When a CPU core is shut down, all of its caches need to be flushed
to prevent stale data from causing errors if the core is resumed.
Current Linux suspend code performs an assignment after the flush,
which can add dirty data back to the cache. On some AMD platforms,
additional speculative reads have caused crashes on resume because
of this dirty data.
Relocate the cache flush to be the very last thing done before
halting. Tie into an assembly line so the compile will not
reorder it. Add some documentation explaining what is going
on and why we're doing this.
Signed-off-by: Mark Langsdorf <mark.langsdorf@amd.com>
Acked-by: Mark Borden <mark.borden@amd.com>
Acked-by: Michael Hohmuth <michael.hohmuth@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Jack Ren and Eric Miao tracked down the following long standing
problem in the NOHZ code:
scheduler switch to idle task
enable interrupts
Window starts here
----> interrupt happens (does not set NEED_RESCHED)
irq_exit() stops the tick
----> interrupt happens (does set NEED_RESCHED)
return from schedule()
cpu_idle(): preempt_disable();
Window ends here
The interrupts can happen at any point inside the race window. The
first interrupt stops the tick, the second one causes the scheduler to
rerun and switch away from idle again and we end up with the tick
disabled.
The fact that it needs two interrupts where the first one does not set
NEED_RESCHED and the second one does made the bug obscure and extremly
hard to reproduce and analyse. Kudos to Jack and Eric.
Solution: Limit the NOHZ functionality to the idle loop to make sure
that we can not run into such a situation ever again.
cpu_idle()
{
preempt_disable();
while(1) {
tick_nohz_stop_sched_tick(1); <- tell NOHZ code that we
are in the idle loop
while (!need_resched())
halt();
tick_nohz_restart_sched_tick(); <- disables NOHZ mode
preempt_enable_no_resched();
schedule();
preempt_disable();
}
}
In hindsight we should have done this forever, but ...
/me grabs a large brown paperbag.
Debugged-by: Jack Ren <jack.ren@marvell.com>,
Debugged-by: eric miao <eric.y.miao@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Take it out of smpboot.c, and move it to process_32.c, closer
to its only user.
Signed-off-by: Glauber Costa <gcosta@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Patrick McHardy reported a crash:
> > I get this oops once a day, its apparently triggered by something
> > run by cron, but the process is a different one each time.
> >
> > Kernel is -git from yesterday shortly before the -rc6 release
> > (last commit is the usb-2.6 merge, the x86 patches are missing),
> > .config is attached.
> >
> > I'll retry with current -git, but the patches that have gone in
> > since I last updated don't look related.
> >
> > [62060.043009] BUG: unable to handle kernel NULL pointer dereference at
> > 000001ff
> > [62060.043009] IP: [<c0102a9b>] __switch_to+0x2f/0x118
> > [62060.043009] *pde = 00000000
> > [62060.043009] Oops: 0002 [#1] PREEMPT
Vegard Nossum analyzed it:
> This decodes to
>
> 0: 0f ae 00 fxsave (%eax)
>
> so it's related to the floating-point context. This is the exact
> location of the crash:
>
> $ addr2line -e arch/x86/kernel/process_32.o -i ab0
> include/asm/i387.h:232
> include/asm/i387.h:262
> arch/x86/kernel/process_32.c:595
>
> ...so it looks like prev_task->thread.xstate->fxsave has become NULL.
> Or maybe it never had any other value.
Somehow (as described below) TS_USEDFPU is set but the fpu is not
allocated or freed.
Another possible FPU pre-emption issue with the sleazy FPU optimization
which was benign before but not so anymore, with the dynamic FPU allocation
patch.
New task is getting exec'd and it is prempted at the below point.
flush_thread() {
...
/*
* Forget coprocessor state..
*/
clear_fpu(tsk);
<----- Preemption point
clear_used_math();
...
}
Now when it context switches in again, as the used_math() is still set
and fpu_counter can be > 5, we will do a math_state_restore() which sets
the task's TS_USEDFPU. After it continues from the above preemption point
it does clear_used_math() and much later free_thread_xstate().
Now, at the next context switch, it is quite possible that xstate is
null, used_math() is not set and TS_USEDFPU is still set. This will
trigger unlazy_fpu() causing kernel oops.
