28 06:31:03 -07:00
if (error[bit]) {
perf_log_lost_samples(event, error[bit]);
if (iregs && perf_event_account_interrupt(event))
perf/x86/intel: Account interrupts for PEBS errors It's possible to set up PEBS events to get only errors and not any data, like on SNB-X (model 45) and IVB-EP (model 62) via 2 perf commands running simultaneously: taskset -c 1 ./perf record -c 4 -e branches:pp -j any -C 10 This leads to a soft lock up, because the error path of the intel_pmu_drain_pebs_nhm() does not account event->hw.interrupt for error PEBS interrupts, so in case you're getting ONLY errors you don't have a way to stop the event when it's over the max_samples_per_tick limit: NMI watchdog: BUG: soft lockup - CPU#22 stuck for 22s! [perf_fuzzer:5816] ... RIP: 0010:[<ffffffff81159232>] [<ffffffff81159232>] smp_call_function_single+0xe2/0x140 ... Call Trace: ? trace_hardirqs_on_caller+0xf5/0x1b0 ? perf_cgroup_attach+0x70/0x70 perf_install_in_context+0x199/0x1b0 ? ctx_resched+0x90/0x90 SYSC_perf_event_open+0x641/0xf90 SyS_perf_event_open+0x9/0x10 do_syscall_64+0x6c/0x1f0 entry_SYSCALL64_slow_path+0x25/0x25 Add perf_event_account_interrupt() which does the interrupt and frequency checks and call it from intel_pmu_drain_pebs_nhm()'s error path. We keep the pending_kill and pending_wakeup logic only in the __perf_event_overflow() path, because they make sense only if there's any data to deliver. Signed-off-by: Jiri Olsa <jolsa@kernel.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vince@deater.net> Cc: Vince Weaver <vincent.weaver@maine.edu> Link: http://lkml.kernel.org/r/1482931866-6018-2-git-send-email-jolsa@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-12-28 06:31:03 -07:00
x86_pmu_stop(event, 0);
}
if (counts[bit]) {
__intel_pmu_pebs_event(event, iregs, data, base,
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
top, bit, counts[bit],
setup_pebs_fixed_sample_data);
}
}
}
static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
{
short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
struct debug_store *ds = cpuc->ds;
struct perf_event *event;
void *base, *at, *top;
int bit;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
u64 mask;
if (!x86_pmu.pebs_active)
return;
base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
ds->pebs_index = ds->pebs_buffer_base;
mask = hybrid(cpuc->pmu, pebs_events_mask) |
(hybrid(cpuc->pmu, fixed_cntr_mask64) << INTEL_PMC_IDX_FIXED);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (unlikely(base >= top)) {
intel_pmu_pebs_event_update_no_drain(cpuc, X86_PMC_IDX_MAX);
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
return;
}
for (at = base; at < top; at += cpuc->pebs_record_size) {
u64 pebs_status;
pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
pebs_status &= mask;
for_each_set_bit(bit, (unsigned long *)&pebs_status, X86_PMC_IDX_MAX)
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
counts[bit]++;
}
for_each_set_bit(bit, (unsigned long *)&mask, X86_PMC_IDX_MAX) {
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (counts[bit] == 0)
continue;
event = cpuc->events[bit];
if (WARN_ON_ONCE(!event))
continue;
if (WARN_ON_ONCE(!event->attr.precise_ip))
continue;
__intel_pmu_pebs_event(event, iregs, data, base,
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
top, bit, counts[bit],
setup_pebs_adaptive_sample_data);
}
}
/*
* BTS, PEBS probe and setup
*/
void __init intel_ds_init(void)
{
/*
* No support for 32bit formats
*/
if (!boot_cpu_has(X86_FEATURE_DTES64))
return;
x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
if (x86_pmu.version <= 4)
perf/x86/kvm: Avoid unnecessary work in guest filtering KVM added a workaround for PEBS events leaking into guests with commit: 26a4f3c08de4 ("perf/x86: disable PEBS on a guest entry.") This uses the VT entry/exit list to add an extra disable of the PEBS_ENABLE MSR. Intel also added a fix for this issue to microcode updates on Haswell/Broadwell/Skylake. It turns out using the MSR entry/exit list makes VM exits significantly slower. The list is only needed for disabling PEBS, because the GLOBAL_CTRL change gets optimized by KVM into changing the VMCS. Check for the microcode updates that have the microcode fix for leaking PEBS, and disable the extra entry/exit list entry for PEBS_ENABLE. In addition we always clear the GLOBAL_CTRL for the PEBS counter while running in the guest, which is enough to make them never fire at the wrong side of the host/guest transition. The overhead for VM exits with the filtering active with the patch is reduced from 8% to 4%. The microcode patch has already been merged into future platforms. This patch is one-off thing. The quirks is used here. For other old platforms which doesn't have microcode patch and quirks, extra disable of the PEBS_ENABLE MSR is still required. Signed-off-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: David Ahern <dsahern@gmail.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Namhyung Kim <namhyung@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: bp@alien8.de Link: https://lkml.kernel.org/r/1549319013-4522-2-git-send-email-kan.liang@linux.intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-04 15:23:30 -07:00
x86_pmu.pebs_no_isolation = 1;
if (x86_pmu.pebs) {
char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
char *pebs_qual = "";
int format = x86_pmu.intel_cap.pebs_format;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
if (format < 4)
x86_pmu.intel_cap.pebs_baseline = 0;
switch (format) {
case 0:
pr_cont("PEBS fmt0%c, ", pebs_type);
x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
/*
* Using >PAGE_SIZE buffers makes the WRMSR to
* PERF_GLOBAL_CTRL in intel_pmu_enable_all()
* mysteriously hang on Core2.
