[Intel-gfx] gem_tiled_fence_blits and OOM

[Daniel Vetter]

On Tue, Apr 14, 2015 at 09:39:13PM +0000, Mason, Michael W wrote:
> We're running i-g-t on Chrome OS and seeing gem_tiled_fence_blits
> failures.  More specifically, the test is getting killed by the OOM
> killer, sometimes leading to hangs and reboots.  This happens
> consistently when the UI is running.  When the UI is stopped, the test
> passes.  This makes sense since there's more memory available.
> 
> I noticed the test uses intel_get_total_ram_mb() in the calculation that
> determines how many buffers to allocate.  Excuse the perhaps naïve
> question, but shouldn't that be intel_get_avail_ram_mb()?  When I change
> the test to use that call it passes consistently with no OOM issues. I
> realize that might be the wrong fix.  I just want to understand why.

igt are bare-metal kernel tests, and if you run them using the piglit
runner (which is the recommended way) it'll bail if any other drm client
is running. All the tests make this assumption, the avail_ram function is
just informational really (since to speed things up we mlock large amounts
of memory). The reason is that we want to make sure that gem bos get
pushed through swap, and not everything else running on your system.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch