Completion markers assume command lists run in recording order

Open
#64 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
15/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
cpp

Research direction

Start from clean/rr_f/runtime/src/rr_runtime.cpp around marker_done (:496), pump (:505) and the slot wait (:2570), then look at the transient and arena release paths at :3448 and :3544. The fix is a design choice between a per-recording marker ring and documenting the submit-order contract. Done means the WARP test that records lists A (serial 1) and B (serial 2), runs only B, and asserts serial 1 is not done passes.

Written by the indexing model from the issue text.

Description

area: ray-reconstruction area: runtime enhancement needs-hardware-test

What happens

Both runtimes use one scalar completion marker and treat "marker observed >= s" as "every recording up to s is done". The marker is written by the recorded list, so this only holds if lists execute in the order they were recorded.

  • RR: clean/rr_f/runtime/src/rr_runtime.cpp:496 (marker_done), pump (:505), slot wait (:2570), frame marker (:3408), transient marker (:3525).
  • Model M: clean/runtime/device_d3d12.cpp:400 (end_external), :421 (external_done), :996 (wait_external).

Model M: the marker belongs to one device, so to one context. If one context's frames ran out of order, its temporal history would already be wrong, so this is not a case that matters in practice.

RR: the marker is runtime-wide (rt.marker, rt.serial) and shared by every instance and by transient copies. If instance B's list runs before instance A's earlier list, B's serial covers A's outstanding work, and defer_release of A's old arena (:3448) or of a transient (:3544) can release early. Out-of-order execution can also move the marker backwards, because the lower serial is written last. That causes slot-wait timeouts (:2572) and failed frames rather than a use-after-free.

When it can happen

Rarely. A use-after-free needs several RR instances (for example two viewports) with lists submitted out of order or on different queues, and an arena regrow, resize or transient release inside that window. Transient copies recorded on a list that is submitted after the evaluate list already break the data dependency.

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

To do

This is a design discussion. Options:

  • Replace the single maximum with a per-recording marker ring. The readback buffer is 256 bytes, so 64 dwords. Each recording writes its serial into dword s % N, and recording s is done when that dword equals s (low 32 bits) or a later serial that maps to the same dword. Bound the recordings in flight to N; the slot waits already bound this. Wrap handling needs care.
  • Or document the contract: one RR instance at a time, or submit in recording order.

Test

WARP. Record list A (serial 1), then list B (serial 2), using only WriteBufferImmediate. Execute B alone and wait for the queue. Assert that "serial 1 done" is false. This fails today and would pass with the ring.

Related: #63, #62, #65.

Dominant language
C++
Stars
12
Forks
2
Avg merge
3h 10m
Merged PRs (30d)
171

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from phantoslayer21/wronggpu

All issues in phantoslayer21/wronggpu

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.