Re: [Intel-gfx] [PATCH 2/5] kernel.h: Add non_block_start/end()

[Daniel Vetter]

On Fri, Aug 16, 2019 at 3:00 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> On Thu, Aug 15, 2019 at 10:49:31PM +0200, Daniel Vetter wrote:
> > On Thu, Aug 15, 2019 at 10:27 PM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
> > > On Thu, Aug 15, 2019 at 10:16:43PM +0200, Daniel Vetter wrote:
> > > > So if someone can explain to me how that works with lockdep I can of
> > > > course implement it. But afaics that doesn't exist (I tried to explain
> > > > that somewhere else already), and I'm no really looking forward to
> > > > hacking also on lockdep for this little series.
> > >
> > > Hmm, kind of looks like it is done by calling preempt_disable()
> >
> > Yup. That was v1, then came the suggestion that disabling preemption
> > is maybe not the best thing (the oom reaper could still run for a long
> > time comparatively, if it's cleaning out gigabytes of process memory
> > or what not, hence this dedicated debug infrastructure).
>
> Oh, I'm coming in late, sorry
>
> Anyhow, I was thinking since we agreed this can trigger on some
> CONFIG_DEBUG flag, something like
>
>     /* This is a sleepable region, but use preempt_disable to get debugging
>      * for calls that are not allowed to block for OOM [.. insert
>      * Michal's explanation.. ] */
>     if (IS_ENABLED(CONFIG_DEBUG_ATOMIC_SLEEP) && !mmu_notifier_range_blockable(range))
>         preempt_disable();
>     ops->invalidate_range_start();

I think we also discussed that, and some expressed concerns it would
change behaviour/timing too much for testing. Since this does does
disable preemption for real, not just for might_sleep.

> And I have also been idly mulling doing something like
>
>    if (IS_ENABLED(CONFIG_DEBUG_NOTIFIERS) &&
>        rand &&
>        mmu_notifier_range_blockable(range)) {
>      range->flags = 0
>      if (!ops->invalidate_range_start(range))
>         continue
>
>      // Failed, try again as blockable
>      range->flags = MMU_NOTIFIER_RANGE_BLOCKABLE
>    }
>    ops->invalidate_range_start(range);
>
> Which would give coverage for this corner case without forcing OOM.

Hm, this sounds like a neat idea to slap on top. The rand is going to
be a bit tricky though, but I guess for this we could stuff another
counter into task_struct and just e.g. do this every 1000th or so
invalidate (well need to pick a prime so we cycle through notifiers in
case there's multiple). I like.

Michal, thoughts?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch