841940 - getUserMedia leaks in crashtests if the permission UI is enabled

Status: RESOLVED WONTFIX

(Core :: WebRTC, defect)

Description

[Randell Jesup [:jesup] (needinfo me)]

Attachment: crashtest output with just 802982 in crashtests.list (1 iteration) and UI enabled

If the permission UI is enabled in crashtests, a single call to getUserMedia({}, {}, {}) will leak a ton of objects, though none appear to be webrtc-related.  Perhaps a document is leaking.  I'll note that crashtests don't have regular chrome, unlike mochitests.


If you replace dom/media/crashtests with this:

default-preferences  pref(media.peerconnection.enabled,true)

load 802892.html

it leaks the world.  You can remove the for loop in 802892.html, and only call mozGetUserMedia() once and it still leaks.

If you add pref(media.navigator.permission.disabled,true) to the default-preferences line, it won't leak (and it doesn't try to contact the UI).

Comment 1

[Randell Jesup [:jesup] (needinfo me)]

Attachment: correct log (previous was 100 iterations)

Comment 2

[Randell Jesup [:jesup] (needinfo me)]

Note: I'm not sure that invoking the permission UI is reasonable in a crashtest.

Comment 3

[Jason Smith [:jsmith]]

(In reply to Randell Jesup [:jesup] from comment #2)
> Note: I'm not sure that invoking the permission UI is reasonable in a
> crashtest.

Probably isn't.

Dao - Any ideas why we leak the world in this bug?

Comment 4

[Jason Smith [:jsmith]]

I actually don't think this one blocks - we are only seeing this happening with fake streams at the moment and may be a crash only issue. If it turns out this ends up being something bigger, then let's reconsider the blocking flag.

Comment 5

[Jason Smith [:jsmith]]

(In reply to Jason Smith [:jsmith] from comment #4)
> I actually don't think this one blocks - we are only seeing this happening
> with fake streams at the moment and may be a crash only issue. If it turns
> out this ends up being something bigger, then let's reconsider the blocking
> flag.

Actually, my wording is wrong here. Let me try again.

I don't think this blocks at the moment because I think this potentially could be an issue that we can't call gUM without fake in CI. If we see this outside of CI or confirm it's possible to be outside of CI, then let's reconsider.

Comment 6

[Randell Jesup [:jesup] (needinfo me)]

dolske said on IRC he didn't believe invoking the permissions UI in a crashtest was a relevant test, and might cause some unexpected issues in that framework.

It would be nice to understand why this happens, but not mandatory.  We may want to consider closing this bug.

Comment 7

[Jason Smith [:jsmith]]

(In reply to Randell Jesup [:jesup] from comment #6)
> dolske said on IRC he didn't believe invoking the permissions UI in a
> crashtest was a relevant test, and might cause some unexpected issues in
> that framework.
> 
> It would be nice to understand why this happens, but not mandatory.  We may
> want to consider closing this bug.

Sounds fine to me.