Re: 08ed4efad6: stress-ng.sigsegv.ops_per_sec -41.9% regression

[Linus Torvalds]

On Thu, Apr 8, 2021 at 1:32 AM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
>
> FYI, we noticed a -41.9% regression of stress-ng.sigsegv.ops_per_sec due to commit
> 08ed4efad684 ("[PATCH v10 6/9] Reimplement RLIMIT_SIGPENDING on top of ucounts")

Ouch.

I *think* this test may be testing "send so many signals that it
triggers the signal queue overflow case".

And I *think* that the performance degradation may be due to lots of
unnecessary allocations, because ity looks like that commit changes
__sigqueue_alloc() to do

        struct sigqueue *q = kmem_cache_alloc(sigqueue_cachep, flags);

*before* checking the signal limit, and then if the signal limit was
exceeded, it will just be free'd instead.

The old code would check the signal count against RLIMIT_SIGPENDING
*first*, and if there were m ore pending signals then it wouldn't do
anything at all (including not incrementing that expensive atomic
count).

Also, the old code was very careful to only do the "get_user()" for
the *first* signal it added to the queue, and do the "put_user()" for
when removing the last signal. Exactly because those atomics are very
expensive.

The new code just does a lot of these atomics unconditionally.

I dunno. The profile data in there is a bit hard to read, but there's
a lot more cachee misses, and a *lot* of node crossers:

>    5961544          +190.4%   17314361        perf-stat.i.cache-misses
>   22107466          +119.2%   48457656        perf-stat.i.cache-references
>     163292 ą  3%   +4582.0%    7645410        perf-stat.i.node-load-misses
>     227388 ą  2%   +3708.8%    8660824        perf-stat.i.node-loads

and (probably as a result) average instruction costs have gone up enormously:

>       3.47           +66.8%       5.79        perf-stat.overall.cpi
>      22849           -65.6%       7866        perf-stat.overall.cycles-between-cache-misses

and it does seem to be at least partly about "put_ucounts()":

>       0.00            +4.5        4.46        perf-profile.calltrace.cycles-pp.put_ucounts.__sigqueue_free.get_signal.arch_do_signal_or_restart.exit_to_user_mode_prepare

and a lot of "get_ucounts()".

But it may also be that the new "get sigpending" is just *so* much
more expensive than it used to be.

               Linus