Interface freeze with kernel 6.11.3 on Fedora Linux 40 (Radeon RX 6600)

Since I updated to kernel 6.11.3, I have been experiencing interface freezes, which require a reboot to resolve. These freezes happen consistently, after every boot, a few minutes after logging into the desktop. I resolved the issue by reverting to kernel 6.10.12 Here are some logs, I hope they will be helpful.

freeze_kernel6-11-3

Sounds very similar to what others are experiencing with the amdgpu. Everything lower than 6.10.13 (so 6.10.12 or earlier) seems to be safe. Anything above 6.10.12 seems to be affected. As I've got issues with a system with a 7900XTX and a 6750XT where the 6750XT shows similar logs and the 7900XTX with most crashes just completely freezes.

Yeah, I tried 6.10.14 & 6.11.2 with RX 7600. Both crash as soon as the amdgpu driver gets initialized. The worst part is that I am unable to get any logs using journalctl. Even Alt+SysRq+S fails.

Can you bisect?

ott 15 21:07:56 fedora kernel: UBSAN: array-index-out-of-bounds in drivers/gpu/drm/amd/amdgpu/../display/amdgpu_dm/amdgpu_dm.c:8929:26 ott 15 21:07:56 fedora kernel: index 3 is out of range for type 'dc_surface_update [3]'

is #3594 (closed)

for me that was enough to crash the system, but I guess randomly it could just destabilize your system and let it live longer just with corrupted memory

but it's unusual that you experience that with any desktop compositor other than Weston, what are you running?

I recall KWin merged a change recently that uses Overlay layer as a fallback Cursor layer under some conditions, maybe that could be relevant

and #3594 (closed) already has patches that were reviewed but not, I think, merged https://lore.kernel.org/amd-gfx/20240925154324.348774-1-mwen@igalia.com/

Hello, thank you for responding. I'm using Cosmic, their compositor is based on Smithay. If you need more information, let me know. Thanks.

I add further information: I have a notebook with AMD hardware, specifically:AMD Ryzen 5 5500U. On this one, I haven't had any freezes so far. The software configuration is the same: Cosmic + Kernel: Linux fedora-note 6.11.4-300.fc41.x86_64 but on Fedora 41. Perhaps the logs from this notebook, where I haven't experienced any freezes, might be useful. grep_amdgpu

grep_drm

Update: I only encounter the issue when starting a session with Cosmic; the graphical session crashes after a few seconds. I can't reproduce it when starting the session with GNOME Shell.

The main cosmic-comp developer has mentioned that COSMIC is one of the only compositors widely utilizing overlay planes. It's possible you're not seeing the crash because GNOME Shell isn't using that.

Bisect according to https://github.com/pop-os/cosmic-comp/issues/887#issuecomment-2408743225: https://github.com/torvalds/linux/commit/1b04dcca4fb10dd3834893a60de74edd99f2bfaf

added 6000 dGPU Series DC bisected labels

Ok, that matches what @melissawen discovered working on her patch for the original issue #3594 (closed).

I am sure that if you pull the patch the issue goes away, but I haven’t checked if there are merge conflicts now.

Right, I've sent a bug fix pointing out that the cursor overlay mode uncovered this array-index-out-of-bounds issue here. AFAIU, a major rework is coming, but I've asked for the bug fix first (the right number of surfaces is 4) without success... I've just pinged again in the same ML thread.

mentioned in issue #3729 (closed)

mentioned in issue #3735

Hi all,

Zaeem from AMD is working on the proper solution which requires a major rework, but it seems this patch can fix the issue in the meanwhile:

• https://lore.kernel.org/amd-gfx/20241114143741.627128-1-zaeem.mohamed@amd.com/

I don't have the hardware and environment to reproduce, so can you verify if this patch helps with the issue?

Please, feel free to provide a direct feedback in the patch e-mail.

Just chiming in to say that I also experience this issue (Framework 13, AMD 7840U).

I applied Zaeem's patch on top of 6.12-rc7, but it didn't help. The only relevant logs captured by journald were:

Nov 14 16:28:02 frmwrk kernel: BUG: unable to handle page fault for address: 000000000174e354
Nov 14 16:28:02 frmwrk kernel: #PF: supervisor read access in kernel mode

which seems to indicate that it's not hitting the new logic.

If it helps any, I've noticed the hang happens when cosmic-panel touches / overlaps with Firefox's window.

I have been running your original suggestion to set MAX_SURFACES and MAX_SURFACE_NUM = 4 which prevents the Firefox issue with the tradeoff that some applets crash when the applet menu is opened.

When this happens dmesg doesn't spit anything out, but cosmic-panel complains with the following when the applet crashes:

Nov 14 16:37:43 frmwrk .cosmic-panel-wrapped[1762]: com.system76.CosmicAppletBattery: Error sending request: Resource temporarily unavailable
Nov 14 16:37:43 frmwrk .cosmic-panel-wrapped[1762]: com.system76.CosmicAppletBattery: exited with code 0

oh, interesting. thanks for the inputs.