Fix this by clearing tsk's fpu_counter before clearing task's fpu.
Reported-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
more unification. Should cause no change in functionality.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
default_idle is selected in cpu_idle(), when no other idle routine is
selected. Select it in select_idle_routine() when mwait is not
selected.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Jürgen Mell reported an FPU state corruption bug under CONFIG_PREEMPT,
and bisected it to commit v2.6.19-1363-gacc2076, "i386: add sleazy FPU
optimization".
Add tsk_used_math() checks to prevent calling math_state_restore()
which can sleep in the case of !tsk_used_math(). This prevents
making a blocking call in __switch_to().
Apparently "fpu_counter > 5" check is not enough, as in some signal handling
and fork/exec scenarios, fpu_counter > 5 and !tsk_used_math() is possible.
It's a side effect though. This is the failing scenario:
process 'A' in save_i387_ia32() just after clear_used_math()
Got an interrupt and pre-empted out.
At the next context switch to process 'A' again, kernel tries to restore
the math state proactively and sees a fpu_counter > 0 and !tsk_used_math()
This results in init_fpu() during the __switch_to()'s math_state_restore()
And resulting in fpu corruption which will be saved/restored
(save_i387_fxsave and restore_i387_fxsave) during the remaining
part of the signal handling after the context switch.
Bisected-by: Jürgen Mell <j.mell@t-online.de>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Tested-by: Jürgen Mell <j.mell@t-online.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@kernel.org
Add preempt off timings. A lot of kernel core code is taken from the RT patch
latency trace that was written by Ingo Molnar.
This adds "preemptoff" and "preemptirqsoff" to /debugfs/tracing/available_tracers
Now instead of just tracing irqs off, preemption off can be selected
to be recorded.
When this is selected, it shares the same files as irqs off timings.
One can either trace preemption off, irqs off, or one or the other off.
By echoing "preemptoff" into /debugfs/tracing/current_tracer, recording
of preempt off only is performed. "irqsoff" will only record the time
irqs are disabled, but "preemptirqsoff" will take the total time irqs
or preemption are disabled. Runtime switching of these options is now
supported by simpling echoing in the appropriate trace name into
/debugfs/tracing/current_tracer.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
OK, so 25-mm1 gave a lockdep error which made me look into this.
The first thing that I noticed was the horrible mess; the second thing I
saw was hacks like: 71e93d1561
The problem is that arch idle routines are somewhat inconsitent with
their IRQ state handling and instead of fixing _that_, we go paper over
the problem.
So the thing I've tried to do is set a standard for idle routines and
fix them all up to adhere to that. So the rules are:
idle routines are entered with IRQs disabled
idle routines will exit with IRQs enabled
Nearly all already did this in one form or another.
Merge the 32 and 64 bit bits so they no longer have different bugs.
As for the actual lockdep warning; __sti_mwait() did a plainly un-annotated
irq-enable.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Bob Copeland <me@bobcopeland.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Only allocate the FPU area when the application actually uses FPU, i.e., in the
first lazy FPU trap. This could save memory for non-fpu using apps.
for example: on my system after boot, there are around 300 processes, with
only 17 using FPU.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Split the FPU save area from the task struct. This allows easy migration
of FPU context, and it's generally cleaner. It also allows the following
two optimizations:
1) only allocate when the application actually uses FPU, so in the first
lazy FPU trap. This could save memory for non-fpu using apps. Next patch
does this lazy allocation.
2) allocate the right size for the actual cpu rather than 512 bytes always.
Patches enabling xsave/xrstor support (coming shortly) will take advantage
of this.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch implements the PR_GET_TSC and PR_SET_TSC prctl()
commands on the x86 platform (both 32 and 64 bit.) These
commands control the ability to read the timestamp counter
from userspace (the RDTSC instruction.)
While the RDTSC instuction is a useful profiling tool,
it is also the source of some non-determinism in ring-3.
For deterministic replay applications it is useful to be
able to trap and emulate (and record the outcome of) this
instruction.
This patch uses code earlier used to disable the timestamp
counter for the SECCOMP framework. A side-effect of this
patch is that the SECCOMP environment will now also disable
the timestamp counter on x86_64 due to the addition of the
TIF_NOTSC define on this platform.
