Merge branch 'for-davem' of git://oss.oracle.com/git/agrover/linux-2.6

include/linux/Kbuild

@@ -302,6 +302,7 @@ header-y += quota.h
 header-y += radeonfb.h
 header-y += random.h
 header-y += raw.h
+header-y += rds.h
 header-y += reboot.h
 header-y += reiserfs_fs.h
 header-y += reiserfs_xattr.h

include/linux/rds.h

@@ -73,6 +73,10 @@
 #define RDS_CMSG_RDMA_MAP		3
 #define RDS_CMSG_RDMA_STATUS		4
 #define RDS_CMSG_CONG_UPDATE		5
+#define RDS_CMSG_ATOMIC_FADD		6
+#define RDS_CMSG_ATOMIC_CSWP		7
+#define RDS_CMSG_MASKED_ATOMIC_FADD	8
+#define RDS_CMSG_MASKED_ATOMIC_CSWP	9
 
 #define RDS_INFO_FIRST			10000
 #define RDS_INFO_COUNTERS		10000
@@ -89,9 +93,9 @@
 #define RDS_INFO_LAST			10010
 
 struct rds_info_counter {
-	u_int8_t	name[32];
-	u_int64_t	value;
-} __packed;
+	uint8_t	name[32];
+	uint64_t	value;
+} __attribute__((packed));
 
 #define RDS_INFO_CONNECTION_FLAG_SENDING	0x01
 #define RDS_INFO_CONNECTION_FLAG_CONNECTING	0x02
@@ -100,56 +104,48 @@ struct rds_info_counter {
 #define TRANSNAMSIZ	16
 
 struct rds_info_connection {
-	u_int64_t	next_tx_seq;
-	u_int64_t	next_rx_seq;
+	uint64_t	next_tx_seq;
+	uint64_t	next_rx_seq;
 	__be32		laddr;
 	__be32		faddr;
-	u_int8_t	transport[TRANSNAMSIZ];		/* null term ascii */
-	u_int8_t	flags;
-} __packed;
-
-struct rds_info_flow {
-	__be32		laddr;
-	__be32		faddr;
-	u_int32_t	bytes;
-	__be16		lport;
-	__be16		fport;
-} __packed;
+	uint8_t	transport[TRANSNAMSIZ];		/* null term ascii */
+	uint8_t	flags;
+} __attribute__((packed));
 
 #define RDS_INFO_MESSAGE_FLAG_ACK               0x01
 #define RDS_INFO_MESSAGE_FLAG_FAST_ACK          0x02
 
 struct rds_info_message {
-	u_int64_t	seq;
-	u_int32_t	len;
+	uint64_t	seq;
+	uint32_t	len;
 	__be32		laddr;
 	__be32		faddr;
 	__be16		lport;
 	__be16		fport;
-	u_int8_t	flags;
-} __packed;
+	uint8_t	flags;
+} __attribute__((packed));
 
 struct rds_info_socket {
-	u_int32_t	sndbuf;
+	uint32_t	sndbuf;
 	__be32		bound_addr;
 	__be32		connected_addr;
 	__be16		bound_port;
 	__be16		connected_port;
-	u_int32_t	rcvbuf;
-	u_int64_t	inum;
-} __packed;
+	uint32_t	rcvbuf;
+	uint64_t	inum;
+} __attribute__((packed));
 
 struct rds_info_tcp_socket {
 	__be32          local_addr;
 	__be16          local_port;
 	__be32          peer_addr;
 	__be16          peer_port;
-	u_int64_t       hdr_rem;
-	u_int64_t       data_rem;
-	u_int32_t       last_sent_nxt;
-	u_int32_t       last_expected_una;
-	u_int32_t       last_seen_una;
-} __packed;
+	uint64_t       hdr_rem;
+	uint64_t       data_rem;
+	uint32_t       last_sent_nxt;
+	uint32_t       last_expected_una;
+	uint32_t       last_seen_una;
+} __attribute__((packed));
 
 #define RDS_IB_GID_LEN	16
 struct rds_info_rdma_connection {
@@ -203,42 +199,69 @@ struct rds_info_rdma_connection {
  * (so that the application does not have to worry about
  * alignment).
  */
-typedef u_int64_t	rds_rdma_cookie_t;
+typedef uint64_t	rds_rdma_cookie_t;
 
 struct rds_iovec {
-	u_int64_t	addr;
-	u_int64_t	bytes;
+	uint64_t	addr;
+	uint64_t	bytes;
 };
 
 struct rds_get_mr_args {
 	struct rds_iovec vec;
-	u_int64_t	cookie_addr;
+	uint64_t	cookie_addr;
 	uint64_t	flags;
 };
 
 struct rds_get_mr_for_dest_args {
 	struct sockaddr_storage	dest_addr;
 	struct rds_iovec 	vec;
-	u_int64_t		cookie_addr;
+	uint64_t		cookie_addr;
 	uint64_t		flags;
 };
 
 struct rds_free_mr_args {
 	rds_rdma_cookie_t cookie;
-	u_int64_t	flags;
+	uint64_t	flags;
 };
 
 struct rds_rdma_args {
 	rds_rdma_cookie_t cookie;
 	struct rds_iovec remote_vec;
-	u_int64_t	local_vec_addr;
-	u_int64_t	nr_local;
-	u_int64_t	flags;
-	u_int64_t	user_token;
+	uint64_t	local_vec_addr;
+	uint64_t	nr_local;
+	uint64_t	flags;
+	uint64_t	user_token;
+};
+
+struct rds_atomic_args {
+	rds_rdma_cookie_t cookie;
+	uint64_t 	local_addr;
+	uint64_t 	remote_addr;
+	union {
+		struct {
+			uint64_t	compare;
+			uint64_t	swap;
+		} cswp;
+		struct {
+			uint64_t	add;
+		} fadd;
+		struct {
+			uint64_t	compare;
+			uint64_t	swap;
+			uint64_t	compare_mask;
+			uint64_t	swap_mask;
+		} m_cswp;
+		struct {
+			uint64_t	add;
+			uint64_t	nocarry_mask;
+		} m_fadd;
+	};
+	uint64_t	flags;
+	uint64_t	user_token;
 };
 
 struct rds_rdma_notify {
-	u_int64_t	user_token;
+	uint64_t	user_token;
 	int32_t		status;
 };
 
@@ -257,5 +280,6 @@ struct rds_rdma_notify {
 #define RDS_RDMA_USE_ONCE	0x0008	/* free MR after use */
 #define RDS_RDMA_DONTWAIT	0x0010	/* Don't wait in SET_BARRIER */
 #define RDS_RDMA_NOTIFY_ME	0x0020	/* Notify when operation completes */
+#define RDS_RDMA_SILENT		0x0040	/* Do not interrupt remote */
 
 #endif /* IB_RDS_H */

net/rds/af_rds.c

@@ -39,7 +39,15 @@
 #include <net/sock.h>
 
 #include "rds.h"
-#include "rdma.h"
+
+char *rds_str_array(char **array, size_t elements, size_t index)
+{
+	if ((index < elements) && array[index])
+		return array[index];
+	else
+		return "unknown";
+}
+EXPORT_SYMBOL(rds_str_array);
 
 /* this is just used for stats gathering :/ */
 static DEFINE_SPINLOCK(rds_sock_lock);
@@ -62,7 +70,7 @@ static int rds_release(struct socket *sock)
 	struct rds_sock *rs;
 	unsigned long flags;
 
-	if (sk == NULL)
+	if (!sk)
 		goto out;
 
 	rs = rds_sk_to_rs(sk);
@@ -73,7 +81,15 @@ static int rds_release(struct socket *sock)
 	 * with the socket. */
 	rds_clear_recv_queue(rs);
 	rds_cong_remove_socket(rs);
+
+	/*
+	 * the binding lookup hash uses rcu, we need to
+	 * make sure we sychronize_rcu before we free our
+	 * entry
+	 */
 	rds_remove_bound(rs);
+	synchronize_rcu();
+
 	rds_send_drop_to(rs, NULL);
 	rds_rdma_drop_keys(rs);
 	rds_notify_queue_get(rs, NULL);
@@ -83,6 +99,8 @@ static int rds_release(struct socket *sock)
 	rds_sock_count--;
 	spin_unlock_irqrestore(&rds_sock_lock, flags);
 
+	rds_trans_put(rs->rs_transport);
+
 	sock->sk = NULL;
 	sock_put(sk);
 out:
@@ -514,7 +532,7 @@ out:
 	spin_unlock_irqrestore(&rds_sock_lock, flags);
 }
 
-static void __exit rds_exit(void)
+static void rds_exit(void)
 {
 	sock_unregister(rds_family_ops.family);
 	proto_unregister(&rds_proto);
@@ -529,7 +547,7 @@ static void __exit rds_exit(void)
 }
 module_exit(rds_exit);
 
-static int __init rds_init(void)
+static int rds_init(void)
 {
 	int ret;
 

net/rds/bind.c

@@ -34,45 +34,52 @@
 #include <net/sock.h>
 #include <linux/in.h>
 #include <linux/if_arp.h>
+#include <linux/jhash.h>
 #include "rds.h"
 
-/*
- * XXX this probably still needs more work.. no INADDR_ANY, and rbtrees aren't
- * particularly zippy.
- *
- * This is now called for every incoming frame so we arguably care much more
- * about it than we used to.
- */
+#define BIND_HASH_SIZE 1024
+static struct hlist_head bind_hash_table[BIND_HASH_SIZE];
 static DEFINE_SPINLOCK(rds_bind_lock);
-static struct rb_root rds_bind_tree = RB_ROOT;
 
-static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port,
-					   struct rds_sock *insert)
+static struct hlist_head *hash_to_bucket(__be32 addr, __be16 port)
+{
+	return bind_hash_table + (jhash_2words((u32)addr, (u32)port, 0) &
+				  (BIND_HASH_SIZE - 1));
+}
+
+static struct rds_sock *rds_bind_lookup(__be32 addr, __be16 port,
+					struct rds_sock *insert)
 {
-	struct rb_node **p = &rds_bind_tree.rb_node;
-	struct rb_node *parent = NULL;
 	struct rds_sock *rs;
+	struct hlist_node *node;
+	struct hlist_head *head = hash_to_bucket(addr, port);
 	u64 cmp;
 	u64 needle = ((u64)be32_to_cpu(addr) << 32) | be16_to_cpu(port);
 
-	while (*p) {
-		parent = *p;
-		rs = rb_entry(parent, struct rds_sock, rs_bound_node);
-
+	rcu_read_lock();
+	hlist_for_each_entry_rcu(rs, node, head, rs_bound_node) {
 		cmp = ((u64)be32_to_cpu(rs->rs_bound_addr) << 32) |
 		      be16_to_cpu(rs->rs_bound_port);
 
-		if (needle < cmp)
-			p = &(*p)->rb_left;
-		else if (needle > cmp)
-			p = &(*p)->rb_right;
-		else
+		if (cmp == needle) {
+			rcu_read_unlock();
 			return rs;
+		}
 	}
+	rcu_read_unlock();
 
 	if (insert) {
-		rb_link_node(&insert->rs_bound_node, parent, p);
-		rb_insert_color(&insert->rs_bound_node, &rds_bind_tree);
+		/*
+		 * make sure our addr and port are set before
+		 * we are added to the list, other people
+		 * in rcu will find us as soon as the
+		 * hlist_add_head_rcu is done
+		 */
+		insert->rs_bound_addr = addr;
+		insert->rs_bound_port = port;
+		rds_sock_addref(insert);
+
+		hlist_add_head_rcu(&insert->rs_bound_node, head);
 	}
 	return NULL;
 }
@@ -86,15 +93,13 @@ static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port,
 struct rds_sock *rds_find_bound(__be32 addr, __be16 port)
 {
 	struct rds_sock *rs;
-	unsigned long flags;
 
-	spin_lock_irqsave(&rds_bind_lock, flags);
-	rs = rds_bind_tree_walk(addr, port, NULL);
+	rs = rds_bind_lookup(addr, port, NULL);
+
 	if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD))
 		rds_sock_addref(rs);
 	else
 		rs = NULL;
-	spin_unlock_irqrestore(&rds_bind_lock, flags);
 
 	rdsdebug("returning rs %p for %pI4:%u\n", rs, &addr,
 		ntohs(port));
@@ -121,22 +126,15 @@ static int rds_add_bound(struct rds_sock *rs, __be32 addr, __be16 *port)
 	do {
 		if (rover == 0)
 			rover++;
-		if (rds_bind_tree_walk(addr, cpu_to_be16(rover), rs) == NULL) {
-			*port = cpu_to_be16(rover);
+		if (!rds_bind_lookup(addr, cpu_to_be16(rover), rs)) {
+			*port = rs->rs_bound_port;
 			ret = 0;
+			rdsdebug("rs %p binding to %pI4:%d\n",
+			  rs, &addr, (int)ntohs(*port));
 			break;
 		}
 	} while (rover++ != last);
 
-	if (ret == 0)  {
-		rs->rs_bound_addr = addr;
-		rs->rs_bound_port = *port;
-		rds_sock_addref(rs);
-
-		rdsdebug("rs %p binding to %pI4:%d\n",
-		  rs, &addr, (int)ntohs(*port));
-	}
-
 	spin_unlock_irqrestore(&rds_bind_lock, flags);
 
 	return ret;
@@ -153,7 +151,7 @@ void rds_remove_bound(struct rds_sock *rs)
 		  rs, &rs->rs_bound_addr,
 		  ntohs(rs->rs_bound_port));
 
-		rb_erase(&rs->rs_bound_node, &rds_bind_tree);
+		hlist_del_init_rcu(&rs->rs_bound_node);
 		rds_sock_put(rs);
 		rs->rs_bound_addr = 0;
 	}
@@ -184,7 +182,7 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 		goto out;
 
 	trans = rds_trans_get_preferred(sin->sin_addr.s_addr);
-	if (trans == NULL) {
+	if (!trans) {
 		ret = -EADDRNOTAVAIL;
 		rds_remove_bound(rs);
 		if (printk_ratelimit())
@@ -198,5 +196,9 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 
 out:
 	release_sock(sk);
+
+	/* we might have called rds_remove_bound on error */
+	if (ret)
+		synchronize_rcu();
 	return ret;
 }

net/rds/cong.c

@@ -141,7 +141,7 @@ static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
 	unsigned long flags;
 
 	map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
-	if (map == NULL)
+	if (!map)
 		return NULL;
 
 	map->m_addr = addr;
@@ -159,7 +159,7 @@ static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
 	ret = rds_cong_tree_walk(addr, map);
 	spin_unlock_irqrestore(&rds_cong_lock, flags);
 
-	if (ret == NULL) {
+	if (!ret) {
 		ret = map;
 		map = NULL;
 	}
@@ -205,7 +205,7 @@ int rds_cong_get_maps(struct rds_connection *conn)
 	conn->c_lcong = rds_cong_from_addr(conn->c_laddr);
 	conn->c_fcong = rds_cong_from_addr(conn->c_faddr);
 
-	if (conn->c_lcong == NULL || conn->c_fcong == NULL)
+	if (!(conn->c_lcong && conn->c_fcong))
 		return -ENOMEM;
 
 	return 0;
@@ -221,7 +221,7 @@ void rds_cong_queue_updates(struct rds_cong_map *map)
 	list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
 		if (!test_and_set_bit(0, &conn->c_map_queued)) {
 			rds_stats_inc(s_cong_update_queued);
-			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+			rds_send_xmit(conn);
 		}
 	}
 

net/rds/connection.c

@@ -37,7 +37,6 @@
 
 #include "rds.h"
 #include "loop.h"
-#include "rdma.h"
 
 #define RDS_CONNECTION_HASH_BITS 12
 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
@@ -63,18 +62,7 @@ static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
 		var |= RDS_INFO_CONNECTION_FLAG_##suffix;	\
 } while (0)
 
-static inline int rds_conn_is_sending(struct rds_connection *conn)
-{
-	int ret = 0;
-
-	if (!mutex_trylock(&conn->c_send_lock))
-		ret = 1;
-	else
-		mutex_unlock(&conn->c_send_lock);
-
-	return ret;
-}
-
+/* rcu read lock must be held or the connection spinlock */
 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
 					      __be32 laddr, __be32 faddr,
 					      struct rds_transport *trans)
@@ -82,7 +70,7 @@ static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
 	struct rds_connection *conn, *ret = NULL;
 	struct hlist_node *pos;
 
-	hlist_for_each_entry(conn, pos, head, c_hash_node) {
+	hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 		if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
 				conn->c_trans == trans) {
 			ret = conn;
@@ -129,10 +117,11 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 {
 	struct rds_connection *conn, *parent = NULL;
 	struct hlist_head *head = rds_conn_bucket(laddr, faddr);
+	struct rds_transport *loop_trans;
 	unsigned long flags;
 	int ret;
 
-	spin_lock_irqsave(&rds_conn_lock, flags);
+	rcu_read_lock();
 	conn = rds_conn_lookup(head, laddr, faddr, trans);
 	if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
 	    !is_outgoing) {
@@ -143,12 +132,12 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 		parent = conn;
 		conn = parent->c_passive;
 	}
-	spin_unlock_irqrestore(&rds_conn_lock, flags);
+	rcu_read_unlock();
 	if (conn)
 		goto out;
 
 	conn = kmem_cache_zalloc(rds_conn_slab, gfp);
-	if (conn == NULL) {
+	if (!conn) {
 		conn = ERR_PTR(-ENOMEM);
 		goto out;
 	}
@@ -159,7 +148,7 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 	spin_lock_init(&conn->c_lock);
 	conn->c_next_tx_seq = 1;
 
-	mutex_init(&conn->c_send_lock);
+	init_waitqueue_head(&conn->c_waitq);
 	INIT_LIST_HEAD(&conn->c_send_queue);
 	INIT_LIST_HEAD(&conn->c_retrans);
 
@@ -175,7 +164,9 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 	 * can bind to the destination address then we'd rather the messages
 	 * flow through loopback rather than either transport.
 	 */
-	if (rds_trans_get_preferred(faddr)) {
+	loop_trans = rds_trans_get_preferred(faddr);
+	if (loop_trans) {
+		rds_trans_put(loop_trans);
 		conn->c_loopback = 1;
 		if (is_outgoing && trans->t_prefer_loopback) {
 			/* "outgoing" connection - and the transport
@@ -238,7 +229,7 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
 			kmem_cache_free(rds_conn_slab, conn);
 			conn = found;
 		} else {
-			hlist_add_head(&conn->c_hash_node, head);
+			hlist_add_head_rcu(&conn->c_hash_node, head);
 			rds_cong_add_conn(conn);
 			rds_conn_count++;
 		}
@@ -263,21 +254,91 @@ struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 }
 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
 
+void rds_conn_shutdown(struct rds_connection *conn)
+{
+	/* shut it down unless it's down already */
+	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
+		/*
+		 * Quiesce the connection mgmt handlers before we start tearing
+		 * things down. We don't hold the mutex for the entire
+		 * duration of the shutdown operation, else we may be
+		 * deadlocking with the CM handler. Instead, the CM event
+		 * handler is supposed to check for state DISCONNECTING
+		 */
+		mutex_lock(&conn->c_cm_lock);
+		if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
+		 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
+			rds_conn_error(conn, "shutdown called in state %d\n",
+					atomic_read(&conn->c_state));
+			mutex_unlock(&conn->c_cm_lock);
+			return;
+		}
+		mutex_unlock(&conn->c_cm_lock);
+
+		wait_event(conn->c_waitq,
+			   !test_bit(RDS_IN_XMIT, &conn->c_flags));
+
+		conn->c_trans->conn_shutdown(conn);
+		rds_conn_reset(conn);
+
+		if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
+			/* This can happen - eg when we're in the middle of tearing
+			 * down the connection, and someone unloads the rds module.
+			 * Quite reproduceable with loopback connections.
+			 * Mostly harmless.
+			 */
+			rds_conn_error(conn,
+				"%s: failed to transition to state DOWN, "
+				"current state is %d\n",
+				__func__,
+				atomic_read(&conn->c_state));
+			return;
+		}
+	}
+
+	/* Then reconnect if it's still live.
+	 * The passive side of an IB loopback connection is never added
+	 * to the conn hash, so we never trigger a reconnect on this
+	 * conn - the reconnect is always triggered by the active peer. */
+	cancel_delayed_work_sync(&conn->c_conn_w);
+	rcu_read_lock();
+	if (!hlist_unhashed(&conn->c_hash_node)) {
+		rcu_read_unlock();
+		rds_queue_reconnect(conn);
+	} else {
+		rcu_read_unlock();
+	}
+}
+
+/*
+ * Stop and free a connection.
+ *
+ * This can only be used in very limited circumstances.  It assumes that once
+ * the conn has been shutdown that no one else is referencing the connection.
+ * We can only ensure this in the rmmod path in the current code.
+ */
 void rds_conn_destroy(struct rds_connection *conn)
 {
 	struct rds_message *rm, *rtmp;
+	unsigned long flags;
 
 	rdsdebug("freeing conn %p for %pI4 -> "
 		 "%pI4\n", conn, &conn->c_laddr,
 		 &conn->c_faddr);
 
-	hlist_del_init(&conn->c_hash_node);
+	/* Ensure conn will not be scheduled for reconnect */
+	spin_lock_irq(&rds_conn_lock);
+	hlist_del_init_rcu(&conn->c_hash_node);
+	spin_unlock_irq(&rds_conn_lock);
+	synchronize_rcu();
 
-	/* wait for the rds thread to shut it down */
-	atomic_set(&conn->c_state, RDS_CONN_ERROR);
-	cancel_delayed_work(&conn->c_conn_w);
-	queue_work(rds_wq, &conn->c_down_w);
-	flush_workqueue(rds_wq);
+	/* shut the connection down */
+	rds_conn_drop(conn);
+	flush_work(&conn->c_down_w);
+
+	/* make sure lingering queued work won't try to ref the conn */
+	cancel_delayed_work_sync(&conn->c_send_w);
+	cancel_delayed_work_sync(&conn->c_recv_w);
 
 	/* tear down queued messages */
 	list_for_each_entry_safe(rm, rtmp,
@@ -302,7 +363,9 @@ void rds_conn_destroy(struct rds_connection *conn)
 	BUG_ON(!list_empty(&conn->c_retrans));
 	kmem_cache_free(rds_conn_slab, conn);
 
+	spin_lock_irqsave(&rds_conn_lock, flags);
 	rds_conn_count--;
+	spin_unlock_irqrestore(&rds_conn_lock, flags);
 }
 EXPORT_SYMBOL_GPL(rds_conn_destroy);
 
@@ -316,23 +379,23 @@ static void rds_conn_message_info(struct socket *sock, unsigned int len,
 	struct list_head *list;
 	struct rds_connection *conn;
 	struct rds_message *rm;
-	unsigned long flags;
 	unsigned int total = 0;
+	unsigned long flags;
 	size_t i;
 
 	len /= sizeof(struct rds_info_message);
 
-	spin_lock_irqsave(&rds_conn_lock, flags);
+	rcu_read_lock();
 
 	for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 	     i++, head++) {
-		hlist_for_each_entry(conn, pos, head, c_hash_node) {
+		hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 			if (want_send)
 				list = &conn->c_send_queue;
 			else
 				list = &conn->c_retrans;
 
-			spin_lock(&conn->c_lock);
+			spin_lock_irqsave(&conn->c_lock, flags);
 
 			/* XXX too lazy to maintain counts.. */
 			list_for_each_entry(rm, list, m_conn_item) {
@@ -343,11 +406,10 @@ static void rds_conn_message_info(struct socket *sock, unsigned int len,
 							  conn->c_faddr, 0);
 			}
 
-			spin_unlock(&conn->c_lock);
+			spin_unlock_irqrestore(&conn->c_lock, flags);
 		}
 	}
-
-	spin_unlock_irqrestore(&rds_conn_lock, flags);
+	rcu_read_unlock();
 
 	lens->nr = total;
 	lens->each = sizeof(struct rds_info_message);
@@ -377,19 +439,17 @@ void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 	uint64_t buffer[(item_len + 7) / 8];
 	struct hlist_head *head;
 	struct hlist_node *pos;
-	struct hlist_node *tmp;
 	struct rds_connection *conn;
-	unsigned long flags;
 	size_t i;
 
-	spin_lock_irqsave(&rds_conn_lock, flags);
+	rcu_read_lock();
 
 	lens->nr = 0;
 	lens->each = item_len;
 
 	for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
 	     i++, head++) {
-		hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
+		hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
 
 			/* XXX no c_lock usage.. */
 			if (!visitor(conn, buffer))
@@ -405,8 +465,7 @@ void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 			lens->nr++;
 		}
 	}
-
-	spin_unlock_irqrestore(&rds_conn_lock, flags);
+	rcu_read_unlock();
 }
 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
 
@@ -423,8 +482,8 @@ static int rds_conn_info_visitor(struct rds_connection *conn,
 		sizeof(cinfo->transport));
 	cinfo->flags = 0;
 
-	rds_conn_info_set(cinfo->flags,
-			  rds_conn_is_sending(conn), SENDING);
+	rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
+			  SENDING);
 	/* XXX Future: return the state rather than these funky bits */
 	rds_conn_info_set(cinfo->flags,
 			  atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
@@ -444,12 +503,12 @@ static void rds_conn_info(struct socket *sock, unsigned int len,
 				sizeof(struct rds_info_connection));
 }
 
-int __init rds_conn_init(void)
+int rds_conn_init(void)
 {
 	rds_conn_slab = kmem_cache_create("rds_connection",
 					  sizeof(struct rds_connection),
 					  0, 0, NULL);
-	if (rds_conn_slab == NULL)
+	if (!rds_conn_slab)
 		return -ENOMEM;
 
 	rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
@@ -486,6 +545,18 @@ void rds_conn_drop(struct rds_connection *conn)
 }
 EXPORT_SYMBOL_GPL(rds_conn_drop);
 
+/*
+ * If the connection is down, trigger a connect. We may have scheduled a
+ * delayed reconnect however - in this case we should not interfere.
+ */
+void rds_conn_connect_if_down(struct rds_connection *conn)
+{
+	if (rds_conn_state(conn) == RDS_CONN_DOWN &&
+	    !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
+		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
+}
+EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
+
 /*
  * An error occurred on the connection
  */

net/rds/ib.c

@@ -53,12 +53,71 @@ MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
 module_param(rds_ib_retry_count, int, 0444);
 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
 
+/*
+ * we have a clumsy combination of RCU and a rwsem protecting this list
+ * because it is used both in the get_mr fast path and while blocking in
+ * the FMR flushing path.
+ */
+DECLARE_RWSEM(rds_ib_devices_lock);
 struct list_head rds_ib_devices;
 
 /* NOTE: if also grabbing ibdev lock, grab this first */
 DEFINE_SPINLOCK(ib_nodev_conns_lock);
 LIST_HEAD(ib_nodev_conns);
 
+void rds_ib_nodev_connect(void)
+{
+	struct rds_ib_connection *ic;
+
+	spin_lock(&ib_nodev_conns_lock);
+	list_for_each_entry(ic, &ib_nodev_conns, ib_node)
+		rds_conn_connect_if_down(ic->conn);
+	spin_unlock(&ib_nodev_conns_lock);
+}
+
+void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
+{
+	struct rds_ib_connection *ic;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rds_ibdev->spinlock, flags);
+	list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
+		rds_conn_drop(ic->conn);
+	spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
+}
+
+/*
+ * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
+ * from interrupt context so we push freing off into a work struct in krdsd.
+ */
+static void rds_ib_dev_free(struct work_struct *work)
+{
+	struct rds_ib_ipaddr *i_ipaddr, *i_next;
+	struct rds_ib_device *rds_ibdev = container_of(work,
+					struct rds_ib_device, free_work);
+
+	if (rds_ibdev->mr_pool)
+		rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
+	if (rds_ibdev->mr)
+		ib_dereg_mr(rds_ibdev->mr);
+	if (rds_ibdev->pd)
+		ib_dealloc_pd(rds_ibdev->pd);
+
+	list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
+		list_del(&i_ipaddr->list);
+		kfree(i_ipaddr);
+	}
+
+	kfree(rds_ibdev);
+}
+
+void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
+{
+	BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
+	if (atomic_dec_and_test(&rds_ibdev->refcount))
+		queue_work(rds_wq, &rds_ibdev->free_work);
+}
+
 void rds_ib_add_one(struct ib_device *device)
 {
 	struct rds_ib_device *rds_ibdev;
@@ -77,11 +136,14 @@ void rds_ib_add_one(struct ib_device *device)
 		goto free_attr;
 	}
 
-	rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL);
+	rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
+				 ibdev_to_node(device));
 	if (!rds_ibdev)
 		goto free_attr;
 
 	spin_lock_init(&rds_ibdev->spinlock);
+	atomic_set(&rds_ibdev->refcount, 1);
+	INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
 
 	rds_ibdev->max_wrs = dev_attr->max_qp_wr;
 	rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
@@ -91,68 +153,107 @@ void rds_ib_add_one(struct ib_device *device)
 			min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
 			fmr_pool_size;
 
+	rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
+	rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
+
 	rds_ibdev->dev = device;
 	rds_ibdev->pd = ib_alloc_pd(device);
-	if (IS_ERR(rds_ibdev->pd))
-		goto free_dev;
+	if (IS_ERR(rds_ibdev->pd)) {
+		rds_ibdev->pd = NULL;
+		goto put_dev;
+	}
 
-	rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd,
-				      IB_ACCESS_LOCAL_WRITE);
-	if (IS_ERR(rds_ibdev->mr))
-		goto err_pd;
+	rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
+	if (IS_ERR(rds_ibdev->mr)) {
+		rds_ibdev->mr = NULL;
+		goto put_dev;
+	}
 
 	rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
 	if (IS_ERR(rds_ibdev->mr_pool)) {
 		rds_ibdev->mr_pool = NULL;
-		goto err_mr;
+		goto put_dev;
 	}
 
 	INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
 	INIT_LIST_HEAD(&rds_ibdev->conn_list);
-	list_add_tail(&rds_ibdev->list, &rds_ib_devices);
+
+	down_write(&rds_ib_devices_lock);
+	list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
+	up_write(&rds_ib_devices_lock);
+	atomic_inc(&rds_ibdev->refcount);
 
 	ib_set_client_data(device, &rds_ib_client, rds_ibdev);
+	atomic_inc(&rds_ibdev->refcount);
 
-	goto free_attr;
+	rds_ib_nodev_connect();
 
-err_mr:
-	ib_dereg_mr(rds_ibdev->mr);
-err_pd:
-	ib_dealloc_pd(rds_ibdev->pd);
-free_dev:
-	kfree(rds_ibdev);
+put_dev:
+	rds_ib_dev_put(rds_ibdev);
 free_attr:
 	kfree(dev_attr);
 }
 
+/*
+ * New connections use this to find the device to associate with the
+ * connection.  It's not in the fast path so we're not concerned about the
+ * performance of the IB call.  (As of this writing, it uses an interrupt
+ * blocking spinlock to serialize walking a per-device list of all registered
+ * clients.)
+ *
+ * RCU is used to handle incoming connections racing with device teardown.
+ * Rather than use a lock to serialize removal from the client_data and
+ * getting a new reference, we use an RCU grace period.  The destruction
+ * path removes the device from client_data and then waits for all RCU
+ * readers to finish.
+ *
+ * A new connection can get NULL from this if its arriving on a
+ * device that is in the process of being removed.
+ */
+struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
+{
+	struct rds_ib_device *rds_ibdev;
+
+	rcu_read_lock();
+	rds_ibdev = ib_get_client_data(device, &rds_ib_client);
+	if (rds_ibdev)
+		atomic_inc(&rds_ibdev->refcount);
+	rcu_read_unlock();
+	return rds_ibdev;
+}
+
+/*
+ * The IB stack is letting us know that a device is going away.  This can
+ * happen if the underlying HCA driver is removed or if PCI hotplug is removing
+ * the pci function, for example.
+ *
+ * This can be called at any time and can be racing with any other RDS path.
+ */
 void rds_ib_remove_one(struct ib_device *device)
 {
 	struct rds_ib_device *rds_ibdev;
-	struct rds_ib_ipaddr *i_ipaddr, *i_next;
 
 	rds_ibdev = ib_get_client_data(device, &rds_ib_client);
 	if (!rds_ibdev)
 		return;
 
-	list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
-		list_del(&i_ipaddr->list);
-		kfree(i_ipaddr);
-	}
+	rds_ib_dev_shutdown(rds_ibdev);
 
-	rds_ib_destroy_conns(rds_ibdev);
+	/* stop connection attempts from getting a reference to this device. */
+	ib_set_client_data(device, &rds_ib_client, NULL);
 
-	if (rds_ibdev->mr_pool)
-		rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
-
-	ib_dereg_mr(rds_ibdev->mr);
-
-	while (ib_dealloc_pd(rds_ibdev->pd)) {
-		rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
-		msleep(1);
-	}
+	down_write(&rds_ib_devices_lock);
+	list_del_rcu(&rds_ibdev->list);
+	up_write(&rds_ib_devices_lock);
 
-	list_del(&rds_ibdev->list);
-	kfree(rds_ibdev);
+	/*
+	 * This synchronize rcu is waiting for readers of both the ib
+	 * client data and the devices list to finish before we drop
+	 * both of those references.
+	 */
+	synchronize_rcu();
+	rds_ib_dev_put(rds_ibdev);
+	rds_ib_dev_put(rds_ibdev);
 }
 
 struct ib_client rds_ib_client = {
@@ -186,7 +287,7 @@ static int rds_ib_conn_info_visitor(struct rds_connection *conn,
 		rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
 		rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
 
-		rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
+		rds_ibdev = ic->rds_ibdev;
 		iinfo->max_send_wr = ic->i_send_ring.w_nr;
 		iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
 		iinfo->max_send_sge = rds_ibdev->max_sge;
@@ -248,29 +349,36 @@ static int rds_ib_laddr_check(__be32 addr)
 	return ret;
 }
 
+static void rds_ib_unregister_client(void)
+{
+	ib_unregister_client(&rds_ib_client);
+	/* wait for rds_ib_dev_free() to complete */
+	flush_workqueue(rds_wq);
+}
+
 void rds_ib_exit(void)
 {
 	rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
+	rds_ib_unregister_client();
 	rds_ib_destroy_nodev_conns();
-	ib_unregister_client(&rds_ib_client);
 	rds_ib_sysctl_exit();
 	rds_ib_recv_exit();
 	rds_trans_unregister(&rds_ib_transport);
+	rds_ib_fmr_exit();
 }
 
 struct rds_transport rds_ib_transport = {
 	.laddr_check		= rds_ib_laddr_check,
 	.xmit_complete		= rds_ib_xmit_complete,
 	.xmit			= rds_ib_xmit,
-	.xmit_cong_map		= NULL,
 	.xmit_rdma		= rds_ib_xmit_rdma,
+	.xmit_atomic		= rds_ib_xmit_atomic,
 	.recv			= rds_ib_recv,
 	.conn_alloc		= rds_ib_conn_alloc,
 	.conn_free		= rds_ib_conn_free,
 	.conn_connect		= rds_ib_conn_connect,
 	.conn_shutdown		= rds_ib_conn_shutdown,
 	.inc_copy_to_user	= rds_ib_inc_copy_to_user,
-	.inc_purge		= rds_ib_inc_purge,
 	.inc_free		= rds_ib_inc_free,
 	.cm_initiate_connect	= rds_ib_cm_initiate_connect,
 	.cm_handle_connect	= rds_ib_cm_handle_connect,
@@ -286,16 +394,20 @@ struct rds_transport rds_ib_transport = {
 	.t_type			= RDS_TRANS_IB
 };
 
-int __init rds_ib_init(void)
+int rds_ib_init(void)
 {
 	int ret;
 
 	INIT_LIST_HEAD(&rds_ib_devices);
 
-	ret = ib_register_client(&rds_ib_client);
+	ret = rds_ib_fmr_init();
 	if (ret)
 		goto out;
 
+	ret = ib_register_client(&rds_ib_client);
+	if (ret)
+		goto out_fmr_exit;
+
 	ret = rds_ib_sysctl_init();
 	if (ret)
 		goto out_ibreg;
@@ -317,7 +429,9 @@ out_recv:
 out_sysctl:
 	rds_ib_sysctl_exit();
 out_ibreg:
-	ib_unregister_client(&rds_ib_client);
+	rds_ib_unregister_client();
+out_fmr_exit:
+	rds_ib_fmr_exit();
 out:
 	return ret;
 }

net/rds/ib.h

@@ -3,11 +3,13 @@
 
 #include <rdma/ib_verbs.h>
 #include <rdma/rdma_cm.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
 #include "rds.h"
 #include "rdma_transport.h"
 
 #define RDS_FMR_SIZE			256
-#define RDS_FMR_POOL_SIZE		4096
+#define RDS_FMR_POOL_SIZE		8192
 
 #define RDS_IB_MAX_SGE			8
 #define RDS_IB_RECV_SGE 		2
@@ -19,6 +21,9 @@
 
 #define RDS_IB_SUPPORTED_PROTOCOLS	0x00000003	/* minor versions supported */
 
+#define RDS_IB_RECYCLE_BATCH_COUNT	32
+
+extern struct rw_semaphore rds_ib_devices_lock;
 extern struct list_head rds_ib_devices;
 
 /*
@@ -26,20 +31,29 @@ extern struct list_head rds_ib_devices;
  * try and minimize the amount of memory tied up both the device and
  * socket receive queues.
  */
-/* page offset of the final full frag that fits in the page */
-#define RDS_PAGE_LAST_OFF (((PAGE_SIZE  / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
 struct rds_page_frag {
 	struct list_head	f_item;
-	struct page		*f_page;
-	unsigned long		f_offset;
-	dma_addr_t 		f_mapped;
+	struct list_head	f_cache_entry;
+	struct scatterlist	f_sg;
 };
 
 struct rds_ib_incoming {
 	struct list_head	ii_frags;
+	struct list_head	ii_cache_entry;
 	struct rds_incoming	ii_inc;
 };
 
+struct rds_ib_cache_head {
+	struct list_head *first;
+	unsigned long count;
+};
+
+struct rds_ib_refill_cache {
+	struct rds_ib_cache_head *percpu;
+	struct list_head	 *xfer;
+	struct list_head	 *ready;
+};
+
 struct rds_ib_connect_private {
 	/* Add new fields at the end, and don't permute existing fields. */
 	__be32			dp_saddr;
@@ -53,8 +67,7 @@ struct rds_ib_connect_private {
 };
 
 struct rds_ib_send_work {
-	struct rds_message	*s_rm;
-	struct rds_rdma_op	*s_op;
+	void			*s_op;
 	struct ib_send_wr	s_wr;
 	struct ib_sge		s_sge[RDS_IB_MAX_SGE];
 	unsigned long		s_queued;
@@ -92,10 +105,11 @@ struct rds_ib_connection {
 
 	/* tx */
 	struct rds_ib_work_ring	i_send_ring;
-	struct rds_message	*i_rm;
+	struct rm_data_op	*i_data_op;
 	struct rds_header	*i_send_hdrs;
 	u64			i_send_hdrs_dma;
 	struct rds_ib_send_work *i_sends;
+	atomic_t		i_signaled_sends;
 
 	/* rx */
 	struct tasklet_struct	i_recv_tasklet;
@@ -106,8 +120,9 @@ struct rds_ib_connection {
 	struct rds_header	*i_recv_hdrs;
 	u64			i_recv_hdrs_dma;
 	struct rds_ib_recv_work *i_recvs;
-	struct rds_page_frag	i_frag;
 	u64			i_ack_recv;	/* last ACK received */
+	struct rds_ib_refill_cache i_cache_incs;
+	struct rds_ib_refill_cache i_cache_frags;
 
 	/* sending acks */
 	unsigned long		i_ack_flags;
@@ -138,7 +153,6 @@ struct rds_ib_connection {
 
 	/* Batched completions */
 	unsigned int		i_unsignaled_wrs;
-	long			i_unsignaled_bytes;
 };
 
 /* This assumes that atomic_t is at least 32 bits */
@@ -164,9 +178,17 @@ struct rds_ib_device {
 	unsigned int		max_fmrs;
 	int			max_sge;
 	unsigned int		max_wrs;
+	unsigned int		max_initiator_depth;
+	unsigned int		max_responder_resources;
 	spinlock_t		spinlock;	/* protect the above */
+	atomic_t		refcount;
+	struct work_struct	free_work;
 };
 
+#define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
+#define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
+#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
+
 /* bits for i_ack_flags */
 #define IB_ACK_IN_FLIGHT	0
 #define IB_ACK_REQUESTED	1
@@ -202,6 +224,8 @@ struct rds_ib_statistics {
 	uint64_t	s_ib_rdma_mr_pool_flush;
 	uint64_t	s_ib_rdma_mr_pool_wait;
 	uint64_t	s_ib_rdma_mr_pool_depleted;
+	uint64_t	s_ib_atomic_cswp;
+	uint64_t	s_ib_atomic_fadd;
 };
 
 extern struct workqueue_struct *rds_ib_wq;
@@ -243,6 +267,8 @@ static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
 extern struct rds_transport rds_ib_transport;
 extern void rds_ib_add_one(struct ib_device *device);
 extern void rds_ib_remove_one(struct ib_device *device);
+struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
+void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
 extern struct ib_client rds_ib_client;
 
 extern unsigned int fmr_pool_size;
@@ -258,7 +284,7 @@ void rds_ib_conn_free(void *arg);
 int rds_ib_conn_connect(struct rds_connection *conn);
 void rds_ib_conn_shutdown(struct rds_connection *conn);
 void rds_ib_state_change(struct sock *sk);
-int __init rds_ib_listen_init(void);
+int rds_ib_listen_init(void);
 void rds_ib_listen_stop(void);
 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
@@ -275,15 +301,7 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn,
 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
-void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock);
-static inline void rds_ib_destroy_nodev_conns(void)
-{
-	__rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
-}
-static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev)
-{
-	__rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
-}
+void rds_ib_destroy_nodev_conns(void);
 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
@@ -292,14 +310,16 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 void rds_ib_sync_mr(void *trans_private, int dir);
 void rds_ib_free_mr(void *trans_private, int invalidate);
 void rds_ib_flush_mrs(void);
+int rds_ib_fmr_init(void);
+void rds_ib_fmr_exit(void);
 
 /* ib_recv.c */
-int __init rds_ib_recv_init(void);
+int rds_ib_recv_init(void);
 void rds_ib_recv_exit(void);
 int rds_ib_recv(struct rds_connection *conn);
-int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
-		       gfp_t page_gfp, int prefill);
-void rds_ib_inc_purge(struct rds_incoming *inc);
+int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
+void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
+void rds_ib_recv_refill(struct rds_connection *conn, int prefill);
 void rds_ib_inc_free(struct rds_incoming *inc);
 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 			     size_t size);
@@ -325,17 +345,19 @@ u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
 extern wait_queue_head_t rds_ib_ring_empty_wait;
 
 /* ib_send.c */
+char *rds_ib_wc_status_str(enum ib_wc_status status);
 void rds_ib_xmit_complete(struct rds_connection *conn);
 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 		unsigned int hdr_off, unsigned int sg, unsigned int off);
 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
-int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
+int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
 			     u32 *adv_credits, int need_posted, int max_posted);
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
 
 /* ib_stats.c */
 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
@@ -344,7 +366,7 @@ unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
 				    unsigned int avail);
 
 /* ib_sysctl.c */
-int __init rds_ib_sysctl_init(void);
+int rds_ib_sysctl_init(void);
 void rds_ib_sysctl_exit(void);
 extern unsigned long rds_ib_sysctl_max_send_wr;
 extern unsigned long rds_ib_sysctl_max_recv_wr;
@@ -354,28 +376,4 @@ extern unsigned long rds_ib_sysctl_max_recv_allocation;
 extern unsigned int rds_ib_sysctl_flow_control;
 extern ctl_table rds_ib_sysctl_table[];
 
-/*
- * Helper functions for getting/setting the header and data SGEs in
- * RDS packets (not RDMA)
- *
- * From version 3.1 onwards, header is in front of data in the sge.
- */
-static inline struct ib_sge *
-rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
-{
-	if (ic->conn->c_version > RDS_PROTOCOL_3_0)
-		return &sge[0];
-	else
-		return &sge[1];
-}
-
-static inline struct ib_sge *
-rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
-{
-	if (ic->conn->c_version > RDS_PROTOCOL_3_0)
-		return &sge[1];
-	else
-		return &sge[0];
-}
-
 #endif

net/rds/ib_cm.c

@@ -38,6 +38,36 @@
 #include "rds.h"
 #include "ib.h"
 
+static char *rds_ib_event_type_strings[] = {
+#define RDS_IB_EVENT_STRING(foo) \
+		[IB_EVENT_##foo] = __stringify(IB_EVENT_##foo)
+	RDS_IB_EVENT_STRING(CQ_ERR),
+	RDS_IB_EVENT_STRING(QP_FATAL),
+	RDS_IB_EVENT_STRING(QP_REQ_ERR),
+	RDS_IB_EVENT_STRING(QP_ACCESS_ERR),
+	RDS_IB_EVENT_STRING(COMM_EST),
+	RDS_IB_EVENT_STRING(SQ_DRAINED),
+	RDS_IB_EVENT_STRING(PATH_MIG),
+	RDS_IB_EVENT_STRING(PATH_MIG_ERR),
+	RDS_IB_EVENT_STRING(DEVICE_FATAL),
+	RDS_IB_EVENT_STRING(PORT_ACTIVE),
+	RDS_IB_EVENT_STRING(PORT_ERR),
+	RDS_IB_EVENT_STRING(LID_CHANGE),
+	RDS_IB_EVENT_STRING(PKEY_CHANGE),
+	RDS_IB_EVENT_STRING(SM_CHANGE),
+	RDS_IB_EVENT_STRING(SRQ_ERR),
+	RDS_IB_EVENT_STRING(SRQ_LIMIT_REACHED),
+	RDS_IB_EVENT_STRING(QP_LAST_WQE_REACHED),
+	RDS_IB_EVENT_STRING(CLIENT_REREGISTER),
+#undef RDS_IB_EVENT_STRING
+};
+
+static char *rds_ib_event_str(enum ib_event_type type)
+{
+	return rds_str_array(rds_ib_event_type_strings,
+			     ARRAY_SIZE(rds_ib_event_type_strings), type);
+};
+
 /*
  * Set the selected protocol version
  */
@@ -95,7 +125,6 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
 {
 	const struct rds_ib_connect_private *dp = NULL;
 	struct rds_ib_connection *ic = conn->c_transport_data;
-	struct rds_ib_device *rds_ibdev;
 	struct ib_qp_attr qp_attr;
 	int err;
 
@@ -111,11 +140,21 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
 		}
 	}
 
-	printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
-			&conn->c_faddr,
-			RDS_PROTOCOL_MAJOR(conn->c_version),
-			RDS_PROTOCOL_MINOR(conn->c_version),
-			ic->i_flowctl ? ", flow control" : "");
+	if (conn->c_version < RDS_PROTOCOL(3,1)) {
+		printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
+		       " no longer supported\n",
+		       &conn->c_faddr,
+		       RDS_PROTOCOL_MAJOR(conn->c_version),
+		       RDS_PROTOCOL_MINOR(conn->c_version));
+		rds_conn_destroy(conn);
+		return;
+	} else {
+		printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
+		       &conn->c_faddr,
+		       RDS_PROTOCOL_MAJOR(conn->c_version),
+		       RDS_PROTOCOL_MINOR(conn->c_version),
+		       ic->i_flowctl ? ", flow control" : "");
+	}
 
 	/*
 	 * Init rings and fill recv. this needs to wait until protocol negotiation
@@ -125,7 +164,7 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
 	rds_ib_recv_init_ring(ic);
 	/* Post receive buffers - as a side effect, this will update
 	 * the posted credit count. */
-	rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
+	rds_ib_recv_refill(conn, 1);
 
 	/* Tune RNR behavior */
 	rds_ib_tune_rnr(ic, &qp_attr);
@@ -135,12 +174,11 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
 	if (err)
 		printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 
-	/* update ib_device with this local ipaddr & conn */
-	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
-	err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
+	/* update ib_device with this local ipaddr */
+	err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
 	if (err)
-		printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
-	rds_ib_add_conn(rds_ibdev, conn);
+		printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
+			err);
 
 	/* If the peer gave us the last packet it saw, process this as if
 	 * we had received a regular ACK. */
@@ -153,18 +191,23 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 			struct rdma_conn_param *conn_param,
 			struct rds_ib_connect_private *dp,
-			u32 protocol_version)
+			u32 protocol_version,
+			u32 max_responder_resources,
+			u32 max_initiator_depth)
 {
+	struct rds_ib_connection *ic = conn->c_transport_data;
+	struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
+
 	memset(conn_param, 0, sizeof(struct rdma_conn_param));
-	/* XXX tune these? */
-	conn_param->responder_resources = 1;
-	conn_param->initiator_depth = 1;
+
+	conn_param->responder_resources =
+		min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
+	conn_param->initiator_depth =
+		min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 	conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 	conn_param->rnr_retry_count = 7;
 
 	if (dp) {
-		struct rds_ib_connection *ic = conn->c_transport_data;
-
 		memset(dp, 0, sizeof(*dp));
 		dp->dp_saddr = conn->c_laddr;
 		dp->dp_daddr = conn->c_faddr;
@@ -189,7 +232,8 @@ static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 
 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 {
-	rdsdebug("event %u data %p\n", event->event, data);
+	rdsdebug("event %u (%s) data %p\n",
+		 event->event, rds_ib_event_str(event->event), data);
 }
 
 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
@@ -197,16 +241,18 @@ static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 	struct rds_connection *conn = data;
 	struct rds_ib_connection *ic = conn->c_transport_data;
 
-	rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
+	rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
+		 rds_ib_event_str(event->event));
 
 	switch (event->event) {
 	case IB_EVENT_COMM_EST:
 		rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 		break;
 	default:
-		rdsdebug("Fatal QP Event %u "
+		rdsdebug("Fatal QP Event %u (%s) "
 			"- connection %pI4->%pI4, reconnecting\n",
-			event->event, &conn->c_laddr, &conn->c_faddr);
+			event->event, rds_ib_event_str(event->event),
+			&conn->c_laddr, &conn->c_faddr);
 		rds_conn_drop(conn);
 		break;
 	}
@@ -224,18 +270,16 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 	struct rds_ib_device *rds_ibdev;
 	int ret;
 
-	/* rds_ib_add_one creates a rds_ib_device object per IB device,
-	 * and allocates a protection domain, memory range and FMR pool
-	 * for each.  If that fails for any reason, it will not register
-	 * the rds_ibdev at all.
+	/*
+	 * It's normal to see a null device if an incoming connection races
+	 * with device removal, so we don't print a warning.
 	 */
-	rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
-	if (rds_ibdev == NULL) {
-		if (printk_ratelimit())
-			printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
-					dev->name);
+	rds_ibdev = rds_ib_get_client_data(dev);
+	if (!rds_ibdev)
 		return -EOPNOTSUPP;
-	}
+
+	/* add the conn now so that connection establishment has the dev */
+	rds_ib_add_conn(rds_ibdev, conn);
 
 	if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 		rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
@@ -306,7 +350,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 					   ic->i_send_ring.w_nr *
 						sizeof(struct rds_header),
 					   &ic->i_send_hdrs_dma, GFP_KERNEL);
-	if (ic->i_send_hdrs == NULL) {
+	if (!ic->i_send_hdrs) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent send failed\n");
 		goto out;
@@ -316,7 +360,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 					   ic->i_recv_ring.w_nr *
 						sizeof(struct rds_header),
 					   &ic->i_recv_hdrs_dma, GFP_KERNEL);
-	if (ic->i_recv_hdrs == NULL) {
+	if (!ic->i_recv_hdrs) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent recv failed\n");
 		goto out;
@@ -324,22 +368,24 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 
 	ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 				       &ic->i_ack_dma, GFP_KERNEL);
-	if (ic->i_ack == NULL) {
+	if (!ic->i_ack) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent ack failed\n");
 		goto out;
 	}
 
-	ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
-	if (ic->i_sends == NULL) {
+	ic->i_sends = vmalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
+				   ibdev_to_node(dev));
+	if (!ic->i_sends) {
 		ret = -ENOMEM;
 		rdsdebug("send allocation failed\n");
 		goto out;
 	}
 	memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 
-	ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
-	if (ic->i_recvs == NULL) {
+	ic->i_recvs = vmalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
+				   ibdev_to_node(dev));
+	if (!ic->i_recvs) {
 		ret = -ENOMEM;
 		rdsdebug("recv allocation failed\n");
 		goto out;
@@ -352,6 +398,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
 		 ic->i_send_cq, ic->i_recv_cq);
 
 out:
+	rds_ib_dev_put(rds_ibdev);
 	return ret;
 }
 
@@ -409,7 +456,7 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 	struct rds_ib_connection *ic = NULL;
 	struct rdma_conn_param conn_param;
 	u32 version;
-	int err, destroy = 1;
+	int err = 1, destroy = 1;
 
 	/* Check whether the remote protocol version matches ours. */
 	version = rds_ib_protocol_compatible(event);
@@ -448,7 +495,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 			/* Wait and see - our connect may still be succeeding */
 			rds_ib_stats_inc(s_ib_connect_raced);
 		}
-		mutex_unlock(&conn->c_cm_lock);
 		goto out;
 	}
 
@@ -479,20 +525,20 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 		goto out;
 	}
 
-	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
+	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
+		event->param.conn.responder_resources,
+		event->param.conn.initiator_depth);
 
 	/* rdma_accept() calls rdma_reject() internally if it fails */
 	err = rdma_accept(cm_id, &conn_param);
-	mutex_unlock(&conn->c_cm_lock);
-	if (err) {
+	if (err)
 		rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
-		goto out;
-	}
-
-	return 0;
 
 out:
-	rdma_reject(cm_id, NULL, 0);
+	if (conn)
+		mutex_unlock(&conn->c_cm_lock);
+	if (err)
+		rdma_reject(cm_id, NULL, 0);
 	return destroy;
 }
 
@@ -516,8 +562,8 @@ int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 		goto out;
 	}
 
-	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
-
+	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
+		UINT_MAX, UINT_MAX);
 	ret = rdma_connect(cm_id, &conn_param);
 	if (ret)
 		rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
@@ -601,9 +647,19 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
 				ic->i_cm_id, err);
 		}
 
+		/*
+		 * We want to wait for tx and rx completion to finish
+		 * before we tear down the connection, but we have to be
+		 * careful not to get stuck waiting on a send ring that
+		 * only has unsignaled sends in it.  We've shutdown new
+		 * sends before getting here so by waiting for signaled
+		 * sends to complete we're ensured that there will be no
+		 * more tx processing.
+		 */
 		wait_event(rds_ib_ring_empty_wait,
-			rds_ib_ring_empty(&ic->i_send_ring) &&
-			rds_ib_ring_empty(&ic->i_recv_ring));
+			   rds_ib_ring_empty(&ic->i_recv_ring) &&
+			   (atomic_read(&ic->i_signaled_sends) == 0));
+		tasklet_kill(&ic->i_recv_tasklet);
 
 		if (ic->i_send_hdrs)
 			ib_dma_free_coherent(dev,
@@ -654,9 +710,12 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
 	BUG_ON(ic->rds_ibdev);
 
 	/* Clear pending transmit */
-	if (ic->i_rm) {
-		rds_message_put(ic->i_rm);
-		ic->i_rm = NULL;
+	if (ic->i_data_op) {
+		struct rds_message *rm;
+
+		rm = container_of(ic->i_data_op, struct rds_message, data);
+		rds_message_put(rm);
+		ic->i_data_op = NULL;
 	}
 
 	/* Clear the ACK state */
@@ -690,12 +749,19 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 {
 	struct rds_ib_connection *ic;
 	unsigned long flags;
+	int ret;
 
 	/* XXX too lazy? */
 	ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
-	if (ic == NULL)
+	if (!ic)
 		return -ENOMEM;
 
+	ret = rds_ib_recv_alloc_caches(ic);
+	if (ret) {
+		kfree(ic);
+		return ret;
+	}
+
 	INIT_LIST_HEAD(&ic->ib_node);
 	tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
 		     (unsigned long) ic);
@@ -703,6 +769,7 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 #ifndef KERNEL_HAS_ATOMIC64
 	spin_lock_init(&ic->i_ack_lock);
 #endif
+	atomic_set(&ic->i_signaled_sends, 0);
 
 	/*
 	 * rds_ib_conn_shutdown() waits for these to be emptied so they
@@ -744,6 +811,8 @@ void rds_ib_conn_free(void *arg)
 	list_del(&ic->ib_node);
 	spin_unlock_irq(lock_ptr);
 
+	rds_ib_recv_free_caches(ic);
+
 	kfree(ic);
 }
 

net/rds/ib_rdma.c

@@ -32,11 +32,16 @@
  */
 #include <linux/kernel.h>
 #include <linux/slab.h>
+#include <linux/rculist.h>
 
 #include "rds.h"
-#include "rdma.h"
 #include "ib.h"
+#include "xlist.h"
 
+struct workqueue_struct *rds_ib_fmr_wq;
+
+static DEFINE_PER_CPU(unsigned long, clean_list_grace);
+#define CLEAN_LIST_BUSY_BIT 0
 
 /*
  * This is stored as mr->r_trans_private.
@@ -45,7 +50,11 @@ struct rds_ib_mr {
 	struct rds_ib_device	*device;
 	struct rds_ib_mr_pool	*pool;
 	struct ib_fmr		*fmr;
-	struct list_head	list;
+
+	struct xlist_head	xlist;
+
+	/* unmap_list is for freeing */
+	struct list_head	unmap_list;
 	unsigned int		remap_count;
 
 	struct scatterlist	*sg;
@@ -59,14 +68,16 @@ struct rds_ib_mr {
  */
 struct rds_ib_mr_pool {
 	struct mutex		flush_lock;		/* serialize fmr invalidate */
-	struct work_struct	flush_worker;		/* flush worker */
+	struct delayed_work	flush_worker;		/* flush worker */
 
-	spinlock_t		list_lock;		/* protect variables below */
 	atomic_t		item_count;		/* total # of MRs */
 	atomic_t		dirty_count;		/* # dirty of MRs */
-	struct list_head	drop_list;		/* MRs that have reached their max_maps limit */
-	struct list_head	free_list;		/* unused MRs */
-	struct list_head	clean_list;		/* unused & unamapped MRs */
+
+	struct xlist_head	drop_list;		/* MRs that have reached their max_maps limit */
+	struct xlist_head	free_list;		/* unused MRs */
+	struct xlist_head	clean_list;		/* global unused & unamapped MRs */
+	wait_queue_head_t	flush_wait;
+
 	atomic_t		free_pinned;		/* memory pinned by free MRs */
 	unsigned long		max_items;
 	unsigned long		max_items_soft;
@@ -74,7 +85,7 @@ struct rds_ib_mr_pool {
 	struct ib_fmr_attr	fmr_attr;
 };
 
-static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all);
+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);
 
@@ -83,16 +94,17 @@ static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 	struct rds_ib_device *rds_ibdev;
 	struct rds_ib_ipaddr *i_ipaddr;
 
-	list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
-		spin_lock_irq(&rds_ibdev->spinlock);
-		list_for_each_entry(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
+		list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 			if (i_ipaddr->ipaddr == ipaddr) {
-				spin_unlock_irq(&rds_ibdev->spinlock);
+				atomic_inc(&rds_ibdev->refcount);
+				rcu_read_unlock();
 				return rds_ibdev;
 			}
 		}
-		spin_unlock_irq(&rds_ibdev->spinlock);
 	}
+	rcu_read_unlock();
 
 	return NULL;
 }
@@ -108,7 +120,7 @@ static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 	i_ipaddr->ipaddr = ipaddr;
 
 	spin_lock_irq(&rds_ibdev->spinlock);
-	list_add_tail(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
+	list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
 	spin_unlock_irq(&rds_ibdev->spinlock);
 
 	return 0;
@@ -116,17 +128,24 @@ static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 
 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
 {
-	struct rds_ib_ipaddr *i_ipaddr, *next;
+	struct rds_ib_ipaddr *i_ipaddr;
+	struct rds_ib_ipaddr *to_free = NULL;
+
 
 	spin_lock_irq(&rds_ibdev->spinlock);
-	list_for_each_entry_safe(i_ipaddr, next, &rds_ibdev->ipaddr_list, list) {
+	list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
 		if (i_ipaddr->ipaddr == ipaddr) {
-			list_del(&i_ipaddr->list);
-			kfree(i_ipaddr);
+			list_del_rcu(&i_ipaddr->list);
+			to_free = i_ipaddr;
 			break;
 		}
 	}
 	spin_unlock_irq(&rds_ibdev->spinlock);
+
+	if (to_free) {
+		synchronize_rcu();
+		kfree(to_free);
+	}
 }
 
 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
@@ -134,8 +153,10 @@ 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) {
 		rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
+		rds_ib_dev_put(rds_ibdev_old);
+	}
 
 	return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
 }
@@ -156,6 +177,7 @@ void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *con
 	spin_unlock_irq(&ib_nodev_conns_lock);
 
 	ic->rds_ibdev = rds_ibdev;
+	atomic_inc(&rds_ibdev->refcount);
 }
 
 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
@@ -175,18 +197,18 @@ void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *
 	spin_unlock(&ib_nodev_conns_lock);
 
 	ic->rds_ibdev = NULL;
+	rds_ib_dev_put(rds_ibdev);
 }
 
-void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock)
+void rds_ib_destroy_nodev_conns(void)
 {
 	struct rds_ib_connection *ic, *_ic;
 	LIST_HEAD(tmp_list);
 
 	/* avoid calling conn_destroy with irqs off */
-	spin_lock_irq(list_lock);
-	list_splice(list, &tmp_list);
-	INIT_LIST_HEAD(list);
-	spin_unlock_irq(list_lock);
+	spin_lock_irq(&ib_nodev_conns_lock);
+	list_splice(&ib_nodev_conns, &tmp_list);
+	spin_unlock_irq(&ib_nodev_conns_lock);
 
 	list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
 		rds_conn_destroy(ic->conn);
@@ -200,12 +222,12 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev)
 	if (!pool)
 		return ERR_PTR(-ENOMEM);
 
-	INIT_LIST_HEAD(&pool->free_list);
-	INIT_LIST_HEAD(&pool->drop_list);
-	INIT_LIST_HEAD(&pool->clean_list);
+	INIT_XLIST_HEAD(&pool->free_list);
+	INIT_XLIST_HEAD(&pool->drop_list);
+	INIT_XLIST_HEAD(&pool->clean_list);
 	mutex_init(&pool->flush_lock);
-	spin_lock_init(&pool->list_lock);
-	INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
+	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;
 	pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
@@ -233,34 +255,60 @@ void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_co
 
 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
 {
-	flush_workqueue(rds_wq);
-	rds_ib_flush_mr_pool(pool, 1);
+	cancel_delayed_work_sync(&pool->flush_worker);
+	rds_ib_flush_mr_pool(pool, 1, NULL);
 	WARN_ON(atomic_read(&pool->item_count));
 	WARN_ON(atomic_read(&pool->free_pinned));
 	kfree(pool);
 }
 
+static void refill_local(struct rds_ib_mr_pool *pool, struct xlist_head *xl,
+			 struct rds_ib_mr **ibmr_ret)
+{
+	struct xlist_head *ibmr_xl;
+	ibmr_xl = xlist_del_head_fast(xl);
+	*ibmr_ret = list_entry(ibmr_xl, struct rds_ib_mr, xlist);
+}
+
 static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
 {
 	struct rds_ib_mr *ibmr = NULL;
-	unsigned long flags;
+	struct xlist_head *ret;
+	unsigned long *flag;
 
-	spin_lock_irqsave(&pool->list_lock, flags);
-	if (!list_empty(&pool->clean_list)) {
-		ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list);
-		list_del_init(&ibmr->list);
-	}
-	spin_unlock_irqrestore(&pool->list_lock, flags);
+	preempt_disable();
+	flag = &__get_cpu_var(clean_list_grace);
+	set_bit(CLEAN_LIST_BUSY_BIT, flag);
+	ret = xlist_del_head(&pool->clean_list);
+	if (ret)
+		ibmr = list_entry(ret, struct rds_ib_mr, xlist);
 
+	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();
+	}
+}
+
 static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 {
 	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
 	struct rds_ib_mr *ibmr = NULL;
 	int err = 0, iter = 0;
 
+	if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
+		queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
+
 	while (1) {
 		ibmr = rds_ib_reuse_fmr(pool);
 		if (ibmr)
@@ -287,19 +335,24 @@ static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
 
 		/* We do have some empty MRs. Flush them out. */
 		rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
-		rds_ib_flush_mr_pool(pool, 0);
+		rds_ib_flush_mr_pool(pool, 0, &ibmr);
+		if (ibmr)
+			return ibmr;
 	}
 
-	ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
+	ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
 	if (!ibmr) {
 		err = -ENOMEM;
 		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 |
-			 IB_ACCESS_REMOTE_WRITE),
+			 IB_ACCESS_REMOTE_WRITE|
+			 IB_ACCESS_REMOTE_ATOMIC),
 			&pool->fmr_attr);
 	if (IS_ERR(ibmr->fmr)) {
 		err = PTR_ERR(ibmr->fmr);
@@ -367,7 +420,8 @@ static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibm
 	if (page_cnt > fmr_message_size)
 		return -EINVAL;
 
-	dma_pages = kmalloc(sizeof(u64) * page_cnt, GFP_ATOMIC);
+	dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
+				 rdsibdev_to_node(rds_ibdev));
 	if (!dma_pages)
 		return -ENOMEM;
 
@@ -441,7 +495,7 @@ 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(in_interrupt());
+			BUG_ON(irqs_disabled());
 			set_page_dirty(page);
 			put_page(page);
 		}
@@ -476,34 +530,110 @@ static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int fr
 	return 0;
 }
 
+/*
+ * given an xlist of mrs, put them all into the list_head for more processing
+ */
+static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
+{
+	struct rds_ib_mr *ibmr;
+	struct xlist_head splice;
+	struct xlist_head *cur;
+	struct xlist_head *next;
+
+	splice.next = NULL;
+	xlist_splice(xlist, &splice);
+	cur = splice.next;
+	while (cur) {
+		next = cur->next;
+		ibmr = list_entry(cur, struct rds_ib_mr, xlist);
+		list_add_tail(&ibmr->unmap_list, list);
+		cur = next;
+	}
+}
+
+/*
+ * this takes a list head of mrs and turns it into an xlist of clusters.
+ * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
+ * reuse.
+ */
+static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
+				struct list_head *list, struct xlist_head *xlist,
+				struct xlist_head **tail_ret)
+{
+	struct rds_ib_mr *ibmr;
+	struct xlist_head *cur_mr = xlist;
+	struct xlist_head *tail_mr = NULL;
+
+	list_for_each_entry(ibmr, list, unmap_list) {
+		tail_mr = &ibmr->xlist;
+		tail_mr->next = NULL;
+		cur_mr->next = tail_mr;
+		cur_mr = tail_mr;
+	}
+	*tail_ret = tail_mr;
+}
+
 /*
  * Flush our pool of MRs.
  * At a minimum, all currently unused MRs are unmapped.
  * If the number of MRs allocated exceeds the limit, we also try
  * 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)
+static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
+			        int free_all, struct rds_ib_mr **ibmr_ret)
 {
 	struct rds_ib_mr *ibmr, *next;
+	struct xlist_head clean_xlist;
+	struct xlist_head *clean_tail;
 	LIST_HEAD(unmap_list);
 	LIST_HEAD(fmr_list);
 	unsigned long unpinned = 0;
-	unsigned long flags;
 	unsigned int nfreed = 0, ncleaned = 0, free_goal;
 	int ret = 0;
 
 	rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
 
-	mutex_lock(&pool->flush_lock);
+	if (ibmr_ret) {
+		DEFINE_WAIT(wait);
+		while(!mutex_trylock(&pool->flush_lock)) {
+			ibmr = rds_ib_reuse_fmr(pool);
+			if (ibmr) {
+				*ibmr_ret = ibmr;
+				finish_wait(&pool->flush_wait, &wait);
+				goto out_nolock;
+			}
+
+			prepare_to_wait(&pool->flush_wait, &wait,
+					TASK_UNINTERRUPTIBLE);
+			if (xlist_empty(&pool->clean_list))
+				schedule();
+
+			ibmr = rds_ib_reuse_fmr(pool);
+			if (ibmr) {
+				*ibmr_ret = ibmr;
+				finish_wait(&pool->flush_wait, &wait);
+				goto out_nolock;
+			}
+		}
+		finish_wait(&pool->flush_wait, &wait);
+	} else
+		mutex_lock(&pool->flush_lock);
+
+	if (ibmr_ret) {
+		ibmr = rds_ib_reuse_fmr(pool);
+		if (ibmr) {
+			*ibmr_ret = ibmr;
+			goto out;
+		}
+	}
 
-	spin_lock_irqsave(&pool->list_lock, flags);
 	/* Get the list of all MRs to be dropped. Ordering matters -
-	 * we want to put drop_list ahead of free_list. */
-	list_splice_init(&pool->free_list, &unmap_list);
-	list_splice_init(&pool->drop_list, &unmap_list);
+	 * we want to put drop_list ahead of free_list.
+	 */
+	xlist_append_to_list(&pool->drop_list, &unmap_list);
+	xlist_append_to_list(&pool->free_list, &unmap_list);
 	if (free_all)
-		list_splice_init(&pool->clean_list, &unmap_list);
-	spin_unlock_irqrestore(&pool->list_lock, flags);
+		xlist_append_to_list(&pool->clean_list, &unmap_list);
 
 	free_goal = rds_ib_flush_goal(pool, free_all);
 
@@ -511,19 +641,20 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
 		goto out;
 
 	/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
-	list_for_each_entry(ibmr, &unmap_list, list)
+	list_for_each_entry(ibmr, &unmap_list, unmap_list)
 		list_add(&ibmr->fmr->list, &fmr_list);
+
 	ret = ib_unmap_fmr(&fmr_list);
 	if (ret)
 		printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
 
 	/* Now we can destroy the DMA mapping and unpin any pages */
-	list_for_each_entry_safe(ibmr, next, &unmap_list, list) {
+	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);
-			list_del(&ibmr->list);
+			list_del(&ibmr->unmap_list);
 			ib_dealloc_fmr(ibmr->fmr);
 			kfree(ibmr);
 			nfreed++;
@@ -531,9 +662,27 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
 		ncleaned++;
 	}
 
-	spin_lock_irqsave(&pool->list_lock, flags);
-	list_splice(&unmap_list, &pool->clean_list);
-	spin_unlock_irqrestore(&pool->list_lock, flags);
+	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 xlist 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 xlist_del_head.
+		 *
+		 * This is pretty unlikely, but just in case  wait for an xlist grace period
+		 * here before adding anything back into the clean list.
+		 */
+		wait_clean_list_grace();
+
+		list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
+		if (ibmr_ret)
+			refill_local(pool, &clean_xlist, ibmr_ret);
+
+		/* refill_local may have emptied our list */
+		if (!xlist_empty(&clean_xlist))
+			xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
+
+	}
 
 	atomic_sub(unpinned, &pool->free_pinned);
 	atomic_sub(ncleaned, &pool->dirty_count);
@@ -541,14 +690,35 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
 
 out:
 	mutex_unlock(&pool->flush_lock);
+	if (waitqueue_active(&pool->flush_wait))
+		wake_up(&pool->flush_wait);
+out_nolock:
 	return ret;
 }
 
+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 void rds_ib_mr_pool_flush_worker(struct work_struct *work)
 {
-	struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker);
+	struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
 
-	rds_ib_flush_mr_pool(pool, 0);
+	rds_ib_flush_mr_pool(pool, 0, NULL);
 }
 
 void rds_ib_free_mr(void *trans_private, int invalidate)
@@ -556,47 +726,49 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
 	struct rds_ib_mr *ibmr = trans_private;
 	struct rds_ib_device *rds_ibdev = ibmr->device;
 	struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
-	unsigned long flags;
 
 	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 
 	/* Return it to the pool's free list */
-	spin_lock_irqsave(&pool->list_lock, flags);
 	if (ibmr->remap_count >= pool->fmr_attr.max_maps)
-		list_add(&ibmr->list, &pool->drop_list);
+		xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
 	else
-		list_add(&ibmr->list, &pool->free_list);
+		xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
 
 	atomic_add(ibmr->sg_len, &pool->free_pinned);
 	atomic_inc(&pool->dirty_count);
-	spin_unlock_irqrestore(&pool->list_lock, flags);
 
 	/* 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)
-		queue_work(rds_wq, &pool->flush_worker);
+		queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
 
 	if (invalidate) {
 		if (likely(!in_interrupt())) {
-			rds_ib_flush_mr_pool(pool, 0);
+			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. */
-			queue_work(rds_wq, &pool->flush_worker);
+			queue_delayed_work(rds_ib_fmr_wq,
+					   &pool->flush_worker, 10);
 		}
 	}
+
+	rds_ib_dev_put(rds_ibdev);
 }
 
 void rds_ib_flush_mrs(void)
 {
 	struct rds_ib_device *rds_ibdev;
 
+	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 (pool)
-			rds_ib_flush_mr_pool(pool, 0);
+			rds_ib_flush_mr_pool(pool, 0, NULL);
 	}
+	up_read(&rds_ib_devices_lock);
 }
 
 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
@@ -628,6 +800,7 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
 
 	ibmr->device = rds_ibdev;
+	rds_ibdev = NULL;
 
  out:
 	if (ret) {
@@ -635,5 +808,8 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 			rds_ib_free_mr(ibmr, 0);
 		ibmr = ERR_PTR(ret);
 	}
+	if (rds_ibdev)
+		rds_ib_dev_put(rds_ibdev);
 	return ibmr;
 }
+

net/rds/ib_recv.c

@@ -43,42 +43,6 @@ static struct kmem_cache *rds_ib_incoming_slab;
 static struct kmem_cache *rds_ib_frag_slab;
 static atomic_t	rds_ib_allocation = ATOMIC_INIT(0);
 
-static void rds_ib_frag_drop_page(struct rds_page_frag *frag)
-{
-	rdsdebug("frag %p page %p\n", frag, frag->f_page);
-	__free_page(frag->f_page);
-	frag->f_page = NULL;
-}
-
-static void rds_ib_frag_free(struct rds_page_frag *frag)
-{
-	rdsdebug("frag %p page %p\n", frag, frag->f_page);
-	BUG_ON(frag->f_page != NULL);
-	kmem_cache_free(rds_ib_frag_slab, frag);
-}
-
-/*
- * We map a page at a time.  Its fragments are posted in order.  This
- * is called in fragment order as the fragments get send completion events.
- * Only the last frag in the page performs the unmapping.
- *
- * It's OK for ring cleanup to call this in whatever order it likes because
- * DMA is not in flight and so we can unmap while other ring entries still
- * hold page references in their frags.
- */
-static void rds_ib_recv_unmap_page(struct rds_ib_connection *ic,
-				   struct rds_ib_recv_work *recv)
-{
-	struct rds_page_frag *frag = recv->r_frag;
-
-	rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page);
-	if (frag->f_mapped)
-		ib_dma_unmap_page(ic->i_cm_id->device,
-			       frag->f_mapped,
-			       RDS_FRAG_SIZE, DMA_FROM_DEVICE);
-	frag->f_mapped = 0;
-}
-
 void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
 {
 	struct rds_ib_recv_work *recv;
@@ -95,16 +59,161 @@ void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
 		recv->r_wr.sg_list = recv->r_sge;
 		recv->r_wr.num_sge = RDS_IB_RECV_SGE;
 
-		sge = rds_ib_data_sge(ic, recv->r_sge);
+		sge = &recv->r_sge[0];
+		sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
+		sge->length = sizeof(struct rds_header);
+		sge->lkey = ic->i_mr->lkey;
+
+		sge = &recv->r_sge[1];
 		sge->addr = 0;
 		sge->length = RDS_FRAG_SIZE;
 		sge->lkey = ic->i_mr->lkey;
+	}
+}
 
-		sge = rds_ib_header_sge(ic, recv->r_sge);
-		sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
-		sge->length = sizeof(struct rds_header);
-		sge->lkey = ic->i_mr->lkey;
+/*
+ * The entire 'from' list, including the from element itself, is put on
+ * to the tail of the 'to' list.
+ */
+static void list_splice_entire_tail(struct list_head *from,
+				    struct list_head *to)
+{
+	struct list_head *from_last = from->prev;
+
+	list_splice_tail(from_last, to);
+	list_add_tail(from_last, to);
+}
+
+static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
+{
+	struct list_head *tmp;
+
+	tmp = xchg(&cache->xfer, NULL);
+	if (tmp) {
+		if (cache->ready)
+			list_splice_entire_tail(tmp, cache->ready);
+		else
+			cache->ready = tmp;
+	}
+}
+
+static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
+{
+	struct rds_ib_cache_head *head;
+	int cpu;
+
+	cache->percpu = alloc_percpu(struct rds_ib_cache_head);
+	if (!cache->percpu)
+	       return -ENOMEM;
+
+	for_each_possible_cpu(cpu) {
+		head = per_cpu_ptr(cache->percpu, cpu);
+		head->first = NULL;
+		head->count = 0;
+	}
+	cache->xfer = NULL;
+	cache->ready = NULL;
+
+	return 0;
+}
+
+int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
+{
+	int ret;
+
+	ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
+	if (!ret) {
+		ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
+		if (ret)
+			free_percpu(ic->i_cache_incs.percpu);
 	}
+
+	return ret;
+}
+
+static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
+					  struct list_head *caller_list)
+{
+	struct rds_ib_cache_head *head;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		head = per_cpu_ptr(cache->percpu, cpu);
+		if (head->first) {
+			list_splice_entire_tail(head->first, caller_list);
+			head->first = NULL;
+		}
+	}
+
+	if (cache->ready) {
+		list_splice_entire_tail(cache->ready, caller_list);
+		cache->ready = NULL;
+	}
+}
+
+void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
+{
+	struct rds_ib_incoming *inc;
+	struct rds_ib_incoming *inc_tmp;
+	struct rds_page_frag *frag;
+	struct rds_page_frag *frag_tmp;
+	LIST_HEAD(list);
+
+	rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
+	rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
+	free_percpu(ic->i_cache_incs.percpu);
+
+	list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
+		list_del(&inc->ii_cache_entry);
+		WARN_ON(!list_empty(&inc->ii_frags));
+		kmem_cache_free(rds_ib_incoming_slab, inc);
+	}
+
+	rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
+	rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
+	free_percpu(ic->i_cache_frags.percpu);
+
+	list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
+		list_del(&frag->f_cache_entry);
+		WARN_ON(!list_empty(&frag->f_item));
+		kmem_cache_free(rds_ib_frag_slab, frag);
+	}
+}
+
+/* fwd decl */
+static void rds_ib_recv_cache_put(struct list_head *new_item,
+				  struct rds_ib_refill_cache *cache);
+static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
+
+
+/* Recycle frag and attached recv buffer f_sg */
+static void rds_ib_frag_free(struct rds_ib_connection *ic,
+			     struct rds_page_frag *frag)
+{
+	rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
+
+	rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
+}
+
+/* Recycle inc after freeing attached frags */
+void rds_ib_inc_free(struct rds_incoming *inc)
+{
+	struct rds_ib_incoming *ibinc;
+	struct rds_page_frag *frag;
+	struct rds_page_frag *pos;
+	struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
+
+	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
+
+	/* Free attached frags */
+	list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
+		list_del_init(&frag->f_item);
+		rds_ib_frag_free(ic, frag);
+	}
+	BUG_ON(!list_empty(&ibinc->ii_frags));
+
+	rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
+	rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
 }
 
 static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
@@ -115,10 +224,8 @@ static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
 		recv->r_ibinc = NULL;
 	}
 	if (recv->r_frag) {
-		rds_ib_recv_unmap_page(ic, recv);
-		if (recv->r_frag->f_page)
-			rds_ib_frag_drop_page(recv->r_frag);
-		rds_ib_frag_free(recv->r_frag);
+		ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
+		rds_ib_frag_free(ic, recv->r_frag);
 		recv->r_frag = NULL;
 	}
 }
@@ -129,84 +236,111 @@ void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
 
 	for (i = 0; i < ic->i_recv_ring.w_nr; i++)
 		rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
-
-	if (ic->i_frag.f_page)
-		rds_ib_frag_drop_page(&ic->i_frag);
 }
 
-static int rds_ib_recv_refill_one(struct rds_connection *conn,
-				  struct rds_ib_recv_work *recv,
-				  gfp_t kptr_gfp, gfp_t page_gfp)
+static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
+						     gfp_t slab_mask)
 {
-	struct rds_ib_connection *ic = conn->c_transport_data;
-	dma_addr_t dma_addr;
-	struct ib_sge *sge;
-	int ret = -ENOMEM;
+	struct rds_ib_incoming *ibinc;
+	struct list_head *cache_item;
+	int avail_allocs;
 
-	if (recv->r_ibinc == NULL) {
-		if (!atomic_add_unless(&rds_ib_allocation, 1, rds_ib_sysctl_max_recv_allocation)) {
+	cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
+	if (cache_item) {
+		ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
+	} else {
+		avail_allocs = atomic_add_unless(&rds_ib_allocation,
+						 1, rds_ib_sysctl_max_recv_allocation);
+		if (!avail_allocs) {
 			rds_ib_stats_inc(s_ib_rx_alloc_limit);
-			goto out;
+			return NULL;
 		}
-		recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab,
-						 kptr_gfp);
-		if (recv->r_ibinc == NULL) {
+		ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
+		if (!ibinc) {
 			atomic_dec(&rds_ib_allocation);
-			goto out;
+			return NULL;
 		}
-		INIT_LIST_HEAD(&recv->r_ibinc->ii_frags);
-		rds_inc_init(&recv->r_ibinc->ii_inc, conn, conn->c_faddr);
 	}
+	INIT_LIST_HEAD(&ibinc->ii_frags);
+	rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
 
-	if (recv->r_frag == NULL) {
-		recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp);
-		if (recv->r_frag == NULL)
-			goto out;
-		INIT_LIST_HEAD(&recv->r_frag->f_item);
-		recv->r_frag->f_page = NULL;
+	return ibinc;
+}
+
+static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
+						    gfp_t slab_mask, gfp_t page_mask)
+{
+	struct rds_page_frag *frag;
+	struct list_head *cache_item;
+	int ret;
+
+	cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
+	if (cache_item) {
+		frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
+	} else {
+		frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
+		if (!frag)
+			return NULL;
+
+		sg_init_table(&frag->f_sg, 1);
+		ret = rds_page_remainder_alloc(&frag->f_sg,
+					       RDS_FRAG_SIZE, page_mask);
+		if (ret) {
+			kmem_cache_free(rds_ib_frag_slab, frag);
+			return NULL;
+		}
 	}
 
-	if (ic->i_frag.f_page == NULL) {
-		ic->i_frag.f_page = alloc_page(page_gfp);
-		if (ic->i_frag.f_page == NULL)
-			goto out;
-		ic->i_frag.f_offset = 0;
+	INIT_LIST_HEAD(&frag->f_item);
+
+	return frag;
+}
+
+static int rds_ib_recv_refill_one(struct rds_connection *conn,
+				  struct rds_ib_recv_work *recv, int prefill)
+{
+	struct rds_ib_connection *ic = conn->c_transport_data;
+	struct ib_sge *sge;
+	int ret = -ENOMEM;
+	gfp_t slab_mask = GFP_NOWAIT;
+	gfp_t page_mask = GFP_NOWAIT;
+
+	if (prefill) {
+		slab_mask = GFP_KERNEL;
+		page_mask = GFP_HIGHUSER;
 	}
 
-	dma_addr = ib_dma_map_page(ic->i_cm_id->device,
-				  ic->i_frag.f_page,
-				  ic->i_frag.f_offset,
-				  RDS_FRAG_SIZE,
-				  DMA_FROM_DEVICE);
-	if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr))
-		goto out;
+	if (!ic->i_cache_incs.ready)
+		rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
+	if (!ic->i_cache_frags.ready)
+		rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
 
 	/*
-	 * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap()
-	 * must be called on this recv.  This happens as completions hit
-	 * in order or on connection shutdown.
+	 * ibinc was taken from recv if recv contained the start of a message.
+	 * recvs that were continuations will still have this allocated.
 	 */
-	recv->r_frag->f_page = ic->i_frag.f_page;
-	recv->r_frag->f_offset = ic->i_frag.f_offset;
-	recv->r_frag->f_mapped = dma_addr;
+	if (!recv->r_ibinc) {
+		recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
+		if (!recv->r_ibinc)
+			goto out;
+	}
 
-	sge = rds_ib_data_sge(ic, recv->r_sge);
-	sge->addr = dma_addr;
-	sge->length = RDS_FRAG_SIZE;
+	WARN_ON(recv->r_frag); /* leak! */
+	recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
+	if (!recv->r_frag)
+		goto out;
+
+	ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
+			    1, DMA_FROM_DEVICE);
+	WARN_ON(ret != 1);
 
-	sge = rds_ib_header_sge(ic, recv->r_sge);
+	sge = &recv->r_sge[0];
 	sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
 	sge->length = sizeof(struct rds_header);
 
-	get_page(recv->r_frag->f_page);
-
-	if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) {
-		ic->i_frag.f_offset += RDS_FRAG_SIZE;
-	} else {
-		put_page(ic->i_frag.f_page);
-		ic->i_frag.f_page = NULL;
-		ic->i_frag.f_offset = 0;
-	}
+	sge = &recv->r_sge[1];
+	sge->addr = sg_dma_address(&recv->r_frag->f_sg);
+	sge->length = sg_dma_len(&recv->r_frag->f_sg);
 
 	ret = 0;
 out:
@@ -216,13 +350,11 @@ out:
 /*
  * This tries to allocate and post unused work requests after making sure that
  * they have all the allocations they need to queue received fragments into
- * sockets.  The i_recv_mutex is held here so that ring_alloc and _unalloc
- * pairs don't go unmatched.
+ * sockets.
  *
  * -1 is returned if posting fails due to temporary resource exhaustion.
  */
-int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
-		       gfp_t page_gfp, int prefill)
+void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct rds_ib_recv_work *recv;
@@ -236,28 +368,25 @@ int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
 		if (pos >= ic->i_recv_ring.w_nr) {
 			printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
 					pos);
-			ret = -EINVAL;
 			break;
 		}
 
 		recv = &ic->i_recvs[pos];
-		ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp);
+		ret = rds_ib_recv_refill_one(conn, recv, prefill);
 		if (ret) {
-			ret = -1;
 			break;
 		}
 
 		/* XXX when can this fail? */
 		ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
 		rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
-			 recv->r_ibinc, recv->r_frag->f_page,
-			 (long) recv->r_frag->f_mapped, ret);
+			 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
+			 (long) sg_dma_address(&recv->r_frag->f_sg), ret);
 		if (ret) {
 			rds_ib_conn_error(conn, "recv post on "
 			       "%pI4 returned %d, disconnecting and "
 			       "reconnecting\n", &conn->c_faddr,
 			       ret);
-			ret = -1;
 			break;
 		}
 
@@ -270,37 +399,73 @@ int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
 
 	if (ret)
 		rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
-	return ret;
 }
 
-void rds_ib_inc_purge(struct rds_incoming *inc)
+/*
+ * We want to recycle several types of recv allocations, like incs and frags.
+ * To use this, the *_free() function passes in the ptr to a list_head within
+ * the recyclee, as well as the cache to put it on.
+ *
+ * First, we put the memory on a percpu list. When this reaches a certain size,
+ * We move it to an intermediate non-percpu list in a lockless manner, with some
+ * xchg/compxchg wizardry.
+ *
+ * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
+ * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
+ * list_empty() will return true with one element is actually present.
+ */
+static void rds_ib_recv_cache_put(struct list_head *new_item,
+				 struct rds_ib_refill_cache *cache)
 {
-	struct rds_ib_incoming *ibinc;
-	struct rds_page_frag *frag;
-	struct rds_page_frag *pos;
+	unsigned long flags;
+	struct rds_ib_cache_head *chp;
+	struct list_head *old;
 
-	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
-	rdsdebug("purging ibinc %p inc %p\n", ibinc, inc);
+	local_irq_save(flags);
 
-	list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
-		list_del_init(&frag->f_item);
-		rds_ib_frag_drop_page(frag);
-		rds_ib_frag_free(frag);
-	}
+	chp = per_cpu_ptr(cache->percpu, smp_processor_id());
+	if (!chp->first)
+		INIT_LIST_HEAD(new_item);
+	else /* put on front */
+		list_add_tail(new_item, chp->first);
+	chp->first = new_item;
+	chp->count++;
+
+	if (chp->count < RDS_IB_RECYCLE_BATCH_COUNT)
+		goto end;
+
+	/*
+	 * Return our per-cpu first list to the cache's xfer by atomically
+	 * grabbing the current xfer list, appending it to our per-cpu list,
+	 * and then atomically returning that entire list back to the
+	 * cache's xfer list as long as it's still empty.
+	 */
+	do {
+		old = xchg(&cache->xfer, NULL);
+		if (old)
+			list_splice_entire_tail(old, chp->first);
+		old = cmpxchg(&cache->xfer, NULL, chp->first);
+	} while (old);
+
+	chp->first = NULL;
+	chp->count = 0;
+end:
+	local_irq_restore(flags);
 }
 
-void rds_ib_inc_free(struct rds_incoming *inc)
+static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
 {
-	struct rds_ib_incoming *ibinc;
-
-	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
+	struct list_head *head = cache->ready;
+
+	if (head) {
+		if (!list_empty(head)) {
+			cache->ready = head->next;
+			list_del_init(head);
+		} else
+			cache->ready = NULL;
+	}
 
-	rds_ib_inc_purge(inc);
-	rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
-	BUG_ON(!list_empty(&ibinc->ii_frags));
-	kmem_cache_free(rds_ib_incoming_slab, ibinc);
-	atomic_dec(&rds_ib_allocation);
-	BUG_ON(atomic_read(&rds_ib_allocation) < 0);
+	return head;
 }
 
 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
@@ -336,13 +501,13 @@ int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
 		to_copy = min_t(unsigned long, to_copy, len - copied);
 
 		rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
-			 "[%p, %lu] + %lu\n",
+			 "[%p, %u] + %lu\n",
 			 to_copy, iov->iov_base, iov->iov_len, iov_off,
-			 frag->f_page, frag->f_offset, frag_off);
+			 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);
 
