drm/ttm: Add a bo list reserve fastpath (v2)

Makes it possible to reserve a list of buffer objects with a single
spin lock / unlock if there is no contention.
Should improve cpu usage on SMP kernels.

v2: Initialize private list members on reserve and don't call
ttm_bo_list_ref_sub() with zero put_count.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 148a322..a586378 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -169,7 +169,7 @@
 }
 EXPORT_SYMBOL(ttm_bo_wait_unreserved);
 
-static void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
+void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
 	struct ttm_mem_type_manager *man;
@@ -191,11 +191,7 @@
 	}
 }
 
-/**
- * Call with the lru_lock held.
- */
-
-static int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
+int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
 {
 	int put_count = 0;
 
@@ -267,6 +263,15 @@
 	BUG();
 }
 
+void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
+			 bool never_free)
+{
+	while (count--)
+		kref_put(&bo->list_kref,
+			 (never_free || (count >= 0)) ? ttm_bo_ref_bug :
+			 ttm_bo_release_list);
+}
+
 int ttm_bo_reserve(struct ttm_buffer_object *bo,
 		   bool interruptible,
 		   bool no_wait, bool use_sequence, uint32_t sequence)
@@ -282,8 +287,7 @@
 		put_count = ttm_bo_del_from_lru(bo);
 	spin_unlock(&glob->lru_lock);
 
-	while (put_count--)
-		kref_put(&bo->list_kref, ttm_bo_ref_bug);
+	ttm_bo_list_ref_sub(bo, put_count, true);
 
 	return ret;
 }
@@ -496,8 +500,7 @@
 		spin_unlock(&glob->lru_lock);
 		ttm_bo_cleanup_memtype_use(bo);
 
-		while (put_count--)
-			kref_put(&bo->list_kref, ttm_bo_ref_bug);
+		ttm_bo_list_ref_sub(bo, put_count, true);
 
 		return;
 	} else {
@@ -580,8 +583,7 @@
 	spin_unlock(&glob->lru_lock);
 	ttm_bo_cleanup_memtype_use(bo);
 
-	while (put_count--)
-		kref_put(&bo->list_kref, ttm_bo_ref_bug);
+	ttm_bo_list_ref_sub(bo, put_count, true);
 
 	return 0;
 }
@@ -802,8 +804,7 @@
 
 	BUG_ON(ret != 0);
 
-	while (put_count--)
-		kref_put(&bo->list_kref, ttm_bo_ref_bug);
+	ttm_bo_list_ref_sub(bo, put_count, true);
 
 	ret = ttm_bo_evict(bo, interruptible, no_wait_reserve, no_wait_gpu);
 	ttm_bo_unreserve(bo);
@@ -1783,8 +1784,7 @@
 	put_count = ttm_bo_del_from_lru(bo);
 	spin_unlock(&glob->lru_lock);
 
-	while (put_count--)
-		kref_put(&bo->list_kref, ttm_bo_ref_bug);
+	ttm_bo_list_ref_sub(bo, put_count, true);
 
 	/**
 	 * Wait for GPU, then move to system cached.
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index c285c29..201a71d 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -32,6 +32,72 @@
 #include <linux/sched.h>
 #include <linux/module.h>
 
+static void ttm_eu_backoff_reservation_locked(struct list_head *list)
+{
+	struct ttm_validate_buffer *entry;
+
+	list_for_each_entry(entry, list, head) {
+		struct ttm_buffer_object *bo = entry->bo;
+		if (!entry->reserved)
+			continue;
+
+		if (entry->removed) {
+			ttm_bo_add_to_lru(bo);
+			entry->removed = false;
+
+		}
+		entry->reserved = false;
+		atomic_set(&bo->reserved, 0);
+		wake_up_all(&bo->event_queue);
+	}
+}
+
+static void ttm_eu_del_from_lru_locked(struct list_head *list)
+{
+	struct ttm_validate_buffer *entry;
+
+	list_for_each_entry(entry, list, head) {
+		struct ttm_buffer_object *bo = entry->bo;
+		if (!entry->reserved)
+			continue;
+
+		if (!entry->removed) {
+			entry->put_count = ttm_bo_del_from_lru(bo);
+			entry->removed = true;
+		}
+	}
+}
+
+static void ttm_eu_list_ref_sub(struct list_head *list)
+{
+	struct ttm_validate_buffer *entry;
+
+	list_for_each_entry(entry, list, head) {
+		struct ttm_buffer_object *bo = entry->bo;
+
+		if (entry->put_count) {
+			ttm_bo_list_ref_sub(bo, entry->put_count, true);
+			entry->put_count = 0;
+		}
+	}
+}
+
+static int ttm_eu_wait_unreserved_locked(struct list_head *list,
+					 struct ttm_buffer_object *bo)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	int ret;
+
+	ttm_eu_del_from_lru_locked(list);
+	spin_unlock(&glob->lru_lock);
+	ret = ttm_bo_wait_unreserved(bo, true);
+	spin_lock(&glob->lru_lock);
+	if (unlikely(ret != 0))
+		ttm_eu_backoff_reservation_locked(list);
+	return ret;
+}
+
+
 void ttm_eu_backoff_reservation(struct list_head *list)
 {
 	struct ttm_validate_buffer *entry;
@@ -61,35 +127,71 @@
 
 int ttm_eu_reserve_buffers(struct list_head *list, uint32_t val_seq)
 {
+	struct ttm_bo_global *glob;
 	struct ttm_validate_buffer *entry;
 	int ret;
 
