Re: [PATCH v4 2/6] perf: Enqueue SIGTRAP always via task_work.

[Oleg Nesterov <oleg@redhat.com>]

On 12/05, Frederic Weisbecker wrote:
>
> Le Thu, Dec 05, 2024 at 10:20:16AM +0100, Oleg Nesterov a écrit :
>
> > > Looking at task_work, it seems that most enqueues happen to the current task.
> > > AFAICT, only io_uring() does remote enqueue. Would it make sense to have a light
> > > version of task_work that is only ever used by current? This would be a very
> > > simple flavour with easy queue and cancellation without locking/atomics/RmW
> > > operations.
> >
> > Perhaps, but we also need to avoid the races with task_work_cancel() from
> > another task. I mean, if a task T does task_work_add_light(work), it can race
> > with task_work_cancel(T, ...) which can change T->task_works on another CPU.
>
> I was thinking about two different lists.

OK... but this needs more thinking/discussion.

> Another alternative is to maintain another head that points to the
> head of the executing list. This way we can have task_work_cancel_current()
> that completely cancels the work. That was my initial proposal here and it
> avoids the lock/xchg for each work:
>
> https://lore.kernel.org/all/Zx-B0wK3xqRQsCOS@localhost.localdomain/

Thanks... Heh, I thought about something like this too ;) Although I thought
that we need a bit more to implement task_work_cancel_sync(). But this is
another story.

> > Hmm. I just noticed that task_work_run() needs a simple fix:
> >
> > 	--- x/kernel/task_work.c
> > 	+++ x/kernel/task_work.c
> > 	@@ -235,7 +235,7 @@
> > 			raw_spin_unlock_irq(&task->pi_lock);
> >
> > 			do {
> > 	-			next = work->next;
> > 	+			next = READ_ONCE(work->next);
> > 				work->func(work);
> > 				work = next;
> > 				cond_resched();
> >
> > Perhaps it makes sense before the patch from Sebastian even if that patch
> > removes this do/while loop ?
>
> Hmm, can work->next be modified concurrently here?

work->func(work) can, say, do kfree(work) or do another task_work_add(X, work).

Oleg.