 		/* XXX needs + offset for multiple recvs per page */
-		ret = rds_page_copy_to_user(frag->f_page,
-					    frag->f_offset + frag_off,
+		ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
+					    frag->f_sg.offset + frag_off,
 					    iov->iov_base + iov_off,
 					    to_copy);
 		if (ret) {
@@ -557,47 +722,6 @@ u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
 	return rds_ib_get_ack(ic);
 }
 
-static struct rds_header *rds_ib_get_header(struct rds_connection *conn,
-					    struct rds_ib_recv_work *recv,
-					    u32 data_len)
-{
-	struct rds_ib_connection *ic = conn->c_transport_data;
-	void *hdr_buff = &ic->i_recv_hdrs[recv - ic->i_recvs];
-	void *addr;
-	u32 misplaced_hdr_bytes;
-
-	/*
-	 * Support header at the front (RDS 3.1+) as well as header-at-end.
-	 *
-	 * Cases:
-	 * 1) header all in header buff (great!)
-	 * 2) header all in data page (copy all to header buff)
-	 * 3) header split across hdr buf + data page
-	 *    (move bit in hdr buff to end before copying other bit from data page)
-	 */
-	if (conn->c_version > RDS_PROTOCOL_3_0 || data_len == RDS_FRAG_SIZE)
-	        return hdr_buff;
-
-	if (data_len <= (RDS_FRAG_SIZE - sizeof(struct rds_header))) {
-		addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0);
-		memcpy(hdr_buff,
-		       addr + recv->r_frag->f_offset + data_len,
-		       sizeof(struct rds_header));
-		kunmap_atomic(addr, KM_SOFTIRQ0);
-		return hdr_buff;
-	}
-
-	misplaced_hdr_bytes = (sizeof(struct rds_header) - (RDS_FRAG_SIZE - data_len));
-
-	memmove(hdr_buff + misplaced_hdr_bytes, hdr_buff, misplaced_hdr_bytes);
-
-	addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0);
-	memcpy(hdr_buff, addr + recv->r_frag->f_offset + data_len,
-	       sizeof(struct rds_header) - misplaced_hdr_bytes);
-	kunmap_atomic(addr, KM_SOFTIRQ0);
-	return hdr_buff;
-}
-
 /*
  * It's kind of lame that we're copying from the posted receive pages into
  * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
@@ -639,7 +763,7 @@ static void rds_ib_cong_recv(struct rds_connection *conn,
 		to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
 		BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
 
-		addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0);
+		addr = kmap_atomic(sg_page(&frag->f_sg), KM_SOFTIRQ0);
 
 		src = addr + frag_off;
 		dst = (void *)map->m_page_addrs[map_page] + map_off;
@@ -710,7 +834,7 @@ static void rds_ib_process_recv(struct rds_connection *conn,
 	}
 	data_len -= sizeof(struct rds_header);
 
-	ihdr = rds_ib_get_header(conn, recv, data_len);
+	ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
 
 	/* Validate the checksum. */
 	if (!rds_message_verify_checksum(ihdr)) {
@@ -742,12 +866,12 @@ static void rds_ib_process_recv(struct rds_connection *conn,
 		 * the inc is freed.  We don't go that route, so we have to drop the
 		 * page ref ourselves.  We can't just leave the page on the recv
 		 * because that confuses the dma mapping of pages and each recv's use
-		 * of a partial page.  We can leave the frag, though, it will be
-		 * reused.
+		 * of a partial page.
 		 *
 		 * FIXME: Fold this into the code path below.
 		 */
-		rds_ib_frag_drop_page(recv->r_frag);
+		rds_ib_frag_free(ic, recv->r_frag);
+		recv->r_frag = NULL;
 		return;
 	}
 
@@ -757,7 +881,7 @@ static void rds_ib_process_recv(struct rds_connection *conn,
 	 * into the inc and save the inc so we can hang upcoming fragments
 	 * off its list.
 	 */
-	if (ibinc == NULL) {
+	if (!ibinc) {
 		ibinc = recv->r_ibinc;
 		recv->r_ibinc = NULL;
 		ic->i_ibinc = ibinc;
@@ -842,32 +966,38 @@ static inline void rds_poll_cq(struct rds_ib_connection *ic,
 	struct rds_ib_recv_work *recv;
 
 	while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
-		rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
-			 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
+		rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
+			 (unsigned long long)wc.wr_id, wc.status,
+			 rds_ib_wc_status_str(wc.status), wc.byte_len,
 			 be32_to_cpu(wc.ex.imm_data));
 		rds_ib_stats_inc(s_ib_rx_cq_event);
 
 		recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
 
-		rds_ib_recv_unmap_page(ic, recv);
+		ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
 
 		/*
 		 * Also process recvs in connecting state because it is possible
 		 * to get a recv completion _before_ the rdmacm ESTABLISHED
 		 * event is processed.
 		 */
-		if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
+		if (wc.status == IB_WC_SUCCESS) {
+			rds_ib_process_recv(conn, recv, wc.byte_len, state);
+		} else {
 			/* We expect errors as the qp is drained during shutdown */
-			if (wc.status == IB_WC_SUCCESS) {
-				rds_ib_process_recv(conn, recv, wc.byte_len, state);
-			} else {
-				rds_ib_conn_error(conn, "recv completion on "
-				       "%pI4 had status %u, disconnecting and "
-				       "reconnecting\n", &conn->c_faddr,
-				       wc.status);
-			}
+			if (rds_conn_up(conn) || rds_conn_connecting(conn))
+				rds_ib_conn_error(conn, "recv completion on %pI4 had "
+						  "status %u (%s), disconnecting and "
+						  "reconnecting\n", &conn->c_faddr,
+						  wc.status,
+						  rds_ib_wc_status_str(wc.status));
 		}
 
+		/*
+		 * It's very important that we only free this ring entry if we've truly
+		 * freed the resources allocated to the entry.  The refilling path can
+		 * leak if we don't.
+		 */
 		rds_ib_ring_free(&ic->i_recv_ring, 1);
 	}
 }
@@ -897,11 +1027,8 @@ void rds_ib_recv_tasklet_fn(unsigned long data)
 	if (rds_ib_ring_empty(&ic->i_recv_ring))
 		rds_ib_stats_inc(s_ib_rx_ring_empty);
 
-	/*
-	 * If the ring is running low, then schedule the thread to refill.
-	 */
 	if (rds_ib_ring_low(&ic->i_recv_ring))
-		queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
+		rds_ib_recv_refill(conn, 0);
 }
 
 int rds_ib_recv(struct rds_connection *conn)
@@ -910,25 +1037,13 @@ int rds_ib_recv(struct rds_connection *conn)
 	int ret = 0;
 
 	rdsdebug("conn %p\n", conn);
-
-	/*
-	 * If we get a temporary posting failure in this context then
-	 * we're really low and we want the caller to back off for a bit.
-	 */
-	mutex_lock(&ic->i_recv_mutex);
-	if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0))
-		ret = -ENOMEM;
-	else
-		rds_ib_stats_inc(s_ib_rx_refill_from_thread);
-	mutex_unlock(&ic->i_recv_mutex);
-
 	if (rds_conn_up(conn))
 		rds_ib_attempt_ack(ic);
 
 	return ret;
 }
 
-int __init rds_ib_recv_init(void)
+int rds_ib_recv_init(void)
 {
 	struct sysinfo si;
 	int ret = -ENOMEM;
@@ -939,14 +1054,14 @@ int __init rds_ib_recv_init(void)
 
 	rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
 					sizeof(struct rds_ib_incoming),
-					0, 0, NULL);
-	if (rds_ib_incoming_slab == NULL)
+					0, SLAB_HWCACHE_ALIGN, NULL);
+	if (!rds_ib_incoming_slab)
 		goto out;
 
 	rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
 					sizeof(struct rds_page_frag),
-					0, 0, NULL);
-	if (rds_ib_frag_slab == NULL)
+					0, SLAB_HWCACHE_ALIGN, NULL);
+	if (!rds_ib_frag_slab)
 		kmem_cache_destroy(rds_ib_incoming_slab);
 	else
 		ret = 0;

net/rds/ib_send.c

@@ -36,11 +36,49 @@
 #include <linux/dmapool.h>
 
 #include "rds.h"
-#include "rdma.h"
 #include "ib.h"
 
-static void rds_ib_send_rdma_complete(struct rds_message *rm,
-				      int wc_status)
+static char *rds_ib_wc_status_strings[] = {
+#define RDS_IB_WC_STATUS_STR(foo) \
+		[IB_WC_##foo] = __stringify(IB_WC_##foo)
+	RDS_IB_WC_STATUS_STR(SUCCESS),
+	RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
+	RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
+	RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
+	RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
+	RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
+	RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
+	RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
+	RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
+	RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
+	RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
+	RDS_IB_WC_STATUS_STR(REM_OP_ERR),
+	RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
+	RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
+	RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
+	RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
+	RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
+	RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
+	RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
+	RDS_IB_WC_STATUS_STR(FATAL_ERR),
+	RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
+	RDS_IB_WC_STATUS_STR(GENERAL_ERR),
+#undef RDS_IB_WC_STATUS_STR
+};
+
+char *rds_ib_wc_status_str(enum ib_wc_status status)
+{
+	return rds_str_array(rds_ib_wc_status_strings,
+			     ARRAY_SIZE(rds_ib_wc_status_strings), status);
+}
+
+/*
+ * Convert IB-specific error message to RDS error message and call core
+ * completion handler.
+ */
+static void rds_ib_send_complete(struct rds_message *rm,
+				 int wc_status,
+				 void (*complete)(struct rds_message *rm, int status))
 {
 	int notify_status;
 
@@ -60,69 +98,125 @@ static void rds_ib_send_rdma_complete(struct rds_message *rm,
 		notify_status = RDS_RDMA_OTHER_ERROR;
 		break;
 	}
-	rds_rdma_send_complete(rm, notify_status);
+	complete(rm, notify_status);
+}
+
+static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
+				   struct rm_data_op *op,
+				   int wc_status)
+{
+	if (op->op_nents)
+		ib_dma_unmap_sg(ic->i_cm_id->device,
+				op->op_sg, op->op_nents,
+				DMA_TO_DEVICE);
 }
 
 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
-				   struct rds_rdma_op *op)
+				   struct rm_rdma_op *op,
+				   int wc_status)
 {
-	if (op->r_mapped) {
+	if (op->op_mapped) {
 		ib_dma_unmap_sg(ic->i_cm_id->device,
-			op->r_sg, op->r_nents,
-			op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
-		op->r_mapped = 0;
+				op->op_sg, op->op_nents,
+				op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		op->op_mapped = 0;
 	}
+
+	/* If the user asked for a completion notification on this
+	 * message, we can implement three different semantics:
+	 *  1.	Notify when we received the ACK on the RDS message
+	 *	that was queued with the RDMA. This provides reliable
+	 *	notification of RDMA status at the expense of a one-way
+	 *	packet delay.
+	 *  2.	Notify when the IB stack gives us the completion event for
+	 *	the RDMA operation.
+	 *  3.	Notify when the IB stack gives us the completion event for
+	 *	the accompanying RDS messages.
+	 * Here, we implement approach #3. To implement approach #2,
+	 * we would need to take an event for the rdma WR. To implement #1,
+	 * don't call rds_rdma_send_complete at all, and fall back to the notify
+	 * handling in the ACK processing code.
+	 *
+	 * Note: There's no need to explicitly sync any RDMA buffers using
+	 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
+	 * operation itself unmapped the RDMA buffers, which takes care
+	 * of synching.
+	 */
+	rds_ib_send_complete(container_of(op, struct rds_message, rdma),
+			     wc_status, rds_rdma_send_complete);
+
+	if (op->op_write)
+		rds_stats_add(s_send_rdma_bytes, op->op_bytes);
+	else
+		rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
 }
 
-static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
-			  struct rds_ib_send_work *send,
-			  int wc_status)
+static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
+				     struct rm_atomic_op *op,
+				     int wc_status)
 {
-	struct rds_message *rm = send->s_rm;
-
-	rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
-
-	ib_dma_unmap_sg(ic->i_cm_id->device,
-		     rm->m_sg, rm->m_nents,
-		     DMA_TO_DEVICE);
-
-	if (rm->m_rdma_op != NULL) {
-		rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
-
-		/* If the user asked for a completion notification on this
-		 * message, we can implement three different semantics:
-		 *  1.	Notify when we received the ACK on the RDS message
-		 *	that was queued with the RDMA. This provides reliable
-		 *	notification of RDMA status at the expense of a one-way
-		 *	packet delay.
-		 *  2.	Notify when the IB stack gives us the completion event for
-		 *	the RDMA operation.
-		 *  3.	Notify when the IB stack gives us the completion event for
-		 *	the accompanying RDS messages.
-		 * Here, we implement approach #3. To implement approach #2,
-		 * call rds_rdma_send_complete from the cq_handler. To implement #1,
-		 * don't call rds_rdma_send_complete at all, and fall back to the notify
-		 * handling in the ACK processing code.
-		 *
-		 * Note: There's no need to explicitly sync any RDMA buffers using
-		 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
-		 * operation itself unmapped the RDMA buffers, which takes care
-		 * of synching.
-		 */
-		rds_ib_send_rdma_complete(rm, wc_status);
+	/* unmap atomic recvbuf */
+	if (op->op_mapped) {
+		ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
+				DMA_FROM_DEVICE);
+		op->op_mapped = 0;
+	}
 
-		if (rm->m_rdma_op->r_write)
-			rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
-		else
-			rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
+	rds_ib_send_complete(container_of(op, struct rds_message, atomic),
+			     wc_status, rds_atomic_send_complete);
+
+	if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
+		rds_ib_stats_inc(s_ib_atomic_cswp);
+	else
+		rds_ib_stats_inc(s_ib_atomic_fadd);
+}
+
+/*
+ * Unmap the resources associated with a struct send_work.
+ *
+ * Returns the rm for no good reason other than it is unobtainable
+ * other than by switching on wr.opcode, currently, and the caller,
+ * the event handler, needs it.
+ */
+static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
+						struct rds_ib_send_work *send,
+						int wc_status)
+{
+	struct rds_message *rm = NULL;
+
+	/* In the error case, wc.opcode sometimes contains garbage */
+	switch (send->s_wr.opcode) {
+	case IB_WR_SEND:
+		if (send->s_op) {
+			rm = container_of(send->s_op, struct rds_message, data);
+			rds_ib_send_unmap_data(ic, send->s_op, wc_status);
+		}
+		break;
+	case IB_WR_RDMA_WRITE:
+	case IB_WR_RDMA_READ:
+		if (send->s_op) {
+			rm = container_of(send->s_op, struct rds_message, rdma);
+			rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
+		}
+		break;
+	case IB_WR_ATOMIC_FETCH_AND_ADD:
+	case IB_WR_ATOMIC_CMP_AND_SWP:
+		if (send->s_op) {
+			rm = container_of(send->s_op, struct rds_message, atomic);
+			rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
+		}
+		break;
+	default:
+		if (printk_ratelimit())
+			printk(KERN_NOTICE
+			       "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
+			       __func__, send->s_wr.opcode);
+		break;
 	}
 
-	/* If anyone waited for this message to get flushed out, wake
-	 * them up now */
-	rds_message_unmapped(rm);
+	send->s_wr.opcode = 0xdead;
 
-	rds_message_put(rm);
-	send->s_rm = NULL;
+	return rm;
 }
 
 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
@@ -133,23 +227,18 @@ void rds_ib_send_init_ring(struct rds_ib_connection *ic)
 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
 		struct ib_sge *sge;
 
-		send->s_rm = NULL;
 		send->s_op = NULL;
 
 		send->s_wr.wr_id = i;
 		send->s_wr.sg_list = send->s_sge;
-		send->s_wr.num_sge = 1;
-		send->s_wr.opcode = IB_WR_SEND;
-		send->s_wr.send_flags = 0;
 		send->s_wr.ex.imm_data = 0;
 
-		sge = rds_ib_data_sge(ic, send->s_sge);
-		sge->lkey = ic->i_mr->lkey;
-
-		sge = rds_ib_header_sge(ic, send->s_sge);
+		sge = &send->s_sge[0];
 		sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
 		sge->length = sizeof(struct rds_header);
 		sge->lkey = ic->i_mr->lkey;
+
+		send->s_sge[1].lkey = ic->i_mr->lkey;
 	}
 }
 
@@ -159,15 +248,23 @@ void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
 	u32 i;
 
 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
-		if (send->s_wr.opcode == 0xdead)
-			continue;
-		if (send->s_rm)
-			rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
-		if (send->s_op)
-			rds_ib_send_unmap_rdma(ic, send->s_op);
+		if (send->s_op && send->s_wr.opcode != 0xdead)
+			rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
 	}
 }
 
+/*
+ * The only fast path caller always has a non-zero nr, so we don't
+ * bother testing nr before performing the atomic sub.
+ */
+static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
+{
+	if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
+	    waitqueue_active(&rds_ib_ring_empty_wait))
+		wake_up(&rds_ib_ring_empty_wait);
+	BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
+}
+
 /*
  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
  * operations performed in the send path.  As the sender allocs and potentially
@@ -178,12 +275,14 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 {
 	struct rds_connection *conn = context;
 	struct rds_ib_connection *ic = conn->c_transport_data;
+	struct rds_message *rm = NULL;
 	struct ib_wc wc;
 	struct rds_ib_send_work *send;
 	u32 completed;
 	u32 oldest;
 	u32 i = 0;
 	int ret;
+	int nr_sig = 0;
 
 	rdsdebug("cq %p conn %p\n", cq, conn);
 	rds_ib_stats_inc(s_ib_tx_cq_call);
@@ -192,8 +291,9 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 
 	while (ib_poll_cq(cq, 1, &wc) > 0) {
-		rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
-			 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
+		rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
+			 (unsigned long long)wc.wr_id, wc.status,
+			 rds_ib_wc_status_str(wc.status), wc.byte_len,
 			 be32_to_cpu(wc.ex.imm_data));
 		rds_ib_stats_inc(s_ib_tx_cq_event);
 
@@ -210,51 +310,30 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 
 		for (i = 0; i < completed; i++) {
 			send = &ic->i_sends[oldest];
+			if (send->s_wr.send_flags & IB_SEND_SIGNALED)
+				nr_sig++;
 
-			/* In the error case, wc.opcode sometimes contains garbage */
-			switch (send->s_wr.opcode) {
-			case IB_WR_SEND:
-				if (send->s_rm)
-					rds_ib_send_unmap_rm(ic, send, wc.status);
-				break;
-			case IB_WR_RDMA_WRITE:
-			case IB_WR_RDMA_READ:
-				/* Nothing to be done - the SG list will be unmapped
-				 * when the SEND completes. */
-				break;
-			default:
-				if (printk_ratelimit())
-					printk(KERN_NOTICE
-						"RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
-						__func__, send->s_wr.opcode);
-				break;
-			}
+			rm = rds_ib_send_unmap_op(ic, send, wc.status);
 
-			send->s_wr.opcode = 0xdead;
-			send->s_wr.num_sge = 1;
 			if (send->s_queued + HZ/2 < jiffies)
 				rds_ib_stats_inc(s_ib_tx_stalled);
 
-			/* If a RDMA operation produced an error, signal this right
-			 * away. If we don't, the subsequent SEND that goes with this
-			 * RDMA will be canceled with ERR_WFLUSH, and the application
-			 * never learn that the RDMA failed. */
-			if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
-				struct rds_message *rm;
-
-				rm = rds_send_get_message(conn, send->s_op);
-				if (rm) {
-					if (rm->m_rdma_op)
-						rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
-					rds_ib_send_rdma_complete(rm, wc.status);
-					rds_message_put(rm);
+			if (send->s_op) {
+				if (send->s_op == rm->m_final_op) {
+					/* If anyone waited for this message to get flushed out, wake
+					 * them up now */
+					rds_message_unmapped(rm);
 				}
+				rds_message_put(rm);
+				send->s_op = NULL;
 			}
 
 			oldest = (oldest + 1) % ic->i_send_ring.w_nr;
 		}
 
 		rds_ib_ring_free(&ic->i_send_ring, completed);
+		rds_ib_sub_signaled(ic, nr_sig);
+		nr_sig = 0;
 
 		if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 		    test_bit(0, &conn->c_map_queued))
@@ -262,10 +341,10 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 
 		/* We expect errors as the qp is drained during shutdown */
 		if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
-			rds_ib_conn_error(conn,
-				"send completion on %pI4 "
-				"had status %u, disconnecting and reconnecting\n",
-				&conn->c_faddr, wc.status);
+			rds_ib_conn_error(conn, "send completion on %pI4 had status "
+					  "%u (%s), disconnecting and reconnecting\n",
+					  &conn->c_faddr, wc.status,
+					  rds_ib_wc_status_str(wc.status));
 		}
 	}
 }
@@ -294,7 +373,7 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
  * credits (see rds_ib_send_add_credits below).
  *
  * The RDS send code is essentially single-threaded; rds_send_xmit
- * grabs c_send_lock to ensure exclusive access to the send ring.
+ * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
  * However, the ACK sending code is independent and can race with
  * message SENDs.
  *
@@ -413,40 +492,21 @@ void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 }
 
-static inline void
-rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
-		struct rds_ib_send_work *send, unsigned int pos,
-		unsigned long buffer, unsigned int length,
-		int send_flags)
+static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
+					     struct rds_ib_send_work *send,
+					     bool notify)
 {
-	struct ib_sge *sge;
-
-	WARN_ON(pos != send - ic->i_sends);
-
-	send->s_wr.send_flags = send_flags;
-	send->s_wr.opcode = IB_WR_SEND;
-	send->s_wr.num_sge = 2;
-	send->s_wr.next = NULL;
-	send->s_queued = jiffies;
-	send->s_op = NULL;
-
-	if (length != 0) {
-		sge = rds_ib_data_sge(ic, send->s_sge);
-		sge->addr = buffer;
-		sge->length = length;
-		sge->lkey = ic->i_mr->lkey;
-
-		sge = rds_ib_header_sge(ic, send->s_sge);
-	} else {
-		/* We're sending a packet with no payload. There is only
-		 * one SGE */
-		send->s_wr.num_sge = 1;
-		sge = &send->s_sge[0];
+	/*
+	 * We want to delay signaling completions just enough to get
+	 * the batching benefits but not so much that we create dead time
+	 * on the wire.
+	 */
+	if (ic->i_unsignaled_wrs-- == 0 || notify) {
+		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+		send->s_wr.send_flags |= IB_SEND_SIGNALED;
+		return 1;
 	}
-
-	sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
-	sge->length = sizeof(struct rds_header);
-	sge->lkey = ic->i_mr->lkey;
+	return 0;
 }
 
 /*
@@ -475,13 +535,14 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	u32 pos;
 	u32 i;
 	u32 work_alloc;
-	u32 credit_alloc;
+	u32 credit_alloc = 0;
 	u32 posted;
 	u32 adv_credits = 0;
 	int send_flags = 0;
-	int sent;
+	int bytes_sent = 0;
 	int ret;
 	int flow_controlled = 0;
+	int nr_sig = 0;
 
 	BUG_ON(off % RDS_FRAG_SIZE);
 	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
@@ -507,14 +568,13 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 		goto out;
 	}
 
-	credit_alloc = work_alloc;
 	if (ic->i_flowctl) {
 		credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
 		adv_credits += posted;
 		if (credit_alloc < work_alloc) {
 			rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
 			work_alloc = credit_alloc;
-			flow_controlled++;
+			flow_controlled = 1;
 		}
 		if (work_alloc == 0) {
 			set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
@@ -525,31 +585,25 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	}
 
 	/* map the message the first time we see it */
-	if (ic->i_rm == NULL) {
-		/*
-		printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
-				be16_to_cpu(rm->m_inc.i_hdr.h_dport),
-				rm->m_inc.i_hdr.h_flags,
-				be32_to_cpu(rm->m_inc.i_hdr.h_len));
-		   */
-		if (rm->m_nents) {
-			rm->m_count = ib_dma_map_sg(dev,
-					 rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
-			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
-			if (rm->m_count == 0) {
+	if (!ic->i_data_op) {
+		if (rm->data.op_nents) {
+			rm->data.op_count = ib_dma_map_sg(dev,
+							  rm->data.op_sg,
+							  rm->data.op_nents,
+							  DMA_TO_DEVICE);
+			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
+			if (rm->data.op_count == 0) {
 				rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 				rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
 				ret = -ENOMEM; /* XXX ? */
 				goto out;
 			}
 		} else {
-			rm->m_count = 0;
+			rm->data.op_count = 0;
 		}
 
-		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
-		ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
 		rds_message_addref(rm);
-		ic->i_rm = rm;
+		ic->i_data_op = &rm->data;
 
 		/* Finalize the header */
 		if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
@@ -559,10 +613,10 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 
 		/* If it has a RDMA op, tell the peer we did it. This is
 		 * used by the peer to release use-once RDMA MRs. */
-		if (rm->m_rdma_op) {
+		if (rm->rdma.op_active) {
 			struct rds_ext_header_rdma ext_hdr;
 
-			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
+			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 			rds_message_add_extension(&rm->m_inc.i_hdr,
 					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 		}
@@ -582,99 +636,77 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 		/*
 		 * Update adv_credits since we reset the ACK_REQUIRED bit.
 		 */
-		rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
-		adv_credits += posted;
-		BUG_ON(adv_credits > 255);
+		if (ic->i_flowctl) {
+			rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
+			adv_credits += posted;
+			BUG_ON(adv_credits > 255);
+		}
 	}
 
-	send = &ic->i_sends[pos];
-	first = send;
-	prev = NULL;
-	scat = &rm->m_sg[sg];
-	sent = 0;
-	i = 0;
-
 	/* Sometimes you want to put a fence between an RDMA
 	 * READ and the following SEND.
 	 * We could either do this all the time
 	 * or when requested by the user. Right now, we let
 	 * the application choose.
 	 */
-	if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
+	if (rm->rdma.op_active && rm->rdma.op_fence)
 		send_flags = IB_SEND_FENCE;
 
-	/*
-	 * We could be copying the header into the unused tail of the page.
-	 * That would need to be changed in the future when those pages might
-	 * be mapped userspace pages or page cache pages.  So instead we always
-	 * use a second sge and our long-lived ring of mapped headers.  We send
-	 * the header after the data so that the data payload can be aligned on
-	 * the receiver.
-	 */
+	/* Each frag gets a header. Msgs may be 0 bytes */
+	send = &ic->i_sends[pos];
+	first = send;
+	prev = NULL;
+	scat = &ic->i_data_op->op_sg[sg];
+	i = 0;
+	do {
+		unsigned int len = 0;
 
-	/* handle a 0-len message */
-	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
-		rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
-		goto add_header;
-	}
+		/* Set up the header */
+		send->s_wr.send_flags = send_flags;
+		send->s_wr.opcode = IB_WR_SEND;
+		send->s_wr.num_sge = 1;
+		send->s_wr.next = NULL;
+		send->s_queued = jiffies;
+		send->s_op = NULL;
 
-	/* if there's data reference it with a chain of work reqs */
-	for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
-		unsigned int len;
+		send->s_sge[0].addr = ic->i_send_hdrs_dma
+			+ (pos * sizeof(struct rds_header));
+		send->s_sge[0].length = sizeof(struct rds_header);
 
-		send = &ic->i_sends[pos];
+		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
 
-		len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
-		rds_ib_xmit_populate_wr(ic, send, pos,
-				ib_sg_dma_address(dev, scat) + off, len,
-				send_flags);
+		/* Set up the data, if present */
+		if (i < work_alloc
+		    && scat != &rm->data.op_sg[rm->data.op_count]) {
+			len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
+			send->s_wr.num_sge = 2;
 
-		/*
-		 * We want to delay signaling completions just enough to get
-		 * the batching benefits but not so much that we create dead time
-		 * on the wire.
-		 */
-		if (ic->i_unsignaled_wrs-- == 0) {
-			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
-			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
-		}
+			send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
+			send->s_sge[1].length = len;
 
-		ic->i_unsignaled_bytes -= len;
-		if (ic->i_unsignaled_bytes <= 0) {
-			ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
-			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+			bytes_sent += len;
+			off += len;
+			if (off == ib_sg_dma_len(dev, scat)) {
+				scat++;
+				off = 0;
+			}
 		}
 
+		rds_ib_set_wr_signal_state(ic, send, 0);
+
 		/*
 		 * Always signal the last one if we're stopping due to flow control.
 		 */
-		if (flow_controlled && i == (work_alloc-1))
+		if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
 			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 
+		if (send->s_wr.send_flags & IB_SEND_SIGNALED)
+			nr_sig++;
+
 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
 			 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
 
-		sent += len;
-		off += len;
-		if (off == ib_sg_dma_len(dev, scat)) {
-			scat++;
-			off = 0;
-		}
-
-add_header:
-		/* Tack on the header after the data. The header SGE should already
-		 * have been set up to point to the right header buffer. */
-		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
-
-		if (0) {
-			struct rds_header *hdr = &ic->i_send_hdrs[pos];
-
-			printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
-				be16_to_cpu(hdr->h_dport),
-				hdr->h_flags,
-				be32_to_cpu(hdr->h_len));
-		}
-		if (adv_credits) {
+		if (ic->i_flowctl && adv_credits) {
 			struct rds_header *hdr = &ic->i_send_hdrs[pos];
 
 			/* add credit and redo the header checksum */
@@ -689,20 +721,25 @@ add_header:
 		prev = send;
 
 		pos = (pos + 1) % ic->i_send_ring.w_nr;
-	}
+		send = &ic->i_sends[pos];
+		i++;
+
+	} while (i < work_alloc
+		 && scat != &rm->data.op_sg[rm->data.op_count]);
 
 	/* Account the RDS header in the number of bytes we sent, but just once.
 	 * The caller has no concept of fragmentation. */
 	if (hdr_off == 0)
-		sent += sizeof(struct rds_header);
+		bytes_sent += sizeof(struct rds_header);
 
 	/* if we finished the message then send completion owns it */
-	if (scat == &rm->m_sg[rm->m_count]) {
-		prev->s_rm = ic->i_rm;
-		prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
-		ic->i_rm = NULL;
+	if (scat == &rm->data.op_sg[rm->data.op_count]) {
+		prev->s_op = ic->i_data_op;
+		prev->s_wr.send_flags |= IB_SEND_SOLICITED;
+		ic->i_data_op = NULL;
 	}
 
