aboutsummaryrefslogtreecommitdiffstats
path: root/net
diff options
context:
space:
mode:
authorDavid S. Miller <davem@davemloft.net>2010-09-09 17:58:11 -0400
committerDavid S. Miller <davem@davemloft.net>2010-09-09 17:58:11 -0400
commitcf0ac2b8a759fecbefd80f890c6dbe80ba65fd95 (patch)
treed5feaeef197dc681d1cdab2e8070ac31f0c43141 /net
parentf27e21a813e2c4ca74b30a5443602e75b146db9b (diff)
parent905d64c89e2a9d71d0606904b7c3908633db6072 (diff)
Merge branch 'for-davem' of git://oss.oracle.com/git/agrover/linux-2.6
Diffstat (limited to 'net')
-rw-r--r--net/rds/af_rds.c26
-rw-r--r--net/rds/bind.c82
-rw-r--r--net/rds/cong.c8
-rw-r--r--net/rds/connection.c157
-rw-r--r--net/rds/ib.c194
-rw-r--r--net/rds/ib.h100
-rw-r--r--net/rds/ib_cm.c183
-rw-r--r--net/rds/ib_rdma.c314
-rw-r--r--net/rds/ib_recv.c549
-rw-r--r--net/rds/ib_send.c682
-rw-r--r--net/rds/ib_stats.c2
-rw-r--r--net/rds/ib_sysctl.c17
-rw-r--r--net/rds/info.c12
-rw-r--r--net/rds/iw.c4
-rw-r--r--net/rds/iw.h11
-rw-r--r--net/rds/iw_cm.c14
-rw-r--r--net/rds/iw_rdma.c1
-rw-r--r--net/rds/iw_recv.c24
-rw-r--r--net/rds/iw_send.c93
-rw-r--r--net/rds/iw_sysctl.c4
-rw-r--r--net/rds/loop.c31
-rw-r--r--net/rds/message.c118
-rw-r--r--net/rds/page.c5
-rw-r--r--net/rds/rdma.c339
-rw-r--r--net/rds/rdma.h85
-rw-r--r--net/rds/rdma_transport.c42
-rw-r--r--net/rds/rds.h187
-rw-r--r--net/rds/recv.c9
-rw-r--r--net/rds/send.c544
-rw-r--r--net/rds/stats.c6
-rw-r--r--net/rds/sysctl.c4
-rw-r--r--net/rds/tcp.c8
-rw-r--r--net/rds/tcp.h9
-rw-r--r--net/rds/tcp_connect.c2
-rw-r--r--net/rds/tcp_listen.c6
-rw-r--r--net/rds/tcp_recv.c14
-rw-r--r--net/rds/tcp_send.c66
-rw-r--r--net/rds/threads.c69
-rw-r--r--net/rds/transport.c19
-rw-r--r--net/rds/xlist.h80
40 files changed, 2548 insertions, 1572 deletions
diff --git a/net/rds/af_rds.c b/net/rds/af_rds.c
index aebfecbdb841..bb6ad81b671d 100644
--- a/net/rds/af_rds.c
+++ b/net/rds/af_rds.c
@@ -39,7 +39,15 @@
39#include <net/sock.h> 39#include <net/sock.h>
40 40
41#include "rds.h" 41#include "rds.h"
42#include "rdma.h" 42
43char *rds_str_array(char **array, size_t elements, size_t index)
44{
45 if ((index < elements) && array[index])
46 return array[index];
47 else
48 return "unknown";
49}
50EXPORT_SYMBOL(rds_str_array);
43 51
44/* this is just used for stats gathering :/ */ 52/* this is just used for stats gathering :/ */
45static DEFINE_SPINLOCK(rds_sock_lock); 53static DEFINE_SPINLOCK(rds_sock_lock);
@@ -62,7 +70,7 @@ static int rds_release(struct socket *sock)
62 struct rds_sock *rs; 70 struct rds_sock *rs;
63 unsigned long flags; 71 unsigned long flags;
64 72
65 if (sk == NULL) 73 if (!sk)
66 goto out; 74 goto out;
67 75
68 rs = rds_sk_to_rs(sk); 76 rs = rds_sk_to_rs(sk);
@@ -73,7 +81,15 @@ static int rds_release(struct socket *sock)
73 * with the socket. */ 81 * with the socket. */
74 rds_clear_recv_queue(rs); 82 rds_clear_recv_queue(rs);
75 rds_cong_remove_socket(rs); 83 rds_cong_remove_socket(rs);
84
85 /*
86 * the binding lookup hash uses rcu, we need to
87 * make sure we sychronize_rcu before we free our
88 * entry
89 */
76 rds_remove_bound(rs); 90 rds_remove_bound(rs);
91 synchronize_rcu();
92
77 rds_send_drop_to(rs, NULL); 93 rds_send_drop_to(rs, NULL);
78 rds_rdma_drop_keys(rs); 94 rds_rdma_drop_keys(rs);
79 rds_notify_queue_get(rs, NULL); 95 rds_notify_queue_get(rs, NULL);
@@ -83,6 +99,8 @@ static int rds_release(struct socket *sock)
83 rds_sock_count--; 99 rds_sock_count--;
84 spin_unlock_irqrestore(&rds_sock_lock, flags); 100 spin_unlock_irqrestore(&rds_sock_lock, flags);
85 101
102 rds_trans_put(rs->rs_transport);
103
86 sock->sk = NULL; 104 sock->sk = NULL;
87 sock_put(sk); 105 sock_put(sk);
88out: 106out:
@@ -514,7 +532,7 @@ out:
514 spin_unlock_irqrestore(&rds_sock_lock, flags); 532 spin_unlock_irqrestore(&rds_sock_lock, flags);
515} 533}
516 534
517static void __exit rds_exit(void) 535static void rds_exit(void)
518{ 536{
519 sock_unregister(rds_family_ops.family); 537 sock_unregister(rds_family_ops.family);
520 proto_unregister(&rds_proto); 538 proto_unregister(&rds_proto);
@@ -529,7 +547,7 @@ static void __exit rds_exit(void)
529} 547}
530module_exit(rds_exit); 548module_exit(rds_exit);
531 549
532static int __init rds_init(void) 550static int rds_init(void)
533{ 551{
534 int ret; 552 int ret;
535 553
diff --git a/net/rds/bind.c b/net/rds/bind.c
index 5d95fc007f1a..2f6b3fcc79f8 100644
--- a/net/rds/bind.c
+++ b/net/rds/bind.c
@@ -34,45 +34,52 @@
34#include <net/sock.h> 34#include <net/sock.h>
35#include <linux/in.h> 35#include <linux/in.h>
36#include <linux/if_arp.h> 36#include <linux/if_arp.h>
37#include <linux/jhash.h>
37#include "rds.h" 38#include "rds.h"
38 39
39/* 40#define BIND_HASH_SIZE 1024
40 * XXX this probably still needs more work.. no INADDR_ANY, and rbtrees aren't 41static struct hlist_head bind_hash_table[BIND_HASH_SIZE];
41 * particularly zippy.
42 *
43 * This is now called for every incoming frame so we arguably care much more
44 * about it than we used to.
45 */
46static DEFINE_SPINLOCK(rds_bind_lock); 42static DEFINE_SPINLOCK(rds_bind_lock);
47static struct rb_root rds_bind_tree = RB_ROOT;
48 43
49static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port, 44static struct hlist_head *hash_to_bucket(__be32 addr, __be16 port)
50 struct rds_sock *insert) 45{
46 return bind_hash_table + (jhash_2words((u32)addr, (u32)port, 0) &
47 (BIND_HASH_SIZE - 1));
48}
49
50static struct rds_sock *rds_bind_lookup(__be32 addr, __be16 port,
51 struct rds_sock *insert)
51{ 52{
52 struct rb_node **p = &rds_bind_tree.rb_node;
53 struct rb_node *parent = NULL;
54 struct rds_sock *rs; 53 struct rds_sock *rs;
54 struct hlist_node *node;
55 struct hlist_head *head = hash_to_bucket(addr, port);
55 u64 cmp; 56 u64 cmp;
56 u64 needle = ((u64)be32_to_cpu(addr) << 32) | be16_to_cpu(port); 57 u64 needle = ((u64)be32_to_cpu(addr) << 32) | be16_to_cpu(port);
57 58
58 while (*p) { 59 rcu_read_lock();
59 parent = *p; 60 hlist_for_each_entry_rcu(rs, node, head, rs_bound_node) {
60 rs = rb_entry(parent, struct rds_sock, rs_bound_node);
61
62 cmp = ((u64)be32_to_cpu(rs->rs_bound_addr) << 32) | 61 cmp = ((u64)be32_to_cpu(rs->rs_bound_addr) << 32) |
63 be16_to_cpu(rs->rs_bound_port); 62 be16_to_cpu(rs->rs_bound_port);
64 63
65 if (needle < cmp) 64 if (cmp == needle) {
66 p = &(*p)->rb_left; 65 rcu_read_unlock();
67 else if (needle > cmp)
68 p = &(*p)->rb_right;
69 else
70 return rs; 66 return rs;
67 }
71 } 68 }
69 rcu_read_unlock();
72 70
73 if (insert) { 71 if (insert) {
74 rb_link_node(&insert->rs_bound_node, parent, p); 72 /*
75 rb_insert_color(&insert->rs_bound_node, &rds_bind_tree); 73 * make sure our addr and port are set before
74 * we are added to the list, other people
75 * in rcu will find us as soon as the
76 * hlist_add_head_rcu is done
77 */
78 insert->rs_bound_addr = addr;
79 insert->rs_bound_port = port;
80 rds_sock_addref(insert);
81
82 hlist_add_head_rcu(&insert->rs_bound_node, head);
76 } 83 }
77 return NULL; 84 return NULL;
78} 85}
@@ -86,15 +93,13 @@ static struct rds_sock *rds_bind_tree_walk(__be32 addr, __be16 port,
86struct rds_sock *rds_find_bound(__be32 addr, __be16 port) 93struct rds_sock *rds_find_bound(__be32 addr, __be16 port)
87{ 94{
88 struct rds_sock *rs; 95 struct rds_sock *rs;
89 unsigned long flags;
90 96
91 spin_lock_irqsave(&rds_bind_lock, flags); 97 rs = rds_bind_lookup(addr, port, NULL);
92 rs = rds_bind_tree_walk(addr, port, NULL); 98
93 if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD)) 99 if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD))
94 rds_sock_addref(rs); 100 rds_sock_addref(rs);
95 else 101 else
96 rs = NULL; 102 rs = NULL;
97 spin_unlock_irqrestore(&rds_bind_lock, flags);
98 103
99 rdsdebug("returning rs %p for %pI4:%u\n", rs, &addr, 104 rdsdebug("returning rs %p for %pI4:%u\n", rs, &addr,
100 ntohs(port)); 105 ntohs(port));
@@ -121,22 +126,15 @@ static int rds_add_bound(struct rds_sock *rs, __be32 addr, __be16 *port)
121 do { 126 do {
122 if (rover == 0) 127 if (rover == 0)
123 rover++; 128 rover++;
124 if (rds_bind_tree_walk(addr, cpu_to_be16(rover), rs) == NULL) { 129 if (!rds_bind_lookup(addr, cpu_to_be16(rover), rs)) {
125 *port = cpu_to_be16(rover); 130 *port = rs->rs_bound_port;
126 ret = 0; 131 ret = 0;
132 rdsdebug("rs %p binding to %pI4:%d\n",
133 rs, &addr, (int)ntohs(*port));
127 break; 134 break;
128 } 135 }
129 } while (rover++ != last); 136 } while (rover++ != last);
130 137
131 if (ret == 0) {
132 rs->rs_bound_addr = addr;
133 rs->rs_bound_port = *port;
134 rds_sock_addref(rs);
135
136 rdsdebug("rs %p binding to %pI4:%d\n",
137 rs, &addr, (int)ntohs(*port));
138 }
139
140 spin_unlock_irqrestore(&rds_bind_lock, flags); 138 spin_unlock_irqrestore(&rds_bind_lock, flags);
141 139
142 return ret; 140 return ret;
@@ -153,7 +151,7 @@ void rds_remove_bound(struct rds_sock *rs)
153 rs, &rs->rs_bound_addr, 151 rs, &rs->rs_bound_addr,
154 ntohs(rs->rs_bound_port)); 152 ntohs(rs->rs_bound_port));
155 153
156 rb_erase(&rs->rs_bound_node, &rds_bind_tree); 154 hlist_del_init_rcu(&rs->rs_bound_node);
157 rds_sock_put(rs); 155 rds_sock_put(rs);
158 rs->rs_bound_addr = 0; 156 rs->rs_bound_addr = 0;
159 } 157 }
@@ -184,7 +182,7 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
184 goto out; 182 goto out;
185 183
186 trans = rds_trans_get_preferred(sin->sin_addr.s_addr); 184 trans = rds_trans_get_preferred(sin->sin_addr.s_addr);
187 if (trans == NULL) { 185 if (!trans) {
188 ret = -EADDRNOTAVAIL; 186 ret = -EADDRNOTAVAIL;
189 rds_remove_bound(rs); 187 rds_remove_bound(rs);
190 if (printk_ratelimit()) 188 if (printk_ratelimit())
@@ -198,5 +196,9 @@ int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
198 196
199out: 197out:
200 release_sock(sk); 198 release_sock(sk);
199
200 /* we might have called rds_remove_bound on error */
201 if (ret)
202 synchronize_rcu();
201 return ret; 203 return ret;
202} 204}
diff --git a/net/rds/cong.c b/net/rds/cong.c
index 0871a29f0780..75ea686f27d5 100644
--- a/net/rds/cong.c
+++ b/net/rds/cong.c
@@ -141,7 +141,7 @@ static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
141 unsigned long flags; 141 unsigned long flags;
142 142
143 map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL); 143 map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
144 if (map == NULL) 144 if (!map)
145 return NULL; 145 return NULL;
146 146
147 map->m_addr = addr; 147 map->m_addr = addr;
@@ -159,7 +159,7 @@ static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
159 ret = rds_cong_tree_walk(addr, map); 159 ret = rds_cong_tree_walk(addr, map);
160 spin_unlock_irqrestore(&rds_cong_lock, flags); 160 spin_unlock_irqrestore(&rds_cong_lock, flags);
161 161
162 if (ret == NULL) { 162 if (!ret) {
163 ret = map; 163 ret = map;
164 map = NULL; 164 map = NULL;
165 } 165 }
@@ -205,7 +205,7 @@ int rds_cong_get_maps(struct rds_connection *conn)
205 conn->c_lcong = rds_cong_from_addr(conn->c_laddr); 205 conn->c_lcong = rds_cong_from_addr(conn->c_laddr);
206 conn->c_fcong = rds_cong_from_addr(conn->c_faddr); 206 conn->c_fcong = rds_cong_from_addr(conn->c_faddr);
207 207
208 if (conn->c_lcong == NULL || conn->c_fcong == NULL) 208 if (!(conn->c_lcong && conn->c_fcong))
209 return -ENOMEM; 209 return -ENOMEM;
210 210
211 return 0; 211 return 0;
@@ -221,7 +221,7 @@ void rds_cong_queue_updates(struct rds_cong_map *map)
221 list_for_each_entry(conn, &map->m_conn_list, c_map_item) { 221 list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
222 if (!test_and_set_bit(0, &conn->c_map_queued)) { 222 if (!test_and_set_bit(0, &conn->c_map_queued)) {
223 rds_stats_inc(s_cong_update_queued); 223 rds_stats_inc(s_cong_update_queued);
224 queue_delayed_work(rds_wq, &conn->c_send_w, 0); 224 rds_send_xmit(conn);
225 } 225 }
226 } 226 }
227 227
diff --git a/net/rds/connection.c b/net/rds/connection.c
index 7619b671ca28..870992e08cae 100644
--- a/net/rds/connection.c
+++ b/net/rds/connection.c
@@ -37,7 +37,6 @@
37 37
38#include "rds.h" 38#include "rds.h"
39#include "loop.h" 39#include "loop.h"
40#include "rdma.h"
41 40
42#define RDS_CONNECTION_HASH_BITS 12 41#define RDS_CONNECTION_HASH_BITS 12
43#define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS) 42#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)
63 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \ 62 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
64} while (0) 63} while (0)
65 64
66static inline int rds_conn_is_sending(struct rds_connection *conn) 65/* rcu read lock must be held or the connection spinlock */
67{
68 int ret = 0;
69
70 if (!mutex_trylock(&conn->c_send_lock))
71 ret = 1;
72 else
73 mutex_unlock(&conn->c_send_lock);
74
75 return ret;
76}
77
78static struct rds_connection *rds_conn_lookup(struct hlist_head *head, 66static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
79 __be32 laddr, __be32 faddr, 67 __be32 laddr, __be32 faddr,
80 struct rds_transport *trans) 68 struct rds_transport *trans)
@@ -82,7 +70,7 @@ static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
82 struct rds_connection *conn, *ret = NULL; 70 struct rds_connection *conn, *ret = NULL;
83 struct hlist_node *pos; 71 struct hlist_node *pos;
84 72
85 hlist_for_each_entry(conn, pos, head, c_hash_node) { 73 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
86 if (conn->c_faddr == faddr && conn->c_laddr == laddr && 74 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
87 conn->c_trans == trans) { 75 conn->c_trans == trans) {
88 ret = conn; 76 ret = conn;
@@ -129,10 +117,11 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
129{ 117{
130 struct rds_connection *conn, *parent = NULL; 118 struct rds_connection *conn, *parent = NULL;
131 struct hlist_head *head = rds_conn_bucket(laddr, faddr); 119 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
120 struct rds_transport *loop_trans;
132 unsigned long flags; 121 unsigned long flags;
133 int ret; 122 int ret;
134 123
135 spin_lock_irqsave(&rds_conn_lock, flags); 124 rcu_read_lock();
136 conn = rds_conn_lookup(head, laddr, faddr, trans); 125 conn = rds_conn_lookup(head, laddr, faddr, trans);
137 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport && 126 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
138 !is_outgoing) { 127 !is_outgoing) {
@@ -143,12 +132,12 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
143 parent = conn; 132 parent = conn;
144 conn = parent->c_passive; 133 conn = parent->c_passive;
145 } 134 }
146 spin_unlock_irqrestore(&rds_conn_lock, flags); 135 rcu_read_unlock();
147 if (conn) 136 if (conn)
148 goto out; 137 goto out;
149 138
150 conn = kmem_cache_zalloc(rds_conn_slab, gfp); 139 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
151 if (conn == NULL) { 140 if (!conn) {
152 conn = ERR_PTR(-ENOMEM); 141 conn = ERR_PTR(-ENOMEM);
153 goto out; 142 goto out;
154 } 143 }
@@ -159,7 +148,7 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
159 spin_lock_init(&conn->c_lock); 148 spin_lock_init(&conn->c_lock);
160 conn->c_next_tx_seq = 1; 149 conn->c_next_tx_seq = 1;
161 150
162 mutex_init(&conn->c_send_lock); 151 init_waitqueue_head(&conn->c_waitq);
163 INIT_LIST_HEAD(&conn->c_send_queue); 152 INIT_LIST_HEAD(&conn->c_send_queue);
164 INIT_LIST_HEAD(&conn->c_retrans); 153 INIT_LIST_HEAD(&conn->c_retrans);
165 154
@@ -175,7 +164,9 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
175 * can bind to the destination address then we'd rather the messages 164 * can bind to the destination address then we'd rather the messages
176 * flow through loopback rather than either transport. 165 * flow through loopback rather than either transport.
177 */ 166 */
178 if (rds_trans_get_preferred(faddr)) { 167 loop_trans = rds_trans_get_preferred(faddr);
168 if (loop_trans) {
169 rds_trans_put(loop_trans);
179 conn->c_loopback = 1; 170 conn->c_loopback = 1;
180 if (is_outgoing && trans->t_prefer_loopback) { 171 if (is_outgoing && trans->t_prefer_loopback) {
181 /* "outgoing" connection - and the transport 172 /* "outgoing" connection - and the transport
@@ -238,7 +229,7 @@ static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
238 kmem_cache_free(rds_conn_slab, conn); 229 kmem_cache_free(rds_conn_slab, conn);
239 conn = found; 230 conn = found;
240 } else { 231 } else {
241 hlist_add_head(&conn->c_hash_node, head); 232 hlist_add_head_rcu(&conn->c_hash_node, head);
242 rds_cong_add_conn(conn); 233 rds_cong_add_conn(conn);
243 rds_conn_count++; 234 rds_conn_count++;
244 } 235 }
@@ -263,21 +254,91 @@ struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
263} 254}
264EXPORT_SYMBOL_GPL(rds_conn_create_outgoing); 255EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
265 256
257void rds_conn_shutdown(struct rds_connection *conn)
258{
259 /* shut it down unless it's down already */
260 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
261 /*
262 * Quiesce the connection mgmt handlers before we start tearing
263 * things down. We don't hold the mutex for the entire
264 * duration of the shutdown operation, else we may be
265 * deadlocking with the CM handler. Instead, the CM event
266 * handler is supposed to check for state DISCONNECTING
267 */
268 mutex_lock(&conn->c_cm_lock);
269 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
270 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
271 rds_conn_error(conn, "shutdown called in state %d\n",
272 atomic_read(&conn->c_state));
273 mutex_unlock(&conn->c_cm_lock);
274 return;
275 }
276 mutex_unlock(&conn->c_cm_lock);
277
278 wait_event(conn->c_waitq,
279 !test_bit(RDS_IN_XMIT, &conn->c_flags));
280
281 conn->c_trans->conn_shutdown(conn);
282 rds_conn_reset(conn);
283
284 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
285 /* This can happen - eg when we're in the middle of tearing
286 * down the connection, and someone unloads the rds module.
287 * Quite reproduceable with loopback connections.
288 * Mostly harmless.
289 */
290 rds_conn_error(conn,
291 "%s: failed to transition to state DOWN, "
292 "current state is %d\n",
293 __func__,
294 atomic_read(&conn->c_state));
295 return;
296 }
297 }
298
299 /* Then reconnect if it's still live.
300 * The passive side of an IB loopback connection is never added
301 * to the conn hash, so we never trigger a reconnect on this
302 * conn - the reconnect is always triggered by the active peer. */
303 cancel_delayed_work_sync(&conn->c_conn_w);
304 rcu_read_lock();
305 if (!hlist_unhashed(&conn->c_hash_node)) {
306 rcu_read_unlock();
307 rds_queue_reconnect(conn);
308 } else {
309 rcu_read_unlock();
310 }
311}
312
313/*
314 * Stop and free a connection.
315 *
316 * This can only be used in very limited circumstances. It assumes that once
317 * the conn has been shutdown that no one else is referencing the connection.
318 * We can only ensure this in the rmmod path in the current code.
319 */
266void rds_conn_destroy(struct rds_connection *conn) 320void rds_conn_destroy(struct rds_connection *conn)
267{ 321{
268 struct rds_message *rm, *rtmp; 322 struct rds_message *rm, *rtmp;
323 unsigned long flags;
269 324
270 rdsdebug("freeing conn %p for %pI4 -> " 325 rdsdebug("freeing conn %p for %pI4 -> "
271 "%pI4\n", conn, &conn->c_laddr, 326 "%pI4\n", conn, &conn->c_laddr,
272 &conn->c_faddr); 327 &conn->c_faddr);
273 328
274 hlist_del_init(&conn->c_hash_node); 329 /* Ensure conn will not be scheduled for reconnect */
330 spin_lock_irq(&rds_conn_lock);
331 hlist_del_init_rcu(&conn->c_hash_node);
332 spin_unlock_irq(&rds_conn_lock);
333 synchronize_rcu();
275 334
276 /* wait for the rds thread to shut it down */ 335 /* shut the connection down */
277 atomic_set(&conn->c_state, RDS_CONN_ERROR); 336 rds_conn_drop(conn);
278 cancel_delayed_work(&conn->c_conn_w); 337 flush_work(&conn->c_down_w);
279 queue_work(rds_wq, &conn->c_down_w); 338
280 flush_workqueue(rds_wq); 339 /* make sure lingering queued work won't try to ref the conn */
340 cancel_delayed_work_sync(&conn->c_send_w);
341 cancel_delayed_work_sync(&conn->c_recv_w);
281 342
282 /* tear down queued messages */ 343 /* tear down queued messages */
283 list_for_each_entry_safe(rm, rtmp, 344 list_for_each_entry_safe(rm, rtmp,
@@ -302,7 +363,9 @@ void rds_conn_destroy(struct rds_connection *conn)
302 BUG_ON(!list_empty(&conn->c_retrans)); 363 BUG_ON(!list_empty(&conn->c_retrans));
303 kmem_cache_free(rds_conn_slab, conn); 364 kmem_cache_free(rds_conn_slab, conn);
304 365
366 spin_lock_irqsave(&rds_conn_lock, flags);
305 rds_conn_count--; 367 rds_conn_count--;
368 spin_unlock_irqrestore(&rds_conn_lock, flags);
306} 369}
307EXPORT_SYMBOL_GPL(rds_conn_destroy); 370EXPORT_SYMBOL_GPL(rds_conn_destroy);
308 371
@@ -316,23 +379,23 @@ static void rds_conn_message_info(struct socket *sock, unsigned int len,
316 struct list_head *list; 379 struct list_head *list;
317 struct rds_connection *conn; 380 struct rds_connection *conn;
318 struct rds_message *rm; 381 struct rds_message *rm;
319 unsigned long flags;
320 unsigned int total = 0; 382 unsigned int total = 0;
383 unsigned long flags;
321 size_t i; 384 size_t i;
322 385
323 len /= sizeof(struct rds_info_message); 386 len /= sizeof(struct rds_info_message);
324 387
325 spin_lock_irqsave(&rds_conn_lock, flags); 388 rcu_read_lock();
326 389
327 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); 390 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
328 i++, head++) { 391 i++, head++) {
329 hlist_for_each_entry(conn, pos, head, c_hash_node) { 392 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
330 if (want_send) 393 if (want_send)
331 list = &conn->c_send_queue; 394 list = &conn->c_send_queue;
332 else 395 else
333 list = &conn->c_retrans; 396 list = &conn->c_retrans;
334 397
335 spin_lock(&conn->c_lock); 398 spin_lock_irqsave(&conn->c_lock, flags);
336 399
337 /* XXX too lazy to maintain counts.. */ 400 /* XXX too lazy to maintain counts.. */
338 list_for_each_entry(rm, list, m_conn_item) { 401 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,
343 conn->c_faddr, 0); 406 conn->c_faddr, 0);
344 } 407 }
345 408
346 spin_unlock(&conn->c_lock); 409 spin_unlock_irqrestore(&conn->c_lock, flags);
347 } 410 }
348 } 411 }
349 412 rcu_read_unlock();
350 spin_unlock_irqrestore(&rds_conn_lock, flags);
351 413
352 lens->nr = total; 414 lens->nr = total;
353 lens->each = sizeof(struct rds_info_message); 415 lens->each = sizeof(struct rds_info_message);
@@ -377,19 +439,17 @@ void rds_for_each_conn_info(struct socket *sock, unsigned int len,
377 uint64_t buffer[(item_len + 7) / 8]; 439 uint64_t buffer[(item_len + 7) / 8];
378 struct hlist_head *head; 440 struct hlist_head *head;
379 struct hlist_node *pos; 441 struct hlist_node *pos;
380 struct hlist_node *tmp;
381 struct rds_connection *conn; 442 struct rds_connection *conn;
382 unsigned long flags;
383 size_t i; 443 size_t i;
384 444
385 spin_lock_irqsave(&rds_conn_lock, flags); 445 rcu_read_lock();
386 446
387 lens->nr = 0; 447 lens->nr = 0;
388 lens->each = item_len; 448 lens->each = item_len;
389 449
390 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); 450 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
391 i++, head++) { 451 i++, head++) {
392 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) { 452 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
393 453
394 /* XXX no c_lock usage.. */ 454 /* XXX no c_lock usage.. */
395 if (!visitor(conn, buffer)) 455 if (!visitor(conn, buffer))
@@ -405,8 +465,7 @@ void rds_for_each_conn_info(struct socket *sock, unsigned int len,
405 lens->nr++; 465 lens->nr++;
406 } 466 }
407 } 467 }
408 468 rcu_read_unlock();
409 spin_unlock_irqrestore(&rds_conn_lock, flags);
410} 469}
411EXPORT_SYMBOL_GPL(rds_for_each_conn_info); 470EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
412 471
@@ -423,8 +482,8 @@ static int rds_conn_info_visitor(struct rds_connection *conn,
423 sizeof(cinfo->transport)); 482 sizeof(cinfo->transport));
424 cinfo->flags = 0; 483 cinfo->flags = 0;
425 484
426 rds_conn_info_set(cinfo->flags, 485 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
427 rds_conn_is_sending(conn), SENDING); 486 SENDING);
428 /* XXX Future: return the state rather than these funky bits */ 487 /* XXX Future: return the state rather than these funky bits */
429 rds_conn_info_set(cinfo->flags, 488 rds_conn_info_set(cinfo->flags,
430 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING, 489 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
@@ -444,12 +503,12 @@ static void rds_conn_info(struct socket *sock, unsigned int len,
444 sizeof(struct rds_info_connection)); 503 sizeof(struct rds_info_connection));
445} 504}
446 505
447int __init rds_conn_init(void) 506int rds_conn_init(void)
448{ 507{
449 rds_conn_slab = kmem_cache_create("rds_connection", 508 rds_conn_slab = kmem_cache_create("rds_connection",
450 sizeof(struct rds_connection), 509 sizeof(struct rds_connection),
451 0, 0, NULL); 510 0, 0, NULL);
452 if (rds_conn_slab == NULL) 511 if (!rds_conn_slab)
453 return -ENOMEM; 512 return -ENOMEM;
454 513
455 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info); 514 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
@@ -487,6 +546,18 @@ void rds_conn_drop(struct rds_connection *conn)
487EXPORT_SYMBOL_GPL(rds_conn_drop); 546EXPORT_SYMBOL_GPL(rds_conn_drop);
488 547
489/* 548/*
549 * If the connection is down, trigger a connect. We may have scheduled a
550 * delayed reconnect however - in this case we should not interfere.
551 */
552void rds_conn_connect_if_down(struct rds_connection *conn)
553{
554 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
555 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
556 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
557}
558EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
559
560/*
490 * An error occurred on the connection 561 * An error occurred on the connection
491 */ 562 */
492void 563void
diff --git a/net/rds/ib.c b/net/rds/ib.c
index 8f2d6dd7700a..b12a3951167d 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -53,12 +53,71 @@ MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
53module_param(rds_ib_retry_count, int, 0444); 53module_param(rds_ib_retry_count, int, 0444);
54MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error"); 54MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
55 55
56/*
57 * we have a clumsy combination of RCU and a rwsem protecting this list
58 * because it is used both in the get_mr fast path and while blocking in
59 * the FMR flushing path.
60 */
61DECLARE_RWSEM(rds_ib_devices_lock);
56struct list_head rds_ib_devices; 62struct list_head rds_ib_devices;
57 63
58/* NOTE: if also grabbing ibdev lock, grab this first */ 64/* NOTE: if also grabbing ibdev lock, grab this first */
59DEFINE_SPINLOCK(ib_nodev_conns_lock); 65DEFINE_SPINLOCK(ib_nodev_conns_lock);
60LIST_HEAD(ib_nodev_conns); 66LIST_HEAD(ib_nodev_conns);
61 67
68void rds_ib_nodev_connect(void)
69{
70 struct rds_ib_connection *ic;
71
72 spin_lock(&ib_nodev_conns_lock);
73 list_for_each_entry(ic, &ib_nodev_conns, ib_node)
74 rds_conn_connect_if_down(ic->conn);
75 spin_unlock(&ib_nodev_conns_lock);
76}
77
78void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
79{
80 struct rds_ib_connection *ic;
81 unsigned long flags;
82
83 spin_lock_irqsave(&rds_ibdev->spinlock, flags);
84 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
85 rds_conn_drop(ic->conn);
86 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
87}
88
89/*
90 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
91 * from interrupt context so we push freing off into a work struct in krdsd.
92 */
93static void rds_ib_dev_free(struct work_struct *work)
94{
95 struct rds_ib_ipaddr *i_ipaddr, *i_next;
96 struct rds_ib_device *rds_ibdev = container_of(work,
97 struct rds_ib_device, free_work);
98
99 if (rds_ibdev->mr_pool)
100 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
101 if (rds_ibdev->mr)
102 ib_dereg_mr(rds_ibdev->mr);
103 if (rds_ibdev->pd)
104 ib_dealloc_pd(rds_ibdev->pd);
105
106 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
107 list_del(&i_ipaddr->list);
108 kfree(i_ipaddr);
109 }
110
111 kfree(rds_ibdev);
112}
113
114void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
115{
116 BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
117 if (atomic_dec_and_test(&rds_ibdev->refcount))
118 queue_work(rds_wq, &rds_ibdev->free_work);
119}
120
62void rds_ib_add_one(struct ib_device *device) 121void rds_ib_add_one(struct ib_device *device)
63{ 122{
64 struct rds_ib_device *rds_ibdev; 123 struct rds_ib_device *rds_ibdev;
@@ -77,11 +136,14 @@ void rds_ib_add_one(struct ib_device *device)
77 goto free_attr; 136 goto free_attr;
78 } 137 }
79 138
80 rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL); 139 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
140 ibdev_to_node(device));
81 if (!rds_ibdev) 141 if (!rds_ibdev)
82 goto free_attr; 142 goto free_attr;
83 143
84 spin_lock_init(&rds_ibdev->spinlock); 144 spin_lock_init(&rds_ibdev->spinlock);
145 atomic_set(&rds_ibdev->refcount, 1);
146 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
85 147
86 rds_ibdev->max_wrs = dev_attr->max_qp_wr; 148 rds_ibdev->max_wrs = dev_attr->max_qp_wr;
87 rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE); 149 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)
91 min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) : 153 min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
92 fmr_pool_size; 154 fmr_pool_size;
93 155
156 rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
157 rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
158
94 rds_ibdev->dev = device; 159 rds_ibdev->dev = device;
95 rds_ibdev->pd = ib_alloc_pd(device); 160 rds_ibdev->pd = ib_alloc_pd(device);
96 if (IS_ERR(rds_ibdev->pd)) 161 if (IS_ERR(rds_ibdev->pd)) {
97 goto free_dev; 162 rds_ibdev->pd = NULL;
163 goto put_dev;
164 }
98 165
99 rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, 166 rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
100 IB_ACCESS_LOCAL_WRITE); 167 if (IS_ERR(rds_ibdev->mr)) {
101 if (IS_ERR(rds_ibdev->mr)) 168 rds_ibdev->mr = NULL;
102 goto err_pd; 169 goto put_dev;
170 }
103 171
104 rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev); 172 rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
105 if (IS_ERR(rds_ibdev->mr_pool)) { 173 if (IS_ERR(rds_ibdev->mr_pool)) {
106 rds_ibdev->mr_pool = NULL; 174 rds_ibdev->mr_pool = NULL;
107 goto err_mr; 175 goto put_dev;
108 } 176 }
109 177
110 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list); 178 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
111 INIT_LIST_HEAD(&rds_ibdev->conn_list); 179 INIT_LIST_HEAD(&rds_ibdev->conn_list);
112 list_add_tail(&rds_ibdev->list, &rds_ib_devices); 180
181 down_write(&rds_ib_devices_lock);
182 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
183 up_write(&rds_ib_devices_lock);
184 atomic_inc(&rds_ibdev->refcount);
113 185
114 ib_set_client_data(device, &rds_ib_client, rds_ibdev); 186 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
187 atomic_inc(&rds_ibdev->refcount);
115 188
116 goto free_attr; 189 rds_ib_nodev_connect();
117 190
118err_mr: 191put_dev:
119 ib_dereg_mr(rds_ibdev->mr); 192 rds_ib_dev_put(rds_ibdev);
120err_pd:
121 ib_dealloc_pd(rds_ibdev->pd);
122free_dev:
123 kfree(rds_ibdev);
124free_attr: 193free_attr:
125 kfree(dev_attr); 194 kfree(dev_attr);
126} 195}
127 196
197/*
198 * New connections use this to find the device to associate with the
199 * connection. It's not in the fast path so we're not concerned about the
200 * performance of the IB call. (As of this writing, it uses an interrupt
201 * blocking spinlock to serialize walking a per-device list of all registered
202 * clients.)
