1027649 - USS increase for the b2g process when running the performance test.

Status: RESOLVED WORKSFORME

(Firefox OS Graveyard :: Gaia, defect, P2)

Description

[Hubert Figuiere [:hub]]

Run |make test-perf| on a device. 

b2g USS increase over time and never comes back.

If you run on Hamachi right now you'll end up getting errors because of that (256MB). See bug 1020557.

Bug 1027232 should alieviate the concerns about the bug failing, but only as a side effect.

Comment 1

[Hubert Figuiere [:hub]]

Note: testing of this will be influenced by bug 1027232 evolution. I'll update the testing procedure.

Comment 2

[Hubert Figuiere [:hub]]

Adding some MTBF folks in Cc. This issue should fall under MTBF, unless the memory leak is relate to our use of marionette.

We haven't tracked the issue yet, but seen the symptoms.

Comment 3

[Cervantes Yu [:cyu] [:cervantes]]

I just encountered this on unagi, and it leaks very quickly as to trigger the OOM killer instead of the expected lowmem killer. The b2g process is leaking data URL strings containing the marketplace icon. The data URL is referenced in JS from:

python ~/git/heapgraph/g/find_roots.py gc-edges.8117.log 0x49e228e0
Parsing gc-edges.8117.log. Done loading graph. Reversing graph. Done.

via nsXPCWrappedJS::mJSObj :
0x44bd38e0 [Object <no private>]
    --[_manifest]--> 0x49e1fa90 [Object <no private>]
    --[icons]--> 0x49e1faf0 [Object <no private>]
    --[objectElements[120]]--> 0x49e228e0 [string <length 28946> data:image/png;base64,iVBORw]

Found and displayed 1 paths.

Comment 4

[Cervantes Yu [:cyu] [:cervantes]]

Attachment: The decoded data URL

Comment 5

[Cervantes Yu [:cyu] [:cervantes]]

Attachment: The memory report

Comment 6

[Hubert Figuiere [:hub]]

The app icons looks a lot like bug 1040341 :
https://bugzilla.mozilla.org/show_bug.cgi?id=1040341

Cervantes, do you think bug 1040341 is what you are seeing?

Comment 7

[Cervantes Yu [:cyu] [:cervantes]]

This doesn't look like bug 1040341. The key difference is in https://bugzilla.mozilla.org/show_bug.cgi?id=1040341#c1 that the images get discarded on memory pressure. In this bug, we see the memory usage increases to the extent of b2g being killed.

On a side note, I didn't see the memory leak in this bug on a recent m-c build. I am not sure if it is fixed elsewhere or I just missed the key to trigger the leak.

Comment 8

[Wander Lairson Costa]

How much is the USS growth? I am testing with a flame 319mb; before make test-perf b2g-info reports around 57~60mb USS memory. When test-perf finishes, the USS is around 68~70mb. I am using RESTART_B2G=0.

Comment 9

[Hubert Figuiere [:hub]]

It used to grow ~230MB and then get OOM-killed . This was on hamachi as well, so other factor might be in effect.

Comment 10

[Hubert Figuiere [:hub]]

Running it on master (2.1) on an hamachi and I can't reproduce. It may have gone away.

Comment 11

[Wander Lairson Costa]

(In reply to Hubert Figuiere [:hub] from comment #10)
> Running it on master (2.1) on an hamachi and I can't reproduce. It may have
> gone away.

Thanks, hub.

Comment 12

[Wander Lairson Costa]

The bug has gone away, closed.