drm/amdgpu: reuse VMIDs assigned to a VM only if there is also a free one
authorChristian König <christian.koenig@amd.com>
Mon, 23 May 2016 14:00:32 +0000 (16:00 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 7 Jul 2016 18:51:23 +0000 (14:51 -0400)
This fixes a fairness problem with the GPU scheduler. VM having lot of
jobs could previously starve VM with less jobs.

Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index cf989cf..f38978d 100644 (file)
@@ -179,75 +179,80 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
        uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
        struct amdgpu_device *adev = ring->adev;
        struct fence *updates = sync->last_vm_update;
-       struct amdgpu_vm_id *id;
+       struct amdgpu_vm_id *id, *idle;
        unsigned i = ring->idx;
        int r;
 
        mutex_lock(&adev->vm_manager.lock);
 
-       /* Check if we can use a VMID already assigned to this VM */
-       do {
-               struct fence *flushed;
-
-               id = vm->ids[i++];
-               if (i == AMDGPU_MAX_RINGS)
-                       i = 0;
-
-               /* Check all the prerequisites to using this VMID */
-               if (!id)
-                       continue;
-
-               if (atomic64_read(&id->owner) != vm->client_id)
-                       continue;
-
-               if (pd_addr != id->pd_gpu_addr)
-                       continue;
+       /* Check if we have an idle VMID */
+       list_for_each_entry(idle, &adev->vm_manager.ids_lru, list) {
+               if (amdgpu_sync_is_idle(&idle->active, ring))
+                       break;
 
-               if (id->last_user != ring &&
-                   (!id->last_flush || !fence_is_signaled(id->last_flush)))
-                       continue;
+       }
 
-               flushed  = id->flushed_updates;
-               if (updates && (!flushed || fence_is_later(updates, flushed)))
-                       continue;
+       /* If we can't find a idle VMID to use, just wait for the oldest */
+       if (&idle->list == &adev->vm_manager.ids_lru) {
+               id = list_first_entry(&adev->vm_manager.ids_lru,
+                                     struct amdgpu_vm_id,
+                                     list);
+       } else {
+               /* Check if we can use a VMID already assigned to this VM */
+               do {
+                       struct fence *flushed;
+
+                       id = vm->ids[i++];
+                       if (i == AMDGPU_MAX_RINGS)
+                               i = 0;
+
+                       /* Check all the prerequisites to using this VMID */
+                       if (!id)
+                               continue;
+
+                       if (atomic64_read(&id->owner) != vm->client_id)
+                               continue;
+
+                       if (pd_addr != id->pd_gpu_addr)
+                               continue;
+
+                       if (id->last_user != ring && (!id->last_flush ||
+                           !fence_is_signaled(id->last_flush)))
+                               continue;
+
+                       flushed  = id->flushed_updates;
+                       if (updates && (!flushed ||
+                           fence_is_later(updates, flushed)))
+                               continue;
+
+                       /* Good we can use this VMID */
+                       if (id->last_user == ring) {
+                               r = amdgpu_sync_fence(ring->adev, sync,
+                                                     id->first);
+                               if (r)
+                                       goto error;
+                       }
 
-               /* Good we can use this VMID */
-               if (id->last_user == ring) {
-                       r = amdgpu_sync_fence(ring->adev, sync,
-                                             id->first);
+                       /* And remember this submission as user of the VMID */
+                       r = amdgpu_sync_fence(ring->adev, &id->active, fence);
                        if (r)
                                goto error;
-               }
-
-               /* And remember this submission as user of the VMID */
-               r = amdgpu_sync_fence(ring->adev, &id->active, fence);
-               if (r)
-                       goto error;
 
-               list_move_tail(&id->list, &adev->vm_manager.ids_lru);
-               vm->ids[ring->idx] = id;
+                       list_move_tail(&id->list, &adev->vm_manager.ids_lru);
+                       vm->ids[ring->idx] = id;
 
-               *vm_id = id - adev->vm_manager.ids;
-               *vm_pd_addr = AMDGPU_VM_NO_FLUSH;
-               trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id, *vm_pd_addr);
+                       *vm_id = id - adev->vm_manager.ids;
+                       *vm_pd_addr = AMDGPU_VM_NO_FLUSH;
+                       trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id,
+                                               *vm_pd_addr);
 
-               mutex_unlock(&adev->vm_manager.lock);
-               return 0;
+                       mutex_unlock(&adev->vm_manager.lock);
+                       return 0;
 
-       } while (i != ring->idx);
+               } while (i != ring->idx);
 
-       /* Check if we have an idle VMID */
-       list_for_each_entry(id, &adev->vm_manager.ids_lru, list) {
-               if (amdgpu_sync_is_idle(&id->active, ring))
-                       break;
-
-       }
-
-       /* If we can't find a idle VMID to use, just wait for the oldest */
-       if (&id->list == &adev->vm_manager.ids_lru) {
-               id = list_first_entry(&adev->vm_manager.ids_lru,
-                                     struct amdgpu_vm_id,
-                                     list);
+               /* Still no ID to use? Then use the idle one found earlier */
+               id = idle;
        }
 
        r = amdgpu_sync_cycle_fences(sync, &id->active, fence);