767479 - Crash in NS_StackWalk

Status: RESOLVED FIXED

(Core :: Gecko Profiler, defect)

Description

[Vladan Djeric (:vladan)]

Breakpad report:

https://crash-stats.mozilla.com/report/index/bp-d76c3a07-8f98-40a4-9bfd-7f4192120622

Comment 1

[Scoobidiver (away)]

It first appeared in 16.0a1/20120619. The regression range might be:
http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=b1a0fb2bdbf7&tochange=373e6f9264e6

Comment 2

[Vladan Djeric (:vladan)]

(In reply to Scoobidiver from comment #1)
> It first appeared in 16.0a1/20120619. The regression range might be:
> http://hg.mozilla.org/mozilla-central/
> pushloghtml?fromchange=b1a0fb2bdbf7&tochange=373e6f9264e6

You're right, this bug depends on frame pointers being turned so it only would have caused an uptick in crash reports when we enabled frame pointers on mozilla-central Nightly. The bug itself has been around for a while though. Ehsan has an idea about the root causes and is working on a patch.

Comment 3

[bomfog]

I hit this reliably on ionmonkey nightly builds of 21 and 22 June (post- bug #764216 merge-to-ion), brand new profile with only Gecko Profiler extension installed and the stack walking option turned on, running kraken benchmark.

That setup crashes every time by/at the end of the benchmark. Stack walking turned off doesn't crash. Mozilla central nightly builds don't crash, regardless of stackwalk option. 

32 bit Windows Vista.

Comment 4

[(no longer active)]

What happens here is that if we have more than 1024 frames on the stack to walk, WalkStackMain64 starts to increment pc_count past that, which then causes the code here: <http://hg.mozilla.org/mozilla-central/diff/b9b9d9f379db/xpcom/base/nsStackWalk.cpp> to read more elements than there are inside the buffer.

I think the right fix is to apply the same checks that we perform when walking the stack from another thread in the case of walking a stack with a thread handle specified.

Patch coming up.

Comment 5

[(no longer active)]

Attachment: Patch (v1)

I also fixed the fallible malloc call which was not being followed by a null check.

Comment 6

[(no longer active)]

http://tbpl.mozilla.org/?tree=Try&rev=e9efe37ec040

Comment 7

[David Baron :dbaron: (⌚️UTC-5, no longer working on Mozilla)]

Comment on attachment 635887 [details] [diff] [review]
Patch (v1)

It seems odd to use alloca for one codepath and moz_xmalloc for the other.  Why not use the same for both?

Otherwise, r=dbaron.

Merging with bug 766579 requires care.

Comment 8

[(no longer active)]

Hmm, you're right.  I see no reason to not use _alloca in both codepaths.

Comment 9

[(no longer active)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/5b03b3e56196

Comment 10

[(no longer active)]

(In reply to Ehsan Akhgari [:ehsan] from comment #9)
> https://hg.mozilla.org/integration/mozilla-inbound/rev/5b03b3e56196

I forgot to fix the commit message to reflect comment 8.  :(

Comment 11

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/5b03b3e56196