1322818 - Reenable the layout/reftests/bugs/1114526-1.html reftest

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect, P3)

Description

[Kartikaya Gupta (email:[e-mail])]

In the interest of getting some reftest jobs running on TreeHerder sooner rather than later, I disabled the layout/reftests/bugs/1114526-1.html reftest in bug 1319188. This bug is to re-enable it after figuring out why it intermittently fails. In the failure cases it seems to loop indefinitely.

Comment 1

[Kartikaya Gupta (email:[e-mail])]

For the record, here's an instance of the failure in question: https://treeherder.mozilla.org/logviewer.html#?job_id=32475943&repo=try#L9079

Comment 2

[Kartikaya Gupta (email:[e-mail])]

I was able to reproduce this locally with rr chaos mode, and it looks like there's an invalid region set on a WebRenderPaintedLayer that seems to never get cleared. As a result, the LayerTreeInvalidation code keeps producing an invalid region from [1] and the harness keeps painting, never actually firing the MozReftestInvalidate event.

I suspect we need to be clearing the invalid region somewhere inside Render(), or something.

[1] http://searchfox.org/mozilla-central/rev/594937fec2e2fc45fa9308ba2fb964816631f017/gfx/layers/LayerTreeInvalidation.cpp#212

Comment 3

[Kartikaya Gupta (email:[e-mail])]

Attachment: Backtrace to setting invalid region

Turns out the invalid region is getting set here (attached stack). My line numbers might be off since they local changes, but frame 3 is at http://searchfox.org/mozilla-central/rev/594937fec2e2fc45fa9308ba2fb964816631f017/layout/painting/FrameLayerBuilder.cpp#4625

Comment 4

[Kartikaya Gupta (email:[e-mail])]

After chasing this some more I think it might boil down to the fact that in non-WR builds, we don't fire a DidComposite message when we do a snapshot into a DrawTarget. In WR builds we do. The DidComposite message triggers the MozAfterPaint event which interferes with the reftest harness.

Comment 5

[Kartikaya Gupta (email:[e-mail])]

Hopefully this fix is solid enough:

https://treeherder.mozilla.org/#/jobs?repo=try&revision=28bab9af267b2cba506a687e72da289d7298ef4e

Comment 6

[Kartikaya Gupta (email:[e-mail])]

^ That made things worse, not better. I'll have to dig into this more, but not right now.

Comment 7

[Kartikaya Gupta (email:[e-mail])]

I tried reproducing this again locally but didn't see it. It might have gotten fixed already. Try push to see if it's still happening on tryserver: https://treeherder.mozilla.org/#/jobs?repo=try&revision=e2694239134b258dc51c1353c5a7a90550b03467

Comment 8

[Kartikaya Gupta (email:[e-mail])]

Attachment: Bug 1322818 - Re-enable skipped reftest that seems to be passing consistently now. ?

Review commit: https://reviewboard.mozilla.org/r/107854/diff/#index_header
See other reviews: https://reviewboard.mozilla.org/r/107854/

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/projects/graphics/rev/6bfcb75c9120
Re-enable skipped reftest that seems to be passing consistently now. r=gfx?

Comment 10

[Kartikaya Gupta (email:[e-mail])]

Comment on attachment 8831298 [details]
Bug 1322818 - Re-enable skipped reftest that seems to be passing consistently now. ?

https://reviewboard.mozilla.org/r/107854/#review131010