[Intel-gfx] [PATCH] drm/i915/sprite: Fix race of vblank irq vs wait in vblank evasion

[Chris Wilson]

On Mon, Aug 22, 2016 at 06:15:53PM +0100, Chris Wilson wrote:
> Hmm, prepare_to_wait() itself has a might_sleep() check, preventing the
> preempt fun. The challenge is that we do not want to be interrupted
> once we decide to apply the update (reading the scanline) - but that
> read has to be after prepare_to_wait to prevent any races before
> sleeping.

Scratch that, didn't read the warning carefully enough and was thrown by
an unreliable frame: 

[   32.616197]  [<ffffffff8104583a>] warn_slowpath_fmt+0x4a/0x50
[   32.616208]  [<ffffffff8107ff6b>] ? prepare_to_wait+0x2b/0x90
[   32.616222]  [<ffffffff8107ff6b>] ? prepare_to_wait+0x2b/0x90
[   32.616229]  [<ffffffff810686b7>] __might_sleep+0x77/0x80

That of course is the wrong way around for prepare_to_wait to be doing
the might_sleep. The issue was just not rearranging the control flow
correctly to finish_wait.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre