Graphical glitches in Firefox with kernel 6.2-rc1 [RX580]

Brief summary of the problem:

After updating to kernel 6.2-rc1 in Fedora Rawhide, there are occasional graphics errors in the upper half of the display when browsing with Firefox. These relatively light glitches often resemble a checkerboard pattern and do not always occur. The system does not freeze or report any errors, it stays perfectly usable, even games work fine. When downgrading to kernel 6.1.0 the glitches disappear completely, and they can easily be reproduced by just booting the rc1 kernel again.

Hardware description:

• CPU: Ryzen 7700X
• GPU: 01:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Ellesmere [Radeon RX 470/480/570/570X/580/580X/590] [1002:67df] (rev e7)
• System Memory: 32 GiB
• Display(s): 2x 4K@60Hz
• Type of Display Connection: DP

System information:

• Distro name and Version: Fedora Rawhide/38
• Kernel version: 6.2.0-0.rc1.14.fc38
• Custom kernel: N/A
• AMD official driver version: N/A

How to reproduce the issue:

• Boot Fedora Rawhide with 6.2-rc1 kernel
• Start Firefox and browse to e.g., https://www.rsync.net/products/borg.html, glitches usually start to appear after a few seconds there

changed the description

Can you bisect?

Have just bisected, on openSUSE Tumbleweed, X, Firefox. Radeon RX580.

the log:

git bisect start
# status: waiting for both good and bad commits
# bad: [1b929c02afd37871d5afb9d498426f83432e71c2] Linux 6.2-rc1
git bisect bad 1b929c02afd37871d5afb9d498426f83432e71c2
# status: waiting for good commit(s), bad commit known
# good: [830b3c68c1fb1e9176028d02ef86f3cf76aa2476] Linux 6.1
git bisect good 830b3c68c1fb1e9176028d02ef86f3cf76aa2476
# bad: [fc4c9f450493daef1c996c9d4b3c647ec3121509] Merge tag 'efi-next-for-v6.2' of git://git.kernel.org/pub/scm/linux/kernel/git/efi/efi
git bisect bad fc4c9f450493daef1c996c9d4b3c647ec3121509
# good: [8ecd28b7a3a4c43a875a8840851f72468a2ca1d7] Merge tag 'ata-6.2-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/dlemoal/libata
git bisect good 8ecd28b7a3a4c43a875a8840851f72468a2ca1d7
# bad: [66efff515a6500d4b4976fbab3bee8b92a1137fb] Merge tag 'amd-drm-next-6.2-2022-12-07' of https://gitlab.freedesktop.org/agd5f/linux into drm-next
git bisect bad 66efff515a6500d4b4976fbab3bee8b92a1137fb
# good: [49e8e6343df688d68b12c2af50791ca37520f0b7] Merge tag 'amd-drm-next-6.2-2022-11-04' of https://gitlab.freedesktop.org/agd5f/linux into drm-next
git bisect good 49e8e6343df688d68b12c2af50791ca37520f0b7
# bad: [fc58764bbf602b65a6f63c53e5fd6feae76c510c] Merge tag 'amd-drm-next-6.2-2022-11-18' of https://gitlab.freedesktop.org/agd5f/linux into drm-next
git bisect bad fc58764bbf602b65a6f63c53e5fd6feae76c510c
# bad: [4e291f2f585313efa5200cce655e17c94906e50a] Merge tag 'drm-misc-next-2022-11-10-1' of git://anongit.freedesktop.org/drm/drm-misc into drm-next
git bisect bad 4e291f2f585313efa5200cce655e17c94906e50a
# good: [78a43c7e3b2ff5aed1809f93b4f87a418355789e] drm/nouveau/gr/gf100-: make global attrib_cb actually global
git bisect good 78a43c7e3b2ff5aed1809f93b4f87a418355789e
# bad: [611fc22c9e5e13276c819a7f7a7d19b794bbed1a] drm/arm/hdlcd: remove calls to drm_mode_config_cleanup()
git bisect bad 611fc22c9e5e13276c819a7f7a7d19b794bbed1a
# bad: [a8d9621b9fc67957b3de334cc1b5f47570fb90a0] drm/ingenic: Don't set struct drm_driver.output_poll_changed
git bisect bad a8d9621b9fc67957b3de334cc1b5f47570fb90a0
# bad: [2cf9886e281678ae9ee57e24a656749071d543bb] drm/scheduler: remove drm_sched_dependency_optimized
git bisect bad 2cf9886e281678ae9ee57e24a656749071d543bb
# skip: [940ca22b7ea9db6857ba7c6adb961b84d8cc28ff] drm/amdgpu: drop amdgpu_sync from amdgpu_vmid_grab v2 - some different kind of corruption and even Firefox/X getting "stuck"
git bisect skip 940ca22b7ea9db6857ba7c6adb961b84d8cc28ff
# bad: [aab9cf7b6954136f4339136a1a7fc0602a2c4d8b] drm/amdgpu: use scheduler dependencies for VM updates
git bisect bad aab9cf7b6954136f4339136a1a7fc0602a2c4d8b
# good: [4f91790b42ffba72d80434d901548979ab41dc7c] drm/amdgpu: use drm_sched_job_add_resv_dependencies for moves - similar corruption like the skipped commit, but no lockup, and not this specific corruption
git bisect good 4f91790b42ffba72d80434d901548979ab41dc7c
# good: [f7d66fb2ea43a3016e78a700a2ca6c77a74579f9] drm/amdgpu: cleanup scheduler job initialization v2
git bisect good f7d66fb2ea43a3016e78a700a2ca6c77a74579f9