203 *
204 * RCU is used to handle incoming connections racing with device teardown.
205 * Rather than use a lock to serialize removal from the client_data and
206 * getting a new reference, we use an RCU grace period. The destruction
207 * path removes the device from client_data and then waits for all RCU
208 * readers to finish.
209 *
210 * A new connection can get NULL from this if its arriving on a
211 * device that is in the process of being removed.
212 */
213struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
214{
215 struct rds_ib_device *rds_ibdev;
216
217 rcu_read_lock();
218 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
219 if (rds_ibdev)
220 atomic_inc(&rds_ibdev->refcount);
221 rcu_read_unlock();
222 return rds_ibdev;
223}
224
225/*
226 * The IB stack is letting us know that a device is going away. This can
227 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
228 * the pci function, for example.
229 *
230 * This can be called at any time and can be racing with any other RDS path.
231 */
128void rds_ib_remove_one(struct ib_device *device) 232void rds_ib_remove_one(struct ib_device *device)
129{ 233{
130 struct rds_ib_device *rds_ibdev; 234 struct rds_ib_device *rds_ibdev;
131 struct rds_ib_ipaddr *i_ipaddr, *i_next;
132 235
133 rds_ibdev = ib_get_client_data(device, &rds_ib_client); 236 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
134 if (!rds_ibdev) 237 if (!rds_ibdev)
135 return; 238 return;
136 239
137 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) { 240 rds_ib_dev_shutdown(rds_ibdev);
138 list_del(&i_ipaddr->list);
139 kfree(i_ipaddr);
140 }
141 241
142 rds_ib_destroy_conns(rds_ibdev); 242 /* stop connection attempts from getting a reference to this device. */
243 ib_set_client_data(device, &rds_ib_client, NULL);
143 244
144 if (rds_ibdev->mr_pool) 245 down_write(&rds_ib_devices_lock);
145 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool); 246 list_del_rcu(&rds_ibdev->list);
146 247 up_write(&rds_ib_devices_lock);
147 ib_dereg_mr(rds_ibdev->mr);
148
149 while (ib_dealloc_pd(rds_ibdev->pd)) {
150 rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
151 msleep(1);
152 }
153 248
154 list_del(&rds_ibdev->list); 249 /*
155 kfree(rds_ibdev); 250 * This synchronize rcu is waiting for readers of both the ib
251 * client data and the devices list to finish before we drop
252 * both of those references.
253 */
254 synchronize_rcu();
255 rds_ib_dev_put(rds_ibdev);
256 rds_ib_dev_put(rds_ibdev);
156} 257}
157 258
158struct ib_client rds_ib_client = { 259struct ib_client rds_ib_client = {
@@ -186,7 +287,7 @@ static int rds_ib_conn_info_visitor(struct rds_connection *conn,
186 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid); 287 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
187 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid); 288 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
188 289
189 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); 290 rds_ibdev = ic->rds_ibdev;
190 iinfo->max_send_wr = ic->i_send_ring.w_nr; 291 iinfo->max_send_wr = ic->i_send_ring.w_nr;
191 iinfo->max_recv_wr = ic->i_recv_ring.w_nr; 292 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
192 iinfo->max_send_sge = rds_ibdev->max_sge; 293 iinfo->max_send_sge = rds_ibdev->max_sge;
@@ -248,29 +349,36 @@ static int rds_ib_laddr_check(__be32 addr)
248 return ret; 349 return ret;
249} 350}
250 351
352static void rds_ib_unregister_client(void)
353{
354 ib_unregister_client(&rds_ib_client);
355 /* wait for rds_ib_dev_free() to complete */
356 flush_workqueue(rds_wq);
357}
358
251void rds_ib_exit(void) 359void rds_ib_exit(void)
252{ 360{
253 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); 361 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
362 rds_ib_unregister_client();
254 rds_ib_destroy_nodev_conns(); 363 rds_ib_destroy_nodev_conns();
255 ib_unregister_client(&rds_ib_client);
256 rds_ib_sysctl_exit(); 364 rds_ib_sysctl_exit();
257 rds_ib_recv_exit(); 365 rds_ib_recv_exit();
258 rds_trans_unregister(&rds_ib_transport); 366 rds_trans_unregister(&rds_ib_transport);
367 rds_ib_fmr_exit();
259} 368}
260 369
261struct rds_transport rds_ib_transport = { 370struct rds_transport rds_ib_transport = {
262 .laddr_check = rds_ib_laddr_check, 371 .laddr_check = rds_ib_laddr_check,
263 .xmit_complete = rds_ib_xmit_complete, 372 .xmit_complete = rds_ib_xmit_complete,
264 .xmit = rds_ib_xmit, 373 .xmit = rds_ib_xmit,
265 .xmit_cong_map = NULL,
266 .xmit_rdma = rds_ib_xmit_rdma, 374 .xmit_rdma = rds_ib_xmit_rdma,
375 .xmit_atomic = rds_ib_xmit_atomic,
267 .recv = rds_ib_recv, 376 .recv = rds_ib_recv,
268 .conn_alloc = rds_ib_conn_alloc, 377 .conn_alloc = rds_ib_conn_alloc,
269 .conn_free = rds_ib_conn_free, 378 .conn_free = rds_ib_conn_free,
270 .conn_connect = rds_ib_conn_connect, 379 .conn_connect = rds_ib_conn_connect,
271 .conn_shutdown = rds_ib_conn_shutdown, 380 .conn_shutdown = rds_ib_conn_shutdown,
272 .inc_copy_to_user = rds_ib_inc_copy_to_user, 381 .inc_copy_to_user = rds_ib_inc_copy_to_user,
273 .inc_purge = rds_ib_inc_purge,
274 .inc_free = rds_ib_inc_free, 382 .inc_free = rds_ib_inc_free,
275 .cm_initiate_connect = rds_ib_cm_initiate_connect, 383 .cm_initiate_connect = rds_ib_cm_initiate_connect,
276 .cm_handle_connect = rds_ib_cm_handle_connect, 384 .cm_handle_connect = rds_ib_cm_handle_connect,
@@ -286,16 +394,20 @@ struct rds_transport rds_ib_transport = {
286 .t_type = RDS_TRANS_IB 394 .t_type = RDS_TRANS_IB
287}; 395};
288 396
289int __init rds_ib_init(void) 397int rds_ib_init(void)
290{ 398{
291 int ret; 399 int ret;
292 400
293 INIT_LIST_HEAD(&rds_ib_devices); 401 INIT_LIST_HEAD(&rds_ib_devices);
294 402
295 ret = ib_register_client(&rds_ib_client); 403 ret = rds_ib_fmr_init();
296 if (ret) 404 if (ret)
297 goto out; 405 goto out;
298 406
407 ret = ib_register_client(&rds_ib_client);
408 if (ret)
409 goto out_fmr_exit;
410
299 ret = rds_ib_sysctl_init(); 411 ret = rds_ib_sysctl_init();
300 if (ret) 412 if (ret)
301 goto out_ibreg; 413 goto out_ibreg;
@@ -317,7 +429,9 @@ out_recv:
317out_sysctl: 429out_sysctl:
318 rds_ib_sysctl_exit(); 430 rds_ib_sysctl_exit();
319out_ibreg: 431out_ibreg:
320 ib_unregister_client(&rds_ib_client); 432 rds_ib_unregister_client();
433out_fmr_exit:
434 rds_ib_fmr_exit();
321out: 435out:
322 return ret; 436 return ret;
323} 437}
diff --git a/net/rds/ib.h b/net/rds/ib.h
index 64df4e79b29f..7ad3d57e06a5 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -3,11 +3,13 @@
3 3
4#include <rdma/ib_verbs.h> 4#include <rdma/ib_verbs.h>
5#include <rdma/rdma_cm.h> 5#include <rdma/rdma_cm.h>
6#include <linux/pci.h>
7#include <linux/slab.h>
6#include "rds.h" 8#include "rds.h"
7#include "rdma_transport.h" 9#include "rdma_transport.h"
8 10
9#define RDS_FMR_SIZE 256 11#define RDS_FMR_SIZE 256
10#define RDS_FMR_POOL_SIZE 4096 12#define RDS_FMR_POOL_SIZE 8192
11 13
12#define RDS_IB_MAX_SGE 8 14#define RDS_IB_MAX_SGE 8
13#define RDS_IB_RECV_SGE 2 15#define RDS_IB_RECV_SGE 2
@@ -19,6 +21,9 @@
19 21
20#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ 22#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
21 23
24#define RDS_IB_RECYCLE_BATCH_COUNT 32
25
26extern struct rw_semaphore rds_ib_devices_lock;
22extern struct list_head rds_ib_devices; 27extern struct list_head rds_ib_devices;
23 28
24/* 29/*
@@ -26,20 +31,29 @@ extern struct list_head rds_ib_devices;
26 * try and minimize the amount of memory tied up both the device and 31 * try and minimize the amount of memory tied up both the device and
27 * socket receive queues. 32 * socket receive queues.
28 */ 33 */
29/* page offset of the final full frag that fits in the page */
30#define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
31struct rds_page_frag { 34struct rds_page_frag {
32 struct list_head f_item; 35 struct list_head f_item;
33 struct page *f_page; 36 struct list_head f_cache_entry;
34 unsigned long f_offset; 37 struct scatterlist f_sg;
35 dma_addr_t f_mapped;
36}; 38};
37 39
38struct rds_ib_incoming { 40struct rds_ib_incoming {
39 struct list_head ii_frags; 41 struct list_head ii_frags;
42 struct list_head ii_cache_entry;
40 struct rds_incoming ii_inc; 43 struct rds_incoming ii_inc;
41}; 44};
42 45
46struct rds_ib_cache_head {
47 struct list_head *first;
48 unsigned long count;
49};
50
51struct rds_ib_refill_cache {
52 struct rds_ib_cache_head *percpu;
53 struct list_head *xfer;
54 struct list_head *ready;
55};
56
43struct rds_ib_connect_private { 57struct rds_ib_connect_private {
44 /* Add new fields at the end, and don't permute existing fields. */ 58 /* Add new fields at the end, and don't permute existing fields. */
45 __be32 dp_saddr; 59 __be32 dp_saddr;
@@ -53,8 +67,7 @@ struct rds_ib_connect_private {
53}; 67};
54 68
55struct rds_ib_send_work { 69struct rds_ib_send_work {
56 struct rds_message *s_rm; 70 void *s_op;
57 struct rds_rdma_op *s_op;
58 struct ib_send_wr s_wr; 71 struct ib_send_wr s_wr;
59 struct ib_sge s_sge[RDS_IB_MAX_SGE]; 72 struct ib_sge s_sge[RDS_IB_MAX_SGE];
60 unsigned long s_queued; 73 unsigned long s_queued;
@@ -92,10 +105,11 @@ struct rds_ib_connection {
92 105
93 /* tx */ 106 /* tx */
94 struct rds_ib_work_ring i_send_ring; 107 struct rds_ib_work_ring i_send_ring;
95 struct rds_message *i_rm; 108 struct rm_data_op *i_data_op;
96 struct rds_header *i_send_hdrs; 109 struct rds_header *i_send_hdrs;
97 u64 i_send_hdrs_dma; 110 u64 i_send_hdrs_dma;
98 struct rds_ib_send_work *i_sends; 111 struct rds_ib_send_work *i_sends;
112 atomic_t i_signaled_sends;
99 113
100 /* rx */ 114 /* rx */
101 struct tasklet_struct i_recv_tasklet; 115 struct tasklet_struct i_recv_tasklet;
@@ -106,8 +120,9 @@ struct rds_ib_connection {
106 struct rds_header *i_recv_hdrs; 120 struct rds_header *i_recv_hdrs;
107 u64 i_recv_hdrs_dma; 121 u64 i_recv_hdrs_dma;
108 struct rds_ib_recv_work *i_recvs; 122 struct rds_ib_recv_work *i_recvs;
109 struct rds_page_frag i_frag;
110 u64 i_ack_recv; /* last ACK received */ 123 u64 i_ack_recv; /* last ACK received */
124 struct rds_ib_refill_cache i_cache_incs;
125 struct rds_ib_refill_cache i_cache_frags;
111 126
112 /* sending acks */ 127 /* sending acks */
113 unsigned long i_ack_flags; 128 unsigned long i_ack_flags;
@@ -138,7 +153,6 @@ struct rds_ib_connection {
138 153
139 /* Batched completions */ 154 /* Batched completions */
140 unsigned int i_unsignaled_wrs; 155 unsigned int i_unsignaled_wrs;
141 long i_unsignaled_bytes;
142}; 156};
143 157
144/* This assumes that atomic_t is at least 32 bits */ 158/* This assumes that atomic_t is at least 32 bits */
@@ -164,9 +178,17 @@ struct rds_ib_device {
164 unsigned int max_fmrs; 178 unsigned int max_fmrs;
165 int max_sge; 179 int max_sge;
166 unsigned int max_wrs; 180 unsigned int max_wrs;
181 unsigned int max_initiator_depth;
182 unsigned int max_responder_resources;
167 spinlock_t spinlock; /* protect the above */ 183 spinlock_t spinlock; /* protect the above */
184 atomic_t refcount;
185 struct work_struct free_work;
168}; 186};
169 187
188#define pcidev_to_node(pcidev) pcibus_to_node(pcidev->bus)
189#define ibdev_to_node(ibdev) pcidev_to_node(to_pci_dev(ibdev->dma_device))
190#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
191
170/* bits for i_ack_flags */ 192/* bits for i_ack_flags */
171#define IB_ACK_IN_FLIGHT 0 193#define IB_ACK_IN_FLIGHT 0
172#define IB_ACK_REQUESTED 1 194#define IB_ACK_REQUESTED 1
@@ -202,6 +224,8 @@ struct rds_ib_statistics {
202 uint64_t s_ib_rdma_mr_pool_flush; 224 uint64_t s_ib_rdma_mr_pool_flush;
203 uint64_t s_ib_rdma_mr_pool_wait; 225 uint64_t s_ib_rdma_mr_pool_wait;
204 uint64_t s_ib_rdma_mr_pool_depleted; 226 uint64_t s_ib_rdma_mr_pool_depleted;
227 uint64_t s_ib_atomic_cswp;
228 uint64_t s_ib_atomic_fadd;
205}; 229};
206 230
207extern struct workqueue_struct *rds_ib_wq; 231extern struct workqueue_struct *rds_ib_wq;
@@ -243,6 +267,8 @@ static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
243extern struct rds_transport rds_ib_transport; 267extern struct rds_transport rds_ib_transport;
244extern void rds_ib_add_one(struct ib_device *device); 268extern void rds_ib_add_one(struct ib_device *device);
245extern void rds_ib_remove_one(struct ib_device *device); 269extern void rds_ib_remove_one(struct ib_device *device);
270struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
271void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
246extern struct ib_client rds_ib_client; 272extern struct ib_client rds_ib_client;
247 273
248extern unsigned int fmr_pool_size; 274extern unsigned int fmr_pool_size;
@@ -258,7 +284,7 @@ void rds_ib_conn_free(void *arg);
258int rds_ib_conn_connect(struct rds_connection *conn); 284int rds_ib_conn_connect(struct rds_connection *conn);
259void rds_ib_conn_shutdown(struct rds_connection *conn); 285void rds_ib_conn_shutdown(struct rds_connection *conn);
260void rds_ib_state_change(struct sock *sk); 286void rds_ib_state_change(struct sock *sk);
261int __init rds_ib_listen_init(void); 287int rds_ib_listen_init(void);
262void rds_ib_listen_stop(void); 288void rds_ib_listen_stop(void);
263void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); 289void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
264int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 290int 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,
275int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); 301int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
276void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 302void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
277void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 303void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
278void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock); 304void rds_ib_destroy_nodev_conns(void);
279static inline void rds_ib_destroy_nodev_conns(void)
280{
281 __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
282}
283static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev)
284{
285 __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
286}
287struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *); 305struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
288void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo); 306void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
289void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); 307void 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,
292void rds_ib_sync_mr(void *trans_private, int dir); 310void rds_ib_sync_mr(void *trans_private, int dir);
293void rds_ib_free_mr(void *trans_private, int invalidate); 311void rds_ib_free_mr(void *trans_private, int invalidate);
294void rds_ib_flush_mrs(void); 312void rds_ib_flush_mrs(void);
313int rds_ib_fmr_init(void);
314void rds_ib_fmr_exit(void);
295 315
296/* ib_recv.c */ 316/* ib_recv.c */
297int __init rds_ib_recv_init(void); 317int rds_ib_recv_init(void);
298void rds_ib_recv_exit(void); 318void rds_ib_recv_exit(void);
299int rds_ib_recv(struct rds_connection *conn); 319int rds_ib_recv(struct rds_connection *conn);
300int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, 320int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic);
301 gfp_t page_gfp, int prefill); 321void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
302void rds_ib_inc_purge(struct rds_incoming *inc); 322void rds_ib_recv_refill(struct rds_connection *conn, int prefill);
303void rds_ib_inc_free(struct rds_incoming *inc); 323void rds_ib_inc_free(struct rds_incoming *inc);
304int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov, 324int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
305 size_t size); 325 size_t size);
@@ -325,17 +345,19 @@ u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
325extern wait_queue_head_t rds_ib_ring_empty_wait; 345extern wait_queue_head_t rds_ib_ring_empty_wait;
326 346
327/* ib_send.c */ 347/* ib_send.c */
348char *rds_ib_wc_status_str(enum ib_wc_status status);
328void rds_ib_xmit_complete(struct rds_connection *conn); 349void rds_ib_xmit_complete(struct rds_connection *conn);
329int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, 350int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
330 unsigned int hdr_off, unsigned int sg, unsigned int off); 351 unsigned int hdr_off, unsigned int sg, unsigned int off);
331void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context); 352void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
332void rds_ib_send_init_ring(struct rds_ib_connection *ic); 353void rds_ib_send_init_ring(struct rds_ib_connection *ic);
333void rds_ib_send_clear_ring(struct rds_ib_connection *ic); 354void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
334int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op); 355int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
335void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); 356void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
336void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); 357void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
337int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, 358int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
338 u32 *adv_credits, int need_posted, int max_posted); 359 u32 *adv_credits, int need_posted, int max_posted);
360int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
339 361
340/* ib_stats.c */ 362/* ib_stats.c */
341DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); 363DECLARE_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,
344 unsigned int avail); 366 unsigned int avail);
345 367
346/* ib_sysctl.c */ 368/* ib_sysctl.c */
347int __init rds_ib_sysctl_init(void); 369int rds_ib_sysctl_init(void);
348void rds_ib_sysctl_exit(void); 370void rds_ib_sysctl_exit(void);
349extern unsigned long rds_ib_sysctl_max_send_wr; 371extern unsigned long rds_ib_sysctl_max_send_wr;
350extern unsigned long rds_ib_sysctl_max_recv_wr; 372extern unsigned long rds_ib_sysctl_max_recv_wr;
@@ -354,28 +376,4 @@ extern unsigned long rds_ib_sysctl_max_recv_allocation;
354extern unsigned int rds_ib_sysctl_flow_control; 376extern unsigned int rds_ib_sysctl_flow_control;
355extern ctl_table rds_ib_sysctl_table[]; 377extern ctl_table rds_ib_sysctl_table[];
356 378
357/*
358 * Helper functions for getting/setting the header and data SGEs in
359 * RDS packets (not RDMA)
360 *
361 * From version 3.1 onwards, header is in front of data in the sge.
362 */
363static inline struct ib_sge *
364rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
365{
366 if (ic->conn->c_version > RDS_PROTOCOL_3_0)
367 return &sge[0];
368 else
369 return &sge[1];
370}
371
372static inline struct ib_sge *
373rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
374{
375 if (ic->conn->c_version > RDS_PROTOCOL_3_0)
376 return &sge[1];
377 else
378 return &sge[0];
379}
380
381#endif 379#endif
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
index f68832798db2..bc3dbc1ba61f 100644
--- a/net/rds/ib_cm.c
+++ b/net/rds/ib_cm.c
@@ -38,6 +38,36 @@
38#include "rds.h" 38#include "rds.h"
39#include "ib.h" 39#include "ib.h"
40 40
41static char *rds_ib_event_type_strings[] = {
42#define RDS_IB_EVENT_STRING(foo) \
43 [IB_EVENT_##foo] = __stringify(IB_EVENT_##foo)
44 RDS_IB_EVENT_STRING(CQ_ERR),
45 RDS_IB_EVENT_STRING(QP_FATAL),
46 RDS_IB_EVENT_STRING(QP_REQ_ERR),
47 RDS_IB_EVENT_STRING(QP_ACCESS_ERR),
48 RDS_IB_EVENT_STRING(COMM_EST),
49 RDS_IB_EVENT_STRING(SQ_DRAINED),
50 RDS_IB_EVENT_STRING(PATH_MIG),
51 RDS_IB_EVENT_STRING(PATH_MIG_ERR),
52 RDS_IB_EVENT_STRING(DEVICE_FATAL),
53 RDS_IB_EVENT_STRING(PORT_ACTIVE),
54 RDS_IB_EVENT_STRING(PORT_ERR),
55 RDS_IB_EVENT_STRING(LID_CHANGE),
56 RDS_IB_EVENT_STRING(PKEY_CHANGE),
57 RDS_IB_EVENT_STRING(SM_CHANGE),
58 RDS_IB_EVENT_STRING(SRQ_ERR),
59 RDS_IB_EVENT_STRING(SRQ_LIMIT_REACHED),
60 RDS_IB_EVENT_STRING(QP_LAST_WQE_REACHED),
61 RDS_IB_EVENT_STRING(CLIENT_REREGISTER),
62#undef RDS_IB_EVENT_STRING
63};
64
65static char *rds_ib_event_str(enum ib_event_type type)
66{
67 return rds_str_array(rds_ib_event_type_strings,
68 ARRAY_SIZE(rds_ib_event_type_strings), type);
69};
70
41/* 71/*
42 * Set the selected protocol version 72 * Set the selected protocol version
43 */ 73 */
@@ -95,7 +125,6 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
95{ 125{
96 const struct rds_ib_connect_private *dp = NULL; 126 const struct rds_ib_connect_private *dp = NULL;
97 struct rds_ib_connection *ic = conn->c_transport_data; 127 struct rds_ib_connection *ic = conn->c_transport_data;
98 struct rds_ib_device *rds_ibdev;
99 struct ib_qp_attr qp_attr; 128 struct ib_qp_attr qp_attr;
100 int err; 129 int err;
101 130
@@ -111,11 +140,21 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
111 } 140 }
112 } 141 }
113 142
114 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", 143 if (conn->c_version < RDS_PROTOCOL(3,1)) {
115 &conn->c_faddr, 144 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
116 RDS_PROTOCOL_MAJOR(conn->c_version), 145 " no longer supported\n",
117 RDS_PROTOCOL_MINOR(conn->c_version), 146 &conn->c_faddr,
118 ic->i_flowctl ? ", flow control" : ""); 147 RDS_PROTOCOL_MAJOR(conn->c_version),
148 RDS_PROTOCOL_MINOR(conn->c_version));
149 rds_conn_destroy(conn);
150 return;
151 } else {
152 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
153 &conn->c_faddr,
154 RDS_PROTOCOL_MAJOR(conn->c_version),
155 RDS_PROTOCOL_MINOR(conn->c_version),
156 ic->i_flowctl ? ", flow control" : "");
157 }
119 158
120 /* 159 /*
121 * Init rings and fill recv. this needs to wait until protocol negotiation 160 * 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
125 rds_ib_recv_init_ring(ic); 164 rds_ib_recv_init_ring(ic);
126 /* Post receive buffers - as a side effect, this will update 165 /* Post receive buffers - as a side effect, this will update
127 * the posted credit count. */ 166 * the posted credit count. */
128 rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1); 167 rds_ib_recv_refill(conn, 1);
129 168
130 /* Tune RNR behavior */ 169 /* Tune RNR behavior */
131 rds_ib_tune_rnr(ic, &qp_attr); 170 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
135 if (err) 174 if (err)
136 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); 175 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
137 176
138 /* update ib_device with this local ipaddr & conn */ 177 /* update ib_device with this local ipaddr */
139 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); 178 err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
140 err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
141 if (err) 179 if (err)
142 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err); 180 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
143 rds_ib_add_conn(rds_ibdev, conn); 181 err);
144 182
145 /* If the peer gave us the last packet it saw, process this as if 183 /* If the peer gave us the last packet it saw, process this as if
146 * we had received a regular ACK. */ 184 * we had received a regular ACK. */
@@ -153,18 +191,23 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
153static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, 191static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
154 struct rdma_conn_param *conn_param, 192 struct rdma_conn_param *conn_param,
155 struct rds_ib_connect_private *dp, 193 struct rds_ib_connect_private *dp,
156 u32 protocol_version) 194 u32 protocol_version,
195 u32 max_responder_resources,
196 u32 max_initiator_depth)
157{ 197{
198 struct rds_ib_connection *ic = conn->c_transport_data;
199 struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
200
158 memset(conn_param, 0, sizeof(struct rdma_conn_param)); 201 memset(conn_param, 0, sizeof(struct rdma_conn_param));
159 /* XXX tune these? */ 202
160 conn_param->responder_resources = 1; 203 conn_param->responder_resources =
161 conn_param->initiator_depth = 1; 204 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
205 conn_param->initiator_depth =
206 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
162 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); 207 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
163 conn_param->rnr_retry_count = 7; 208 conn_param->rnr_retry_count = 7;
164 209
165 if (dp) { 210 if (dp) {
166 struct rds_ib_connection *ic = conn->c_transport_data;
167
168 memset(dp, 0, sizeof(*dp)); 211 memset(dp, 0, sizeof(*dp));
169 dp->dp_saddr = conn->c_laddr; 212 dp->dp_saddr = conn->c_laddr;
170 dp->dp_daddr = conn->c_faddr; 213 dp->dp_daddr = conn->c_faddr;
@@ -189,7 +232,8 @@ static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
189 232
190static void rds_ib_cq_event_handler(struct ib_event *event, void *data) 233static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
191{ 234{
192 rdsdebug("event %u data %p\n", event->event, data); 235 rdsdebug("event %u (%s) data %p\n",
236 event->event, rds_ib_event_str(event->event), data);
193} 237}
194 238
195static void rds_ib_qp_event_handler(struct ib_event *event, void *data) 239static 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)
197 struct rds_connection *conn = data; 241 struct rds_connection *conn = data;
198 struct rds_ib_connection *ic = conn->c_transport_data; 242 struct rds_ib_connection *ic = conn->c_transport_data;
199 243
200 rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event); 244 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
245 rds_ib_event_str(event->event));
201 246
202 switch (event->event) { 247 switch (event->event) {
203 case IB_EVENT_COMM_EST: 248 case IB_EVENT_COMM_EST:
204 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); 249 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
205 break; 250 break;
206 default: 251 default:
207 rdsdebug("Fatal QP Event %u " 252 rdsdebug("Fatal QP Event %u (%s) "
208 "- connection %pI4->%pI4, reconnecting\n", 253 "- connection %pI4->%pI4, reconnecting\n",
209 event->event, &conn->c_laddr, &conn->c_faddr); 254 event->event, rds_ib_event_str(event->event),
255 &conn->c_laddr, &conn->c_faddr);
210 rds_conn_drop(conn); 256 rds_conn_drop(conn);
211 break; 257 break;
212 } 258 }
@@ -224,18 +270,16 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
224 struct rds_ib_device *rds_ibdev; 270 struct rds_ib_device *rds_ibdev;
225 int ret; 271 int ret;
226 272
227 /* rds_ib_add_one creates a rds_ib_device object per IB device, 273 /*
228 * and allocates a protection domain, memory range and FMR pool 274 * It's normal to see a null device if an incoming connection races
229 * for each. If that fails for any reason, it will not register 275 * with device removal, so we don't print a warning.
230 * the rds_ibdev at all.
231 */ 276 */
232 rds_ibdev = ib_get_client_data(dev, &rds_ib_client); 277 rds_ibdev = rds_ib_get_client_data(dev);
233 if (rds_ibdev == NULL) { 278 if (!rds_ibdev)
234 if (printk_ratelimit())
235 printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
236 dev->name);
237 return -EOPNOTSUPP; 279 return -EOPNOTSUPP;
238 } 280
281 /* add the conn now so that connection establishment has the dev */
282 rds_ib_add_conn(rds_ibdev, conn);
239 283
240 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) 284 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
241 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); 285 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)
306 ic->i_send_ring.w_nr * 350 ic->i_send_ring.w_nr *
307 sizeof(struct rds_header), 351 sizeof(struct rds_header),
308 &ic->i_send_hdrs_dma, GFP_KERNEL); 352 &ic->i_send_hdrs_dma, GFP_KERNEL);
309 if (ic->i_send_hdrs == NULL) { 353 if (!ic->i_send_hdrs) {
310 ret = -ENOMEM; 354 ret = -ENOMEM;
311 rdsdebug("ib_dma_alloc_coherent send failed\n"); 355 rdsdebug("ib_dma_alloc_coherent send failed\n");
312 goto out; 356 goto out;
@@ -316,7 +360,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
316 ic->i_recv_ring.w_nr * 360 ic->i_recv_ring.w_nr *
317 sizeof(struct rds_header), 361 sizeof(struct rds_header),
318 &ic->i_recv_hdrs_dma, GFP_KERNEL); 362 &ic->i_recv_hdrs_dma, GFP_KERNEL);
319 if (ic->i_recv_hdrs == NULL) { 363 if (!ic->i_recv_hdrs) {
320 ret = -ENOMEM; 364 ret = -ENOMEM;
321 rdsdebug("ib_dma_alloc_coherent recv failed\n"); 365 rdsdebug("ib_dma_alloc_coherent recv failed\n");
322 goto out; 366 goto out;
@@ -324,22 +368,24 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
324 368
325 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), 369 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
326 &ic->i_ack_dma, GFP_KERNEL); 370 &ic->i_ack_dma, GFP_KERNEL);
327 if (ic->i_ack == NULL) { 371 if (!ic->i_ack) {
328 ret = -ENOMEM; 372 ret = -ENOMEM;
329 rdsdebug("ib_dma_alloc_coherent ack failed\n"); 373 rdsdebug("ib_dma_alloc_coherent ack failed\n");
330 goto out; 374 goto out;
331 } 375 }
332 376
333 ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); 377 ic->i_sends = vmalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
334 if (ic->i_sends == NULL) { 378 ibdev_to_node(dev));
379 if (!ic->i_sends) {
335 ret = -ENOMEM; 380 ret = -ENOMEM;
336 rdsdebug("send allocation failed\n"); 381 rdsdebug("send allocation failed\n");
337 goto out; 382 goto out;
338 } 383 }
339 memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); 384 memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
340 385
341 ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); 386 ic->i_recvs = vmalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
342 if (ic->i_recvs == NULL) { 387 ibdev_to_node(dev));
388 if (!ic->i_recvs) {
343 ret = -ENOMEM; 389 ret = -ENOMEM;
344 rdsdebug("recv allocation failed\n"); 390 rdsdebug("recv allocation failed\n");
345 goto out; 391 goto out;
@@ -352,6 +398,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
352 ic->i_send_cq, ic->i_recv_cq); 398 ic->i_send_cq, ic->i_recv_cq);
353 399
354out: 400out:
401 rds_ib_dev_put(rds_ibdev);
355 return ret; 402 return ret;
356} 403}
357 404
@@ -409,7 +456,7 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
409 struct rds_ib_connection *ic = NULL; 456 struct rds_ib_connection *ic = NULL;
410 struct rdma_conn_param conn_param; 457 struct rdma_conn_param conn_param;
411 u32 version; 458 u32 version;
412 int err, destroy = 1; 459 int err = 1, destroy = 1;
413 460
414 /* Check whether the remote protocol version matches ours. */ 461 /* Check whether the remote protocol version matches ours. */
415 version = rds_ib_protocol_compatible(event); 462 version = rds_ib_protocol_compatible(event);
@@ -448,7 +495,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
448 /* Wait and see - our connect may still be succeeding */ 495 /* Wait and see - our connect may still be succeeding */
449 rds_ib_stats_inc(s_ib_connect_raced); 496 rds_ib_stats_inc(s_ib_connect_raced);
450 } 497 }
451 mutex_unlock(&conn->c_cm_lock);
452 goto out; 498 goto out;
453 } 499 }
454 500
@@ -479,20 +525,20 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
479 goto out; 525 goto out;
480 } 526 }
481 527
482 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version); 528 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
529 event->param.conn.responder_resources,
530 event->param.conn.initiator_depth);
483 531
484 /* rdma_accept() calls rdma_reject() internally if it fails */ 532 /* rdma_accept() calls rdma_reject() internally if it fails */
485 err = rdma_accept(cm_id, &conn_param); 533 err = rdma_accept(cm_id, &conn_param);
486 mutex_unlock(&conn->c_cm_lock); 534 if (err)
487 if (err) {
488 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); 535 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
489 goto out;
490 }
491
492 return 0;
493 536
494out: 537out:
495 rdma_reject(cm_id, NULL, 0); 538 if (conn)
539 mutex_unlock(&conn->c_cm_lock);
540 if (err)
541 rdma_reject(cm_id, NULL, 0);
496 return destroy; 542 return destroy;
497} 543}
498 544
@@ -516,8 +562,8 @@ int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
516 goto out; 562 goto out;
517 } 563 }
518 564
519 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION); 565 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
520 566 UINT_MAX, UINT_MAX);
521 ret = rdma_connect(cm_id, &conn_param); 567 ret = rdma_connect(cm_id, &conn_param);
522 if (ret) 568 if (ret)
523 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); 569 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
@@ -601,9 +647,19 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
601 ic->i_cm_id, err); 647 ic->i_cm_id, err);
602 } 648 }
603 649
650 /*
651 * We want to wait for tx and rx completion to finish
652 * before we tear down the connection, but we have to be
653 * careful not to get stuck waiting on a send ring that
654 * only has unsignaled sends in it. We've shutdown new
655 * sends before getting here so by waiting for signaled
656 * sends to complete we're ensured that there will be no
657 * more tx processing.
658 */
604 wait_event(rds_ib_ring_empty_wait, 659 wait_event(rds_ib_ring_empty_wait,
605 rds_ib_ring_empty(&ic->i_send_ring) && 660 rds_ib_ring_empty(&ic->i_recv_ring) &&
606 rds_ib_ring_empty(&ic->i_recv_ring)); 661 (atomic_read(&ic->i_signaled_sends) == 0));
662 tasklet_kill(&ic->i_recv_tasklet);
607 663
608 if (ic->i_send_hdrs) 664 if (ic->i_send_hdrs)
609 ib_dma_free_coherent(dev, 665 ib_dma_free_coherent(dev,
@@ -654,9 +710,12 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
654 BUG_ON(ic->rds_ibdev); 710 BUG_ON(ic->rds_ibdev);
655 711
656 /* Clear pending transmit */ 712 /* Clear pending transmit */
657 if (ic->i_rm) { 713 if (ic->i_data_op) {
658 rds_message_put(ic->i_rm); 714 struct rds_message *rm;
659 ic->i_rm = NULL; 715
716 rm = container_of(ic->i_data_op, struct rds_message, data);
717 rds_message_put(rm);
718 ic->i_data_op = NULL;
660 } 719 }
661 720
662 /* Clear the ACK state */ 721 /* Clear the ACK state */
@@ -690,12 +749,19 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
690{ 749{
691 struct rds_ib_connection *ic; 750 struct rds_ib_connection *ic;
692 unsigned long flags; 751 unsigned long flags;
752 int ret;
693 753
694 /* XXX too lazy? */ 754 /* XXX too lazy? */
695 ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL); 755 ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
696 if (ic == NULL) 756 if (!ic)
697 return -ENOMEM; 757 return -ENOMEM;
698 758
759 ret = rds_ib_recv_alloc_caches(ic);
760 if (ret) {
761 kfree(ic);
762 return ret;
763 }
764
699 INIT_LIST_HEAD(&ic->ib_node); 765 INIT_LIST_HEAD(&ic->ib_node);
700 tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn, 766 tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
701 (unsigned long) ic); 767 (unsigned long) ic);
@@ -703,6 +769,7 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
703#ifndef KERNEL_HAS_ATOMIC64 769#ifndef KERNEL_HAS_ATOMIC64
704 spin_lock_init(&ic->i_ack_lock); 770 spin_lock_init(&ic->i_ack_lock);
705#endif 771#endif
772 atomic_set(&ic->i_signaled_sends, 0);
706 773
707 /* 774 /*
708 * rds_ib_conn_shutdown() waits for these to be emptied so they 775 * rds_ib_conn_shutdown() waits for these to be emptied so they
@@ -744,6 +811,8 @@ void rds_ib_conn_free(void *arg)
744 list_del(&ic->ib_node); 811 list_del(&ic->ib_node);
745 spin_unlock_irq(lock_ptr); 812 spin_unlock_irq(lock_ptr);
746 813
814 rds_ib_recv_free_caches(ic);
815
747 kfree(ic); 816 kfree(ic);
748} 817}
749 818
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index a54cd63f9e35..8f6e221c9f78 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -32,11 +32,16 @@
32 */ 32 */
33#include <linux/kernel.h> 33#include <linux/kernel.h>
34#include <linux/slab.h> 34#include <linux/slab.h>
35#include <linux/rculist.h>
35 36
36#include "rds.h" 37#include "rds.h"
37#include "rdma.h"
38#include "ib.h" 38#include "ib.h"
39#include "xlist.h"
39 40
41struct workqueue_struct *rds_ib_fmr_wq;
42
43static DEFINE_PER_CPU(unsigned long, clean_list_grace);
44#define CLEAN_LIST_BUSY_BIT 0
40 45
41/* 46/*
42 * This is stored as mr->r_trans_private. 47 * This is stored as mr->r_trans_private.
@@ -45,7 +50,11 @@ struct rds_ib_mr {
45 struct rds_ib_device *device; 50 struct rds_ib_device *device;
46 struct rds_ib_mr_pool *pool; 51 struct rds_ib_mr_pool *pool;
47 struct ib_fmr *fmr; 52 struct ib_fmr *fmr;
48 struct list_head list; 53
54 struct xlist_head xlist;
55
56 /* unmap_list is for freeing */
57 struct list_head unmap_list;
49 unsigned int remap_count; 58 unsigned int remap_count;
50 59
51 struct scatterlist *sg; 60 struct scatterlist *sg;
@@ -59,14 +68,16 @@ struct rds_ib_mr {
59 */ 68 */
60struct rds_ib_mr_pool { 69struct rds_ib_mr_pool {
61 struct mutex flush_lock; /* serialize fmr invalidate */ 70 struct mutex flush_lock; /* serialize fmr invalidate */
62 struct work_struct flush_worker; /* flush worker */ 71 struct delayed_work flush_worker; /* flush worker */
63 72
64 spinlock_t list_lock; /* protect variables below */
65 atomic_t item_count; /* total # of MRs */ 73 atomic_t item_count; /* total # of MRs */
66 atomic_t dirty_count; /* # dirty of MRs */ 74 atomic_t dirty_count; /* # dirty of MRs */
67 struct list_head drop_list; /* MRs that have reached their max_maps limit */ 75
68 struct list_head free_list; /* unused MRs */ 76 struct xlist_head drop_list; /* MRs that have reached their max_maps limit */
69 struct list_head clean_list; /* unused & unamapped MRs */ 77 struct xlist_head free_list; /* unused MRs */
78 struct xlist_head clean_list; /* global unused & unamapped MRs */
79 wait_queue_head_t flush_wait;
80
70 atomic_t free_pinned; /* memory pinned by free MRs */ 81 atomic_t free_pinned; /* memory pinned by free MRs */
71 unsigned long max_items; 82 unsigned long max_items;
72 unsigned long max_items_soft; 83 unsigned long max_items_soft;
@@ -74,7 +85,7 @@ struct rds_ib_mr_pool {
74 struct ib_fmr_attr fmr_attr; 85 struct ib_fmr_attr fmr_attr;
75}; 86};
76 87
77static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all); 88static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
78static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr); 89static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
79static void rds_ib_mr_pool_flush_worker(struct work_struct *work); 90static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
80 91
@@ -83,16 +94,17 @@ static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
83 struct rds_ib_device *rds_ibdev; 94 struct rds_ib_device *rds_ibdev;
84 struct rds_ib_ipaddr *i_ipaddr; 95 struct rds_ib_ipaddr *i_ipaddr;
85 96
86 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) { 97 rcu_read_lock();
87 spin_lock_irq(&rds_ibdev->spinlock); 98 list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
88 list_for_each_entry(i_ipaddr, &rds_ibdev->ipaddr_list, list) { 99 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
89 if (i_ipaddr->ipaddr == ipaddr) { 100 if (i_ipaddr->ipaddr == ipaddr) {
90 spin_unlock_irq(&rds_ibdev->spinlock); 101 atomic_inc(&rds_ibdev->refcount);
102 rcu_read_unlock();
91 return rds_ibdev; 103 return rds_ibdev;
92 } 104 }
93 } 105 }
94 spin_unlock_irq(&rds_ibdev->spinlock);
95 } 106 }
107 rcu_read_unlock();
96 108
97 return NULL; 109 return NULL;
98} 110}
@@ -108,7 +120,7 @@ static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
108 i_ipaddr->ipaddr = ipaddr; 120 i_ipaddr->ipaddr = ipaddr;
109 121
110 spin_lock_irq(&rds_ibdev->spinlock); 122 spin_lock_irq(&rds_ibdev->spinlock);
111 list_add_tail(&i_ipaddr->list, &rds_ibdev->ipaddr_list); 123 list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
112 spin_unlock_irq(&rds_ibdev->spinlock); 124 spin_unlock_irq(&rds_ibdev->spinlock);
113 125
114 return 0; 126 return 0;
@@ -116,17 +128,24 @@ static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
116 128
117static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) 129static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
118{ 130{
119 struct rds_ib_ipaddr *i_ipaddr, *next; 131 struct rds_ib_ipaddr *i_ipaddr;
132 struct rds_ib_ipaddr *to_free = NULL;
133
120 134
121 spin_lock_irq(&rds_ibdev->spinlock); 135 spin_lock_irq(&rds_ibdev->spinlock);
122 list_for_each_entry_safe(i_ipaddr, next, &rds_ibdev->ipaddr_list, list) { 136 list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
123 if (i_ipaddr->ipaddr == ipaddr) { 137 if (i_ipaddr->ipaddr == ipaddr) {
124 list_del(&i_ipaddr->list); 138 list_del_rcu(&i_ipaddr->list);
125 kfree(i_ipaddr); 139 to_free = i_ipaddr;
126 break; 140 break;
127 } 141 }
128 } 142 }
129 spin_unlock_irq(&rds_ibdev->spinlock); 143 spin_unlock_irq(&rds_ibdev->spinlock);
144
145 if (to_free) {
146 synchronize_rcu();
147 kfree(to_free);
148 }
130} 149}
131 150
132int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) 151int 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)
134 struct rds_ib_device *rds_ibdev_old; 153 struct rds_ib_device *rds_ibdev_old;
135 154
136 rds_ibdev_old = rds_ib_get_device(ipaddr); 155 rds_ibdev_old = rds_ib_get_device(ipaddr);
137 if (rds_ibdev_old) 156 if (rds_ibdev_old) {
138 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr); 157 rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
158 rds_ib_dev_put(rds_ibdev_old);
159 }
139 160
140 return rds_ib_add_ipaddr(rds_ibdev, ipaddr); 161 return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
141} 162}
@@ -156,6 +177,7 @@ void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *con
156 spin_unlock_irq(&ib_nodev_conns_lock); 177 spin_unlock_irq(&ib_nodev_conns_lock);
157 178
158 ic->rds_ibdev = rds_ibdev; 179 ic->rds_ibdev = rds_ibdev;
180 atomic_inc(&rds_ibdev->refcount);
159} 181}
160 182
161void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) 183void 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 *
175 spin_unlock(&ib_nodev_conns_lock); 197 spin_unlock(&ib_nodev_conns_lock);
176 198
177 ic->rds_ibdev = NULL; 199 ic->rds_ibdev = NULL;
200 rds_ib_dev_put(rds_ibdev);
178} 201}
179 202
180void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock) 203void rds_ib_destroy_nodev_conns(void)
181{ 204{
182 struct rds_ib_connection *ic, *_ic; 205 struct rds_ib_connection *ic, *_ic;
183 LIST_HEAD(tmp_list); 206 LIST_HEAD(tmp_list);
184 207
185 /* avoid calling conn_destroy with irqs off */ 208 /* avoid calling conn_destroy with irqs off */
186 spin_lock_irq(list_lock); 209 spin_lock_irq(&ib_nodev_conns_lock);
187 list_splice(list, &tmp_list); 210 list_splice(&ib_nodev_conns, &tmp_list);
188 INIT_LIST_HEAD(list); 211 spin_unlock_irq(&ib_nodev_conns_lock);
189 spin_unlock_irq(list_lock);
190 212
191 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) 213 list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
192 rds_conn_destroy(ic->conn); 214 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)
200 if (!pool) 222 if (!pool)
201 return ERR_PTR(-ENOMEM); 223 return ERR_PTR(-ENOMEM);
202 224
203 INIT_LIST_HEAD(&pool->free_list); 225 INIT_XLIST_HEAD(&pool->free_list);
204 INIT_LIST_HEAD(&pool->drop_list); 226 INIT_XLIST_HEAD(&pool->drop_list);
205 INIT_LIST_HEAD(&pool->clean_list); 227 INIT_XLIST_HEAD(&pool->clean_list);
206 mutex_init(&pool->flush_lock); 228 mutex_init(&pool->flush_lock);
207 spin_lock_init(&pool->list_lock); 229 init_waitqueue_head(&pool->flush_wait);
208 INIT_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker); 230 INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
209 231
210 pool->fmr_attr.max_pages = fmr_message_size; 232 pool->fmr_attr.max_pages = fmr_message_size;
211 pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps; 233 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
233 255
234void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool) 256void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
235{ 257{
236 flush_workqueue(rds_wq); 258 cancel_delayed_work_sync(&pool->flush_worker);
237 rds_ib_flush_mr_pool(pool, 1); 259 rds_ib_flush_mr_pool(pool, 1, NULL);
238 WARN_ON(atomic_read(&pool->item_count)); 260 WARN_ON(atomic_read(&pool->item_count));
239 WARN_ON(atomic_read(&pool->free_pinned)); 261 WARN_ON(atomic_read(&pool->free_pinned));
240 kfree(pool); 262 kfree(pool);
241} 263}
242 264
265static void refill_local(struct rds_ib_mr_pool *pool, struct xlist_head *xl,
266 struct rds_ib_mr **ibmr_ret)
267{
268 struct xlist_head *ibmr_xl;
269 ibmr_xl = xlist_del_head_fast(xl);
270 *ibmr_ret = list_entry(ibmr_xl, struct rds_ib_mr, xlist);
271}
272
243static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool) 273static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
244{ 274{
245 struct rds_ib_mr *ibmr = NULL; 275 struct rds_ib_mr *ibmr = NULL;
246 unsigned long flags; 276 struct xlist_head *ret;
277 unsigned long *flag;
247 278
248 spin_lock_irqsave(&pool->list_lock, flags); 279 preempt_disable();
249 if (!list_empty(&pool->clean_list)) { 280 flag = &__get_cpu_var(clean_list_grace);
250 ibmr = list_entry(pool->clean_list.next, struct rds_ib_mr, list); 281 set_bit(CLEAN_LIST_BUSY_BIT, flag);
251 list_del_init(&ibmr->list); 282 ret = xlist_del_head(&pool->clean_list);
252 } 283 if (ret)
253 spin_unlock_irqrestore(&pool->list_lock, flags); 284 ibmr = list_entry(ret, struct rds_ib_mr, xlist);
254 285
286 clear_bit(CLEAN_LIST_BUSY_BIT, flag);
287 preempt_enable();
255 return ibmr; 288 return ibmr;
256} 289}
257 290
291static inline void wait_clean_list_grace(void)
292{
293 int cpu;
294 unsigned long *flag;
295
296 for_each_online_cpu(cpu) {
297 flag = &per_cpu(clean_list_grace, cpu);
298 while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
299 cpu_relax();
300 }
301}
302
258static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev) 303static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
259{ 304{
260 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; 305 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
261 struct rds_ib_mr *ibmr = NULL; 306 struct rds_ib_mr *ibmr = NULL;
262 int err = 0, iter = 0; 307 int err = 0, iter = 0;
263 308
309 if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
310 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
311
264 while (1) { 312 while (1) {
265 ibmr = rds_ib_reuse_fmr(pool); 313 ibmr = rds_ib_reuse_fmr(pool);
266 if (ibmr) 314 if (ibmr)
@@ -287,19 +335,24 @@ static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev)
287 335
288 /* We do have some empty MRs. Flush them out. */ 336 /* We do have some empty MRs. Flush them out. */
289 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait); 337 rds_ib_stats_inc(s_ib_rdma_mr_pool_wait);
290 rds_ib_flush_mr_pool(pool, 0); 338 rds_ib_flush_mr_pool(pool, 0, &ibmr);
339 if (ibmr)
340 return ibmr;
291 } 341 }
292 342
293 ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL); 343 ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
294 if (!ibmr) { 344 if (!ibmr) {
295 err = -ENOMEM; 345 err = -ENOMEM;
296 goto out_no_cigar; 346 goto out_no_cigar;
297 } 347 }
298 348
349 memset(ibmr, 0, sizeof(*ibmr));
350
299 ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd, 351 ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
300 (IB_ACCESS_LOCAL_WRITE | 352 (IB_ACCESS_LOCAL_WRITE |
301 IB_ACCESS_REMOTE_READ | 353 IB_ACCESS_REMOTE_READ |
302 IB_ACCESS_REMOTE_WRITE), 354 IB_ACCESS_REMOTE_WRITE|
355 IB_ACCESS_REMOTE_ATOMIC),
303 &pool->fmr_attr); 356 &pool->fmr_attr);
304 if (IS_ERR(ibmr->fmr)) { 357 if (IS_ERR(ibmr->fmr)) {
305 err = PTR_ERR(ibmr->fmr); 358 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
367 if (page_cnt > fmr_message_size) 420 if (page_cnt > fmr_message_size)
368 return -EINVAL; 421 return -EINVAL;
369 422
370 dma_pages = kmalloc(sizeof(u64) * page_cnt, GFP_ATOMIC); 423 dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
424 rdsibdev_to_node(rds_ibdev));
371 if (!dma_pages) 425 if (!dma_pages)
372 return -ENOMEM; 426 return -ENOMEM;
373 427
@@ -441,7 +495,7 @@ static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
441 495
442 /* FIXME we need a way to tell a r/w MR 496 /* FIXME we need a way to tell a r/w MR
443 * from a r/o MR */ 497 * from a r/o MR */
444 BUG_ON(in_interrupt()); 498 BUG_ON(irqs_disabled());
445 set_page_dirty(page); 499 set_page_dirty(page);
446 put_page(page); 500 put_page(page);
447 } 501 }
@@ -477,33 +531,109 @@ static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int fr
477} 531}
478 532
479/* 533/*
534 * given an xlist of mrs, put them all into the list_head for more processing
535 */
536static void xlist_append_to_list(struct xlist_head *xlist, struct list_head *list)
537{
538 struct rds_ib_mr *ibmr;
539 struct xlist_head splice;
540 struct xlist_head *cur;
541 struct xlist_head *next;
542
543 splice.next = NULL;
544 xlist_splice(xlist, &splice);
545 cur = splice.next;
546 while (cur) {
547 next = cur->next;
548 ibmr = list_entry(cur, struct rds_ib_mr, xlist);
549 list_add_tail(&ibmr->unmap_list, list);
550 cur = next;
551 }
552}
553
554/*
555 * this takes a list head of mrs and turns it into an xlist of clusters.
556 * each cluster has an xlist of MR_CLUSTER_SIZE mrs that are ready for
557 * reuse.
558 */
559static void list_append_to_xlist(struct rds_ib_mr_pool *pool,
560 struct list_head *list, struct xlist_head *xlist,
561 struct xlist_head **tail_ret)
562{
563 struct rds_ib_mr *ibmr;
564 struct xlist_head *cur_mr = xlist;
565 struct xlist_head *tail_mr = NULL;
566
567 list_for_each_entry(ibmr, list, unmap_list) {
568 tail_mr = &ibmr->xlist;
569 tail_mr->next = NULL;
570 cur_mr->next = tail_mr;
571 cur_mr = tail_mr;
572 }
573 *tail_ret = tail_mr;
574}
575
576/*
480 * Flush our pool of MRs. 577 * Flush our pool of MRs.
481 * At a minimum, all currently unused MRs are unmapped. 578 * At a minimum, all currently unused MRs are unmapped.
482 * If the number of MRs allocated exceeds the limit, we also try 579 * If the number of MRs allocated exceeds the limit, we also try
483 * to free as many MRs as needed to get back to this limit. 580 * to free as many MRs as needed to get back to this limit.
484 */ 581 */
485static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all) 582static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
583 int free_all, struct rds_ib_mr **ibmr_ret)
486{ 584{
487 struct rds_ib_mr *ibmr, *next; 585 struct rds_ib_mr *ibmr, *next;
586 struct xlist_head clean_xlist;
587 struct xlist_head *clean_tail;
488 LIST_HEAD(unmap_list); 588 LIST_HEAD(unmap_list);
489 LIST_HEAD(fmr_list); 589 LIST_HEAD(fmr_list);
490 unsigned long unpinned = 0; 590 unsigned long unpinned = 0;
491 unsigned long flags;
492 unsigned int nfreed = 0, ncleaned = 0, free_goal; 591 unsigned int nfreed = 0, ncleaned = 0, free_goal;
493 int ret = 0; 592 int ret = 0;
494 593
495 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush); 594 rds_ib_stats_inc(s_ib_rdma_mr_pool_flush);
496 595
497 mutex_lock(&pool->flush_lock); 596 if (ibmr_ret) {
597 DEFINE_WAIT(wait);
598 while(!mutex_trylock(&pool->flush_lock)) {
599 ibmr = rds_ib_reuse_fmr(pool);
600 if (ibmr) {
601 *ibmr_ret = ibmr;
602 finish_wait(&pool->flush_wait, &wait);
603 goto out_nolock;
604 }
605
606 prepare_to_wait(&pool->flush_wait, &wait,
607 TASK_UNINTERRUPTIBLE);
608 if (xlist_empty(&pool->clean_list))
609 schedule();
610
611 ibmr = rds_ib_reuse_fmr(pool);
612 if (ibmr) {
613 *ibmr_ret = ibmr;
614 finish_wait(&pool->flush_wait, &wait);
615 goto out_nolock;
616 }
617 }
618 finish_wait(&pool->flush_wait, &wait);
619 } else
620 mutex_lock(&pool->flush_lock);
621
622 if (ibmr_ret) {
623 ibmr = rds_ib_reuse_fmr(pool);
624 if (ibmr) {
625 *ibmr_ret = ibmr;
626 goto out;
627 }
628 }
498 629
499 spin_lock_irqsave(&pool->list_lock, flags);
500 /* Get the list of all MRs to be dropped. Ordering matters - 630 /* Get the list of all MRs to be dropped. Ordering matters -
501 * we want to put drop_list ahead of free_list. */ 631 * we want to put drop_list ahead of free_list.
502 list_splice_init(&pool->free_list, &unmap_list); 632 */
503 list_splice_init(&pool->drop_list, &unmap_list); 633 xlist_append_to_list(&pool->drop_list, &unmap_list);
634 xlist_append_to_list(&pool->free_list, &unmap_list);
504 if (free_all) 635 if (free_all)
505 list_splice_init(&pool->clean_list, &unmap_list); 636 xlist_append_to_list(&pool->clean_list, &unmap_list);
506 spin_unlock_irqrestore(&pool->list_lock, flags);
507 637
508 free_goal = rds_ib_flush_goal(pool, free_all); 638 free_goal = rds_ib_flush_goal(pool, free_all);
509 639
@@ -511,19 +641,20 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
511 goto out; 641 goto out;
512 642
513 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */ 643 /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
514 list_for_each_entry(ibmr, &unmap_list, list) 644 list_for_each_entry(ibmr, &unmap_list, unmap_list)
515 list_add(&ibmr->fmr->list, &fmr_list); 645 list_add(&ibmr->fmr->list, &fmr_list);
646
516 ret = ib_unmap_fmr(&fmr_list); 647 ret = ib_unmap_fmr(&fmr_list);
517 if (ret) 648 if (ret)
518 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret); 649 printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
519 650
520 /* Now we can destroy the DMA mapping and unpin any pages */ 651 /* Now we can destroy the DMA mapping and unpin any pages */
521 list_for_each_entry_safe(ibmr, next, &unmap_list, list) { 652 list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
522 unpinned += ibmr->sg_len; 653 unpinned += ibmr->sg_len;
523 __rds_ib_teardown_mr(ibmr); 654 __rds_ib_teardown_mr(ibmr);
524 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) { 655 if (nfreed < free_goal || ibmr->remap_count >= pool->fmr_attr.max_maps) {
525 rds_ib_stats_inc(s_ib_rdma_mr_free); 656 rds_ib_stats_inc(s_ib_rdma_mr_free);
526 list_del(&ibmr->list); 657 list_del(&ibmr->unmap_list);
527 ib_dealloc_fmr(ibmr->fmr); 658 ib_dealloc_fmr(ibmr->fmr);
528 kfree(ibmr); 659 kfree(ibmr);
529 nfreed++; 660 nfreed++;
@@ -531,9 +662,27 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all)
531 ncleaned++; 662 ncleaned++;
532 } 663 }
533 664
534 spin_lock_irqsave(&pool->list_lock, flags); 665 if (!list_empty(&unmap_list)) {
535 list_splice(&unmap_list, &pool->clean_list); 666 /* we have to make sure that none of the things we're about
536 spin_unlock_irqrestore(&pool->list_lock, flags); 667 * to put on the clean list would race with other cpus trying
668 * to pull items off. The xlist would explode if we managed to
669 * remove something from the clean list and then add it back again
670 * while another CPU was spinning on that same item in xlist_del_head.
671 *
672 * This is pretty unlikely, but just in case wait for an xlist grace period
673 * here before adding anything back into the clean list.
674 */
675 wait_clean_list_grace();
676
677 list_append_to_xlist(pool, &unmap_list, &clean_xlist, &clean_tail);
678 if (ibmr_ret)
679 refill_local(pool, &clean_xlist, ibmr_ret);
680
681 /* refill_local may have emptied our list */
682 if (!xlist_empty(&clean_xlist))
683 xlist_add(clean_xlist.next, clean_tail, &pool->clean_list);
684
685 }
537 686
538 atomic_sub(unpinned, &pool->free_pinned); 687 atomic_sub(unpinned, &pool->free_pinned);
539 atomic_sub(ncleaned, &pool->dirty_count); 688 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)
541 690
542out: 691out:
543 mutex_unlock(&pool->flush_lock); 692 mutex_unlock(&pool->flush_lock);
693 if (waitqueue_active(&pool->flush_wait))
694 wake_up(&pool->flush_wait);
695out_nolock:
544 return ret; 696 return ret;
545} 697}
546 698
699int rds_ib_fmr_init(void)
700{
701 rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
702 if (!rds_ib_fmr_wq)
703 return -ENOMEM;
704 return 0;
705}
706
707/*
708 * By the time this is called all the IB devices should have been torn down and
709 * had their pools freed. As each pool is freed its work struct is waited on,
710 * so the pool flushing work queue should be idle by the time we get here.
711 */
712void rds_ib_fmr_exit(void)
713{
714 destroy_workqueue(rds_ib_fmr_wq);
715}
716
547static void rds_ib_mr_pool_flush_worker(struct work_struct *work) 717static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
548{ 718{
549 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker); 719 struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
550 720
551 rds_ib_flush_mr_pool(pool, 0); 721 rds_ib_flush_mr_pool(pool, 0, NULL);
552} 722}
553 723
554void rds_ib_free_mr(void *trans_private, int invalidate) 724void rds_ib_free_mr(void *trans_private, int invalidate)
@@ -556,47 +726,49 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
556 struct rds_ib_mr *ibmr = trans_private; 726 struct rds_ib_mr *ibmr = trans_private;
557 struct rds_ib_device *rds_ibdev = ibmr->device; 727 struct rds_ib_device *rds_ibdev = ibmr->device;
558 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; 728 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
559 unsigned long flags;
560 729
561 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len); 730 rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
562 731
563 /* Return it to the pool's free list */ 732 /* Return it to the pool's free list */
564 spin_lock_irqsave(&pool->list_lock, flags);
565 if (ibmr->remap_count >= pool->fmr_attr.max_maps) 733 if (ibmr->remap_count >= pool->fmr_attr.max_maps)
566 list_add(&ibmr->list, &pool->drop_list); 734 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->drop_list);
567 else 735 else
568 list_add(&ibmr->list, &pool->free_list); 736 xlist_add(&ibmr->xlist, &ibmr->xlist, &pool->free_list);
569 737
570 atomic_add(ibmr->sg_len, &pool->free_pinned); 738 atomic_add(ibmr->sg_len, &pool->free_pinned);
571 atomic_inc(&pool->dirty_count); 739 atomic_inc(&pool->dirty_count);
572 spin_unlock_irqrestore(&pool->list_lock, flags);
573 740
574 /* If we've pinned too many pages, request a flush */ 741 /* If we've pinned too many pages, request a flush */
575 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || 742 if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
576 atomic_read(&pool->dirty_count) >= pool->max_items / 10) 743 atomic_read(&pool->dirty_count) >= pool->max_items / 10)
577 queue_work(rds_wq, &pool->flush_worker); 744 queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
578 745
579 if (invalidate) { 746 if (invalidate) {
580 if (likely(!in_interrupt())) { 747 if (likely(!in_interrupt())) {
581 rds_ib_flush_mr_pool(pool, 0); 748 rds_ib_flush_mr_pool(pool, 0, NULL);
582 } else { 749 } else {
583 /* We get here if the user created a MR marked 750 /* We get here if the user created a MR marked
584 * as use_once and invalidate at the same time. */ 751 * as use_once and invalidate at the same time. */
585 queue_work(rds_wq, &pool->flush_worker); 752 queue_delayed_work(rds_ib_fmr_wq,
753 &pool->flush_worker, 10);
586 } 754 }
587 } 755 }
756
757 rds_ib_dev_put(rds_ibdev);
588} 758}
589 759
590void rds_ib_flush_mrs(void) 760void rds_ib_flush_mrs(void)
591{ 761{
592 struct rds_ib_device *rds_ibdev; 762 struct rds_ib_device *rds_ibdev;
593 763
764 down_read(&rds_ib_devices_lock);
594 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) { 765 list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
595 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool; 766 struct rds_ib_mr_pool *pool = rds_ibdev->mr_pool;
596 767
597 if (pool) 768 if (pool)
598 rds_ib_flush_mr_pool(pool, 0); 769 rds_ib_flush_mr_pool(pool, 0, NULL);
599 } 770 }
771 up_read(&rds_ib_devices_lock);
600} 772}
601 773
602void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, 774void *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,
628 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret); 800 printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
629 801
630 ibmr->device = rds_ibdev; 802 ibmr->device = rds_ibdev;
803 rds_ibdev = NULL;
631 804
632 out: 805 out:
633 if (ret) { 806 if (ret) {
@@ -635,5 +808,8 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
635 rds_ib_free_mr(ibmr, 0); 808 rds_ib_free_mr(ibmr, 0);
636 ibmr = ERR_PTR(ret); 809 ibmr = ERR_PTR(ret);
637 } 810 }
811 if (rds_ibdev)
812 rds_ib_dev_put(rds_ibdev);
638 return ibmr; 813 return ibmr;
639} 814}
815
diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index c74e9904a6b2..e29e0ca32f74 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -43,42 +43,6 @@ static struct kmem_cache *rds_ib_incoming_slab;
43static struct kmem_cache *rds_ib_frag_slab; 43static struct kmem_cache *rds_ib_frag_slab;
44static atomic_t rds_ib_allocation = ATOMIC_INIT(0); 44static atomic_t rds_ib_allocation = ATOMIC_INIT(0);
45 45
46static void rds_ib_frag_drop_page(struct rds_page_frag *frag)
47{
48 rdsdebug("frag %p page %p\n", frag, frag->f_page);
49 __free_page(frag->f_page);
50 frag->f_page = NULL;
51}
52
53static void rds_ib_frag_free(struct rds_page_frag *frag)
54{
55 rdsdebug("frag %p page %p\n", frag, frag->f_page);
56 BUG_ON(frag->f_page != NULL);
57 kmem_cache_free(rds_ib_frag_slab, frag);
58}
59
60/*
61 * We map a page at a time. Its fragments are posted in order. This
62 * is called in fragment order as the fragments get send completion events.
63 * Only the last frag in the page performs the unmapping.
64 *
65 * It's OK for ring cleanup to call this in whatever order it likes because
66 * DMA is not in flight and so we can unmap while other ring entries still
67 * hold page references in their frags.
68 */
69static void rds_ib_recv_unmap_page(struct rds_ib_connection *ic,
70 struct rds_ib_recv_work *recv)
71{
72 struct rds_page_frag *frag = recv->r_frag;
73
74 rdsdebug("recv %p frag %p page %p\n", recv, frag, frag->f_page);
75 if (frag->f_mapped)
76 ib_dma_unmap_page(ic->i_cm_id->device,
77 frag->f_mapped,
78 RDS_FRAG_SIZE, DMA_FROM_DEVICE);
79 frag->f_mapped = 0;
80}
81
82void rds_ib_recv_init_ring(struct rds_ib_connection *ic) 46void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
83{ 47{
84 struct rds_ib_recv_work *recv; 48 struct rds_ib_recv_work *recv;
@@ -95,16 +59,161 @@ void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
95 recv->r_wr.sg_list = recv->r_sge; 59 recv->r_wr.sg_list = recv->r_sge;
96 recv->r_wr.num_sge = RDS_IB_RECV_SGE; 60 recv->r_wr.num_sge = RDS_IB_RECV_SGE;
97 61
98 sge = rds_ib_data_sge(ic, recv->r_sge); 62 sge = &recv->r_sge[0];
63 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
64 sge->length = sizeof(struct rds_header);
65 sge->lkey = ic->i_mr->lkey;
66
67 sge = &recv->r_sge[1];
99 sge->addr = 0; 68 sge->addr = 0;
100 sge->length = RDS_FRAG_SIZE; 69 sge->length = RDS_FRAG_SIZE;
101 sge->lkey = ic->i_mr->lkey; 70 sge->lkey = ic->i_mr->lkey;
71 }
72}
102 73
103 sge = rds_ib_header_sge(ic, recv->r_sge); 74/*
104 sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); 75 * The entire 'from' list, including the from element itself, is put on
105 sge->length = sizeof(struct rds_header); 76 * to the tail of the 'to' list.
