302737 - [FIX]Plugins leak ns4xPluginStreamListeners and whatever they entrain

Status: RESOLVED FIXED

(Core Graveyard :: Plug-ins, defect, P2)

Description

[Boris Zbarsky [:bzbarsky]]

We end up leaking listeners that are allocated when a plugin calls NPN_GetURL,
with the following stack to the unbalanced addref:

NPN_GetURL (/home/bzbarsky/plugins/libflashplayer.so)
_geturl (../../../../../mozilla/modules/plugin/base/src/ns4xPlugin.cpp:964)
MakeNew4xStreamInternal(_NPP*, char const*, char const*, eNPPStreamTypeInternal,
int, void*, unsigned int, char const*, unsigned char)
(../../../../../mozilla/modules/plugin/base/src/ns4xPlugin.cpp:928)
ns4xPluginInstance::NewNotifyStream(nsIPluginStreamListener**, void*, int, char
const*) (../../../../../mozilla/modules/plugin/base/src/ns4xPluginInstance.cpp:1357)

And indeed, NewNotifyStream addrefs its our param and the caller never releases
it here.  Fixing that makes us crash, since
ns4xPluginStreamListener::CallURLNotify helpfully calls NS_RELEASE_THIS() with
the comment:

235   // Let's not leak this stream listener. Release the reference to the
stream listener 
236   // added for the notify callback in NewNotifyStream. 

Problem is, CallURLNotify has several early returns (in non-failure cases,
even).  So do the methods that call CallURLNotify.  The net result is that it's
pretty easy for us to leak this whole mess.

What I think should happen is that this NS_RELEASE_THIS() should be removed, and
that MakeNew4xStreamInternal should be just release its ptr.  To make sure this
is safe, though, have to trace through the ownership model between this class
and ns4xPluginInstance...

Comment 1

[Boris Zbarsky [:bzbarsky]]

Attachment: Like this, perhaps?

I'm guessing this is 1.9-type stuff at this point.... :(

Comment 2

[Johnny Stenback (:jst)]

Comment on attachment 191025 [details] [diff] [review]
Like this, perhaps?

I'm not going to claim I completely understand the ownership model for the
objects in question here, but from looking at this change and the surrounding
code it *seems* like this is indeed what we want here.

r=jst to give this some mileage on the trunk...

Comment 3

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 191025 [details] [diff] [review]
Like this, perhaps?

jst, want to just make this an r+sr?

Comment 4

[Johnny Stenback (:jst)]

Comment on attachment 191025 [details] [diff] [review]
Like this, perhaps?

Sure, sr=jst too.

Comment 5

[Boris Zbarsky [:bzbarsky]]

Fixed on trunk.

Comment 6

[Daniel Veditz [:dveditz] out until Nov 10]

Is this leak worth fixing for 1.8.0.1 or 1.8.1?

Comment 7

[Boris Zbarsky [:bzbarsky]]

No idea.  More importantly, I can't give a risk evaluation here; I have no idea what this patch could break.

Comment 8

[Mike Schroepfer]

JST - thoughts on a branch patch for this?

Comment 9

[Johnny Stenback (:jst)]

Seems worth doing. It's been on the trunk long enough IMO, and this patch applies cleanly (though with some offset) to the branch as well so IMO we could just land this.

Comment 10

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

jst, want to land it now that it's approved?

Comment 11

[Boris Zbarsky [:bzbarsky]]

Fixed on branch.  Hadn't realized this got approved...

Comment 12

[Daniel Veditz [:dveditz] out until Nov 10]

Comment on attachment 191025 [details] [diff] [review]
Like this, perhaps?

approved for 1.8.0 branch, a=dveditz for drivers

Comment 13

[Boris Zbarsky [:bzbarsky]]

Fixed for 1.8.0.7

Comment 14

[Boris Zbarsky [:bzbarsky]]

So this caused regression bug 354124...  No idea why.  :(

Comment 15

[Boris Zbarsky [:bzbarsky]]

If my analysis in bug 354124 is correct, we probably need to back this out of branches....  and add more plug-in tests to whatever pre-release test suites we run, or not touch plug-in code on branches anymore.

I also feel that this bug is just another indication that trunk is really not getting enough testing in the various modes we support to base decisions on "there haven't been any problems on trunk".