Re: [x86/signal] 3aac3ebea0: will-it-scale.per_thread_ops -11.9% regression

[Yin Fengwei]

On 12/8/2021 6:17 AM, Bae, Chang Seok wrote:
> On Dec 7, 2021, at 12:36, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>> On Tue, Dec 07 2021 at 18:49, Chang Seok Bae wrote:
>>> On Dec 7, 2021, at 05:38, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>>>>
>>>> Does that use sigaltstack() ?
>>>
>>> FWIW, I was also wondering about this with:
>>>
>>> $ git clone https://github.com/antonblanchard/will-it-scale.git
>>> $ cd will-it-scale/
>>> $ git grep sigaltstack
>>> $
>>>
>>> But, the test seems to use python via runtest.py. And the python code has
>>> sigaltstack():
>>>    https://github.com/python/cpython/blob/main/Modules/faulthandler.c#L454
>>
>> But how does that affect the test written in C? Mysterious!
> 
> Indeed, I can only see the sigaltstack() trace via the python script. 
> 
> $ strace -f python3.7 ./runtest.py signal1 1>out 2>err
> $ grep -r sigaltstack err
> sigaltstack({ss_sp=0xe13f50, ss_flags=0, ss_size=16384}, {ss_sp=NULL,
> ss_flags=SS_DISABLE, ss_size=0}) = 0
> …
> 
> $ strace -f ./signal1_processes 1>out 2>err
> $ grep -r sigaltstack err
> $ strace -f ./signal1_threads 1>out 2>err
> $ grep -r sigaltstack err
> $

Here is what I got:

>From the strace output, we could see:
[pid 3171703] rt_sigreturn({mask=[]})   = 0


Code in kernel:
SYSCALL_DEFINE0(rt_sigreturn)
  restore_altstack
    do_sigaltstack
      sigaltstack_lock


And this matches the perf callstack in the report:

      0.00            +6.3        6.33 ±  8%  perf-profile.calltrace.cycles-pp.native_queued_spin_lock_slowpath._raw_spin_lock_irq.do_sigaltstack.restore_altstack.__x64_sys_rt_sigreturn

The rt_sigreturn was set as signal action restorer by libc during sigaction call. Thanks.

Regards
Yin, Fengwei      

> 
> I don’t get how this syscall could contribute 11% degradation in this test.
> 
> BTW, the current code rejects the reported instruction here:
> 
> $ python3.7 ./runtest.py signal1 295 thread 16
> Usage: runtest.py <testcase>
> 
> Thanks,
> Chang
>