1 files changed, 118 insertions, 53 deletions

diff --git a/kernel/net/rds/ib_rdma.c b/kernel/net/rds/ib_rdma.c
index c8faaf364..19123a97b 100644
--- a/kernel/net/rds/ib_rdma.c
+++ b/kernel/net/rds/ib_rdma.c
@@ -66,6 +66,7 @@ struct rds_ib_mr {
  * Our own little FMR pool
  */
 struct rds_ib_mr_pool {
+	unsigned int            pool_type;
 	struct mutex		flush_lock;		/* serialize fmr invalidate */
 	struct delayed_work	flush_worker;		/* flush worker */
 
@@ -84,6 +85,25 @@ struct rds_ib_mr_pool {
 	struct ib_fmr_attr	fmr_attr;
 };
 
+static struct workqueue_struct *rds_ib_fmr_wq;
+
+int rds_ib_fmr_init(void)
+{
+	rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
+	if (!rds_ib_fmr_wq)
+		return -ENOMEM;
+	return 0;
+}
+
+/* By the time this is called all the IB devices should have been torn down and
+ * had their pools freed.  As each pool is freed its work struct is waited on,
+ * so the pool flushing work queue should be idle by the time we get here.
+ */
+void rds_ib_fmr_exit(void)
+{
+	destroy_workqueue(rds_ib_fmr_wq);
+}
+
 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
 static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
 static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
@@ -141,10 +161,8 @@ static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 	}
 	spin_unlock_irq(&rds_ibdev->spinlock);
 
-	if (to_free) {
-		synchronize_rcu();
-		kfree(to_free);
-	}
+	if (to_free)
+		kfree_rcu(to_free, rcu);
 }
 
 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
@@ -152,12 +170,17 @@ int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 	struct rds_ib_device *rds_ibdev_old;
 
 	rds_ibdev_old = rds_ib_get_device(ipaddr);
-	if (rds_ibdev_old) {
+	if (!rds_ibdev_old)
+		return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
+
+	if (rds_ibdev_old != rds_ibdev) {
 		rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
 		rds_ib_dev_put(rds_ibdev_old);
+		return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 	}
+	rds_ib_dev_put(rds_ibdev_old);
 
-	return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
+	return 0;
 }
 
 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
@@ -213,7 +236,8 @@ void rds_ib_destroy_nodev_conns(void)
 		rds_conn_destroy(ic->conn);
 }
 
-struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
+struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
+					     int pool_type)
 {
 	struct rds_ib_mr_pool *pool;
 
@@ -221,6 +245,7 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 	if (!pool)
 		return ERR_PTR(-ENOMEM);
 
+	pool->pool_type = pool_type;
 	init_llist_head(&pool->free_list);
 	init_llist_head(&pool->drop_list);
 	init_llist_head(&pool->clean_list);
@@ -228,28 +253,30 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 	init_waitqueue_head(&pool->flush_wait);
 	INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
 
-	pool->fmr_attr.max_pages = fmr_message_size;
+	if (pool_type == RDS_IB_MR_1M_POOL) {
+		/* +1 allows for unaligned MRs */
+		pool->fmr_attr.max_pages = RDS_FMR_1M_MSG_SIZE + 1;
+		pool->max_items = RDS_FMR_1M_POOL_SIZE;
+	} else {
+		/* pool_type == RDS_IB_MR_8K_POOL */
+		pool->fmr_attr.max_pages = RDS_FMR_8K_MSG_SIZE + 1;
+		pool->max_items = RDS_FMR_8K_POOL_SIZE;
+	}
+
+	pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
 	pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
 	pool->fmr_attr.page_shift = PAGE_SHIFT;
-	pool->max_free_pinned = rds_ibdev->max_fmrs * fmr_message_size / 4;
-
-	/* We never allow more than max_items MRs to be allocated.
-	 * When we exceed more than max_items_soft, we start freeing
-	 * items more aggressively.
-	 * Make sure that max_items > max_items_soft > max_items / 2
-	 */
 	pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
-	pool->max_items = rds_ibdev->max_fmrs;
 
 	return pool;
 }
 
 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
 {
-	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
+	struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
 
-	iinfo->rdma_mr_max = pool->max_items;
-	iinfo->rdma_mr_size = pool->fmr_attr.max_pages;
+	iinfo->rdma_mr_max = pool_1m->max_items;
+	iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
 }
 
 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
@@ -291,14 +318,28 @@ static inline void wait_clean_list_grace(void)
 	}
 }
 
