Model M releases its resources after a drain timeout, though game lists may use them

Aperta
#62 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
22/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
cpp
Ambito
devtools

Direzione di ricerca

Start in clean/runtime/interop/executor_d3d12.cpp: drain() around line 329 and the destructor around line 52, which releases proxy16_, proxy32_, mpoint_ and retired_. Then check device_d3d12.cpp at lines 146-185 for the retain-on-failure path. The test seam is in dx12/tests/native_direct_test.cpp using the WARP Gpu helper near line 46. Done means drain() returns bool, the destructor retains the runtime on timeout when the device is not removed, and the WARP test confirms a probe object is not released.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

area: runtime bug needs-hardware-test

What happens

D3d12Executor::drain() (clean/runtime/interop/executor_d3d12.cpp:329) polls external_done() < external_recorded() for up to 5000 × Sleep(1). It returns void. On timeout it only logs.

The destructor (:52) then calls begin_teardown() and runtime_destroy(rt_) unconditionally, and releases proxy16_, proxy32_, mpoint_ and retired_ (:59-:61). The game's recorded list references these objects (df.output, df.motion, the copy16 source).

runtime_destroy ends in the device destructor (clean/runtime/device_d3d12.cpp:146). Its "retain on failure" path (:148-:163) depends only on finish_queue() (:869), which submits to and waits on the device's own create-time queue. It does not look at external_done(). When that succeeds, :166-:185 releases the arena, PSOs, descriptor heap, root signatures, direct textures and the marker.

shutdown (clean/runtime/interop/backend.cpp:1933) calls exec_->drain() and then exec_.reset(). Nothing else waits.

When it can happen

Only if the game's list runs more than 5 s after the destroy call. Two cases:

  • The game records a dispatch and destroys the context before it submits that list, on the same thread (for example a settings or resize handler in the middle of a frame). The wait can never succeed, so the game thread freezes for 5 s, and the game then submits a list whose resources have been freed.
  • The queue is stalled for more than 5 s without device removal. This is rare, because a driver timeout normally fires first.

In the common case the front end parks destroyed contexts (clean/fsr/src/frontend.h:156), so the executor is not torn down when the game calls destroy. Most engines also flush the GPU before releasing an upscaler. The backend's own contract says destroy "waits for in-flight work" (clean/fsr/include/wronggpu_upscaler.h:210); it does not say unsubmitted work is safe.

The impact is device removal. The probability is low.

Found by reading the code. It has not been reproduced.

To do

This is the minimal fix, matching the device's existing policy at device_d3d12.cpp:157-:163:

  • Make drain() return bool.
  • In ~D3d12Executor, if the drain timed out and GetDeviceRemovedReason() is S_OK: log it and retain the whole runtime. Skip begin_teardown, runtime_destroy and the proxy, mpoint_ and retired_ releases, and deliberately leak rt_.
  • If the device is removed, release as now.

The cost is a leak of the arena (the size of the network) plus the proxies, per occurrence, until the process exits. A new context on the same device then has less VRAM; the OOM wait at device_d3d12.cpp:1170 copes by waiting and failing cleanly. A later improvement would be a process-wide list that keeps the executor and releases it once external_done() catches up.

Test

CPU / WARP, with a seam so external_done() can be faked. dx12/tests/native_direct_test.cpp has a WARP Gpu helper (:46). Call begin_external and end_external on a list that is never executed, destroy the device, and assert that a tracked probe object was not released. A GPU test is needed for the full executor path.

Related: #63, #58 (objects released while a recorded list may still use them), #68.

Lingua principale
C++
Stelle
12
Fork
2
Merge medio
3h 10m
PR unite (30g)
171

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di phantoslayer21/wronggpu

Tutte le issue di phantoslayer21/wronggpu

Issue simili

Altre issue su C++

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.