Re: [PATCH v2 1/4] sched/task: Add the put_task_struct_atomic_safe function

[Valentin Schneider <vschneid@redhat.com>]

On 30/01/23 11:58, Wander Lairson Costa wrote:
> On Mon, Jan 30, 2023 at 11:47 AM Valentin Schneider <vschneid@redhat.com> wrote:
>>
>> On 30/01/23 08:49, Wander Lairson Costa wrote:
>> > On Fri, Jan 27, 2023 at 12:55 PM Valentin Schneider <vschneid@redhat.com> wrote:
>> >>
>> >> On 23/01/23 14:24, Wander Lairson Costa wrote:
>> >> > Therefore (if I am correct in my assumption), it would make sense for
>> >> > only some call sites to pay the overhead price for it. But this is
>> >> > just a guess, and I have no evidence to support my claim.
>> >>
>> >> My worry here is that it's easy to miss problematic callgraphs, and it's
>> >> potentially easy for new ones to creep in. Having a solution within
>> >> put_task_struct() itself would prevent that.
>> >>
>> >
>> > We could add a WARN_ON statement in put_task_struct() to detect such cases.
>> >
>>
>> Anyone running their kernel with DEBUG_ATOMIC_SLEEP should be able to
>> detect misuse, but it doesn't change that some callgraphs will only
>> materialize under certain hardware/configuration combos.
>>
>
> If we put a WARN_ON in put_task_struct(), we catch cases where the
> reference count didn't reach zero.
>

True, that'd be an improvement.

>> >> Another thing, if you look at release_task_stack(), it either caches the
>> >> outgoing stack for later use, or frees it via RCU (regardless of
>> >> PREEMPT_RT). Perhaps we could follow that and just always punt the freeing
>> >> of the task struct to RCU?
>> >>
>> >
>> > That's a point. Do you mean doing that even for !PREEMPT_RT?
>>
>> Could be worth a try?
>
> Sure. But I would do it only for PREEMPT_RT.
>
>> I think because of the cache thing the task stack is
>> a bit less aggressive wrt RCU callback processing, but at a quick glance I
>> don't see any fundamental reason why the task_struct itself can't be given
>> the same treatment.
>>
>
> Any idea about tests to catch performance regressions?
>

I would wager anything fork-heavy with short-lived tasks, say loops of
short hackbench runs, I belive stress-ng also has a fork test case.