106 sge->lkey = ic->i_mr->lkey; 77 */
78static void list_splice_entire_tail(struct list_head *from,
79 struct list_head *to)
80{
81 struct list_head *from_last = from->prev;
82
83 list_splice_tail(from_last, to);
84 list_add_tail(from_last, to);
85}
86
87static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
88{
89 struct list_head *tmp;
90
91 tmp = xchg(&cache->xfer, NULL);
92 if (tmp) {
93 if (cache->ready)
94 list_splice_entire_tail(tmp, cache->ready);
95 else
96 cache->ready = tmp;
97 }
98}
99
100static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
101{
102 struct rds_ib_cache_head *head;
103 int cpu;
104
105 cache->percpu = alloc_percpu(struct rds_ib_cache_head);
106 if (!cache->percpu)
107 return -ENOMEM;
108
109 for_each_possible_cpu(cpu) {
110 head = per_cpu_ptr(cache->percpu, cpu);
111 head->first = NULL;
112 head->count = 0;
113 }
114 cache->xfer = NULL;
115 cache->ready = NULL;
116
117 return 0;
118}
119
120int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
121{
122 int ret;
123
124 ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
125 if (!ret) {
126 ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
127 if (ret)
128 free_percpu(ic->i_cache_incs.percpu);
107 } 129 }
130
131 return ret;
132}
133
134static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
135 struct list_head *caller_list)
136{
137 struct rds_ib_cache_head *head;
138 int cpu;
139
140 for_each_possible_cpu(cpu) {
141 head = per_cpu_ptr(cache->percpu, cpu);
142 if (head->first) {
143 list_splice_entire_tail(head->first, caller_list);
144 head->first = NULL;
145 }
146 }
147
148 if (cache->ready) {
149 list_splice_entire_tail(cache->ready, caller_list);
150 cache->ready = NULL;
151 }
152}
153
154void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
155{
156 struct rds_ib_incoming *inc;
157 struct rds_ib_incoming *inc_tmp;
158 struct rds_page_frag *frag;
159 struct rds_page_frag *frag_tmp;
160 LIST_HEAD(list);
161
162 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
163 rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
164 free_percpu(ic->i_cache_incs.percpu);
165
166 list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
167 list_del(&inc->ii_cache_entry);
168 WARN_ON(!list_empty(&inc->ii_frags));
169 kmem_cache_free(rds_ib_incoming_slab, inc);
170 }
171
172 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
173 rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
174 free_percpu(ic->i_cache_frags.percpu);
175
176 list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
177 list_del(&frag->f_cache_entry);
178 WARN_ON(!list_empty(&frag->f_item));
179 kmem_cache_free(rds_ib_frag_slab, frag);
180 }
181}
182
183/* fwd decl */
184static void rds_ib_recv_cache_put(struct list_head *new_item,
185 struct rds_ib_refill_cache *cache);
186static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);
187
188
189/* Recycle frag and attached recv buffer f_sg */
190static void rds_ib_frag_free(struct rds_ib_connection *ic,
191 struct rds_page_frag *frag)
192{
193 rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));
194
195 rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
196}
197
198/* Recycle inc after freeing attached frags */
199void rds_ib_inc_free(struct rds_incoming *inc)
200{
201 struct rds_ib_incoming *ibinc;
202 struct rds_page_frag *frag;
203 struct rds_page_frag *pos;
204 struct rds_ib_connection *ic = inc->i_conn->c_transport_data;
205
206 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
207
208 /* Free attached frags */
209 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
210 list_del_init(&frag->f_item);
211 rds_ib_frag_free(ic, frag);
212 }
213 BUG_ON(!list_empty(&ibinc->ii_frags));
214
215 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
216 rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
108} 217}
109 218
110static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, 219static 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,
115 recv->r_ibinc = NULL; 224 recv->r_ibinc = NULL;
116 } 225 }
117 if (recv->r_frag) { 226 if (recv->r_frag) {
118 rds_ib_recv_unmap_page(ic, recv); 227 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
119 if (recv->r_frag->f_page) 228 rds_ib_frag_free(ic, recv->r_frag);
120 rds_ib_frag_drop_page(recv->r_frag);
121 rds_ib_frag_free(recv->r_frag);
122 recv->r_frag = NULL; 229 recv->r_frag = NULL;
123 } 230 }
124} 231}
@@ -129,84 +236,111 @@ void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
129 236
130 for (i = 0; i < ic->i_recv_ring.w_nr; i++) 237 for (i = 0; i < ic->i_recv_ring.w_nr; i++)
131 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); 238 rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
132
133 if (ic->i_frag.f_page)
134 rds_ib_frag_drop_page(&ic->i_frag);
135} 239}
136 240
137static int rds_ib_recv_refill_one(struct rds_connection *conn, 241static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic,
138 struct rds_ib_recv_work *recv, 242 gfp_t slab_mask)
139 gfp_t kptr_gfp, gfp_t page_gfp)
140{ 243{
141 struct rds_ib_connection *ic = conn->c_transport_data; 244 struct rds_ib_incoming *ibinc;
142 dma_addr_t dma_addr; 245 struct list_head *cache_item;
143 struct ib_sge *sge; 246 int avail_allocs;
144 int ret = -ENOMEM;
145 247
146 if (recv->r_ibinc == NULL) { 248 cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
147 if (!atomic_add_unless(&rds_ib_allocation, 1, rds_ib_sysctl_max_recv_allocation)) { 249 if (cache_item) {
250 ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
251 } else {
252 avail_allocs = atomic_add_unless(&rds_ib_allocation,
253 1, rds_ib_sysctl_max_recv_allocation);
254 if (!avail_allocs) {
148 rds_ib_stats_inc(s_ib_rx_alloc_limit); 255 rds_ib_stats_inc(s_ib_rx_alloc_limit);
149 goto out; 256 return NULL;
150 } 257 }
151 recv->r_ibinc = kmem_cache_alloc(rds_ib_incoming_slab, 258 ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask);
152 kptr_gfp); 259 if (!ibinc) {
153 if (recv->r_ibinc == NULL) {
154 atomic_dec(&rds_ib_allocation); 260 atomic_dec(&rds_ib_allocation);
155 goto out; 261 return NULL;
156 } 262 }
157 INIT_LIST_HEAD(&recv->r_ibinc->ii_frags);
158 rds_inc_init(&recv->r_ibinc->ii_inc, conn, conn->c_faddr);
159 } 263 }
264 INIT_LIST_HEAD(&ibinc->ii_frags);
265 rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);
160 266
161 if (recv->r_frag == NULL) { 267 return ibinc;
162 recv->r_frag = kmem_cache_alloc(rds_ib_frag_slab, kptr_gfp); 268}
163 if (recv->r_frag == NULL) 269
164 goto out; 270static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic,
165 INIT_LIST_HEAD(&recv->r_frag->f_item); 271 gfp_t slab_mask, gfp_t page_mask)
166 recv->r_frag->f_page = NULL; 272{
273 struct rds_page_frag *frag;
274 struct list_head *cache_item;
275 int ret;
276
277 cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
278 if (cache_item) {
279 frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
280 } else {
281 frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask);
282 if (!frag)
283 return NULL;
284
285 sg_init_table(&frag->f_sg, 1);
286 ret = rds_page_remainder_alloc(&frag->f_sg,
287 RDS_FRAG_SIZE, page_mask);
288 if (ret) {
289 kmem_cache_free(rds_ib_frag_slab, frag);
290 return NULL;
291 }
167 } 292 }
168 293
169 if (ic->i_frag.f_page == NULL) { 294 INIT_LIST_HEAD(&frag->f_item);
170 ic->i_frag.f_page = alloc_page(page_gfp); 295
171 if (ic->i_frag.f_page == NULL) 296 return frag;
172 goto out; 297}
173 ic->i_frag.f_offset = 0; 298
299static int rds_ib_recv_refill_one(struct rds_connection *conn,
300 struct rds_ib_recv_work *recv, int prefill)
301{
302 struct rds_ib_connection *ic = conn->c_transport_data;
303 struct ib_sge *sge;
304 int ret = -ENOMEM;
305 gfp_t slab_mask = GFP_NOWAIT;
306 gfp_t page_mask = GFP_NOWAIT;
307
308 if (prefill) {
309 slab_mask = GFP_KERNEL;
310 page_mask = GFP_HIGHUSER;
174 } 311 }
175 312
176 dma_addr = ib_dma_map_page(ic->i_cm_id->device, 313 if (!ic->i_cache_incs.ready)
177 ic->i_frag.f_page, 314 rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
178 ic->i_frag.f_offset, 315 if (!ic->i_cache_frags.ready)
179 RDS_FRAG_SIZE, 316 rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
180 DMA_FROM_DEVICE);
181 if (ib_dma_mapping_error(ic->i_cm_id->device, dma_addr))
182 goto out;
183 317
184 /* 318 /*
185 * Once we get the RDS_PAGE_LAST_OFF frag then rds_ib_frag_unmap() 319 * ibinc was taken from recv if recv contained the start of a message.
186 * must be called on this recv. This happens as completions hit 320 * recvs that were continuations will still have this allocated.
187 * in order or on connection shutdown.
188 */ 321 */
189 recv->r_frag->f_page = ic->i_frag.f_page; 322 if (!recv->r_ibinc) {
190 recv->r_frag->f_offset = ic->i_frag.f_offset; 323 recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask);
191 recv->r_frag->f_mapped = dma_addr; 324 if (!recv->r_ibinc)
325 goto out;
326 }
192 327
193 sge = rds_ib_data_sge(ic, recv->r_sge); 328 WARN_ON(recv->r_frag); /* leak! */
194 sge->addr = dma_addr; 329 recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask);
195 sge->length = RDS_FRAG_SIZE; 330 if (!recv->r_frag)
331 goto out;
332
333 ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
334 1, DMA_FROM_DEVICE);
335 WARN_ON(ret != 1);
196 336
197 sge = rds_ib_header_sge(ic, recv->r_sge); 337 sge = &recv->r_sge[0];
198 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); 338 sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
199 sge->length = sizeof(struct rds_header); 339 sge->length = sizeof(struct rds_header);
200 340
201 get_page(recv->r_frag->f_page); 341 sge = &recv->r_sge[1];
202 342 sge->addr = sg_dma_address(&recv->r_frag->f_sg);
203 if (ic->i_frag.f_offset < RDS_PAGE_LAST_OFF) { 343 sge->length = sg_dma_len(&recv->r_frag->f_sg);
204 ic->i_frag.f_offset += RDS_FRAG_SIZE;
205 } else {
206 put_page(ic->i_frag.f_page);
207 ic->i_frag.f_page = NULL;
208 ic->i_frag.f_offset = 0;
209 }
210 344
211 ret = 0; 345 ret = 0;
212out: 346out:
@@ -216,13 +350,11 @@ out:
216/* 350/*
217 * This tries to allocate and post unused work requests after making sure that 351 * This tries to allocate and post unused work requests after making sure that
218 * they have all the allocations they need to queue received fragments into 352 * they have all the allocations they need to queue received fragments into
219 * sockets. The i_recv_mutex is held here so that ring_alloc and _unalloc 353 * sockets.
220 * pairs don't go unmatched.
221 * 354 *
222 * -1 is returned if posting fails due to temporary resource exhaustion. 355 * -1 is returned if posting fails due to temporary resource exhaustion.
223 */ 356 */
224int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, 357void rds_ib_recv_refill(struct rds_connection *conn, int prefill)
225 gfp_t page_gfp, int prefill)
226{ 358{
227 struct rds_ib_connection *ic = conn->c_transport_data; 359 struct rds_ib_connection *ic = conn->c_transport_data;
228 struct rds_ib_recv_work *recv; 360 struct rds_ib_recv_work *recv;
@@ -236,28 +368,25 @@ int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
236 if (pos >= ic->i_recv_ring.w_nr) { 368 if (pos >= ic->i_recv_ring.w_nr) {
237 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", 369 printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
238 pos); 370 pos);
239 ret = -EINVAL;
240 break; 371 break;
241 } 372 }
242 373
243 recv = &ic->i_recvs[pos]; 374 recv = &ic->i_recvs[pos];
244 ret = rds_ib_recv_refill_one(conn, recv, kptr_gfp, page_gfp); 375 ret = rds_ib_recv_refill_one(conn, recv, prefill);
245 if (ret) { 376 if (ret) {
246 ret = -1;
247 break; 377 break;
248 } 378 }
249 379
250 /* XXX when can this fail? */ 380 /* XXX when can this fail? */
251 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr); 381 ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
252 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv, 382 rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
253 recv->r_ibinc, recv->r_frag->f_page, 383 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
254 (long) recv->r_frag->f_mapped, ret); 384 (long) sg_dma_address(&recv->r_frag->f_sg), ret);
255 if (ret) { 385 if (ret) {
256 rds_ib_conn_error(conn, "recv post on " 386 rds_ib_conn_error(conn, "recv post on "
257 "%pI4 returned %d, disconnecting and " 387 "%pI4 returned %d, disconnecting and "
258 "reconnecting\n", &conn->c_faddr, 388 "reconnecting\n", &conn->c_faddr,
259 ret); 389 ret);
260 ret = -1;
261 break; 390 break;
262 } 391 }
263 392
@@ -270,37 +399,73 @@ int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
270 399
271 if (ret) 400 if (ret)
272 rds_ib_ring_unalloc(&ic->i_recv_ring, 1); 401 rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
273 return ret;
274} 402}
275 403
276void rds_ib_inc_purge(struct rds_incoming *inc) 404/*
405 * We want to recycle several types of recv allocations, like incs and frags.
406 * To use this, the *_free() function passes in the ptr to a list_head within
407 * the recyclee, as well as the cache to put it on.
408 *
409 * First, we put the memory on a percpu list. When this reaches a certain size,
410 * We move it to an intermediate non-percpu list in a lockless manner, with some
411 * xchg/compxchg wizardry.
412 *
413 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
414 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
415 * list_empty() will return true with one element is actually present.
416 */
417static void rds_ib_recv_cache_put(struct list_head *new_item,
418 struct rds_ib_refill_cache *cache)
277{ 419{
278 struct rds_ib_incoming *ibinc; 420 unsigned long flags;
279 struct rds_page_frag *frag; 421 struct rds_ib_cache_head *chp;
280 struct rds_page_frag *pos; 422 struct list_head *old;
281 423
282 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 424 local_irq_save(flags);
283 rdsdebug("purging ibinc %p inc %p\n", ibinc, inc);
284 425
285 list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { 426 chp = per_cpu_ptr(cache->percpu, smp_processor_id());
286 list_del_init(&frag->f_item); 427 if (!chp->first)
287 rds_ib_frag_drop_page(frag); 428 INIT_LIST_HEAD(new_item);
288 rds_ib_frag_free(frag); 429 else /* put on front */
289 } 430 list_add_tail(new_item, chp->first);
431 chp->first = new_item;
432 chp->count++;
433
434 if (chp->count < RDS_IB_RECYCLE_BATCH_COUNT)
435 goto end;
436
437 /*
438 * Return our per-cpu first list to the cache's xfer by atomically
439 * grabbing the current xfer list, appending it to our per-cpu list,
440 * and then atomically returning that entire list back to the
441 * cache's xfer list as long as it's still empty.
442 */
443 do {
444 old = xchg(&cache->xfer, NULL);
445 if (old)
446 list_splice_entire_tail(old, chp->first);
447 old = cmpxchg(&cache->xfer, NULL, chp->first);
448 } while (old);
449
450 chp->first = NULL;
451 chp->count = 0;
452end:
453 local_irq_restore(flags);
290} 454}
291 455
292void rds_ib_inc_free(struct rds_incoming *inc) 456static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
293{ 457{
294 struct rds_ib_incoming *ibinc; 458 struct list_head *head = cache->ready;
295 459
296 ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); 460 if (head) {
461 if (!list_empty(head)) {
462 cache->ready = head->next;
463 list_del_init(head);
464 } else
465 cache->ready = NULL;
466 }
297 467
298 rds_ib_inc_purge(inc); 468 return head;
299 rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
300 BUG_ON(!list_empty(&ibinc->ii_frags));
301 kmem_cache_free(rds_ib_incoming_slab, ibinc);
302 atomic_dec(&rds_ib_allocation);
303 BUG_ON(atomic_read(&rds_ib_allocation) < 0);
304} 469}
305 470
306int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov, 471int 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,
336 to_copy = min_t(unsigned long, to_copy, len - copied); 501 to_copy = min_t(unsigned long, to_copy, len - copied);
337 502
338 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag " 503 rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
339 "[%p, %lu] + %lu\n", 504 "[%p, %u] + %lu\n",
340 to_copy, iov->iov_base, iov->iov_len, iov_off, 505 to_copy, iov->iov_base, iov->iov_len, iov_off,
341 frag->f_page, frag->f_offset, frag_off); 506 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);
342 507
343 /* XXX needs + offset for multiple recvs per page */ 508 /* XXX needs + offset for multiple recvs per page */
344 ret = rds_page_copy_to_user(frag->f_page, 509 ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
345 frag->f_offset + frag_off, 510 frag->f_sg.offset + frag_off,
346 iov->iov_base + iov_off, 511 iov->iov_base + iov_off,
347 to_copy); 512 to_copy);
348 if (ret) { 513 if (ret) {
@@ -557,47 +722,6 @@ u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
557 return rds_ib_get_ack(ic); 722 return rds_ib_get_ack(ic);
558} 723}
559 724
560static struct rds_header *rds_ib_get_header(struct rds_connection *conn,
561 struct rds_ib_recv_work *recv,
562 u32 data_len)
563{
564 struct rds_ib_connection *ic = conn->c_transport_data;
565 void *hdr_buff = &ic->i_recv_hdrs[recv - ic->i_recvs];
566 void *addr;
567 u32 misplaced_hdr_bytes;
568
569 /*
570 * Support header at the front (RDS 3.1+) as well as header-at-end.
571 *
572 * Cases:
573 * 1) header all in header buff (great!)
574 * 2) header all in data page (copy all to header buff)
575 * 3) header split across hdr buf + data page
576 * (move bit in hdr buff to end before copying other bit from data page)
577 */
578 if (conn->c_version > RDS_PROTOCOL_3_0 || data_len == RDS_FRAG_SIZE)
579 return hdr_buff;
580
581 if (data_len <= (RDS_FRAG_SIZE - sizeof(struct rds_header))) {
582 addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0);
583 memcpy(hdr_buff,
584 addr + recv->r_frag->f_offset + data_len,
585 sizeof(struct rds_header));
586 kunmap_atomic(addr, KM_SOFTIRQ0);
587 return hdr_buff;
588 }
589
590 misplaced_hdr_bytes = (sizeof(struct rds_header) - (RDS_FRAG_SIZE - data_len));
591
592 memmove(hdr_buff + misplaced_hdr_bytes, hdr_buff, misplaced_hdr_bytes);
593
594 addr = kmap_atomic(recv->r_frag->f_page, KM_SOFTIRQ0);
595 memcpy(hdr_buff, addr + recv->r_frag->f_offset + data_len,
596 sizeof(struct rds_header) - misplaced_hdr_bytes);
597 kunmap_atomic(addr, KM_SOFTIRQ0);
598 return hdr_buff;
599}
600
601/* 725/*
602 * It's kind of lame that we're copying from the posted receive pages into 726 * It's kind of lame that we're copying from the posted receive pages into
603 * long-lived bitmaps. We could have posted the bitmaps and rdma written into 727 * 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,
639 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); 763 to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
640 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ 764 BUG_ON(to_copy & 7); /* Must be 64bit aligned. */
641 765
642 addr = kmap_atomic(frag->f_page, KM_SOFTIRQ0); 766 addr = kmap_atomic(sg_page(&frag->f_sg), KM_SOFTIRQ0);
643 767
644 src = addr + frag_off; 768 src = addr + frag_off;
645 dst = (void *)map->m_page_addrs[map_page] + map_off; 769 dst = (void *)map->m_page_addrs[map_page] + map_off;
@@ -710,7 +834,7 @@ static void rds_ib_process_recv(struct rds_connection *conn,
710 } 834 }
711 data_len -= sizeof(struct rds_header); 835 data_len -= sizeof(struct rds_header);
712 836
713 ihdr = rds_ib_get_header(conn, recv, data_len); 837 ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];
714 838
715 /* Validate the checksum. */ 839 /* Validate the checksum. */
716 if (!rds_message_verify_checksum(ihdr)) { 840 if (!rds_message_verify_checksum(ihdr)) {
@@ -742,12 +866,12 @@ static void rds_ib_process_recv(struct rds_connection *conn,
742 * the inc is freed. We don't go that route, so we have to drop the 866 * the inc is freed. We don't go that route, so we have to drop the
743 * page ref ourselves. We can't just leave the page on the recv 867 * page ref ourselves. We can't just leave the page on the recv
744 * because that confuses the dma mapping of pages and each recv's use 868 * because that confuses the dma mapping of pages and each recv's use
745 * of a partial page. We can leave the frag, though, it will be 869 * of a partial page.
746 * reused.
747 * 870 *
748 * FIXME: Fold this into the code path below. 871 * FIXME: Fold this into the code path below.
749 */ 872 */
750 rds_ib_frag_drop_page(recv->r_frag); 873 rds_ib_frag_free(ic, recv->r_frag);
874 recv->r_frag = NULL;
751 return; 875 return;
752 } 876 }
753 877
@@ -757,7 +881,7 @@ static void rds_ib_process_recv(struct rds_connection *conn,
757 * into the inc and save the inc so we can hang upcoming fragments 881 * into the inc and save the inc so we can hang upcoming fragments
758 * off its list. 882 * off its list.
759 */ 883 */
760 if (ibinc == NULL) { 884 if (!ibinc) {
761 ibinc = recv->r_ibinc; 885 ibinc = recv->r_ibinc;
762 recv->r_ibinc = NULL; 886 recv->r_ibinc = NULL;
763 ic->i_ibinc = ibinc; 887 ic->i_ibinc = ibinc;
@@ -842,32 +966,38 @@ static inline void rds_poll_cq(struct rds_ib_connection *ic,
842 struct rds_ib_recv_work *recv; 966 struct rds_ib_recv_work *recv;
843 967
844 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) { 968 while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
845 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 969 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
846 (unsigned long long)wc.wr_id, wc.status, wc.byte_len, 970 (unsigned long long)wc.wr_id, wc.status,
971 rds_ib_wc_status_str(wc.status), wc.byte_len,
847 be32_to_cpu(wc.ex.imm_data)); 972 be32_to_cpu(wc.ex.imm_data));
848 rds_ib_stats_inc(s_ib_rx_cq_event); 973 rds_ib_stats_inc(s_ib_rx_cq_event);
849 974
850 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; 975 recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];
851 976
852 rds_ib_recv_unmap_page(ic, recv); 977 ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
853 978
854 /* 979 /*
855 * Also process recvs in connecting state because it is possible 980 * Also process recvs in connecting state because it is possible
856 * to get a recv completion _before_ the rdmacm ESTABLISHED 981 * to get a recv completion _before_ the rdmacm ESTABLISHED
857 * event is processed. 982 * event is processed.
858 */ 983 */
859 if (rds_conn_up(conn) || rds_conn_connecting(conn)) { 984 if (wc.status == IB_WC_SUCCESS) {
985 rds_ib_process_recv(conn, recv, wc.byte_len, state);
986 } else {
860 /* We expect errors as the qp is drained during shutdown */ 987 /* We expect errors as the qp is drained during shutdown */
861 if (wc.status == IB_WC_SUCCESS) { 988 if (rds_conn_up(conn) || rds_conn_connecting(conn))
862 rds_ib_process_recv(conn, recv, wc.byte_len, state); 989 rds_ib_conn_error(conn, "recv completion on %pI4 had "
863 } else { 990 "status %u (%s), disconnecting and "
864 rds_ib_conn_error(conn, "recv completion on " 991 "reconnecting\n", &conn->c_faddr,
865 "%pI4 had status %u, disconnecting and " 992 wc.status,
866 "reconnecting\n", &conn->c_faddr, 993 rds_ib_wc_status_str(wc.status));
867 wc.status);
868 }
869 } 994 }
870 995
996 /*
997 * It's very important that we only free this ring entry if we've truly
998 * freed the resources allocated to the entry. The refilling path can
999 * leak if we don't.
1000 */
871 rds_ib_ring_free(&ic->i_recv_ring, 1); 1001 rds_ib_ring_free(&ic->i_recv_ring, 1);
872 } 1002 }
873} 1003}
@@ -897,11 +1027,8 @@ void rds_ib_recv_tasklet_fn(unsigned long data)
897 if (rds_ib_ring_empty(&ic->i_recv_ring)) 1027 if (rds_ib_ring_empty(&ic->i_recv_ring))
898 rds_ib_stats_inc(s_ib_rx_ring_empty); 1028 rds_ib_stats_inc(s_ib_rx_ring_empty);
899 1029
900 /*
901 * If the ring is running low, then schedule the thread to refill.
902 */
903 if (rds_ib_ring_low(&ic->i_recv_ring)) 1030 if (rds_ib_ring_low(&ic->i_recv_ring))
904 queue_delayed_work(rds_wq, &conn->c_recv_w, 0); 1031 rds_ib_recv_refill(conn, 0);
905} 1032}
906 1033
907int rds_ib_recv(struct rds_connection *conn) 1034int rds_ib_recv(struct rds_connection *conn)
@@ -910,25 +1037,13 @@ int rds_ib_recv(struct rds_connection *conn)
910 int ret = 0; 1037 int ret = 0;
911 1038
912 rdsdebug("conn %p\n", conn); 1039 rdsdebug("conn %p\n", conn);
913
914 /*
915 * If we get a temporary posting failure in this context then
916 * we're really low and we want the caller to back off for a bit.
917 */
918 mutex_lock(&ic->i_recv_mutex);
919 if (rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 0))
920 ret = -ENOMEM;
921 else
922 rds_ib_stats_inc(s_ib_rx_refill_from_thread);
923 mutex_unlock(&ic->i_recv_mutex);
924
925 if (rds_conn_up(conn)) 1040 if (rds_conn_up(conn))
926 rds_ib_attempt_ack(ic); 1041 rds_ib_attempt_ack(ic);
927 1042
928 return ret; 1043 return ret;
929} 1044}
930 1045
931int __init rds_ib_recv_init(void) 1046int rds_ib_recv_init(void)
932{ 1047{
933 struct sysinfo si; 1048 struct sysinfo si;
934 int ret = -ENOMEM; 1049 int ret = -ENOMEM;
@@ -939,14 +1054,14 @@ int __init rds_ib_recv_init(void)
939 1054
940 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", 1055 rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
941 sizeof(struct rds_ib_incoming), 1056 sizeof(struct rds_ib_incoming),
942 0, 0, NULL); 1057 0, SLAB_HWCACHE_ALIGN, NULL);
943 if (rds_ib_incoming_slab == NULL) 1058 if (!rds_ib_incoming_slab)
944 goto out; 1059 goto out;
945 1060
946 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", 1061 rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
947 sizeof(struct rds_page_frag), 1062 sizeof(struct rds_page_frag),
948 0, 0, NULL); 1063 0, SLAB_HWCACHE_ALIGN, NULL);
949 if (rds_ib_frag_slab == NULL) 1064 if (!rds_ib_frag_slab)
950 kmem_cache_destroy(rds_ib_incoming_slab); 1065 kmem_cache_destroy(rds_ib_incoming_slab);
951 else 1066 else
952 ret = 0; 1067 ret = 0;
diff --git a/net/rds/ib_send.c b/net/rds/ib_send.c
index 17fa80803ab0..71f373c421bc 100644
--- a/net/rds/ib_send.c
+++ b/net/rds/ib_send.c
@@ -36,11 +36,49 @@
36#include <linux/dmapool.h> 36#include <linux/dmapool.h>
37 37
38#include "rds.h" 38#include "rds.h"
39#include "rdma.h"
40#include "ib.h" 39#include "ib.h"
41 40
42static void rds_ib_send_rdma_complete(struct rds_message *rm, 41static char *rds_ib_wc_status_strings[] = {
43 int wc_status) 42#define RDS_IB_WC_STATUS_STR(foo) \
43 [IB_WC_##foo] = __stringify(IB_WC_##foo)
44 RDS_IB_WC_STATUS_STR(SUCCESS),
45 RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
46 RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
47 RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
48 RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
49 RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
50 RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
51 RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
52 RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
53 RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
54 RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
55 RDS_IB_WC_STATUS_STR(REM_OP_ERR),
56 RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
57 RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
58 RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
59 RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
60 RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
61 RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
62 RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
63 RDS_IB_WC_STATUS_STR(FATAL_ERR),
64 RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
65 RDS_IB_WC_STATUS_STR(GENERAL_ERR),
66#undef RDS_IB_WC_STATUS_STR
67};
68
69char *rds_ib_wc_status_str(enum ib_wc_status status)
70{
71 return rds_str_array(rds_ib_wc_status_strings,
72 ARRAY_SIZE(rds_ib_wc_status_strings), status);
73}
74
75/*
76 * Convert IB-specific error message to RDS error message and call core
77 * completion handler.
78 */
79static void rds_ib_send_complete(struct rds_message *rm,
80 int wc_status,
81 void (*complete)(struct rds_message *rm, int status))
44{ 82{
45 int notify_status; 83 int notify_status;
46 84
@@ -60,69 +98,125 @@ static void rds_ib_send_rdma_complete(struct rds_message *rm,
60 notify_status = RDS_RDMA_OTHER_ERROR; 98 notify_status = RDS_RDMA_OTHER_ERROR;
61 break; 99 break;
62 } 100 }
63 rds_rdma_send_complete(rm, notify_status); 101 complete(rm, notify_status);
102}
103
104static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
105 struct rm_data_op *op,
106 int wc_status)
107{
108 if (op->op_nents)
109 ib_dma_unmap_sg(ic->i_cm_id->device,
110 op->op_sg, op->op_nents,
111 DMA_TO_DEVICE);
64} 112}
65 113
66static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic, 114static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
67 struct rds_rdma_op *op) 115 struct rm_rdma_op *op,
116 int wc_status)
68{ 117{
69 if (op->r_mapped) { 118 if (op->op_mapped) {
70 ib_dma_unmap_sg(ic->i_cm_id->device, 119 ib_dma_unmap_sg(ic->i_cm_id->device,
71 op->r_sg, op->r_nents, 120 op->op_sg, op->op_nents,
72 op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE); 121 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
73 op->r_mapped = 0; 122 op->op_mapped = 0;
74 } 123 }
124
125 /* If the user asked for a completion notification on this
126 * message, we can implement three different semantics:
127 * 1. Notify when we received the ACK on the RDS message
128 * that was queued with the RDMA. This provides reliable
129 * notification of RDMA status at the expense of a one-way
130 * packet delay.
131 * 2. Notify when the IB stack gives us the completion event for
132 * the RDMA operation.
133 * 3. Notify when the IB stack gives us the completion event for
134 * the accompanying RDS messages.
135 * Here, we implement approach #3. To implement approach #2,
136 * we would need to take an event for the rdma WR. To implement #1,
137 * don't call rds_rdma_send_complete at all, and fall back to the notify
138 * handling in the ACK processing code.
139 *
140 * Note: There's no need to explicitly sync any RDMA buffers using
141 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
142 * operation itself unmapped the RDMA buffers, which takes care
143 * of synching.
144 */
145 rds_ib_send_complete(container_of(op, struct rds_message, rdma),
146 wc_status, rds_rdma_send_complete);
147
148 if (op->op_write)
149 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
150 else
151 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
75} 152}
76 153
77static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic, 154static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
78 struct rds_ib_send_work *send, 155 struct rm_atomic_op *op,
79 int wc_status) 156 int wc_status)
80{ 157{
81 struct rds_message *rm = send->s_rm; 158 /* unmap atomic recvbuf */
82 159 if (op->op_mapped) {
83 rdsdebug("ic %p send %p rm %p\n", ic, send, rm); 160 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
84 161 DMA_FROM_DEVICE);
85 ib_dma_unmap_sg(ic->i_cm_id->device, 162 op->op_mapped = 0;
86 rm->m_sg, rm->m_nents, 163 }
87 DMA_TO_DEVICE);
88
89 if (rm->m_rdma_op != NULL) {
90 rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
91
92 /* If the user asked for a completion notification on this
93 * message, we can implement three different semantics:
94 * 1. Notify when we received the ACK on the RDS message
95 * that was queued with the RDMA. This provides reliable
96 * notification of RDMA status at the expense of a one-way
97 * packet delay.
98 * 2. Notify when the IB stack gives us the completion event for
99 * the RDMA operation.
100 * 3. Notify when the IB stack gives us the completion event for
101 * the accompanying RDS messages.
102 * Here, we implement approach #3. To implement approach #2,
103 * call rds_rdma_send_complete from the cq_handler. To implement #1,
104 * don't call rds_rdma_send_complete at all, and fall back to the notify
105 * handling in the ACK processing code.
106 *
107 * Note: There's no need to explicitly sync any RDMA buffers using
108 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
109 * operation itself unmapped the RDMA buffers, which takes care
110 * of synching.
111 */
112 rds_ib_send_rdma_complete(rm, wc_status);
113 164
114 if (rm->m_rdma_op->r_write) 165 rds_ib_send_complete(container_of(op, struct rds_message, atomic),
115 rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes); 166 wc_status, rds_atomic_send_complete);
116 else 167
117 rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes); 168 if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
169 rds_ib_stats_inc(s_ib_atomic_cswp);
170 else
171 rds_ib_stats_inc(s_ib_atomic_fadd);
172}
173
174/*
175 * Unmap the resources associated with a struct send_work.
176 *
177 * Returns the rm for no good reason other than it is unobtainable
178 * other than by switching on wr.opcode, currently, and the caller,
179 * the event handler, needs it.
180 */
181static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
182 struct rds_ib_send_work *send,
183 int wc_status)
184{
185 struct rds_message *rm = NULL;
186
187 /* In the error case, wc.opcode sometimes contains garbage */
188 switch (send->s_wr.opcode) {
189 case IB_WR_SEND:
190 if (send->s_op) {
191 rm = container_of(send->s_op, struct rds_message, data);
192 rds_ib_send_unmap_data(ic, send->s_op, wc_status);
193 }
194 break;
195 case IB_WR_RDMA_WRITE:
196 case IB_WR_RDMA_READ:
197 if (send->s_op) {
198 rm = container_of(send->s_op, struct rds_message, rdma);
199 rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
200 }
201 break;
202 case IB_WR_ATOMIC_FETCH_AND_ADD:
203 case IB_WR_ATOMIC_CMP_AND_SWP:
204 if (send->s_op) {
205 rm = container_of(send->s_op, struct rds_message, atomic);
206 rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
207 }
208 break;
209 default:
210 if (printk_ratelimit())
211 printk(KERN_NOTICE
212 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
213 __func__, send->s_wr.opcode);
214 break;
118 } 215 }
119 216
120 /* If anyone waited for this message to get flushed out, wake 217 send->s_wr.opcode = 0xdead;
121 * them up now */
122 rds_message_unmapped(rm);
123 218
124 rds_message_put(rm); 219 return rm;
125 send->s_rm = NULL;
126} 220}
127 221
128void rds_ib_send_init_ring(struct rds_ib_connection *ic) 222void 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)
133 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { 227 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
134 struct ib_sge *sge; 228 struct ib_sge *sge;
135 229
136 send->s_rm = NULL;
137 send->s_op = NULL; 230 send->s_op = NULL;
138 231
139 send->s_wr.wr_id = i; 232 send->s_wr.wr_id = i;
140 send->s_wr.sg_list = send->s_sge; 233 send->s_wr.sg_list = send->s_sge;
141 send->s_wr.num_sge = 1;
142 send->s_wr.opcode = IB_WR_SEND;
143 send->s_wr.send_flags = 0;
144 send->s_wr.ex.imm_data = 0; 234 send->s_wr.ex.imm_data = 0;
145 235
146 sge = rds_ib_data_sge(ic, send->s_sge); 236 sge = &send->s_sge[0];
147 sge->lkey = ic->i_mr->lkey;
148
149 sge = rds_ib_header_sge(ic, send->s_sge);
150 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header)); 237 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
151 sge->length = sizeof(struct rds_header); 238 sge->length = sizeof(struct rds_header);
152 sge->lkey = ic->i_mr->lkey; 239 sge->lkey = ic->i_mr->lkey;
240
241 send->s_sge[1].lkey = ic->i_mr->lkey;
153 } 242 }
154} 243}
155 244
@@ -159,16 +248,24 @@ void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
159 u32 i; 248 u32 i;
160 249
161 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { 250 for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
162 if (send->s_wr.opcode == 0xdead) 251 if (send->s_op && send->s_wr.opcode != 0xdead)
163 continue; 252 rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
164 if (send->s_rm)
165 rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
166 if (send->s_op)
167 rds_ib_send_unmap_rdma(ic, send->s_op);
168 } 253 }
169} 254}
170 255
171/* 256/*
257 * The only fast path caller always has a non-zero nr, so we don't
258 * bother testing nr before performing the atomic sub.
259 */
260static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
261{
262 if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
263 waitqueue_active(&rds_ib_ring_empty_wait))
264 wake_up(&rds_ib_ring_empty_wait);
265 BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
266}
267
268/*
172 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc 269 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
173 * operations performed in the send path. As the sender allocs and potentially 270 * operations performed in the send path. As the sender allocs and potentially
174 * unallocs the next free entry in the ring it doesn't alter which is 271 * unallocs the next free entry in the ring it doesn't alter which is
@@ -178,12 +275,14 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
178{ 275{
179 struct rds_connection *conn = context; 276 struct rds_connection *conn = context;
180 struct rds_ib_connection *ic = conn->c_transport_data; 277 struct rds_ib_connection *ic = conn->c_transport_data;
278 struct rds_message *rm = NULL;
181 struct ib_wc wc; 279 struct ib_wc wc;
182 struct rds_ib_send_work *send; 280 struct rds_ib_send_work *send;
183 u32 completed; 281 u32 completed;
184 u32 oldest; 282 u32 oldest;
185 u32 i = 0; 283 u32 i = 0;
186 int ret; 284 int ret;
285 int nr_sig = 0;
187 286
188 rdsdebug("cq %p conn %p\n", cq, conn); 287 rdsdebug("cq %p conn %p\n", cq, conn);
189 rds_ib_stats_inc(s_ib_tx_cq_call); 288 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)
192 rdsdebug("ib_req_notify_cq send failed: %d\n", ret); 291 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
193 292
194 while (ib_poll_cq(cq, 1, &wc) > 0) { 293 while (ib_poll_cq(cq, 1, &wc) > 0) {
195 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 294 rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
196 (unsigned long long)wc.wr_id, wc.status, wc.byte_len, 295 (unsigned long long)wc.wr_id, wc.status,
296 rds_ib_wc_status_str(wc.status), wc.byte_len,
197 be32_to_cpu(wc.ex.imm_data)); 297 be32_to_cpu(wc.ex.imm_data));
198 rds_ib_stats_inc(s_ib_tx_cq_event); 298 rds_ib_stats_inc(s_ib_tx_cq_event);
199 299
@@ -210,51 +310,30 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
210 310
211 for (i = 0; i < completed; i++) { 311 for (i = 0; i < completed; i++) {
212 send = &ic->i_sends[oldest]; 312 send = &ic->i_sends[oldest];
313 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
314 nr_sig++;
213 315
214 /* In the error case, wc.opcode sometimes contains garbage */ 316 rm = rds_ib_send_unmap_op(ic, send, wc.status);
215 switch (send->s_wr.opcode) {
216 case IB_WR_SEND:
217 if (send->s_rm)
218 rds_ib_send_unmap_rm(ic, send, wc.status);
219 break;
220 case IB_WR_RDMA_WRITE:
221 case IB_WR_RDMA_READ:
222 /* Nothing to be done - the SG list will be unmapped
223 * when the SEND completes. */
224 break;
225 default:
226 if (printk_ratelimit())
227 printk(KERN_NOTICE
228 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
229 __func__, send->s_wr.opcode);
230 break;
231 }
232 317
233 send->s_wr.opcode = 0xdead;
234 send->s_wr.num_sge = 1;
235 if (send->s_queued + HZ/2 < jiffies) 318 if (send->s_queued + HZ/2 < jiffies)
236 rds_ib_stats_inc(s_ib_tx_stalled); 319 rds_ib_stats_inc(s_ib_tx_stalled);
237 320
238 /* If a RDMA operation produced an error, signal this right 321 if (send->s_op) {
239 * away. If we don't, the subsequent SEND that goes with this 322 if (send->s_op == rm->m_final_op) {
240 * RDMA will be canceled with ERR_WFLUSH, and the application 323 /* If anyone waited for this message to get flushed out, wake
241 * never learn that the RDMA failed. */ 324 * them up now */
242 if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) { 325 rds_message_unmapped(rm);
243 struct rds_message *rm;
244
245 rm = rds_send_get_message(conn, send->s_op);
246 if (rm) {
247 if (rm->m_rdma_op)
248 rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
249 rds_ib_send_rdma_complete(rm, wc.status);
250 rds_message_put(rm);
251 } 326 }
327 rds_message_put(rm);
328 send->s_op = NULL;
252 } 329 }
253 330
254 oldest = (oldest + 1) % ic->i_send_ring.w_nr; 331 oldest = (oldest + 1) % ic->i_send_ring.w_nr;
255 } 332 }
256 333
257 rds_ib_ring_free(&ic->i_send_ring, completed); 334 rds_ib_ring_free(&ic->i_send_ring, completed);
335 rds_ib_sub_signaled(ic, nr_sig);
336 nr_sig = 0;
258 337
259 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) || 338 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
260 test_bit(0, &conn->c_map_queued)) 339 test_bit(0, &conn->c_map_queued))
@@ -262,10 +341,10 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
262 341
263 /* We expect errors as the qp is drained during shutdown */ 342 /* We expect errors as the qp is drained during shutdown */
264 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) { 343 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
265 rds_ib_conn_error(conn, 344 rds_ib_conn_error(conn, "send completion on %pI4 had status "
266 "send completion on %pI4 " 345 "%u (%s), disconnecting and reconnecting\n",
267 "had status %u, disconnecting and reconnecting\n", 346 &conn->c_faddr, wc.status,
268 &conn->c_faddr, wc.status); 347 rds_ib_wc_status_str(wc.status));
269 } 348 }
270 } 349 }
271} 350}
@@ -294,7 +373,7 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
294 * credits (see rds_ib_send_add_credits below). 373 * credits (see rds_ib_send_add_credits below).
