Filed by: cbook [at] mozilla.com https://treeherder.mozilla.org/logviewer.html#?job_id=78182273&repo=mozilla-inbound https://archive.mozilla.org/pub/firefox/tinderbox-builds/mozilla-inbound-win64/1487303480/mozilla-inbound_win8_64_test-crashtest-e10s-bm109-tests1-windows-build184.txt.gz https://hg.mozilla.org/mozilla-central/raw-file/tip/layout/tools/reftest/reftest-analyzer.xhtml#logurl=https://archive.mozilla.org/pub/firefox/tinderbox-builds/mozilla-inbound-win64/1487303480/mozilla-inbound_win8_64_test-crashtest-e10s-bm109-tests1-windows-build184.txt.gz&only_show_unexpected=1

It looks like this started failing intermittently as soon as it was added. :dholbert - Can you make this test more reliable?

Sorry about that! My first guess is that the onload event must be firing before we've done layout (which is allowed to happen) -- so we make the size-tweak before the first layout ever happens, which means the size-tweak doesn't count as a resize, which means we don't fire the resize handler & never clear "reftest-wait". I'll adjust the test to use MozReftestInvalidate instead of an onload handler -- that should avoid this problem.

Pushed by dholbert@mozilla.com: https://hg.mozilla.org/integration/mozilla-inbound/rev/ce729354cba0 Try to make crashtest 1338772-1.html more reliable. (no review, test-only)

(In reply to Daniel Holbert [:dholbert] from comment #5) > I'll adjust the test to use MozReftestInvalidate instead of an onload > handler -- that should avoid this problem. Turns out this hypothetical-change wasn't good -- it neutered the crashtest such that it no longer triggered the ASAN abort, in pre-fixed builds! So, there must be a semi-race-condition that we're depending on for the bad behavior here, and the MozReftestInvalidate fire-time is late enough that we "win" the race and don't trigger the crash. So I made a slightly different tweak, which I think should still make the resize event fire reliably (and hence should get the "reftest-wait" attribute removed reliably). I verified this tweak still produces a crashy crashtest in pre-fixed builds. (I used the parent of the fix cset.)

This appears to still be happening.

So there still seems to be a race condition here that makes the resize handler fail to get triggered, some fraction of the time. Not sure why. I tried to capture the failure using rr on a local "linux64-qr" configuration (ac_add_options --enable-webrender) -- that's the linux configuration where this fails reliably, it seems. But unfortunately, that build won't run under rr. I filed bug 1343102 on that. Rather than investing more time in investigating & crafting the perfect crashtest here, I'm just going to add a setTimeout() with a failsafe reftest-wait removal, to make us proceed after ~500ms if we happen to go that long without invoking the resize handler.

Pushed by dholbert@mozilla.com: https://hg.mozilla.org/integration/mozilla-inbound/rev/e0565c2ea34c part 2: Add a failsafe to remove "reftest-wait" in crashtest 1338772-1.html, after some time has elapsed. (test-only, no review)