477066 - nsPluginHostImpl leaks when running mochitest in the "console" session as opposed to an RDP one

Status: RESOLVED INCOMPLETE

(Core Graveyard :: Plug-ins, defect)

Description

[[e-mail] (:bhearsum)]

I'm doing some work on our win32 build machines right now that may result in us running tests from the "console" session rather than an RDP one. During initial tests I've seen a consistent leak in nsPluginHostImpl happening on mozilla-1.9.1, mozilla-central, and tracemonkey unit tests. Here's the details:
== BloatView: ALL (cumulative) LEAK AND BLOAT STATISTICS

     |<----------------Class--------------->|<-----Bytes------>|<----------------Objects---------------->|<--------------References-------------->|
                                              Per-Inst   Leaked    Total      Rem      Mean       StdDev     Total      Rem      Mean       StdDev
   0 TOTAL                                          17      188 1555149935        4 (53352.96 +/-     0.00) 346121445        3 (20019.70 +/-     0.00)
 667 nsLocalFile                                    88       88    33989        1 (  238.51 +/-    43.25)   123213        1 (  369.50 +/-    84.46)
 760 nsPluginHostImpl                               88       88        1        1 (    1.00 +/-     0.00)     4354        1 (   11.77 +/-     3.71)
 943 nsStringBuffer                                  8        8  7347552        1 (25130.85 +/-  4469.28) 13495209        1 (27574.96 +/-  3435.31)
1033 nsVoidArray                                     4        4 67579154        1 (10810.39 +/-  2643.72)        0        0 (    0.00 +/-     0.00)

nsTraceRefcntImpl::DumpStatistics: 1158 entries
TEST-UNEXPECTED-FAIL | runtests-leaks | leaked 188 bytes during test execution
TEST-UNEXPECTED-FAIL | runtests-leaks | leaked 1 instance of nsLocalFile with size 88 bytes
TEST-UNEXPECTED-FAIL | runtests-leaks | leaked 1 instance of nsPluginHostImpl with size 88 bytes
TEST-UNEXPECTED-FAIL | runtests-leaks | leaked 1 instance of nsStringBuffer with size 8 bytes
TEST-UNEXPECTED-FAIL | runtests-leaks | leaked 1 instance of nsVoidArray with size 4 bytes

On 1.9.1 there's an additional leak of nsWeakReference:
1034 nsWeakReference                                12       12    64516        1 (  799.24 +/-   272.73)   419883        1 ( 1944.96 +/-   831.27)

I can reproduce this 100% of the time so I'm happy to help track it down.

Comment 1

[[e-mail] (:bhearsum)]

Any chance someone can look at this? When bug 428123 gets rolled out this bug is going to have to be fixed, or we're going to have to adjust the mochitest leak threshold.

Comment 2

[[e-mail] (:bhearsum)]

Attachment: balance tree from a leaking run

Comment 3

[[e-mail] (:bhearsum)]

Attachment: proper balance tree

Comment 4

[[e-mail] (:bhearsum)]

Attachment: balance tree with proper symbols

Comment 5

[[e-mail] (:bhearsum)]

Attachment: latest balance tree

I don't think the previous balance tree is quite right either, this is my most recent one, with --disable-optimize and --disable-debug. Generated after leaking on test_HTMLDocument07.html.

Comment 6

[[e-mail] (:bhearsum)]

Attachment: balance tree, again

bent suggested trying to find this leak without mochitest. This is the balance tree I got after doing so.

Comment 7

[[e-mail] (:bhearsum)]

Interestingly, I can't reproduce this on mozilla-central. I had a look over the plugin related checkins since the start of the year and a few popped out to me:
bug 476409 - http://hg.mozilla.org/mozilla-central/rev/a7853b4dbd34
bug 474022 - http://hg.mozilla.org/mozilla-central/rev/011d2208412c and http://hg.mozilla.org/mozilla-central/rev/bf1556e3ed4b
bug 474866 - http://hg.mozilla.org/mozilla-central/rev/0932e7c4d976 and http://hg.mozilla.org/mozilla-central/rev/abd5a41c0b70

The last two look particularly suspicious given that we sometimes leak an nsWeakReference. But then again, they look like they're on 1.9.1 already, so probably not!

Comment 8

[[e-mail] (:bhearsum)]

I don't think we care about this anymore.