https://bugs.kde.org/show_bug.cgi?id=525182
Bug ID: 525182
Summary: KWin/Wayland: cascading amdgpu GPU reset crash loop on
RDNA2 (RX 6700 XT) — regression starting Plasma 6.7.0,
absent on 6.6.6
Classification: Plasma
Product: kwin
Version First 6.7.4
Reported In:
Platform: Manjaro
OS: Linux
Status: REPORTED
Severity: crash
Priority: NOR
Component: wayland-generic
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---
Created attachment 195740
--> https://bugs.kde.org/attachment.cgi?id=195740&action=edit
Daily monitoring script output with all 7 signatures
Since upgrading from KDE Plasma 6.6.6 to 6.7.0 (and confirmed still present
through 6.7.3 and 6.7.4), our system on AMD RDNA2 hardware (Radeon RX 6700 XT)
experiences repeated KWin/amdgpu GPU reset crash loops under Wayland. The
desktop becomes fully unresponsive; in most cases the compositor recovers
within under a minute (either automatically or after restarting the plasma
session from a TTY), but on at least one occasion the GPU entered an
unrecoverable MODE1 reset loop that required a hard power cycle.
Rolling back to Plasma 6.6.6 (same kernel, same Mesa version, same hardware)
eliminates the issue completely — it has not reproduced on 6.6.6 under any
circumstance across several weeks of use.
An independent report with identical GPU hardware (RX 6700 XT / RDNA2) on
CachyOS describes the same regression window (works on 6.6.5, breaks on
6.7.0/6.7.1/6.7.2) with the same elimination of Mesa and kernel as causes:
https://discuss.cachyos.org/t/plasma-6-7-x-wayland-hard-freeze-at-login-on-rx-6700-xt-rdna2-with-4-monitors-6-6-5-works-6-7-0-6-7-1-6-7-2-all-freeze/32650
That report remains unresolved as well.
There is no single 100% deterministic repro step; the bug is reliably
triggered — sooner or later — by any of the following on a Plasma 6.7.x +
RDNA2 Wayland session:
1. Log in to a Wayland session (the single most consistent trigger — 3 of 7
confirmed occurrences happened at login).
2. Let the screen dim via the power-saving timeout, then resume activity
(wake from DPMS/screen-off).
3. Enable HDR in System Settings → Display and Monitor, then use the desktop
normally for an extended session.
4. Launch a GPU-accelerated application after some uptime — observed with
Brave, Firefox, and Harmonoid (a Flutter-based media app), so this does
not appear to be specific to any one browser or rendering engine.
The kernel log fills with "ring gfx_0.X.0 timeout" and
"GCVM_L2_PROTECTION_FAULT_STATUS" faults, generally attributed to the
kwin_wayland and/or plasmashell processes. The driver attempts a ring reset,
sometimes succeeding within seconds (desktop briefly stutters/blacks out,
though the session UI usually remains unresponsive and needs
"systemctl restart plasmalogin" from a TTY to fully recover). In the worst
observed case (2026-08-31), the GPU entered a continuous MODE1 reset loop
("Ring reset failed", "VRAM is lost due to GPU reset!",
"Failed to initialize parser -125", no auto-stabilization) that did not
resolve until a hard power cycle.
Representative log excerpt (most recent event, 2026-09-02):
amdgpu 0000:0b:00.0: amdgpu: ring gfx_0.1.0 timeout, signaled seq=4242, emitted
seq=4244
amdgpu 0000:0b:00.0: amdgpu: Starting gfx_0.1.0 ring reset
amdgpu 0000:0b:00.0: amdgpu: ring gfx_0.0.0 timeout, signaled seq=2205, emitted
seq=2208
amdgpu 0000:0b:00.0: amdgpu: Starting gfx_0.0.0 ring reset
[...]
amdgpu 0000:0b:00.0: amdgpu: Process kwin_wayland pid 1543 thread
kwin_wayla:cs0 pid 1552
amdgpu 0000:0b:00.0: amdgpu: Ring gfx_0.1.0 reset succeeded
amdgpu 0000:0b:00.0: amdgpu: Process plasmashell pid 1738 thread plasmashel:cs0
pid 1832
amdgpu 0000:0b:00.0: amdgpu: Ring gfx_0.0.0 reset succeeded
Full logs for all seven confirmed occurrences are available and attached
(gpu_monitor.log).