1. Does cosmic-panel still complain if MAX_SURFACES = 6 as in this series?

2. Can you share the outputs of drm_info and # cat /sys/kernel/debug/dri/0/amdgpu_dm_dtn_log?

1. Does cosmic-panel still complain if MAX_SURFACES = 6 as in this series?

So after applying that patch series the applets behave (no crashing, no complaining). Firefox doesn't, however and I got some logs when it crashed this time, here's the relevant bit: journald

2. Can you share the outputs of drm_info and # cat /sys/kernel/debug/dri/0/amdgpu_dm_dtn_log?

drm_info dtn_log

It's worth noting here that I'm also applying a patch reverting 338567d1762 for another issue with DSC / bandwidth calc (#3735) for my own QOL. I can revert that to isolate if you think it might be interfering.

Hi @cfrenette,

Thanks for sharing your logs.

I didn't manage to reproduce the index-out-of-bounds on Cosmic yet, but I got this page fault with amd-staging-drm-next kernel with this commit reverted: .

1. I wonder if the index-out-of-bounds is still present on 6.12-rc7 (without Zaeem's patch) or only the page-fault error.
2. If so, maybe dmesg shows both many index-out-of-bounds errors and the page-fault as the last one? Can you double check your dmesg?

I've since moved to using 6.12.0 and I can't seem to reproduce the freeze with cosmic-comp commit 7e8cb91 or at least not in the same way. I'll keep using 6.12.0 for now and I'll report back with logs if I see the issue.

I can also switch back to 6.12-rc7 and compare / bisect to determine what might have fixed it if you'd like that info.

It seems I spoke too soon. I got a freeze, here are the logs:

freeze_6-12_journald.log

Steps to reproduce:

1. Set cosmic-panel to automatically hide, turn off gap and extend to edges and set opacity < 1
2. Tile Firefox fullscreen
3. Unhide (mouse over) the panel to cause it to overlap with Firefox

I looked back in dmesg further and I do not see the index oob in this particular case (though I was seeing it on 6.11).

Did anyone here check if the issue is present in the amd-staging-drm-next kernel?

Compiling from the "amd-staging-drm-next" branch [1], I don't have any issues. However, I don't know if it's expected, but from that branch, I have version 6.10.* of the kernel. The problem occurs with kernel 6.11. I'm not sure if I'm doing something wrong or if this is expected. If you need, let me know what I can do to help with testing. Thanks.

The Fedora kernel: 6.10.10-200.fc40

Compiled kernel: Linux fedora 6.10.0+ #2 SMP PREEMPT_DYNAMIC Tue Nov 26 23:58:43 CET 2024 x86_64 GNU/Linux

[1] git clone --depth=1 -b amd-staging-drm-next https://gitlab.freedesktop.org/agd5f/linux.git

Hi @emanuc, thanks for checking it out and reporting!

amd-staging-drm-next is the upstream branch for AMD driver and it's currently kernel 6.10 + all AMD driver changes from that kernel version until now. It doesn't include core kernel branches, but you get almost all AMD driver changes.

---

I just managed to consistently reproduce the page fault and I think it's causing the interface freeze since the beginning (not the array-index-out-of-bounds).

On kernel Fedora 41, 6.11.4 kernel, I see both array-index-out-of-bounds and the last error is in fact the page fault:

• dmesg-index-out-of-bounds-plus-page-fault-fedora41-cosmic-6-11-kernel.txt

On amd-staging-drm-next kernel I only see the page fault:

• dmesg-page-fault-fedora41-cosmic-asdn-kernel.txt

Steps to reproduce:

• Open a firefox window and drag it to the right side (multitasking active screen edges)
• Open file manager and drag it to the left-top side. Here the system freezes due to page fault.

I'll check mainline kernel to see the status tomorrow.

Hi,

I just sent another proposal to fix this issue, following a different approach since I noticed the page fault is the current cause of the system freeze.

• https://lore.kernel.org/dri-devel/20241204210929.1994522-1-mwen@igalia.com/

I'd really appreciate it if you could try it out and give any feedback.

Here are the patches attachments, it's based on upstream amd-staging-drm-next branch.

• 0001-drm-amd-display-fix-page-fault-due-to-max-surface-de.patch
• 0002-drm-amd-display-increase-MAX_SURFACES-to-the-value-s.patch

I can backport to 6.11 if you need for validation, just let me know.

just tested:

• 6.11: broken
• 6.12: broken
• amd-staging-drm-next tip (7a41d27dc5dbb66ff5caed0203b8ef89ec3e61b8): broken
• these two patches on tip: fixed

thanks :)

fwiw i use niri, so the actual trigger for the issue is probably smithay specific and not something in cosmic itself, but that's a tiny detail

Just tested this on top of stable (6.12.1) and I can't seem to reproduce the freeze now.

I'll continue using this patch and report back if I encounter any issues.

Thanks for the fix!

added Rembrandt Vangogh labels

added 7000 dGPU series label

marked this issue as related to #3520 (closed)

marked this issue as related to #3594 (closed)