drm/radeon: fix VM flush handlingvm-flushing

Because of possible hardware race conditions
we need to flush each ring individually.

Signed-off-by: Christian König <christian.koenig@amd.com>

3 files changed, 17 insertions, 30 deletions

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 99605c96a9b2..4a5f326d9461 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -895,12 +895,13 @@ struct radeon_vm {
 	struct radeon_bo_va		*ib_bo_va;
 
 	struct mutex			mutex;
-	/* last fence for cs using this vm */
-	struct radeon_fence		*fence;
-	/* last flush or NULL if we still need to flush */
-	struct radeon_fence		*last_flush;
+
+	/* last PT update */
+	struct radeon_fence		*last_pt_update;
 	/* last use of vmid */
 	struct radeon_fence		*last_id_use;
+	/* last time we flushed on each ring */
+	struct radeon_fence		*last_flushes[RADEON_NUM_RINGS];
 };
 
 struct radeon_vm_manager {
diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index cb12df784d83..aab06acd1c6c 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -538,7 +538,7 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 		goto out;
 	}
 	radeon_cs_sync_rings(parser);
-	radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence);
+	radeon_semaphore_sync_to(parser->ib.semaphore, vm->last_pt_update);
 
 	if ((rdev->family >= CHIP_TAHITI) &&
 	    (parser->chunk_const_ib_idx != -1)) {
diff --git a/drivers/gpu/drm/radeon/radeon_vm.c b/drivers/gpu/drm/radeon/radeon_vm.c
index 8d15ce6187da..82479c4da40e 100644
--- a/drivers/gpu/drm/radeon/radeon_vm.c
+++ b/drivers/gpu/drm/radeon/radeon_vm.c
@@ -187,9 +187,6 @@ struct radeon_fence *radeon_vm_grab_id(struct radeon_device *rdev,
 	if (vm->last_id_use && vm->last_id_use == rdev->vm_manager.active[vm->id])
 		return NULL;
 
-	/* we definately need to flush */
-	radeon_fence_unref(&vm->last_flush);
-
 	/* skip over VMID 0, since it is the system VM */
 	for (i = 1; i < rdev->vm_manager.nvm; ++i) {
 		struct radeon_fence *fence = rdev->vm_manager.active[i];
@@ -236,14 +233,14 @@ void radeon_vm_flush(struct radeon_device *rdev,
 		     int ring)
 {
 	uint64_t pd_addr = radeon_bo_gpu_offset(vm->page_directory);
+	struct radeon_fence **last_flush = &vm->last_flushes[ring];
 
-	/* if we can't remember our last VM flush then flush now! */
-	/* flushing only when needed doesn't work because of a hardware race
-	   condition between the DMA and the GFX engines */
-	if (!vm->last_flush || true || pd_addr != vm->pd_gpu_addr) {
+	if (*last_flush != vm->last_pt_update || pd_addr != vm->pd_gpu_addr) {
 		trace_radeon_vm_flush(pd_addr, ring, vm->id);
 		vm->pd_gpu_addr = pd_addr;
 		radeon_ring_vm_flush(rdev, ring, vm);
+		radeon_fence_unref(last_flush);
+		*last_flush = radeon_fence_ref(vm->last_pt_update);
 	}
 }
 
@@ -263,18 +260,11 @@ void radeon_vm_fence(struct radeon_device *rdev,
 		     struct radeon_vm *vm,
 		     struct radeon_fence *fence)
 {
-	radeon_fence_unref(&vm->fence);
-	vm->fence = radeon_fence_ref(fence);
-
 	radeon_fence_unref(&rdev->vm_manager.active[vm->id]);
 	rdev->vm_manager.active[vm->id] = radeon_fence_ref(fence);
 
 	radeon_fence_unref(&vm->last_id_use);
 	vm->last_id_use = radeon_fence_ref(fence);
-
-        /* we just flushed the VM, remember that */
-        if (!vm->last_flush)
-                vm->last_flush = radeon_fence_ref(fence);
 }
 
 /**
@@ -704,9 +694,8 @@ int radeon_vm_update_page_directory(struct radeon_device *rdev,
 			radeon_ib_free(rdev, &ib);
 			return r;
 		}
-		radeon_fence_unref(&vm->fence);
-		vm->fence = radeon_fence_ref(ib.fence);
-		radeon_fence_unref(&vm->last_flush);
+		radeon_fence_unref(&vm->last_pt_update);
+		vm->last_pt_update = radeon_fence_ref(ib.fence);
 	}
 	radeon_ib_free(rdev, &ib);
 
@@ -962,16 +951,15 @@ int radeon_vm_bo_update(struct radeon_device *rdev,
 	radeon_asic_vm_pad_ib(rdev, &ib);
 	WARN_ON(ib.length_dw > ndw);
 
-	radeon_semaphore_sync_to(ib.semaphore, vm->fence);
+	radeon_semaphore_sync_to(ib.semaphore, vm->last_id_use);
 	r = radeon_ib_schedule(rdev, &ib, NULL);
 	if (r) {
 		radeon_ib_free(rdev, &ib);
 		return r;
 	}
-	radeon_fence_unref(&vm->fence);
-	vm->fence = radeon_fence_ref(ib.fence);
+	radeon_fence_unref(&vm->last_pt_update);
+	vm->last_pt_update = radeon_fence_ref(ib.fence);
 	radeon_ib_free(rdev, &ib);
-	radeon_fence_unref(&vm->last_flush);
 
 	return 0;
 }
@@ -1101,8 +1089,7 @@ int radeon_vm_init(struct radeon_device *rdev, struct radeon_vm *vm)
 
 	vm->id = 0;
 	vm->ib_bo_va = NULL;
-	vm->fence = NULL;
-	vm->last_flush = NULL;
+	vm->last_pt_update = NULL;
 	vm->last_id_use = NULL;
 	mutex_init(&vm->mutex);
 	vm->va = RB_ROOT;
@@ -1173,8 +1160,7 @@ void radeon_vm_fini(struct radeon_device *rdev, struct radeon_vm *vm)
 
 	radeon_bo_unref(&vm->page_directory);
 
-	radeon_fence_unref(&vm->fence);
-	radeon_fence_unref(&vm->last_flush);
+	radeon_fence_unref(&vm->last_pt_update);
 	radeon_fence_unref(&vm->last_id_use);
 
 	mutex_destroy(&vm->mutex);