The code which enables/disables the RDTSC instruction during
context switches is in the __switch_to_xtra function, which
already handles other unusual conditions, so normal
performance should not have to suffer from this change.
Signed-off-by: Erik Bosman <ejbosman@cs.vu.nl>
Acked-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
I've made a small investigation about vm86.h inclusion rules and it
looks like everything is more or less ok.
Files that rely on asm/vm86.h symbols are:
- kprobes.c
- process_32.c
- signal_32.c
- traps_32.c
- vm86_32.c
File process_32.c includes vm86.h explicitly. We can remove that
include and it won't break anything.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This matches x86_64 behaviour, which is a superior one IMHO
Signed-off-by: Glauber Costa <gcosta@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Otherwise, enabling (or better, subsequent disabling) of single
stepping would cause a kernel oops on CPUs not having this MSR.
The patch could have been added a conditional to the MSR write in
user_disable_single_step(), but centralizing the updates seems safer
and (looking forward) better manageable.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Cc: Markus Metzger <markus.t.metzger@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch also resolves hangs on boot:
http://lkml.org/lkml/2008/2/23/263http://bugzilla.kernel.org/show_bug.cgi?id=10093
The bug was causing once-in-few-reboots 10-15 sec wait during boot on
certain laptops.
Earlier commit 40d6a14662 added
smp_call_function in cpu_idle_wait() to kick cpus that are in tickless
idle. Looking at cpu_idle_wait code at that time, code seemed to be
over-engineered for a case which is rarely used (while changing idle
handler).
Below is a simplified version of cpu_idle_wait, which just makes a dummy
smp_call_function to all cpus, to make them come out of old idle handler
and start using the new idle handler. It eliminates code in the idle
loop to handle cpu_idle_wait.
Signed-off-by: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
revert the BTS ptrace extension for now.
based on general objections from Roland McGrath:
http://lkml.org/lkml/2008/2/21/323
we'll let the BTS functionality cook some more and re-enable
it in v2.6.26. We'll leave the dead code around to help the
development of this code.
(X86_BTS is not defined at the moment)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Suppress A.OUT library support if CONFIG_ARCH_SUPPORTS_AOUT is not set.
Not all architectures support the A.OUT binfmt, so the ELF binfmt should not
be permitted to go looking for A.OUT libraries to load in such a case. Not
only that, but under such conditions A.OUT core dumps are not produced either.
To make this work, this patch also does the following:
(1) Makes the existence of the contents of linux/a.out.h contingent on
CONFIG_ARCH_SUPPORTS_AOUT.
(2) Renames dump_thread() to aout_dump_thread() as it's only called by A.OUT
core dumping code.
(3) Moves aout_dump_thread() into asm/a.out-core.h and makes it inline. This
is then included only where needed. This means that this bit of arch
code will be stored in the appropriate A.OUT binfmt module rather than
the core kernel.
(4) Drops A.OUT support for Blackfin (according to Mike Frysinger it's not
needed) and FRV.
This patch depends on the previous patch to move STACK_TOP[_MAX] out of
asm/a.out.h and into asm/processor.h as they're required whether or not A.OUT
format is available.
[jdike@addtoit.com: uml: re-remove accidentally restored code]
Signed-off-by: David Howells <dhowells@redhat.com>
Cc: <linux-arch@vger.kernel.org>
Signed-off-by: Jeff Dike <jdike@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
arch/x86/kernel/process_32.c:254:43: warning: Using plain integer as NULL pointer
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Fix the following warning:
WARNING: arch/x86/kernel/built-in.o(.text+0x3): Section mismatch: reference to .cpuinit.data:force_mwait in 'mwait_usable'
[Seen on 64 bit only but similar pattern exist on 32 bit so fix it there too]
mwait_usable() were only used by a function annotated __cpuinit
so annotate mwait_usable() with __cpuinit to fix the warning.
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The warning message at idle_setup() is never shown because
smp_num_sibling hasn't been updated at this point yet.
Move this polling idle and HT enabled warning to select_idle_routine().
I also implement this warning on 64-bit kernel.
Signed-off-by: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Previously there was a AMD specific quirk to handle the case of
AMD Fam10h MWAIT not supporting any C states. But it turns out
that CPUID already has ways to detectly detect that without
using special quirks.
The new code simply checks if MWAIT supports at least C1 and doesn't
use it if it doesn't. No more vendor specific code.
Note this is does not simply clear MWAIT because MWAIT can be still
useful even without C states.
Credit goes to Ben Serebrin for pointing out the (nearly) obvious.
Cc: "Andreas Herrmann" <andreas.herrmann3@amd.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
an otherwise idle system takes about 3 ticks per network
interface in unregister_netdev() due to multiple calls to synchronize_rcu(),
which adds up to quite a few seconds for tearing down thousands of
interfaces. By flushing pending rcu callbacks in the idle loop, the system
makes progress hundreds of times faster. If this is indeed a sane thing to,
it probably needs to be done for other architectures than x86. And yes, the
network stack shouldn't call synchronize_rcu() quite so much, but fixing that
is a little more involved.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Right now, we take the stack pointer early during the backtrace path, but
only calculate bp several functions deep later, making it hard to reconcile
the stack and bp backtraces (as well as showing several internal backtrace
functions on the stack with bp based backtracing).
This patch moves the bp taking to the same place we take the stack pointer;
sadly this ripples through several layers of the back tracing stack,
but it's not all that bad in the end I hope.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This removes a bunch of dead code that is no longer needed now
that the user_regset interfaces are being used for all these jobs.
Signed-off-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
fastcall is always defined to be empty, remove it from arch/x86
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Resend using different mail client
Changes to the last version:
- split implementation into two layers: ds/bts and ptrace
- renamed TIF's
- save/restore ds save area msr in __switch_to_xtra()
- make block-stepping only look at BTF bit
Signed-off-by: Markus Metzger <markus.t.metzger@intel.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
White space and coding style clean up.
Make process_32/64.c similar.
Signed-off-by: Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This changes size-specific register names (eip/rip, esp/rsp, etc.) to
generic names in the thread and tss structures.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This replaces the debugreg[7] member of thread_struct with individual
members debugreg0, etc. This saves two words for the dummies 4 and 5,
and harmonizes the code between 32 and 64.
Signed-off-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Switch struct user_regs_struct (defined in <asm/user.h>, which is no
longer exported to userspace) to using register names without e- or
r-prefixes for both 32 and 64 bit x86. This is intended as a
preliminary step in unifying this code between architectures.
Also, be a bit more strict in truncating 32-bit "extended" segment
register values to 16 bits.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
We have a lot of code which differs only by the naming of specific
members of structures that contain registers. In order to enable
additional unifications, this patch drops the e- or r- size prefix
from the register names in struct pt_regs, and drops the x- prefixes
for segment registers on the 32-bit side.
This patch also performs the equivalent renames in some additional
places that might be candidates for unification in the future.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This adds low-level support for a per-thread value of MSR_IA32_DEBUGCTLMSR.
The per-thread value is switched in when TIF_DEBUGCTLMSR is set.
Signed-off-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This changes the single-step support to use a new thread_info flag
TIF_FORCED_TF instead of the PT_DTRACE flag in task_struct.ptrace.
This keeps arch implementation uses out of this non-arch field.
This changes the ptrace access to eflags to mask TF and maintain
the TIF_FORCED_TF flag directly if userland sets TF, instead of
relying on ptrace_signal_deliver. The 64-bit and 32-bit kernels
are harmonized on this same behavior. The ptrace_signal_deliver
approach works now, but this change makes the low-level register
access code reliable when called from different contexts than a
ptrace stop, which will be possible in the future.
The 64-bit do_debug exception handler is also changed not to clear TF
from user-mode registers. This matches the 32-bit kernel's behavior.
Signed-off-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This consolidates the four different places that implemented the same
encoding magic for the GDT-slot 32-bit TLS support. The old tls32.c was
renamed and is now only slightly modified to be the shared implementation.