295 * 374 *
296 * The RDS send code is essentially single-threaded; rds_send_xmit 375 * The RDS send code is essentially single-threaded; rds_send_xmit
297 * grabs c_send_lock to ensure exclusive access to the send ring. 376 * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
298 * However, the ACK sending code is independent and can race with 377 * However, the ACK sending code is independent and can race with
299 * message SENDs. 378 * message SENDs.
300 * 379 *
@@ -413,40 +492,21 @@ void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
413 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); 492 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
414} 493}
415 494
416static inline void 495static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
417rds_ib_xmit_populate_wr(struct rds_ib_connection *ic, 496 struct rds_ib_send_work *send,
418 struct rds_ib_send_work *send, unsigned int pos, 497 bool notify)
419 unsigned long buffer, unsigned int length,
420 int send_flags)
421{ 498{
422 struct ib_sge *sge; 499 /*
423 500 * We want to delay signaling completions just enough to get
424 WARN_ON(pos != send - ic->i_sends); 501 * the batching benefits but not so much that we create dead time
425 502 * on the wire.
426 send->s_wr.send_flags = send_flags; 503 */
427 send->s_wr.opcode = IB_WR_SEND; 504 if (ic->i_unsignaled_wrs-- == 0 || notify) {
428 send->s_wr.num_sge = 2; 505 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
429 send->s_wr.next = NULL; 506 send->s_wr.send_flags |= IB_SEND_SIGNALED;
430 send->s_queued = jiffies; 507 return 1;
431 send->s_op = NULL;
432
433 if (length != 0) {
434 sge = rds_ib_data_sge(ic, send->s_sge);
435 sge->addr = buffer;
436 sge->length = length;
437 sge->lkey = ic->i_mr->lkey;
438
439 sge = rds_ib_header_sge(ic, send->s_sge);
440 } else {
441 /* We're sending a packet with no payload. There is only
442 * one SGE */
443 send->s_wr.num_sge = 1;
444 sge = &send->s_sge[0];
445 } 508 }
446 509 return 0;
447 sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
448 sge->length = sizeof(struct rds_header);
449 sge->lkey = ic->i_mr->lkey;
450} 510}
451 511
452/* 512/*
@@ -475,13 +535,14 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
475 u32 pos; 535 u32 pos;
476 u32 i; 536 u32 i;
477 u32 work_alloc; 537 u32 work_alloc;
478 u32 credit_alloc; 538 u32 credit_alloc = 0;
479 u32 posted; 539 u32 posted;
480 u32 adv_credits = 0; 540 u32 adv_credits = 0;
481 int send_flags = 0; 541 int send_flags = 0;
482 int sent; 542 int bytes_sent = 0;
483 int ret; 543 int ret;
484 int flow_controlled = 0; 544 int flow_controlled = 0;
545 int nr_sig = 0;
485 546
486 BUG_ON(off % RDS_FRAG_SIZE); 547 BUG_ON(off % RDS_FRAG_SIZE);
487 BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header)); 548 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,
507 goto out; 568 goto out;
508 } 569 }
509 570
510 credit_alloc = work_alloc;
511 if (ic->i_flowctl) { 571 if (ic->i_flowctl) {
512 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT); 572 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
513 adv_credits += posted; 573 adv_credits += posted;
514 if (credit_alloc < work_alloc) { 574 if (credit_alloc < work_alloc) {
515 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc); 575 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
516 work_alloc = credit_alloc; 576 work_alloc = credit_alloc;
517 flow_controlled++; 577 flow_controlled = 1;
518 } 578 }
519 if (work_alloc == 0) { 579 if (work_alloc == 0) {
520 set_bit(RDS_LL_SEND_FULL, &conn->c_flags); 580 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,
525 } 585 }
526 586
527 /* map the message the first time we see it */ 587 /* map the message the first time we see it */
528 if (ic->i_rm == NULL) { 588 if (!ic->i_data_op) {
529 /* 589 if (rm->data.op_nents) {
530 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n", 590 rm->data.op_count = ib_dma_map_sg(dev,
531 be16_to_cpu(rm->m_inc.i_hdr.h_dport), 591 rm->data.op_sg,
532 rm->m_inc.i_hdr.h_flags, 592 rm->data.op_nents,
533 be32_to_cpu(rm->m_inc.i_hdr.h_len)); 593 DMA_TO_DEVICE);
534 */ 594 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
535 if (rm->m_nents) { 595 if (rm->data.op_count == 0) {
536 rm->m_count = ib_dma_map_sg(dev,
537 rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
538 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
539 if (rm->m_count == 0) {
540 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); 596 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
541 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 597 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
542 ret = -ENOMEM; /* XXX ? */ 598 ret = -ENOMEM; /* XXX ? */
543 goto out; 599 goto out;
544 } 600 }
545 } else { 601 } else {
546 rm->m_count = 0; 602 rm->data.op_count = 0;
547 } 603 }
548 604
549 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
550 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
551 rds_message_addref(rm); 605 rds_message_addref(rm);
552 ic->i_rm = rm; 606 ic->i_data_op = &rm->data;
553 607
554 /* Finalize the header */ 608 /* Finalize the header */
555 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags)) 609 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,
559 613
560 /* If it has a RDMA op, tell the peer we did it. This is 614 /* If it has a RDMA op, tell the peer we did it. This is
561 * used by the peer to release use-once RDMA MRs. */ 615 * used by the peer to release use-once RDMA MRs. */
562 if (rm->m_rdma_op) { 616 if (rm->rdma.op_active) {
563 struct rds_ext_header_rdma ext_hdr; 617 struct rds_ext_header_rdma ext_hdr;
564 618
565 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key); 619 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
566 rds_message_add_extension(&rm->m_inc.i_hdr, 620 rds_message_add_extension(&rm->m_inc.i_hdr,
567 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr)); 621 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
568 } 622 }
@@ -582,99 +636,77 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
582 /* 636 /*
583 * Update adv_credits since we reset the ACK_REQUIRED bit. 637 * Update adv_credits since we reset the ACK_REQUIRED bit.
584 */ 638 */
585 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits); 639 if (ic->i_flowctl) {
586 adv_credits += posted; 640 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
587 BUG_ON(adv_credits > 255); 641 adv_credits += posted;
642 BUG_ON(adv_credits > 255);
643 }
588 } 644 }
589 645
590 send = &ic->i_sends[pos];
591 first = send;
592 prev = NULL;
593 scat = &rm->m_sg[sg];
594 sent = 0;
595 i = 0;
596
597 /* Sometimes you want to put a fence between an RDMA 646 /* Sometimes you want to put a fence between an RDMA
598 * READ and the following SEND. 647 * READ and the following SEND.
599 * We could either do this all the time 648 * We could either do this all the time
600 * or when requested by the user. Right now, we let 649 * or when requested by the user. Right now, we let
601 * the application choose. 650 * the application choose.
602 */ 651 */
603 if (rm->m_rdma_op && rm->m_rdma_op->r_fence) 652 if (rm->rdma.op_active && rm->rdma.op_fence)
604 send_flags = IB_SEND_FENCE; 653 send_flags = IB_SEND_FENCE;
605 654
606 /* 655 /* Each frag gets a header. Msgs may be 0 bytes */
607 * We could be copying the header into the unused tail of the page. 656 send = &ic->i_sends[pos];
608 * That would need to be changed in the future when those pages might 657 first = send;
609 * be mapped userspace pages or page cache pages. So instead we always 658 prev = NULL;
610 * use a second sge and our long-lived ring of mapped headers. We send 659 scat = &ic->i_data_op->op_sg[sg];
611 * the header after the data so that the data payload can be aligned on 660 i = 0;
612 * the receiver. 661 do {
613 */ 662 unsigned int len = 0;
614 663
615 /* handle a 0-len message */ 664 /* Set up the header */
616 if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) { 665 send->s_wr.send_flags = send_flags;
617 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags); 666 send->s_wr.opcode = IB_WR_SEND;
618 goto add_header; 667 send->s_wr.num_sge = 1;
619 } 668 send->s_wr.next = NULL;
669 send->s_queued = jiffies;
670 send->s_op = NULL;
620 671
621 /* if there's data reference it with a chain of work reqs */ 672 send->s_sge[0].addr = ic->i_send_hdrs_dma
622 for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) { 673 + (pos * sizeof(struct rds_header));
623 unsigned int len; 674 send->s_sge[0].length = sizeof(struct rds_header);
624 675
625 send = &ic->i_sends[pos]; 676 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
626 677
627 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off); 678 /* Set up the data, if present */
628 rds_ib_xmit_populate_wr(ic, send, pos, 679 if (i < work_alloc
629 ib_sg_dma_address(dev, scat) + off, len, 680 && scat != &rm->data.op_sg[rm->data.op_count]) {
630 send_flags); 681 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
682 send->s_wr.num_sge = 2;
631 683
632 /* 684 send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
633 * We want to delay signaling completions just enough to get 685 send->s_sge[1].length = len;
634 * the batching benefits but not so much that we create dead time
635 * on the wire.
636 */
637 if (ic->i_unsignaled_wrs-- == 0) {
638 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
639 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
640 }
641 686
642 ic->i_unsignaled_bytes -= len; 687 bytes_sent += len;
643 if (ic->i_unsignaled_bytes <= 0) { 688 off += len;
644 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes; 689 if (off == ib_sg_dma_len(dev, scat)) {
645 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED; 690 scat++;
691 off = 0;
692 }
646 } 693 }
647 694
695 rds_ib_set_wr_signal_state(ic, send, 0);
696
648 /* 697 /*
649 * Always signal the last one if we're stopping due to flow control. 698 * Always signal the last one if we're stopping due to flow control.
650 */ 699 */
651 if (flow_controlled && i == (work_alloc-1)) 700 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
652 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED; 701 send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
653 702
703 if (send->s_wr.send_flags & IB_SEND_SIGNALED)
704 nr_sig++;
705
654 rdsdebug("send %p wr %p num_sge %u next %p\n", send, 706 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
655 &send->s_wr, send->s_wr.num_sge, send->s_wr.next); 707 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
656 708
657 sent += len; 709 if (ic->i_flowctl && adv_credits) {
658 off += len;
659 if (off == ib_sg_dma_len(dev, scat)) {
660 scat++;
661 off = 0;
662 }
663
664add_header:
665 /* Tack on the header after the data. The header SGE should already
666 * have been set up to point to the right header buffer. */
667 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
668
669 if (0) {
670 struct rds_header *hdr = &ic->i_send_hdrs[pos];
671
672 printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
673 be16_to_cpu(hdr->h_dport),
674 hdr->h_flags,
675 be32_to_cpu(hdr->h_len));
676 }
677 if (adv_credits) {
678 struct rds_header *hdr = &ic->i_send_hdrs[pos]; 710 struct rds_header *hdr = &ic->i_send_hdrs[pos];
679 711
680 /* add credit and redo the header checksum */ 712 /* add credit and redo the header checksum */
@@ -689,20 +721,25 @@ add_header:
689 prev = send; 721 prev = send;
690 722
691 pos = (pos + 1) % ic->i_send_ring.w_nr; 723 pos = (pos + 1) % ic->i_send_ring.w_nr;
692 } 724 send = &ic->i_sends[pos];
725 i++;
726
727 } while (i < work_alloc
728 && scat != &rm->data.op_sg[rm->data.op_count]);
693 729
694 /* Account the RDS header in the number of bytes we sent, but just once. 730 /* Account the RDS header in the number of bytes we sent, but just once.
695 * The caller has no concept of fragmentation. */ 731 * The caller has no concept of fragmentation. */
696 if (hdr_off == 0) 732 if (hdr_off == 0)
697 sent += sizeof(struct rds_header); 733 bytes_sent += sizeof(struct rds_header);
698 734
699 /* if we finished the message then send completion owns it */ 735 /* if we finished the message then send completion owns it */
700 if (scat == &rm->m_sg[rm->m_count]) { 736 if (scat == &rm->data.op_sg[rm->data.op_count]) {
701 prev->s_rm = ic->i_rm; 737 prev->s_op = ic->i_data_op;
702 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED; 738 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
703 ic->i_rm = NULL; 739 ic->i_data_op = NULL;
704 } 740 }
705 741
742 /* Put back wrs & credits we didn't use */
706 if (i < work_alloc) { 743 if (i < work_alloc) {
707 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); 744 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
708 work_alloc = i; 745 work_alloc = i;
@@ -710,6 +747,9 @@ add_header:
710 if (ic->i_flowctl && i < credit_alloc) 747 if (ic->i_flowctl && i < credit_alloc)
711 rds_ib_send_add_credits(conn, credit_alloc - i); 748 rds_ib_send_add_credits(conn, credit_alloc - i);
712 749
750 if (nr_sig)
751 atomic_add(nr_sig, &ic->i_signaled_sends);
752
713 /* XXX need to worry about failed_wr and partial sends. */ 753 /* XXX need to worry about failed_wr and partial sends. */
714 failed_wr = &first->s_wr; 754 failed_wr = &first->s_wr;
715 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr); 755 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
@@ -720,32 +760,127 @@ add_header:
720 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 " 760 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
721 "returned %d\n", &conn->c_faddr, ret); 761 "returned %d\n", &conn->c_faddr, ret);
722 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 762 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
723 if (prev->s_rm) { 763 rds_ib_sub_signaled(ic, nr_sig);
724 ic->i_rm = prev->s_rm; 764 if (prev->s_op) {
725 prev->s_rm = NULL; 765 ic->i_data_op = prev->s_op;
766 prev->s_op = NULL;
726 } 767 }
727 768
728 rds_ib_conn_error(ic->conn, "ib_post_send failed\n"); 769 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
729 goto out; 770 goto out;
730 } 771 }
731 772
732 ret = sent; 773 ret = bytes_sent;
733out: 774out:
734 BUG_ON(adv_credits); 775 BUG_ON(adv_credits);
735 return ret; 776 return ret;
736} 777}
737 778
738int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op) 779/*
780 * Issue atomic operation.
781 * A simplified version of the rdma case, we always map 1 SG, and
782 * only 8 bytes, for the return value from the atomic operation.
783 */
784int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
785{
786 struct rds_ib_connection *ic = conn->c_transport_data;
787 struct rds_ib_send_work *send = NULL;
788 struct ib_send_wr *failed_wr;
789 struct rds_ib_device *rds_ibdev;
790 u32 pos;
791 u32 work_alloc;
792 int ret;
793 int nr_sig = 0;
794
795 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
796
797 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
798 if (work_alloc != 1) {
799 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
800 rds_ib_stats_inc(s_ib_tx_ring_full);
801 ret = -ENOMEM;
802 goto out;
803 }
804
805 /* address of send request in ring */
806 send = &ic->i_sends[pos];
807 send->s_queued = jiffies;
808
809 if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
810 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
811 send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
812 send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
813 send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
814 send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
815 } else { /* FADD */
816 send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
817 send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
818 send->s_wr.wr.atomic.swap = 0;
819 send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
820 send->s_wr.wr.atomic.swap_mask = 0;
821 }
822 nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
823 send->s_wr.num_sge = 1;
824 send->s_wr.next = NULL;
825 send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
826 send->s_wr.wr.atomic.rkey = op->op_rkey;
827 send->s_op = op;
828 rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
829
830 /* map 8 byte retval buffer to the device */
831 ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
832 rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
833 if (ret != 1) {
834 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
835 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
836 ret = -ENOMEM; /* XXX ? */
837 goto out;
838 }
839
840 /* Convert our struct scatterlist to struct ib_sge */
841 send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
842 send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
843 send->s_sge[0].lkey = ic->i_mr->lkey;
844
845 rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
846 send->s_sge[0].addr, send->s_sge[0].length);
847
848 if (nr_sig)
849 atomic_add(nr_sig, &ic->i_signaled_sends);
850
851 failed_wr = &send->s_wr;
852 ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
853 rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
854 send, &send->s_wr, ret, failed_wr);
855 BUG_ON(failed_wr != &send->s_wr);
856 if (ret) {
857 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
858 "returned %d\n", &conn->c_faddr, ret);
859 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
860 rds_ib_sub_signaled(ic, nr_sig);
861 goto out;
862 }
863
864 if (unlikely(failed_wr != &send->s_wr)) {
865 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
866 BUG_ON(failed_wr != &send->s_wr);
867 }
868
869out:
870 return ret;
871}
872
873int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
739{ 874{
740 struct rds_ib_connection *ic = conn->c_transport_data; 875 struct rds_ib_connection *ic = conn->c_transport_data;
741 struct rds_ib_send_work *send = NULL; 876 struct rds_ib_send_work *send = NULL;
742 struct rds_ib_send_work *first; 877 struct rds_ib_send_work *first;
743 struct rds_ib_send_work *prev; 878 struct rds_ib_send_work *prev;
744 struct ib_send_wr *failed_wr; 879 struct ib_send_wr *failed_wr;
745 struct rds_ib_device *rds_ibdev;
746 struct scatterlist *scat; 880 struct scatterlist *scat;
747 unsigned long len; 881 unsigned long len;
748 u64 remote_addr = op->r_remote_addr; 882 u64 remote_addr = op->op_remote_addr;
883 u32 max_sge = ic->rds_ibdev->max_sge;
749 u32 pos; 884 u32 pos;
750 u32 work_alloc; 885 u32 work_alloc;
751 u32 i; 886 u32 i;
@@ -753,29 +888,28 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
753 int sent; 888 int sent;
754 int ret; 889 int ret;
755 int num_sge; 890 int num_sge;
756 891 int nr_sig = 0;
757 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); 892
758 893 /* map the op the first time we see it */
759 /* map the message the first time we see it */ 894 if (!op->op_mapped) {
760 if (!op->r_mapped) { 895 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
761 op->r_count = ib_dma_map_sg(ic->i_cm_id->device, 896 op->op_sg, op->op_nents, (op->op_write) ?
762 op->r_sg, op->r_nents, (op->r_write) ? 897 DMA_TO_DEVICE : DMA_FROM_DEVICE);
763 DMA_TO_DEVICE : DMA_FROM_DEVICE); 898 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
764 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count); 899 if (op->op_count == 0) {
765 if (op->r_count == 0) {
766 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); 900 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
767 ret = -ENOMEM; /* XXX ? */ 901 ret = -ENOMEM; /* XXX ? */
768 goto out; 902 goto out;
769 } 903 }
770 904
771 op->r_mapped = 1; 905 op->op_mapped = 1;
772 } 906 }
773 907
774 /* 908 /*
775 * Instead of knowing how to return a partial rdma read/write we insist that there 909 * Instead of knowing how to return a partial rdma read/write we insist that there
776 * be enough work requests to send the entire message. 910 * be enough work requests to send the entire message.
777 */ 911 */
778 i = ceil(op->r_count, rds_ibdev->max_sge); 912 i = ceil(op->op_count, max_sge);
779 913
780 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos); 914 work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
781 if (work_alloc != i) { 915 if (work_alloc != i) {
@@ -788,30 +922,24 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
788 send = &ic->i_sends[pos]; 922 send = &ic->i_sends[pos];
789 first = send; 923 first = send;
790 prev = NULL; 924 prev = NULL;
791 scat = &op->r_sg[0]; 925 scat = &op->op_sg[0];
792 sent = 0; 926 sent = 0;
793 num_sge = op->r_count; 927 num_sge = op->op_count;
794 928
795 for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) { 929 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
796 send->s_wr.send_flags = 0; 930 send->s_wr.send_flags = 0;
797 send->s_queued = jiffies; 931 send->s_queued = jiffies;
798 /* 932 send->s_op = NULL;
799 * We want to delay signaling completions just enough to get 933
800 * the batching benefits but not so much that we create dead time on the wire. 934 nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
801 */
802 if (ic->i_unsignaled_wrs-- == 0) {
803 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
804 send->s_wr.send_flags = IB_SEND_SIGNALED;
805 }
806 935
807 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ; 936 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
808 send->s_wr.wr.rdma.remote_addr = remote_addr; 937 send->s_wr.wr.rdma.remote_addr = remote_addr;
809 send->s_wr.wr.rdma.rkey = op->r_key; 938 send->s_wr.wr.rdma.rkey = op->op_rkey;
810 send->s_op = op;
811 939
812 if (num_sge > rds_ibdev->max_sge) { 940 if (num_sge > max_sge) {
813 send->s_wr.num_sge = rds_ibdev->max_sge; 941 send->s_wr.num_sge = max_sge;
814 num_sge -= rds_ibdev->max_sge; 942 num_sge -= max_sge;
815 } else { 943 } else {
816 send->s_wr.num_sge = num_sge; 944 send->s_wr.num_sge = num_sge;
817 } 945 }
@@ -821,7 +949,7 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
821 if (prev) 949 if (prev)
822 prev->s_wr.next = &send->s_wr; 950 prev->s_wr.next = &send->s_wr;
823 951
824 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) { 952 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
825 len = ib_sg_dma_len(ic->i_cm_id->device, scat); 953 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
826 send->s_sge[j].addr = 954 send->s_sge[j].addr =
827 ib_sg_dma_address(ic->i_cm_id->device, scat); 955 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)
843 send = ic->i_sends; 971 send = ic->i_sends;
844 } 972 }
845 973
846 /* if we finished the message then send completion owns it */ 974 /* give a reference to the last op */
847 if (scat == &op->r_sg[op->r_count]) 975 if (scat == &op->op_sg[op->op_count]) {
848 prev->s_wr.send_flags = IB_SEND_SIGNALED; 976 prev->s_op = op;
977 rds_message_addref(container_of(op, struct rds_message, rdma));
978 }
849 979
850 if (i < work_alloc) { 980 if (i < work_alloc) {
851 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); 981 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
852 work_alloc = i; 982 work_alloc = i;
853 } 983 }
854 984
985 if (nr_sig)
986 atomic_add(nr_sig, &ic->i_signaled_sends);
987
855 failed_wr = &first->s_wr; 988 failed_wr = &first->s_wr;
856 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr); 989 ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
857 rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic, 990 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)
861 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 " 994 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
862 "returned %d\n", &conn->c_faddr, ret); 995 "returned %d\n", &conn->c_faddr, ret);
863 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); 996 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
997 rds_ib_sub_signaled(ic, nr_sig);
864 goto out; 998 goto out;
865 } 999 }
866 1000
diff --git a/net/rds/ib_stats.c b/net/rds/ib_stats.c
index d2c904dd6fbc..2d5965d6e97c 100644
--- a/net/rds/ib_stats.c
+++ b/net/rds/ib_stats.c
@@ -67,6 +67,8 @@ static const char *const rds_ib_stat_names[] = {
67 "ib_rdma_mr_pool_flush", 67 "ib_rdma_mr_pool_flush",
68 "ib_rdma_mr_pool_wait", 68 "ib_rdma_mr_pool_wait",
69 "ib_rdma_mr_pool_depleted", 69 "ib_rdma_mr_pool_depleted",
70 "ib_atomic_cswp",
71 "ib_atomic_fadd",
70}; 72};
71 73
72unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, 74unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
diff --git a/net/rds/ib_sysctl.c b/net/rds/ib_sysctl.c
index 03f01cb4e0fe..fc3da37220fd 100644
--- a/net/rds/ib_sysctl.c
+++ b/net/rds/ib_sysctl.c
@@ -49,10 +49,6 @@ unsigned long rds_ib_sysctl_max_unsig_wrs = 16;
49static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1; 49static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1;
50static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64; 50static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64;
51 51
52unsigned long rds_ib_sysctl_max_unsig_bytes = (16 << 20);
53static unsigned long rds_ib_sysctl_max_unsig_bytes_min = 1;
54static unsigned long rds_ib_sysctl_max_unsig_bytes_max = ~0UL;
55
56/* 52/*
57 * This sysctl does nothing. 53 * This sysctl does nothing.
58 * 54 *
@@ -94,15 +90,6 @@ ctl_table rds_ib_sysctl_table[] = {
94 .extra2 = &rds_ib_sysctl_max_unsig_wr_max, 90 .extra2 = &rds_ib_sysctl_max_unsig_wr_max,
95 }, 91 },
96 { 92 {
97 .procname = "max_unsignaled_bytes",
98 .data = &rds_ib_sysctl_max_unsig_bytes,
99 .maxlen = sizeof(unsigned long),
100 .mode = 0644,
101 .proc_handler = proc_doulongvec_minmax,
102 .extra1 = &rds_ib_sysctl_max_unsig_bytes_min,
103 .extra2 = &rds_ib_sysctl_max_unsig_bytes_max,
104 },
105 {
106 .procname = "max_recv_allocation", 93 .procname = "max_recv_allocation",
107 .data = &rds_ib_sysctl_max_recv_allocation, 94 .data = &rds_ib_sysctl_max_recv_allocation,
108 .maxlen = sizeof(unsigned long), 95 .maxlen = sizeof(unsigned long),
@@ -132,10 +119,10 @@ void rds_ib_sysctl_exit(void)
132 unregister_sysctl_table(rds_ib_sysctl_hdr); 119 unregister_sysctl_table(rds_ib_sysctl_hdr);
133} 120}
134 121
135int __init rds_ib_sysctl_init(void) 122int rds_ib_sysctl_init(void)
136{ 123{
137 rds_ib_sysctl_hdr = register_sysctl_paths(rds_ib_sysctl_path, rds_ib_sysctl_table); 124 rds_ib_sysctl_hdr = register_sysctl_paths(rds_ib_sysctl_path, rds_ib_sysctl_table);
138 if (rds_ib_sysctl_hdr == NULL) 125 if (!rds_ib_sysctl_hdr)
139 return -ENOMEM; 126 return -ENOMEM;
140 return 0; 127 return 0;
141} 128}
diff --git a/net/rds/info.c b/net/rds/info.c
index c45c4173a44d..4fdf1b6e84ff 100644
--- a/net/rds/info.c
+++ b/net/rds/info.c
@@ -76,7 +76,7 @@ void rds_info_register_func(int optname, rds_info_func func)
76 BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST); 76 BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
77 77
78 spin_lock(&rds_info_lock); 78 spin_lock(&rds_info_lock);
79 BUG_ON(rds_info_funcs[offset] != NULL); 79 BUG_ON(rds_info_funcs[offset]);
80 rds_info_funcs[offset] = func; 80 rds_info_funcs[offset] = func;
81 spin_unlock(&rds_info_lock); 81 spin_unlock(&rds_info_lock);
82} 82}
@@ -102,7 +102,7 @@ EXPORT_SYMBOL_GPL(rds_info_deregister_func);
102 */ 102 */
103void rds_info_iter_unmap(struct rds_info_iterator *iter) 103void rds_info_iter_unmap(struct rds_info_iterator *iter)
104{ 104{
105 if (iter->addr != NULL) { 105 if (iter->addr) {
106 kunmap_atomic(iter->addr, KM_USER0); 106 kunmap_atomic(iter->addr, KM_USER0);
107 iter->addr = NULL; 107 iter->addr = NULL;
108 } 108 }
@@ -117,7 +117,7 @@ void rds_info_copy(struct rds_info_iterator *iter, void *data,
117 unsigned long this; 117 unsigned long this;
118 118
119 while (bytes) { 119 while (bytes) {
120 if (iter->addr == NULL) 120 if (!iter->addr)
121 iter->addr = kmap_atomic(*iter->pages, KM_USER0); 121 iter->addr = kmap_atomic(*iter->pages, KM_USER0);
122 122
123 this = min(bytes, PAGE_SIZE - iter->offset); 123 this = min(bytes, PAGE_SIZE - iter->offset);
@@ -188,7 +188,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval,
188 >> PAGE_SHIFT; 188 >> PAGE_SHIFT;
189 189
190 pages = kmalloc(nr_pages * sizeof(struct page *), GFP_KERNEL); 190 pages = kmalloc(nr_pages * sizeof(struct page *), GFP_KERNEL);
191 if (pages == NULL) { 191 if (!pages) {
192 ret = -ENOMEM; 192 ret = -ENOMEM;
193 goto out; 193 goto out;
194 } 194 }
@@ -206,7 +206,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval,
206 206
207call_func: 207call_func:
208 func = rds_info_funcs[optname - RDS_INFO_FIRST]; 208 func = rds_info_funcs[optname - RDS_INFO_FIRST];
209 if (func == NULL) { 209 if (!func) {
210 ret = -ENOPROTOOPT; 210 ret = -ENOPROTOOPT;
211 goto out; 211 goto out;
212 } 212 }
@@ -234,7 +234,7 @@ call_func:
234 ret = -EFAULT; 234 ret = -EFAULT;
235 235
236out: 236out:
237 for (i = 0; pages != NULL && i < nr_pages; i++) 237 for (i = 0; pages && i < nr_pages; i++)
238 put_page(pages[i]); 238 put_page(pages[i]);
239 kfree(pages); 239 kfree(pages);
240 240
diff --git a/net/rds/iw.c b/net/rds/iw.c
index c8f3d3525cb9..56808cac0fc7 100644
--- a/net/rds/iw.c
+++ b/net/rds/iw.c
@@ -264,7 +264,6 @@ struct rds_transport rds_iw_transport = {
264 .laddr_check = rds_iw_laddr_check, 264 .laddr_check = rds_iw_laddr_check,
265 .xmit_complete = rds_iw_xmit_complete, 265 .xmit_complete = rds_iw_xmit_complete,
266 .xmit = rds_iw_xmit, 266 .xmit = rds_iw_xmit,
267 .xmit_cong_map = NULL,
268 .xmit_rdma = rds_iw_xmit_rdma, 267 .xmit_rdma = rds_iw_xmit_rdma,
269 .recv = rds_iw_recv, 268 .recv = rds_iw_recv,
270 .conn_alloc = rds_iw_conn_alloc, 269 .conn_alloc = rds_iw_conn_alloc,
@@ -272,7 +271,6 @@ struct rds_transport rds_iw_transport = {
272 .conn_connect = rds_iw_conn_connect, 271 .conn_connect = rds_iw_conn_connect,
273 .conn_shutdown = rds_iw_conn_shutdown, 272 .conn_shutdown = rds_iw_conn_shutdown,
274 .inc_copy_to_user = rds_iw_inc_copy_to_user, 273 .inc_copy_to_user = rds_iw_inc_copy_to_user,
275 .inc_purge = rds_iw_inc_purge,
276 .inc_free = rds_iw_inc_free, 274 .inc_free = rds_iw_inc_free,
277 .cm_initiate_connect = rds_iw_cm_initiate_connect, 275 .cm_initiate_connect = rds_iw_cm_initiate_connect,
278 .cm_handle_connect = rds_iw_cm_handle_connect, 276 .cm_handle_connect = rds_iw_cm_handle_connect,
@@ -289,7 +287,7 @@ struct rds_transport rds_iw_transport = {
289 .t_prefer_loopback = 1, 287 .t_prefer_loopback = 1,
290}; 288};
291 289
292int __init rds_iw_init(void) 290int rds_iw_init(void)
293{ 291{
294 int ret; 292 int ret;
295 293
diff --git a/net/rds/iw.h b/net/rds/iw.h
index eef2f0c28476..543e665fafe3 100644
--- a/net/rds/iw.h
+++ b/net/rds/iw.h
@@ -70,7 +70,7 @@ struct rds_iw_send_work {
70 struct rds_message *s_rm; 70 struct rds_message *s_rm;
71 71
72 /* We should really put these into a union: */ 72 /* We should really put these into a union: */
73 struct rds_rdma_op *s_op; 73 struct rm_rdma_op *s_op;
74 struct rds_iw_mapping *s_mapping; 74 struct rds_iw_mapping *s_mapping;
75 struct ib_mr *s_mr; 75 struct ib_mr *s_mr;
76 struct ib_fast_reg_page_list *s_page_list; 76 struct ib_fast_reg_page_list *s_page_list;
@@ -284,7 +284,7 @@ void rds_iw_conn_free(void *arg);
284int rds_iw_conn_connect(struct rds_connection *conn); 284int rds_iw_conn_connect(struct rds_connection *conn);
285void rds_iw_conn_shutdown(struct rds_connection *conn); 285void rds_iw_conn_shutdown(struct rds_connection *conn);
286void rds_iw_state_change(struct sock *sk); 286void rds_iw_state_change(struct sock *sk);
287int __init rds_iw_listen_init(void); 287int rds_iw_listen_init(void);
288void rds_iw_listen_stop(void); 288void rds_iw_listen_stop(void);
289void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...); 289void __rds_iw_conn_error(struct rds_connection *conn, const char *, ...);
290int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id, 290int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
@@ -321,12 +321,11 @@ void rds_iw_flush_mrs(void);
321void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id); 321void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
322 322
323/* ib_recv.c */ 323/* ib_recv.c */
324int __init rds_iw_recv_init(void); 324int rds_iw_recv_init(void);
325void rds_iw_recv_exit(void); 325void rds_iw_recv_exit(void);
326int rds_iw_recv(struct rds_connection *conn); 326int rds_iw_recv(struct rds_connection *conn);
327int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, 327int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
328 gfp_t page_gfp, int prefill); 328 gfp_t page_gfp, int prefill);
329void rds_iw_inc_purge(struct rds_incoming *inc);
330void rds_iw_inc_free(struct rds_incoming *inc); 329void rds_iw_inc_free(struct rds_incoming *inc);
331int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov, 330int rds_iw_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
332 size_t size); 331 size_t size);
@@ -358,7 +357,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
358void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context); 357void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context);
359void rds_iw_send_init_ring(struct rds_iw_connection *ic); 358void rds_iw_send_init_ring(struct rds_iw_connection *ic);
360void rds_iw_send_clear_ring(struct rds_iw_connection *ic); 359void rds_iw_send_clear_ring(struct rds_iw_connection *ic);
361int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op); 360int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
362void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits); 361void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits);
363void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted); 362void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted);
364int rds_iw_send_grab_credits(struct rds_iw_connection *ic, u32 wanted, 363int 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,
371 unsigned int avail); 370 unsigned int avail);
372 371
373/* ib_sysctl.c */ 372/* ib_sysctl.c */
374int __init rds_iw_sysctl_init(void); 373int rds_iw_sysctl_init(void);
375void rds_iw_sysctl_exit(void); 374void rds_iw_sysctl_exit(void);
376extern unsigned long rds_iw_sysctl_max_send_wr; 375extern unsigned long rds_iw_sysctl_max_send_wr;
377extern unsigned long rds_iw_sysctl_max_recv_wr; 376extern unsigned long rds_iw_sysctl_max_recv_wr;
diff --git a/net/rds/iw_cm.c b/net/rds/iw_cm.c
index b5dd6ac39be8..712cf2d1f28e 100644
--- a/net/rds/iw_cm.c
+++ b/net/rds/iw_cm.c
@@ -257,7 +257,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
257 * the rds_iwdev at all. 257 * the rds_iwdev at all.
