drm/amdgpu: stop adding VM updates fences to the resv obj

Don't add the VM update fences to the resv object and remove
the handling to stop implicitely syncing to them.

Ongoing updates prevent page tables from being evicted and we manually
block for all updates to complete before releasing PDs and PTS.

This way we can do updates even without the resv obj locked.

Signed-off-by: Christian König <[e-mail]>
Reviewed-by: Felix Kuehling <[e-mail]>
Signed-off-by: Alex Deucher <[e-mail]>

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

···
 			continue;
 
 		if (amdgpu_sync_same_dev(adev, f)) {
-			/* VM updates are only interesting
-			 * for other VM updates and moves.
+			/* VM updates only sync with moves but not with user
+			 * command submissions or KFD evictions fences
 			 */
-			if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
-			    (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
-			    ((owner == AMDGPU_FENCE_OWNER_VM) !=
-			     (fence_owner == AMDGPU_FENCE_OWNER_VM)))
+			if (owner == AMDGPU_FENCE_OWNER_VM &&
+			    fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED)
 				continue;
 
 			/* Ignore fence from the same owner and explicit one as

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

···
 {
 	entry->priority = 0;
 	entry->tv.bo = &vm->root.base.bo->tbo;
-	/* One for the VM updates, one for TTM and one for the CS job */
-	entry->tv.num_shared = 3;
+	/* One for TTM and one for the CS job */
+	entry->tv.num_shared = 2;
 	entry->user_pages = NULL;
 	list_add(&entry->tv.head, validated);
 }
···
 	if (!dma_resv_test_signaled_rcu(bo->tbo.base.resv, true))
 		return false;
 
+	/* Don't evict VM page tables while they are updated */
+	if (!dma_fence_is_signaled(bo_base->vm->last_direct) ||
+	    !dma_fence_is_signaled(bo_base->vm->last_delayed))
+		return false;
+
 	return true;
 }
 
···
  */
 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
 {
-	return dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
-						   true, true, timeout);
+	timeout = dma_resv_wait_timeout_rcu(vm->root.base.bo->tbo.base.resv,
+					    true, true, timeout);
+	if (timeout <= 0)
+		return timeout;
+
+	timeout = dma_fence_wait_timeout(vm->last_direct, true, timeout);
+	if (timeout <= 0)
+		return timeout;
+
+	return dma_fence_wait_timeout(vm->last_delayed, true, timeout);
 }
 
 /**
···
 	else
 		vm->update_funcs = &amdgpu_vm_sdma_funcs;
 	vm->last_update = NULL;
+	vm->last_direct = dma_fence_get_stub();
+	vm->last_delayed = dma_fence_get_stub();
 
 	amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, false, &bp);
 	if (vm_context == AMDGPU_VM_CONTEXT_COMPUTE)
···
 	vm->root.base.bo = NULL;
 
 error_free_delayed:
+	dma_fence_put(vm->last_direct);
+	dma_fence_put(vm->last_delayed);
 	drm_sched_entity_destroy(&vm->delayed);
 
 error_free_direct:
···
 		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
 		vm->pasid = 0;
 	}
+
+	dma_fence_wait(vm->last_direct, false);
+	dma_fence_put(vm->last_direct);
+	dma_fence_wait(vm->last_delayed, false);
+	dma_fence_put(vm->last_delayed);
 
 	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
 		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {

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

···
 	struct drm_sched_entity	direct;
 	struct drm_sched_entity	delayed;
 
+	/* Last submission to the scheduler entities */
+	struct dma_fence	*last_direct;
+	struct dma_fence	*last_delayed;
+
 	unsigned int		pasid;
 	/* dedicated to vm */
 	struct amdgpu_vmid	*reserved_vmid[AMDGPU_MAX_VMHUBS];

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

···
 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
 				 struct dma_fence **fence)
 {
-	struct amdgpu_bo *root = p->vm->root.base.bo;
 	struct amdgpu_ib *ib = p->job->ibs;
 	struct drm_sched_entity *entity;
+	struct dma_fence *f, *tmp;
 	struct amdgpu_ring *ring;
-	struct dma_fence *f;
 	int r;
 
 	entity = p->direct ? &p->vm->direct : &p->vm->delayed;
···
 	if (r)
 		goto error;
 
-	amdgpu_bo_fence(root, f, true);
+	tmp = dma_fence_get(f);
+	if (p->direct)
+		swap(p->vm->last_direct, tmp);
+	else
+		swap(p->vm->last_delayed, tmp);
+	dma_fence_put(tmp);
+
 	if (fence && !p->direct)
 		swap(*fence, f);
 	dma_fence_put(f);