Eagerly release GPU resources when we lose the device. (#4851)

Co-authored-by: Connor Fitzgerald <connorwadefitzgerald@gmail.com>

CHANGELOG.md

@@ -65,6 +65,7 @@ Wgpu now exposes backend feature for the Direct3D 12 (`dx12`) and Metal (`metal`
 - Added `DownlevelFlags::VERTEX_AND_INSTANCE_INDEX_RESPECTS_RESPECTIVE_FIRST_VALUE_IN_INDIRECT_DRAW` to know if `@builtin(vertex_index)` and `@builtin(instance_index)` will respect the `first_vertex` / `first_instance` in indirect calls. If this is not present, both will always start counting from 0. Currently enabled on all backends except DX12. By @cwfitzgerald in [#4722](https://github.com/gfx-rs/wgpu/pull/4722)
 - No longer validate surfaces against their allowed extent range on configure. This caused warnings that were almost impossible to avoid. As before, the resulting behavior depends on the compositor. By @wumpf in [#4796](https://github.com/gfx-rs/wgpu/pull/4796)
 - Added support for the float32-filterable feature. By @almarklein in [#4759](https://github.com/gfx-rs/wgpu/pull/4759)
+- GPU buffer memory is released during "lose the device". By @bradwerth in [#4851](https://github.com/gfx-rs/wgpu/pull/4851)
 - wgpu and wgpu-core features are now documented on docs.rs. By @wumpf in [#4886](https://github.com/gfx-rs/wgpu/pull/4886)
 - DeviceLostClosure is guaranteed to be invoked exactly once. By @bradwerth in [#4862](https://github.com/gfx-rs/wgpu/pull/4862)
 

wgpu-core/src/device/life.rs

@@ -241,7 +241,7 @@ pub(crate) struct LifetimeTracker<A: HalApi> {
     ready_to_map: Vec<Arc<Buffer<A>>>,
 
     /// Queue "on_submitted_work_done" closures that were initiated for while there is no
-    /// currently pending submissions. These cannot be immeidately invoked as they
+    /// currently pending submissions. These cannot be immediately invoked as they
     /// must happen _after_ all mapped buffer callbacks are mapped, so we defer them
     /// here until the next time the device is maintained.
     work_done_closures: SmallVec<[SubmittedWorkDoneClosure; 1]>,
@@ -1017,4 +1017,47 @@ impl<A: HalApi> LifetimeTracker<A> {
         }
         pending_callbacks
     }
+
+    pub(crate) fn release_gpu_resources(&mut self) {
+        // This is called when the device is lost, which makes every associated
+        // resource invalid and unusable. This is an opportunity to release all of
+        // the underlying gpu resources, even though the objects remain visible to
+        // the user agent. We purge this memory naturally when resources have been
+        // moved into the appropriate buckets, so this function just needs to
+        // initiate movement into those buckets, and it can do that by calling
+        // "destroy" on all the resources we know about which aren't already marked
+        // for cleanup.
+
+        // During these iterations, we discard all errors. We don't care!
+
+        // Destroy all the mapped buffers.
+        for buffer in &self.mapped {
+            let _ = buffer.destroy();
+        }
+
+        // Destroy all the unmapped buffers.
+        for buffer in &self.ready_to_map {
+            let _ = buffer.destroy();
+        }
+
+        // Destroy all the future_suspected_buffers.
+        for buffer in &self.future_suspected_buffers {
+            let _ = buffer.destroy();
+        }
+
+        // Destroy the buffers in all active submissions.
+        for submission in &self.active {
+            for buffer in &submission.mapped {
+                let _ = buffer.destroy();
+            }
+        }
+
+        // Destroy all the future_suspected_textures.
+        // TODO: texture.destroy is not implemented
+        /*
+        for texture in &self.future_suspected_textures {
+            let _ = texture.destroy();
+        }
+        */
+    }
 }

wgpu-core/src/device/resource.rs

@@ -381,15 +381,19 @@ impl<A: HalApi> Device<A> {
         // our caller. This will complete the steps for both destroy and for
         // "lose the device".
         let mut device_lost_invocations = SmallVec::new();
-        if !self.is_valid()
-            && life_tracker.queue_empty()
-            && life_tracker.device_lost_closure.is_some()
-        {
-            device_lost_invocations.push(DeviceLostInvocation {
-                closure: life_tracker.device_lost_closure.take().unwrap(),
-                reason: DeviceLostReason::Destroyed,
-                message: String::new(),
-            });
+        if !self.is_valid() && life_tracker.queue_empty() {
+            // We can release gpu resources associated with this device.
+            life_tracker.release_gpu_resources();
+
+            // If we have a DeviceLostClosure, build an invocation with the
+            // reason DeviceLostReason::Destroyed and no message.
+            if life_tracker.device_lost_closure.is_some() {
+                device_lost_invocations.push(DeviceLostInvocation {
+                    closure: life_tracker.device_lost_closure.take().unwrap(),
+                    reason: DeviceLostReason::Destroyed,
+                    message: String::new(),
+                });
+            }
         }
 
         let closures = UserClosures {
@@ -3362,9 +3366,13 @@ impl<A: HalApi> Device<A> {
         self.valid.store(false, Ordering::Release);
 
         // 1) Resolve the GPUDevice device.lost promise.
-        let closure = self.lock_life().device_lost_closure.take();
+        let mut life_lock = self.lock_life();
+        let closure = life_lock.device_lost_closure.take();
         if let Some(device_lost_closure) = closure {
+            // It's important to not hold the lock while calling the closure.
+            drop(life_lock);
             device_lost_closure.call(DeviceLostReason::Unknown, message.to_string());
+            life_lock = self.lock_life();
         }
 
         // 2) Complete any outstanding mapAsync() steps.
@@ -3374,6 +3382,9 @@ impl<A: HalApi> Device<A> {
         // since that will prevent any new work from being added to the queues.
         // Future calls to poll_devices will continue to check the work queues
         // until they are cleared, and then drop the device.
+
+        // Eagerly release GPU resources.
+        life_lock.release_gpu_resources();
     }
 }