870528 - DMD reports 3.2mb of unreported memory in Layer::mAnimationData

Status: RESOLVED WORKSFORME

(Core :: Graphics: Layers, defect)

Description

[Andrew McCreight [:mccr8]]

Attachment: DMD report for Layer::SetAnimations

In fact, Layer appears to have no memory reporter at all.

DMD reports two allocation stacks including the following line representing about 2.4mb of memory on B2G (on a long-running session):
  css::ComputedTimingFunction* ctf = new css::ComputedTimingFunction();

A slice of the DMD report:
  mozilla::layers::Layer::SetAnimations(...)
  mozilla::layers::ShadowLayersParent::RecvUpdate(...)
  nsTArray_base<nsTArrayInfallibleAllocator>::Length() const
  mozilla::layers::PLayersParent::OnMessageReceived(IPC::Message const&)
  mozilla::layers::PCompositorParent::OnMessageReceived(IPC::Message const&)

(Length() isn't present in the second stack, but otherwise it looks the same.)

This timing function is getting put into Layer::mAnimationData.

In bug 870436, there's also about 6mb of unreported image data sitting around.  Could we somehow be leaking animation somehow (not knowing anything about layers or images)?  I don't have the about:memory information for that bug yet, so I'm not sure where the rest of the memory is going.

This is based on a DMD report from bug 870319, where a Unagi run for 24 hours with the monkey tester isn't able to open the home screen any more, probably indicating a leak.

Comment 1

[Andrew McCreight [:mccr8]]

jlebar said there was a recent patch that fixed leaks with these data structures, so we should figure out if that was fixed here, because if it was, we probably have another leak.

Comment 2

[David Zbarsky (:dzbarsky)]

Bug 856080 landed on b2g18 and b2g18-1.0.1 on April 30th.  I think this build is May 8th.

The report claims
Build: sp8810eabase_512x256_hvga-eng 4.0.4.0.4.0.4 OPENMASTER eng.apuser.20130508.210210 test-keys

Comment 3

[Justin Lebar (not reading bugmail)]

> I think this build is May 8th.

It is /possible/ that they didn't build the tip of branch.  I agree, they would have had to pull a pretty old cset, but the DMD report is exactly the same...

Anyway if this is still leaking, that's very bad, I think.

Comment 4

[Justin Lebar (not reading bugmail)]

Per bug 870319 comment 9, they were using an old Gecko.

One bug down...