1633661 - Crash in [@ radv_CmdBindDescriptorSets]

Status: RESOLVED FIXED

(Core :: Graphics: WebGPU, defect, P2)

Description

[Gabriele Svelto [:gsvelto] (PTO)]

This bug is for crash report bp-a7d742fb-a97c-4711-bd78-8c3ee0200424.

Top 10 frames of crashing thread:

0 libvulkan_radeon.so radv_CmdBindDescriptorSets ./build/../src/amd/vulkan/radv_radeon_winsys.h:362
1 libxul.so wgpu_core::command::render::<impl wgpu_core::hub::Global<G>>::command_encoder_run_render_pass third_party/rust/ash/src/vk.rs:5714
2 libxul.so mozilla::webgpu::WebGPUParent::RecvCommandEncoderRunRenderPass dom/webgpu/ipc/WebGPUParent.cpp:350
3 libxul.so mozilla::webgpu::PWebGPUParent::OnMessageReceived ipc/ipdl/PWebGPUParent.cpp:1503
4 libxul.so mozilla::layers::PCompositorManagerParent::OnMessageReceived ipc/ipdl/PCompositorManagerParent.cpp:197
5 libxul.so mozilla::ipc::MessageChannel::DispatchMessage ipc/glue/MessageChannel.cpp:2186
6 libxul.so mozilla::ipc::MessageChannel::MessageTask::Run ipc/glue/MessageChannel.cpp:1989
7 libxul.so base::MessagePumpDefault::Run ipc/chromium/src/base/message_pump_default.cc:35
8 libxul.so MessageLoop::Run ipc/chromium/src/base/message_loop.cc:290
9 libxul.so base::Thread::ThreadMain ipc/chromium/src/base/thread.cc:192

We now have debug symbols for Vulkan libraries on Linux and this is the first crash that popped up. This is happening on multiple Linux distros (Arch, Debian, Ubuntu and derivatives) and for the time being seems to happen only on Mesa 19.2.8.0 and 19.3.3.0. Unfortunately many of the crash reports are orphaned, i.e. they miss the crash metadata information.

Comment 1

[Dzmitry Malyshau [:kvark]]

This could be caused by misformed applications that aren't correctly binding the GPUBindGroup entries. We still don't have full validation of shader requirements implemented, so some of the invalid use (like that) may come through. It would help to confirm this if we knew the URLs crashing.

Comment 2

[Gabriele Svelto [:gsvelto] (PTO)]

Here you go:

https://wgpu.rs/examples/?example=boids
https://wgpu.rs/examples/?example=cube
https://wgpu.rs/examples/?example=mipmap
https://wgpu.rs/examples/?example=shadow
https://wgpu.rs/examples/?example=skybox

Comment 3

[Dzmitry Malyshau [:kvark]]

Interesting, thank you! I'll see if I can find a Linux AMD machine to reproduce this in isolation.

Comment 4

[Bob Clary [:bc] (inactive)]

I have a Ryzen 9 3950X with a RX 5700 XT gpu and can be remotely driven as a debugger if you need it.

Comment 5

[Dzmitry Malyshau [:kvark]]

I have a fix ready. The problem here is that AMD driver expects the descriptor set layout to be alive. The Vulkan spec is not clear on that, and I requested clarification from them in the meantime.

Comment 6

[Dzmitry Malyshau [:kvark]]

Attachment: Keep wgpu bind group layouts alive

Comment 7

[BugBot [:suhaib / :marco/ :calixte]]

The severity field is not set for this bug.
:kvark, could you have a look please?

For more information, please visit auto_nag documentation.

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/170cb6b6d830
Keep wgpu bind group layouts alive r=groves

Comment 9

[Noemi Erli[:noemi_erli]]

https://hg.mozilla.org/mozilla-central/rev/170cb6b6d830