last step produced the culprit: 1b2d5eda5ad7 ("drm/amdgpu: move explicit sync check into the CS")

Built that as well and confirmed it's bad. So despite some issues towards the end of the bisect, the last good seemed fine, and the first bad commit directly follows that in git log. Will also try revert on top rc2 if possible.

I tried a revert of commit 1b2d5eda5ad7 on top of the openSUSE 6.2-rc2 kernel (I can see the glitches with Radeon WX2100 since 6.2 merge window snapshots). It seems to behave better, e.g. the issue with "konsole" windows blinking between previous and current content seems to be gone and the glitches seem to only appear in specific firefox windows. But I can still see the zigzag patterns in some firefox windows; e.g. https://www.astro.cz/ seems to be quite reliable reproducer, especially when the cursor moves on top of that window.

Edit: after some time, the "konsole" windows blinking is back as well.

I investigated the issue some more. First I tried a bisect and ended up with the same commit as Vlastimil. Then I retried revert of commit 1b2d5eda5ad7 on top of clean mainline v6.2-rc1 and it still did show the glitches. So I tried another bisect, this time between 1b2d5eda5ad7 and v6.2-rc1 with a revert of 1b2d5eda5ad7 in each step. This bisect resulted in commit 1728baa7e4e6 ("drm/amdgpu: use scheduler dependencies for CS") as first bad.

Finally, I tried v6.2-rc2 with revert of

• 2cf9886e2816 ("drm/scheduler: remove drm_sched_dependency_optimized")
• 1728baa7e4e6 ("drm/amdgpu: use scheduler dependencies for CS")
• 1b2d5eda5ad7 ("drm/amdgpu: move explicit sync check into the CS")

and the glitches are gone. So apparently either of 1b2d5eda5ad7 and 1728baa7e4e6 causes the problem. (Commits aab9cf7b6954 and 46e0270c71d3 might deserve a check too.)

Just to confirm: You need to revert 1b2d5eda5ad7 ("drm/amdgpu: move explicit sync check into the CS") as well to make this issue go away?

Yes, I need to revert both of them. I was able to reproduce the issue with

• commit 1b2d5eda5ad7
• commit 1728baa7e4e6 plus a revert of 1b2d5eda5ad7

but not with

• f7d66fb2ea43 (= 1b2d5eda5ad7^)
• 46e0270c71d3 (= 1728baa7e4e6^) plus a revert of 1b2d5eda5ad7
• 6.2-rc2 plus revert of both 1b2d5eda5ad7 and 1728baa7e4e6

Can you try the following patch and see if it fixes the problem:

0001-drm-amdgpu-fix-pipeline-sync.patch

I can still see the glitches in some firefox windows with this patch on top of current mainline (commit 1fe4fd6f5cad).

Same here, still happens.

Quite persistent problem. Can you check this one here as well:

0001-drm-amdgpu-fix-pipeline-sync-v2.patch

Thanks in advance for the help.

Yes, this one looks good so far (on top of current mainline). I've been running with it only for a few minutes but I already tried tests that triggered the issue in any "bad" kernel before.

Looks good here as well, I seem unable to reproduce the issue with this patch applied.

Nice, thanks a lot guys. Just one more test. I want to see if I it's sufficient if I reduce this to the explicit sync or if we somehow have a dependency for the implicit syncs as well.

0001-drm-amdgpu-fix-pipeline-sync-v3.patch

Nice, thanks a lot guys. Just one more test. I want to see if I it's sufficient if I reduce this to the explicit sync or if we somehow have a dependency for the implicit syncs as well.

v3 still produces artifacts on my end.

Ok, thanks. Then I will use for v2 as bug fix for this.

Should I add any Tested-by tags?

I'll pass unless it helps you somehow with submitting the patch. Thanks for the fix!

I can see glitches with v3.

For v2, feel free to use Tested-by: Michal Kubecek <mkubecek@suse.cz>

Haven't tried v3, seems unnecessary at this point. For v2, Tested-by: Vlastimil Babka <vbabka@suse.cz> Thanks!

@ckoenig Any ideas? Possibly related to #2309 (closed).

No idea of hand, but going to look into this with priority. Thanks for the notice.

mentioned in commit FireBurn/linux@3bd68b32

mentioned in commit agd5f/linux@31974c5f

closed