Looks like tests time out more frequently with GL layers enabled; they don't appear to be hanging, however (hooray!). I'm betting things are just slower. https://treeherder.mozilla.org/#/jobs?repo=mozilla-inbound&revision=cafa7213ef23c20fd5774db249db63b07314dcd9&selectedJob=30273630

Wow, looks like the talos numbers got shaken up quite a bit by that push; https://treeherder.mozilla.org/perf.html#/compare?originalProject=mozilla-inbound&originalRevision=85825c55516f39c0f3763db8036a47055791936a&newProject=mozilla-inbound&newRevision=cafa7213ef23c20fd5774db249db63b07314dcd9&framework=1&filter=linux Something recent is likely to have regressed our performance with GL layers on Linux significantly; the numbers weren't this dismal the last time I did a push with talos jobs. I wouldn't be surprised if this is related to these timeouts.

Some investigation has revealed that the talos regressions are more than likely because our GLX buffer swaps block until the next VBlank, even under ASAP mode. The unit tests are more than likely running slower due to this as well. Adding support for SGI_swap_control and calling glXSwapIntervalSGI(0) under ASAP mode should resolve this issue (like we do for CGL contexts in https://dxr.mozilla.org/mozilla-central/rev/5f95858f8ddf21ea2271a12810332efd09eff138/gfx/gl/GLContextProviderCGL.mm#122).

Correction; we'll have to use MESA_swap_control or GLX_swap_control. SGI_swap_control prohibits an interval of zero, which is what we want in ASAP mode (no blocking).

Review commit: https://reviewboard.mozilla.org/r/59574/diff/#index_header See other reviews: https://reviewboard.mozilla.org/r/59574/

This patch fixes many of the regressions with ASAP mode talos suites. Comparison using GL layers with ASAP vblanks vs. without ASAP vblanks: https://treeherder.mozilla.org/perf.html#/compare?originalProject=mozilla-inbound&originalRevision=cafa7213ef23&newProject=try&newRevision=2581d23f1890&framework=1&filter=linux Comparison using GL layers with ASAP vblanks vs. basic composition: https://treeherder.mozilla.org/perf.html#/compare?originalProject=mozilla-central&originalRevision=3ce53bd1e25b&newProject=try&newRevision=2581d23f1890&framework=1&filter=linux Many of the remaining non-ASAP test regressions appear to be drawing related, such as tart- in which case, GL layers provides very little benefit. We also get some additional wins from never waiting for vblanks (even with ASAP off), particularly in tresize; https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=2581d23f1890&newProject=try&newRevision=88178bb543f5&framework=1 Lastly, there are a few cases where e10s performance has regressed to be in line with non-e10s performance. This is worth some further investigation, as the e10s numbers appear to be suspiciously low with basic composition.

I didn't realize this before, but the test slaves (i.e. tst-linux64-spot) are running llvmpipe, which I believe does not block on VBlank for buffer swaps. It's quite likely then that we're looking at two separate issues, the other possibly being caused by the additional slowness of software rendering. I'm going to repurpose this bug to fix ASAP mode on the talos machines.

Comment on attachment 8763411 [details] Bug 1280744 - Set swap interval to 0 when running in ASAP mode on GLX. https://reviewboard.mozilla.org/r/59574/#review57748 It's fine, but these fixes would be nice. ::: gfx/gl/GLXLibrary.h:145 (Diff revision 1) > bool UseTextureFromPixmap() { return mUseTextureFromPixmap; } > bool HasRobustness() { return mHasRobustness; } > bool HasCreateContextAttribs() { return mHasCreateContextAttribs; } > bool SupportsTextureFromPixmap(gfxASurface* aSurface); > bool SupportsVideoSync(); > + bool SupportsSwapControl(); bool SupportsSwapControl() const { return bool(xSwapIntervalInternal); } ::: gfx/gl/GLXLibrary.h:245 (Diff revision 1) > PFNGLXGETVIDEOSYNCSGI xGetVideoSyncInternal; > > typedef int (GLAPIENTRY *PFNGLXWAITVIDEOSYNCSGI) (int divisor, int remainder, unsigned int *count); > PFNGLXWAITVIDEOSYNCSGI xWaitVideoSyncInternal; > > + typedef void (GLAPIENTRY *PFNGLXSWAPINTERVALEXT) (Display *dpy, GLXDrawable drawable, int interval); star to left ::: gfx/gl/GLXLibrary.h:260 (Diff revision 1) > bool mUseTextureFromPixmap; > bool mDebug; > bool mHasRobustness; > bool mHasCreateContextAttribs; > bool mHasVideoSync; > + bool mHasSwapControl; We shouldn't need this member.

Comment on attachment 8763411 [details] Bug 1280744 - Set swap interval to 0 when running in ASAP mode on GLX. Review request updated; see interdiff: https://reviewboard.mozilla.org/r/59574/diff/1-2/

Pushed by acomminos@mozilla.com: https://hg.mozilla.org/integration/mozilla-inbound/rev/3c96cf7b1b9c Set swap interval to 0 when running in ASAP mode on GLX. r=jgilbert