Closed Bug 1716254 Opened 3 years ago Closed 3 years ago

AMD Radeon Vega 8: WebRender causes graphical lockups etc.

Categories

(Core :: Graphics: WebRender, defect)

Firefox 89
defect

Tracking

()

RESOLVED WORKSFORME

People

(Reporter: hikaph+mozilla, Unassigned)

References

(Blocks 1 open bug)

Details

Attachments

(1 file)

User Agent: Mozilla/5.0 (X11; Fedora; Linux x86_64; rv:89.0) Gecko/20100101 Firefox/89.0

Steps to reproduce:

  1. I upgraded to Firefox 89. Being greedy and wanting to see if WebRender would work, I set MOZ_WEBRENDER=1 (among some other environment variables) when running Firefox.
  2. I verified that about:support shows Compositing: WebRender.
  3. I open YouTube and watch some videos, using my browser normally.

Platform Information

I'm running Firefox 89 packaged for Fedora 34 with mesa 21.1.1.

Copied from about:support:

GPU #1
Active	Yes
Description	AMD Radeon(TM) Vega 8 Graphics (RAVEN, DRM 3.40.0, 5.12.10-300.fc34.x86_64, LLVM 12.0.0)
Vendor ID	0x1002
Device ID	0x15d8
Driver Vendor	mesa/radeonsi
Driver Version	21.1.1.0
RAM	0

Actual results:

After a random amount of time (usually less than 2 hours), my graphics stack freezes Assorted visual artifacts (glitched boxes, garbage textures, etc.) may be drawn over the entire display. Kernel logs show lots of amdgpu and/or drm messages, and usually things recover within 30 seconds. Sometimes (not deterministically), the lockups may blank the screen and leave my device unresponsive, requiring a forcible reboot to recover.

Expected results:

The graphical stack should not have done any of the preceding, allowing for normal usage.

  • I downgraded WebRender to use the software backend and this appears to fix (or at least greatly lessen the prevalence of) the issue.
  • I fully expect this to be a Mesa issue or similar, so I reckon there won't be a fix from the Firefox side. I'm only filing in case the feedback is useful for planning WebRender rollout, because even after unsetting MOZ_WEBRENDER in my environment, I still had it enabled by default (so I'm guessing that it was switched on for me at 89.0).
Blocks: wr-linux

:stransky, :rmader, can you comment to the bug?

Flags: needinfo?(stransky)
Flags: needinfo?(robert.mader)

I agree with Kalvin that this is most likely a mesa/kernel issue.
Kalvin, mind also sharing which desktop environment and protocol(Wayland/X11) you use? Or preferably you whole about:support? :)

Flags: needinfo?(robert.mader) → needinfo?(hikaph+mozilla)
Flags: needinfo?(stransky)
Flags: needinfo?(hikaph+mozilla)

Robert — I'm using sway 1.6 and I usually unset DISPLAY when running Firefox in an attempt to nudge it to use pure Wayland.

I tried enabling vanilla WebRender again with the combination of

  • Kernel 5.12.14 and
  • Mesa 21.1.4 (contains two-ish AMD fixes)

and the worst of the graphical lockups has disappeared. In the worst case, this happens, which freezes the compositor for a few seconds:

Jul 04 01:52:36 flaglock6 kernel: amdgpu 0000:06:00.0: amdgpu: [gfxhub0] retry page fault (src_id:0 ring:0 vmid:6 pasid:32783, for process firefox-bin pid 12977 thread firefox-bi:cs0 pid 13071)
[snip]
Jul 04 01:52:43 flaglock6 kernel: [drm:amdgpu_dm_atomic_commit_tail [amdgpu]] *ERROR* Waiting for fences timed out!
Jul 04 01:52:47 flaglock6 kernel: [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring gfx timeout, signaled seq=2600195, emitted seq=2600198
Jul 04 01:52:47 flaglock6 kernel: [drm:amdgpu_job_timedout [amdgpu]] *ERROR* Process information: process firefox-bin pid 12977 thread firefox-bi:cs0 pid 13071

which could be any of several outstanding bugs on the drm/amd or Mesa bug trackers. In the snippet above, I simply waited a few seconds and resumed normal usage of Firefox — though I did notice that the compositing downgraded to WebRender (Software). This comes after a few hours of usage that I would have expected to reproduce the issue, so this bug should change to "WorksForMe."

Status: UNCONFIRMED → RESOLVED
Closed: 3 years ago
Resolution: --- → WORKSFORME
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Creator:
Created:
Updated:
Size: