Re: [PATCH] task_work: Consume only item at a time while invoking the callbacks.

[Sebastian Andrzej Siewior]

On 2025-02-23 23:40:15 [+0100], Oleg Nesterov wrote:
> Well... I won't really argue because I can't suggest a better fix at
> least right now. Most probably never.
> 
> However, let me say that this patch doesn't make me happy ;) See below.
> 
> On 02/21, Sebastian Andrzej Siewior wrote:
> >
> > Oleg pointed out that this might be problematic if one closes 2.000.000
> > files at once. While testing this scenario by opening that many files
> > following by exit() to ensure that all files are closed at once, I did
> > not observe anything outside of noise.
> 
> and this probably means that we can revert c82199061009 ("task_work: remove
> fifo ordering guarantee") and restore the fifo ordering which IMO makes much
> more sense.

So assume that turning around will fix the problem because the cancel
callback is run first followed by the clean up.

> But:
> 
> > Fixes: c5d93d23a2601 ("perf: Enqueue SIGTRAP always via task_work.")
> 
> Yes. So, to fix this specific problem in perf this patch changes task_work.c
> 
> And after this change we can never enforce a "clear" ordering, fifo or even lifo.
> The ordering is simply "unpredictable/random".
> 
> I'll try to find and read the previous discussions tomorrow, but iirc Frederic
> had another solution?

Two at least:
- having a pointer to the next item
- avoiding the wait in the task_work callback. I think this is the
  unfortunate part. I think he had something but was very unhappy with
  it.

> Oleg.

Sebastian