931455 - WebAudio: "ASSERTION: Ran out of data but track not ended?"

Status: RESOLVED WORKSFORME

(Core :: Web Audio, defect, P3)

Description

[Jesse Ruderman]

Attachment: testcase

###!!! ASSERTION: Ran out of data but track not ended?: 'endTicks == sliceEnd || track->IsEnded()', file content/media/MediaStreamGraph.cpp, line 855

Comment 1

[Jesse Ruderman]

Attachment: stack

Comment 2

[Maire Reavy [:mreavy]]

Karl --- Can you look at this?

Comment 3

[Karl Tomlinson (:karlt)]

Might be another case like bug 926606, where it is expected that upstream has already produced output, which won't necessarily be the case in a cycle.

I need to work out whether there is a suitable general strategy here, such as providing a silent upstream source, or we need to patch up all the possible downstream code.

Another option to consider is just not processing streams that are in a cycle without a DelayNode.

Comment 4

[Robert O'Callahan (:roc) (email my personal email if necessary)]

(In reply to Karl Tomlinson (:karlt) from comment #3)
> Another option to consider is just not processing streams that are in a
> cycle without a DelayNode.

I guess this is slightly underspecified. We have said that all nodes in a cycle without a DelayNode produce silence. However it's not clear what this means for the internal state of such nodes. Maybe we should change it to every *edge* in the cycle --- i.e. if node A is connected into node B, and that connection is part of a cycle which does not contain a DelayNode, then node A's input to node B is treated as silence.

Comment 5

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Actually, let's say in that case node A's output to all nodes is treated as silence --- but nodes A and B still keep doing all their processing, it's just that most of that processing will be processing silence and the output will be ignored.

Comment 6

[Robert O'Callahan (:roc) (email my personal email if necessary)]

We talked some more. We decided to go with something more like comment #4: if a connection from node A to node B is part of a zero-delay cycle, then we ignore that connection for the purposes of processing. This is easy to implement for all kinds of processing nodes; we just skip such an InputPort when we iterate through the incoming InputPorts of MediaStream B.

We'll have to modify the cycle-detection logic so that instead of marking MediaStreams as belonging to a zero-delay cycle, we mark InputPorts instead. If we find any such InputPorts then we may need to rerun MediaStream sorting ignoring the marked InputPorts. (It may be possible to implement the cycle-detection and sorting so that this can be done in one pass, but it sounds complex.)

Comment 7

[Jesse Ruderman]

Still happens after https://hg.mozilla.org/mozilla-central/rev/2bf5b49d85e7, but now fatal:

Assertion failure: track->IsEnded() (Not enough data, and track not ended.), at content/media/MediaStreamGraph.cpp:951

Comment 8

[Karl Tomlinson (:karlt)]

Like bug 926606, this no longer reproduces.