1732677 - [M1] OS crashes (kernel panic) upon toggling full screen for software-decoded video

Status: VERIFIED FIXED

(Core :: Graphics, defect, P1)

Description

[Timea Cernea [:tbabos][inactive]]

Affected Version:

• latest Nightly 94.0a1 (2021-09-26) (64-bit)

Reproducible on:

• MacBook Air M1 - MacOS Big Sur

Steps to reproduce:

1. Go to Youtube and select any video (or any other site with videos)
2. Click on the FullScreen button

Expected Result:
The video is played back without any issues

Actual Result:
The whole OS crashes and the device is restarted

Regression-range:
Not reproducible if the following pref is disabled: gfx.core-animation.specialize-video
Regression of: Bug 1653417

Comment 1

[Markus Stange [:mstange]]

Which exact macOS version are you using?

This does not happen for me on macOS 11.6 (20G165) on an M1 MacBook Pro 13". But it was happening on this machine with an earlier version of macOS when I was running a test build with these patches, two or three weeks ago.

Comment 2

[Timea Cernea [:tbabos][inactive]]

Hey Markus, I am on MacOS Big Sur (Version 11.2.2) - MacBook Air (M1, 2020). Please be aware that this is only reproducible on the M1 device (Apple chip).

Comment 3

[Gian-Carlo Pascutto [:gcp]]

I'm on macOS Big Sur 11.6, also a Macbook Air M1.

Comment 5

[Markus Stange [:mstange]]

I have filed FB9662075 about this.

Comment 6

[Romain Testard [:RT]]

We have about 350k DAU on aarch64 on macOS (https://sql.telemetry.mozilla.org/queries/76596/source#191266) so an acceptable option for 94 could be to pref off bug 1653417 for these clients if no fix can be made in time for beta merge?

Comment 7

[Brad Werth [:bradwerth]]

Attachment: WIP: Bug 1732677: Split CATransactions to separate layer properties and layer relationships.

Since the M1 Macs panic only after the video layer has been isolated, that
means that the AVSampleBufferDisplayLayer has been successfully enqueueing
buffers up to that point. It's possible that our manipulation of the layer
tree is the source of the panic.

Comment 8

[Brad Werth [:bradwerth]]

(In reply to Romain Testard [:RT] from comment #6)

We have about 350k DAU on aarch64 on macOS (https://sql.telemetry.mozilla.org/queries/76596/source#191266) so an acceptable option for 94 could be to pref off bug 1653417 for these clients if no fix can be made in time for beta merge?

Yes, I have opened Bug 1733148 to do that.

Comment 9

[Markus Stange [:mstange]]

I can reproduce this 100% of the time on the following software-decoded video, if I have no external screen and no power cable connected: https://upload.wikimedia.org/wikipedia/commons/2/22/Volcano_Lava_Sample.webm

I cannot reproduce this at all on the following hardware-decoded video: https://chromanomicon.com/TestPatterns/chip-chart-1080-untagged.mp4

Comment 10

[Anna Sobiepanek]

I am seeing the same thing when loading https://firebase.google.com/docs/firestore/data-model and trying to watch the video full screen.

macOS Big Sur Version 11.6
Hardware Overview:

  Model Name:	MacBook Pro
  Model Identifier:	MacBookPro17,1
  Chip:	Apple M1
  Total Number of Cores:	8 (4 performance and 4 efficiency)
  Memory:	16 GB
  System Firmware Version:	6723.140.2
  OS Loader Version:	6723.140.2
  Activation Lock Status:	Disabled

Please also note that turning off gfx.core-animation.specialize-video fixes the issue

Comment 11

[Markus Stange [:mstange]]

We've landed the patch to turn off the pref on M1 machines in bug 1733148, so this should be fixed in the next Nightly.

We'll use this bug to find a way to get low-power software-decoded video on M1 machines without kernel panicking.

Comment 13

[Markus Stange [:mstange]]

Using NV12 instead of YUV422 (by partially reverting bug 1657107) gets rid of the kernel panic, but doesn't hit detached mode. So it probably doesn't kernel panic because it doesn't hit detached mode.

We don't yet have a good idea of the set of surface formats that let us hit detached mode with software-decoded video. Adding a software-decoded video path to the test app would help us gain a better understanding.

Comment 14

[Markus Stange [:mstange]]

Actually let's close this bug, so that the tracking flags etc make sense. I've filed two bugs to re-enable the power-efficient path on M1 machines:

• bug 1733500 for hardware-decoded video
• bug 1733499 for software-decoded video

Comment 15

[Markus Stange [:mstange]]

Fixed by bug 1733148.

Comment 16

[Timea Cernea [:tbabos][inactive]]

Verified-fixed on the latest Firefox Nightly 94.0a1 (2021-09-30) ( 20210930215026) on MacOS 11.2.2 - M1 Apple. There were no kernel panics encountered since the preference that caused it is now disabled by default.