Closed Bug 1441640 Opened 3 years ago Closed 3 years ago
Frequent reftest-with-gpu failures in css-writing-mode tests on Windows 7 debug when Gecko 60 merges to Beta on 2018-03-01
css-writing-mode reftests fail frequently in central-as-beta simulations but not on central. If there is a test failure, there seem to be multiple test failures in the same chunk, but only in the central-as-beta simulations, not with mozilla-central. central-as-beta simulation: https://treeherder.mozilla.org/#/jobs?repo=try&revision=902ef48cdb6fbdf82a8fc6a0bd50f8f35b705038&filter-resultStatus=testfailed&filter-resultStatus=busted&filter-resultStatus=exception&filter-resultStatus=retry&filter-resultStatus=usercancel&filter-resultStatus=runnable&filter-resultStatus=success&group_state=expanded&filter-searchStr=Rg7&selectedJob=164598966 Failure log: https://treeherder.mozilla.org/logviewer.html#?job_id=164648584&repo=try same revision on central without failures: https://treeherder.mozilla.org/#/jobs?repo=mozilla-central&revision=b184be59874080e96903183176c0f88dcbfafe25&filter-searchStr=windows%20debug%20rg7&group_state=expanded The issue also reproduced with one of yesterday's uplift simulations: https://treeherder.mozilla.org/#/jobs?repo=try&revision=0c6b6d66cbfb42a470dda027a9104d9c24f7ee60&group_state=expanded&filter-searchStr=windows%20debug%20rg&selectedJob=164677722 Any idea what is causing this different behavior between central and beta?
(In reply to Sebastian Hengst [:aryx][:archaeopteryx] (needinfo on intermittent or backout) from comment #0) > css-writing-mode reftests fail frequently in central-as-beta simulations but > not on central. > > If there is a test failure, there seem to be multiple test failures in the > same chunk, but only in the central-as-beta simulations, not with > mozilla-central. > > central-as-beta simulation: > https://treeherder.mozilla.org/#/ > jobs?repo=try&revision=902ef48cdb6fbdf82a8fc6a0bd50f8f35b705038&filter- > resultStatus=testfailed&filter-resultStatus=busted&filter- > resultStatus=exception&filter-resultStatus=retry&filter- > resultStatus=usercancel&filter-resultStatus=runnable&filter- > resultStatus=success&group_state=expanded&filter- > searchStr=Rg7&selectedJob=164598966 > Failure log: > https://treeherder.mozilla.org/logviewer.html#?job_id=164648584&repo=try This is an example of bug 1392106; the characteristic sign of this in the log is the presence of the skia "Access is denied" errors: 16:36:15 INFO - Could not create glyph run analysis. 16:36:15 INFO - z:/build/build/src/gfx/skia/skia/src/ports/SkScalerContext_win_dw.cpp(508) : error 0x80070005: Access is denied. 16:36:15 INFO - Requested bounding box could not be determined. 16:36:15 INFO - z:/build/build/src/gfx/skia/skia/src/ports/SkScalerContext_win_dw.cpp(583) : error 0x80070005: Access is denied. It's quite common that when this happens, a whole series of tests (that all use the same font) may be affected. > > same revision on central without failures: > https://treeherder.mozilla.org/#/jobs?repo=mozilla- > central&revision=b184be59874080e96903183176c0f88dcbfafe25&filter- > searchStr=windows%20debug%20rg7&group_state=expanded > > The issue also reproduced with one of yesterday's uplift simulations: > https://treeherder.mozilla.org/#/ > jobs?repo=try&revision=0c6b6d66cbfb42a470dda027a9104d9c24f7ee60&group_state=e > xpanded&filter-searchStr=windows%20debug%20rg&selectedJob=164677722 I didn't go through all the failures there, but the ones I checked looked like bug 1392106 again. > > Any idea what is causing this different behavior between central and beta? Not really. Does reftest chunking work out differently, perhaps? My current suspicion is that there may be some kind of temporary resource exhaustion in the windows graphics world that triggers the failure, and so any change to chunking may affect how likely it is to happen within a given chunk, and which tests are most likely to be the victims. But that's little more than hand-wavy speculation, really.
Component: Layout: Text → Graphics: Text
Depends on: missing-letter-win7
one large difference between beta/central is that on central we run reftests per manifest now instead of as a whole and chunk based on number of tests. This means we restart the browser often in each chunk. It is odd we do not see these failures on central. Possibly what :jfkthame mentioned related to Skia or other graphics- maybe there is a difference in the beta configs which cause issues.
3 years ago
There are bugs for the individual failures e.g. bug 1442637, closing this one.
Status: NEW → RESOLVED
Closed: 3 years ago
Resolution: --- → WORKSFORME
You need to log in before you can comment on or make changes to this bug.