Signed-off-by: Roland McGrath <roland@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Zachary Amsden <zach@vmware.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Randomize the location of the heap (brk) for i386 and x86_64. The range is
randomized in the range starting at current brk location up to 0x02000000
offset for both architectures. This, together with
pie-executable-randomization.patch and
pie-executable-randomization-fix.patch, should make the address space
randomization on i386 and x86_64 complete.
Arjan says:
This is known to break older versions of some emacs variants, whose dumper
code assumed that the last variable declared in the program is equal to the
start of the dynamically allocated memory region.
(The dumper is the code where emacs effectively dumps core at the end of it's
compilation stage; this coredump is then loaded as the main program during
normal use)
iirc this was 5 years or so; we found this way back when I was at RH and we
first did the security stuff there (including this brk randomization). It
wasn't all variants of emacs, and it got fixed as a result (I vaguely remember
that emacs already had code to deal with it for other archs/oses, just
ifdeffed wrongly).
It's a rare and wrong assumption as a general thing, just on x86 it mostly
happened to be true (but to be honest, it'll break too if gcc does
something fancy or if the linker does a non-standard order). Still its
something we should at least document.
Note 2: afaik it only broke the emacs *build*. I'm not 100% sure about that
(it IS 5 years ago) though.
[ akpm@linux-foundation.org: deuglification ]
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Roland McGrath <roland@redhat.com>
Cc: Jakub Jelinek <jakub@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
do a proper idle-wakeup event on HLT as well - some CPUs stop the TSC
in HLT too, not just when going through the ACPI methods.
(the ACPI idle code already does this.)
[ update the 64-bit side too, as noticed by Jiri Slaby. ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Sometimes cpu_idle_wait gets stuck because it might miss CPUS that are
already in idle, have no tasks waiting to run and have no interrupts going
to them. This is common on bootup when switching cpu idle governors.
This patch gives those CPUS that don't check in an IPI kick.
Background:
-----------
I notice this while developing the mcount patches, that every once in a
while the system would hang. Looking deeper, the hang was always at boot
up when registering init_menu of the cpu_idle menu governor. Talking
with Thomas Gliexner, we discovered that one of the CPUS had no timer
events scheduled for it and it was in idle (running with NO_HZ). So the
CPU would not set the cpu_idle_state bit.
Hitting sysrq-t a few times would eventually route the interrupt to the
stuck CPU and the system would continue.
Note, I would have used the PDA isidle but that is set after the
cpu_idle_state bit is cleared, and would leave a window open where we
may miss being kicked.
hmm, looking closer at this, we still have a small race window between
clearing the cpu_idle_state and disabling interrupts (hence the RFC).
CPU0: CPU 1:
--------- ---------
cpu_idle_wait(): cpu_idle():
| __cpu_cpu_var(is_idle) = 1;
| if (__get_cpu_var(cpu_idle_state)) /* == 0 */
per_cpu(cpu_idle_state, 1) = 1; |
if (per_cpu(is_idle, 1)) /* == 1 */ |
smp_call_function(1) |
| receives ipi and runs do_nothing.
wait on map == empty idle();
/* waits forever */
So really we need interrupts off for most of this then. One might think
that we could simply clear the cpu_idle_state from do_nothing, but I'm
assuming that cpu_idle governors can be removed, and this might cause a
race that a governor might be used after the module was removed.
Venki said:
I think your RFC patch is the right solution here. As I see it, there is
no race with your RFC patch. As long as you call a dummy smp_call_function
on all CPUs, we should be OK. We can get rid of cpu_idle_state and the
current wait forever logic altogether with dummy smp_call_function. And so
there wont be any wait forever scenario.
The whole point of cpu_idle_wait() is to make all CPUs come out of idle
loop atleast once. The caller will use cpu_idle_wait something like this.
// Want to change idle handler
- Switch global idle handler to always present default_idle
- call cpu_idle_wait so that all cpus come out of idle for an instant
and stop using old idle pointer and start using default idle
- Change the idle handler to a new handler
- optional cpu_idle_wait if you want all cpus to start using the new
handler immediately.
Maybe the below 1s patch is safe bet for .24. But for .25, I would say we
just replace all complicated logic by simple dummy smp_call_function and
remove cpu_idle_state altogether.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Cc: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Andi Kleen <ak@suse.de>
Cc: Len Brown <lenb@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
