1888634 - OffscreenCanvas Codepen demo sometimes completely freezes the tab for 10-15 seconds

Status: RESOLVED FIXED

(Core :: Graphics: Canvas2D, defect, P3)

Description

[Mayank Bansal]

This may be easier to repro with gpu-canvas enabled.

1. Download the attached demo and save it to your local machine (original demo at https://codepen.io/rroyerrivard/pen/jOwBLbB may not work)
   1.5 Open the testcase on your machine. .
2. Allow the page access to your camera permanently
3. Click on the first button "Start Worker"
4. Switch to another tab for 10-15 seconds
5. Switch back to the demo tab and wait for 3-4 seconds
6. Rapidly click on the second button "Add Canvas to page" 10-15 times
   If it doesnt repro, close the browser and open the testcase again.

AR: The page freezez
ER: Not so

Profile with all threads, IPC and File IO: https://share.firefox.dev/3VIWuxH
Profile with gfx preset logging: https://share.firefox.dev/4cBrBAX
Profile with webrtc preset logging: https://share.firefox.dev/3x8VxUU

Can only repro about 50% of the time so cannot find a regression range. Not even sure if this is a canvas or webrtc bug or something else.

Comment 1

[Mayank Bansal]

Attachment: Standalone_Testcase.html

Comment 2

[Mayank Bansal]

Attachment: about:support

Comment 3

[Mayank Bansal]

FWIW, I also got drive crash/reset on two occasions while testing this.
Here is a crash report generated on force crashing : https://crash-stats.mozilla.org/report/index/1c14fcbe-ef0a-432a-9966-423d50240329

gfx-log:
(#0) Error Killing GPU process due to IPC reply timeout
(#1) Error Failed as lost WebRenderBridgeChild.
(#2) Error Fallback WR to SW-WR + D3D11
(#3) Error Fallback SW-WR + D3D11 to SW-WR
(#4) Error CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#5) Error CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#6) Error CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#7) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#8) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#9) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#10) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#11) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown
(#12) CP+[GFX1-]: Failed as lost WebRenderBridgeChild.
(#13) CP+[GFX1-]: CompositorBridgeChild receives IPC close with reason=AbnormalShutdown

Comment 4

[Mayank Bansal]

Alternate STR2:

Download the attached demo and save it to your local machine
Open the testcase on your machine. Allow the page access to your camera permanently
Press-and-hold the "f5" key on your keyboard so that the the page reloads rextremely rapidly. Let this continue for 4-5 seconds.
Click on the first button "Start Worker"
Rapidly click on the second button "Add Canvas to page" 10-15 times

Alternate STR3:
Open the testcase from your local machine
Click on start the worker and click on add a canvas
Start rapidly resizing the browser window diagonally. Keep on resizing till you notice that the video has frozen.

Comment 5

[Andrew Osmond [:aosmond] (he/him)]

Does this still happen now that bug 1887729 landed?

Comment 6

[Andrew Osmond [:aosmond] (he/him)]

Note that the codepen doesn't need to use oncontextlost/restored to benefit from the improvement where we avoid the freezing after a GPU process crash.

Comment 7

[Mayank Bansal]

(In reply to Andrew Osmond [:aosmond] (he/him) from comment #5)

Does this still happen now that bug 1887729 landed?

I have the latest Nightly from this: https://hg.mozilla.org/mozilla-central/rev/4120fb3d12f57159ca2e6809f555312e00adda5d and I can still repro, I just tried the STR3 and the tab froze.
(bug 1887729 appears to have landed about 2 nightlies prior)

Comment 8

[Brad Werth [:bradwerth]]

I can reproduce on macOS using STR in comment 0, and I get the hang the very first time I click "Add canvas to page". In fact, the highlight on the button is never cleared. I will attempt a mozregression.

Comment 9

[Brad Werth [:bradwerth]]

Weirdly, mozregression points to Bug 1806400. I'll do another mozregression to double-check.

Comment 10

[Brad Werth [:bradwerth]]

A more careful mozregression indicates Bug 1855742, which makes a lot more sense.

Comment 11

[BugBot [:suhaib / :marco/ :calixte]]

Set release status flags based on info from the regressing bug 1855742

Comment 12

[Andrew Osmond [:aosmond] (he/him)]

I am speculating that this is caused by us clearing out the tab cache (i.e. the ClearCachedResources path), we end up using a stale remote texture ID somehow, and block in the RemoteTextureMap code:
https://searchfox.org/mozilla-central/rev/c09764753ea40725eb50decad2c51edecbd33308/gfx/layers/RemoteTextureMap.cpp#988

Comment 13

[Andrew Osmond [:aosmond] (he/him)]

Actually this is related to capture stream. The test case is a bit contrived in that it feeds the capture stream back into the canvas.

Comment 14

[Andrew Osmond [:aosmond] (he/him)]

Attachment: Bug 1888634 - Disallow captureStream with canvas transferred to DOM worker.

This behaviour has yet to be defined by the spec and/or WPT, see:

https://github.com/w3c/mediacapture-fromelement/issues/65
https://github.com/w3c/mediacapture-extensions/pull/26
https://github.com/web-platform-tests/wpt/issues/21102

Comment 15

[BugBot [:suhaib / :marco/ :calixte]]

Set release status flags based on info from the regressing bug 1855742

Comment 16

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/59e2426d20ae
Disallow captureStream with canvas transferred to DOM worker. r=gfx-reviewers,lsalzman

Comment 17

[Ryan VanderMeulen [:RyanVM]]

Was the demo in comment 0 reduced from a live site or...? Would be nice to better understand how the testcase was found/created to assess the real-world impact of this bug.

Comment 18

[Mayank Bansal]

(In reply to Ryan VanderMeulen [:RyanVM] from comment #17)

Was the demo in comment 0 reduced from a live site or...? Would be nice to better understand how the testcase was found/created to assess the real-world impact of this bug.

I just search for "offscreen-canvas" on Codepen, and test all the results that come up.
I am not the author of the demo in comment 0, and I dont know if this is related to any real site or not.

Comment 19

[Andrew Osmond [:aosmond] (he/him)]

My understanding is that use of captureStream is fairly niche.

Comment 20

[Andrew Osmond [:aosmond] (he/him)]

Comment on attachment 9396923 [details]
Bug 1888634 - Disallow captureStream with canvas transferred to DOM worker.

Beta/Release Uplift Approval Request

• User impact if declined: Sites using captureStream with canvases transferred to a DOM worker will encounter bugs. Should be uncommon. Behaviour not defined by spec, although there is meeting notes indicating this will change in the future.
• Is this code covered by automated tests?: Yes
• Has the fix been verified in Nightly?: Yes
• Needs manual test from QE?: No
• If yes, steps to reproduce:
• List of other uplifts needed: None
• Risk to taking this patch: Low
• Why is the change risky/not risky? (and alternatives if risky): Just disables captureStream from working in conjunction with a DOM worker transferred canvas. This configuration should be rare as its behaviour is not specified by the spec yet.
• String changes made/needed:
• Is Android affected?: Yes

Comment 21

[Aron Cseh]

https://hg.mozilla.org/mozilla-central/rev/59e2426d20ae

Comment 22

[Donal Meehan [:dmeehan]]

Comment on attachment 9396923 [details]
Bug 1888634 - Disallow captureStream with canvas transferred to DOM worker.

Approved for 126.0b3