[PATCH] drm/ttm: Fix two race conditions

[Thomas Hellstrom]

This fixes a race pointed out by Dave Airlie where we don't take a buffer
object about to be destroyed off the LRU lists properly. It also fixes a rare
case where a buffer object could be destroyed in the middle of an
accelerated eviction.

The patch also adds a utility function that can be used to prematurely
release GPU memory space usage of an object waiting to be destroyed.
For example during eviction or swapout.

Signed-off-by: Thomas Hellstrom <thellstrom at vmware.com>
---
 drivers/gpu/drm/ttm/ttm_bo.c |   74 +++++++++++++++++++++++++++++++++++------
 include/drm/ttm/ttm_bo_api.h |    4 ++-
 2 files changed, 66 insertions(+), 12 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index cb4cf7e..b98cb2c 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -442,6 +442,43 @@ out_err:
 }
 
 /**
+ * Call bo::reserved and with the lru lock held.
+ * Will release GPU memory type usage on destruction.
+ * This is the place to put in driver specific hooks.
+ * Will release the bo::reserved lock and the
+ * lru lock on exit.
+ */
+
+static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
+{
+	struct ttm_bo_global *glob = bo->glob;
+
+	if (bo->ttm) {
+
+		/**
+		 * Release the lru_lock, since we don't want to have
+		 * an atomic requirement on ttm_tt[unbind|destroy].
+		 */
+
+		spin_unlock(&glob->lru_lock);
+		ttm_tt_unbind(bo->ttm);
+		ttm_tt_destroy(bo->ttm);
+		bo->ttm = NULL;
+		spin_lock(&glob->lru_lock);
+	}
+
+	if (bo->mem.mm_node) {
+		drm_mm_put_block(bo->mem.mm_node);
+		bo->mem.mm_node = NULL;
+	}
+
+	atomic_set(&bo->reserved, 0);
+	wake_up_all(&bo->event_queue);
+	spin_unlock(&glob->lru_lock);
+}
+
+
+/**
  * If bo idle, remove from delayed- and lru lists, and unref.
  * If not idle, and already on delayed list, do nothing.
  * If not idle, and not on delayed list, put on delayed list,
@@ -464,24 +501,39 @@ static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, bool remove_all)
 		spin_unlock(&bo->lock);
 
 		spin_lock(&glob->lru_lock);
-		put_count = ttm_bo_del_from_lru(bo);
+		ret = ttm_bo_reserve_locked(bo, false, true, false, 0);
+
+		/**
+		 * Someone else has the object reserved. Bail and retry.
+		 */
+
+		if (unlikely(ret != 0)) {
+			spin_unlock(&glob->lru_lock);
+			return 0;
+		}
 
-		ret = ttm_bo_reserve_locked(bo, false, false, false, 0);
-		BUG_ON(ret);
-		if (bo->ttm)
-			ttm_tt_unbind(bo->ttm);
+		/**
+		 * We can re-check for sync object without taking
+		 * the bo::lock since setting the sync object requires
+		 * also bo::reserved. A busy object at this point may
+		 * be caused by another thread starting an accelerated
+		 * eviction.
+		 */
+
+		if (unlikely(bo->sync_obj)) {
+			atomic_set(&bo->reserved, 0);
+			wake_up_all(&bo->event_queue);
+			spin_unlock(&glob->lru_lock);
+		}
+
+		put_count = ttm_bo_del_from_lru(bo);
 
 		if (!list_empty(&bo->ddestroy)) {
 			list_del_init(&bo->ddestroy);
 			++put_count;
 		}
-		if (bo->mem.mm_node) {
-			drm_mm_put_block(bo->mem.mm_node);
-			bo->mem.mm_node = NULL;
-		}
-		spin_unlock(&glob->lru_lock);
 
-		atomic_set(&bo->reserved, 0);
+		ttm_bo_cleanup_memtype_use(bo);
 
 		while (put_count--)
 			kref_put(&bo->list_kref, ttm_bo_ref_bug);
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 267a86c..2040e6c 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -246,9 +246,11 @@ struct ttm_buffer_object {
 
 	atomic_t reserved;
 
-
 	/**
 	 * Members protected by the bo::lock
+	 * In addition, setting sync_obj to anything else
+	 * than NULL requires bo::reserved to be held. This allows for
+	 * checking NULL while reserved but not holding bo::lock.
 	 */
 
 	void *sync_obj_arg;
-- 
1.6.2.5