Prefer GL_ARB_sync/ES3 fences

Prefer GL_ARB_sync/ES3 fences

They are a better option since server-side wait is supported by default.
It also avoids the broken EGL_KHR_fence_sync implementation
on Qualcomm drivers.

BUG=371530, 278606, 341867
R=piman@chromium.org, reveman@chromium.org

Committed: https://src.chromium.org/viewvc/chrome?view=rev&revision=269863

[no sievers, 2014-05-12 18:03:40]

ptal

[reveman, 2014-05-12 18:20:22]

can we add a unit test that check if the preferred fence type works properly?

[no sievers, 2014-05-12 18:27:20]

On 2014/05/12 18:20:22, reveman wrote:
> can we add a unit test that check if the preferred fence type works properly?

There is a gl_test that tests async readback with testing fences.

We also do run gl_tests on the waterfall, but we don't have a waterfall with a
wide variety of relevant devices+drivers.
Right now we seem to be running N4 with KOT49H. It might have the fix. I'd
expect that when we previously enabled async readbacks, since the waterfall did
not catch this it must have the fix. Or we were running an older driver back
then.

[piman, 2014-05-12 18:49:57]

lgtm

[reveman, 2014-05-12 18:58:54]

On 2014/05/12 18:27:20, sievers wrote:
> On 2014/05/12 18:20:22, reveman wrote:
> > can we add a unit test that check if the preferred fence type works
properly?
> 
> There is a gl_test that tests async readback with testing fences.
> 
> We also do run gl_tests on the waterfall, but we don't have a waterfall with a
> wide variety of relevant devices+drivers.
> Right now we seem to be running N4 with KOT49H. It might have the fix. I'd
> expect that when we previously enabled async readbacks, since the waterfall
did
> not catch this it must have the fix. Or we were running an older driver back
> then.

Ok, lgtm. I assume those async readback tests are failing on affected devices
prior to this fix then.

[no sievers, 2014-05-12 19:35:34]

On 2014/05/12 18:58:54, reveman wrote:
> On 2014/05/12 18:27:20, sievers wrote:
> > On 2014/05/12 18:20:22, reveman wrote:
> > > can we add a unit test that check if the preferred fence type works
> properly?
> > 
> > There is a gl_test that tests async readback with testing fences.
> > 
> > We also do run gl_tests on the waterfall, but we don't have a waterfall with
a
> > wide variety of relevant devices+drivers.
> > Right now we seem to be running N4 with KOT49H. It might have the fix. I'd
> > expect that when we previously enabled async readbacks, since the waterfall
> did
> > not catch this it must have the fix. Or we were running an older driver back
> > then.
> 
> Ok, lgtm. I assume those async readback tests are failing on affected devices
> prior to this fix then.

Yes, when I originally discovered the bug, the existing test was suitable to
reproduce it.

[no sievers, 2014-05-12 20:12:08]

Committed patchset #2 manually as r269863 (presubmit successful).