drm/ttm: add a way to bo_wait for either the last read or last write

Sometimes we want to know whether a buffer is busy and wait for it (bo_wait).
However, sometimes it would be more useful to be able to query whether
a buffer is busy and being either read or written, and wait until it's stopped
being either read or written. The point of this is to be able to avoid
unnecessary waiting, e.g. if a GPU has written something to a buffer and is now
reading that buffer, and a CPU wants to map that buffer for read, it needs to
only wait for the last write. If there were no write, there wouldn't be any
waiting needed.

This, or course, requires user space drivers to send read/write flags
with each relocation (like we have read/write domains in radeon, so we can
actually use those for something useful now).

Now how this patch works:

The read/write flags should passed to ttm_validate_buffer. TTM maintains
separate sync objects of the last read and write for each buffer, in addition
to the sync object of the last use of a buffer. ttm_bo_wait then operates
with one the sync objects.

Signed-off-by: Marek Olšák <maraeo@gmail.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

4 files changed, 111 insertions, 33 deletions

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index a4d38d85909..b824d9bdd87 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -498,7 +498,7 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 	int ret;
 
 	spin_lock(&bdev->fence_lock);
-	(void) ttm_bo_wait(bo, false, false, true);
+	(void) ttm_bo_wait(bo, false, false, true, TTM_USAGE_READWRITE);
 	if (!bo->sync_obj) {
 
 		spin_lock(&glob->lru_lock);
@@ -566,7 +566,8 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
 
 retry:
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0))
@@ -725,7 +726,8 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
 	int ret = 0;
 
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0)) {
@@ -1072,7 +1074,8 @@ int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 	 * instead of doing it here.
 	 */
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 	if (ret)
 		return ret;
@@ -1692,34 +1695,83 @@ out_unlock:
 	return ret;
 }
 
+static void ttm_bo_unref_sync_obj_locked(struct ttm_buffer_object *bo,
+					 void *sync_obj,
+					 void **extra_sync_obj)
+{
+	struct ttm_bo_device *bdev = bo->bdev;
+	struct ttm_bo_driver *driver = bdev->driver;
+	void *tmp_obj = NULL, *tmp_obj_read = NULL, *tmp_obj_write = NULL;
+
+	/* We must unref the sync obj wherever it's ref'd.
+	 * Note that if we unref bo->sync_obj, we can unref both the read
+	 * and write sync objs too, because they can't be newer than
+	 * bo->sync_obj, so they are no longer relevant. */
+	if (sync_obj == bo->sync_obj ||
+	    sync_obj == bo->sync_obj_read) {
+		tmp_obj_read = bo->sync_obj_read;
+		bo->sync_obj_read = NULL;
+	}
+	if (sync_obj == bo->sync_obj ||
+	    sync_obj == bo->sync_obj_write) {
+		tmp_obj_write = bo->sync_obj_write;
+		bo->sync_obj_write = NULL;
+	}
+	if (sync_obj == bo->sync_obj) {
+		tmp_obj = bo->sync_obj;
+		bo->sync_obj = NULL;
+	}
+
+	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
+	spin_unlock(&bdev->fence_lock);
+	if (tmp_obj)
+		driver->sync_obj_unref(&tmp_obj);
+	if (tmp_obj_read)
+		driver->sync_obj_unref(&tmp_obj_read);
+	if (tmp_obj_write)
+		driver->sync_obj_unref(&tmp_obj_write);
+	if (extra_sync_obj)
+		driver->sync_obj_unref(extra_sync_obj);
+	spin_lock(&bdev->fence_lock);
+}
+
 int ttm_bo_wait(struct ttm_buffer_object *bo,
-		bool lazy, bool interruptible, bool no_wait)
+		bool lazy, bool interruptible, bool no_wait,
+		enum ttm_buffer_usage usage)
 {
 	struct ttm_bo_driver *driver = bo->bdev->driver;
 	struct ttm_bo_device *bdev = bo->bdev;
 	void *sync_obj;
 	void *sync_obj_arg;
 	int ret = 0;
+	void **bo_sync_obj;
 
-	if (likely(bo->sync_obj == NULL))
+	switch (usage) {
+	case TTM_USAGE_READ:
+		bo_sync_obj = &bo->sync_obj_read;
+		break;
+	case TTM_USAGE_WRITE:
+		bo_sync_obj = &bo->sync_obj_write;
+		break;
+	case TTM_USAGE_READWRITE:
+	default:
+		bo_sync_obj = &bo->sync_obj;
+	}
+
+	if (likely(*bo_sync_obj == NULL))
 		return 0;
 
-	while (bo->sync_obj) {
+	while (*bo_sync_obj) {
 
-		if (driver->sync_obj_signaled(bo->sync_obj, bo->sync_obj_arg)) {
-			void *tmp_obj = bo->sync_obj;
-			bo->sync_obj = NULL;
-			clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-			spin_unlock(&bdev->fence_lock);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
+		if (driver->sync_obj_signaled(*bo_sync_obj, bo->sync_obj_arg)) {
+			ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, NULL);
 			continue;
 		}
 
 		if (no_wait)
 			return -EBUSY;
 
