1
linux/kernel/trace/fprobe.c

387 lines
9.2 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* fprobe - Simple ftrace probe wrapper for function entry.
*/
#define pr_fmt(fmt) "fprobe: " fmt
#include <linux/err.h>
#include <linux/fprobe.h>
#include <linux/kallsyms.h>
#include <linux/kprobes.h>
#include <linux/rethook.h>
#include <linux/slab.h>
#include <linux/sort.h>
#include "trace.h"
struct fprobe_rethook_node {
struct rethook_node node;
unsigned long entry_ip;
unsigned long entry_parent_ip;
char data[];
};
static inline void __fprobe_handler(unsigned long ip, unsigned long parent_ip,
struct ftrace_ops *ops, struct ftrace_regs *fregs)
{
struct fprobe_rethook_node *fpr;
struct rethook_node *rh = NULL;
struct fprobe *fp;
void *entry_data = NULL;
int ret = 0;
fp = container_of(ops, struct fprobe, ops);
if (fp->exit_handler) {
rh = rethook_try_get(fp->rethook);
if (!rh) {
fp->nmissed++;
return;
}
fpr = container_of(rh, struct fprobe_rethook_node, node);
fpr->entry_ip = ip;
fpr->entry_parent_ip = parent_ip;
if (fp->entry_data_size)
entry_data = fpr->data;
}
if (fp->entry_handler)
ret = fp->entry_handler(fp, ip, parent_ip, ftrace_get_regs(fregs), entry_data);
/* If entry_handler returns !0, nmissed is not counted. */
if (rh) {
if (ret)
rethook_recycle(rh);
else
rethook_hook(rh, ftrace_get_regs(fregs), true);
}
}
static void fprobe_handler(unsigned long ip, unsigned long parent_ip,
struct ftrace_ops *ops, struct ftrace_regs *fregs)
{
struct fprobe *fp;
int bit;
fp = container_of(ops, struct fprobe, ops);
if (fprobe_disabled(fp))
return;
/* recursion detection has to go before any traceable function and
* all functions before this point should be marked as notrace
*/
bit = ftrace_test_recursion_trylock(ip, parent_ip);
if (bit < 0) {
fp->nmissed++;
return;
}
__fprobe_handler(ip, parent_ip, ops, fregs);
ftrace_test_recursion_unlock(bit);
}
NOKPROBE_SYMBOL(fprobe_handler);
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
static void fprobe_kprobe_handler(unsigned long ip, unsigned long parent_ip,
struct ftrace_ops *ops, struct ftrace_regs *fregs)
{
struct fprobe *fp;
int bit;
fp = container_of(ops, struct fprobe, ops);
if (fprobe_disabled(fp))
return;
/* recursion detection has to go before any traceable function and
* all functions called before this point should be marked as notrace
*/
bit = ftrace_test_recursion_trylock(ip, parent_ip);
if (bit < 0) {
fp->nmissed++;
return;
}
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
/*
* This user handler is shared with other kprobes and is not expected to be
* called recursively. So if any other kprobe handler is running, this will
* exit as kprobe does. See the section 'Share the callbacks with kprobes'
* in Documentation/trace/fprobe.rst for more information.
*/
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
if (unlikely(kprobe_running())) {
fp->nmissed++;
goto recursion_unlock;
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
}
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
kprobe_busy_begin();
__fprobe_handler(ip, parent_ip, ops, fregs);
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
kprobe_busy_end();
recursion_unlock:
ftrace_test_recursion_unlock(bit);
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
}
static void fprobe_exit_handler(struct rethook_node *rh, void *data,
unsigned long ret_ip, struct pt_regs *regs)
{
struct fprobe *fp = (struct fprobe *)data;
struct fprobe_rethook_node *fpr;
int bit;
if (!fp || fprobe_disabled(fp))
return;
fpr = container_of(rh, struct fprobe_rethook_node, node);
/*
* we need to assure no calls to traceable functions in-between the
* end of fprobe_handler and the beginning of fprobe_exit_handler.
