drm/ttm: Wait for a BO to become idle before unbinding it from GTT

Fixes hangs under memory pressure, e.g. running the piglit test
tex3d-maxsize concurrently with other tests.

Fixes: 17d33bc9d6ef ("drm/ttm: drop waiting for idle in ttm_bo_evict.")
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

6 files changed, 19 insertions, 9 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 819c62d27bf8..105bd229fdad 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -336,7 +336,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
-	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
+	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, new_mem);
 out_cleanup:
 	ttm_bo_mem_put(bo, &tmp_mem);
 	return r;
@@ -369,7 +369,7 @@ static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 	if (unlikely(r)) {
 		return r;
 	}
-	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
+	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, &tmp_mem);
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 528bdeffb339..6190035edfea 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -1151,7 +1151,7 @@ nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
 	if (ret)
 		goto out;
 
-	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
+	ret = ttm_bo_move_ttm(bo, true, intr, no_wait_gpu, new_mem);
 out:
 	ttm_bo_mem_put(bo, &tmp_mem);
 	return ret;
@@ -1179,7 +1179,7 @@ nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
 	if (ret)
 		return ret;
 
-	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
+	ret = ttm_bo_move_ttm(bo, true, intr, no_wait_gpu, &tmp_mem);
 	if (ret)
 		goto out;
 
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index ffdad81ef964..0c00e192c845 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -346,7 +346,7 @@ static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
-	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
+	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, new_mem);
 out_cleanup:
 	ttm_bo_mem_put(bo, &tmp_mem);
 	return r;
@@ -379,7 +379,7 @@ static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
 	if (unlikely(r)) {
 		return r;
 	}
-	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
+	r = ttm_bo_move_ttm(bo, true, interruptible, no_wait_gpu, &tmp_mem);
 	if (unlikely(r)) {
 		goto out_cleanup;
 	}
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 4054d804fe06..42c074a9c955 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -354,7 +354,8 @@ static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
 
 	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
 	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
-		ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
+		ret = ttm_bo_move_ttm(bo, evict, interruptible, no_wait_gpu,
+				      mem);
 	else if (bdev->driver->move)
 		ret = bdev->driver->move(bo, evict, interruptible,
 					 no_wait_gpu, mem);
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index 2df602a35f92..f157a9efd220 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -45,7 +45,7 @@ void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
 }
 
 int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
-		    bool evict,
+		    bool evict, bool interruptible,
 		    bool no_wait_gpu, struct ttm_mem_reg *new_mem)
 {
 	struct ttm_tt *ttm = bo->ttm;
@@ -53,6 +53,14 @@ int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
 	int ret;
 
 	if (old_mem->mem_type != TTM_PL_SYSTEM) {
+		ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
+
+		if (unlikely(ret != 0)) {
+			if (ret != -ERESTARTSYS)
+				pr_err("Failed to expire sync object before unbinding TTM\n");
+			return ret;
+		}
+
 		ttm_tt_unbind(ttm);
 		ttm_bo_free_old_node(bo);
 		ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM,
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 4348d6d5877a..99c6d01d24f2 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -962,6 +962,7 @@ void ttm_mem_io_free(struct ttm_bo_device *bdev,
  *
  * @bo: A pointer to a struct ttm_buffer_object.
  * @evict: 1: This is an eviction. Don't try to pipeline.
+ * @interruptible: Sleep interruptible if waiting.
  * @no_wait_gpu: Return immediately if the GPU is busy.
  * @new_mem: struct ttm_mem_reg indicating where to move.
  *
@@ -976,7 +977,7 @@ void ttm_mem_io_free(struct ttm_bo_device *bdev,
  */
 
 extern int ttm_bo_move_ttm(struct ttm_buffer_object *bo,
-			   bool evict, bool no_wait_gpu,
+			   bool evict, bool interruptible, bool no_wait_gpu,
 			   struct ttm_mem_reg *new_mem);
 
 /**