Commit dfadbbdb authored by Marek Olšák's avatar Marek Olšák Committed by Dave Airlie
Browse files

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: default avatarMarek Olšák <maraeo@gmail.com>
Reviewed-by: default avatarJerome Glisse <jglisse@redhat.com>
Signed-off-by: default avatarDave Airlie <airlied@redhat.com>
parent b464e9a2

drivers/gpu/drm/nouveau/nouveau_bo.c

+2 −1
Original line number Diff line number Diff line
@@ -1104,7 +1104,8 @@ nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma) 
	if (vma->node) {

		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {

			spin_lock(&nvbo->bo.bdev->fence_lock);

			ttm_bo_wait(&nvbo->bo, false, false, false);

			ttm_bo_wait(&nvbo->bo, false, false, false,

				    TTM_USAGE_READWRITE);

			spin_unlock(&nvbo->bo.bdev->fence_lock);

			nouveau_vm_unmap(vma);

		}

drivers/gpu/drm/nouveau/nouveau_gem.c

+3 −2
Original line number Diff line number Diff line
@@ -589,7 +589,8 @@ nouveau_gem_pushbuf_reloc_apply(struct drm_device *dev, 
		}



		spin_lock(&nvbo->bo.bdev->fence_lock);

		ret = ttm_bo_wait(&nvbo->bo, false, false, false);

		ret = ttm_bo_wait(&nvbo->bo, false, false, false,

				  TTM_USAGE_READWRITE);

		spin_unlock(&nvbo->bo.bdev->fence_lock);

		if (ret) {

			NV_ERROR(dev, "reloc wait_idle failed: %d\n", ret);
@@ -825,7 +826,7 @@ nouveau_gem_ioctl_cpu_prep(struct drm_device *dev, void *data, 
	nvbo = nouveau_gem_object(gem);



	spin_lock(&nvbo->bo.bdev->fence_lock);

	ret = ttm_bo_wait(&nvbo->bo, true, true, no_wait);

	ret = ttm_bo_wait(&nvbo->bo, true, true, no_wait, TTM_USAGE_READWRITE);

	spin_unlock(&nvbo->bo.bdev->fence_lock);

	drm_gem_object_unreference_unlocked(gem);

	return ret;

drivers/gpu/drm/radeon/radeon_cs.c

+1 −0
Original line number Diff line number Diff line
@@ -80,6 +80,7 @@ int radeon_cs_parser_relocs(struct radeon_cs_parser *p) 
			p->relocs[i].lobj.wdomain = r->write_domain;

			p->relocs[i].lobj.rdomain = r->read_domains;

			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;

			p->relocs[i].lobj.tv.usage = TTM_USAGE_READWRITE;

			p->relocs[i].handle = r->handle;

			p->relocs[i].flags = r->flags;

			radeon_bo_list_add_object(&p->relocs[i].lobj,

drivers/gpu/drm/radeon/radeon_object.h

+1 −1
Original line number Diff line number Diff line
@@ -130,7 +130,7 @@ static inline int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, 
	if (mem_type)

		*mem_type = bo->tbo.mem.mem_type;

	if (bo->tbo.sync_obj)

		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);

		r = ttm_bo_wait(&bo->tbo, true, true, no_wait, TTM_USAGE_READWRITE);

	spin_unlock(&bo->tbo.bdev->fence_lock);

	ttm_bo_unreserve(&bo->tbo);

	return r;

drivers/gpu/drm/ttm/ttm_bo.c

+71 −26
Original line number Diff line number Diff line
@@ -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))

CRA Git | Maintained and supported by SUSTech CRA and CCSE