add a retry mechanism for waiting on the last submission in Queue::drop

wgpu-core/src/device/queue.rs

@@ -149,21 +149,33 @@ impl Drop for Queue {
             .load(Ordering::Acquire);
 
         let fence = self.device.fence.read();
+
+        // Try waiting on the last submission using the following sequence of timeouts
+        let timeouts_in_ms = [100, 200, 400, 800, 1600, 3200];
+
+        for (i, timeout_ms) in timeouts_in_ms.into_iter().enumerate() {
+            let is_last_iter = i == timeouts_in_ms.len() - 1;
+
+            api_log!(
+                "Waiting on last submission. try: {}/{}. timeout: {}ms",
+                i + 1,
+                timeouts_in_ms.len(),
+                timeout_ms
+            );
+
             let wait_res = unsafe {
                 self.device.raw().wait(
                     fence.as_ref(),
                     last_successful_submission_index,
                     #[cfg(not(target_arch = "wasm32"))]
-                crate::device::CLEANUP_WAIT_MS,
+                    timeout_ms,
                     #[cfg(target_arch = "wasm32")]
                     0, // WebKit and Chromium don't support a non-0 timeout
                 )
             };
-        drop(fence);
+            // Note: If we don't panic below we are in UB land (destroying resources while they are still in use by the GPU).
-
             match wait_res {
-            Ok(true) => {}
+                Ok(true) => break,
-            // Note: If we don't panic here we are in UB land (destroying resources while they are still in use by the GPU).
                 Ok(false) => {
                     // It's fine that we timed out on WebGL; GL objects can be deleted early as they
                     // will be kept around by the driver if GPU work hasn't finished.
@@ -171,15 +183,41 @@ impl Drop for Queue {
                     // backends since getBufferSubData is synchronous with respect to the other previously enqueued GL commands.
                     // Relying on this behavior breaks the clean abstraction wgpu-hal tries to maintain and
                     // we should find ways to improve this. See https://github.com/gfx-rs/wgpu/issues/6538.
-                #[cfg(not(target_arch = "wasm32"))]
+                    #[cfg(target_arch = "wasm32")]
-                panic!("We timed out while waiting on the last successful submission to complete!");
+                    {
+                        break;
                     }
-            Err(e) => {
+                    #[cfg(not(target_arch = "wasm32"))]
+                    {
+                        if is_last_iter {
                             panic!(
-                    "We ran into an error while waiting on the last successful submission to complete! - {e}"
+                                "We timed out while waiting on the last successful submission to complete!"
                             );
                         }
                     }
+                }
+                Err(e) => match e {
+                    hal::DeviceError::OutOfMemory => {
+                        if is_last_iter {
+                            panic!(
+                                "We ran into an OOM error while waiting on the last successful submission to complete!"
+                            );
+                        }
+                    }
+                    hal::DeviceError::Lost => {
+                        self.device.handle_hal_error(e); // will lose the device
+                        break;
+                    }
+                    hal::DeviceError::ResourceCreationFailed => unreachable!(),
+                    hal::DeviceError::Unexpected => {
+                        panic!(
+                            "We ran into an unexpected error while waiting on the last successful submission to complete!"
+                        );
+                    }
+                },
+            }
+        }
+        drop(fence);
 
         let snatch_guard = self.device.snatchable_lock.read();
         let (submission_closures, mapping_closures, queue_empty) =