I can reliably reproduce this error. Running the composer crashtests under valgrind also reliably reproduce crashtest hanging on exit, while running other crashtest files does not reproduce the hang.
Created attachment 328259 [details] valgrind log To reproduce: 1) ./run-mozilla.sh /bin/bash 2) valgrind ./firefox-bin --reftest=/home/sayrer/dev/mozilla-central/editor/composer/src/crashtests/crashtests.list &> valgrind-composer.log
The errors of interest are ==6038== Conditional jump or move depends on uninitialised value(s) ==6038== at 0x5D1F583: nsIHTMLDocument::IsEditingOn() (nsIHTMLDocument.h:159) and the assertions nearby.
Peter can you take a look since hg-blame pins those lines on you? We've been having crashtest hangs repeatedly on linux so would be great to figure this out..
That's a bit odd, IsEditingOn returns GetEditingState() == eDesignMode || GetEditingState() == eContentEditable and GetEditingState returns nsHTMLDocument's mEditingState, which should be set to 0 by the nsIDocument's zeroing new() operator.
Created attachment 328320 [details] [diff] [review] Fix valgrind warning This probably fixes the valgrind warnings (Robert, can you confirm that?), but it shouldn't make any difference for the hang unless our zeroing operator new isn't working. The assertions are a problem and we should fix them at some point, but they shouldn't cause hangs either.
I can reproduce the valgrind error using a stock jemalloc-enabled build. In a build with "ac_add_options --with-valgrind" I don't get the error, so I think this is a bogus error from valgrind, see bug 424040 comment 10.
(In reply to comment #5) > This probably fixes the valgrind warnings It doesn't help. I think --with-valgrind is required for reliable valgrind results on jemalloc builds.
(In reply to comment #7) > It doesn't help. I think --with-valgrind is required for reliable valgrind > results on jemalloc builds. Robert, did you set that option? If not could you rerun with it set? I'm going to be away till Wednesday next week, so if this is urgent you should look for a better owner.
yep, that's it. sorry.
I'm confused. Is peterv's patch in comment 5 bogus? This crashtest is causing crashes for me, btw.
Jesse: it shouldn't make a difference. AFAIK Robert was building without the required --with-valgrind. I think it'd be better to file a new bug on the crash.