*/
bit = ftrace_test_recursion_trylock(fpr->entry_ip, fpr->entry_parent_ip);
if (bit < 0) {
fp->nmissed++;
return;
}
fp->exit_handler(fp, fpr->entry_ip, ret_ip, regs,
fp->entry_data_size ? (void *)fpr->data : NULL);
ftrace_test_recursion_unlock(bit);
}
NOKPROBE_SYMBOL(fprobe_exit_handler);
static int symbols_cmp(const void *a, const void *b)
{
const char **str_a = (const char **) a;
const char **str_b = (const char **) b;
return strcmp(*str_a, *str_b);
}
/* Convert ftrace location address from symbols */
static unsigned long *get_ftrace_locations(const char **syms, int num)
{
unsigned long *addrs;
/* Convert symbols to symbol address */
addrs = kcalloc(num, sizeof(*addrs), GFP_KERNEL);
if (!addrs)
return ERR_PTR(-ENOMEM);
/* ftrace_lookup_symbols expects sorted symbols */
sort(syms, num, sizeof(*syms), symbols_cmp, NULL);
if (!ftrace_lookup_symbols(syms, num, addrs))
return addrs;
kfree(addrs);
return ERR_PTR(-ENOENT);
}
static void fprobe_init(struct fprobe *fp)
{
fp->nmissed = 0;
fprobe: Introduce FPROBE_FL_KPROBE_SHARED flag for fprobe Introduce FPROBE_FL_KPROBE_SHARED flag for sharing fprobe callback with kprobes safely from the viewpoint of recursion. Since the recursion safety of the fprobe (and ftrace) is a bit different from the kprobes, this may cause an issue if user wants to run the same code from the fprobe and the kprobes. The kprobes has per-cpu 'current_kprobe' variable which protects the kprobe handler from recursion in any case. On the other hand, the fprobe uses only ftrace_test_recursion_trylock(), which will allow interrupt context calls another (or same) fprobe during the fprobe user handler is running. This is not a matter in cases if the common callback shared among the kprobes and the fprobe has its own recursion detection, or it can handle the recursion in the different contexts (normal/interrupt/NMI.) But if it relies on the 'current_kprobe' recursion lock, it has to check kprobe_running() and use kprobe_busy_*() APIs. Fprobe has FPROBE_FL_KPROBE_SHARED flag to do this. If your common callback code will be shared with kprobes, please set FPROBE_FL_KPROBE_SHARED *before* registering the fprobe, like; fprobe.flags = FPROBE_FL_KPROBE_SHARED; register_fprobe(&fprobe, "func*", NULL); This will protect your common callback from the nested call. Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Tested-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Alexei Starovoitov <ast@kernel.org> Link: https://lore.kernel.org/bpf/164735293127.1084943.15687374237275817599.stgit@devnote2
2022-03-15 07:02:11 -07:00
if (fprobe_shared_with_kprobes(fp))
fp->ops.func = fprobe_kprobe_handler;
else
fp->ops.func = fprobe_handler;
fp->ops.flags |= FTRACE_OPS_FL_SAVE_REGS;
}
static int fprobe_init_rethook(struct fprobe *fp, int num)
{
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
int size;
if (!fp->exit_handler) {
fp->rethook = NULL;
return 0;
}
/* Initialize rethook if needed */
if (fp->nr_maxactive)
num = fp->nr_maxactive;
else
num *= num_possible_cpus() * 2;
if (num <= 0)
return -EINVAL;
size = sizeof(struct fprobe_rethook_node) + fp->entry_data_size;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
/* Initialize rethook */
fp->rethook = rethook_alloc((void *)fp, fprobe_exit_handler, size, num);
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
if (IS_ERR(fp->rethook))
return PTR_ERR(fp->rethook);
return 0;
}
static void fprobe_fail_cleanup(struct fprobe *fp)
{
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
if (!IS_ERR_OR_NULL(fp->rethook)) {
/* Don't need to cleanup rethook->handler because this is not used. */
rethook_free(fp->rethook);
fp->rethook = NULL;
}
ftrace_free_filter(&fp->ops);
}
/**
* register_fprobe() - Register fprobe to ftrace by pattern.
* @fp: A fprobe data structure to be registered.
* @filter: A wildcard pattern of probed symbols.
* @notfilter: A wildcard pattern of NOT probed symbols.
*
* Register @fp to ftrace for enabling the probe on the symbols matched to @filter.
