Re: psi_trigger_poll() is completely broken

[Suren Baghdasaryan]

On Mon, Jan 10, 2022 at 10:19 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Mon, Jan 10, 2022 at 9:42 AM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, Jan 10, 2022 at 9:25 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > >
> > > About the issue of serializing concurrent writes for
> > > cgroup_pressure_write() similar to how psi_write() does. Doesn't
> > > of->mutex inside kernfs_fop_write_iter() serialize the writes to the
> > > same file?
> >
> > Ahh, yes, it looks like that does solve the serialization issue.
> > Sorry, I missed that because I'm not actually all that familiar with
> > the kernfs 'of' code.
> >
> > So the only issue is the trigger lifetime one, and if a single trigger
> > is sufficient and returning -EBUSY for trying to replace an existing
> > one is good, then I think that's the proper fix.
> >
> > I'm very busy with the merge window (and some upcoming travel and
> > family events), so I'm hoping somebody will write and test such a
> > patch. Please?
>
> Yes, definitely. I'm on it. Will try posting it later today or
> tomorrow morning if testing reveals something unexpected.

My first attempt to fix this issue is posted at:
https://lore.kernel.org/all/20220111025138.1071848-1-surenb@xxxxxxxxxx/
Couple notes:
- I don't think we need psi_trigger::refcount anymore, therefore it's removed.
- synchronize_rcu is kept to ensure we do not free group->poll_task
while psi_schedule_poll_work is using it.
- Documentation needed minimal changes because it did not clearly
specify how trigger overwrite should work. Now it does.

I ran as many test cases as I could find/create. I'll work on adding
some kselftests for psi triggers to test different usage patterns.
Thanks,
Suren.

> Thanks!
>
> >
> >                    Linus