+	if (list_empty(list))
+		return 0;
+
+	list_for_each_entry(entry, list, head) {
+		entry->reserved = false;
+		entry->put_count = 0;
+		entry->removed = false;
+	}
+
+	entry = list_first_entry(list, struct ttm_validate_buffer, head);
+	glob = entry->bo->glob;
+
 retry:
+	spin_lock(&glob->lru_lock);
 	list_for_each_entry(entry, list, head) {
 		struct ttm_buffer_object *bo = entry->bo;
 
-		entry->reserved = false;
-		ret = ttm_bo_reserve(bo, true, false, true, val_seq);
-		if (ret != 0) {
-			ttm_eu_backoff_reservation(list);
-			if (ret == -EAGAIN) {
-				ret = ttm_bo_wait_unreserved(bo, true);
-				if (unlikely(ret != 0))
-					return ret;
-				goto retry;
-			} else
+retry_this_bo:
+		ret = ttm_bo_reserve_locked(bo, true, true, true, val_seq);
+		switch (ret) {
+		case 0:
+			break;
+		case -EBUSY:
+			ret = ttm_eu_wait_unreserved_locked(list, bo);
+			if (unlikely(ret != 0)) {
+				spin_unlock(&glob->lru_lock);
+				ttm_eu_list_ref_sub(list);
 				return ret;
+			}
+			goto retry_this_bo;
+		case -EAGAIN:
+			ttm_eu_backoff_reservation_locked(list);
+			spin_unlock(&glob->lru_lock);
+			ttm_eu_list_ref_sub(list);
+			ret = ttm_bo_wait_unreserved(bo, true);
+			if (unlikely(ret != 0))
+				return ret;
+			goto retry;
+		default:
+			ttm_eu_backoff_reservation_locked(list);
+			spin_unlock(&glob->lru_lock);
+			ttm_eu_list_ref_sub(list);
+			return ret;
 		}
 
 		entry->reserved = true;
 		if (unlikely(atomic_read(&bo->cpu_writers) > 0)) {
-			ttm_eu_backoff_reservation(list);
+			ttm_eu_backoff_reservation_locked(list);
+			spin_unlock(&glob->lru_lock);
+			ttm_eu_list_ref_sub(list);
 			ret = ttm_bo_wait_cpu(bo, false);
 			if (ret)
 				return ret;
 			goto retry;
 		}
 	}
+
+	ttm_eu_del_from_lru_locked(list);
+	spin_unlock(&glob->lru_lock);
+	ttm_eu_list_ref_sub(list);
+
 	return 0;
 }
 EXPORT_SYMBOL(ttm_eu_reserve_buffers);
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index beafc15..b0fc9c1 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -364,6 +364,44 @@
  */
 extern void ttm_bo_unref(struct ttm_buffer_object **bo);
 
+
+/**
+ * ttm_bo_list_ref_sub
+ *
+ * @bo: The buffer object.
+ * @count: The number of references with which to decrease @bo::list_kref;
+ * @never_free: The refcount should not reach zero with this operation.
+ *
+ * Release @count lru list references to this buffer object.
+ */
+extern void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
+				bool never_free);
+
+/**
+ * ttm_bo_add_to_lru
+ *
+ * @bo: The buffer object.
+ *
+ * Add this bo to the relevant mem type lru and, if it's backed by
+ * system pages (ttms) to the swap list.
+ * This function must be called with struct ttm_bo_global::lru_lock held, and
+ * is typically called immediately prior to unreserving a bo.
+ */
+extern void ttm_bo_add_to_lru(struct ttm_buffer_object *bo);
+
+/**
+ * ttm_bo_del_from_lru
+ *
+ * @bo: The buffer object.
+ *
+ * Remove this bo from all lru lists used to lookup and reserve an object.
+ * This function must be called with struct ttm_bo_global::lru_lock held,
+ * and is usually called just immediately after the bo has been reserved to
+ * avoid recursive reservation from lru lists.
+ */
+extern int ttm_bo_del_from_lru(struct ttm_buffer_object *bo);
+
+
 /**
  * ttm_bo_lock_delayed_workqueue
  *
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 8e0c848..95068e6 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -864,6 +864,20 @@
 			  bool interruptible,
 			  bool no_wait, bool use_sequence, uint32_t sequence);
 
+
+/**
+ * ttm_bo_reserve_locked:
+ *
+ * Similar to ttm_bo_reserve, but must be called with the glob::lru_lock
+ * spinlock held, and will not remove reserved buffers from the lru lists.
+ * The function may release the LRU spinlock if it needs to sleep.
+ */
+
+extern int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
+				 bool interruptible,
+				 bool no_wait, bool use_sequence,
+				 uint32_t sequence);
+
 /**
  * ttm_bo_unreserve
  *
diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h
index cd2c475..fd09b84 100644
--- a/include/drm/ttm/ttm_execbuf_util.h
+++ b/include/drm/ttm/ttm_execbuf_util.h
@@ -41,7 +41,9 @@
  * @bo:             refcounted buffer object pointer.
  * @new_sync_obj_arg: New sync_obj_arg for @bo, to be used once
  * adding a new sync object.
- * @reservied:      Indicates whether @bo has been reserved for validation.
+ * @reserved:       Indicates whether @bo has been reserved for validation.
+ * @removed:        Indicates whether @bo has been removed from lru lists.
+ * @put_count:      Number of outstanding references on bo::list_kref.
  */
 
 struct ttm_validate_buffer {
@@ -49,6 +51,8 @@
 	struct ttm_buffer_object *bo;
 	void *new_sync_obj_arg;
 	bool reserved;
+	bool removed;
+	int put_count;
 };
 
 /**