gpu: Add WebGL robustness telemetry test

gpu: Add WebGL robustness telemetry test

This patch adds a telemetry-based WebGL robustness test. It executes an
optional test from the WebGL conformance test suite to trigger a GPU
context loss and verifies this doesn't crash the browser.

To run the test for example on Android, execute:

$ content/test/gpu/run_gpu_test --browser=android-chrome webgl_robustness

BUG=232449
Committed: https://src.chromium.org/viewvc/chrome?view=rev&revision=215929

[Ken Russell (switch to Gerrit), 2013-07-30 19:39:27]

Please wait for Brandon's review. Overall this LGTM.

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
File content/test/gpu/gpu_tests/webgl_robustness.py (right):

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
content/test/gpu/gpu_tests/webgl_robustness.py:15:
webglTestHarness.notifyFinished();
If the GPU happens to be fast enough to run that test without losing the
context, won't the test be considered to fail?

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
content/test/gpu/gpu_tests/webgl_robustness.py:28: 'description': 'Test cases
for WebGL robustness',
Just checking: this test won't be run by default, will it? The concern is
running it on devices which don't support the robustness extension.

[bajones, 2013-07-30 20:51:03]

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
File content/test/gpu/gpu_tests/webgl_robustness.py (right):

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
content/test/gpu/gpu_tests/webgl_robustness.py:15:
webglTestHarness.notifyFinished();
On 2013/07/30 19:39:27, Ken Russell wrote:
> If the GPU happens to be fast enough to run that test without losing the
> context, won't the test be considered to fail?

Yeah, I worry that this is a race condition as well. The test you are using does
several fairly slow operations, but it's still possible that they could complete
pretty quickly. It would be better to have the script spin in a
requestAnimationFrame or similar loop at the end to ensure it doesn't exit too
quickly.

https://codereview.chromium.org/21150015/diff/1/content/test/gpu/gpu_tests/we...
content/test/gpu/gpu_tests/webgl_robustness.py:33: 'url':
'file:///extra/lots-of-polys-example.html',
Perhaps I'm missing something, but I don't see where this triggers a context
loss. If this script is triggering a context loss spontaneously that sounds like
a bug and we shouldn't be relying on it in tests.

Would it be reasonable to use the WEBGL_lose_context extension instead?
http://www.khronos.org/registry/webgl/extensions/WEBGL_lose_context/

[Ken Russell (switch to Gerrit), 2013-07-30 22:25:23]

On 2013/07/30 20:51:03, bajones wrote:
> Perhaps I'm missing something, but I don't see where this triggers a context
> loss. If this script is triggering a context loss spontaneously that sounds
like
> a bug and we shouldn't be relying on it in tests.

This test submits so much work to the GPU that it's generally expected that it
causes a "real" context lost via GL_ARB_robustness or GL_EXT_robustness.

> Would it be reasonable to use the WEBGL_lose_context extension instead?
> http://www.khronos.org/registry/webgl/extensions/WEBGL_lose_context/

No, that would defeat the purpose of the test.

[bajones, 2013-07-30 22:31:58]

On 2013/07/30 22:25:23, Ken Russell wrote:
> On 2013/07/30 20:51:03, bajones wrote:
> > Perhaps I'm missing something, but I don't see where this triggers a context
> > loss. If this script is triggering a context loss spontaneously that sounds
> like
> > a bug and we shouldn't be relying on it in tests.
> 
> This test submits so much work to the GPU that it's generally expected that it
> causes a "real" context lost via GL_ARB_robustness or GL_EXT_robustness.
> 
> > Would it be reasonable to use the WEBGL_lose_context extension instead?
> > http://www.khronos.org/registry/webgl/extensions/WEBGL_lose_context/
> 
> No, that would defeat the purpose of the test.

Okay, given that then the only suggestion I have is using a rAF loop to prevent
the script from exiting too fast.

LGTM once you've accounted for the race condition.

[Sami, 2013-08-06 15:13:10]

Great idea about running the test in a loop until we lose the context. I've now
implemented that and tested this on Nexus 4 and 7. Nexus 7 passes the test
reliably in about 16 seconds. The new driver for the Nexus 4 (and the new Nexus
7) seems to have improved so that it can run the test loop a couple of times,
but when it actually loses the context it internally calls exit(). I'll chase
that separately.

On 2013/07/30 19:39:27, Ken Russell wrote:
> Just checking: this test won't be run by default, will it? The concern is
> running it on devices which don't support the robustness extension.

We need to add the test to the buildbot config to actually start running it.
I'll do that for Android once we have dedicated GPU bots (crbug.com/256576).
Those devices will all have ARB_robustness support.

[commit-bot: I haz the power, 2013-08-06 15:13:47]

CQ is trying da patch. Follow status at
https://chromium-status.appspot.com/cq/skyostil@chromium.org/21150015/9001

[Sami, 2013-08-06 16:07:18]

FYI, the driver regression is tracked in https://b/10203430.

[commit-bot: I haz the power, 2013-08-06 18:01:44]

Change committed as 215929