*
* As a workaround, we don't do this.
*/
x86_pmu.pebs_buffer_size = PAGE_SIZE;
x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
break;
case 1:
pr_cont("PEBS fmt1%c, ", pebs_type);
x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
break;
case 2:
pr_cont("PEBS fmt2%c, ", pebs_type);
x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
break;
case 3:
pr_cont("PEBS fmt3%c, ", pebs_type);
x86_pmu.pebs_record_size =
sizeof(struct pebs_record_skl);
x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
break;
case 5:
x86_pmu.pebs_ept = 1;
fallthrough;
case 4:
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
if (x86_pmu.intel_cap.pebs_baseline) {
x86_pmu.large_pebs_flags |=
PERF_SAMPLE_BRANCH_STACK |
PERF_SAMPLE_TIME;
x86_pmu.flags |= PMU_FL_PEBS_ALL;
x86_pmu.pebs_capable = ~0ULL;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
pebs_qual = "-baseline";
x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
} else {
/* Only basic record supported */
x86_pmu.large_pebs_flags &=
~(PERF_SAMPLE_ADDR |
PERF_SAMPLE_TIME |
PERF_SAMPLE_DATA_SRC |
PERF_SAMPLE_TRANSACTION |
PERF_SAMPLE_REGS_USER |
PERF_SAMPLE_REGS_INTR);
}
pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
perf/x86/intel: Hybrid PMU support for perf capabilities Some platforms, e.g. Alder Lake, have hybrid architecture. Although most PMU capabilities are the same, there are still some unique PMU capabilities for different hybrid PMUs. Perf should register a dedicated pmu for each hybrid PMU. Add a new struct x86_hybrid_pmu, which saves the dedicated pmu and capabilities for each hybrid PMU. The architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicates the architecture features which are available on all hybrid PMUs. The architecture features are stored in the global x86_pmu.intel_cap. For Alder Lake, the model-specific features are perf metrics and PEBS-via-PT. The corresponding bits of the global x86_pmu.intel_cap should be 0 for these two features. Perf should not use the global intel_cap to check the features on a hybrid system. Add a dedicated intel_cap in the x86_hybrid_pmu to store the model-specific capabilities. Use the dedicated intel_cap to replace the global intel_cap for thse two features. The dedicated intel_cap will be set in the following "Add Alder Lake Hybrid support" patch. Add is_hybrid() to distinguish a hybrid system. ADL may have an alternative configuration. With that configuration, the X86_FEATURE_HYBRID_CPU is not set. Perf cannot rely on the feature bit. Add a new static_key_false, perf_is_hybrid, to indicate a hybrid system. It will be assigned in the following "Add Alder Lake Hybrid support" patch as well. Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/1618237865-33448-5-git-send-email-kan.liang@linux.intel.com
2021-04-12 07:30:44 -07:00
if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) {
pr_cont("PEBS-via-PT, ");
x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
}
perf/x86/intel: Support adaptive PEBS v4 Adaptive PEBS is a new way to report PEBS sampling information. Instead of a fixed size record for all PEBS events it allows to configure the PEBS record to only include the information needed. Events can then opt in to use such an extended record, or stay with a basic record which only contains the IP. The major new feature is to support LBRs in PEBS record. Besides normal LBR, this allows (much faster) large PEBS, while still supporting callstacks through callstack LBR. So essentially a lot of profiling can now be done without frequent interrupts, dropping the overhead significantly. The main requirement still is to use a period, and not use frequency mode, because frequency mode requires reevaluating the frequency on each overflow. The floating point state (XMM) is also supported, which allows efficient profiling of FP function arguments. Introduce specific drain function to handle variable length records. Use a new callback to parse the new record format, and also handle the STATUS field now being at a different offset. Add code to set up the configuration register. Since there is only a single register, all events either get the full super set of all events, or only the basic record. Originally-by: Andi Kleen <ak@linux.intel.com> Signed-off-by: Kan Liang <kan.liang@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Jiri Olsa <jolsa@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vince Weaver <vincent.weaver@maine.edu> Cc: acme@kernel.org Cc: jolsa@kernel.org Link: https://lkml.kernel.org/r/20190402194509.2832-6-kan.liang@linux.intel.com [ Renamed GPRS => GP. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-04-02 12:45:02 -07:00
break;
default:
pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
x86_pmu.pebs = 0;
}
}
}
void perf_restore_debug_store(void)
{
struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
if (!x86_pmu.bts && !x86_pmu.pebs)
return;
wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
}