Fix deadlock between snatchable_lock and trackers in Queue::write_texture #7004
Conversation
|
Its actually not completely fixing the issue mentioned. Just one occurrence that I ran into. |
|
I just noticed another easy-to-fix deadlock. I will add another commit soon. (Sorry) |
|
I am not sure how feasible it is but I would suggest adding some internal documentation that exactly specifies in which order locks should be acquired if you need multiple. Like it was done for snatchable_lock at wgpu-core/src/device/resource.rs:88 The ordering that I detected from reading code that references the mutexes: |
|
This should be defined by https://github.com/gfx-rs/wgpu/blob/trunk/wgpu-core/src/lock/rank.rs - I'm not sure what the current status of mutex rank checking. |
- reimplement LAYERWAIT - fix for_layer_in_range to use inclusive [x; y] ranges instead of semi-internvals [x; y) - fix SEWAIT racing with the first set of the status by the audio loop - add deadlock detection to catch the elusive deadlock in wgpu (though it's probably fixed by gfx-rs/wgpu#7004) - fix yet another dynamic buffer-related crash
- reimplement LAYERWAIT - fix for_layer_in_range to use inclusive [x; y] ranges instead of semi-internvals [x; y) - fix SEWAIT racing with the first set of the status by the audio loop - add deadlock detection to catch the elusive deadlock in wgpu (though it's probably fixed by gfx-rs/wgpu#7004) - fix yet another dynamic buffer-related crash
|
Would it be possible to backport this as 24.0.x bugfix release? I am reproducing this deadlock quite reliably in my engine :(. In my case it's between |
|
If you could backport it (make a PR against the v24 branch) , we'll be doing a release next week and can release it |
…ture (gfx-rs#7004) * Fix deadlock between snatchable_lock and trackers in Queue::write_texture * Fix another deadlock in write_texture between pending_writes and snatchable_lock. # Conflicts: # CHANGELOG.md
…ture (#7004) * Fix deadlock between snatchable_lock and trackers in Queue::write_texture * Fix another deadlock in write_texture between pending_writes and snatchable_lock. # Conflicts: # CHANGELOG.md
…ture (#7004) * Fix deadlock between snatchable_lock and trackers in Queue::write_texture * Fix another deadlock in write_texture between pending_writes and snatchable_lock. # Conflicts: # CHANGELOG.md
Connections
Related to #5737
Description
The
Device.snatchable_lockis usually acquired beforeDevice.trackersis. OnlyQueue.write_texture()acquired them in reverse order, introducing deadlocks. This PR swaps the order there to align with other usages of those mutexes.Testing
Manually checked all references to trackers mutex for snatchable_lock guards in the same scope.
Checklist
cargo fmt.taplo format.cargo clippy. If applicable, add:--target wasm32-unknown-unknown--target wasm32-unknown-emscriptencargo xtask testto run tests. (Some tests failed here, but those fail without my change too)CHANGELOG.md. See simple instructions inside file.