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
+		sync_obj = driver->sync_obj_ref(*bo_sync_obj);
 		sync_obj_arg = bo->sync_obj_arg;
 		spin_unlock(&bdev->fence_lock);
 		ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
@@ -1730,16 +1782,9 @@ int ttm_bo_wait(struct ttm_buffer_object *bo,
 			return ret;
 		}
 		spin_lock(&bdev->fence_lock);
-		if (likely(bo->sync_obj == sync_obj &&
+		if (likely(*bo_sync_obj == sync_obj &&
 			   bo->sync_obj_arg == sync_obj_arg)) {
-			void *tmp_obj = bo->sync_obj;
-			bo->sync_obj = NULL;
-			clear_bit(TTM_BO_PRIV_FLAG_MOVING,
-				  &bo->priv_flags);
-			spin_unlock(&bdev->fence_lock);
-			driver->sync_obj_unref(&sync_obj);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
+			ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, &sync_obj);
 		} else {
 			spin_unlock(&bdev->fence_lock);
 			driver->sync_obj_unref(&sync_obj);
@@ -1763,7 +1808,7 @@ int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
 	if (unlikely(ret != 0))
 		return ret;
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, true, no_wait);
+	ret = ttm_bo_wait(bo, false, true, no_wait, TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 	if (likely(ret == 0))
 		atomic_inc(&bo->cpu_writers);
@@ -1837,7 +1882,7 @@ static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
 	 */
 
 	spin_lock(&bo->bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, false, false);
+	ret = ttm_bo_wait(bo, false, false, false, TTM_USAGE_READWRITE);
 	spin_unlock(&bo->bdev->fence_lock);
 
 	if (unlikely(ret != 0))
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index ae3c6f5dd2b..6135f58169c 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -436,6 +436,8 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 	atomic_set(&fbo->cpu_writers, 0);
 
 	fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
+	fbo->sync_obj_read = driver->sync_obj_ref(bo->sync_obj_read);
+	fbo->sync_obj_write = driver->sync_obj_ref(bo->sync_obj_write);
 	kref_init(&fbo->list_kref);
 	kref_init(&fbo->kref);
 	fbo->destroy = &ttm_transfered_destroy;
@@ -618,20 +620,30 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 	struct ttm_mem_reg *old_mem = &bo->mem;
 	int ret;
 	struct ttm_buffer_object *ghost_obj;
-	void *tmp_obj = NULL;
+	void *tmp_obj = NULL, *tmp_obj_read = NULL, *tmp_obj_write = NULL;
 
 	spin_lock(&bdev->fence_lock);
-	if (bo->sync_obj) {
+	if (bo->sync_obj)
 		tmp_obj = bo->sync_obj;
-		bo->sync_obj = NULL;
-	}
+	if (bo->sync_obj_read)
+		tmp_obj_read = bo->sync_obj_read;
+	if (bo->sync_obj_write)
+		tmp_obj_write = bo->sync_obj_write;
+
 	bo->sync_obj = driver->sync_obj_ref(sync_obj);
+	bo->sync_obj_read = driver->sync_obj_ref(sync_obj);
+	bo->sync_obj_write = driver->sync_obj_ref(sync_obj);
 	bo->sync_obj_arg = sync_obj_arg;
 	if (evict) {
-		ret = ttm_bo_wait(bo, false, false, false);
+		ret = ttm_bo_wait(bo, false, false, false,
+				  TTM_USAGE_READWRITE);
 		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
+		if (tmp_obj_read)
+			driver->sync_obj_unref(&tmp_obj_read);
+		if (tmp_obj_write)
+			driver->sync_obj_unref(&tmp_obj_write);
 		if (ret)
 			return ret;
 
@@ -655,6 +667,10 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
+		if (tmp_obj_read)
+			driver->sync_obj_unref(&tmp_obj_read);
+		if (tmp_obj_write)
+			driver->sync_obj_unref(&tmp_obj_write);
 
 		ret = ttm_buffer_object_transfer(bo, &ghost_obj);
 		if (ret)
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
index 221b924aceb..ff1e26f4b09 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
@@ -122,7 +122,7 @@ static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 
 	spin_lock(&bdev->fence_lock);
 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
-		ret = ttm_bo_wait(bo, false, true, false);
+		ret = ttm_bo_wait(bo, false, true, false, TTM_USAGE_READWRITE);
 		spin_unlock(&bdev->fence_lock);
 		if (unlikely(ret != 0)) {
 			retval = (ret != -ERESTARTSYS) ?
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index 3832fe10b4d..36d111a8823 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -221,8 +221,18 @@ void ttm_eu_fence_buffer_objects(struct list_head *list, void *sync_obj)
 
 	list_for_each_entry(entry, list, head) {
 		bo = entry->bo;
+		entry->old_sync_obj_read = NULL;
+		entry->old_sync_obj_write = NULL;
 		entry->old_sync_obj = bo->sync_obj;
 		bo->sync_obj = driver->sync_obj_ref(sync_obj);
+		if (entry->usage & TTM_USAGE_READ) {
+			entry->old_sync_obj_read = bo->sync_obj_read;
+			bo->sync_obj_read = driver->sync_obj_ref(sync_obj);
+		}
+		if (entry->usage & TTM_USAGE_WRITE) {
+			entry->old_sync_obj_write = bo->sync_obj_write;
+			bo->sync_obj_write = driver->sync_obj_ref(sync_obj);
+		}
 		bo->sync_obj_arg = entry->new_sync_obj_arg;
 		ttm_bo_unreserve_locked(bo);
 		entry->reserved = false;
@@ -231,8 +241,15 @@ void ttm_eu_fence_buffer_objects(struct list_head *list, void *sync_obj)
 	spin_unlock(&bdev->fence_lock);
 
 	list_for_each_entry(entry, list, head) {
-		if (entry->old_sync_obj)
+		if (entry->old_sync_obj) {
 			driver->sync_obj_unref(&entry->old_sync_obj);
+		}
+		if (entry->old_sync_obj_read) {
+			driver->sync_obj_unref(&entry->old_sync_obj_read);
+		}
+		if (entry->old_sync_obj_write) {
+			driver->sync_obj_unref(&entry->old_sync_obj_write);
+		}
 	}
 }
 EXPORT_SYMBOL(ttm_eu_fence_buffer_objects);