-static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
+static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev,
+					  int npages)
 {
-	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
+	struct rds_ib_mr_pool *pool;
 	struct rds_ib_mr *ibmr = NULL;
 	int err = 0, iter = 0;
 
+	if (npages <= RDS_FMR_8K_MSG_SIZE)
+		pool = rds_ibdev->mr_8k_pool;
+	else
+		pool = rds_ibdev->mr_1m_pool;
+
 	if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
-		schedule_delayed_work(&pool->flush_worker, 10);
+		queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
+
+	/* Switch pools if one of the pool is reaching upper limit */
+	if (atomic_read(&pool->dirty_count) >=  pool->max_items * 9 / 10) {
+		if (pool->pool_type == RDS_IB_MR_8K_POOL)
+			pool = rds_ibdev->mr_1m_pool;
+		else
+			pool = rds_ibdev->mr_8k_pool;
+	}
 
 	while (1) {
 		ibmr = rds_ib_reuse_fmr(pool);
@@ -320,12 +361,18 @@ static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 		atomic_dec(&pool->item_count);
 
 		if (++iter > 2) {
-			rds_ib_stats_inc(s_ib_rdma_mr_pool_depleted);
+			if (pool->pool_type == RDS_IB_MR_8K_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);
 		}
 
 		/* We do have some empty MRs. Flush them out. */
-		rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
+		if (pool->pool_type == RDS_IB_MR_8K_POOL)
+			rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
+		else
+			rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
 		rds_ib_flush_mr_pool(pool, 0, &ibmr);
 		if (ibmr)
 			return ibmr;
@@ -337,8 +384,6 @@ static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 		goto out_no_cigar;
 	}
 
-	memset(ibmr, 0, sizeof(*ibmr));
-
 	ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
 			(IB_ACCESS_LOCAL_WRITE |
 			 IB_ACCESS_REMOTE_READ |
@@ -352,7 +397,12 @@ static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 		goto out_no_cigar;
 	}
 
-	rds_ib_stats_inc(s_ib_rdma_mr_alloc);
+	ibmr->pool = pool;
+	if (pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
+
 	return ibmr;
 
 out_no_cigar:
@@ -408,7 +458,7 @@ static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibm
 	}
 
 	page_cnt += len >> PAGE_SHIFT;
-	if (page_cnt > fmr_message_size)
+	if (page_cnt > ibmr->pool->fmr_attr.max_pages)
 		return -EINVAL;
 
 	dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
@@ -440,7 +490,10 @@ static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibm
 	ibmr->sg_dma_len = sg_dma_len;
 	ibmr->remap_count++;
 
-	rds_ib_stats_inc(s_ib_rdma_mr_used);
+	if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
 	ret = 0;
 
 out:
@@ -486,7 +539,7 @@ static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 
 			/* FIXME we need a way to tell a r/w MR
 			 * from a r/o MR */
-			BUG_ON(irqs_disabled());
+			WARN_ON(!page->mapping && irqs_disabled());
 			set_page_dirty(page);
 			put_page(page);
 		}
@@ -503,8 +556,7 @@ static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
 
 	__rds_ib_teardown_mr(ibmr);
 	if (pinned) {
-		struct rds_ib_device *rds_ibdev = ibmr->device;
-		struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
+		struct rds_ib_mr_pool *pool = ibmr->pool;
 
 		atomic_sub(pinned, &pool->free_pinned);
 	}
@@ -524,11 +576,13 @@ static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int fr
 /*
  * given an llist of mrs, put them all into the list_head for more processing
  */
-static void llist_append_to_list(struct llist_head *llist, struct list_head *list)
+static unsigned int llist_append_to_list(struct llist_head *llist,
+					 struct list_head *list)
 {
 	struct rds_ib_mr *ibmr;
 	struct llist_node *node;
 	struct llist_node *next;
+	unsigned int count = 0;
 
 	node = llist_del_all(llist);
 	while (node) {
@@ -536,7 +590,9 @@ static void llist_append_to_list(struct llist_head *llist, struct list_head *lis
 		ibmr = llist_entry(node, struct rds_ib_mr, llnode);
 		list_add_tail(&ibmr->unmap_list, list);
 		node = next;
+		count++;
 	}
+	return count;
 }
 