+	/* Put back wrs & credits we didn't use */
 	if (i < work_alloc) {
 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 		work_alloc = i;
@@ -710,6 +747,9 @@ add_header:
 	if (ic->i_flowctl && i < credit_alloc)
 		rds_ib_send_add_credits(conn, credit_alloc - i);
 
+	if (nr_sig)
+		atomic_add(nr_sig, &ic->i_signaled_sends);
+
 	/* XXX need to worry about failed_wr and partial sends. */
 	failed_wr = &first->s_wr;
 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
@@ -720,32 +760,127 @@ add_header:
 		printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
 		       "returned %d\n", &conn->c_faddr, ret);
 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
-		if (prev->s_rm) {
-			ic->i_rm = prev->s_rm;
-			prev->s_rm = NULL;
+		rds_ib_sub_signaled(ic, nr_sig);
+		if (prev->s_op) {
+			ic->i_data_op = prev->s_op;
+			prev->s_op = NULL;
 		}
 
 		rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
 		goto out;
 	}
 
-	ret = sent;
+	ret = bytes_sent;
 out:
 	BUG_ON(adv_credits);
 	return ret;
 }
 
-int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
+/*
+ * Issue atomic operation.
+ * A simplified version of the rdma case, we always map 1 SG, and
+ * only 8 bytes, for the return value from the atomic operation.
+ */
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
+{
+	struct rds_ib_connection *ic = conn->c_transport_data;
+	struct rds_ib_send_work *send = NULL;
+	struct ib_send_wr *failed_wr;
+	struct rds_ib_device *rds_ibdev;
+	u32 pos;
+	u32 work_alloc;
+	int ret;
+	int nr_sig = 0;
+
+	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
+
+	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
+	if (work_alloc != 1) {
+		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+		rds_ib_stats_inc(s_ib_tx_ring_full);
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	/* address of send request in ring */
+	send = &ic->i_sends[pos];
+	send->s_queued = jiffies;
+
+	if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
+		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
+		send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
+		send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
+		send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
+		send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
+	} else { /* FADD */
+		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
+		send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
+		send->s_wr.wr.atomic.swap = 0;
+		send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
+		send->s_wr.wr.atomic.swap_mask = 0;
+	}
+	nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
+	send->s_wr.num_sge = 1;
+	send->s_wr.next = NULL;
+	send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
+	send->s_wr.wr.atomic.rkey = op->op_rkey;
+	send->s_op = op;
+	rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
+
+	/* map 8 byte retval buffer to the device */
+	ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
+	rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
+	if (ret != 1) {
+		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+		rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+		ret = -ENOMEM; /* XXX ? */
+		goto out;
+	}
+
+	/* Convert our struct scatterlist to struct ib_sge */
+	send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
+	send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
+	send->s_sge[0].lkey = ic->i_mr->lkey;
+
+	rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
+		 send->s_sge[0].addr, send->s_sge[0].length);
+
+	if (nr_sig)
+		atomic_add(nr_sig, &ic->i_signaled_sends);
+
+	failed_wr = &send->s_wr;
+	ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
+	rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
+		 send, &send->s_wr, ret, failed_wr);
+	BUG_ON(failed_wr != &send->s_wr);
+	if (ret) {
+		printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
+		       "returned %d\n", &conn->c_faddr, ret);
+		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+		rds_ib_sub_signaled(ic, nr_sig);
+		goto out;
+	}
+
+	if (unlikely(failed_wr != &send->s_wr)) {
+		printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
+		BUG_ON(failed_wr != &send->s_wr);
+	}
+
+out:
+	return ret;
+}
+
+int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct rds_ib_send_work *send = NULL;
 	struct rds_ib_send_work *first;
 	struct rds_ib_send_work *prev;
 	struct ib_send_wr *failed_wr;
-	struct rds_ib_device *rds_ibdev;
 	struct scatterlist *scat;
 	unsigned long len;
-	u64 remote_addr = op->r_remote_addr;
+	u64 remote_addr = op->op_remote_addr;
+	u32 max_sge = ic->rds_ibdev->max_sge;
 	u32 pos;
 	u32 work_alloc;
 	u32 i;
@@ -753,29 +888,28 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	int sent;
 	int ret;
 	int num_sge;
-
-	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
-
-	/* map the message the first time we see it */
-	if (!op->r_mapped) {
-		op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
-					op->r_sg, op->r_nents, (op->r_write) ?
-					DMA_TO_DEVICE : DMA_FROM_DEVICE);
-		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
-		if (op->r_count == 0) {
+	int nr_sig = 0;
+
+	/* map the op the first time we see it */
+	if (!op->op_mapped) {
+		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
+					     op->op_sg, op->op_nents, (op->op_write) ?
+					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
+		if (op->op_count == 0) {
 			rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
 			ret = -ENOMEM; /* XXX ? */
 			goto out;
 		}
 
-		op->r_mapped = 1;
+		op->op_mapped = 1;
 	}
 
 	/*
 	 * Instead of knowing how to return a partial rdma read/write we insist that there
 	 * be enough work requests to send the entire message.
 	 */
-	i = ceil(op->r_count, rds_ibdev->max_sge);
+	i = ceil(op->op_count, max_sge);
 
 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
 	if (work_alloc != i) {
@@ -788,30 +922,24 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	send = &ic->i_sends[pos];
 	first = send;
 	prev = NULL;
-	scat = &op->r_sg[0];
+	scat = &op->op_sg[0];
 	sent = 0;
-	num_sge = op->r_count;
+	num_sge = op->op_count;
 
-	for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
+	for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 		send->s_wr.send_flags = 0;
 		send->s_queued = jiffies;
-		/*
-		 * We want to delay signaling completions just enough to get
-		 * the batching benefits but not so much that we create dead time on the wire.
-		 */
-		if (ic->i_unsignaled_wrs-- == 0) {
-			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
-			send->s_wr.send_flags = IB_SEND_SIGNALED;
-		}
+		send->s_op = NULL;
+
+		nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 
-		send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
+		send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 		send->s_wr.wr.rdma.remote_addr = remote_addr;
-		send->s_wr.wr.rdma.rkey = op->r_key;
-		send->s_op = op;
+		send->s_wr.wr.rdma.rkey = op->op_rkey;
 
-		if (num_sge > rds_ibdev->max_sge) {
-			send->s_wr.num_sge = rds_ibdev->max_sge;
-			num_sge -= rds_ibdev->max_sge;
+		if (num_sge > max_sge) {
+			send->s_wr.num_sge = max_sge;
+			num_sge -= max_sge;
 		} else {
 			send->s_wr.num_sge = num_sge;
 		}
@@ -821,7 +949,7 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 		if (prev)
 			prev->s_wr.next = &send->s_wr;
 
-		for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
+		for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 			send->s_sge[j].addr =
 				 ib_sg_dma_address(ic->i_cm_id->device, scat);
@@ -843,15 +971,20 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 			send = ic->i_sends;
 	}
 
-	/* if we finished the message then send completion owns it */
-	if (scat == &op->r_sg[op->r_count])
-		prev->s_wr.send_flags = IB_SEND_SIGNALED;
+	/* give a reference to the last op */
+	if (scat == &op->op_sg[op->op_count]) {
+		prev->s_op = op;
+		rds_message_addref(container_of(op, struct rds_message, rdma));
+	}
 
 	if (i < work_alloc) {
 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 		work_alloc = i;
 	}
 
+	if (nr_sig)
+		atomic_add(nr_sig, &ic->i_signaled_sends);
+
 	failed_wr = &first->s_wr;
 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
@@ -861,6 +994,7 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 		printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
 		       "returned %d\n", &conn->c_faddr, ret);
 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
+		rds_ib_sub_signaled(ic, nr_sig);
 		goto out;
 	}
 

net/rds/ib_stats.c

@@ -67,6 +67,8 @@ static const char *const rds_ib_stat_names[] = {
 	"ib_rdma_mr_pool_flush",
 	"ib_rdma_mr_pool_wait",
 	"ib_rdma_mr_pool_depleted",
+	"ib_atomic_cswp",
+	"ib_atomic_fadd",
 };
 
 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,

net/rds/ib_sysctl.c

@@ -49,10 +49,6 @@ unsigned long rds_ib_sysctl_max_unsig_wrs = 16;
 static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1;
 static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64;
 
-unsigned long rds_ib_sysctl_max_unsig_bytes = (16 << 20);
-static unsigned long rds_ib_sysctl_max_unsig_bytes_min = 1;
-static unsigned long rds_ib_sysctl_max_unsig_bytes_max = ~0UL;
-
 /*
  * This sysctl does nothing.
  *
@@ -93,15 +89,6 @@ ctl_table rds_ib_sysctl_table[] = {
 		.extra1		= &rds_ib_sysctl_max_unsig_wr_min,
 		.extra2		= &rds_ib_sysctl_max_unsig_wr_max,
 	},
-	{
-		.procname       = "max_unsignaled_bytes",
-		.data		= &rds_ib_sysctl_max_unsig_bytes,
-		.maxlen         = sizeof(unsigned long),
-		.mode           = 0644,
-		.proc_handler   = proc_doulongvec_minmax,
-		.extra1		= &rds_ib_sysctl_max_unsig_bytes_min,
-		.extra2		= &rds_ib_sysctl_max_unsig_bytes_max,
-	},
 	{
 		.procname       = "max_recv_allocation",
 		.data		= &rds_ib_sysctl_max_recv_allocation,
@@ -132,10 +119,10 @@ void rds_ib_sysctl_exit(void)
 		unregister_sysctl_table(rds_ib_sysctl_hdr);
 }
 
-int __init rds_ib_sysctl_init(void)
+int rds_ib_sysctl_init(void)
 {
 	rds_ib_sysctl_hdr = register_sysctl_paths(rds_ib_sysctl_path, rds_ib_sysctl_table);
-	if (rds_ib_sysctl_hdr == NULL)
+	if (!rds_ib_sysctl_hdr)
 		return -ENOMEM;
 	return 0;
 }

net/rds/info.c

@@ -76,7 +76,7 @@ void rds_info_register_func(int optname, rds_info_func func)
 	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
 
 	spin_lock(&rds_info_lock);
-	BUG_ON(rds_info_funcs[offset] != NULL);
+	BUG_ON(rds_info_funcs[offset]);
 	rds_info_funcs[offset] = func;
 	spin_unlock(&rds_info_lock);
 }
@@ -102,7 +102,7 @@ EXPORT_SYMBOL_GPL(rds_info_deregister_func);
  */
 void rds_info_iter_unmap(struct rds_info_iterator *iter)
 {
-	if (iter->addr != NULL) {
+	if (iter->addr) {
 		kunmap_atomic(iter->addr, KM_USER0);
 		iter->addr = NULL;
 	}
@@ -117,7 +117,7 @@ void rds_info_copy(struct rds_info_iterator *iter, void *data,
 	unsigned long this;
 
 	while (bytes) {
-		if (iter->addr == NULL)
+		if (!iter->addr)
 			iter->addr = kmap_atomic(*iter->pages, KM_USER0);
 
 		this = min(bytes, PAGE_SIZE - iter->offset);
@@ -188,7 +188,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval,
 			>> PAGE_SHIFT;
 
 	pages = kmalloc(nr_pages * sizeof(struct page *), GFP_KERNEL);
-	if (pages == NULL) {
+	if (!pages) {
 		ret = -ENOMEM;
 		goto out;
 	}
@@ -206,7 +206,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval,
 
 call_func:
 	func = rds_info_funcs[optname - RDS_INFO_FIRST];
-	if (func == NULL) {
+	if (!func) {
 		ret = -ENOPROTOOPT;
 		goto out;
 	}
@@ -234,7 +234,7 @@ call_func:
 		ret = -EFAULT;
 
 out:
-	for (i = 0; pages != NULL && i < nr_pages; i++)
+	for (i = 0; pages && i < nr_pages; i++)
 		put_page(pages[i]);
 	kfree(pages);
 

net/rds/iw.c

@@ -264,7 +264,6 @@ struct rds_transport rds_iw_transport = {
 	.laddr_check		= rds_iw_laddr_check,
 	.xmit_complete		= rds_iw_xmit_complete,
 	.xmit			= rds_iw_xmit,
-	.xmit_cong_map		= NULL,
 	.xmit_rdma		= rds_iw_xmit_rdma,
 	.recv			= rds_iw_recv,
 	.conn_alloc		= rds_iw_conn_alloc,
@@ -272,7 +271,6 @@ struct rds_transport rds_iw_transport = {
 	.conn_connect		= rds_iw_conn_connect,
 	.conn_shutdown		= rds_iw_conn_shutdown,
 	.inc_copy_to_user	= rds_iw_inc_copy_to_user,
-	.inc_purge		= rds_iw_inc_purge,
 	.inc_free		= rds_iw_inc_free,
 	.cm_initiate_connect	= rds_iw_cm_initiate_connect,
 	.cm_handle_connect	= rds_iw_cm_handle_connect,
@@ -289,7 +287,7 @@ struct rds_transport rds_iw_transport = {
 	.t_prefer_loopback	= 1,
 };
 
-int __init rds_iw_init(void)
+int rds_iw_init(void)
 {
 	int ret;
 

net/rds/iw.h

@@ -70,7 +70,7 @@ struct rds_iw_send_work {
 	struct rds_message	*s_rm;
 
 	/* We should really put these into a union: */
-	struct rds_rdma_op	*s_op;
+	struct rm_rdma_op	*s_op;
 	struct rds_iw_mapping	*s_mapping;
 	struct ib_mr		*s_mr;
 	struct ib_fast_reg_page_list *s_page_list;
@@ -284,7 +284,7 @@ void rds_iw_conn_free(void *arg);
 int rds_iw_conn_connect(struct rds_connection *conn);
 void rds_iw_conn_shutdown(struct rds_connection *conn);
 void rds_iw_state_change(struct sock *sk);
-int __init rds_iw_listen_init(void);
+int rds_iw_listen_init(void);
 void rds_iw_listen_stop(void);
 void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
@@ -321,12 +321,11 @@ void rds_iw_flush_mrs(void);
 void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
 
 /* ib_recv.c */
-int __init rds_iw_recv_init(void);
+int rds_iw_recv_init(void);
 void rds_iw_recv_exit(void);
 int rds_iw_recv(struct rds_connection *conn);
 int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
 		       gfp_t page_gfp, int prefill);
-void rds_iw_inc_purge(struct rds_incoming *inc);
 void rds_iw_inc_free(struct rds_incoming *inc);
 int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 			     size_t size);
@@ -358,7 +357,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
 void rds_iw_send_init_ring(struct rds_iw_connection *ic);
 void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
-int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
+int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
 void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
 void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
 int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted,
@@ -371,7 +370,7 @@ unsigned int rds_iw_stats_info_copy(struct rds_info_iterator *iter,
 				    unsigned int avail);
 
 /* ib_sysctl.c */
-int __init rds_iw_sysctl_init(void);
+int rds_iw_sysctl_init(void);
 void rds_iw_sysctl_exit(void);
 extern unsigned long rds_iw_sysctl_max_send_wr;
 extern unsigned long rds_iw_sysctl_max_recv_wr;

net/rds/iw_cm.c

@@ -257,7 +257,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
 	 * the rds_iwdev at all.
 	 */
 	rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
-	if (rds_iwdev == NULL) {
+	if (!rds_iwdev) {
 		if (printk_ratelimit())
 			printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
 					dev->name);
@@ -292,7 +292,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
 					   ic->i_send_ring.w_nr *
 						sizeof(struct rds_header),
 					   &ic->i_send_hdrs_dma, GFP_KERNEL);
-	if (ic->i_send_hdrs == NULL) {
+	if (!ic->i_send_hdrs) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent send failed\n");
 		goto out;
@@ -302,7 +302,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
 					   ic->i_recv_ring.w_nr *
 						sizeof(struct rds_header),
 					   &ic->i_recv_hdrs_dma, GFP_KERNEL);
-	if (ic->i_recv_hdrs == NULL) {
+	if (!ic->i_recv_hdrs) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent recv failed\n");
 		goto out;
@@ -310,14 +310,14 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
 
 	ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 				       &ic->i_ack_dma, GFP_KERNEL);
-	if (ic->i_ack == NULL) {
+	if (!ic->i_ack) {
 		ret = -ENOMEM;
 		rdsdebug("ib_dma_alloc_coherent ack failed\n");
 		goto out;
 	}
 
 	ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
-	if (ic->i_sends == NULL) {
+	if (!ic->i_sends) {
 		ret = -ENOMEM;
 		rdsdebug("send allocation failed\n");
 		goto out;
@@ -325,7 +325,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
 	rds_iw_send_init_ring(ic);
 
 	ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
-	if (ic->i_recvs == NULL) {
+	if (!ic->i_recvs) {
 		ret = -ENOMEM;
 		rdsdebug("recv allocation failed\n");
 		goto out;
@@ -696,7 +696,7 @@ int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 
 	/* XXX too lazy? */
 	ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL);
-	if (ic == NULL)
+	if (!ic)
 		return -ENOMEM;
 
 	INIT_LIST_HEAD(&ic->iw_node);

net/rds/iw_rdma.c

@@ -34,7 +34,6 @@
 #include <linux/slab.h>
 
 #include "rds.h"
-#include "rdma.h"
 #include "iw.h"
 
 

net/rds/iw_recv.c

@@ -53,7 +53,7 @@ static void rds_iw_frag_drop_page(struct rds_page_frag *frag)
 static void rds_iw_frag_free(struct rds_page_frag *frag)
 {
 	rdsdebug("frag %p page %p\n", frag, frag->f_page);
-	BUG_ON(frag->f_page != NULL);
+	BUG_ON(frag->f_page);
 	kmem_cache_free(rds_iw_frag_slab, frag);
 }
 
@@ -143,14 +143,14 @@ static int rds_iw_recv_refill_one(struct rds_connection *conn,
 	struct ib_sge *sge;
 	int ret = -ENOMEM;
 
-	if (recv->r_iwinc == NULL) {
+	if (!recv->r_iwinc) {
 		if (!atomic_add_unless(&rds_iw_allocation, 1, rds_iw_sysctl_max_recv_allocation)) {
 			rds_iw_stats_inc(s_iw_rx_alloc_limit);
 			goto out;
 		}
 		recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab,
 						 kptr_gfp);
-		if (recv->r_iwinc == NULL) {
+		if (!recv->r_iwinc) {
 			atomic_dec(&rds_iw_allocation);
 			goto out;
 		}
@@ -158,17 +158,17 @@ static int rds_iw_recv_refill_one(struct rds_connection *conn,
 		rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr);
 	}
 
-	if (recv->r_frag == NULL) {
+	if (!recv->r_frag) {
 		recv->r_frag = kmem_cache_alloc(rds_iw_frag_slab, kptr_gfp);
-		if (recv->r_frag == NULL)
+		if (!recv->r_frag)
 			goto out;
 		INIT_LIST_HEAD(&recv->r_frag->f_item);
 		recv->r_frag->f_page = NULL;
 	}
 
-	if (ic->i_frag.f_page == NULL) {
+	if (!ic->i_frag.f_page) {
 		ic->i_frag.f_page = alloc_page(page_gfp);
-		if (ic->i_frag.f_page == NULL)
+		if (!ic->i_frag.f_page)
 			goto out;
 		ic->i_frag.f_offset = 0;
 	}
@@ -273,7 +273,7 @@ int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
 	return ret;
 }
 
-void rds_iw_inc_purge(struct rds_incoming *inc)
+static void rds_iw_inc_purge(struct rds_incoming *inc)
 {
 	struct rds_iw_incoming *iwinc;
 	struct rds_page_frag *frag;
@@ -716,7 +716,7 @@ static void rds_iw_process_recv(struct rds_connection *conn,
 	 * into the inc and save the inc so we can hang upcoming fragments
 	 * off its list.
 	 */
-	if (iwinc == NULL) {
+	if (!iwinc) {
 		iwinc = recv->r_iwinc;
 		recv->r_iwinc = NULL;
 		ic->i_iwinc = iwinc;
@@ -887,7 +887,7 @@ int rds_iw_recv(struct rds_connection *conn)
 	return ret;
 }
 
-int __init rds_iw_recv_init(void)
+int rds_iw_recv_init(void)
 {
 	struct sysinfo si;
 	int ret = -ENOMEM;
@@ -899,13 +899,13 @@ int __init rds_iw_recv_init(void)
 	rds_iw_incoming_slab = kmem_cache_create("rds_iw_incoming",
 					sizeof(struct rds_iw_incoming),
 					0, 0, NULL);
-	if (rds_iw_incoming_slab == NULL)
+	if (!rds_iw_incoming_slab)
 		goto out;
 
 	rds_iw_frag_slab = kmem_cache_create("rds_iw_frag",
 					sizeof(struct rds_page_frag),
 					0, 0, NULL);
-	if (rds_iw_frag_slab == NULL)
+	if (!rds_iw_frag_slab)
 		kmem_cache_destroy(rds_iw_incoming_slab);
 	else
 		ret = 0;

net/rds/iw_send.c

@@ -36,7 +36,6 @@
 #include <linux/dmapool.h>
 
 #include "rds.h"
-#include "rdma.h"
 #include "iw.h"
 
 static void rds_iw_send_rdma_complete(struct rds_message *rm,
@@ -64,13 +63,13 @@ static void rds_iw_send_rdma_complete(struct rds_message *rm,
 }
 
 static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic,
-				   struct rds_rdma_op *op)
+				   struct rm_rdma_op *op)
 {
-	if (op->r_mapped) {
+	if (op->op_mapped) {
 		ib_dma_unmap_sg(ic->i_cm_id->device,
-			op->r_sg, op->r_nents,
-			op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
-		op->r_mapped = 0;
+			op->op_sg, op->op_nents,
+			op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		op->op_mapped = 0;
 	}
 }
 
@@ -83,11 +82,11 @@ static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
 	rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
 
 	ib_dma_unmap_sg(ic->i_cm_id->device,
-		     rm->m_sg, rm->m_nents,
+		     rm->data.op_sg, rm->data.op_nents,
 		     DMA_TO_DEVICE);
 
-	if (rm->m_rdma_op != NULL) {
-		rds_iw_send_unmap_rdma(ic, rm->m_rdma_op);
+	if (rm->rdma.op_active) {
+		rds_iw_send_unmap_rdma(ic, &rm->rdma);
 
 		/* If the user asked for a completion notification on this
 		 * message, we can implement three different semantics:
@@ -111,10 +110,10 @@ static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
 		 */
 		rds_iw_send_rdma_complete(rm, wc_status);
 
-		if (rm->m_rdma_op->r_write)
-			rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
+		if (rm->rdma.op_write)
+			rds_stats_add(s_send_rdma_bytes, rm->rdma.op_bytes);
 		else
-			rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
+			rds_stats_add(s_recv_rdma_bytes, rm->rdma.op_bytes);
 	}
 
 	/* If anyone waited for this message to get flushed out, wake
@@ -556,25 +555,27 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 	}
 
 	/* map the message the first time we see it */
-	if (ic->i_rm == NULL) {
+	if (!ic->i_rm) {
 		/*
 		printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
 				be16_to_cpu(rm->m_inc.i_hdr.h_dport),
 				rm->m_inc.i_hdr.h_flags,
 				be32_to_cpu(rm->m_inc.i_hdr.h_len));
 		   */
-		if (rm->m_nents) {
-			rm->m_count = ib_dma_map_sg(dev,
-					 rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
-			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
-			if (rm->m_count == 0) {
+		if (rm->data.op_nents) {
+			rm->data.op_count = ib_dma_map_sg(dev,
+							  rm->data.op_sg,
+							  rm->data.op_nents,
+							  DMA_TO_DEVICE);
+			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
+			if (rm->data.op_count == 0) {
 				rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 				rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
 				ret = -ENOMEM; /* XXX ? */
 				goto out;
 			}
 		} else {
-			rm->m_count = 0;
+			rm->data.op_count = 0;
 		}
 
 		ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
@@ -590,10 +591,10 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 
 		/* If it has a RDMA op, tell the peer we did it. This is
 		 * used by the peer to release use-once RDMA MRs. */
-		if (rm->m_rdma_op) {
+		if (rm->rdma.op_active) {
 			struct rds_ext_header_rdma ext_hdr;
 
-			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
+			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
 			rds_message_add_extension(&rm->m_inc.i_hdr,
 					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
 		}
@@ -621,7 +622,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 	send = &ic->i_sends[pos];
 	first = send;
 	prev = NULL;
-	scat = &rm->m_sg[sg];
+	scat = &rm->data.op_sg[sg];
 	sent = 0;
 	i = 0;
 
@@ -631,7 +632,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 	 * or when requested by the user. Right now, we let
 	 * the application choose.
 	 */
-	if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
+	if (rm->rdma.op_active && rm->rdma.op_fence)
 		send_flags = IB_SEND_FENCE;
 
 	/*
@@ -650,7 +651,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
 	}
 
 	/* if there's data reference it with a chain of work reqs */
-	for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
+	for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
 		unsigned int len;
 
 		send = &ic->i_sends[pos];
@@ -728,7 +729,7 @@ add_header:
 		sent += sizeof(struct rds_header);
 
 	/* if we finished the message then send completion owns it */
-	if (scat == &rm->m_sg[rm->m_count]) {
+	if (scat == &rm->data.op_sg[rm->data.op_count]) {
 		prev->s_rm = ic->i_rm;
 		prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
 		ic->i_rm = NULL;
@@ -784,7 +785,7 @@ static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rd
 	ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
 }
 
-int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
+int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 {
 	struct rds_iw_connection *ic = conn->c_transport_data;
 	struct rds_iw_send_work *send = NULL;
@@ -794,7 +795,7 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	struct rds_iw_device *rds_iwdev;
 	struct scatterlist *scat;
 	unsigned long len;
-	u64 remote_addr = op->r_remote_addr;
+	u64 remote_addr = op->op_remote_addr;
 	u32 pos, fr_pos;
 	u32 work_alloc;
 	u32 i;
@@ -806,21 +807,21 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 
 	/* map the message the first time we see it */
-	if (!op->r_mapped) {
-		op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
-					op->r_sg, op->r_nents, (op->r_write) ?
-					DMA_TO_DEVICE : DMA_FROM_DEVICE);
-		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
-		if (op->r_count == 0) {
+	if (!op->op_mapped) {
+		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
+					     op->op_sg, op->op_nents, (op->op_write) ?
+					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
+		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
+		if (op->op_count == 0) {
 			rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
 			ret = -ENOMEM; /* XXX ? */
 			goto out;
 		}
 
-		op->r_mapped = 1;
+		op->op_mapped = 1;
 	}
 
-	if (!op->r_write) {
+	if (!op->op_write) {
 		/* Alloc space on the send queue for the fastreg */
 		work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
 		if (work_alloc != 1) {
@@ -835,7 +836,7 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	 * Instead of knowing how to return a partial rdma read/write we insist that there
 	 * be enough work requests to send the entire message.
 	 */
-	i = ceil(op->r_count, rds_iwdev->max_sge);
+	i = ceil(op->op_count, rds_iwdev->max_sge);
 
 	work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
 	if (work_alloc != i) {
@@ -846,17 +847,17 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	}
 
 	send = &ic->i_sends[pos];
-	if (!op->r_write) {
+	if (!op->op_write) {
 		first = prev = &ic->i_sends[fr_pos];
 	} else {
 		first = send;
 		prev = NULL;
 	}
-	scat = &op->r_sg[0];
+	scat = &op->op_sg[0];
 	sent = 0;
-	num_sge = op->r_count;
+	num_sge = op->op_count;
 
-	for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
+	for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
 		send->s_wr.send_flags = 0;
 		send->s_queued = jiffies;
 
@@ -873,13 +874,13 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 		 * for local access after RDS is finished with it, using
 		 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
 		 */
-		if (op->r_write)
+		if (op->op_write)
 			send->s_wr.opcode = IB_WR_RDMA_WRITE;
 		else
 			send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
 
 		send->s_wr.wr.rdma.remote_addr = remote_addr;
-		send->s_wr.wr.rdma.rkey = op->r_key;
+		send->s_wr.wr.rdma.rkey = op->op_rkey;
 		send->s_op = op;
 
 		if (num_sge > rds_iwdev->max_sge) {
@@ -893,7 +894,7 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 		if (prev)
 			prev->s_wr.next = &send->s_wr;
 
-		for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
+		for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
 			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
 
 			if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
@@ -927,7 +928,7 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	}
 
 	/* if we finished the message then send completion owns it */
-	if (scat == &op->r_sg[op->r_count])
+	if (scat == &op->op_sg[op->op_count])
 		first->s_wr.send_flags = IB_SEND_SIGNALED;
 
 	if (i < work_alloc) {
@@ -941,9 +942,9 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	 * adapters do not allow using the lkey for this at all.  To bypass this use a
 	 * fastreg_mr (or possibly a dma_mr)
 	 */
-	if (!op->r_write) {
+	if (!op->op_write) {
 		rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
-			op->r_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
+			op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
 		work_alloc++;
 	}
 

net/rds/iw_sysctl.c

@@ -122,10 +122,10 @@ void rds_iw_sysctl_exit(void)
 		unregister_sysctl_table(rds_iw_sysctl_hdr);
 }
 
-int __init rds_iw_sysctl_init(void)
+int rds_iw_sysctl_init(void)
 {
 	rds_iw_sysctl_hdr = register_sysctl_paths(rds_iw_sysctl_path, rds_iw_sysctl_table);
-	if (rds_iw_sysctl_hdr == NULL)
+	if (!rds_iw_sysctl_hdr)
 		return -ENOMEM;
 	return 0;
 }

net/rds/loop.c

@@ -61,10 +61,17 @@ static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
 			 unsigned int hdr_off, unsigned int sg,
 			 unsigned int off)
 {
+	/* Do not send cong updates to loopback */
+	if (rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
+		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
+		return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
+	}
+
 	BUG_ON(hdr_off || sg || off);
 
 	rds_inc_init(&rm->m_inc, conn, conn->c_laddr);
-	rds_message_addref(rm); /* for the inc */
+	/* For the embedded inc. Matching put is in loop_inc_free() */
+	rds_message_addref(rm);
 
 	rds_recv_incoming(conn, conn->c_laddr, conn->c_faddr, &rm->m_inc,
 			  GFP_KERNEL, KM_USER0);
@@ -77,16 +84,14 @@ static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
 	return sizeof(struct rds_header) + be32_to_cpu(rm->m_inc.i_hdr.h_len);
 }
 
-static int rds_loop_xmit_cong_map(struct rds_connection *conn,
-				  struct rds_cong_map *map,
-				  unsigned long offset)
+/*
+ * See rds_loop_xmit(). Since our inc is embedded in the rm, we
+ * make sure the rm lives at least until the inc is done.
+ */
+static void rds_loop_inc_free(struct rds_incoming *inc)
 {
-	BUG_ON(offset);
-	BUG_ON(map != conn->c_lcong);
-
-	rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
-
-	return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
+        struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
+        rds_message_put(rm);
 }
 
 /* we need to at least give the thread something to succeed */
@@ -112,7 +117,7 @@ static int rds_loop_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 	unsigned long flags;
 
 	lc = kzalloc(sizeof(struct rds_loop_connection), GFP_KERNEL);
-	if (lc == NULL)
+	if (!lc)
 		return -ENOMEM;
 
 	INIT_LIST_HEAD(&lc->loop_node);
@@ -169,14 +174,12 @@ void rds_loop_exit(void)
  */
 struct rds_transport rds_loop_transport = {
 	.xmit			= rds_loop_xmit,
-	.xmit_cong_map		= rds_loop_xmit_cong_map,
 	.recv			= rds_loop_recv,
 	.conn_alloc		= rds_loop_conn_alloc,
 	.conn_free		= rds_loop_conn_free,
 	.conn_connect		= rds_loop_conn_connect,
 	.conn_shutdown		= rds_loop_conn_shutdown,
 	.inc_copy_to_user	= rds_message_inc_copy_to_user,
-	.inc_purge		= rds_message_inc_purge,
-	.inc_free		= rds_message_inc_free,
+	.inc_free		= rds_loop_inc_free,
 	.t_name			= "loopback",
 };

net/rds/message.c

@@ -34,9 +34,6 @@
 #include <linux/slab.h>
 
 #include "rds.h"
-#include "rdma.h"
-
-static DECLARE_WAIT_QUEUE_HEAD(rds_message_flush_waitq);
 
 static unsigned int	rds_exthdr_size[__RDS_EXTHDR_MAX] = {
 [RDS_EXTHDR_NONE]	= 0,
@@ -63,29 +60,31 @@ static void rds_message_purge(struct rds_message *rm)
 	if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
 		return;
 
-	for (i = 0; i < rm->m_nents; i++) {
-		rdsdebug("putting data page %p\n", (void *)sg_page(&rm->m_sg[i]));
+	for (i = 0; i < rm->data.op_nents; i++) {
+		rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.op_sg[i]));
 		/* XXX will have to put_page for page refs */
-		__free_page(sg_page(&rm->m_sg[i]));
+		__free_page(sg_page(&rm->data.op_sg[i]));
 	}
-	rm->m_nents = 0;
+	rm->data.op_nents = 0;
 
-	if (rm->m_rdma_op)
-		rds_rdma_free_op(rm->m_rdma_op);
-	if (rm->m_rdma_mr)
-		rds_mr_put(rm->m_rdma_mr);
-}
+	if (rm->rdma.op_active)
+		rds_rdma_free_op(&rm->rdma);
+	if (rm->rdma.op_rdma_mr)
+		rds_mr_put(rm->rdma.op_rdma_mr);
 
-void rds_message_inc_purge(struct rds_incoming *inc)
-{
-	struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
-	rds_message_purge(rm);
+	if (rm->atomic.op_active)
+		rds_atomic_free_op(&rm->atomic);
+	if (rm->atomic.op_rdma_mr)
+		rds_mr_put(rm->atomic.op_rdma_mr);
 }
 
 void rds_message_put(struct rds_message *rm)
 {
 	rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
-
+	if (atomic_read(&rm->m_refcount) == 0) {
+printk(KERN_CRIT "danger refcount zero on %p\n", rm);
+WARN_ON(1);
+	}
 	if (atomic_dec_and_test(&rm->m_refcount)) {
 		BUG_ON(!list_empty(&rm->m_sock_item));
 		BUG_ON(!list_empty(&rm->m_conn_item));
@@ -96,12 +95,6 @@ void rds_message_put(struct rds_message *rm)
 }
 EXPORT_SYMBOL_GPL(rds_message_put);
 
-void rds_message_inc_free(struct rds_incoming *inc)
-{
-	struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
-	rds_message_put(rm);
-}
-
 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
 				 __be16 dport, u64 seq)
 {
@@ -214,41 +207,68 @@ int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 o
 }
 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
 
-struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp)
+/*
+ * Each rds_message is allocated with extra space for the scatterlist entries
+ * rds ops will need. This is to minimize memory allocation count. Then, each rds op
+ * can grab SGs when initializing its part of the rds_message.
+ */
+struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
 {
 	struct rds_message *rm;
 
-	rm = kzalloc(sizeof(struct rds_message) +
-		     (nents * sizeof(struct scatterlist)), gfp);
+	rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
 	if (!rm)
 		goto out;
 
-	if (nents)
-		sg_init_table(rm->m_sg, nents);
+	rm->m_used_sgs = 0;
+	rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
+
 	atomic_set(&rm->m_refcount, 1);
 	INIT_LIST_HEAD(&rm->m_sock_item);
 	INIT_LIST_HEAD(&rm->m_conn_item);
 	spin_lock_init(&rm->m_rs_lock);
+	init_waitqueue_head(&rm->m_flush_wait);
 
 out:
 	return rm;
 }
 
+/*
+ * RDS ops use this to grab SG entries from the rm's sg pool.
+ */
+struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
+{
+	struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
+	struct scatterlist *sg_ret;
+
+	WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
+	WARN_ON(!nents);
+
+	sg_ret = &sg_first[rm->m_used_sgs];
+	sg_init_table(sg_ret, nents);
+	rm->m_used_sgs += nents;
+
+	return sg_ret;
+}
+
 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
 {
 	struct rds_message *rm;
 	unsigned int i;
+	int num_sgs = ceil(total_len, PAGE_SIZE);
+	int extra_bytes = num_sgs * sizeof(struct scatterlist);
 
-	rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
-	if (rm == NULL)
+	rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
+	if (!rm)
 		return ERR_PTR(-ENOMEM);
 
 	set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
 	rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
-	rm->m_nents = ceil(total_len, PAGE_SIZE);
+	rm->data.op_nents = ceil(total_len, PAGE_SIZE);
+	rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
 
-	for (i = 0; i < rm->m_nents; ++i) {
-		sg_set_page(&rm->m_sg[i],
+	for (i = 0; i < rm->data.op_nents; ++i) {
+		sg_set_page(&rm->data.op_sg[i],
 				virt_to_page(page_addrs[i]),
 				PAGE_SIZE, 0);
 	}
@@ -256,40 +276,33 @@ struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned in
 	return rm;
 }
 
-struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
+int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
 					       size_t total_len)
 {
 	unsigned long to_copy;
 	unsigned long iov_off;
 	unsigned long sg_off;
-	struct rds_message *rm;
 	struct iovec *iov;
 	struct scatterlist *sg;
-	int ret;
-
-	rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
-	if (rm == NULL) {
-		ret = -ENOMEM;
-		goto out;
-	}
+	int ret = 0;
 
 	rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
 
 	/*
 	 * now allocate and copy in the data payload.
 	 */
-	sg = rm->m_sg;
+	sg = rm->data.op_sg;
 	iov = first_iov;
 	iov_off = 0;
 	sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
 
 	while (total_len) {
-		if (sg_page(sg) == NULL) {
+		if (!sg_page(sg)) {
 			ret = rds_page_remainder_alloc(sg, total_len,
 						       GFP_HIGHUSER);
 			if (ret)
 				goto out;
-			rm->m_nents++;
+			rm->data.op_nents++;
 			sg_off = 0;
 		}
 
@@ -320,14 +333,8 @@ struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
 			sg++;
 	}
 
-	ret = 0;
 out:
-	if (ret) {
-		if (rm)
-			rds_message_put(rm);
-		rm = ERR_PTR(ret);
-	}
-	return rm;
+	return ret;
 }
 
 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
@@ -348,7 +355,7 @@ int rds_message_inc_copy_to_user(struct rds_incoming *inc,
 
 	iov = first_iov;
 	iov_off = 0;
-	sg = rm->m_sg;
+	sg = rm->data.op_sg;
 	vec_off = 0;
 	copied = 0;
 
@@ -394,15 +401,14 @@ int rds_message_inc_copy_to_user(struct rds_incoming *inc,
  */
 void rds_message_wait(struct rds_message *rm)
 {
-	wait_event(rds_message_flush_waitq,
+	wait_event_interruptible(rm->m_flush_wait,
 			!test_bit(RDS_MSG_MAPPED, &rm->m_flags));
 }
 
 void rds_message_unmapped(struct rds_message *rm)
 {
 	clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
-	if (waitqueue_active(&rds_message_flush_waitq))
-		wake_up(&rds_message_flush_waitq);
+	wake_up_interruptible(&rm->m_flush_wait);
 }
 EXPORT_SYMBOL_GPL(rds_message_unmapped);
 

net/rds/page.c

@@ -116,7 +116,7 @@ int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
 	/* jump straight to allocation if we're trying for a huge page */
 	if (bytes >= PAGE_SIZE) {
 		page = alloc_page(gfp);
-		if (page == NULL) {
+		if (!page) {
 			ret = -ENOMEM;
 		} else {
 			sg_set_page(scat, page, PAGE_SIZE, 0);
@@ -162,7 +162,7 @@ int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
 		rem = &per_cpu(rds_page_remainders, get_cpu());
 		local_irq_save(flags);
 
-		if (page == NULL) {
+		if (!page) {
 			ret = -ENOMEM;
 			break;
 		}
@@ -186,6 +186,7 @@ out:
 		 ret ? 0 : scat->length);
 	return ret;
 }
+EXPORT_SYMBOL_GPL(rds_page_remainder_alloc);
 
 static int rds_page_remainder_cpu_notify(struct notifier_block *self,
 					 unsigned long action, void *hcpu)

net/rds/rdma.c

@@ -35,7 +35,7 @@
 #include <linux/rbtree.h>
 #include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
 
-#include "rdma.h"
+#include "rds.h"
 
 /*
  * XXX
@@ -130,14 +130,22 @@ void rds_rdma_drop_keys(struct rds_sock *rs)
 {
 	struct rds_mr *mr;
 	struct rb_node *node;
+	unsigned long flags;
 
 	/* Release any MRs associated with this socket */
+	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
 	while ((node = rb_first(&rs->rs_rdma_keys))) {
 		mr = container_of(node, struct rds_mr, r_rb_node);
 		if (mr->r_trans == rs->rs_transport)
 			mr->r_invalidate = 0;
+		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
+		RB_CLEAR_NODE(&mr->r_rb_node);
+		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+		rds_destroy_mr(mr);
 		rds_mr_put(mr);
+		spin_lock_irqsave(&rs->rs_rdma_lock, flags);
 	}
+	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
 
 	if (rs->rs_transport && rs->rs_transport->flush_mrs)
 		rs->rs_transport->flush_mrs();
@@ -181,7 +189,7 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 		goto out;
 	}
 
-	if (rs->rs_transport->get_mr == NULL) {
+	if (!rs->rs_transport->get_mr) {
 		ret = -EOPNOTSUPP;
 		goto out;
 	}
@@ -197,13 +205,13 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 
 	/* XXX clamp nr_pages to limit the size of this alloc? */
 	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
-	if (pages == NULL) {
+	if (!pages) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
 	mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
-	if (mr == NULL) {
+	if (!mr) {
 		ret = -ENOMEM;
 		goto out;
 	}
@@ -230,13 +238,13 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 	 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
 	 * the zero page.
 	 */
-	ret = rds_pin_pages(args->vec.addr & PAGE_MASK, nr_pages, pages, 1);
+	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
 	if (ret < 0)
 		goto out;
 
 	nents = ret;
 	sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
-	if (sg == NULL) {
+	if (!sg) {
 		ret = -ENOMEM;
 		goto out;
 	}
@@ -406,68 +414,127 @@ void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
 
 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
 	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
-	if (mr && (mr->r_use_once || force)) {
+	if (!mr) {
+		printk(KERN_ERR "rds: trying to unuse MR with unknown r_key %u!\n", r_key);
+		spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
+		return;
+	}
+
+	if (mr->r_use_once || force) {
 		rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
 		RB_CLEAR_NODE(&mr->r_rb_node);
 		zot_me = 1;
-	} else if (mr)
-		atomic_inc(&mr->r_refcount);
+	}
 	spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
 
 	/* May have to issue a dma_sync on this memory region.
 	 * Note we could avoid this if the operation was a RDMA READ,
 	 * but at this point we can't tell. */
-	if (mr != NULL) {
-		if (mr->r_trans->sync_mr)
-			mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
-
-		/* If the MR was marked as invalidate, this will
-		 * trigger an async flush. */
-		if (zot_me)
-			rds_destroy_mr(mr);
-		rds_mr_put(mr);
-	}
+	if (mr->r_trans->sync_mr)
+		mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
+
+	/* If the MR was marked as invalidate, this will
+	 * trigger an async flush. */
+	if (zot_me)
+		rds_destroy_mr(mr);
+	rds_mr_put(mr);
 }
 
-void rds_rdma_free_op(struct rds_rdma_op *ro)
+void rds_rdma_free_op(struct rm_rdma_op *ro)
 {
 	unsigned int i;
 
-	for (i = 0; i < ro->r_nents; i++) {
-		struct page *page = sg_page(&ro->r_sg[i]);
+	for (i = 0; i < ro->op_nents; i++) {
+		struct page *page = sg_page(&ro->op_sg[i]);
 
 		/* Mark page dirty if it was possibly modified, which
 		 * is the case for a RDMA_READ which copies from remote
 		 * to local memory */
-		if (!ro->r_write) {
-			BUG_ON(in_interrupt());
+		if (!ro->op_write) {
+			BUG_ON(irqs_disabled());
 			set_page_dirty(page);
 		}
 		put_page(page);
 	}
 
-	kfree(ro->r_notifier);
-	kfree(ro);
+	kfree(ro->op_notifier);
+	ro->op_notifier = NULL;
+	ro->op_active = 0;
+}
+
+void rds_atomic_free_op(struct rm_atomic_op *ao)
+{
+	struct page *page = sg_page(ao->op_sg);
+
+	/* Mark page dirty if it was possibly modified, which
+	 * is the case for a RDMA_READ which copies from remote
+	 * to local memory */
+	set_page_dirty(page);
+	put_page(page);
+
+	kfree(ao->op_notifier);
+	ao->op_notifier = NULL;
+	ao->op_active = 0;
+}
+
+
+/*
+ * Count the number of pages needed to describe an incoming iovec.
+ */
+static int rds_rdma_pages(struct rds_rdma_args *args)
+{
+	struct rds_iovec vec;
+	struct rds_iovec __user *local_vec;
+	unsigned int tot_pages = 0;
+	unsigned int nr_pages;
+	unsigned int i;
+
+	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
+
+	/* figure out the number of pages in the vector */
+	for (i = 0; i < args->nr_local; i++) {
+		if (copy_from_user(&vec, &local_vec[i],
+				   sizeof(struct rds_iovec)))
+			return -EFAULT;
+
+		nr_pages = rds_pages_in_vec(&vec);
+		if (nr_pages == 0)
+			return -EINVAL;
+
+		tot_pages += nr_pages;
+	}
+
+	return tot_pages;
+}
+
+int rds_rdma_extra_size(struct rds_rdma_args *args)
+{
+	return rds_rdma_pages(args) * sizeof(struct scatterlist);
 }
 
 /*
- * args is a pointer to an in-kernel copy in the sendmsg cmsg.
+ * The application asks for a RDMA transfer.
+ * Extract all arguments and set up the rdma_op
  */
-static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
-					    struct rds_rdma_args *args)
+int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg)
 {
+	struct rds_rdma_args *args;
 	struct rds_iovec vec;
-	struct rds_rdma_op *op = NULL;
+	struct rm_rdma_op *op = &rm->rdma;
 	unsigned int nr_pages;
-	unsigned int max_pages;
 	unsigned int nr_bytes;
 	struct page **pages = NULL;
 	struct rds_iovec __user *local_vec;
-	struct scatterlist *sg;
 	unsigned int nr;
 	unsigned int i, j;
-	int ret;
+	int ret = 0;
+
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
+	    || rm->rdma.op_active)
+		return -EINVAL;
 
+	args = CMSG_DATA(cmsg);
 
 	if (rs->rs_bound_addr == 0) {
 		ret = -ENOTCONN; /* XXX not a great errno */
@@ -479,61 +546,38 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 		goto out;
 	}
 
-	nr_pages = 0;
-	max_pages = 0;
-
-	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
-
-	/* figure out the number of pages in the vector */
-	for (i = 0; i < args->nr_local; i++) {
-		if (copy_from_user(&vec, &local_vec[i],
-				   sizeof(struct rds_iovec))) {
-			ret = -EFAULT;
-			goto out;
-		}
-
-		nr = rds_pages_in_vec(&vec);
-		if (nr == 0) {
-			ret = -EINVAL;
-			goto out;
-		}
-
-		max_pages = max(nr, max_pages);
-		nr_pages += nr;
-	}
-
-	pages = kcalloc(max_pages, sizeof(struct page *), GFP_KERNEL);
-	if (pages == NULL) {
-		ret = -ENOMEM;
+	nr_pages = rds_rdma_pages(args);
+	if (nr_pages < 0)
 		goto out;
-	}
 
-	op = kzalloc(offsetof(struct rds_rdma_op, r_sg[nr_pages]), GFP_KERNEL);
-	if (op == NULL) {
+	pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
+	if (!pages) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
-	op->r_write = !!(args->flags & RDS_RDMA_READWRITE);
-	op->r_fence = !!(args->flags & RDS_RDMA_FENCE);
-	op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
-	op->r_recverr = rs->rs_recverr;
+	op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
+	op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
+	op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
+	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
+	op->op_active = 1;
+	op->op_recverr = rs->rs_recverr;
 	WARN_ON(!nr_pages);
-	sg_init_table(op->r_sg, nr_pages);
+	op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
 
-	if (op->r_notify || op->r_recverr) {
+	if (op->op_notify || op->op_recverr) {
 		/* We allocate an uninitialized notifier here, because
 		 * we don't want to do that in the completion handler. We
 		 * would have to use GFP_ATOMIC there, and don't want to deal
 		 * with failed allocations.
 		 */
-		op->r_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
-		if (!op->r_notifier) {
+		op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
+		if (!op->op_notifier) {
 			ret = -ENOMEM;
 			goto out;
 		}
-		op->r_notifier->n_user_token = args->user_token;
-		op->r_notifier->n_status = RDS_RDMA_SUCCESS;
+		op->op_notifier->n_user_token = args->user_token;
+		op->op_notifier->n_status = RDS_RDMA_SUCCESS;
 	}
 
 	/* The cookie contains the R_Key of the remote memory region, and
@@ -543,15 +587,17 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 	 * destination address (which is really an offset into the MR)
 	 * FIXME: We may want to move this into ib_rdma.c
 	 */
-	op->r_key = rds_rdma_cookie_key(args->cookie);
-	op->r_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
+	op->op_rkey = rds_rdma_cookie_key(args->cookie);
+	op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
 
 	nr_bytes = 0;
 
 	rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
 	       (unsigned long long)args->nr_local,
 	       (unsigned long long)args->remote_vec.addr,
-	       op->r_key);
+	       op->op_rkey);
+
+	local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
 
 	for (i = 0; i < args->nr_local; i++) {
 		if (copy_from_user(&vec, &local_vec[i],
@@ -569,15 +615,10 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 		rs->rs_user_addr = vec.addr;
 		rs->rs_user_bytes = vec.bytes;
 
-		/* did the user change the vec under us? */
-		if (nr > max_pages || op->r_nents + nr > nr_pages) {
-			ret = -EINVAL;
-			goto out;
-		}
 		/* If it's a WRITE operation, we want to pin the pages for reading.
 		 * If it's a READ operation, we need to pin the pages for writing.
 		 */
-		ret = rds_pin_pages(vec.addr & PAGE_MASK, nr, pages, !op->r_write);
+		ret = rds_pin_pages(vec.addr, nr, pages, !op->op_write);
 		if (ret < 0)
 			goto out;
 
@@ -588,8 +629,9 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 
 		for (j = 0; j < nr; j++) {
 			unsigned int offset = vec.addr & ~PAGE_MASK;
+			struct scatterlist *sg;
 
-			sg = &op->r_sg[op->r_nents + j];
+			sg = &op->op_sg[op->op_nents + j];
 			sg_set_page(sg, pages[j],
 					min_t(unsigned int, vec.bytes, PAGE_SIZE - offset),
 					offset);
@@ -601,10 +643,9 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 			vec.bytes -= sg->length;
 		}
 
-		op->r_nents += nr;
+		op->op_nents += nr;
 	}
 
-
 	if (nr_bytes > args->remote_vec.bytes) {
 		rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
 				nr_bytes,
@@ -612,38 +653,17 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
 		ret = -EINVAL;
 		goto out;
 	}
-	op->r_bytes = nr_bytes;
+	op->op_bytes = nr_bytes;
 
 	ret = 0;
 out:
 	kfree(pages);
-	if (ret) {
-		if (op)
-			rds_rdma_free_op(op);
-		op = ERR_PTR(ret);
-	}
-	return op;
-}
-
-/*
- * The application asks for a RDMA transfer.
- * Extract all arguments and set up the rdma_op
- */
-int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
-			  struct cmsghdr *cmsg)
-{
-	struct rds_rdma_op *op;
-
-	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) ||
-	    rm->m_rdma_op != NULL)
-		return -EINVAL;
+	if (ret)
+		rds_rdma_free_op(op);
 
-	op = rds_rdma_prepare(rs, CMSG_DATA(cmsg));
-	if (IS_ERR(op))
-		return PTR_ERR(op);
 	rds_stats_inc(s_send_rdma);
-	rm->m_rdma_op = op;
-	return 0;
+
+	return ret;
 }
 
 /*
@@ -673,7 +693,7 @@ int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
 
 	spin_lock_irqsave(&rs->rs_rdma_lock, flags);
 	mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
-	if (mr == NULL)
+	if (!mr)
 		err = -EINVAL;	/* invalid r_key */
 	else
 		atomic_inc(&mr->r_refcount);
@@ -681,7 +701,7 @@ int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
 
 	if (mr) {
 		mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
-		rm->m_rdma_mr = mr;
+		rm->rdma.op_rdma_mr = mr;
 	}
 	return err;
 }
@@ -699,5 +719,98 @@ int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
 	    rm->m_rdma_cookie != 0)
 		return -EINVAL;
 
-	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->m_rdma_mr);
+	return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr);
+}
+
+/*
+ * Fill in rds_message for an atomic request.
+ */
+int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
+		    struct cmsghdr *cmsg)
+{
+	struct page *page = NULL;
+	struct rds_atomic_args *args;
+	int ret = 0;
+
+	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
+	 || rm->atomic.op_active)
+		return -EINVAL;
+
+	args = CMSG_DATA(cmsg);
+
+	/* Nonmasked & masked cmsg ops converted to masked hw ops */
+	switch (cmsg->cmsg_type) {
+	case RDS_CMSG_ATOMIC_FADD:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
+		rm->atomic.op_m_fadd.add = args->fadd.add;
+		rm->atomic.op_m_fadd.nocarry_mask = 0;
+		break;
+	case RDS_CMSG_MASKED_ATOMIC_FADD:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
+		rm->atomic.op_m_fadd.add = args->m_fadd.add;
+		rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
+		break;
+	case RDS_CMSG_ATOMIC_CSWP:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
+		rm->atomic.op_m_cswp.compare = args->cswp.compare;
+		rm->atomic.op_m_cswp.swap = args->cswp.swap;
+		rm->atomic.op_m_cswp.compare_mask = ~0;
+		rm->atomic.op_m_cswp.swap_mask = ~0;
+		break;
+	case RDS_CMSG_MASKED_ATOMIC_CSWP:
+		rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
+		rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
+		rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
+		rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
+		rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
+		break;
+	default:
+		BUG(); /* should never happen */
+	}
+
+	rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
+	rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
+	rm->atomic.op_active = 1;
+	rm->atomic.op_recverr = rs->rs_recverr;
+	rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
+
+	/* verify 8 byte-aligned */
+	if (args->local_addr & 0x7) {
+		ret = -EFAULT;
+		goto err;
+	}
+
+	ret = rds_pin_pages(args->local_addr, 1, &page, 1);
+	if (ret != 1)
+		goto err;
+	ret = 0;
+
+	sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
+
+	if (rm->atomic.op_notify || rm->atomic.op_recverr) {
+		/* We allocate an uninitialized notifier here, because
+		 * we don't want to do that in the completion handler. We
+		 * would have to use GFP_ATOMIC there, and don't want to deal
+		 * with failed allocations.
+		 */
+		rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
+		if (!rm->atomic.op_notifier) {
+			ret = -ENOMEM;
+			goto err;
+		}
+
+		rm->atomic.op_notifier->n_user_token = args->user_token;
+		rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
+	}
+
+	rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
+	rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
+
+	return ret;
+err:
+	if (page)
+		put_page(page);
+	kfree(rm->atomic.op_notifier);
+
+	return ret;
 }

net/rds/rdma.h

@@ -1,85 +0,0 @@
-#ifndef _RDS_RDMA_H
-#define _RDS_RDMA_H
-
-#include <linux/rbtree.h>
-#include <linux/spinlock.h>
-#include <linux/scatterlist.h>
-
-#include "rds.h"
-
-struct rds_mr {
-	struct rb_node		r_rb_node;
-	atomic_t		r_refcount;
-	u32			r_key;
-
-	/* A copy of the creation flags */
-	unsigned int		r_use_once:1;
-	unsigned int		r_invalidate:1;
-	unsigned int		r_write:1;
-
-	/* This is for RDS_MR_DEAD.
-	 * It would be nice & consistent to make this part of the above
-	 * bit field here, but we need to use test_and_set_bit.
-	 */
-	unsigned long		r_state;
-	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
-	struct rds_transport	*r_trans;
-	void			*r_trans_private;
-};
-
-/* Flags for mr->r_state */
-#define RDS_MR_DEAD		0
-
-struct rds_rdma_op {
-	u32			r_key;
-	u64			r_remote_addr;
-	unsigned int		r_write:1;
-	unsigned int		r_fence:1;
-	unsigned int		r_notify:1;
-	unsigned int		r_recverr:1;
-	unsigned int		r_mapped:1;
-	struct rds_notifier	*r_notifier;
-	unsigned int		r_bytes;
-	unsigned int		r_nents;
-	unsigned int		r_count;
-	struct scatterlist	r_sg[0];
-};
-
-static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
-{
-	return r_key | (((u64) offset) << 32);
-}
-
-static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
-{
-	return cookie;
-}
-
-static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
-{
-	return cookie >> 32;
-}
-
-int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
-int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
-int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
-void rds_rdma_drop_keys(struct rds_sock *rs);
-int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
-			  struct cmsghdr *cmsg);
-int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
-			  struct cmsghdr *cmsg);
-int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
-			  struct cmsghdr *cmsg);
-int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
-			  struct cmsghdr *cmsg);
-void rds_rdma_free_op(struct rds_rdma_op *ro);
-void rds_rdma_send_complete(struct rds_message *rm, int);
-
-extern void __rds_put_mr_final(struct rds_mr *mr);
-static inline void rds_mr_put(struct rds_mr *mr)
-{
-	if (atomic_dec_and_test(&mr->r_refcount))
-		__rds_put_mr_final(mr);
-}
-
-#endif

net/rds/rdma_transport.c

@@ -36,6 +36,34 @@
 
 static struct rdma_cm_id *rds_rdma_listen_id;
 
+static char *rds_cm_event_strings[] = {
+#define RDS_CM_EVENT_STRING(foo) \
+		[RDMA_CM_EVENT_##foo] = __stringify(RDMA_CM_EVENT_##foo)
+	RDS_CM_EVENT_STRING(ADDR_RESOLVED),
+	RDS_CM_EVENT_STRING(ADDR_ERROR),
+	RDS_CM_EVENT_STRING(ROUTE_RESOLVED),
+	RDS_CM_EVENT_STRING(ROUTE_ERROR),
+	RDS_CM_EVENT_STRING(CONNECT_REQUEST),
+	RDS_CM_EVENT_STRING(CONNECT_RESPONSE),
+	RDS_CM_EVENT_STRING(CONNECT_ERROR),
+	RDS_CM_EVENT_STRING(UNREACHABLE),
+	RDS_CM_EVENT_STRING(REJECTED),
+	RDS_CM_EVENT_STRING(ESTABLISHED),
+	RDS_CM_EVENT_STRING(DISCONNECTED),
+	RDS_CM_EVENT_STRING(DEVICE_REMOVAL),
+	RDS_CM_EVENT_STRING(MULTICAST_JOIN),
+	RDS_CM_EVENT_STRING(MULTICAST_ERROR),
+	RDS_CM_EVENT_STRING(ADDR_CHANGE),
+	RDS_CM_EVENT_STRING(TIMEWAIT_EXIT),
+#undef RDS_CM_EVENT_STRING
+};
+
+static char *rds_cm_event_str(enum rdma_cm_event_type type)
+{
+	return rds_str_array(rds_cm_event_strings,
+			     ARRAY_SIZE(rds_cm_event_strings), type);
+};
+
 int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
 			      struct rdma_cm_event *event)
 {
@@ -44,8 +72,8 @@ int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
 	struct rds_transport *trans;
 	int ret = 0;
 
-	rdsdebug("conn %p id %p handling event %u\n", conn, cm_id,
-		 event->event);
+	rdsdebug("conn %p id %p handling event %u (%s)\n", conn, cm_id,
+		 event->event, rds_cm_event_str(event->event));
 
 	if (cm_id->device->node_type == RDMA_NODE_RNIC)
 		trans = &rds_iw_transport;
@@ -109,7 +137,8 @@ int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
 
 	default:
 		/* things like device disconnect? */
-		printk(KERN_ERR "RDS: unknown event %u!\n", event->event);
+		printk(KERN_ERR "RDS: unknown event %u (%s)!\n",
+		       event->event, rds_cm_event_str(event->event));
 		break;
 	}
 
