[Intel-gfx] [PATCH] drm/i915: Do not invalidate obj->pages under mempressure

[Chris Wilson]

On Thu, Jan 15, 2015 at 01:55:52AM +0100, Daniel Vetter wrote:
> On Wed, Jan 14, 2015 at 08:34:31PM +0000, Chris Wilson wrote:
> > This (partially) reverts
> >
> > commit 5537252b6b6d71fb1a8ed7395a8e5babf91953fd
> > Author: Chris Wilson <chris at chris-wilson.co.uk>
> > Date:   Tue Mar 25 13:23:06 2014 +0000
> >
> >     drm/i915: Invalidate our pages under memory pressure
> >
> > It appears given the right workload, that pages which are swapped out
> > more than once are incorrectly invalidated and discarded. I had presumed
> > that the swapin would mark the pages dirty again and so preserve them
> > against the next cycle of invalidation - that appears to be false, and
> > leads to memory corruption (even leak of stale pages to userspace).
> >
> > Reported-by: Sean V Kelley <sean.v.kelley at intel.com>
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Sean V Kelley <sean.v.kelley at intel.com>
> > Cc: stable at vger.kernel.org
> 
> Hm, scary. Do we have a testcase for this (might need a correctness
> version for some of the swap thrashing tests maybe).

Right, based on my theory about how this screws up, a few passes through
gem_tiled_swapping should be enough to hit the invalidate.

fill_bo -> GTT write, obj->dirty = true -> shinker -> GTT unbind, obj->dirty = false
-> check_bo -> GTT read, obj->dirty = false -> shrinker -> GTT unbind, invalidate
without us calling page_set_dirty()

i.e. a few loops of check_bo should be enough to cycle every object
through swap a couple of times and for subsequent swapin/swapouts not to
cause us to explicity mark the pages as dirty, thus prone to
invalidate_mapping_range() actually removing unwanted pages.

The error does not not show up with igt/gem_tiled_swapping.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre