[Intel-gfx] [PATCH v2] drm/i915: Recover batch pool caches from shrinker

[Chris Wilson]

Quoting Matthew Auld (2018-09-19 19:34:47)
> On Sun, 16 Sep 2018 at 21:59, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >
> > Discard all of our batch pools under mempressure to make their pages
> > available to the shrinker. We will quickly reacquire them when necessary
> > for more GPU relocations or for the command parser.
> >
> > v2: Init the lists for mock_engine
> >
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107936
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> What stops shrink_caches being called while we have a batch pool
> object with its pages still pinned, like in the middle of
> __reloc_gpu_alloc? Maybe I'm missing something...

Once it is pinned for use, it is fine, it cannot be shrunk underneath
us. The problem lies in the zombie reference as we may find ourselves
without an active reference. The only way to be certain would be to
return a strong ref. That just felt like more upheaval than it was
worth (or rather I was hoping that the active-ref would be earlier).
-Chris