* If @notfilter is not NULL, the symbols matched the @notfilter are not probed.
*
* Return 0 if @fp is registered successfully, -errno if not.
*/
int register_fprobe(struct fprobe *fp, const char *filter, const char *notfilter)
{
struct ftrace_hash *hash;
unsigned char *str;
int ret, len;
if (!fp || !filter)
return -EINVAL;
fprobe_init(fp);
len = strlen(filter);
str = kstrdup(filter, GFP_KERNEL);
ret = ftrace_set_filter(&fp->ops, str, len, 0);
kfree(str);
if (ret)
return ret;
if (notfilter) {
len = strlen(notfilter);
str = kstrdup(notfilter, GFP_KERNEL);
ret = ftrace_set_notrace(&fp->ops, str, len, 0);
kfree(str);
if (ret)
goto out;
}
/* TODO:
* correctly calculate the total number of filtered symbols
* from both filter and notfilter.
*/
hash = rcu_access_pointer(fp->ops.local_hash.filter_hash);
if (WARN_ON_ONCE(!hash))
goto out;
ret = fprobe_init_rethook(fp, (int)hash->count);
if (!ret)
ret = register_ftrace_function(&fp->ops);
out:
if (ret)
fprobe_fail_cleanup(fp);
return ret;
}
EXPORT_SYMBOL_GPL(register_fprobe);
/**
* register_fprobe_ips() - Register fprobe to ftrace by address.
* @fp: A fprobe data structure to be registered.
* @addrs: An array of target ftrace location addresses.
* @num: The number of entries of @addrs.
*
* Register @fp to ftrace for enabling the probe on the address given by @addrs.
* The @addrs must be the addresses of ftrace location address, which may be
* the symbol address + arch-dependent offset.
* If you unsure what this mean, please use other registration functions.
*
* Return 0 if @fp is registered successfully, -errno if not.
*/
int register_fprobe_ips(struct fprobe *fp, unsigned long *addrs, int num)
{
int ret;
if (!fp || !addrs || num <= 0)
return -EINVAL;
fprobe_init(fp);
ret = ftrace_set_filter_ips(&fp->ops, addrs, num, 0, 0);
if (ret)
return ret;
ret = fprobe_init_rethook(fp, num);
if (!ret)
ret = register_ftrace_function(&fp->ops);
if (ret)
fprobe_fail_cleanup(fp);
return ret;
}
EXPORT_SYMBOL_GPL(register_fprobe_ips);
/**
* register_fprobe_syms() - Register fprobe to ftrace by symbols.
* @fp: A fprobe data structure to be registered.
* @syms: An array of target symbols.
* @num: The number of entries of @syms.
*
* Register @fp to the symbols given by @syms array. This will be useful if
* you are sure the symbols exist in the kernel.
*
* Return 0 if @fp is registered successfully, -errno if not.
*/
int register_fprobe_syms(struct fprobe *fp, const char **syms, int num)
{
unsigned long *addrs;
int ret;
if (!fp || !syms || num <= 0)
return -EINVAL;
addrs = get_ftrace_locations(syms, num);
if (IS_ERR(addrs))
return PTR_ERR(addrs);
ret = register_fprobe_ips(fp, addrs, num);
kfree(addrs);
return ret;
}
EXPORT_SYMBOL_GPL(register_fprobe_syms);
tracing/probes: Add fprobe events for tracing function entry and exit. Add fprobe events for tracing function entry and exit instead of kprobe events. With this change, we can continue to trace function entry/exit even if the CONFIG_KPROBES_ON_FTRACE is not available. Since CONFIG_KPROBES_ON_FTRACE requires the CONFIG_DYNAMIC_FTRACE_WITH_REGS, it is not available if the architecture only supports CONFIG_DYNAMIC_FTRACE_WITH_ARGS. And that means kprobe events can not probe function entry/exit effectively on such architecture. But this can be solved if the dynamic events supports fprobe events. The fprobe event is a new dynamic events which is only for the function (symbol) entry and exit. This event accepts non register fetch arguments so that user can trace the function arguments and return values. The fprobe events syntax is here; f[:[GRP/][EVENT]] FUNCTION [FETCHARGS] f[MAXACTIVE][:[GRP/][EVENT]] FUNCTION%return [FETCHARGS] E.g. # echo 'f vfs_read $arg1' >> dynamic_events # echo 'f vfs_read%return $retval' >> dynamic_events # cat dynamic_events f:fprobes/vfs_read__entry vfs_read arg1=$arg1 f:fprobes/vfs_read__exit vfs_read%return arg1=$retval # echo 1 > events/fprobes/enable # head -n 20 trace | tail # TASK-PID CPU# ||||| TIMESTAMP FUNCTION # | | | ||||| | | sh-142 [005] ...1. 448.386420: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386436: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.386451: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386458: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.386469: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386476: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.602073: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.602089: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 Link: https://lore.kernel.org/all/168507469754.913472.6112857614708350210.stgit@mhiramat.roam.corp.google.com/ Reported-by: kernel test robot <lkp@intel.com> Link: https://lore.kernel.org/all/202302011530.7vm4O8Ro-lkp@intel.com/ Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-06-06 05:39:55 -07:00
bool fprobe_is_registered(struct fprobe *fp)
{
if (!fp || (fp->ops.saved_func != fprobe_handler &&
fp->ops.saved_func != fprobe_kprobe_handler))
return false;
return true;
}
/**
* unregister_fprobe() - Unregister fprobe from ftrace
* @fp: A fprobe data structure to be unregistered.
*
* Unregister fprobe (and remove ftrace hooks from the function entries).
*
* Return 0 if @fp is unregistered successfully, -errno if not.
*/
int unregister_fprobe(struct fprobe *fp)
{
int ret;
tracing/probes: Add fprobe events for tracing function entry and exit. Add fprobe events for tracing function entry and exit instead of kprobe events. With this change, we can continue to trace function entry/exit even if the CONFIG_KPROBES_ON_FTRACE is not available. Since CONFIG_KPROBES_ON_FTRACE requires the CONFIG_DYNAMIC_FTRACE_WITH_REGS, it is not available if the architecture only supports CONFIG_DYNAMIC_FTRACE_WITH_ARGS. And that means kprobe events can not probe function entry/exit effectively on such architecture. But this can be solved if the dynamic events supports fprobe events. The fprobe event is a new dynamic events which is only for the function (symbol) entry and exit. This event accepts non register fetch arguments so that user can trace the function arguments and return values. The fprobe events syntax is here; f[:[GRP/][EVENT]] FUNCTION [FETCHARGS] f[MAXACTIVE][:[GRP/][EVENT]] FUNCTION%return [FETCHARGS] E.g. # echo 'f vfs_read $arg1' >> dynamic_events # echo 'f vfs_read%return $retval' >> dynamic_events # cat dynamic_events f:fprobes/vfs_read__entry vfs_read arg1=$arg1 f:fprobes/vfs_read__exit vfs_read%return arg1=$retval # echo 1 > events/fprobes/enable # head -n 20 trace | tail # TASK-PID CPU# ||||| TIMESTAMP FUNCTION # | | | ||||| | | sh-142 [005] ...1. 448.386420: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386436: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.386451: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386458: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.386469: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.386476: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 sh-142 [005] ...1. 448.602073: vfs_read__entry: (vfs_read+0x4/0x340) arg1=0xffff888007f7c540 sh-142 [005] ..... 448.602089: vfs_read__exit: (ksys_read+0x75/0x100 <- vfs_read) arg1=0x1 Link: https://lore.kernel.org/all/168507469754.913472.6112857614708350210.stgit@mhiramat.roam.corp.google.com/ Reported-by: kernel test robot <lkp@intel.com> Link: https://lore.kernel.org/all/202302011530.7vm4O8Ro-lkp@intel.com/ Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-06-06 05:39:55 -07:00
if (!fprobe_is_registered(fp))
return -EINVAL;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
if (!IS_ERR_OR_NULL(fp->rethook))