258 */ 258 */
259 rds_iwdev = ib_get_client_data(dev, &rds_iw_client); 259 rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
260 if (rds_iwdev == NULL) { 260 if (!rds_iwdev) {
261 if (printk_ratelimit()) 261 if (printk_ratelimit())
262 printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n", 262 printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
263 dev->name); 263 dev->name);
@@ -292,7 +292,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
292 ic->i_send_ring.w_nr * 292 ic->i_send_ring.w_nr *
293 sizeof(struct rds_header), 293 sizeof(struct rds_header),
294 &ic->i_send_hdrs_dma, GFP_KERNEL); 294 &ic->i_send_hdrs_dma, GFP_KERNEL);
295 if (ic->i_send_hdrs == NULL) { 295 if (!ic->i_send_hdrs) {
296 ret = -ENOMEM; 296 ret = -ENOMEM;
297 rdsdebug("ib_dma_alloc_coherent send failed\n"); 297 rdsdebug("ib_dma_alloc_coherent send failed\n");
298 goto out; 298 goto out;
@@ -302,7 +302,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
302 ic->i_recv_ring.w_nr * 302 ic->i_recv_ring.w_nr *
303 sizeof(struct rds_header), 303 sizeof(struct rds_header),
304 &ic->i_recv_hdrs_dma, GFP_KERNEL); 304 &ic->i_recv_hdrs_dma, GFP_KERNEL);
305 if (ic->i_recv_hdrs == NULL) { 305 if (!ic->i_recv_hdrs) {
306 ret = -ENOMEM; 306 ret = -ENOMEM;
307 rdsdebug("ib_dma_alloc_coherent recv failed\n"); 307 rdsdebug("ib_dma_alloc_coherent recv failed\n");
308 goto out; 308 goto out;
@@ -310,14 +310,14 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
310 310
311 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), 311 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
312 &ic->i_ack_dma, GFP_KERNEL); 312 &ic->i_ack_dma, GFP_KERNEL);
313 if (ic->i_ack == NULL) { 313 if (!ic->i_ack) {
314 ret = -ENOMEM; 314 ret = -ENOMEM;
315 rdsdebug("ib_dma_alloc_coherent ack failed\n"); 315 rdsdebug("ib_dma_alloc_coherent ack failed\n");
316 goto out; 316 goto out;
317 } 317 }
318 318
319 ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work)); 319 ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
320 if (ic->i_sends == NULL) { 320 if (!ic->i_sends) {
321 ret = -ENOMEM; 321 ret = -ENOMEM;
322 rdsdebug("send allocation failed\n"); 322 rdsdebug("send allocation failed\n");
323 goto out; 323 goto out;
@@ -325,7 +325,7 @@ static int rds_iw_setup_qp(struct rds_connection *conn)
325 rds_iw_send_init_ring(ic); 325 rds_iw_send_init_ring(ic);
326 326
327 ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work)); 327 ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
328 if (ic->i_recvs == NULL) { 328 if (!ic->i_recvs) {
329 ret = -ENOMEM; 329 ret = -ENOMEM;
330 rdsdebug("recv allocation failed\n"); 330 rdsdebug("recv allocation failed\n");
331 goto out; 331 goto out;
@@ -696,7 +696,7 @@ int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
696 696
697 /* XXX too lazy? */ 697 /* XXX too lazy? */
698 ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL); 698 ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL);
699 if (ic == NULL) 699 if (!ic)
700 return -ENOMEM; 700 return -ENOMEM;
701 701
702 INIT_LIST_HEAD(&ic->iw_node); 702 INIT_LIST_HEAD(&ic->iw_node);
diff --git a/net/rds/iw_rdma.c b/net/rds/iw_rdma.c
index 13dc1862d862..4e152e2daa3d 100644
--- a/net/rds/iw_rdma.c
+++ b/net/rds/iw_rdma.c
@@ -34,7 +34,6 @@
34#include <linux/slab.h> 34#include <linux/slab.h>
35 35
36#include "rds.h" 36#include "rds.h"
37#include "rdma.h"
38#include "iw.h" 37#include "iw.h"
39 38
40 39
diff --git a/net/rds/iw_recv.c b/net/rds/iw_recv.c
index 3d479067d54d..5e57347f49ff 100644
--- a/net/rds/iw_recv.c
+++ b/net/rds/iw_recv.c
@@ -53,7 +53,7 @@ static void rds_iw_frag_drop_page(struct rds_page_frag *frag)
53static void rds_iw_frag_free(struct rds_page_frag *frag) 53static void rds_iw_frag_free(struct rds_page_frag *frag)
54{ 54{
55 rdsdebug("frag %p page %p\n", frag, frag->f_page); 55 rdsdebug("frag %p page %p\n", frag, frag->f_page);
56 BUG_ON(frag->f_page != NULL); 56 BUG_ON(frag->f_page);
57 kmem_cache_free(rds_iw_frag_slab, frag); 57 kmem_cache_free(rds_iw_frag_slab, frag);
58} 58}
59 59
@@ -143,14 +143,14 @@ static int rds_iw_recv_refill_one(struct rds_connection *conn,
143 struct ib_sge *sge; 143 struct ib_sge *sge;
144 int ret = -ENOMEM; 144 int ret = -ENOMEM;
145 145
146 if (recv->r_iwinc == NULL) { 146 if (!recv->r_iwinc) {
147 if (!atomic_add_unless(&rds_iw_allocation, 1, rds_iw_sysctl_max_recv_allocation)) { 147 if (!atomic_add_unless(&rds_iw_allocation, 1, rds_iw_sysctl_max_recv_allocation)) {
148 rds_iw_stats_inc(s_iw_rx_alloc_limit); 148 rds_iw_stats_inc(s_iw_rx_alloc_limit);
149 goto out; 149 goto out;
150 } 150 }
151 recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab, 151 recv->r_iwinc = kmem_cache_alloc(rds_iw_incoming_slab,
152 kptr_gfp); 152 kptr_gfp);
153 if (recv->r_iwinc == NULL) { 153 if (!recv->r_iwinc) {
154 atomic_dec(&rds_iw_allocation); 154 atomic_dec(&rds_iw_allocation);
155 goto out; 155 goto out;
156 } 156 }
@@ -158,17 +158,17 @@ static int rds_iw_recv_refill_one(struct rds_connection *conn,
158 rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr); 158 rds_inc_init(&recv->r_iwinc->ii_inc, conn, conn->c_faddr);
159 } 159 }
160 160
161 if (recv->r_frag == NULL) { 161 if (!recv->r_frag) {
162 recv->r_frag = kmem_cache_alloc(rds_iw_frag_slab, kptr_gfp); 162 recv->r_frag = kmem_cache_alloc(rds_iw_frag_slab, kptr_gfp);
163 if (recv->r_frag == NULL) 163 if (!recv->r_frag)
164 goto out; 164 goto out;
165 INIT_LIST_HEAD(&recv->r_frag->f_item); 165 INIT_LIST_HEAD(&recv->r_frag->f_item);
166 recv->r_frag->f_page = NULL; 166 recv->r_frag->f_page = NULL;
167 } 167 }
168 168
169 if (ic->i_frag.f_page == NULL) { 169 if (!ic->i_frag.f_page) {
170 ic->i_frag.f_page = alloc_page(page_gfp); 170 ic->i_frag.f_page = alloc_page(page_gfp);
171 if (ic->i_frag.f_page == NULL) 171 if (!ic->i_frag.f_page)
172 goto out; 172 goto out;
173 ic->i_frag.f_offset = 0; 173 ic->i_frag.f_offset = 0;
174 } 174 }
@@ -273,7 +273,7 @@ int rds_iw_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
273 return ret; 273 return ret;
274} 274}
275 275
276void rds_iw_inc_purge(struct rds_incoming *inc) 276static void rds_iw_inc_purge(struct rds_incoming *inc)
277{ 277{
278 struct rds_iw_incoming *iwinc; 278 struct rds_iw_incoming *iwinc;
279 struct rds_page_frag *frag; 279 struct rds_page_frag *frag;
@@ -716,7 +716,7 @@ static void rds_iw_process_recv(struct rds_connection *conn,
716 * into the inc and save the inc so we can hang upcoming fragments 716 * into the inc and save the inc so we can hang upcoming fragments
717 * off its list. 717 * off its list.
718 */ 718 */
719 if (iwinc == NULL) { 719 if (!iwinc) {
720 iwinc = recv->r_iwinc; 720 iwinc = recv->r_iwinc;
721 recv->r_iwinc = NULL; 721 recv->r_iwinc = NULL;
722 ic->i_iwinc = iwinc; 722 ic->i_iwinc = iwinc;
@@ -887,7 +887,7 @@ int rds_iw_recv(struct rds_connection *conn)
887 return ret; 887 return ret;
888} 888}
889 889
890int __init rds_iw_recv_init(void) 890int rds_iw_recv_init(void)
891{ 891{
892 struct sysinfo si; 892 struct sysinfo si;
893 int ret = -ENOMEM; 893 int ret = -ENOMEM;
@@ -899,13 +899,13 @@ int __init rds_iw_recv_init(void)
899 rds_iw_incoming_slab = kmem_cache_create("rds_iw_incoming", 899 rds_iw_incoming_slab = kmem_cache_create("rds_iw_incoming",
900 sizeof(struct rds_iw_incoming), 900 sizeof(struct rds_iw_incoming),
901 0, 0, NULL); 901 0, 0, NULL);
902 if (rds_iw_incoming_slab == NULL) 902 if (!rds_iw_incoming_slab)
903 goto out; 903 goto out;
904 904
905 rds_iw_frag_slab = kmem_cache_create("rds_iw_frag", 905 rds_iw_frag_slab = kmem_cache_create("rds_iw_frag",
906 sizeof(struct rds_page_frag), 906 sizeof(struct rds_page_frag),
907 0, 0, NULL); 907 0, 0, NULL);
908 if (rds_iw_frag_slab == NULL) 908 if (!rds_iw_frag_slab)
909 kmem_cache_destroy(rds_iw_incoming_slab); 909 kmem_cache_destroy(rds_iw_incoming_slab);
910 else 910 else
911 ret = 0; 911 ret = 0;
diff --git a/net/rds/iw_send.c b/net/rds/iw_send.c
index 52182ff7519e..6280ea020d4e 100644
--- a/net/rds/iw_send.c
+++ b/net/rds/iw_send.c
@@ -36,7 +36,6 @@
36#include <linux/dmapool.h> 36#include <linux/dmapool.h>
37 37
38#include "rds.h" 38#include "rds.h"
39#include "rdma.h"
40#include "iw.h" 39#include "iw.h"
41 40
42static void rds_iw_send_rdma_complete(struct rds_message *rm, 41static 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,
64} 63}
65 64
66static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic, 65static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic,
67 struct rds_rdma_op *op) 66 struct rm_rdma_op *op)
68{ 67{
69 if (op->r_mapped) { 68 if (op->op_mapped) {
70 ib_dma_unmap_sg(ic->i_cm_id->device, 69 ib_dma_unmap_sg(ic->i_cm_id->device,
71 op->r_sg, op->r_nents, 70 op->op_sg, op->op_nents,
72 op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE); 71 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
73 op->r_mapped = 0; 72 op->op_mapped = 0;
74 } 73 }
75} 74}
76 75
@@ -83,11 +82,11 @@ static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
83 rdsdebug("ic %p send %p rm %p\n", ic, send, rm); 82 rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
84 83
85 ib_dma_unmap_sg(ic->i_cm_id->device, 84 ib_dma_unmap_sg(ic->i_cm_id->device,
86 rm->m_sg, rm->m_nents, 85 rm->data.op_sg, rm->data.op_nents,
87 DMA_TO_DEVICE); 86 DMA_TO_DEVICE);
88 87
89 if (rm->m_rdma_op != NULL) { 88 if (rm->rdma.op_active) {
90 rds_iw_send_unmap_rdma(ic, rm->m_rdma_op); 89 rds_iw_send_unmap_rdma(ic, &rm->rdma);
91 90
92 /* If the user asked for a completion notification on this 91 /* If the user asked for a completion notification on this
93 * message, we can implement three different semantics: 92 * message, we can implement three different semantics:
@@ -111,10 +110,10 @@ static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
111 */ 110 */
112 rds_iw_send_rdma_complete(rm, wc_status); 111 rds_iw_send_rdma_complete(rm, wc_status);
113 112
114 if (rm->m_rdma_op->r_write) 113 if (rm->rdma.op_write)
115 rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes); 114 rds_stats_add(s_send_rdma_bytes, rm->rdma.op_bytes);
116 else 115 else
117 rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes); 116 rds_stats_add(s_recv_rdma_bytes, rm->rdma.op_bytes);
118 } 117 }
119 118
120 /* If anyone waited for this message to get flushed out, wake 119 /* 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,
556 } 555 }
557 556
558 /* map the message the first time we see it */ 557 /* map the message the first time we see it */
559 if (ic->i_rm == NULL) { 558 if (!ic->i_rm) {
560 /* 559 /*
561 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n", 560 printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
562 be16_to_cpu(rm->m_inc.i_hdr.h_dport), 561 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
563 rm->m_inc.i_hdr.h_flags, 562 rm->m_inc.i_hdr.h_flags,
564 be32_to_cpu(rm->m_inc.i_hdr.h_len)); 563 be32_to_cpu(rm->m_inc.i_hdr.h_len));
565 */ 564 */
566 if (rm->m_nents) { 565 if (rm->data.op_nents) {
567 rm->m_count = ib_dma_map_sg(dev, 566 rm->data.op_count = ib_dma_map_sg(dev,
568 rm->m_sg, rm->m_nents, DMA_TO_DEVICE); 567 rm->data.op_sg,
569 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count); 568 rm->data.op_nents,
570 if (rm->m_count == 0) { 569 DMA_TO_DEVICE);
570 rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
571 if (rm->data.op_count == 0) {
571 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure); 572 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
572 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc); 573 rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
573 ret = -ENOMEM; /* XXX ? */ 574 ret = -ENOMEM; /* XXX ? */
574 goto out; 575 goto out;
575 } 576 }
576 } else { 577 } else {
577 rm->m_count = 0; 578 rm->data.op_count = 0;
578 } 579 }
579 580
580 ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs; 581 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,
590 591
591 /* If it has a RDMA op, tell the peer we did it. This is 592 /* If it has a RDMA op, tell the peer we did it. This is
592 * used by the peer to release use-once RDMA MRs. */ 593 * used by the peer to release use-once RDMA MRs. */
593 if (rm->m_rdma_op) { 594 if (rm->rdma.op_active) {
594 struct rds_ext_header_rdma ext_hdr; 595 struct rds_ext_header_rdma ext_hdr;
595 596
596 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key); 597 ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
597 rds_message_add_extension(&rm->m_inc.i_hdr, 598 rds_message_add_extension(&rm->m_inc.i_hdr,
598 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr)); 599 RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
599 } 600 }
@@ -621,7 +622,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
621 send = &ic->i_sends[pos]; 622 send = &ic->i_sends[pos];
622 first = send; 623 first = send;
623 prev = NULL; 624 prev = NULL;
624 scat = &rm->m_sg[sg]; 625 scat = &rm->data.op_sg[sg];
625 sent = 0; 626 sent = 0;
626 i = 0; 627 i = 0;
627 628
@@ -631,7 +632,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
631 * or when requested by the user. Right now, we let 632 * or when requested by the user. Right now, we let
632 * the application choose. 633 * the application choose.
633 */ 634 */
634 if (rm->m_rdma_op && rm->m_rdma_op->r_fence) 635 if (rm->rdma.op_active && rm->rdma.op_fence)
635 send_flags = IB_SEND_FENCE; 636 send_flags = IB_SEND_FENCE;
636 637
637 /* 638 /*
@@ -650,7 +651,7 @@ int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
650 } 651 }
651 652
652 /* if there's data reference it with a chain of work reqs */ 653 /* if there's data reference it with a chain of work reqs */
653 for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) { 654 for (; i < work_alloc && scat != &rm->data.op_sg[rm->data.op_count]; i++) {
654 unsigned int len; 655 unsigned int len;
655 656
656 send = &ic->i_sends[pos]; 657 send = &ic->i_sends[pos];
@@ -728,7 +729,7 @@ add_header:
728 sent += sizeof(struct rds_header); 729 sent += sizeof(struct rds_header);
729 730
730 /* if we finished the message then send completion owns it */ 731 /* if we finished the message then send completion owns it */
731 if (scat == &rm->m_sg[rm->m_count]) { 732 if (scat == &rm->data.op_sg[rm->data.op_count]) {
732 prev->s_rm = ic->i_rm; 733 prev->s_rm = ic->i_rm;
733 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED; 734 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
734 ic->i_rm = NULL; 735 ic->i_rm = NULL;
@@ -784,7 +785,7 @@ static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rd
784 ib_update_fast_reg_key(send->s_mr, send->s_remap_count++); 785 ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
785} 786}
786 787
787int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op) 788int rds_iw_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
788{ 789{
789 struct rds_iw_connection *ic = conn->c_transport_data; 790 struct rds_iw_connection *ic = conn->c_transport_data;
790 struct rds_iw_send_work *send = NULL; 791 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)
794 struct rds_iw_device *rds_iwdev; 795 struct rds_iw_device *rds_iwdev;
795 struct scatterlist *scat; 796 struct scatterlist *scat;
796 unsigned long len; 797 unsigned long len;
797 u64 remote_addr = op->r_remote_addr; 798 u64 remote_addr = op->op_remote_addr;
798 u32 pos, fr_pos; 799 u32 pos, fr_pos;
799 u32 work_alloc; 800 u32 work_alloc;
800 u32 i; 801 u32 i;
@@ -806,21 +807,21 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
806 rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client); 807 rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
807 808
808 /* map the message the first time we see it */ 809 /* map the message the first time we see it */
809 if (!op->r_mapped) { 810 if (!op->op_mapped) {
810 op->r_count = ib_dma_map_sg(ic->i_cm_id->device, 811 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
811 op->r_sg, op->r_nents, (op->r_write) ? 812 op->op_sg, op->op_nents, (op->op_write) ?
812 DMA_TO_DEVICE : DMA_FROM_DEVICE); 813 DMA_TO_DEVICE : DMA_FROM_DEVICE);
813 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count); 814 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
814 if (op->r_count == 0) { 815 if (op->op_count == 0) {
815 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure); 816 rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
816 ret = -ENOMEM; /* XXX ? */ 817 ret = -ENOMEM; /* XXX ? */
817 goto out; 818 goto out;
818 } 819 }
819 820
820 op->r_mapped = 1; 821 op->op_mapped = 1;
821 } 822 }
822 823
823 if (!op->r_write) { 824 if (!op->op_write) {
824 /* Alloc space on the send queue for the fastreg */ 825 /* Alloc space on the send queue for the fastreg */
825 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos); 826 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
826 if (work_alloc != 1) { 827 if (work_alloc != 1) {
@@ -835,7 +836,7 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
835 * Instead of knowing how to return a partial rdma read/write we insist that there 836 * Instead of knowing how to return a partial rdma read/write we insist that there
836 * be enough work requests to send the entire message. 837 * be enough work requests to send the entire message.
837 */ 838 */
838 i = ceil(op->r_count, rds_iwdev->max_sge); 839 i = ceil(op->op_count, rds_iwdev->max_sge);
839 840
840 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos); 841 work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
841 if (work_alloc != i) { 842 if (work_alloc != i) {
@@ -846,17 +847,17 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
846 } 847 }
847 848
848 send = &ic->i_sends[pos]; 849 send = &ic->i_sends[pos];
849 if (!op->r_write) { 850 if (!op->op_write) {
850 first = prev = &ic->i_sends[fr_pos]; 851 first = prev = &ic->i_sends[fr_pos];
851 } else { 852 } else {
852 first = send; 853 first = send;
853 prev = NULL; 854 prev = NULL;
854 } 855 }
855 scat = &op->r_sg[0]; 856 scat = &op->op_sg[0];
856 sent = 0; 857 sent = 0;
857 num_sge = op->r_count; 858 num_sge = op->op_count;
858 859
859 for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) { 860 for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
860 send->s_wr.send_flags = 0; 861 send->s_wr.send_flags = 0;
861 send->s_queued = jiffies; 862 send->s_queued = jiffies;
862 863
@@ -873,13 +874,13 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
873 * for local access after RDS is finished with it, using 874 * for local access after RDS is finished with it, using
874 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed. 875 * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
875 */ 876 */
876 if (op->r_write) 877 if (op->op_write)
877 send->s_wr.opcode = IB_WR_RDMA_WRITE; 878 send->s_wr.opcode = IB_WR_RDMA_WRITE;
878 else 879 else
879 send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV; 880 send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
880 881
881 send->s_wr.wr.rdma.remote_addr = remote_addr; 882 send->s_wr.wr.rdma.remote_addr = remote_addr;
882 send->s_wr.wr.rdma.rkey = op->r_key; 883 send->s_wr.wr.rdma.rkey = op->op_rkey;
883 send->s_op = op; 884 send->s_op = op;
884 885
885 if (num_sge > rds_iwdev->max_sge) { 886 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)
893 if (prev) 894 if (prev)
894 prev->s_wr.next = &send->s_wr; 895 prev->s_wr.next = &send->s_wr;
895 896
896 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) { 897 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
897 len = ib_sg_dma_len(ic->i_cm_id->device, scat); 898 len = ib_sg_dma_len(ic->i_cm_id->device, scat);
898 899
899 if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) 900 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)
927 } 928 }
928 929
929 /* if we finished the message then send completion owns it */ 930 /* if we finished the message then send completion owns it */
930 if (scat == &op->r_sg[op->r_count]) 931 if (scat == &op->op_sg[op->op_count])
931 first->s_wr.send_flags = IB_SEND_SIGNALED; 932 first->s_wr.send_flags = IB_SEND_SIGNALED;
932 933
933 if (i < work_alloc) { 934 if (i < work_alloc) {
@@ -941,9 +942,9 @@ int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
941 * adapters do not allow using the lkey for this at all. To bypass this use a 942 * adapters do not allow using the lkey for this at all. To bypass this use a
942 * fastreg_mr (or possibly a dma_mr) 943 * fastreg_mr (or possibly a dma_mr)
943 */ 944 */
944 if (!op->r_write) { 945 if (!op->op_write) {
945 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos], 946 rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
946 op->r_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr); 947 op->op_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
947 work_alloc++; 948 work_alloc++;
948 } 949 }
949 950
diff --git a/net/rds/iw_sysctl.c b/net/rds/iw_sysctl.c
index 1c4428a61a02..23e3a9a26aaf 100644
--- a/net/rds/iw_sysctl.c
+++ b/net/rds/iw_sysctl.c
@@ -122,10 +122,10 @@ void rds_iw_sysctl_exit(void)
122 unregister_sysctl_table(rds_iw_sysctl_hdr); 122 unregister_sysctl_table(rds_iw_sysctl_hdr);
123} 123}
124 124
125int __init rds_iw_sysctl_init(void) 125int rds_iw_sysctl_init(void)
126{ 126{
127 rds_iw_sysctl_hdr = register_sysctl_paths(rds_iw_sysctl_path, rds_iw_sysctl_table); 127 rds_iw_sysctl_hdr = register_sysctl_paths(rds_iw_sysctl_path, rds_iw_sysctl_table);
128 if (rds_iw_sysctl_hdr == NULL) 128 if (!rds_iw_sysctl_hdr)
129 return -ENOMEM; 129 return -ENOMEM;
130 return 0; 130 return 0;
131} 131}
diff --git a/net/rds/loop.c b/net/rds/loop.c
index dd9879379457..c390156b426f 100644
--- a/net/rds/loop.c
+++ b/net/rds/loop.c
@@ -61,10 +61,17 @@ static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
61 unsigned int hdr_off, unsigned int sg, 61 unsigned int hdr_off, unsigned int sg,
62 unsigned int off) 62 unsigned int off)
63{ 63{
64 /* Do not send cong updates to loopback */
65 if (rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
66 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
67 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
68 }
69
64 BUG_ON(hdr_off || sg || off); 70 BUG_ON(hdr_off || sg || off);
65 71
66 rds_inc_init(&rm->m_inc, conn, conn->c_laddr); 72 rds_inc_init(&rm->m_inc, conn, conn->c_laddr);
67 rds_message_addref(rm); /* for the inc */ 73 /* For the embedded inc. Matching put is in loop_inc_free() */
74 rds_message_addref(rm);
68 75
69 rds_recv_incoming(conn, conn->c_laddr, conn->c_faddr, &rm->m_inc, 76 rds_recv_incoming(conn, conn->c_laddr, conn->c_faddr, &rm->m_inc,
70 GFP_KERNEL, KM_USER0); 77 GFP_KERNEL, KM_USER0);
@@ -77,16 +84,14 @@ static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
77 return sizeof(struct rds_header) + be32_to_cpu(rm->m_inc.i_hdr.h_len); 84 return sizeof(struct rds_header) + be32_to_cpu(rm->m_inc.i_hdr.h_len);
78} 85}
79 86
80static int rds_loop_xmit_cong_map(struct rds_connection *conn, 87/*
81 struct rds_cong_map *map, 88 * See rds_loop_xmit(). Since our inc is embedded in the rm, we
82 unsigned long offset) 89 * make sure the rm lives at least until the inc is done.
90 */
91static void rds_loop_inc_free(struct rds_incoming *inc)
83{ 92{
84 BUG_ON(offset); 93 struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
85 BUG_ON(map != conn->c_lcong); 94 rds_message_put(rm);
86
87 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
88
89 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
90} 95}
91 96
92/* we need to at least give the thread something to succeed */ 97/* 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)
112 unsigned long flags; 117 unsigned long flags;
113 118
114 lc = kzalloc(sizeof(struct rds_loop_connection), GFP_KERNEL); 119 lc = kzalloc(sizeof(struct rds_loop_connection), GFP_KERNEL);
115 if (lc == NULL) 120 if (!lc)
116 return -ENOMEM; 121 return -ENOMEM;
117 122
118 INIT_LIST_HEAD(&lc->loop_node); 123 INIT_LIST_HEAD(&lc->loop_node);
@@ -169,14 +174,12 @@ void rds_loop_exit(void)
169 */ 174 */
170struct rds_transport rds_loop_transport = { 175struct rds_transport rds_loop_transport = {
171 .xmit = rds_loop_xmit, 176 .xmit = rds_loop_xmit,
172 .xmit_cong_map = rds_loop_xmit_cong_map,
173 .recv = rds_loop_recv, 177 .recv = rds_loop_recv,
174 .conn_alloc = rds_loop_conn_alloc, 178 .conn_alloc = rds_loop_conn_alloc,
175 .conn_free = rds_loop_conn_free, 179 .conn_free = rds_loop_conn_free,
176 .conn_connect = rds_loop_conn_connect, 180 .conn_connect = rds_loop_conn_connect,
177 .conn_shutdown = rds_loop_conn_shutdown, 181 .conn_shutdown = rds_loop_conn_shutdown,
178 .inc_copy_to_user = rds_message_inc_copy_to_user, 182 .inc_copy_to_user = rds_message_inc_copy_to_user,
179 .inc_purge = rds_message_inc_purge, 183 .inc_free = rds_loop_inc_free,
180 .inc_free = rds_message_inc_free,
181 .t_name = "loopback", 184 .t_name = "loopback",
182}; 185};
diff --git a/net/rds/message.c b/net/rds/message.c
index 9a1d67e001ba..84f937f11d47 100644
--- a/net/rds/message.c
+++ b/net/rds/message.c
@@ -34,9 +34,6 @@
34#include <linux/slab.h> 34#include <linux/slab.h>
35 35
36#include "rds.h" 36#include "rds.h"
37#include "rdma.h"
38
39static DECLARE_WAIT_QUEUE_HEAD(rds_message_flush_waitq);
40 37
41static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = { 38static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = {
42[RDS_EXTHDR_NONE] = 0, 39[RDS_EXTHDR_NONE] = 0,
@@ -63,29 +60,31 @@ static void rds_message_purge(struct rds_message *rm)
63 if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags))) 60 if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
64 return; 61 return;
65 62
66 for (i = 0; i < rm->m_nents; i++) { 63 for (i = 0; i < rm->data.op_nents; i++) {
67 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->m_sg[i])); 64 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.op_sg[i]));
68 /* XXX will have to put_page for page refs */ 65 /* XXX will have to put_page for page refs */
69 __free_page(sg_page(&rm->m_sg[i])); 66 __free_page(sg_page(&rm->data.op_sg[i]));
70 } 67 }
71 rm->m_nents = 0; 68 rm->data.op_nents = 0;
72 69
73 if (rm->m_rdma_op) 70 if (rm->rdma.op_active)
74 rds_rdma_free_op(rm->m_rdma_op); 71 rds_rdma_free_op(&rm->rdma);
75 if (rm->m_rdma_mr) 72 if (rm->rdma.op_rdma_mr)
76 rds_mr_put(rm->m_rdma_mr); 73 rds_mr_put(rm->rdma.op_rdma_mr);
77}
78 74
79void rds_message_inc_purge(struct rds_incoming *inc) 75 if (rm->atomic.op_active)
80{ 76 rds_atomic_free_op(&rm->atomic);
81 struct rds_message *rm = container_of(inc, struct rds_message, m_inc); 77 if (rm->atomic.op_rdma_mr)
82 rds_message_purge(rm); 78 rds_mr_put(rm->atomic.op_rdma_mr);
83} 79}
84 80
85void rds_message_put(struct rds_message *rm) 81void rds_message_put(struct rds_message *rm)
86{ 82{
87 rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount)); 83 rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
88 84 if (atomic_read(&rm->m_refcount) == 0) {
85printk(KERN_CRIT "danger refcount zero on %p\n", rm);
86WARN_ON(1);
87 }
89 if (atomic_dec_and_test(&rm->m_refcount)) { 88 if (atomic_dec_and_test(&rm->m_refcount)) {
90 BUG_ON(!list_empty(&rm->m_sock_item)); 89 BUG_ON(!list_empty(&rm->m_sock_item));
91 BUG_ON(!list_empty(&rm->m_conn_item)); 90 BUG_ON(!list_empty(&rm->m_conn_item));
@@ -96,12 +95,6 @@ void rds_message_put(struct rds_message *rm)
96} 95}
97EXPORT_SYMBOL_GPL(rds_message_put); 96EXPORT_SYMBOL_GPL(rds_message_put);
98 97
99void rds_message_inc_free(struct rds_incoming *inc)
100{
101 struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
102 rds_message_put(rm);
103}
104
105void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 98void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
106 __be16 dport, u64 seq) 99 __be16 dport, u64 seq)
107{ 100{
@@ -214,41 +207,68 @@ int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 o
214} 207}
215EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension); 208EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
216 209
217struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp) 210/*
211 * Each rds_message is allocated with extra space for the scatterlist entries
212 * rds ops will need. This is to minimize memory allocation count. Then, each rds op
213 * can grab SGs when initializing its part of the rds_message.
214 */
215struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
218{ 216{
219 struct rds_message *rm; 217 struct rds_message *rm;
220 218
221 rm = kzalloc(sizeof(struct rds_message) + 219 rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
222 (nents * sizeof(struct scatterlist)), gfp);
223 if (!rm) 220 if (!rm)
224 goto out; 221 goto out;
225 222
226 if (nents) 223 rm->m_used_sgs = 0;
227 sg_init_table(rm->m_sg, nents); 224 rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
225
228 atomic_set(&rm->m_refcount, 1); 226 atomic_set(&rm->m_refcount, 1);
229 INIT_LIST_HEAD(&rm->m_sock_item); 227 INIT_LIST_HEAD(&rm->m_sock_item);
230 INIT_LIST_HEAD(&rm->m_conn_item); 228 INIT_LIST_HEAD(&rm->m_conn_item);
231 spin_lock_init(&rm->m_rs_lock); 229 spin_lock_init(&rm->m_rs_lock);
230 init_waitqueue_head(&rm->m_flush_wait);
232 231
233out: 232out:
234 return rm; 233 return rm;
235} 234}
236 235
236/*
237 * RDS ops use this to grab SG entries from the rm's sg pool.
238 */
239struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
240{
241 struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
242 struct scatterlist *sg_ret;
243
244 WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
245 WARN_ON(!nents);
246
247 sg_ret = &sg_first[rm->m_used_sgs];
248 sg_init_table(sg_ret, nents);
249 rm->m_used_sgs += nents;
250
251 return sg_ret;
252}
253
237struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len) 254struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
238{ 255{
239 struct rds_message *rm; 256 struct rds_message *rm;
240 unsigned int i; 257 unsigned int i;
258 int num_sgs = ceil(total_len, PAGE_SIZE);
259 int extra_bytes = num_sgs * sizeof(struct scatterlist);
241 260
242 rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL); 261 rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
243 if (rm == NULL) 262 if (!rm)
244 return ERR_PTR(-ENOMEM); 263 return ERR_PTR(-ENOMEM);
245 264
246 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags); 265 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
247 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len); 266 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
248 rm->m_nents = ceil(total_len, PAGE_SIZE); 267 rm->data.op_nents = ceil(total_len, PAGE_SIZE);
268 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
249 269
250 for (i = 0; i < rm->m_nents; ++i) { 270 for (i = 0; i < rm->data.op_nents; ++i) {
251 sg_set_page(&rm->m_sg[i], 271 sg_set_page(&rm->data.op_sg[i],
252 virt_to_page(page_addrs[i]), 272 virt_to_page(page_addrs[i]),
253 PAGE_SIZE, 0); 273 PAGE_SIZE, 0);
254 } 274 }
@@ -256,40 +276,33 @@ struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned in
256 return rm; 276 return rm;
257} 277}
258 278
259struct rds_message *rds_message_copy_from_user(struct iovec *first_iov, 279int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
260 size_t total_len) 280 size_t total_len)
261{ 281{
262 unsigned long to_copy; 282 unsigned long to_copy;
263 unsigned long iov_off; 283 unsigned long iov_off;
264 unsigned long sg_off; 284 unsigned long sg_off;
265 struct rds_message *rm;
266 struct iovec *iov; 285 struct iovec *iov;
267 struct scatterlist *sg; 286 struct scatterlist *sg;
268 int ret; 287 int ret = 0;
269
270 rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
271 if (rm == NULL) {
272 ret = -ENOMEM;
273 goto out;
274 }
275 288
276 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len); 289 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
277 290
278 /* 291 /*
279 * now allocate and copy in the data payload. 292 * now allocate and copy in the data payload.
