653026 - GC hazard with NoSuchMethod

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

Attachment: testcase

This script crashes in a shell with no jit options.

Comment 1

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

Attachment: patch

The basic problem is that the new object being allocated in js_OnUnknownMethod is the only thing that roots the method. This object is not scanned during a GC because it has a NULL shape, so it's treated as a newborn object. This was apparently sort of intentional, to ensure that these objects don't escape into user code. However, I think it's more important that we scan them.

I set the shape so that the object looks non-native, and then I added a trace hook to the class. I thought about making it a native object, but I think then we'd need a proto for it, and I didn't want to deal with that. Please correct me if I'm wrong, Jeff.

Comment 2

[Jeff Walden [:Waldo]]

Comment on attachment 528630 [details] [diff] [review]
patch

Sounds reasonable.

Comment 3

[Johnny Stenback (:jst)]

Can this land now that it's reviewed n' all? Would this patch be safe for beta?

Comment 4

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

Sorry for the delay. I meant to land it yesterday but I forgot. I'll make sure it gets into the beta.

Comment 5

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

http://hg.mozilla.org/tracemonkey/rev/10ba4918f08e

I pushed without the test. That needs to go in later.

Comment 6

[Chris Leary [:cdleary] (not checking bugmail)]

cdleary-bot mozilla-central merge info:
http://hg.mozilla.org/mozilla-central/rev/10ba4918f08e

Comment 7

[David Mandelin [:dmandelin]]

http://hg.mozilla.org/mozilla-beta/rev/cb0b3bc9c53b

Comment 8

[David Mandelin [:dmandelin]]

This made it to mozilla-aurora as

http://hg.mozilla.org/mozilla-aurora/rev/10ba4918f08e

Comment 9

[Daniel Veditz [:dveditz]]

On the 1.9.2 branch I get "TypeError: obj.x(a()) is not a function" rather than a crash. Does that mean that branch is not affected by this bug? Do we know what regressed it?

Comment 10

[Bill McCloskey [inactive unless it's an emergency] (:billm)]

It looks like the testcase in the bug regressed in bug 584917, so 3.6 should be safe. Before that patch, we rooted the function to call via the NoSuchMethod object's parent (or proto, I forget the argument order).

I'm still a bit worried about whether the other thing stored in the NoSuchMethod object in rooted. It can only be a jsid, but it seems like e4x might somehow allow an object to be stashed there. I don't really understand the code too well though. It would be good if someone who understands e4x better could take a look. I'm guessing this means Brendan or Igor?

Comment 11

[Daniel Veditz [:dveditz]]

Thanks, for now we'll not track this on the old branch.

Comment 12

[Virgil Dicu [:virgil] [QA]]

Mozilla/5.0 (X11; Linux x86_64; rv:6.0) Gecko/20100101 Firefox/6.0

Could you please provide some clear STR in order to verify this issue?