@@ -117,12 +146,13 @@ out:
 	if (conn)
 		mutex_unlock(&conn->c_cm_lock);
 
-	rdsdebug("id %p event %u handling ret %d\n", cm_id, event->event, ret);
+	rdsdebug("id %p event %u (%s) handling ret %d\n", cm_id, event->event,
+		 rds_cm_event_str(event->event), ret);
 
 	return ret;
 }
 
-static int __init rds_rdma_listen_init(void)
+static int rds_rdma_listen_init(void)
 {
 	struct sockaddr_in sin;
 	struct rdma_cm_id *cm_id;
@@ -177,7 +207,7 @@ static void rds_rdma_listen_stop(void)
 	}
 }
 
-int __init rds_rdma_init(void)
+int rds_rdma_init(void)
 {
 	int ret;
 

net/rds/rds.h

@@ -80,6 +80,7 @@ enum {
 /* Bits for c_flags */
 #define RDS_LL_SEND_FULL	0
 #define RDS_RECONNECT_PENDING	1
+#define RDS_IN_XMIT		2
 
 struct rds_connection {
 	struct hlist_node	c_hash_node;
@@ -91,12 +92,13 @@ struct rds_connection {
 	struct rds_cong_map	*c_lcong;
 	struct rds_cong_map	*c_fcong;
 
-	struct mutex		c_send_lock;	/* protect send ring */
 	struct rds_message	*c_xmit_rm;
 	unsigned long		c_xmit_sg;
 	unsigned int		c_xmit_hdr_off;
 	unsigned int		c_xmit_data_off;
+	unsigned int		c_xmit_atomic_sent;
 	unsigned int		c_xmit_rdma_sent;
+	unsigned int		c_xmit_data_sent;
 
 	spinlock_t		c_lock;		/* protect msg queues */
 	u64			c_next_tx_seq;
@@ -116,11 +118,10 @@ struct rds_connection {
 	struct delayed_work	c_conn_w;
 	struct work_struct	c_down_w;
 	struct mutex		c_cm_lock;	/* protect conn state & cm */
+	wait_queue_head_t	c_waitq;
 
 	struct list_head	c_map_item;
 	unsigned long		c_map_queued;
-	unsigned long		c_map_offset;
-	unsigned long		c_map_bytes;
 
 	unsigned int		c_unacked_packets;
 	unsigned int		c_unacked_bytes;
@@ -206,6 +207,48 @@ struct rds_incoming {
 	rds_rdma_cookie_t	i_rdma_cookie;
 };
 
+struct rds_mr {
+	struct rb_node		r_rb_node;
+	atomic_t		r_refcount;
+	u32			r_key;
+
+	/* A copy of the creation flags */
+	unsigned int		r_use_once:1;
+	unsigned int		r_invalidate:1;
+	unsigned int		r_write:1;
+
+	/* This is for RDS_MR_DEAD.
+	 * It would be nice & consistent to make this part of the above
+	 * bit field here, but we need to use test_and_set_bit.
+	 */
+	unsigned long		r_state;
+	struct rds_sock		*r_sock; /* back pointer to the socket that owns us */
+	struct rds_transport	*r_trans;
+	void			*r_trans_private;
+};
+
+/* Flags for mr->r_state */
+#define RDS_MR_DEAD		0
+
+static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
+{
+	return r_key | (((u64) offset) << 32);
+}
+
+static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
+{
+	return cookie;
+}
+
+static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
+{
+	return cookie >> 32;
+}
+
+/* atomic operation types */
+#define RDS_ATOMIC_TYPE_CSWP		0
+#define RDS_ATOMIC_TYPE_FADD		1
+
 /*
  * m_sock_item and m_conn_item are on lists that are serialized under
  * conn->c_lock.  m_sock_item has additional meaning in that once it is empty
@@ -258,13 +301,71 @@ struct rds_message {
 	 *   -> rs->rs_lock
 	 */
 	spinlock_t		m_rs_lock;
+	wait_queue_head_t	m_flush_wait;
+
 	struct rds_sock		*m_rs;
-	struct rds_rdma_op	*m_rdma_op;
+
+	/* cookie to send to remote, in rds header */
 	rds_rdma_cookie_t	m_rdma_cookie;
-	struct rds_mr		*m_rdma_mr;
-	unsigned int		m_nents;
-	unsigned int		m_count;
-	struct scatterlist	m_sg[0];
+
+	unsigned int		m_used_sgs;
+	unsigned int		m_total_sgs;
+
+	void			*m_final_op;
+
+	struct {
+		struct rm_atomic_op {
+			int			op_type;
+			union {
+				struct {
+					uint64_t	compare;
+					uint64_t	swap;
+					uint64_t	compare_mask;
+					uint64_t	swap_mask;
+				} op_m_cswp;
+				struct {
+					uint64_t	add;
+					uint64_t	nocarry_mask;
+				} op_m_fadd;
+			};
+
+			u32			op_rkey;
+			u64			op_remote_addr;
+			unsigned int		op_notify:1;
+			unsigned int		op_recverr:1;
+			unsigned int		op_mapped:1;
+			unsigned int		op_silent:1;
+			unsigned int		op_active:1;
+			struct scatterlist	*op_sg;
+			struct rds_notifier	*op_notifier;
+
+			struct rds_mr		*op_rdma_mr;
+		} atomic;
+		struct rm_rdma_op {
+			u32			op_rkey;
+			u64			op_remote_addr;
+			unsigned int		op_write:1;
+			unsigned int		op_fence:1;
+			unsigned int		op_notify:1;
+			unsigned int		op_recverr:1;
+			unsigned int		op_mapped:1;
+			unsigned int		op_silent:1;
+			unsigned int		op_active:1;
+			unsigned int		op_bytes;
+			unsigned int		op_nents;
+			unsigned int		op_count;
+			struct scatterlist	*op_sg;
+			struct rds_notifier	*op_notifier;
+
+			struct rds_mr		*op_rdma_mr;
+		} rdma;
+		struct rm_data_op {
+			unsigned int		op_active:1;
+			unsigned int		op_nents;
+			unsigned int		op_count;
+			struct scatterlist	*op_sg;
+		} data;
+	};
 };
 
 /*
@@ -305,10 +406,6 @@ struct rds_notifier {
  *                 transport is responsible for other serialization, including
  *                 rds_recv_incoming().  This is called in process context but
  *                 should try hard not to block.
- *
- * @xmit_cong_map: This asks the transport to send the local bitmap down the
- * 		   given connection.  XXX get a better story about the bitmap
- * 		   flag and header.
  */
 
 #define RDS_TRANS_IB	0
@@ -332,13 +429,11 @@ struct rds_transport {
 	void (*xmit_complete)(struct rds_connection *conn);
 	int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
 		    unsigned int hdr_off, unsigned int sg, unsigned int off);
-	int (*xmit_cong_map)(struct rds_connection *conn,
-			     struct rds_cong_map *map, unsigned long offset);
-	int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
+	int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
+	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
 	int (*recv)(struct rds_connection *conn);
 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
 				size_t size);
-	void (*inc_purge)(struct rds_incoming *inc);
 	void (*inc_free)(struct rds_incoming *inc);
 
 	int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
@@ -367,17 +462,11 @@ struct rds_sock {
 	 * bound_addr used for both incoming and outgoing, no INADDR_ANY
 	 * support.
 	 */
-	struct rb_node		rs_bound_node;
+	struct hlist_node	rs_bound_node;
 	__be32			rs_bound_addr;
 	__be32			rs_conn_addr;
 	__be16			rs_bound_port;
 	__be16			rs_conn_port;
-
-	/*
-	 * This is only used to communicate the transport between bind and
-	 * initiating connections.  All other trans use is referenced through
-	 * the connection.
-	 */
 	struct rds_transport    *rs_transport;
 
 	/*
@@ -466,8 +555,8 @@ struct rds_statistics {
 	uint64_t	s_recv_ping;
 	uint64_t	s_send_queue_empty;
 	uint64_t	s_send_queue_full;
-	uint64_t	s_send_sem_contention;
-	uint64_t	s_send_sem_queue_raced;
+	uint64_t	s_send_lock_contention;
+	uint64_t	s_send_lock_queue_raced;
 	uint64_t	s_send_immediate_retry;
 	uint64_t	s_send_delayed_retry;
 	uint64_t	s_send_drop_acked;
@@ -487,6 +576,7 @@ struct rds_statistics {
 };
 
 /* af_rds.c */
+char *rds_str_array(char **array, size_t elements, size_t index);
 void rds_sock_addref(struct rds_sock *rs);
 void rds_sock_put(struct rds_sock *rs);
 void rds_wake_sk_sleep(struct rds_sock *rs);
@@ -521,15 +611,17 @@ void rds_cong_exit(void);
 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
 
 /* conn.c */
-int __init rds_conn_init(void);
+int rds_conn_init(void);
 void rds_conn_exit(void);
 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
 				       struct rds_transport *trans, gfp_t gfp);
 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
 			       struct rds_transport *trans, gfp_t gfp);
+void rds_conn_shutdown(struct rds_connection *conn);
 void rds_conn_destroy(struct rds_connection *conn);
 void rds_conn_reset(struct rds_connection *conn);
 void rds_conn_drop(struct rds_connection *conn);
+void rds_conn_connect_if_down(struct rds_connection *conn);
 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
 			  struct rds_info_iterator *iter,
 			  struct rds_info_lengths *lens,
@@ -566,7 +658,8 @@ rds_conn_connecting(struct rds_connection *conn)
 
 /* message.c */
 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
-struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
+struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
+int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
 					       size_t total_len);
 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
@@ -580,7 +673,6 @@ int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *vers
 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
 				 struct iovec *first_iov, size_t size);
-void rds_message_inc_purge(struct rds_incoming *inc);
 void rds_message_inc_free(struct rds_incoming *inc);
 void rds_message_addref(struct rds_message *rm);
 void rds_message_put(struct rds_message *rm);
@@ -636,14 +728,39 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
 typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
 			 is_acked_func is_acked);
-int rds_send_acked_before(struct rds_connection *conn, u64 seq);
 void rds_send_remove_from_sock(struct list_head *messages, int status);
 int rds_send_pong(struct rds_connection *conn, __be16 dport);
 struct rds_message *rds_send_get_message(struct rds_connection *,
-					 struct rds_rdma_op *);
+					 struct rm_rdma_op *);
 
 /* rdma.c */
 void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
+int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
+int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
+int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
+void rds_rdma_drop_keys(struct rds_sock *rs);
+int rds_rdma_extra_size(struct rds_rdma_args *args);
+int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg);
+int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg);
+int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg);
+int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
+			  struct cmsghdr *cmsg);
+void rds_rdma_free_op(struct rm_rdma_op *ro);
+void rds_atomic_free_op(struct rm_atomic_op *ao);
+void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
+void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
+int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
+		    struct cmsghdr *cmsg);
+
+extern void __rds_put_mr_final(struct rds_mr *mr);
+static inline void rds_mr_put(struct rds_mr *mr)
+{
+	if (atomic_dec_and_test(&mr->r_refcount))
+		__rds_put_mr_final(mr);
+}
 
 /* stats.c */
 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
@@ -657,14 +774,14 @@ DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
 	put_cpu();					\
 } while (0)
 #define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
-int __init rds_stats_init(void);
+int rds_stats_init(void);
 void rds_stats_exit(void);
 void rds_stats_info_copy(struct rds_info_iterator *iter,
 			 uint64_t *values, const char *const *names,
 			 size_t nr);
 
 /* sysctl.c */
-int __init rds_sysctl_init(void);
+int rds_sysctl_init(void);
 void rds_sysctl_exit(void);
 extern unsigned long rds_sysctl_sndbuf_min;
 extern unsigned long rds_sysctl_sndbuf_default;
@@ -678,9 +795,10 @@ extern unsigned long rds_sysctl_trace_flags;
 extern unsigned int  rds_sysctl_trace_level;
 
 /* threads.c */
-int __init rds_threads_init(void);
+int rds_threads_init(void);
 void rds_threads_exit(void);
 extern struct workqueue_struct *rds_wq;
+void rds_queue_reconnect(struct rds_connection *conn);
 void rds_connect_worker(struct work_struct *);
 void rds_shutdown_worker(struct work_struct *);
 void rds_send_worker(struct work_struct *);
@@ -691,9 +809,10 @@ void rds_connect_complete(struct rds_connection *conn);
 int rds_trans_register(struct rds_transport *trans);
 void rds_trans_unregister(struct rds_transport *trans);
 struct rds_transport *rds_trans_get_preferred(__be32 addr);
+void rds_trans_put(struct rds_transport *trans);
 unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
 				       unsigned int avail);
-int __init rds_trans_init(void);
+int rds_trans_init(void);
 void rds_trans_exit(void);
 
 #endif

net/rds/recv.c

@@ -36,7 +36,6 @@
 #include <linux/in.h>
 
 #include "rds.h"
-#include "rdma.h"
 
 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
 		  __be32 saddr)
@@ -210,7 +209,7 @@ void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
 	}
 
 	rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
-	if (rs == NULL) {
+	if (!rs) {
 		rds_stats_inc(s_recv_drop_no_sock);
 		goto out;
 	}
@@ -251,7 +250,7 @@ static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
 {
 	unsigned long flags;
 
-	if (*inc == NULL) {
+	if (!*inc) {
 		read_lock_irqsave(&rs->rs_recv_lock, flags);
 		if (!list_empty(&rs->rs_recv_queue)) {
 			*inc = list_entry(rs->rs_recv_queue.next,
@@ -334,10 +333,10 @@ int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
 
 		if (msghdr) {
 			cmsg.user_token = notifier->n_user_token;
-			cmsg.status  = notifier->n_status;
+			cmsg.status = notifier->n_status;
 
 			err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
-					sizeof(cmsg), &cmsg);
+				       sizeof(cmsg), &cmsg);
 			if (err)
 				break;
 		}

net/rds/send.c

@@ -37,7 +37,6 @@
 #include <linux/list.h>
 
 #include "rds.h"
-#include "rdma.h"
 
 /* When transmitting messages in rds_send_xmit, we need to emerge from
  * time to time and briefly release the CPU. Otherwise the softlock watchdog
@@ -54,7 +53,8 @@ module_param(send_batch_count, int, 0444);
 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
 
 /*
- * Reset the send state. Caller must hold c_send_lock when calling here.
+ * Reset the send state.  Callers must ensure that this doesn't race with
+ * rds_send_xmit().
  */
 void rds_send_reset(struct rds_connection *conn)
 {
@@ -62,18 +62,22 @@ void rds_send_reset(struct rds_connection *conn)
 	unsigned long flags;
 
 	if (conn->c_xmit_rm) {
+		rm = conn->c_xmit_rm;
+		conn->c_xmit_rm = NULL;
 		/* Tell the user the RDMA op is no longer mapped by the
 		 * transport. This isn't entirely true (it's flushed out
 		 * independently) but as the connection is down, there's
 		 * no ongoing RDMA to/from that memory */
-		rds_message_unmapped(conn->c_xmit_rm);
-		rds_message_put(conn->c_xmit_rm);
-		conn->c_xmit_rm = NULL;
+		rds_message_unmapped(rm);
+		rds_message_put(rm);
 	}
+
 	conn->c_xmit_sg = 0;
 	conn->c_xmit_hdr_off = 0;
 	conn->c_xmit_data_off = 0;
+	conn->c_xmit_atomic_sent = 0;
 	conn->c_xmit_rdma_sent = 0;
+	conn->c_xmit_data_sent = 0;
 
 	conn->c_map_queued = 0;
 
@@ -90,6 +94,25 @@ void rds_send_reset(struct rds_connection *conn)
 	spin_unlock_irqrestore(&conn->c_lock, flags);
 }
 
+static int acquire_in_xmit(struct rds_connection *conn)
+{
+	return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
+}
+
+static void release_in_xmit(struct rds_connection *conn)
+{
+	clear_bit(RDS_IN_XMIT, &conn->c_flags);
+	smp_mb__after_clear_bit();
+	/*
+	 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
+	 * hot path and finding waiters is very rare.  We don't want to walk
+	 * the system-wide hashed waitqueue buckets in the fast path only to
+	 * almost never find waiters.
+	 */
+	if (waitqueue_active(&conn->c_waitq))
+		wake_up_all(&conn->c_waitq);
+}
+
 /*
  * We're making the concious trade-off here to only send one message
  * down the connection at a time.
@@ -109,102 +132,69 @@ int rds_send_xmit(struct rds_connection *conn)
 	struct rds_message *rm;
 	unsigned long flags;
 	unsigned int tmp;
-	unsigned int send_quota = send_batch_count;
 	struct scatterlist *sg;
 	int ret = 0;
-	int was_empty = 0;
 	LIST_HEAD(to_be_dropped);
 
+restart:
+
 	/*
 	 * sendmsg calls here after having queued its message on the send
 	 * queue.  We only have one task feeding the connection at a time.  If
 	 * another thread is already feeding the queue then we back off.  This
 	 * avoids blocking the caller and trading per-connection data between
 	 * caches per message.
-	 *
-	 * The sem holder will issue a retry if they notice that someone queued
-	 * a message after they stopped walking the send queue but before they
-	 * dropped the sem.
 	 */
-	if (!mutex_trylock(&conn->c_send_lock)) {
-		rds_stats_inc(s_send_sem_contention);
+	if (!acquire_in_xmit(conn)) {
+		rds_stats_inc(s_send_lock_contention);
 		ret = -ENOMEM;
 		goto out;
 	}
 
+	/*
+	 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
+	 * we do the opposite to avoid races.
+	 */
+	if (!rds_conn_up(conn)) {
+		release_in_xmit(conn);
+		ret = 0;
+		goto out;
+	}
+
 	if (conn->c_trans->xmit_prepare)
 		conn->c_trans->xmit_prepare(conn);
 
 	/*
 	 * spin trying to push headers and data down the connection until
-	 * the connection doens't make forward progress.
+	 * the connection doesn't make forward progress.
 	 */
-	while (--send_quota) {
-		/*
-		 * See if need to send a congestion map update if we're
-		 * between sending messages.  The send_sem protects our sole
-		 * use of c_map_offset and _bytes.
-		 * Note this is used only by transports that define a special
-		 * xmit_cong_map function. For all others, we create allocate
-		 * a cong_map message and treat it just like any other send.
-		 */
-		if (conn->c_map_bytes) {
-			ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
-						conn->c_map_offset);
-			if (ret <= 0)
-				break;
+	while (1) {
 
-			conn->c_map_offset += ret;
-			conn->c_map_bytes -= ret;
-			if (conn->c_map_bytes)
-				continue;
-		}
-
-		/* If we're done sending the current message, clear the
-		 * offset and S/G temporaries.
-		 */
 		rm = conn->c_xmit_rm;
-		if (rm != NULL &&
-		    conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
-		    conn->c_xmit_sg == rm->m_nents) {
-			conn->c_xmit_rm = NULL;
-			conn->c_xmit_sg = 0;
-			conn->c_xmit_hdr_off = 0;
-			conn->c_xmit_data_off = 0;
-			conn->c_xmit_rdma_sent = 0;
 
-			/* Release the reference to the previous message. */
-			rds_message_put(rm);
-			rm = NULL;
-		}
-
-		/* If we're asked to send a cong map update, do so.
+		/*
+		 * If between sending messages, we can send a pending congestion
+		 * map update.
 		 */
-		if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) {
-			if (conn->c_trans->xmit_cong_map != NULL) {
-				conn->c_map_offset = 0;
-				conn->c_map_bytes = sizeof(struct rds_header) +
-					RDS_CONG_MAP_BYTES;
-				continue;
-			}
-
+		if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
 			rm = rds_cong_update_alloc(conn);
 			if (IS_ERR(rm)) {
 				ret = PTR_ERR(rm);
 				break;
 			}
+			rm->data.op_active = 1;
 
 			conn->c_xmit_rm = rm;
 		}
 
 		/*
-		 * Grab the next message from the send queue, if there is one.
+		 * If not already working on one, grab the next message.
 		 *
 		 * c_xmit_rm holds a ref while we're sending this message down
 		 * the connction.  We can use this ref while holding the
 		 * send_sem.. rds_send_reset() is serialized with it.
 		 */
-		if (rm == NULL) {
+		if (!rm) {
 			unsigned int len;
 
 			spin_lock_irqsave(&conn->c_lock, flags);
@@ -224,10 +214,8 @@ int rds_send_xmit(struct rds_connection *conn)
 
 			spin_unlock_irqrestore(&conn->c_lock, flags);
 
-			if (rm == NULL) {
-				was_empty = 1;
+			if (!rm)
 				break;
-			}
 
 			/* Unfortunately, the way Infiniband deals with
 			 * RDMA to a bad MR key is by moving the entire
@@ -236,13 +224,12 @@ int rds_send_xmit(struct rds_connection *conn)
 			 * connection.
 			 * Therefore, we never retransmit messages with RDMA ops.
 			 */
-			if (rm->m_rdma_op &&
+			if (rm->rdma.op_active &&
 			    test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
 				spin_lock_irqsave(&conn->c_lock, flags);
 				if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
 					list_move(&rm->m_conn_item, &to_be_dropped);
 				spin_unlock_irqrestore(&conn->c_lock, flags);
-				rds_message_put(rm);
 				continue;
 			}
 
@@ -263,23 +250,55 @@ int rds_send_xmit(struct rds_connection *conn)
 			conn->c_xmit_rm = rm;
 		}
 
-		/*
-		 * Try and send an rdma message.  Let's see if we can
-		 * keep this simple and require that the transport either
-		 * send the whole rdma or none of it.
-		 */
-		if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
-			ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
+		/* The transport either sends the whole rdma or none of it */
+		if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
+			rm->m_final_op = &rm->rdma;
+			ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
 			if (ret)
 				break;
 			conn->c_xmit_rdma_sent = 1;
+
 			/* The transport owns the mapped memory for now.
 			 * You can't unmap it while it's on the send queue */
 			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 		}
 
-		if (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
-		    conn->c_xmit_sg < rm->m_nents) {
+		if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
+			rm->m_final_op = &rm->atomic;
+			ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
+			if (ret)
+				break;
+			conn->c_xmit_atomic_sent = 1;
+
+			/* The transport owns the mapped memory for now.
+			 * You can't unmap it while it's on the send queue */
+			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
+		}
+
+		/*
+		 * A number of cases require an RDS header to be sent
+		 * even if there is no data.
+		 * We permit 0-byte sends; rds-ping depends on this.
+		 * However, if there are exclusively attached silent ops,
+		 * we skip the hdr/data send, to enable silent operation.
+		 */
+		if (rm->data.op_nents == 0) {
+			int ops_present;
+			int all_ops_are_silent = 1;
+
+			ops_present = (rm->atomic.op_active || rm->rdma.op_active);
+			if (rm->atomic.op_active && !rm->atomic.op_silent)
+				all_ops_are_silent = 0;
+			if (rm->rdma.op_active && !rm->rdma.op_silent)
+				all_ops_are_silent = 0;
+
+			if (ops_present && all_ops_are_silent
+			    && !rm->m_rdma_cookie)
+				rm->data.op_active = 0;
+		}
+
+		if (rm->data.op_active && !conn->c_xmit_data_sent) {
+			rm->m_final_op = &rm->data;
 			ret = conn->c_trans->xmit(conn, rm,
 						  conn->c_xmit_hdr_off,
 						  conn->c_xmit_sg,
@@ -295,7 +314,7 @@ int rds_send_xmit(struct rds_connection *conn)
 				ret -= tmp;
 			}
 
-			sg = &rm->m_sg[conn->c_xmit_sg];
+			sg = &rm->data.op_sg[conn->c_xmit_sg];
 			while (ret) {
 				tmp = min_t(int, ret, sg->length -
 						      conn->c_xmit_data_off);
@@ -306,49 +325,63 @@ int rds_send_xmit(struct rds_connection *conn)
 					sg++;
 					conn->c_xmit_sg++;
 					BUG_ON(ret != 0 &&
-					       conn->c_xmit_sg == rm->m_nents);
+					       conn->c_xmit_sg == rm->data.op_nents);
 				}
 			}
+
+			if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
+			    (conn->c_xmit_sg == rm->data.op_nents))
+				conn->c_xmit_data_sent = 1;
 		}
-	}
 
-	/* Nuke any messages we decided not to retransmit. */
-	if (!list_empty(&to_be_dropped))
-		rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
+		/*
+		 * A rm will only take multiple times through this loop
+		 * if there is a data op. Thus, if the data is sent (or there was
+		 * none), then we're done with the rm.
+		 */
+		if (!rm->data.op_active || conn->c_xmit_data_sent) {
+			conn->c_xmit_rm = NULL;
+			conn->c_xmit_sg = 0;
+			conn->c_xmit_hdr_off = 0;
+			conn->c_xmit_data_off = 0;
+			conn->c_xmit_rdma_sent = 0;
+			conn->c_xmit_atomic_sent = 0;
+			conn->c_xmit_data_sent = 0;
+
+			rds_message_put(rm);
+		}
+	}
 
 	if (conn->c_trans->xmit_complete)
 		conn->c_trans->xmit_complete(conn);
 
-	/*
-	 * We might be racing with another sender who queued a message but
-	 * backed off on noticing that we held the c_send_lock.  If we check
-	 * for queued messages after dropping the sem then either we'll
-	 * see the queued message or the queuer will get the sem.  If we
-	 * notice the queued message then we trigger an immediate retry.
-	 *
-	 * We need to be careful only to do this when we stopped processing
-	 * the send queue because it was empty.  It's the only way we
-	 * stop processing the loop when the transport hasn't taken
-	 * responsibility for forward progress.
-	 */
-	mutex_unlock(&conn->c_send_lock);
+	release_in_xmit(conn);
 
-	if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
-		/* We exhausted the send quota, but there's work left to
-		 * do. Return and (re-)schedule the send worker.
-		 */
-		ret = -EAGAIN;
+	/* Nuke any messages we decided not to retransmit. */
+	if (!list_empty(&to_be_dropped)) {
+		/* irqs on here, so we can put(), unlike above */
+		list_for_each_entry(rm, &to_be_dropped, m_conn_item)
+			rds_message_put(rm);
+		rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
 	}
 
-	if (ret == 0 && was_empty) {
-		/* A simple bit test would be way faster than taking the
-		 * spin lock */
-		spin_lock_irqsave(&conn->c_lock, flags);
+	/*
+	 * Other senders can queue a message after we last test the send queue
+	 * but before we clear RDS_IN_XMIT.  In that case they'd back off and
+	 * not try and send their newly queued message.  We need to check the
+	 * send queue after having cleared RDS_IN_XMIT so that their message
+	 * doesn't get stuck on the send queue.
+	 *
+	 * If the transport cannot continue (i.e ret != 0), then it must
+	 * call us when more room is available, such as from the tx
+	 * completion handler.
+	 */
+	if (ret == 0) {
+		smp_mb();
 		if (!list_empty(&conn->c_send_queue)) {
-			rds_stats_inc(s_send_sem_queue_raced);
-			ret = -EAGAIN;
+			rds_stats_inc(s_send_lock_queue_raced);
+			goto restart;
 		}
-		spin_unlock_irqrestore(&conn->c_lock, flags);
 	}
 out:
 	return ret;
@@ -376,52 +409,60 @@ static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
 }
 
 /*
- * Returns true if there are no messages on the send and retransmit queues
- * which have a sequence number greater than or equal to the given sequence
- * number.
+ * This is pretty similar to what happens below in the ACK
+ * handling code - except that we call here as soon as we get
+ * the IB send completion on the RDMA op and the accompanying
+ * message.
  */
-int rds_send_acked_before(struct rds_connection *conn, u64 seq)
+void rds_rdma_send_complete(struct rds_message *rm, int status)
 {
-	struct rds_message *rm, *tmp;
-	int ret = 1;
+	struct rds_sock *rs = NULL;
+	struct rm_rdma_op *ro;
+	struct rds_notifier *notifier;
+	unsigned long flags;
 
-	spin_lock(&conn->c_lock);
+	spin_lock_irqsave(&rm->m_rs_lock, flags);
 
-	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
-		if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
-			ret = 0;
-		break;
-	}
+	ro = &rm->rdma;
+	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
+	    ro->op_active && ro->op_notify && ro->op_notifier) {
+		notifier = ro->op_notifier;
+		rs = rm->m_rs;
+		sock_hold(rds_rs_to_sk(rs));
 
-	list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
-		if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
-			ret = 0;
-		break;
+		notifier->n_status = status;
+		spin_lock(&rs->rs_lock);
+		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
+		spin_unlock(&rs->rs_lock);
+
+		ro->op_notifier = NULL;
 	}
 
-	spin_unlock(&conn->c_lock);
+	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 
-	return ret;
+	if (rs) {
+		rds_wake_sk_sleep(rs);
+		sock_put(rds_rs_to_sk(rs));
+	}
 }
+EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
 
 /*
- * This is pretty similar to what happens below in the ACK
- * handling code - except that we call here as soon as we get
- * the IB send completion on the RDMA op and the accompanying
- * message.
+ * Just like above, except looks at atomic op
  */
-void rds_rdma_send_complete(struct rds_message *rm, int status)
+void rds_atomic_send_complete(struct rds_message *rm, int status)
 {
 	struct rds_sock *rs = NULL;
-	struct rds_rdma_op *ro;
+	struct rm_atomic_op *ao;
 	struct rds_notifier *notifier;
+	unsigned long flags;
 
-	spin_lock(&rm->m_rs_lock);
+	spin_lock_irqsave(&rm->m_rs_lock, flags);
 
-	ro = rm->m_rdma_op;
-	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
-	    ro && ro->r_notify && ro->r_notifier) {
-		notifier = ro->r_notifier;
+	ao = &rm->atomic;
+	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
+	    && ao->op_active && ao->op_notify && ao->op_notifier) {
+		notifier = ao->op_notifier;
 		rs = rm->m_rs;
 		sock_hold(rds_rs_to_sk(rs));
 
@@ -430,17 +471,17 @@ void rds_rdma_send_complete(struct rds_message *rm, int status)
 		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
 		spin_unlock(&rs->rs_lock);
 