280 */ 293 */
281 sg = rm->m_sg; 294 sg = rm->data.op_sg;
282 iov = first_iov; 295 iov = first_iov;
283 iov_off = 0; 296 iov_off = 0;
284 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */ 297 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
285 298
286 while (total_len) { 299 while (total_len) {
287 if (sg_page(sg) == NULL) { 300 if (!sg_page(sg)) {
288 ret = rds_page_remainder_alloc(sg, total_len, 301 ret = rds_page_remainder_alloc(sg, total_len,
289 GFP_HIGHUSER); 302 GFP_HIGHUSER);
290 if (ret) 303 if (ret)
291 goto out; 304 goto out;
292 rm->m_nents++; 305 rm->data.op_nents++;
293 sg_off = 0; 306 sg_off = 0;
294 } 307 }
295 308
@@ -320,14 +333,8 @@ struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
320 sg++; 333 sg++;
321 } 334 }
322 335
323 ret = 0;
324out: 336out:
325 if (ret) { 337 return ret;
326 if (rm)
327 rds_message_put(rm);
328 rm = ERR_PTR(ret);
329 }
330 return rm;
331} 338}
332 339
333int rds_message_inc_copy_to_user(struct rds_incoming *inc, 340int 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,
348 355
349 iov = first_iov; 356 iov = first_iov;
350 iov_off = 0; 357 iov_off = 0;
351 sg = rm->m_sg; 358 sg = rm->data.op_sg;
352 vec_off = 0; 359 vec_off = 0;
353 copied = 0; 360 copied = 0;
354 361
@@ -394,15 +401,14 @@ int rds_message_inc_copy_to_user(struct rds_incoming *inc,
394 */ 401 */
395void rds_message_wait(struct rds_message *rm) 402void rds_message_wait(struct rds_message *rm)
396{ 403{
397 wait_event(rds_message_flush_waitq, 404 wait_event_interruptible(rm->m_flush_wait,
398 !test_bit(RDS_MSG_MAPPED, &rm->m_flags)); 405 !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
399} 406}
400 407
401void rds_message_unmapped(struct rds_message *rm) 408void rds_message_unmapped(struct rds_message *rm)
402{ 409{
403 clear_bit(RDS_MSG_MAPPED, &rm->m_flags); 410 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
404 if (waitqueue_active(&rds_message_flush_waitq)) 411 wake_up_interruptible(&rm->m_flush_wait);
405 wake_up(&rds_message_flush_waitq);
406} 412}
407EXPORT_SYMBOL_GPL(rds_message_unmapped); 413EXPORT_SYMBOL_GPL(rds_message_unmapped);
408 414
diff --git a/net/rds/page.c b/net/rds/page.c
index 595a952d4b17..5e44f5ae7898 100644
--- a/net/rds/page.c
+++ b/net/rds/page.c
@@ -116,7 +116,7 @@ int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
116 /* jump straight to allocation if we're trying for a huge page */ 116 /* jump straight to allocation if we're trying for a huge page */
117 if (bytes >= PAGE_SIZE) { 117 if (bytes >= PAGE_SIZE) {
118 page = alloc_page(gfp); 118 page = alloc_page(gfp);
119 if (page == NULL) { 119 if (!page) {
120 ret = -ENOMEM; 120 ret = -ENOMEM;
121 } else { 121 } else {
122 sg_set_page(scat, page, PAGE_SIZE, 0); 122 sg_set_page(scat, page, PAGE_SIZE, 0);
@@ -162,7 +162,7 @@ int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
162 rem = &per_cpu(rds_page_remainders, get_cpu()); 162 rem = &per_cpu(rds_page_remainders, get_cpu());
163 local_irq_save(flags); 163 local_irq_save(flags);
164 164
165 if (page == NULL) { 165 if (!page) {
166 ret = -ENOMEM; 166 ret = -ENOMEM;
167 break; 167 break;
168 } 168 }
@@ -186,6 +186,7 @@ out:
186 ret ? 0 : scat->length); 186 ret ? 0 : scat->length);
187 return ret; 187 return ret;
188} 188}
189EXPORT_SYMBOL_GPL(rds_page_remainder_alloc);
189 190
190static int rds_page_remainder_cpu_notify(struct notifier_block *self, 191static int rds_page_remainder_cpu_notify(struct notifier_block *self,
191 unsigned long action, void *hcpu) 192 unsigned long action, void *hcpu)
diff --git a/net/rds/rdma.c b/net/rds/rdma.c
index 75fd13bb631b..48064673fc76 100644
--- a/net/rds/rdma.c
+++ b/net/rds/rdma.c
@@ -35,7 +35,7 @@
35#include <linux/rbtree.h> 35#include <linux/rbtree.h>
36#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */ 36#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */
37 37
38#include "rdma.h" 38#include "rds.h"
39 39
40/* 40/*
41 * XXX 41 * XXX
@@ -130,14 +130,22 @@ void rds_rdma_drop_keys(struct rds_sock *rs)
130{ 130{
131 struct rds_mr *mr; 131 struct rds_mr *mr;
132 struct rb_node *node; 132 struct rb_node *node;
133 unsigned long flags;
133 134
134 /* Release any MRs associated with this socket */ 135 /* Release any MRs associated with this socket */
136 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
135 while ((node = rb_first(&rs->rs_rdma_keys))) { 137 while ((node = rb_first(&rs->rs_rdma_keys))) {
136 mr = container_of(node, struct rds_mr, r_rb_node); 138 mr = container_of(node, struct rds_mr, r_rb_node);
137 if (mr->r_trans == rs->rs_transport) 139 if (mr->r_trans == rs->rs_transport)
138 mr->r_invalidate = 0; 140 mr->r_invalidate = 0;
141 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
142 RB_CLEAR_NODE(&mr->r_rb_node);
143 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
144 rds_destroy_mr(mr);
139 rds_mr_put(mr); 145 rds_mr_put(mr);
146 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
140 } 147 }
148 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
141 149
142 if (rs->rs_transport && rs->rs_transport->flush_mrs) 150 if (rs->rs_transport && rs->rs_transport->flush_mrs)
143 rs->rs_transport->flush_mrs(); 151 rs->rs_transport->flush_mrs();
@@ -181,7 +189,7 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
181 goto out; 189 goto out;
182 } 190 }
183 191
184 if (rs->rs_transport->get_mr == NULL) { 192 if (!rs->rs_transport->get_mr) {
185 ret = -EOPNOTSUPP; 193 ret = -EOPNOTSUPP;
186 goto out; 194 goto out;
187 } 195 }
@@ -197,13 +205,13 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
197 205
198 /* XXX clamp nr_pages to limit the size of this alloc? */ 206 /* XXX clamp nr_pages to limit the size of this alloc? */
199 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); 207 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
200 if (pages == NULL) { 208 if (!pages) {
201 ret = -ENOMEM; 209 ret = -ENOMEM;
202 goto out; 210 goto out;
203 } 211 }
204 212
205 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL); 213 mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL);
206 if (mr == NULL) { 214 if (!mr) {
207 ret = -ENOMEM; 215 ret = -ENOMEM;
208 goto out; 216 goto out;
209 } 217 }
@@ -230,13 +238,13 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
230 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to 238 * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
231 * the zero page. 239 * the zero page.
232 */ 240 */
233 ret = rds_pin_pages(args->vec.addr & PAGE_MASK, nr_pages, pages, 1); 241 ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
234 if (ret < 0) 242 if (ret < 0)
235 goto out; 243 goto out;
236 244
237 nents = ret; 245 nents = ret;
238 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL); 246 sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
239 if (sg == NULL) { 247 if (!sg) {
240 ret = -ENOMEM; 248 ret = -ENOMEM;
241 goto out; 249 goto out;
242 } 250 }
@@ -406,68 +414,127 @@ void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
406 414
407 spin_lock_irqsave(&rs->rs_rdma_lock, flags); 415 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
408 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); 416 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
409 if (mr && (mr->r_use_once || force)) { 417 if (!mr) {
418 printk(KERN_ERR "rds: trying to unuse MR with unknown r_key %u!\n", r_key);
419 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
420 return;
421 }
422
423 if (mr->r_use_once || force) {
410 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); 424 rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
411 RB_CLEAR_NODE(&mr->r_rb_node); 425 RB_CLEAR_NODE(&mr->r_rb_node);
412 zot_me = 1; 426 zot_me = 1;
413 } else if (mr) 427 }
414 atomic_inc(&mr->r_refcount);
415 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); 428 spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
416 429
417 /* May have to issue a dma_sync on this memory region. 430 /* May have to issue a dma_sync on this memory region.
418 * Note we could avoid this if the operation was a RDMA READ, 431 * Note we could avoid this if the operation was a RDMA READ,
419 * but at this point we can't tell. */ 432 * but at this point we can't tell. */
420 if (mr != NULL) { 433 if (mr->r_trans->sync_mr)
421 if (mr->r_trans->sync_mr) 434 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
422 mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE); 435
423 436 /* If the MR was marked as invalidate, this will
424 /* If the MR was marked as invalidate, this will 437 * trigger an async flush. */
425 * trigger an async flush. */ 438 if (zot_me)
426 if (zot_me) 439 rds_destroy_mr(mr);
427 rds_destroy_mr(mr); 440 rds_mr_put(mr);
428 rds_mr_put(mr);
429 }
430} 441}
431 442
432void rds_rdma_free_op(struct rds_rdma_op *ro) 443void rds_rdma_free_op(struct rm_rdma_op *ro)
433{ 444{
434 unsigned int i; 445 unsigned int i;
435 446
436 for (i = 0; i < ro->r_nents; i++) { 447 for (i = 0; i < ro->op_nents; i++) {
437 struct page *page = sg_page(&ro->r_sg[i]); 448 struct page *page = sg_page(&ro->op_sg[i]);
438 449
439 /* Mark page dirty if it was possibly modified, which 450 /* Mark page dirty if it was possibly modified, which
440 * is the case for a RDMA_READ which copies from remote 451 * is the case for a RDMA_READ which copies from remote
441 * to local memory */ 452 * to local memory */
442 if (!ro->r_write) { 453 if (!ro->op_write) {
443 BUG_ON(in_interrupt()); 454 BUG_ON(irqs_disabled());
444 set_page_dirty(page); 455 set_page_dirty(page);
445 } 456 }
446 put_page(page); 457 put_page(page);
447 } 458 }
448 459
449 kfree(ro->r_notifier); 460 kfree(ro->op_notifier);
450 kfree(ro); 461 ro->op_notifier = NULL;
462 ro->op_active = 0;
463}
464
465void rds_atomic_free_op(struct rm_atomic_op *ao)
466{
467 struct page *page = sg_page(ao->op_sg);
468
469 /* Mark page dirty if it was possibly modified, which
470 * is the case for a RDMA_READ which copies from remote
471 * to local memory */
472 set_page_dirty(page);
473 put_page(page);
474
475 kfree(ao->op_notifier);
476 ao->op_notifier = NULL;
477 ao->op_active = 0;
478}
479
480
481/*
482 * Count the number of pages needed to describe an incoming iovec.
483 */
484static int rds_rdma_pages(struct rds_rdma_args *args)
485{
486 struct rds_iovec vec;
487 struct rds_iovec __user *local_vec;
488 unsigned int tot_pages = 0;
489 unsigned int nr_pages;
490 unsigned int i;
491
492 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
493
494 /* figure out the number of pages in the vector */
495 for (i = 0; i < args->nr_local; i++) {
496 if (copy_from_user(&vec, &local_vec[i],
497 sizeof(struct rds_iovec)))
498 return -EFAULT;
499
500 nr_pages = rds_pages_in_vec(&vec);
501 if (nr_pages == 0)
502 return -EINVAL;
503
504 tot_pages += nr_pages;
505 }
506
507 return tot_pages;
508}
509
510int rds_rdma_extra_size(struct rds_rdma_args *args)
511{
512 return rds_rdma_pages(args) * sizeof(struct scatterlist);
451} 513}
452 514
453/* 515/*
454 * args is a pointer to an in-kernel copy in the sendmsg cmsg. 516 * The application asks for a RDMA transfer.
517 * Extract all arguments and set up the rdma_op
455 */ 518 */
456static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs, 519int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
457 struct rds_rdma_args *args) 520 struct cmsghdr *cmsg)
458{ 521{
522 struct rds_rdma_args *args;
459 struct rds_iovec vec; 523 struct rds_iovec vec;
460 struct rds_rdma_op *op = NULL; 524 struct rm_rdma_op *op = &rm->rdma;
461 unsigned int nr_pages; 525 unsigned int nr_pages;
462 unsigned int max_pages;
463 unsigned int nr_bytes; 526 unsigned int nr_bytes;
464 struct page **pages = NULL; 527 struct page **pages = NULL;
465 struct rds_iovec __user *local_vec; 528 struct rds_iovec __user *local_vec;
466 struct scatterlist *sg;
467 unsigned int nr; 529 unsigned int nr;
468 unsigned int i, j; 530 unsigned int i, j;
469 int ret; 531 int ret = 0;
532
533 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
534 || rm->rdma.op_active)
535 return -EINVAL;
470 536
537 args = CMSG_DATA(cmsg);
471 538
472 if (rs->rs_bound_addr == 0) { 539 if (rs->rs_bound_addr == 0) {
473 ret = -ENOTCONN; /* XXX not a great errno */ 540 ret = -ENOTCONN; /* XXX not a great errno */
@@ -479,61 +546,38 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
479 goto out; 546 goto out;
480 } 547 }
481 548
482 nr_pages = 0; 549 nr_pages = rds_rdma_pages(args);
483 max_pages = 0; 550 if (nr_pages < 0)
484
485 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
486
487 /* figure out the number of pages in the vector */
488 for (i = 0; i < args->nr_local; i++) {
489 if (copy_from_user(&vec, &local_vec[i],
490 sizeof(struct rds_iovec))) {
491 ret = -EFAULT;
492 goto out;
493 }
494
495 nr = rds_pages_in_vec(&vec);
496 if (nr == 0) {
497 ret = -EINVAL;
498 goto out;
499 }
500
501 max_pages = max(nr, max_pages);
502 nr_pages += nr;
503 }
504
505 pages = kcalloc(max_pages, sizeof(struct page *), GFP_KERNEL);
506 if (pages == NULL) {
507 ret = -ENOMEM;
508 goto out; 551 goto out;
509 }
510 552
511 op = kzalloc(offsetof(struct rds_rdma_op, r_sg[nr_pages]), GFP_KERNEL); 553 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
512 if (op == NULL) { 554 if (!pages) {
513 ret = -ENOMEM; 555 ret = -ENOMEM;
514 goto out; 556 goto out;
515 } 557 }
516 558
517 op->r_write = !!(args->flags & RDS_RDMA_READWRITE); 559 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
518 op->r_fence = !!(args->flags & RDS_RDMA_FENCE); 560 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
519 op->r_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); 561 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
520 op->r_recverr = rs->rs_recverr; 562 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
563 op->op_active = 1;
564 op->op_recverr = rs->rs_recverr;
521 WARN_ON(!nr_pages); 565 WARN_ON(!nr_pages);
522 sg_init_table(op->r_sg, nr_pages); 566 op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
523 567
524 if (op->r_notify || op->r_recverr) { 568 if (op->op_notify || op->op_recverr) {
525 /* We allocate an uninitialized notifier here, because 569 /* We allocate an uninitialized notifier here, because
526 * we don't want to do that in the completion handler. We 570 * we don't want to do that in the completion handler. We
527 * would have to use GFP_ATOMIC there, and don't want to deal 571 * would have to use GFP_ATOMIC there, and don't want to deal
528 * with failed allocations. 572 * with failed allocations.
529 */ 573 */
530 op->r_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL); 574 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
531 if (!op->r_notifier) { 575 if (!op->op_notifier) {
532 ret = -ENOMEM; 576 ret = -ENOMEM;
533 goto out; 577 goto out;
534 } 578 }
535 op->r_notifier->n_user_token = args->user_token; 579 op->op_notifier->n_user_token = args->user_token;
536 op->r_notifier->n_status = RDS_RDMA_SUCCESS; 580 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
537 } 581 }
538 582
539 /* The cookie contains the R_Key of the remote memory region, and 583 /* 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,
543 * destination address (which is really an offset into the MR) 587 * destination address (which is really an offset into the MR)
544 * FIXME: We may want to move this into ib_rdma.c 588 * FIXME: We may want to move this into ib_rdma.c
545 */ 589 */
546 op->r_key = rds_rdma_cookie_key(args->cookie); 590 op->op_rkey = rds_rdma_cookie_key(args->cookie);
547 op->r_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie); 591 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
548 592
549 nr_bytes = 0; 593 nr_bytes = 0;
550 594
551 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n", 595 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
552 (unsigned long long)args->nr_local, 596 (unsigned long long)args->nr_local,
553 (unsigned long long)args->remote_vec.addr, 597 (unsigned long long)args->remote_vec.addr,
554 op->r_key); 598 op->op_rkey);
599
600 local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
555 601
556 for (i = 0; i < args->nr_local; i++) { 602 for (i = 0; i < args->nr_local; i++) {
557 if (copy_from_user(&vec, &local_vec[i], 603 if (copy_from_user(&vec, &local_vec[i],
@@ -569,15 +615,10 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
569 rs->rs_user_addr = vec.addr; 615 rs->rs_user_addr = vec.addr;
570 rs->rs_user_bytes = vec.bytes; 616 rs->rs_user_bytes = vec.bytes;
571 617
572 /* did the user change the vec under us? */
573 if (nr > max_pages || op->r_nents + nr > nr_pages) {
574 ret = -EINVAL;
575 goto out;
576 }
577 /* If it's a WRITE operation, we want to pin the pages for reading. 618 /* If it's a WRITE operation, we want to pin the pages for reading.
578 * If it's a READ operation, we need to pin the pages for writing. 619 * If it's a READ operation, we need to pin the pages for writing.
579 */ 620 */
580 ret = rds_pin_pages(vec.addr & PAGE_MASK, nr, pages, !op->r_write); 621 ret = rds_pin_pages(vec.addr, nr, pages, !op->op_write);
581 if (ret < 0) 622 if (ret < 0)
582 goto out; 623 goto out;
583 624
@@ -588,8 +629,9 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
588 629
589 for (j = 0; j < nr; j++) { 630 for (j = 0; j < nr; j++) {
590 unsigned int offset = vec.addr & ~PAGE_MASK; 631 unsigned int offset = vec.addr & ~PAGE_MASK;
632 struct scatterlist *sg;
591 633
592 sg = &op->r_sg[op->r_nents + j]; 634 sg = &op->op_sg[op->op_nents + j];
593 sg_set_page(sg, pages[j], 635 sg_set_page(sg, pages[j],
594 min_t(unsigned int, vec.bytes, PAGE_SIZE - offset), 636 min_t(unsigned int, vec.bytes, PAGE_SIZE - offset),
595 offset); 637 offset);
@@ -601,10 +643,9 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
601 vec.bytes -= sg->length; 643 vec.bytes -= sg->length;
602 } 644 }
603 645
604 op->r_nents += nr; 646 op->op_nents += nr;
605 } 647 }
606 648
607
608 if (nr_bytes > args->remote_vec.bytes) { 649 if (nr_bytes > args->remote_vec.bytes) {
609 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n", 650 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
610 nr_bytes, 651 nr_bytes,
@@ -612,38 +653,17 @@ static struct rds_rdma_op *rds_rdma_prepare(struct rds_sock *rs,
612 ret = -EINVAL; 653 ret = -EINVAL;
613 goto out; 654 goto out;
614 } 655 }
615 op->r_bytes = nr_bytes; 656 op->op_bytes = nr_bytes;
616 657
617 ret = 0; 658 ret = 0;
618out: 659out:
619 kfree(pages); 660 kfree(pages);
620 if (ret) { 661 if (ret)
621 if (op) 662 rds_rdma_free_op(op);
622 rds_rdma_free_op(op);
623 op = ERR_PTR(ret);
624 }
625 return op;
626}
627
628/*
629 * The application asks for a RDMA transfer.
630 * Extract all arguments and set up the rdma_op
631 */
632int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
633 struct cmsghdr *cmsg)
634{
635 struct rds_rdma_op *op;
636
637 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) ||
638 rm->m_rdma_op != NULL)
639 return -EINVAL;
640 663
641 op = rds_rdma_prepare(rs, CMSG_DATA(cmsg));
642 if (IS_ERR(op))
643 return PTR_ERR(op);
644 rds_stats_inc(s_send_rdma); 664 rds_stats_inc(s_send_rdma);
645 rm->m_rdma_op = op; 665
646 return 0; 666 return ret;
647} 667}
648 668
649/* 669/*
@@ -673,7 +693,7 @@ int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
673 693
674 spin_lock_irqsave(&rs->rs_rdma_lock, flags); 694 spin_lock_irqsave(&rs->rs_rdma_lock, flags);
675 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); 695 mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
676 if (mr == NULL) 696 if (!mr)
677 err = -EINVAL; /* invalid r_key */ 697 err = -EINVAL; /* invalid r_key */
678 else 698 else
679 atomic_inc(&mr->r_refcount); 699 atomic_inc(&mr->r_refcount);
@@ -681,7 +701,7 @@ int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
681 701
682 if (mr) { 702 if (mr) {
683 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE); 703 mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE);
684 rm->m_rdma_mr = mr; 704 rm->rdma.op_rdma_mr = mr;
685 } 705 }
686 return err; 706 return err;
687} 707}
@@ -699,5 +719,98 @@ int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
699 rm->m_rdma_cookie != 0) 719 rm->m_rdma_cookie != 0)
700 return -EINVAL; 720 return -EINVAL;
701 721
702 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->m_rdma_mr); 722 return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, &rm->rdma.op_rdma_mr);
723}
724
725/*
726 * Fill in rds_message for an atomic request.
727 */
728int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
729 struct cmsghdr *cmsg)
730{
731 struct page *page = NULL;
732 struct rds_atomic_args *args;
733 int ret = 0;
734
735 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
736 || rm->atomic.op_active)
737 return -EINVAL;
738
739 args = CMSG_DATA(cmsg);
740
741 /* Nonmasked & masked cmsg ops converted to masked hw ops */
742 switch (cmsg->cmsg_type) {
743 case RDS_CMSG_ATOMIC_FADD:
744 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
745 rm->atomic.op_m_fadd.add = args->fadd.add;
746 rm->atomic.op_m_fadd.nocarry_mask = 0;
747 break;
748 case RDS_CMSG_MASKED_ATOMIC_FADD:
749 rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
750 rm->atomic.op_m_fadd.add = args->m_fadd.add;
751 rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
752 break;
753 case RDS_CMSG_ATOMIC_CSWP:
754 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
755 rm->atomic.op_m_cswp.compare = args->cswp.compare;
756 rm->atomic.op_m_cswp.swap = args->cswp.swap;
757 rm->atomic.op_m_cswp.compare_mask = ~0;
758 rm->atomic.op_m_cswp.swap_mask = ~0;
759 break;
760 case RDS_CMSG_MASKED_ATOMIC_CSWP:
761 rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
762 rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
763 rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
764 rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
765 rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
766 break;
767 default:
768 BUG(); /* should never happen */
769 }
770
771 rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
772 rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
773 rm->atomic.op_active = 1;
774 rm->atomic.op_recverr = rs->rs_recverr;
775 rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
776
777 /* verify 8 byte-aligned */
778 if (args->local_addr & 0x7) {
779 ret = -EFAULT;
780 goto err;
781 }
782
783 ret = rds_pin_pages(args->local_addr, 1, &page, 1);
784 if (ret != 1)
785 goto err;
786 ret = 0;
787
788 sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
789
790 if (rm->atomic.op_notify || rm->atomic.op_recverr) {
791 /* We allocate an uninitialized notifier here, because
792 * we don't want to do that in the completion handler. We
793 * would have to use GFP_ATOMIC there, and don't want to deal
794 * with failed allocations.
795 */
796 rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL);
797 if (!rm->atomic.op_notifier) {
798 ret = -ENOMEM;
799 goto err;
800 }
801
802 rm->atomic.op_notifier->n_user_token = args->user_token;
803 rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
804 }
805
806 rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
807 rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
808
809 return ret;
810err:
811 if (page)
812 put_page(page);
813 kfree(rm->atomic.op_notifier);
814
815 return ret;
703} 816}
diff --git a/net/rds/rdma.h b/net/rds/rdma.h
deleted file mode 100644
index 909c39835a5d..000000000000
--- a/net/rds/rdma.h
+++ /dev/null
@@ -1,85 +0,0 @@
1#ifndef _RDS_RDMA_H
2#define _RDS_RDMA_H
3
4#include <linux/rbtree.h>
5#include <linux/spinlock.h>
6#include <linux/scatterlist.h>
7
8#include "rds.h"
9
10struct rds_mr {
11 struct rb_node r_rb_node;
12 atomic_t r_refcount;
13 u32 r_key;
14
15 /* A copy of the creation flags */
16 unsigned int r_use_once:1;
17 unsigned int r_invalidate:1;
18 unsigned int r_write:1;
19
20 /* This is for RDS_MR_DEAD.
21 * It would be nice & consistent to make this part of the above
22 * bit field here, but we need to use test_and_set_bit.
23 */
24 unsigned long r_state;
25 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
26 struct rds_transport *r_trans;
27 void *r_trans_private;
28};
29
30/* Flags for mr->r_state */
31#define RDS_MR_DEAD 0
32
33struct rds_rdma_op {
34 u32 r_key;
35 u64 r_remote_addr;
36 unsigned int r_write:1;
37 unsigned int r_fence:1;
38 unsigned int r_notify:1;
39 unsigned int r_recverr:1;
40 unsigned int r_mapped:1;
41 struct rds_notifier *r_notifier;
42 unsigned int r_bytes;
43 unsigned int r_nents;
44 unsigned int r_count;
45 struct scatterlist r_sg[0];
46};
47
48static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
49{
50 return r_key | (((u64) offset) << 32);
51}
52
53static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
54{
55 return cookie;
56}
57
58static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
59{
60 return cookie >> 32;
61}
62
63int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
64int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
65int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
66void rds_rdma_drop_keys(struct rds_sock *rs);
67int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
68 struct cmsghdr *cmsg);
69int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
70 struct cmsghdr *cmsg);
71int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
72 struct cmsghdr *cmsg);
73int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
74 struct cmsghdr *cmsg);
75void rds_rdma_free_op(struct rds_rdma_op *ro);
76void rds_rdma_send_complete(struct rds_message *rm, int);
77
78extern void __rds_put_mr_final(struct rds_mr *mr);
79static inline void rds_mr_put(struct rds_mr *mr)
80{
81 if (atomic_dec_and_test(&mr->r_refcount))
82 __rds_put_mr_final(mr);
83}
84
85#endif
diff --git a/net/rds/rdma_transport.c b/net/rds/rdma_transport.c
index e599ba2f950d..e6ed10aee190 100644
--- a/net/rds/rdma_transport.c
+++ b/net/rds/rdma_transport.c
@@ -36,6 +36,34 @@
36 36
37static struct rdma_cm_id *rds_rdma_listen_id; 37static struct rdma_cm_id *rds_rdma_listen_id;
38 38
39static char *rds_cm_event_strings[] = {
40#define RDS_CM_EVENT_STRING(foo) \
41 [RDMA_CM_EVENT_##foo] = __stringify(RDMA_CM_EVENT_##foo)
42 RDS_CM_EVENT_STRING(ADDR_RESOLVED),
43 RDS_CM_EVENT_STRING(ADDR_ERROR),
44 RDS_CM_EVENT_STRING(ROUTE_RESOLVED),
45 RDS_CM_EVENT_STRING(ROUTE_ERROR),
46 RDS_CM_EVENT_STRING(CONNECT_REQUEST),
47 RDS_CM_EVENT_STRING(CONNECT_RESPONSE),
48 RDS_CM_EVENT_STRING(CONNECT_ERROR),
49 RDS_CM_EVENT_STRING(UNREACHABLE),
50 RDS_CM_EVENT_STRING(REJECTED),
51 RDS_CM_EVENT_STRING(ESTABLISHED),
52 RDS_CM_EVENT_STRING(DISCONNECTED),
53 RDS_CM_EVENT_STRING(DEVICE_REMOVAL),
54 RDS_CM_EVENT_STRING(MULTICAST_JOIN),
55 RDS_CM_EVENT_STRING(MULTICAST_ERROR),
56 RDS_CM_EVENT_STRING(ADDR_CHANGE),
57 RDS_CM_EVENT_STRING(TIMEWAIT_EXIT),
58#undef RDS_CM_EVENT_STRING
59};
60
61static char *rds_cm_event_str(enum rdma_cm_event_type type)
62{
63 return rds_str_array(rds_cm_event_strings,
64 ARRAY_SIZE(rds_cm_event_strings), type);
65};
66
39int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id, 67int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
40 struct rdma_cm_event *event) 68 struct rdma_cm_event *event)
41{ 69{
@@ -44,8 +72,8 @@ int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
44 struct rds_transport *trans; 72 struct rds_transport *trans;
45 int ret = 0; 73 int ret = 0;
46 74
47 rdsdebug("conn %p id %p handling event %u\n", conn, cm_id, 75 rdsdebug("conn %p id %p handling event %u (%s)\n", conn, cm_id,
48 event->event); 76 event->event, rds_cm_event_str(event->event));
49 77
50 if (cm_id->device->node_type == RDMA_NODE_RNIC) 78 if (cm_id->device->node_type == RDMA_NODE_RNIC)
51 trans = &rds_iw_transport; 79 trans = &rds_iw_transport;
@@ -109,7 +137,8 @@ int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
109 137
110 default: 138 default:
111 /* things like device disconnect? */ 139 /* things like device disconnect? */
112 printk(KERN_ERR "RDS: unknown event %u!\n", event->event); 140 printk(KERN_ERR "RDS: unknown event %u (%s)!\n",
141 event->event, rds_cm_event_str(event->event));
113 break; 142 break;
114 } 143 }
115 144
@@ -117,12 +146,13 @@ out:
117 if (conn) 146 if (conn)
118 mutex_unlock(&conn->c_cm_lock); 147 mutex_unlock(&conn->c_cm_lock);
119 148
120 rdsdebug("id %p event %u handling ret %d\n", cm_id, event->event, ret); 149 rdsdebug("id %p event %u (%s) handling ret %d\n", cm_id, event->event,
150 rds_cm_event_str(event->event), ret);
121 151
122 return ret; 152 return ret;
123} 153}
124 154
125static int __init rds_rdma_listen_init(void) 155static int rds_rdma_listen_init(void)
126{ 156{
127 struct sockaddr_in sin; 157 struct sockaddr_in sin;
128 struct rdma_cm_id *cm_id; 158 struct rdma_cm_id *cm_id;
@@ -177,7 +207,7 @@ static void rds_rdma_listen_stop(void)
177 } 207 }
178} 208}
179 209
180int __init rds_rdma_init(void) 210int rds_rdma_init(void)
181{ 211{
182 int ret; 212 int ret;
183 213
diff --git a/net/rds/rds.h b/net/rds/rds.h
index c224b5bb3ba9..8103dcf8b976 100644
--- a/net/rds/rds.h
+++ b/net/rds/rds.h
@@ -80,6 +80,7 @@ enum {
80/* Bits for c_flags */ 80/* Bits for c_flags */
81#define RDS_LL_SEND_FULL 0 81#define RDS_LL_SEND_FULL 0
82#define RDS_RECONNECT_PENDING 1 82#define RDS_RECONNECT_PENDING 1
83#define RDS_IN_XMIT 2
83 84
84struct rds_connection { 85struct rds_connection {
85 struct hlist_node c_hash_node; 86 struct hlist_node c_hash_node;
@@ -91,12 +92,13 @@ struct rds_connection {
91 struct rds_cong_map *c_lcong; 92 struct rds_cong_map *c_lcong;
92 struct rds_cong_map *c_fcong; 93 struct rds_cong_map *c_fcong;
93 94
94 struct mutex c_send_lock; /* protect send ring */
95 struct rds_message *c_xmit_rm; 95 struct rds_message *c_xmit_rm;
96 unsigned long c_xmit_sg; 96 unsigned long c_xmit_sg;
97 unsigned int c_xmit_hdr_off; 97 unsigned int c_xmit_hdr_off;
98 unsigned int c_xmit_data_off; 98 unsigned int c_xmit_data_off;
99 unsigned int c_xmit_atomic_sent;
99 unsigned int c_xmit_rdma_sent; 100 unsigned int c_xmit_rdma_sent;
101 unsigned int c_xmit_data_sent;
100 102
101 spinlock_t c_lock; /* protect msg queues */ 103 spinlock_t c_lock; /* protect msg queues */
102 u64 c_next_tx_seq; 104 u64 c_next_tx_seq;
@@ -116,11 +118,10 @@ struct rds_connection {
116 struct delayed_work c_conn_w; 118 struct delayed_work c_conn_w;
117 struct work_struct c_down_w; 119 struct work_struct c_down_w;
118 struct mutex c_cm_lock; /* protect conn state & cm */ 120 struct mutex c_cm_lock; /* protect conn state & cm */
121 wait_queue_head_t c_waitq;
119 122
120 struct list_head c_map_item; 123 struct list_head c_map_item;
121 unsigned long c_map_queued; 124 unsigned long c_map_queued;
122 unsigned long c_map_offset;
123 unsigned long c_map_bytes;
124 125
125 unsigned int c_unacked_packets; 126 unsigned int c_unacked_packets;
126 unsigned int c_unacked_bytes; 127 unsigned int c_unacked_bytes;
@@ -206,6 +207,48 @@ struct rds_incoming {
206 rds_rdma_cookie_t i_rdma_cookie; 207 rds_rdma_cookie_t i_rdma_cookie;
207}; 208};
208 209
210struct rds_mr {
211 struct rb_node r_rb_node;
212 atomic_t r_refcount;
213 u32 r_key;
214
215 /* A copy of the creation flags */
216 unsigned int r_use_once:1;
217 unsigned int r_invalidate:1;
218 unsigned int r_write:1;
219
220 /* This is for RDS_MR_DEAD.
221 * It would be nice & consistent to make this part of the above
222 * bit field here, but we need to use test_and_set_bit.
223 */
224 unsigned long r_state;
225 struct rds_sock *r_sock; /* back pointer to the socket that owns us */
226 struct rds_transport *r_trans;
227 void *r_trans_private;
228};
229
230/* Flags for mr->r_state */
231#define RDS_MR_DEAD 0
232
233static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset)
234{
235 return r_key | (((u64) offset) << 32);
236}
237
238static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie)
239{
240 return cookie;
241}
242
243static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie)
244{
245 return cookie >> 32;
246}
247
248/* atomic operation types */
249#define RDS_ATOMIC_TYPE_CSWP 0
250#define RDS_ATOMIC_TYPE_FADD 1
251
209/* 252/*
210 * m_sock_item and m_conn_item are on lists that are serialized under 253 * m_sock_item and m_conn_item are on lists that are serialized under
211 * conn->c_lock. m_sock_item has additional meaning in that once it is empty 254 * conn->c_lock. m_sock_item has additional meaning in that once it is empty
@@ -258,13 +301,71 @@ struct rds_message {
258 * -> rs->rs_lock 301 * -> rs->rs_lock
259 */ 302 */
260 spinlock_t m_rs_lock; 303 spinlock_t m_rs_lock;
304 wait_queue_head_t m_flush_wait;
305
261 struct rds_sock *m_rs; 306 struct rds_sock *m_rs;
262 struct rds_rdma_op *m_rdma_op; 307
308 /* cookie to send to remote, in rds header */
263 rds_rdma_cookie_t m_rdma_cookie; 309 rds_rdma_cookie_t m_rdma_cookie;
264 struct rds_mr *m_rdma_mr; 310
265 unsigned int m_nents; 311 unsigned int m_used_sgs;
266 unsigned int m_count; 312 unsigned int m_total_sgs;
267 struct scatterlist m_sg[0]; 313
314 void *m_final_op;
315
316 struct {
317 struct rm_atomic_op {
318 int op_type;
319 union {
320 struct {
321 uint64_t compare;
322 uint64_t swap;
323 uint64_t compare_mask;
324 uint64_t swap_mask;
325 } op_m_cswp;
326 struct {
327 uint64_t add;
328 uint64_t nocarry_mask;
329 } op_m_fadd;
330 };
331
332 u32 op_rkey;
333 u64 op_remote_addr;
334 unsigned int op_notify:1;
335 unsigned int op_recverr:1;
336 unsigned int op_mapped:1;
337 unsigned int op_silent:1;
338 unsigned int op_active:1;
339 struct scatterlist *op_sg;
340 struct rds_notifier *op_notifier;
341
342 struct rds_mr *op_rdma_mr;
343 } atomic;
344 struct rm_rdma_op {
345 u32 op_rkey;
346 u64 op_remote_addr;
347 unsigned int op_write:1;
348 unsigned int op_fence:1;
349 unsigned int op_notify:1;
350 unsigned int op_recverr:1;
351 unsigned int op_mapped:1;
352 unsigned int op_silent:1;
353 unsigned int op_active:1;
354 unsigned int op_bytes;
355 unsigned int op_nents;
356 unsigned int op_count;
357 struct scatterlist *op_sg;
358 struct rds_notifier *op_notifier;
359
360 struct rds_mr *op_rdma_mr;
361 } rdma;
362 struct rm_data_op {
363 unsigned int op_active:1;
364 unsigned int op_nents;
365 unsigned int op_count;
366 struct scatterlist *op_sg;
367 } data;
368 };
268}; 369};
269 370
270/* 371/*
@@ -305,10 +406,6 @@ struct rds_notifier {
305 * transport is responsible for other serialization, including 406 * transport is responsible for other serialization, including
306 * rds_recv_incoming(). This is called in process context but 407 * rds_recv_incoming(). This is called in process context but
307 * should try hard not to block. 408 * should try hard not to block.