 /*
@@ -569,7 +625,7 @@ static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
  * to free as many MRs as needed to get back to this limit.
  */
 static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
-			        int free_all, struct rds_ib_mr **ibmr_ret)
+				int free_all, struct rds_ib_mr **ibmr_ret)
 {
 	struct rds_ib_mr *ibmr, *next;
 	struct llist_node *clean_nodes;
@@ -577,14 +633,17 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	LIST_HEAD(unmap_list);
 	LIST_HEAD(fmr_list);
 	unsigned long unpinned = 0;
-	unsigned int nfreed = 0, ncleaned = 0, free_goal;
+	unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
 	int ret = 0;
 
-	rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
+	if (pool->pool_type == RDS_IB_MR_8K_POOL)
+		rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
+	else
+		rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
 
 	if (ibmr_ret) {
 		DEFINE_WAIT(wait);
-		while(!mutex_trylock(&pool->flush_lock)) {
+		while (!mutex_trylock(&pool->flush_lock)) {
 			ibmr = rds_ib_reuse_fmr(pool);
 			if (ibmr) {
 				*ibmr_ret = ibmr;
@@ -619,8 +678,8 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	/* Get the list of all MRs to be dropped. Ordering matters -
 	 * we want to put drop_list ahead of free_list.
 	 */
-	llist_append_to_list(&pool->drop_list, &unmap_list);
-	llist_append_to_list(&pool->free_list, &unmap_list);
+	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)
 		llist_append_to_list(&pool->clean_list, &unmap_list);
 
@@ -641,14 +700,17 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
 		unpinned += ibmr->sg_len;
 		__rds_ib_teardown_mr(ibmr);
-		if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
-			rds_ib_stats_inc(s_ib_rdma_mr_free);
+		if (nfreed < free_goal ||
+		    ibmr->remap_count >= pool->fmr_attr.max_maps) {
+			if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+				rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
+			else
+				rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
 			list_del(&ibmr->unmap_list);
 			ib_dealloc_fmr(ibmr->fmr);
 			kfree(ibmr);
 			nfreed++;
 		}
-		ncleaned++;
 	}
 
 	if (!list_empty(&unmap_list)) {
@@ -674,7 +736,7 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
 	}
 
 	atomic_sub(unpinned, &pool->free_pinned);
-	atomic_sub(ncleaned, &pool->dirty_count);
+	atomic_sub(dirty_to_clean, &pool->dirty_count);
 	atomic_sub(nfreed, &pool->item_count);
 
 out:
@@ -695,8 +757,8 @@ static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 void rds_ib_free_mr(void *trans_private, int invalidate)
 {
 	struct rds_ib_mr *ibmr = trans_private;
+	struct rds_ib_mr_pool *pool = ibmr->pool;
 	struct rds_ib_device *rds_ibdev = ibmr->device;
-	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 
 	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 
@@ -711,16 +773,18 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
 
 	/* If we've pinned too many pages, request a flush */
 	if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
-	    atomic_read(&pool->dirty_count) >= pool->max_items / 10)
-		schedule_delayed_work(&pool->flush_worker, 10);
+	    atomic_read(&pool->dirty_count) >= pool->max_items / 5)
+		queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 
 	if (invalidate) {
 		if (likely(!in_interrupt())) {
 			rds_ib_flush_mr_pool(pool, 0, NULL);
 		} else {
 			/* We get here if the user created a MR marked
-			 * as use_once and invalidate at the same time. */
-			schedule_delayed_work(&pool->flush_worker, 10);
+			 * as use_once and invalidate at the same time.
+			 */
+			queue_delayed_work(rds_ib_fmr_wq,
+					   &pool->flush_worker, 10);
 		}
 	}
 
@@ -733,10 +797,11 @@ void rds_ib_flush_mrs(void)
 
 	down_read(&rds_ib_devices_lock);
 	list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
-		struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
+		if (rds_ibdev->mr_8k_pool)
+			rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
 
-		if (pool)
-			rds_ib_flush_mr_pool(pool, 0, NULL);
+		if (rds_ibdev->mr_1m_pool)
+			rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
 	}
 	up_read(&rds_ib_devices_lock);
 }
@@ -754,12 +819,12 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		goto out;
 	}
 
-	if (!rds_ibdev->mr_pool) {
+	if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
 		ret = -ENODEV;
 		goto out;
 	}
 
-	ibmr = rds_ib_alloc_fmr(rds_ibdev);
+	ibmr = rds_ib_alloc_fmr(rds_ibdev, nents);
 	if (IS_ERR(ibmr)) {
 		rds_ib_dev_put(rds_ibdev);
 		return ibmr;