-		ro->r_notifier = NULL;
+		ao->op_notifier = NULL;
 	}
 
-	spin_unlock(&rm->m_rs_lock);
+	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
 
 	if (rs) {
 		rds_wake_sk_sleep(rs);
 		sock_put(rds_rs_to_sk(rs));
 	}
 }
-EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
+EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
 
 /*
  * This is the same as rds_rdma_send_complete except we
@@ -448,15 +489,23 @@ EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
  * socket, socket lock) and can just move the notifier.
  */
 static inline void
-__rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
+__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
 {
-	struct rds_rdma_op *ro;
+	struct rm_rdma_op *ro;
+	struct rm_atomic_op *ao;
+
+	ro = &rm->rdma;
+	if (ro->op_active && ro->op_notify && ro->op_notifier) {
+		ro->op_notifier->n_status = status;
+		list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
+		ro->op_notifier = NULL;
+	}
 
-	ro = rm->m_rdma_op;
-	if (ro && ro->r_notify && ro->r_notifier) {
-		ro->r_notifier->n_status = status;
-		list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue);
-		ro->r_notifier = NULL;
+	ao = &rm->atomic;
+	if (ao->op_active && ao->op_notify && ao->op_notifier) {
+		ao->op_notifier->n_status = status;
+		list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
+		ao->op_notifier = NULL;
 	}
 
 	/* No need to wake the app - caller does this */
@@ -468,7 +517,7 @@ __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status
  * So speed is not an issue here.
  */
 struct rds_message *rds_send_get_message(struct rds_connection *conn,
-					 struct rds_rdma_op *op)
+					 struct rm_rdma_op *op)
 {
 	struct rds_message *rm, *tmp, *found = NULL;
 	unsigned long flags;
@@ -476,7 +525,7 @@ struct rds_message *rds_send_get_message(struct rds_connection *conn,
 	spin_lock_irqsave(&conn->c_lock, flags);
 
 	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
-		if (rm->m_rdma_op == op) {
+		if (&rm->rdma == op) {
 			atomic_inc(&rm->m_refcount);
 			found = rm;
 			goto out;
@@ -484,7 +533,7 @@ struct rds_message *rds_send_get_message(struct rds_connection *conn,
 	}
 
 	list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
-		if (rm->m_rdma_op == op) {
+		if (&rm->rdma == op) {
 			atomic_inc(&rm->m_refcount);
 			found = rm;
 			break;
@@ -544,19 +593,20 @@ void rds_send_remove_from_sock(struct list_head *messages, int status)
 		spin_lock(&rs->rs_lock);
 
 		if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
-			struct rds_rdma_op *ro = rm->m_rdma_op;
+			struct rm_rdma_op *ro = &rm->rdma;
 			struct rds_notifier *notifier;
 
 			list_del_init(&rm->m_sock_item);
 			rds_send_sndbuf_remove(rs, rm);
 
-			if (ro && ro->r_notifier && (status || ro->r_notify)) {
-				notifier = ro->r_notifier;
+			if (ro->op_active && ro->op_notifier &&
+			       (ro->op_notify || (ro->op_recverr && status))) {
+				notifier = ro->op_notifier;
 				list_add_tail(&notifier->n_list,
 						&rs->rs_notify_queue);
 				if (!notifier->n_status)
 					notifier->n_status = status;
-				rm->m_rdma_op->r_notifier = NULL;
+				rm->rdma.op_notifier = NULL;
 			}
 			was_on_sock = 1;
 			rm->m_rs = NULL;
@@ -619,9 +669,8 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 {
 	struct rds_message *rm, *tmp;
 	struct rds_connection *conn;
-	unsigned long flags, flags2;
+	unsigned long flags;
 	LIST_HEAD(list);
-	int wake = 0;
 
 	/* get all the messages we're dropping under the rs lock */
 	spin_lock_irqsave(&rs->rs_lock, flags);
@@ -631,59 +680,54 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 			     dest->sin_port != rm->m_inc.i_hdr.h_dport))
 			continue;
 
-		wake = 1;
 		list_move(&rm->m_sock_item, &list);
 		rds_send_sndbuf_remove(rs, rm);
 		clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
 	}
 
 	/* order flag updates with the rs lock */
-	if (wake)
-		smp_mb__after_clear_bit();
+	smp_mb__after_clear_bit();
 
 	spin_unlock_irqrestore(&rs->rs_lock, flags);
 
-	conn = NULL;
+	if (list_empty(&list))
+		return;
 
-	/* now remove the messages from the conn list as needed */
+	/* Remove the messages from the conn */
 	list_for_each_entry(rm, &list, m_sock_item) {
-		/* We do this here rather than in the loop above, so that
-		 * we don't have to nest m_rs_lock under rs->rs_lock */
-		spin_lock_irqsave(&rm->m_rs_lock, flags2);
-		/* If this is a RDMA operation, notify the app. */
-		spin_lock(&rs->rs_lock);
-		__rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
-		spin_unlock(&rs->rs_lock);
-		rm->m_rs = NULL;
-		spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
 
+		conn = rm->m_inc.i_conn;
+
+		spin_lock_irqsave(&conn->c_lock, flags);
 		/*
-		 * If we see this flag cleared then we're *sure* that someone
-		 * else beat us to removing it from the conn.  If we race
-		 * with their flag update we'll get the lock and then really
-		 * see that the flag has been cleared.
+		 * Maybe someone else beat us to removing rm from the conn.
+		 * If we race with their flag update we'll get the lock and
+		 * then really see that the flag has been cleared.
 		 */
-		if (!test_bit(RDS_MSG_ON_CONN, &rm->m_flags))
+		if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
+			spin_unlock_irqrestore(&conn->c_lock, flags);
 			continue;
-
-		if (conn != rm->m_inc.i_conn) {
-			if (conn)
-				spin_unlock_irqrestore(&conn->c_lock, flags);
-			conn = rm->m_inc.i_conn;
-			spin_lock_irqsave(&conn->c_lock, flags);
 		}
+		list_del_init(&rm->m_conn_item);
+		spin_unlock_irqrestore(&conn->c_lock, flags);
 
-		if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
-			list_del_init(&rm->m_conn_item);
-			rds_message_put(rm);
-		}
-	}
+		/*
+		 * Couldn't grab m_rs_lock in top loop (lock ordering),
+		 * but we can now.
+		 */
+		spin_lock_irqsave(&rm->m_rs_lock, flags);
 
-	if (conn)
-		spin_unlock_irqrestore(&conn->c_lock, flags);
+		spin_lock(&rs->rs_lock);
+		__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
+		spin_unlock(&rs->rs_lock);
 
-	if (wake)
-		rds_wake_sk_sleep(rs);
+		rm->m_rs = NULL;
+		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
+
+		rds_message_put(rm);
+	}
+
+	rds_wake_sk_sleep(rs);
 
 	while (!list_empty(&list)) {
 		rm = list_entry(list.next, struct rds_message, m_sock_item);
@@ -763,6 +807,63 @@ out:
 	return *queued;
 }
 
+/*
+ * rds_message is getting to be quite complicated, and we'd like to allocate
+ * it all in one go. This figures out how big it needs to be up front.
+ */
+static int rds_rm_size(struct msghdr *msg, int data_len)
+{
+	struct cmsghdr *cmsg;
+	int size = 0;
+	int cmsg_groups = 0;
+	int retval;
+
+	for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
+		if (!CMSG_OK(msg, cmsg))
+			return -EINVAL;
+
+		if (cmsg->cmsg_level != SOL_RDS)
+			continue;
+
+		switch (cmsg->cmsg_type) {
+		case RDS_CMSG_RDMA_ARGS:
+			cmsg_groups |= 1;
+			retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
+			if (retval < 0)
+				return retval;
+			size += retval;
+
+			break;
+
+		case RDS_CMSG_RDMA_DEST:
+		case RDS_CMSG_RDMA_MAP:
+			cmsg_groups |= 2;
+			/* these are valid but do no add any size */
+			break;
+
+		case RDS_CMSG_ATOMIC_CSWP:
+		case RDS_CMSG_ATOMIC_FADD:
+		case RDS_CMSG_MASKED_ATOMIC_CSWP:
+		case RDS_CMSG_MASKED_ATOMIC_FADD:
+			cmsg_groups |= 1;
+			size += sizeof(struct scatterlist);
+			break;
+
+		default:
+			return -EINVAL;
+		}
+
+	}
+
+	size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
+
+	/* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
+	if (cmsg_groups == 3)
+		return -EINVAL;
+
+	return size;
+}
+
 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 			 struct msghdr *msg, int *allocated_mr)
 {
@@ -777,7 +878,7 @@ static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 			continue;
 
 		/* As a side effect, RDMA_DEST and RDMA_MAP will set
-		 * rm->m_rdma_cookie and rm->m_rdma_mr.
+		 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
 		 */
 		switch (cmsg->cmsg_type) {
 		case RDS_CMSG_RDMA_ARGS:
@@ -793,6 +894,12 @@ static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
 			if (!ret)
 				*allocated_mr = 1;
 			break;
+		case RDS_CMSG_ATOMIC_CSWP:
+		case RDS_CMSG_ATOMIC_FADD:
+		case RDS_CMSG_MASKED_ATOMIC_CSWP:
+		case RDS_CMSG_MASKED_ATOMIC_FADD:
+			ret = rds_cmsg_atomic(rs, rm, cmsg);
+			break;
 
 		default:
 			return -EINVAL;
@@ -850,13 +957,26 @@ int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 		goto out;
 	}
 
-	rm = rds_message_copy_from_user(msg->msg_iov, payload_len);
-	if (IS_ERR(rm)) {
-		ret = PTR_ERR(rm);
-		rm = NULL;
+	/* size of rm including all sgs */
+	ret = rds_rm_size(msg, payload_len);
+	if (ret < 0)
+		goto out;
+
+	rm = rds_message_alloc(ret, GFP_KERNEL);
+	if (!rm) {
+		ret = -ENOMEM;
 		goto out;
 	}
 
+	/* Attach data to the rm */
+	if (payload_len) {
+		rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
+		ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
+		if (ret)
+			goto out;
+	}
+	rm->data.op_active = 1;
+
 	rm->m_daddr = daddr;
 
 	/* rds_conn_create has a spinlock that runs with IRQ off.
@@ -879,22 +999,23 @@ int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 	if (ret)
 		goto out;
 
-	if ((rm->m_rdma_cookie || rm->m_rdma_op) &&
-	    conn->c_trans->xmit_rdma == NULL) {
+	if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
 		if (printk_ratelimit())
 			printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
-				rm->m_rdma_op, conn->c_trans->xmit_rdma);
+			       &rm->rdma, conn->c_trans->xmit_rdma);
 		ret = -EOPNOTSUPP;
 		goto out;
 	}
 
-	/* If the connection is down, trigger a connect. We may
-	 * have scheduled a delayed reconnect however - in this case
-	 * we should not interfere.
-	 */
-	if (rds_conn_state(conn) == RDS_CONN_DOWN &&
-	    !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
-		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
+	if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
+		if (printk_ratelimit())
+			printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
+			       &rm->atomic, conn->c_trans->xmit_atomic);
+		ret = -EOPNOTSUPP;
+		goto out;
+	}
+
+	rds_conn_connect_if_down(conn);
 
 	ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
 	if (ret) {
@@ -938,7 +1059,7 @@ int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
 	rds_stats_inc(s_send_queued);
 
 	if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
-		rds_send_worker(&conn->c_send_w.work);
+		rds_send_xmit(conn);
 
 	rds_message_put(rm);
 	return payload_len;
@@ -966,20 +1087,15 @@ rds_send_pong(struct rds_connection *conn, __be16 dport)
 	int ret = 0;
 
 	rm = rds_message_alloc(0, GFP_ATOMIC);
-	if (rm == NULL) {
+	if (!rm) {
 		ret = -ENOMEM;
 		goto out;
 	}
 
 	rm->m_daddr = conn->c_faddr;
+	rm->data.op_active = 1;
 
-	/* If the connection is down, trigger a connect. We may
-	 * have scheduled a delayed reconnect however - in this case
-	 * we should not interfere.
-	 */
-	if (rds_conn_state(conn) == RDS_CONN_DOWN &&
-	    !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
-		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
+	rds_conn_connect_if_down(conn);
 
 	ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
 	if (ret)
@@ -999,7 +1115,9 @@ rds_send_pong(struct rds_connection *conn, __be16 dport)
 	rds_stats_inc(s_send_queued);
 	rds_stats_inc(s_send_pong);
 
-	queue_delayed_work(rds_wq, &conn->c_send_w, 0);
+	if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
+		rds_send_xmit(conn);
+
 	rds_message_put(rm);
 	return 0;
 

net/rds/stats.c

@@ -57,8 +57,8 @@ static const char *const rds_stat_names[] = {
 	"recv_ping",
 	"send_queue_empty",
 	"send_queue_full",
-	"send_sem_contention",
-	"send_sem_queue_raced",
+	"send_lock_contention",
+	"send_lock_queue_raced",
 	"send_immediate_retry",
 	"send_delayed_retry",
 	"send_drop_acked",
@@ -143,7 +143,7 @@ void rds_stats_exit(void)
 	rds_info_deregister_func(RDS_INFO_COUNTERS, rds_stats_info);
 }
 
-int __init rds_stats_init(void)
+int rds_stats_init(void)
 {
 	rds_info_register_func(RDS_INFO_COUNTERS, rds_stats_info);
 	return 0;

net/rds/sysctl.c

@@ -105,13 +105,13 @@ void rds_sysctl_exit(void)
 		unregister_sysctl_table(rds_sysctl_reg_table);
 }
 
-int __init rds_sysctl_init(void)
+int rds_sysctl_init(void)
 {
 	rds_sysctl_reconnect_min = msecs_to_jiffies(1);
 	rds_sysctl_reconnect_min_jiffies = rds_sysctl_reconnect_min;
 
 	rds_sysctl_reg_table = register_sysctl_paths(rds_sysctl_path, rds_sysctl_rds_table);
-	if (rds_sysctl_reg_table == NULL)
+	if (!rds_sysctl_reg_table)
 		return -ENOMEM;
 	return 0;
 }

net/rds/tcp.c

@@ -200,7 +200,7 @@ static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 	struct rds_tcp_connection *tc;
 
 	tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
-	if (tc == NULL)
+	if (!tc)
 		return -ENOMEM;
 
 	tc->t_sock = NULL;
@@ -258,7 +258,6 @@ struct rds_transport rds_tcp_transport = {
 	.laddr_check		= rds_tcp_laddr_check,
 	.xmit_prepare		= rds_tcp_xmit_prepare,
 	.xmit_complete		= rds_tcp_xmit_complete,
-	.xmit_cong_map		= rds_tcp_xmit_cong_map,
 	.xmit			= rds_tcp_xmit,
 	.recv			= rds_tcp_recv,
 	.conn_alloc		= rds_tcp_conn_alloc,
@@ -266,7 +265,6 @@ struct rds_transport rds_tcp_transport = {
 	.conn_connect		= rds_tcp_conn_connect,
 	.conn_shutdown		= rds_tcp_conn_shutdown,
 	.inc_copy_to_user	= rds_tcp_inc_copy_to_user,
-	.inc_purge		= rds_tcp_inc_purge,
 	.inc_free		= rds_tcp_inc_free,
 	.stats_info_copy	= rds_tcp_stats_info_copy,
 	.exit			= rds_tcp_exit,
@@ -276,14 +274,14 @@ struct rds_transport rds_tcp_transport = {
 	.t_prefer_loopback	= 1,
 };
 
-int __init rds_tcp_init(void)
+int rds_tcp_init(void)
 {
 	int ret;
 
 	rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 					      sizeof(struct rds_tcp_connection),
 					      0, 0, NULL);
-	if (rds_tcp_conn_slab == NULL) {
+	if (!rds_tcp_conn_slab) {
 		ret = -ENOMEM;
 		goto out;
 	}

net/rds/tcp.h

@@ -43,7 +43,7 @@ struct rds_tcp_statistics {
 };
 
 /* tcp.c */
-int __init rds_tcp_init(void);
+int rds_tcp_init(void);
 void rds_tcp_exit(void);
 void rds_tcp_tune(struct socket *sock);
 void rds_tcp_nonagle(struct socket *sock);
@@ -61,16 +61,15 @@ void rds_tcp_conn_shutdown(struct rds_connection *conn);
 void rds_tcp_state_change(struct sock *sk);
 
 /* tcp_listen.c */
-int __init rds_tcp_listen_init(void);
+int rds_tcp_listen_init(void);
 void rds_tcp_listen_stop(void);
 void rds_tcp_listen_data_ready(struct sock *sk, int bytes);
 
 /* tcp_recv.c */
-int __init rds_tcp_recv_init(void);
+int rds_tcp_recv_init(void);
 void rds_tcp_recv_exit(void);
 void rds_tcp_data_ready(struct sock *sk, int bytes);
 int rds_tcp_recv(struct rds_connection *conn);
-void rds_tcp_inc_purge(struct rds_incoming *inc);
 void rds_tcp_inc_free(struct rds_incoming *inc);
 int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
 			     size_t size);
@@ -81,8 +80,6 @@ void rds_tcp_xmit_complete(struct rds_connection *conn);
 int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
 	         unsigned int hdr_off, unsigned int sg, unsigned int off);
 void rds_tcp_write_space(struct sock *sk);
-int rds_tcp_xmit_cong_map(struct rds_connection *conn,
-			  struct rds_cong_map *map, unsigned long offset);
 
 /* tcp_stats.c */
 DECLARE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats);

net/rds/tcp_connect.c

@@ -45,7 +45,7 @@ void rds_tcp_state_change(struct sock *sk)
 
 	read_lock(&sk->sk_callback_lock);
 	conn = sk->sk_user_data;
-	if (conn == NULL) {
+	if (!conn) {
 		state_change = sk->sk_state_change;
 		goto out;
 	}

net/rds/tcp_listen.c

@@ -116,7 +116,7 @@ void rds_tcp_listen_data_ready(struct sock *sk, int bytes)
 
 	read_lock(&sk->sk_callback_lock);
 	ready = sk->sk_user_data;
-	if (ready == NULL) { /* check for teardown race */
+	if (!ready) { /* check for teardown race */
 		ready = sk->sk_data_ready;
 		goto out;
 	}
@@ -135,7 +135,7 @@ out:
 	ready(sk, bytes);
 }
 
-int __init rds_tcp_listen_init(void)
+int rds_tcp_listen_init(void)
 {
 	struct sockaddr_in sin;
 	struct socket *sock = NULL;
@@ -178,7 +178,7 @@ void rds_tcp_listen_stop(void)
 	struct socket *sock = rds_tcp_listen_sock;
 	struct sock *sk;
 
-	if (sock == NULL)
+	if (!sock)
 		return;
 
 	sk = sock->sk;

net/rds/tcp_recv.c

@@ -39,7 +39,7 @@
 
 static struct kmem_cache *rds_tcp_incoming_slab;
 
-void rds_tcp_inc_purge(struct rds_incoming *inc)
+static void rds_tcp_inc_purge(struct rds_incoming *inc)
 {
 	struct rds_tcp_incoming *tinc;
 	tinc = container_of(inc, struct rds_tcp_incoming, ti_inc);
@@ -190,10 +190,10 @@ static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb,
 	 * processing.
 	 */
 	while (left) {
-		if (tinc == NULL) {
+		if (!tinc) {
 			tinc = kmem_cache_alloc(rds_tcp_incoming_slab,
 					        arg->gfp);
-			if (tinc == NULL) {
+			if (!tinc) {
 				desc->error = -ENOMEM;
 				goto out;
 			}
@@ -229,7 +229,7 @@ static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb,
 
 		if (left && tc->t_tinc_data_rem) {
 			clone = skb_clone(skb, arg->gfp);
-			if (clone == NULL) {
+			if (!clone) {
 				desc->error = -ENOMEM;
 				goto out;
 			}
@@ -326,7 +326,7 @@ void rds_tcp_data_ready(struct sock *sk, int bytes)
 
 	read_lock(&sk->sk_callback_lock);
 	conn = sk->sk_user_data;
-	if (conn == NULL) { /* check for teardown race */
+	if (!conn) { /* check for teardown race */
 		ready = sk->sk_data_ready;
 		goto out;
 	}
@@ -342,12 +342,12 @@ out:
 	ready(sk, bytes);
 }
 
-int __init rds_tcp_recv_init(void)
+int rds_tcp_recv_init(void)
 {
 	rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming",
 					sizeof(struct rds_tcp_incoming),
 					0, 0, NULL);
-	if (rds_tcp_incoming_slab == NULL)
+	if (!rds_tcp_incoming_slab)
 		return -ENOMEM;
 	return 0;
 }

net/rds/tcp_send.c

@@ -76,56 +76,6 @@ int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
 	return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len);
 }
 
-/* the core send_sem serializes this with other xmit and shutdown */
-int rds_tcp_xmit_cong_map(struct rds_connection *conn,
-			  struct rds_cong_map *map, unsigned long offset)
-{
-	static struct rds_header rds_tcp_map_header = {
-		.h_flags = RDS_FLAG_CONG_BITMAP,
-	};
-	struct rds_tcp_connection *tc = conn->c_transport_data;
-	unsigned long i;
-	int ret;
-	int copied = 0;
-
-	/* Some problem claims cpu_to_be32(constant) isn't a constant. */
-	rds_tcp_map_header.h_len = cpu_to_be32(RDS_CONG_MAP_BYTES);
-
-	if (offset < sizeof(struct rds_header)) {
-		ret = rds_tcp_sendmsg(tc->t_sock,
-				      (void *)&rds_tcp_map_header + offset,
-				      sizeof(struct rds_header) - offset);
-		if (ret <= 0)
-			return ret;
-		offset += ret;
-		copied = ret;
-		if (offset < sizeof(struct rds_header))
-			return ret;
-	}
-
-	offset -= sizeof(struct rds_header);
-	i = offset / PAGE_SIZE;
-	offset = offset % PAGE_SIZE;
-	BUG_ON(i >= RDS_CONG_MAP_PAGES);
-
-	do {
-		ret = tc->t_sock->ops->sendpage(tc->t_sock,
-					virt_to_page(map->m_page_addrs[i]),
-					offset, PAGE_SIZE - offset,
-					MSG_DONTWAIT);
-		if (ret <= 0)
-			break;
-		copied += ret;
-		offset += ret;
-		if (offset == PAGE_SIZE) {
-			offset = 0;
-			i++;
-		}
-	} while (i < RDS_CONG_MAP_PAGES);
-
-        return copied ? copied : ret;
-}
-
 /* the core send_sem serializes this with other xmit and shutdown */
 int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
 	         unsigned int hdr_off, unsigned int sg, unsigned int off)
@@ -166,21 +116,21 @@ int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
 			goto out;
 	}
 
-	while (sg < rm->m_nents) {
+	while (sg < rm->data.op_nents) {
 		ret = tc->t_sock->ops->sendpage(tc->t_sock,
-						sg_page(&rm->m_sg[sg]),
-						rm->m_sg[sg].offset + off,
-						rm->m_sg[sg].length - off,
+						sg_page(&rm->data.op_sg[sg]),
+						rm->data.op_sg[sg].offset + off,
+						rm->data.op_sg[sg].length - off,
 						MSG_DONTWAIT|MSG_NOSIGNAL);
-		rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->m_sg[sg]),
-			 rm->m_sg[sg].offset + off, rm->m_sg[sg].length - off,
+		rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
+			 rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
 			 ret);
 		if (ret <= 0)
 			break;
 
 		off += ret;
 		done += ret;
-		if (off == rm->m_sg[sg].length) {
+		if (off == rm->data.op_sg[sg].length) {
 			off = 0;
 			sg++;
 		}
@@ -226,7 +176,7 @@ void rds_tcp_write_space(struct sock *sk)
 
 	read_lock(&sk->sk_callback_lock);
 	conn = sk->sk_user_data;
-	if (conn == NULL) {
+	if (!conn) {
 		write_space = sk->sk_write_space;
 		goto out;
 	}

net/rds/threads.c

@@ -61,7 +61,7 @@
  *
  * Transition to state DISCONNECTING/DOWN:
  *  -	Inside the shutdown worker; synchronizes with xmit path
- *	through c_send_lock, and with connection management callbacks
+ *	through RDS_IN_XMIT, and with connection management callbacks
  *	via c_cm_lock.
  *
  *	For receive callbacks, we rely on the underlying transport
@@ -110,7 +110,7 @@ EXPORT_SYMBOL_GPL(rds_connect_complete);
  * We should *always* start with a random backoff; otherwise a broken connection
  * will always take several iterations to be re-established.
  */
-static void rds_queue_reconnect(struct rds_connection *conn)
+void rds_queue_reconnect(struct rds_connection *conn)
 {
 	unsigned long rand;
 
@@ -156,58 +156,6 @@ void rds_connect_worker(struct work_struct *work)
 	}
 }
 
-void rds_shutdown_worker(struct work_struct *work)
-{
-	struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
-
-	/* shut it down unless it's down already */
-	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
-		/*
-		 * Quiesce the connection mgmt handlers before we start tearing
-		 * things down. We don't hold the mutex for the entire
-		 * duration of the shutdown operation, else we may be
-		 * deadlocking with the CM handler. Instead, the CM event
-		 * handler is supposed to check for state DISCONNECTING
-		 */
-		mutex_lock(&conn->c_cm_lock);
-		if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING) &&
-		    !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
-			rds_conn_error(conn, "shutdown called in state %d\n",
-					atomic_read(&conn->c_state));
-			mutex_unlock(&conn->c_cm_lock);
-			return;
-		}
-		mutex_unlock(&conn->c_cm_lock);
-
-		mutex_lock(&conn->c_send_lock);
-		conn->c_trans->conn_shutdown(conn);
-		rds_conn_reset(conn);
-		mutex_unlock(&conn->c_send_lock);
-
-		if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
-			/* This can happen - eg when we're in the middle of tearing
-			 * down the connection, and someone unloads the rds module.
-			 * Quite reproduceable with loopback connections.
-			 * Mostly harmless.
-			 */
-			rds_conn_error(conn,
-				"%s: failed to transition to state DOWN, "
-				"current state is %d\n",
-				__func__,
-				atomic_read(&conn->c_state));
-			return;
-		}
-	}
-
-	/* Then reconnect if it's still live.
-	 * The passive side of an IB loopback connection is never added
-	 * to the conn hash, so we never trigger a reconnect on this
-	 * conn - the reconnect is always triggered by the active peer. */
-	cancel_delayed_work(&conn->c_conn_w);
-	if (!hlist_unhashed(&conn->c_hash_node))
-		rds_queue_reconnect(conn);
-}
-
 void rds_send_worker(struct work_struct *work)
 {
 	struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
@@ -252,15 +200,22 @@ void rds_recv_worker(struct work_struct *work)
 	}
 }
 
+void rds_shutdown_worker(struct work_struct *work)
+{
+	struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
+
+	rds_conn_shutdown(conn);
+}
+
 void rds_threads_exit(void)
 {
 	destroy_workqueue(rds_wq);
 }
 
-int __init rds_threads_init(void)
+int rds_threads_init(void)
 {
-	rds_wq = create_workqueue("krdsd");
-	if (rds_wq == NULL)
+	rds_wq = create_singlethread_workqueue("krdsd");
+	if (!rds_wq)
 		return -ENOMEM;
 
 	return 0;

net/rds/transport.c

@@ -71,19 +71,28 @@ void rds_trans_unregister(struct rds_transport *trans)
 }
 EXPORT_SYMBOL_GPL(rds_trans_unregister);
 
+void rds_trans_put(struct rds_transport *trans)
+{
+	if (trans && trans->t_owner)
+		module_put(trans->t_owner);
+}
+
 struct rds_transport *rds_trans_get_preferred(__be32 addr)
 {
 	struct rds_transport *ret = NULL;
-	int i;
+	struct rds_transport *trans;
+	unsigned int i;
 
 	if (IN_LOOPBACK(ntohl(addr)))
 		return &rds_loop_transport;
 
 	down_read(&rds_trans_sem);
-	for (i = 0; i < RDS_TRANS_COUNT; i++)
-	{
-		if (transports[i] && (transports[i]->laddr_check(addr) == 0)) {
-			ret = transports[i];
+	for (i = 0; i < RDS_TRANS_COUNT; i++) {
+		trans = transports[i];
+
+		if (trans && (trans->laddr_check(addr) == 0) &&
+		    (!trans->t_owner || try_module_get(trans->t_owner))) {
+			ret = trans;
 			break;
 		}
 	}

net/rds/xlist.h

@@ -0,0 +1,80 @@
+#ifndef _LINUX_XLIST_H
+#define _LINUX_XLIST_H
+
+#include <linux/stddef.h>
+#include <linux/poison.h>
+#include <linux/prefetch.h>
+#include <asm/system.h>
+
+struct xlist_head {
+	struct xlist_head *next;
+};
+
+static inline void INIT_XLIST_HEAD(struct xlist_head *list)
+{
+	list->next = NULL;
+}
+
+static inline int xlist_empty(struct xlist_head *head)
+{
+	return head->next == NULL;
+}
+
+static inline void xlist_add(struct xlist_head *new, struct xlist_head *tail,
+			     struct xlist_head *head)
+{
+	struct xlist_head *cur;
+	struct xlist_head *check;
+
+	while (1) {
+		cur = head->next;
+		tail->next = cur;
+		check = cmpxchg(&head->next, cur, new);
+		if (check == cur)
+			break;
+	}
+}
+
+static inline struct xlist_head *xlist_del_head(struct xlist_head *head)
+{
+	struct xlist_head *cur;
+	struct xlist_head *check;
+	struct xlist_head *next;
+
+	while (1) {
+		cur = head->next;
+		if (!cur)
+			goto out;
+
+		next = cur->next;
+		check = cmpxchg(&head->next, cur, next);
+		if (check == cur)
+			goto out;
+	}
+out:
+	return cur;
+}
+
+static inline struct xlist_head *xlist_del_head_fast(struct xlist_head *head)
+{
+	struct xlist_head *cur;
+
+	cur = head->next;
+	if (!cur)
+		return NULL;
+
+	head->next = cur->next;
+	return cur;
+}
+
+static inline void xlist_splice(struct xlist_head *list,
+				struct xlist_head *head)
+{
+	struct xlist_head *cur;
+
+	WARN_ON(head->next);
+	cur = xchg(&list->next, NULL);
+	head->next = cur;
+}
+
+#endif