fprobe: Ensure running fprobe_exit_handler() finished before calling rethook_free() Ensure running fprobe_exit_handler() has finished before calling rethook_free() in the unregister_fprobe() so that caller can free the fprobe right after unregister_fprobe(). unregister_fprobe() ensured that all running fprobe_entry/exit_handler() have finished by calling unregister_ftrace_function() which synchronizes RCU. But commit 5f81018753df ("fprobe: Release rethook after the ftrace_ops is unregistered") changed to call rethook_free() after unregister_ftrace_function(). So call rethook_stop() to make rethook disabled before unregister_ftrace_function() and ensure it again. Here is the possible code flow that can call the exit handler after unregister_fprobe(). ------ CPU1 CPU2 call unregister_fprobe(fp) ... __fprobe_handler() rethook_hook() on probed function unregister_ftrace_function() return from probed function rethook hooks find rh->handler == fprobe_exit_handler call fprobe_exit_handler() rethook_free(): set rh->handler = NULL; return from unreigster_fprobe; call fp->exit_handler() <- (*) ------ (*) At this point, the exit handler is called after returning from unregister_fprobe(). This fixes it as following; ------ CPU1 CPU2 call unregister_fprobe() ... rethook_stop(): set rh->handler = NULL; __fprobe_handler() rethook_hook() on probed function unregister_ftrace_function() return from probed function rethook hooks find rh->handler == NULL return from rethook rethook_free() return from unreigster_fprobe; ------ Link: https://lore.kernel.org/all/168873859949.156157.13039240432299335849.stgit@devnote2/ Fixes: 5f81018753df ("fprobe: Release rethook after the ftrace_ops is unregistered") Cc: stable@vger.kernel.org Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
2023-07-07 07:03:19 -07:00
rethook_stop(fp->rethook);
ret = unregister_ftrace_function(&fp->ops);
if (ret < 0)
return ret;
kprobes: kretprobe scalability improvement kretprobe is using freelist to manage return-instances, but freelist, as LIFO queue based on singly linked list, scales badly and reduces the overall throughput of kretprobed routines, especially for high contention scenarios. Here's a typical throughput test of sys_prctl (counts in 10 seconds, measured with perf stat -a -I 10000 -e syscalls:sys_enter_prctl): OS: Debian 10 X86_64, Linux 6.5rc7 with freelist HW: XEON 8336C x 2, 64 cores/128 threads, DDR4 3200MT/s 1T 2T 4T 8T 16T 24T 24150045 29317964 15446741 12494489 18287272 17708768 32T 48T 64T 72T 96T 128T 16200682 13737658 11645677 11269858 10470118 9931051 This patch introduces objpool to replace freelist. objpool is a high performance queue, which can bring near-linear scalability to kretprobed routines. Tests of kretprobe throughput show the biggest ratio as 159x of original freelist. Here's the result: 1T 2T 4T 8T 16T native: 41186213 82336866 164250978 328662645 658810299 freelist: 24150045 29317964 15446741 12494489 18287272 objpool: 23926730 48010314 96125218 191782984 385091769 32T 48T 64T 96T 128T native: 1330338351 1969957941 2512291791 2615754135 2671040914 freelist: 16200682 13737658 11645677 10470118 9931051 objpool: 764481096 1147149781 1456220214 1502109662 1579015050 Testings on 96-core ARM64 output similarly, but with the biggest ratio up to 448x: OS: Debian 10 AARCH64, Linux 6.5rc7 HW: Kunpeng-920 96 cores/2 sockets/4 NUMA nodes, DDR4 2933 MT/s 1T 2T 4T 8T 16T native: . 30066096 63569843 126194076 257447289 505800181 freelist: 16152090 11064397 11124068 7215768 5663013 objpool: 13997541 28032100 55726624 110099926 221498787 24T 32T 48T 64T 96T native: 763305277 1015925192 1521075123 2033009392 3021013752 freelist: 5015810 4602893 3766792 3382478 2945292 objpool: 328192025 439439564 668534502 887401381 1319972072 Link: https://lore.kernel.org/all/20231017135654.82270-4-wuqiang.matt@bytedance.com/ Signed-off-by: wuqiang.matt <wuqiang.matt@bytedance.com> Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
2023-10-17 06:56:52 -07:00
if (!IS_ERR_OR_NULL(fp->rethook))
rethook_free(fp->rethook);
ftrace_free_filter(&fp->ops);
return ret;
}
EXPORT_SYMBOL_GPL(unregister_fprobe);