The system remains stable under normal use, screen-dimming, HDR activation,
and GPU-accelerated application launches — matching the behavior observed on
Plasma 6.6.6 with identical hardware, kernel, and Mesa version.
```
Distro: Manjaro Linux (stable branch)
Kernel: 6.18.45-1-MANJARO (also reproduced on 6.18.44 and previously on
6.12.x — confirmed independent of kernel version)
KDE Plasma: 6.7.4 (also reproduced on 6.7.3; first appeared at 6.7.0;
NOT present on 6.6.6)
Mesa: 26.1.7 (also reproduced on 26.1.5 and 26.1.6 — confirmed independent
of Mesa version)
GPU: AMD Radeon RX 6700 XT (Navi 22, RDNA2), amdgpu/RADV
CPU: AMD Ryzen 7 3700X
RAM: 32 GiB DDR4
Display: LG UltraGear 27GS85Q, 2560x1440 @144Hz via DisplayPort
Session: KDE Plasma on Wayland (kwin_wayland)
```
Timeline of confirmed occurrences:
2026-08-16 (morning) — HDR active + Firefox. Signature: kwin_wayland.
Recovered via TTY + systemctl restart plasmalogin.
2026-08-16 (evening) — Second event same day. Signature: kwin_wayland.
Self-stabilized within under a minute.
2026-08-18 — Signature: kwin_wayland. Self-stabilized within under a minute.
2026-08-19 — Reset loop repeating roughly every 10s across several cycles
before stabilizing. Signature: kwin_wayland.
2026-08-27 — Both plasmashell AND kwin_wayland flagged in the same event,
suggesting multiple Plasma processes submit GPU work concurrently at
the point of failure. Self-stabilized within under a minute.
2026-08-31 — First time Firefox identified as a direct trigger (previously
only kwin_wayland/plasmashell were implicated). kwin_wayland crashed
mid-loop and restarted under a new PID, immediately re-entering the
loop. THE ONLY EVENT with "Ring reset failed" / "VRAM is lost due to
GPU reset!" / "Failed to initialize parser -125" — no auto-recovery,
required a hard power cycle.
2026-09-02 — First time Harmonoid (Flutter-based app) identified as
trigger. Signature: kwin_wayland + plasmashell. Ring resets reported
as "succeeded" at the driver level, but the desktop UI remained
unresponsive — required TTY diagnostic + reboot.
Mitigations tested (none eliminates the problem; some reduce frequency):
- amdgpu.gpu_recovery=1 amdgpu.noretry=0 (kernel param) — no observable
effect.
- amdgpu.dcdebugmask=0x10 (kernel param, disables Panel Self Refresh) —
reduces frequency but does not eliminate; best clean stretch was ~8 days,
but the bug has recurred twice since first applying it.
- KWIN_DRM_NO_DIRECT_SCANOUT=1 (env var) — same partial improvement,
combined with the above.
- KWIN_DRM_NO_AMS=1 (env var, disables atomic mode setting) — no
improvement observed. This matches what is reported in the CachyOS
thread referenced above, where the same env var also failed to help on
identical GPU hardware.
- Disabling HDR — reduces frequency but does not eliminate it (the
2026-09-02 event occurred with HDR off).
- Full rollback to Plasma 6.6.6 — ELIMINATES the problem completely, with
weeks of use and zero recurrence on identical hardware/kernel/Mesa.
Related reports:
- Independent report, identical GPU hardware (RX 6700 XT / RDNA2), same
regression window (6.6.5 healthy -> 6.7.0+ broken), unresolved:
https://discuss.cachyos.org/t/plasma-6-7-x-wayland-hard-freeze-at-login-on-rx-6700-xt-rdna2-with-4-monitors-6-6-5-works-6-7-0-6-7-1-6-7-2-all-freeze/32650
- Possibly related (similar "amdgpu GPU reset crash loop in KWin" pattern
from earlier Plasma cycles; not confirmed as an exact duplicate of this
regression): bugs.kde.org #453147, #513151
I wrote this report using the AI Claude. My English is very poor—please
excuse that. I hope this helps resolve the bug.
--
You are receiving this mail because:
You are watching all bug changes.
