RADV: compute ring timeout (comp_1.x) + page fault at 0x0 — Path of Exile 2, 7900 XTX, regression in 224226→224535

[cwtopel]

Summary

GPU hang / MODE1 reset while playing Path of Exile 2 (Vulkan, via gamescope) on a 7900 XTX. Compute ring times out with a null-address [gfxhub] page fault. Happened twice on consecutive days after ~100 hours of prior stability, immediately following a mesa-tkg-git update.

System Info

• GPU: AMD Radeon RX 7900 XTX (Navi 31, gfx1100)
• Kernel: 7.0.12-zen1-1-zen
• Distro: Arch Linux
• mesa-tkg-git (last known-good): 26.2.0_devel.224226.a07ded8e8cf-1
• mesa-tkg-git (suspect): 26.2.0_devel.224535.67a88ee9c8e-1
• lib32-mesa-tkg-git: same versions, in lockstep
• linux-firmware: 20260519-1 (current, SMU 78.131.0 — confirmed unrelated to this issue)
• Game: Path of Exile 2, run via gamescope

Timeline

• Stable for ~100 hours of PoE2 play prior to first crash, across several prior mesa-tkg-git builds.
• 2026-06-18 04:50 — mesa-tkg-git upgraded 224226.a07ded8e8cf → 224535.67a88ee9c8e.
• Same day, first crash occurred during a PoE2 session (GPU reset, recovered).
• Second crash the following day, same symptoms, different ring instance.

Crash 1 (dmesg)

amdgpu 0000:03:00.0: ring comp_1.1.0 timeout, signaled seq=343701, emitted seq=343702
amdgpu 0000:03:00.0:  Process gamescope-wl pid 8836 thread gamescope-xwm pid 8864
amdgpu 0000:03:00.0: Starting comp_1.1.0 ring reset
amdgpu 0000:03:00.0: reset compute queue (1:1:0)
amdgpu 0000:03:00.0: failed to wait on hqd deactivate
amdgpu 0000:03:00.0: failed to reset legacy queue
amdgpu 0000:03:00.0: fail(-110) to reset kcq and try pipe reset
amdgpu 0000:03:00.0: The CPFW hasn't support pipe reset yet.
amdgpu 0000:03:00.0: Ring comp_1.1.0 reset failed
amdgpu 0000:03:00.0: GPU reset begin!. Source:  1
...
amdgpu 0000:03:00.0: MODE1 reset
amdgpu 0000:03:00.0: GPU smu mode1 reset
amdgpu 0000:03:00.0: GPU reset succeeded, trying to resume
amdgpu 0000:03:00.0: VRAM is lost due to GPU reset!
...
amdgpu 0000:03:00.0: GPU reset(1) succeeded!
amdgpu 0000:03:00.0: [drm] *ERROR* Failed to initialize parser -125!
amdgpu 0000:03:00.0: [drm] device wedged, but recovered through reset
amdgpu 0000:03:00.0: [drm] *ERROR* [CRTC:363:crtc-0] flip_done timed out
amdgpu 0000:03:00.0: [drm] *ERROR* [CRTC:363:crtc-0] hw_done or flip_done timed out

GPU reset succeeded; system stayed up. Display flips briefly timed out after recovery before stabilizing.

Crash 2 (devcoredump excerpt — umr/strings decoded)

Ring timed out details
IP Type: 1 Ring Name: comp_1.0.1

[gfxhub] Page fault observed
Faulty page starting at address: 0x0000000000000000
Protection fault status register: 0x0

Note: different ring instance than crash 1 (comp_1.0.1 vs comp_1.1.0), both in the compute family, both faulting on address 0x0 — looks like a null/unbound resource being dereferenced from a compute dispatch, not a hardware fault (same fault signature recurring on two different ring instances across two sessions points to software, not silicon).

Process name in the coredump header is generic (Main PID: ...); the dmesg from crash 1 explicitly names gamescope-wl/gamescope-xwm as the offending process, consistent with the Vulkan compute work being submitted through gamescope's compositor path while PoE2 is running underneath it.

What I've ruled out

• SMU/firmware mismatch (smu driver if version not matched, 0x3d vs 0x40) — present in dmesg on every boot regardless of crash, confirmed to be a long-standing, unrelated cosmetic warning. Same firmware version present both before and after the mesa-tkg-git update; not a contributing factor.
• Kernel regression — same kernel (7.0.12-zen1-1-zen) was running stable for the prior ~100 hours before the mesa-tkg-git update landed.

Suspected cause

A regression somewhere between 224226.a07ded8e8cf and 224535.67a88ee9c8e affecting RADV's compute pipeline/descriptor handling, causing a compute dispatch to reference an unbound (null) GPU virtual address under PoE2 + gamescope's workload pattern.