drm/amdgpu: grab VMID before submitting job v5

This allows the scheduler to handle the dependencies on ID contention as well.

v2: grab id only once
v3: use a separate lock for the VMIDs
v4: cleanup after semaphore removal
v5: minor coding style change
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -777,6 +777,7 @@ struct amdgpu_ib {
 	struct amdgpu_ring		*ring;
 	struct amdgpu_fence		*fence;
 	struct amdgpu_user_fence        *user;
+	bool				grabbed_vmid;
 	struct amdgpu_vm		*vm;
 	struct amdgpu_ctx		*ctx;
 	struct amdgpu_sync		sync;
@@ -925,6 +926,9 @@ struct amdgpu_vm {
 };
 
 struct amdgpu_vm_manager {
+	/* protecting IDs */
+	struct mutex				lock;
+
 	struct {
 		struct fence	*active;
 		atomic_long_t	owner;

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

@@ -1456,6 +1456,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	/* mutex initialization are all done here so we
 	 * can recall function without having locking issues */
 	mutex_init(&adev->ring_lock);
+	mutex_init(&adev->vm_manager.lock);
 	atomic_set(&adev->irq.ih.lock, 0);
 	mutex_init(&adev->gem.mutex);
 	mutex_init(&adev->pm.mutex);

drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c

@@ -142,20 +142,16 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 		return -EINVAL;
 	}
 
-	r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
-	if (r) {
-		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
-		return r;
+	if (vm && !ibs->grabbed_vmid) {
+		dev_err(adev->dev, "VM IB without ID\n");
+		return -EINVAL;
 	}
 
-	if (vm) {
-		/* grab a vm id if necessary */
-		r = amdgpu_vm_grab_id(ibs->vm, ibs->ring, &ibs->sync);
+	r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
 	if (r) {
-			amdgpu_ring_unlock_undo(ring);
+		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
 		return r;
 	}
-	}
 
 	r = amdgpu_sync_wait(&ibs->sync);
 	if (r) {
@@ -207,9 +203,6 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 				       AMDGPU_FENCE_FLAG_64BIT);
 	}
 
-	if (ib->vm)
-		amdgpu_vm_fence(adev, ib->vm, &ib->fence->base);
-
 	amdgpu_ring_unlock_commit(ring);
 	return 0;
 }

drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c

@@ -31,7 +31,31 @@
 static struct fence *amdgpu_sched_dependency(struct amd_sched_job *sched_job)
 {
 	struct amdgpu_job *job = to_amdgpu_job(sched_job);
-	return amdgpu_sync_get_fence(&job->ibs->sync);
+	struct amdgpu_sync *sync = &job->ibs->sync;
+	struct amdgpu_vm *vm = job->ibs->vm;
+
+	struct fence *fence = amdgpu_sync_get_fence(sync);
+
+	if (fence == NULL && vm && !job->ibs->grabbed_vmid) {
+		struct amdgpu_ring *ring = job->ibs->ring;
+		struct amdgpu_device *adev = ring->adev;
+		int r;
+
+		mutex_lock(&adev->vm_manager.lock);
+		r = amdgpu_vm_grab_id(vm, ring, sync);
+		if (r) {
+			DRM_ERROR("Error getting VM ID (%d)\n", r);
+		} else {
+			fence = &job->base.s_fence->base;
+			amdgpu_vm_fence(ring->adev, vm, fence);
+			job->ibs->grabbed_vmid = true;
+		}
+		mutex_unlock(&adev->vm_manager.lock);
+
+		fence = amdgpu_sync_get_fence(sync);
+	}
+
+	return fence;
 }
 
 static struct fence *amdgpu_sched_run_job(struct amd_sched_job *sched_job)