Re: [RESEND PATCH v13 2/3] sched: Move task_mm_cid_work to mm work_struct

[Gabriele Monaco <gmonaco@redhat.com>]

On Wed, 2025-06-25 at 09:57 -0400, Mathieu Desnoyers wrote:
> On 2025-06-25 04:01, kernel test robot wrote:
> > 
> > Hello,
> > 
> > kernel test robot noticed a 10.1% regression of
> > hackbench.throughput on:
> 
> Hi Gabriele,
> 
> This is a significant regression. Can you investigate before it gets
> merged ?
> 

Hi Mathieu,

I run some tests, the culprit for this performance regression seems to
be the interference due to more consistent `mm_cid` scans and them
running in `work_struct`, which brings some more scheduling overhead.

One solution could be to reduce the frequency: now they run
(sporadically) about every 100ms, if the minimum delay is 1s, the test
results seem ok.

However, I tried another approach that seems promising: work_struct get
scheduled relatively fast and this ends up giving a lot of contention
with kworkers, however something like timer_list seems less aggressive
and we obtain a similar reliability with respect to calls to the mm_cid
scan, without the same performance impact.

At the moment I just kept roughly the same structure of the patch and
used a timer delayed by 1 jiffy in place of the work_struct.
It may look cleaner if we use the timer directly for the 100ms delay
instead of storing and checking the time, in fact running a scan about
100ms after every rseq_handle_notify_resume.

What do you think?

Thanks,
Gabriele