-
Notifications
You must be signed in to change notification settings - Fork 0
Crashes GPU and Memory
A game closing, a frozen frame, a black screen, and a complete session restart may look similar but leave different evidence. Use the incident time to correlate the game/Steam process, Gamescope, kernel GPU, memory, and core-dump layers.
sudo ./deckdoc.shRead these together:
module_coredump.logmodule_steam.logmodule_gamescope.logmodule_gpu.logmodule_dxvk.logmodule_memory.logmodule_thermal.log
The core-dump module separates retained history, the current boot, and a rolling last-24-hour window.
It breaks that recent window into steamwebhelper, Gamescope, and Wine/Proton families. A helper
SIGTRAP needs UI/incident correlation; a Gamescope SIGABRT or SIGSEGV is a different, higher-value
session failure signal.
A current-boot SIGSEGV/SIGABRT for the game, Wine/Proton, Steam helper, MangoApp, or Gamescope
identifies the process that terminated. It does not alone explain why. Check surrounding logs and
whether only one title reproduces.
A current Gamescope dump, Vulkan descriptor failure, or repeated session restart can return the user
to Library/login or blank the whole session. A gamescope-wl dump recorded only at normal session exit
may be cosmetic; time and restart count matter.
The MangoApp FDINFO_PERMISSION_ABORT is a separate helper signature. Do not call it a GPU or panel
failure without additional evidence.
amdgpu_job_timedout, ring timeout, VRAM is lost, and reset messages can explain a frozen game or
session. Classify the outcome:
- reset succeeded: the driver recovered, but clients may still exit;
- reset failed/skipped: higher risk of a hard lock and orderly reboot may be required;
- no reset: inspect process crashes, VM faults, memory, and display-specific branches.
DeckDoc does not trigger the kernel's debugfs GPU-reset control because it discards in-flight work and can destroy the failure state.
The DXVK/VKD3D module recognizes selected GCVM_L2_PROTECTION, UTCL2, CB/DB/CPF/CPD, mapping, and
walker strings. These labels classify where the kernel reported a fault; they do not uniquely prove a
specific translation layer or hardware defect. Correlate process attribution, game/API, Mesa/Proton
version, and whether other titles reproduce.
Current-boot OOM-killer/page-allocation messages are strong evidence. Below 1 GB MemAvailable is a
warning and below 512 MB is critical. Swap over 50% and cumulative swapped pages are historical context;
live vmstat swap I/O is stronger evidence of current pressure. None reconstructs a past incident by
itself.
The Deck's memory is shared with graphics, so pressure can cascade into stutter, allocation failure, or GPU/client instability. Close the workload and preserve the process/memory context; do not clear arbitrary caches as a first response.
| Scope | First hypotheses |
|---|---|
| One title/save/scene | game bug, mod, launch option, Proton/API path, corrupt title data |
| All D3D12 but not Vulkan/D3D11 | VKD3D/driver/API path |
| All titles after an update | SteamOS/Mesa/Gamescope/Proton regression |
| Desktop and games | compositor, kernel/GPU, memory, thermal, hardware |
| Only after wake | device/resume/firmware state |
Change one variable at a time: remove per-game modifications, select a known-good Proton version, verify game files, or test a clean boot. Keep the exact result.
- GPU reset fails or the Deck hard-locks repeatedly;
- multiple unrelated titles reproduce on a clean, updated system;
- page faults/timeouts recur at ordinary temperatures and memory use;
- filesystem/SMART errors accompany crashes;
- a core dump and logs identify a reproducible upstream component regression.
DeckDoc records evidence before action. Preserve the incident, change one variable at a time, verify the result, and use Steam Support for safety, warranty, and hardware concerns.
- Home
- Getting started
- Triage flow
- Reading reports
- Collect and share evidence
- Continuous incident probe
- Privileged authorization