1 files changed, 20 insertions, 40 deletions

diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index 0b347f46b2f4..c8c1e3ae8d84 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -40,9 +40,6 @@
 
 struct workqueue_struct *rds_ib_mr_wq;
 
-static DEFINE_PER_CPU(unsigned long, clean_list_grace);
-#define CLEAN_LIST_BUSY_BIT 0
-
 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 {
 	struct rds_ib_device *rds_ibdev;
@@ -195,12 +192,11 @@ struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 {
 	struct rds_ib_mr *ibmr = NULL;
 	struct llist_node *ret;
-	unsigned long *flag;
+	unsigned long flags;
 
-	preempt_disable();
-	flag = this_cpu_ptr(&clean_list_grace);
-	set_bit(CLEAN_LIST_BUSY_BIT, flag);
+	spin_lock_irqsave(&pool->clean_lock, flags);
 	ret = llist_del_first(&pool->clean_list);
+	spin_unlock_irqrestore(&pool->clean_lock, flags);
 	if (ret) {
 		ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
 		if (pool->pool_type == RDS_IB_MR_8K_POOL)
@@ -209,23 +205,9 @@ struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
 			rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
 	}
 
-	clear_bit(CLEAN_LIST_BUSY_BIT, flag);
-	preempt_enable();
 	return ibmr;
 }
 
-static inline void wait_clean_list_grace(void)
-{
-	int cpu;
-	unsigned long *flag;
-
-	for_each_online_cpu(cpu) {
-		flag = &per_cpu(clean_list_grace, cpu);
-		while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
-			cpu_relax();
-	}
-}
-
 void rds_ib_sync_mr(void *trans_private, int direction)
 {
 	struct rds_ib_mr *ibmr = trans_private;
@@ -324,8 +306,7 @@ static unsigned int llist_append_to_list(struct llist_head *llist,
  * of clusters.  Each cluster has linked llist nodes of
  * MR_CLUSTER_SIZE mrs that are ready for reuse.
  */
-static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
-				struct list_head *list,
+static void list_to_llist_nodes(struct list_head *list,
 				struct llist_node **nodes_head,
 				struct llist_node **nodes_tail)
 {
@@ -402,8 +383,13 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	 */
 	dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
 	dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
-	if (free_all)
+	if (free_all) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&pool->clean_lock, flags);
 		llist_append_to_list(&pool->clean_list, &unmap_list);
+		spin_unlock_irqrestore(&pool->clean_lock, flags);
+	}
 
 	free_goal = rds_ib_flush_goal(pool, free_all);
 
@@ -416,27 +402,20 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 		rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
 
 	if (!list_empty(&unmap_list)) {
-		/* we have to make sure that none of the things we're about
-		 * to put on the clean list would race with other cpus trying
-		 * to pull items off.  The llist would explode if we managed to
-		 * remove something from the clean list and then add it back again
-		 * while another CPU was spinning on that same item in llist_del_first.
-		 *
-		 * This is pretty unlikely, but just in case  wait for an llist grace period
-		 * here before adding anything back into the clean list.
-		 */
-		wait_clean_list_grace();
-
-		list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
+		unsigned long flags;
+
+		list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
 		if (ibmr_ret) {
 			*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
 			clean_nodes = clean_nodes->next;
 		}
 		/* more than one entry in llist nodes */
-		if (clean_nodes)
+		if (clean_nodes) {
+			spin_lock_irqsave(&pool->clean_lock, flags);
 			llist_add_batch(clean_nodes, clean_tail,
 					&pool->clean_list);
-
+			spin_unlock_irqrestore(&pool->clean_lock, flags);
+		}
 	}
 
 	atomic_sub(unpinned, &pool->free_pinned);
@@ -471,7 +450,7 @@ struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 				rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
 			else
 				rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
-			return ERR_PTR(-EAGAIN);
+			break;
 		}
 
 		/* We do have some empty MRs. Flush them out. */
@@ -485,7 +464,7 @@ struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
 			return ibmr;
 	}
 
-	return ibmr;
+	return NULL;
 }
 
 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
@@ -610,6 +589,7 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
 	init_llist_head(&pool->free_list);
 	init_llist_head(&pool->drop_list);
 	init_llist_head(&pool->clean_list);
+	spin_lock_init(&pool->clean_lock);
 	mutex_init(&pool->flush_lock);
 	init_waitqueue_head(&pool->flush_wait);
 	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);