1883399 - Canvas demo can spend 13s in HTMLCanvasElementObserver::UnregisterObserverEvents() (4m+ in extreme case)

Status: NEW

(Core :: Graphics: Canvas2D, defect)

Description

[Mayank Bansal]

1. Go to https://ondras.github.io/primitive.js/
2. Select the attached image file as the input
3. Set the box (3rd column, 2 row) slider to maximum value - 1000
   Extreme 3. (For the "extreme" version, do the above and then set the first slider to maximum value - 500)
4. Click "Lets Go"

AR:
Extreme version: https://share.firefox.dev/3T0sGJF
Normal version (skia-canvas): https://share.firefox.dev/48Nao4M
Normal version (gpu-canvas): https://share.firefox.dev/3TkYhXZ

ER: Maybe something to improve?

(This sounds like it may be bug 1870765 and thus may end up being a dupe.)

cc smaug and continuation
Can repro on a build from Jan2022, so not a recent regression

Comment 1

[Mayank Bansal]

Attachment: Light Bulb - 1024x768.jpg

Comment 2

[Mayank Bansal]

Attachment: about:support

Comment 3

[Andrew McCreight [:mccr8]]

It looks like the time is all under HTMLCanvasElementObserver::UnregisterObserverEvents(). The cycle collector itself can't do anything about it. I'm guessing there's quadratic behavior in that code somehow.

Comment 4

[Timothy Nikkel (:tnikkel)]

Is there a lot of canvas objects? Removing the observer removes from an unordered list so it could be quadratic. We had a similar problem with ImageBitmaps, so we just used a hastset to keep track of imagebitmaps that observed shutdown.

Comment 5

[Timothy Nikkel (:tnikkel)]

Yeah, I see 116000 calls to HTMLCanvasElementObserver::UnregisterObserverEvents.

Comment 6

[Ashley Hale [:ahale]]

Triage - rating as S3, if this appears on websites that aren't merely demos I would consider S2.

Comment 7

[Mayank Bansal]

With gpu-canvas, this also manifests as a "system-level" exponential increase in memory use that will OOM the browser/machine.
Example profile : https://share.firefox.dev/48SgsZw (system-level memory was about 10GB from a baseline of 5.5GB)