308 *
309 * @xmit_cong_map: This asks the transport to send the local bitmap down the
310 * given connection. XXX get a better story about the bitmap
311 * flag and header.
312 */ 409 */
313 410
314#define RDS_TRANS_IB 0 411#define RDS_TRANS_IB 0
@@ -332,13 +429,11 @@ struct rds_transport {
332 void (*xmit_complete)(struct rds_connection *conn); 429 void (*xmit_complete)(struct rds_connection *conn);
333 int (*xmit)(struct rds_connection *conn, struct rds_message *rm, 430 int (*xmit)(struct rds_connection *conn, struct rds_message *rm,
334 unsigned int hdr_off, unsigned int sg, unsigned int off); 431 unsigned int hdr_off, unsigned int sg, unsigned int off);
335 int (*xmit_cong_map)(struct rds_connection *conn, 432 int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op);
336 struct rds_cong_map *map, unsigned long offset); 433 int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
337 int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
338 int (*recv)(struct rds_connection *conn); 434 int (*recv)(struct rds_connection *conn);
339 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov, 435 int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
340 size_t size); 436 size_t size);
341 void (*inc_purge)(struct rds_incoming *inc);
342 void (*inc_free)(struct rds_incoming *inc); 437 void (*inc_free)(struct rds_incoming *inc);
343 438
344 int (*cm_handle_connect)(struct rdma_cm_id *cm_id, 439 int (*cm_handle_connect)(struct rdma_cm_id *cm_id,
@@ -367,17 +462,11 @@ struct rds_sock {
367 * bound_addr used for both incoming and outgoing, no INADDR_ANY 462 * bound_addr used for both incoming and outgoing, no INADDR_ANY
368 * support. 463 * support.
369 */ 464 */
370 struct rb_node rs_bound_node; 465 struct hlist_node rs_bound_node;
371 __be32 rs_bound_addr; 466 __be32 rs_bound_addr;
372 __be32 rs_conn_addr; 467 __be32 rs_conn_addr;
373 __be16 rs_bound_port; 468 __be16 rs_bound_port;
374 __be16 rs_conn_port; 469 __be16 rs_conn_port;
375
376 /*
377 * This is only used to communicate the transport between bind and
378 * initiating connections. All other trans use is referenced through
379 * the connection.
380 */
381 struct rds_transport *rs_transport; 470 struct rds_transport *rs_transport;
382 471
383 /* 472 /*
@@ -466,8 +555,8 @@ struct rds_statistics {
466 uint64_t s_recv_ping; 555 uint64_t s_recv_ping;
467 uint64_t s_send_queue_empty; 556 uint64_t s_send_queue_empty;
468 uint64_t s_send_queue_full; 557 uint64_t s_send_queue_full;
469 uint64_t s_send_sem_contention; 558 uint64_t s_send_lock_contention;
470 uint64_t s_send_sem_queue_raced; 559 uint64_t s_send_lock_queue_raced;
471 uint64_t s_send_immediate_retry; 560 uint64_t s_send_immediate_retry;
472 uint64_t s_send_delayed_retry; 561 uint64_t s_send_delayed_retry;
473 uint64_t s_send_drop_acked; 562 uint64_t s_send_drop_acked;
@@ -487,6 +576,7 @@ struct rds_statistics {
487}; 576};
488 577
489/* af_rds.c */ 578/* af_rds.c */
579char *rds_str_array(char **array, size_t elements, size_t index);
490void rds_sock_addref(struct rds_sock *rs); 580void rds_sock_addref(struct rds_sock *rs);
491void rds_sock_put(struct rds_sock *rs); 581void rds_sock_put(struct rds_sock *rs);
492void rds_wake_sk_sleep(struct rds_sock *rs); 582void rds_wake_sk_sleep(struct rds_sock *rs);
@@ -521,15 +611,17 @@ void rds_cong_exit(void);
521struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); 611struct rds_message *rds_cong_update_alloc(struct rds_connection *conn);
522 612
523/* conn.c */ 613/* conn.c */
524int __init rds_conn_init(void); 614int rds_conn_init(void);
525void rds_conn_exit(void); 615void rds_conn_exit(void);
526struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr, 616struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
527 struct rds_transport *trans, gfp_t gfp); 617 struct rds_transport *trans, gfp_t gfp);
528struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr, 618struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
529 struct rds_transport *trans, gfp_t gfp); 619 struct rds_transport *trans, gfp_t gfp);
620void rds_conn_shutdown(struct rds_connection *conn);
530void rds_conn_destroy(struct rds_connection *conn); 621void rds_conn_destroy(struct rds_connection *conn);
531void rds_conn_reset(struct rds_connection *conn); 622void rds_conn_reset(struct rds_connection *conn);
532void rds_conn_drop(struct rds_connection *conn); 623void rds_conn_drop(struct rds_connection *conn);
624void rds_conn_connect_if_down(struct rds_connection *conn);
533void rds_for_each_conn_info(struct socket *sock, unsigned int len, 625void rds_for_each_conn_info(struct socket *sock, unsigned int len,
534 struct rds_info_iterator *iter, 626 struct rds_info_iterator *iter,
535 struct rds_info_lengths *lens, 627 struct rds_info_lengths *lens,
@@ -566,7 +658,8 @@ rds_conn_connecting(struct rds_connection *conn)
566 658
567/* message.c */ 659/* message.c */
568struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); 660struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp);
569struct rds_message *rds_message_copy_from_user(struct iovec *first_iov, 661struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents);
662int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
570 size_t total_len); 663 size_t total_len);
571struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); 664struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len);
572void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 665void 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
580int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); 673int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset);
581int rds_message_inc_copy_to_user(struct rds_incoming *inc, 674int rds_message_inc_copy_to_user(struct rds_incoming *inc,
582 struct iovec *first_iov, size_t size); 675 struct iovec *first_iov, size_t size);
583void rds_message_inc_purge(struct rds_incoming *inc);
584void rds_message_inc_free(struct rds_incoming *inc); 676void rds_message_inc_free(struct rds_incoming *inc);
585void rds_message_addref(struct rds_message *rm); 677void rds_message_addref(struct rds_message *rm);
586void rds_message_put(struct rds_message *rm); 678void rds_message_put(struct rds_message *rm);
@@ -636,14 +728,39 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest);
636typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); 728typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack);
637void rds_send_drop_acked(struct rds_connection *conn, u64 ack, 729void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
638 is_acked_func is_acked); 730 is_acked_func is_acked);
639int rds_send_acked_before(struct rds_connection *conn, u64 seq);
640void rds_send_remove_from_sock(struct list_head *messages, int status); 731void rds_send_remove_from_sock(struct list_head *messages, int status);
641int rds_send_pong(struct rds_connection *conn, __be16 dport); 732int rds_send_pong(struct rds_connection *conn, __be16 dport);
642struct rds_message *rds_send_get_message(struct rds_connection *, 733struct rds_message *rds_send_get_message(struct rds_connection *,
643 struct rds_rdma_op *); 734 struct rm_rdma_op *);
644 735
645/* rdma.c */ 736/* rdma.c */
646void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); 737void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force);
738int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen);
739int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen);
740int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen);
741void rds_rdma_drop_keys(struct rds_sock *rs);
742int rds_rdma_extra_size(struct rds_rdma_args *args);
743int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
744 struct cmsghdr *cmsg);
745int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
746 struct cmsghdr *cmsg);
747int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
748 struct cmsghdr *cmsg);
749int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
750 struct cmsghdr *cmsg);
751void rds_rdma_free_op(struct rm_rdma_op *ro);
752void rds_atomic_free_op(struct rm_atomic_op *ao);
753void rds_rdma_send_complete(struct rds_message *rm, int wc_status);
754void rds_atomic_send_complete(struct rds_message *rm, int wc_status);
755int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
756 struct cmsghdr *cmsg);
757
758extern void __rds_put_mr_final(struct rds_mr *mr);
759static inline void rds_mr_put(struct rds_mr *mr)
760{
761 if (atomic_dec_and_test(&mr->r_refcount))
762 __rds_put_mr_final(mr);
763}
647 764
648/* stats.c */ 765/* stats.c */
649DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); 766DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
@@ -657,14 +774,14 @@ DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
657 put_cpu(); \ 774 put_cpu(); \
658} while (0) 775} while (0)
659#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) 776#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count)
660int __init rds_stats_init(void); 777int rds_stats_init(void);
661void rds_stats_exit(void); 778void rds_stats_exit(void);
662void rds_stats_info_copy(struct rds_info_iterator *iter, 779void rds_stats_info_copy(struct rds_info_iterator *iter,
663 uint64_t *values, const char *const *names, 780 uint64_t *values, const char *const *names,
664 size_t nr); 781 size_t nr);
665 782
666/* sysctl.c */ 783/* sysctl.c */
667int __init rds_sysctl_init(void); 784int rds_sysctl_init(void);
668void rds_sysctl_exit(void); 785void rds_sysctl_exit(void);
669extern unsigned long rds_sysctl_sndbuf_min; 786extern unsigned long rds_sysctl_sndbuf_min;
670extern unsigned long rds_sysctl_sndbuf_default; 787extern unsigned long rds_sysctl_sndbuf_default;
@@ -678,9 +795,10 @@ extern unsigned long rds_sysctl_trace_flags;
678extern unsigned int rds_sysctl_trace_level; 795extern unsigned int rds_sysctl_trace_level;
679 796
680/* threads.c */ 797/* threads.c */
681int __init rds_threads_init(void); 798int rds_threads_init(void);
682void rds_threads_exit(void); 799void rds_threads_exit(void);
683extern struct workqueue_struct *rds_wq; 800extern struct workqueue_struct *rds_wq;
801void rds_queue_reconnect(struct rds_connection *conn);
684void rds_connect_worker(struct work_struct *); 802void rds_connect_worker(struct work_struct *);
685void rds_shutdown_worker(struct work_struct *); 803void rds_shutdown_worker(struct work_struct *);
686void rds_send_worker(struct work_struct *); 804void rds_send_worker(struct work_struct *);
@@ -691,9 +809,10 @@ void rds_connect_complete(struct rds_connection *conn);
691int rds_trans_register(struct rds_transport *trans); 809int rds_trans_register(struct rds_transport *trans);
692void rds_trans_unregister(struct rds_transport *trans); 810void rds_trans_unregister(struct rds_transport *trans);
693struct rds_transport *rds_trans_get_preferred(__be32 addr); 811struct rds_transport *rds_trans_get_preferred(__be32 addr);
812void rds_trans_put(struct rds_transport *trans);
694unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, 813unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter,
695 unsigned int avail); 814 unsigned int avail);
696int __init rds_trans_init(void); 815int rds_trans_init(void);
697void rds_trans_exit(void); 816void rds_trans_exit(void);
698 817
699#endif 818#endif
diff --git a/net/rds/recv.c b/net/rds/recv.c
index c93588c2d553..68800f02aa30 100644
--- a/net/rds/recv.c
+++ b/net/rds/recv.c
@@ -36,7 +36,6 @@
36#include <linux/in.h> 36#include <linux/in.h>
37 37
38#include "rds.h" 38#include "rds.h"
39#include "rdma.h"
40 39
41void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, 40void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
42 __be32 saddr) 41 __be32 saddr)
@@ -210,7 +209,7 @@ void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
210 } 209 }
211 210
212 rs = rds_find_bound(daddr, inc->i_hdr.h_dport); 211 rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
213 if (rs == NULL) { 212 if (!rs) {
214 rds_stats_inc(s_recv_drop_no_sock); 213 rds_stats_inc(s_recv_drop_no_sock);
215 goto out; 214 goto out;
216 } 215 }
@@ -251,7 +250,7 @@ static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
251{ 250{
252 unsigned long flags; 251 unsigned long flags;
253 252
254 if (*inc == NULL) { 253 if (!*inc) {
255 read_lock_irqsave(&rs->rs_recv_lock, flags); 254 read_lock_irqsave(&rs->rs_recv_lock, flags);
256 if (!list_empty(&rs->rs_recv_queue)) { 255 if (!list_empty(&rs->rs_recv_queue)) {
257 *inc = list_entry(rs->rs_recv_queue.next, 256 *inc = list_entry(rs->rs_recv_queue.next,
@@ -334,10 +333,10 @@ int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
334 333
335 if (msghdr) { 334 if (msghdr) {
336 cmsg.user_token = notifier->n_user_token; 335 cmsg.user_token = notifier->n_user_token;
337 cmsg.status = notifier->n_status; 336 cmsg.status = notifier->n_status;
338 337
339 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS, 338 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
340 sizeof(cmsg), &cmsg); 339 sizeof(cmsg), &cmsg);
341 if (err) 340 if (err)
342 break; 341 break;
343 } 342 }
diff --git a/net/rds/send.c b/net/rds/send.c
index 9c1c6bcaa6c9..9b951a0ab6b7 100644
--- a/net/rds/send.c
+++ b/net/rds/send.c
@@ -37,7 +37,6 @@
37#include <linux/list.h> 37#include <linux/list.h>
38 38
39#include "rds.h" 39#include "rds.h"
40#include "rdma.h"
41 40
42/* When transmitting messages in rds_send_xmit, we need to emerge from 41/* When transmitting messages in rds_send_xmit, we need to emerge from
43 * time to time and briefly release the CPU. Otherwise the softlock watchdog 42 * time to time and briefly release the CPU. Otherwise the softlock watchdog
@@ -54,7 +53,8 @@ module_param(send_batch_count, int, 0444);
54MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); 53MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
55 54
56/* 55/*
57 * Reset the send state. Caller must hold c_send_lock when calling here. 56 * Reset the send state. Callers must ensure that this doesn't race with
57 * rds_send_xmit().
58 */ 58 */
59void rds_send_reset(struct rds_connection *conn) 59void rds_send_reset(struct rds_connection *conn)
60{ 60{
@@ -62,18 +62,22 @@ void rds_send_reset(struct rds_connection *conn)
62 unsigned long flags; 62 unsigned long flags;
63 63
64 if (conn->c_xmit_rm) { 64 if (conn->c_xmit_rm) {
65 rm = conn->c_xmit_rm;
66 conn->c_xmit_rm = NULL;
65 /* Tell the user the RDMA op is no longer mapped by the 67 /* Tell the user the RDMA op is no longer mapped by the
66 * transport. This isn't entirely true (it's flushed out 68 * transport. This isn't entirely true (it's flushed out
67 * independently) but as the connection is down, there's 69 * independently) but as the connection is down, there's
68 * no ongoing RDMA to/from that memory */ 70 * no ongoing RDMA to/from that memory */
69 rds_message_unmapped(conn->c_xmit_rm); 71 rds_message_unmapped(rm);
70 rds_message_put(conn->c_xmit_rm); 72 rds_message_put(rm);
71 conn->c_xmit_rm = NULL;
72 } 73 }
74
73 conn->c_xmit_sg = 0; 75 conn->c_xmit_sg = 0;
74 conn->c_xmit_hdr_off = 0; 76 conn->c_xmit_hdr_off = 0;
75 conn->c_xmit_data_off = 0; 77 conn->c_xmit_data_off = 0;
78 conn->c_xmit_atomic_sent = 0;
76 conn->c_xmit_rdma_sent = 0; 79 conn->c_xmit_rdma_sent = 0;
80 conn->c_xmit_data_sent = 0;
77 81
78 conn->c_map_queued = 0; 82 conn->c_map_queued = 0;
79 83
@@ -90,6 +94,25 @@ void rds_send_reset(struct rds_connection *conn)
90 spin_unlock_irqrestore(&conn->c_lock, flags); 94 spin_unlock_irqrestore(&conn->c_lock, flags);
91} 95}
92 96
97static int acquire_in_xmit(struct rds_connection *conn)
98{
99 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
100}
101
102static void release_in_xmit(struct rds_connection *conn)
103{
104 clear_bit(RDS_IN_XMIT, &conn->c_flags);
105 smp_mb__after_clear_bit();
106 /*
107 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
108 * hot path and finding waiters is very rare. We don't want to walk
109 * the system-wide hashed waitqueue buckets in the fast path only to
110 * almost never find waiters.
111 */
112 if (waitqueue_active(&conn->c_waitq))
113 wake_up_all(&conn->c_waitq);
114}
115
93/* 116/*
94 * We're making the concious trade-off here to only send one message 117 * We're making the concious trade-off here to only send one message
95 * down the connection at a time. 118 * down the connection at a time.
@@ -109,102 +132,69 @@ int rds_send_xmit(struct rds_connection *conn)
109 struct rds_message *rm; 132 struct rds_message *rm;
110 unsigned long flags; 133 unsigned long flags;
111 unsigned int tmp; 134 unsigned int tmp;
112 unsigned int send_quota = send_batch_count;
113 struct scatterlist *sg; 135 struct scatterlist *sg;
114 int ret = 0; 136 int ret = 0;
115 int was_empty = 0;
116 LIST_HEAD(to_be_dropped); 137 LIST_HEAD(to_be_dropped);
117 138
139restart:
140
118 /* 141 /*
119 * sendmsg calls here after having queued its message on the send 142 * sendmsg calls here after having queued its message on the send
120 * queue. We only have one task feeding the connection at a time. If 143 * queue. We only have one task feeding the connection at a time. If
121 * another thread is already feeding the queue then we back off. This 144 * another thread is already feeding the queue then we back off. This
122 * avoids blocking the caller and trading per-connection data between 145 * avoids blocking the caller and trading per-connection data between
123 * caches per message. 146 * caches per message.
124 *
125 * The sem holder will issue a retry if they notice that someone queued
126 * a message after they stopped walking the send queue but before they
127 * dropped the sem.
128 */ 147 */
129 if (!mutex_trylock(&conn->c_send_lock)) { 148 if (!acquire_in_xmit(conn)) {
130 rds_stats_inc(s_send_sem_contention); 149 rds_stats_inc(s_send_lock_contention);
131 ret = -ENOMEM; 150 ret = -ENOMEM;
132 goto out; 151 goto out;
133 } 152 }
134 153
154 /*
155 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
156 * we do the opposite to avoid races.
157 */
158 if (!rds_conn_up(conn)) {
159 release_in_xmit(conn);
160 ret = 0;
161 goto out;
162 }
163
135 if (conn->c_trans->xmit_prepare) 164 if (conn->c_trans->xmit_prepare)
136 conn->c_trans->xmit_prepare(conn); 165 conn->c_trans->xmit_prepare(conn);
137 166
138 /* 167 /*
139 * spin trying to push headers and data down the connection until 168 * spin trying to push headers and data down the connection until
140 * the connection doens't make forward progress. 169 * the connection doesn't make forward progress.
141 */ 170 */
142 while (--send_quota) { 171 while (1) {
143 /*
144 * See if need to send a congestion map update if we're
145 * between sending messages. The send_sem protects our sole
146 * use of c_map_offset and _bytes.
147 * Note this is used only by transports that define a special
148 * xmit_cong_map function. For all others, we create allocate
149 * a cong_map message and treat it just like any other send.
150 */
151 if (conn->c_map_bytes) {
152 ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
153 conn->c_map_offset);
154 if (ret <= 0)
155 break;
156 172
157 conn->c_map_offset += ret;
158 conn->c_map_bytes -= ret;
159 if (conn->c_map_bytes)
160 continue;
161 }
162
163 /* If we're done sending the current message, clear the
164 * offset and S/G temporaries.
165 */
166 rm = conn->c_xmit_rm; 173 rm = conn->c_xmit_rm;
167 if (rm != NULL &&
168 conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
169 conn->c_xmit_sg == rm->m_nents) {
170 conn->c_xmit_rm = NULL;
171 conn->c_xmit_sg = 0;
172 conn->c_xmit_hdr_off = 0;
173 conn->c_xmit_data_off = 0;
174 conn->c_xmit_rdma_sent = 0;
175 174
176 /* Release the reference to the previous message. */ 175 /*
177 rds_message_put(rm); 176 * If between sending messages, we can send a pending congestion
178 rm = NULL; 177 * map update.
179 }
180
181 /* If we're asked to send a cong map update, do so.
182 */ 178 */
183 if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) { 179 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
184 if (conn->c_trans->xmit_cong_map != NULL) {
185 conn->c_map_offset = 0;
186 conn->c_map_bytes = sizeof(struct rds_header) +
187 RDS_CONG_MAP_BYTES;
188 continue;
189 }
190
191 rm = rds_cong_update_alloc(conn); 180 rm = rds_cong_update_alloc(conn);
192 if (IS_ERR(rm)) { 181 if (IS_ERR(rm)) {
193 ret = PTR_ERR(rm); 182 ret = PTR_ERR(rm);
194 break; 183 break;
195 } 184 }
185 rm->data.op_active = 1;
196 186
197 conn->c_xmit_rm = rm; 187 conn->c_xmit_rm = rm;
198 } 188 }
199 189
200 /* 190 /*
201 * Grab the next message from the send queue, if there is one. 191 * If not already working on one, grab the next message.
202 * 192 *
203 * c_xmit_rm holds a ref while we're sending this message down 193 * c_xmit_rm holds a ref while we're sending this message down
204 * the connction. We can use this ref while holding the 194 * the connction. We can use this ref while holding the
205 * send_sem.. rds_send_reset() is serialized with it. 195 * send_sem.. rds_send_reset() is serialized with it.
206 */ 196 */
207 if (rm == NULL) { 197 if (!rm) {
208 unsigned int len; 198 unsigned int len;
209 199
210 spin_lock_irqsave(&conn->c_lock, flags); 200 spin_lock_irqsave(&conn->c_lock, flags);
@@ -224,10 +214,8 @@ int rds_send_xmit(struct rds_connection *conn)
224 214
225 spin_unlock_irqrestore(&conn->c_lock, flags); 215 spin_unlock_irqrestore(&conn->c_lock, flags);
226 216
227 if (rm == NULL) { 217 if (!rm)
228 was_empty = 1;
229 break; 218 break;
230 }
231 219
232 /* Unfortunately, the way Infiniband deals with 220 /* Unfortunately, the way Infiniband deals with
233 * RDMA to a bad MR key is by moving the entire 221 * RDMA to a bad MR key is by moving the entire
@@ -236,13 +224,12 @@ int rds_send_xmit(struct rds_connection *conn)
236 * connection. 224 * connection.
237 * Therefore, we never retransmit messages with RDMA ops. 225 * Therefore, we never retransmit messages with RDMA ops.
238 */ 226 */
239 if (rm->m_rdma_op && 227 if (rm->rdma.op_active &&
240 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) { 228 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
241 spin_lock_irqsave(&conn->c_lock, flags); 229 spin_lock_irqsave(&conn->c_lock, flags);
242 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) 230 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
243 list_move(&rm->m_conn_item, &to_be_dropped); 231 list_move(&rm->m_conn_item, &to_be_dropped);
244 spin_unlock_irqrestore(&conn->c_lock, flags); 232 spin_unlock_irqrestore(&conn->c_lock, flags);
245 rds_message_put(rm);
246 continue; 233 continue;
247 } 234 }
248 235
@@ -263,23 +250,55 @@ int rds_send_xmit(struct rds_connection *conn)
263 conn->c_xmit_rm = rm; 250 conn->c_xmit_rm = rm;
264 } 251 }
265 252
266 /* 253 /* The transport either sends the whole rdma or none of it */
267 * Try and send an rdma message. Let's see if we can 254 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
268 * keep this simple and require that the transport either 255 rm->m_final_op = &rm->rdma;
269 * send the whole rdma or none of it. 256 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
270 */
271 if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
272 ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
273 if (ret) 257 if (ret)
274 break; 258 break;
275 conn->c_xmit_rdma_sent = 1; 259 conn->c_xmit_rdma_sent = 1;
260
276 /* The transport owns the mapped memory for now. 261 /* The transport owns the mapped memory for now.
277 * You can't unmap it while it's on the send queue */ 262 * You can't unmap it while it's on the send queue */
278 set_bit(RDS_MSG_MAPPED, &rm->m_flags); 263 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
279 } 264 }
280 265
281 if (conn->c_xmit_hdr_off < sizeof(struct rds_header) || 266 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
282 conn->c_xmit_sg < rm->m_nents) { 267 rm->m_final_op = &rm->atomic;
268 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
269 if (ret)
270 break;
271 conn->c_xmit_atomic_sent = 1;
272
273 /* The transport owns the mapped memory for now.
274 * You can't unmap it while it's on the send queue */
275 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
276 }
277
278 /*
279 * A number of cases require an RDS header to be sent
280 * even if there is no data.
281 * We permit 0-byte sends; rds-ping depends on this.
282 * However, if there are exclusively attached silent ops,
283 * we skip the hdr/data send, to enable silent operation.
284 */
285 if (rm->data.op_nents == 0) {
286 int ops_present;
287 int all_ops_are_silent = 1;
288
289 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
290 if (rm->atomic.op_active && !rm->atomic.op_silent)
291 all_ops_are_silent = 0;
292 if (rm->rdma.op_active && !rm->rdma.op_silent)
293 all_ops_are_silent = 0;
294
295 if (ops_present && all_ops_are_silent
296 && !rm->m_rdma_cookie)
297 rm->data.op_active = 0;
298 }
299
300 if (rm->data.op_active && !conn->c_xmit_data_sent) {
301 rm->m_final_op = &rm->data;
283 ret = conn->c_trans->xmit(conn, rm, 302 ret = conn->c_trans->xmit(conn, rm,
284 conn->c_xmit_hdr_off, 303 conn->c_xmit_hdr_off,
285 conn->c_xmit_sg, 304 conn->c_xmit_sg,
@@ -295,7 +314,7 @@ int rds_send_xmit(struct rds_connection *conn)
295 ret -= tmp; 314 ret -= tmp;
296 } 315 }
297 316
298 sg = &rm->m_sg[conn->c_xmit_sg]; 317 sg = &rm->data.op_sg[conn->c_xmit_sg];
299 while (ret) { 318 while (ret) {
300 tmp = min_t(int, ret, sg->length - 319 tmp = min_t(int, ret, sg->length -
301 conn->c_xmit_data_off); 320 conn->c_xmit_data_off);
@@ -306,49 +325,63 @@ int rds_send_xmit(struct rds_connection *conn)
306 sg++; 325 sg++;
307 conn->c_xmit_sg++; 326 conn->c_xmit_sg++;
308 BUG_ON(ret != 0 && 327 BUG_ON(ret != 0 &&
309 conn->c_xmit_sg == rm->m_nents); 328 conn->c_xmit_sg == rm->data.op_nents);
310 } 329 }
311 } 330 }
331
332 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
333 (conn->c_xmit_sg == rm->data.op_nents))
334 conn->c_xmit_data_sent = 1;
312 } 335 }
313 }
314 336
315 /* Nuke any messages we decided not to retransmit. */ 337 /*
316 if (!list_empty(&to_be_dropped)) 338 * A rm will only take multiple times through this loop
317 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); 339 * if there is a data op. Thus, if the data is sent (or there was
340 * none), then we're done with the rm.
341 */
342 if (!rm->data.op_active || conn->c_xmit_data_sent) {
343 conn->c_xmit_rm = NULL;
344 conn->c_xmit_sg = 0;
345 conn->c_xmit_hdr_off = 0;
346 conn->c_xmit_data_off = 0;
347 conn->c_xmit_rdma_sent = 0;
348 conn->c_xmit_atomic_sent = 0;
349 conn->c_xmit_data_sent = 0;
350
351 rds_message_put(rm);
352 }
353 }
318 354
319 if (conn->c_trans->xmit_complete) 355 if (conn->c_trans->xmit_complete)
320 conn->c_trans->xmit_complete(conn); 356 conn->c_trans->xmit_complete(conn);
321 357
322 /* 358 release_in_xmit(conn);
323 * We might be racing with another sender who queued a message but
324 * backed off on noticing that we held the c_send_lock. If we check
325 * for queued messages after dropping the sem then either we'll
326 * see the queued message or the queuer will get the sem. If we
327 * notice the queued message then we trigger an immediate retry.
328 *
329 * We need to be careful only to do this when we stopped processing
330 * the send queue because it was empty. It's the only way we
331 * stop processing the loop when the transport hasn't taken
332 * responsibility for forward progress.
333 */
334 mutex_unlock(&conn->c_send_lock);
335 359
336 if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) { 360 /* Nuke any messages we decided not to retransmit. */
337 /* We exhausted the send quota, but there's work left to 361 if (!list_empty(&to_be_dropped)) {
338 * do. Return and (re-)schedule the send worker. 362 /* irqs on here, so we can put(), unlike above */
339 */ 363 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
340 ret = -EAGAIN; 364 rds_message_put(rm);
365 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
341 } 366 }
342 367
343 if (ret == 0 && was_empty) { 368 /*
344 /* A simple bit test would be way faster than taking the 369 * Other senders can queue a message after we last test the send queue
345 * spin lock */ 370 * but before we clear RDS_IN_XMIT. In that case they'd back off and
346 spin_lock_irqsave(&conn->c_lock, flags); 371 * not try and send their newly queued message. We need to check the
372 * send queue after having cleared RDS_IN_XMIT so that their message
373 * doesn't get stuck on the send queue.
374 *
375 * If the transport cannot continue (i.e ret != 0), then it must
376 * call us when more room is available, such as from the tx
377 * completion handler.
378 */
379 if (ret == 0) {
380 smp_mb();
347 if (!list_empty(&conn->c_send_queue)) { 381 if (!list_empty(&conn->c_send_queue)) {
348 rds_stats_inc(s_send_sem_queue_raced); 382 rds_stats_inc(s_send_lock_queue_raced);
349 ret = -EAGAIN; 383 goto restart;
350 } 384 }
351 spin_unlock_irqrestore(&conn->c_lock, flags);
352 } 385 }
353out: 386out:
354 return ret; 387 return ret;
@@ -376,52 +409,60 @@ static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
376} 409}
377 410
378/* 411/*
379 * Returns true if there are no messages on the send and retransmit queues 412 * This is pretty similar to what happens below in the ACK
380 * which have a sequence number greater than or equal to the given sequence 413 * handling code - except that we call here as soon as we get
381 * number. 414 * the IB send completion on the RDMA op and the accompanying
415 * message.
382 */ 416 */
383int rds_send_acked_before(struct rds_connection *conn, u64 seq) 417void rds_rdma_send_complete(struct rds_message *rm, int status)
384{ 418{
385 struct rds_message *rm, *tmp; 419 struct rds_sock *rs = NULL;
386 int ret = 1; 420 struct rm_rdma_op *ro;
421 struct rds_notifier *notifier;
422 unsigned long flags;
387 423
388 spin_lock(&conn->c_lock); 424 spin_lock_irqsave(&rm->m_rs_lock, flags);
389 425
390 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 426 ro = &rm->rdma;
391 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) 427 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
392 ret = 0; 428 ro->op_active && ro->op_notify && ro->op_notifier) {
393 break; 429 notifier = ro->op_notifier;
394 } 430 rs = rm->m_rs;
431 sock_hold(rds_rs_to_sk(rs));
395 432
396 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { 433 notifier->n_status = status;
397 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq) 434 spin_lock(&rs->rs_lock);
398 ret = 0; 435 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
399 break; 436 spin_unlock(&rs->rs_lock);
437
438 ro->op_notifier = NULL;
400 } 439 }
401 440
402 spin_unlock(&conn->c_lock); 441 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
403 442
404 return ret; 443 if (rs) {
444 rds_wake_sk_sleep(rs);
445 sock_put(rds_rs_to_sk(rs));
446 }
405} 447}
448EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
406 449
407/* 450/*
408 * This is pretty similar to what happens below in the ACK 451 * Just like above, except looks at atomic op
409 * handling code - except that we call here as soon as we get
410 * the IB send completion on the RDMA op and the accompanying
411 * message.
412 */ 452 */
413void rds_rdma_send_complete(struct rds_message *rm, int status) 453void rds_atomic_send_complete(struct rds_message *rm, int status)
414{ 454{
415 struct rds_sock *rs = NULL; 455 struct rds_sock *rs = NULL;
416 struct rds_rdma_op *ro; 456 struct rm_atomic_op *ao;
417 struct rds_notifier *notifier; 457 struct rds_notifier *notifier;
458 unsigned long flags;
418 459
419 spin_lock(&rm->m_rs_lock); 460 spin_lock_irqsave(&rm->m_rs_lock, flags);
420 461
421 ro = rm->m_rdma_op; 462 ao = &rm->atomic;
422 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && 463 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
423 ro && ro->r_notify && ro->r_notifier) { 464 && ao->op_active && ao->op_notify && ao->op_notifier) {
424 notifier = ro->r_notifier; 465 notifier = ao->op_notifier;
425 rs = rm->m_rs; 466 rs = rm->m_rs;
426 sock_hold(rds_rs_to_sk(rs)); 467 sock_hold(rds_rs_to_sk(rs));
427 468
@@ -430,17 +471,17 @@ void rds_rdma_send_complete(struct rds_message *rm, int status)
430 list_add_tail(&notifier->n_list, &rs->rs_notify_queue); 471 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
431 spin_unlock(&rs->rs_lock); 472 spin_unlock(&rs->rs_lock);
432 473
433 ro->r_notifier = NULL; 474 ao->op_notifier = NULL;
434 } 475 }
435 476
436 spin_unlock(&rm->m_rs_lock); 477 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
437 478
438 if (rs) { 479 if (rs) {
439 rds_wake_sk_sleep(rs); 480 rds_wake_sk_sleep(rs);
440 sock_put(rds_rs_to_sk(rs)); 481 sock_put(rds_rs_to_sk(rs));
441 } 482 }
442} 483}
443EXPORT_SYMBOL_GPL(rds_rdma_send_complete); 484EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
444 485
445/* 486/*
446 * This is the same as rds_rdma_send_complete except we 487 * This is the same as rds_rdma_send_complete except we
@@ -448,15 +489,23 @@ EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
448 * socket, socket lock) and can just move the notifier. 489 * socket, socket lock) and can just move the notifier.
449 */ 490 */
450static inline void 491static inline void
451__rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) 492__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
452{ 493{
453 struct rds_rdma_op *ro; 494 struct rm_rdma_op *ro;
495 struct rm_atomic_op *ao;
496
497 ro = &rm->rdma;
498 if (ro->op_active && ro->op_notify && ro->op_notifier) {
499 ro->op_notifier->n_status = status;
500 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
501 ro->op_notifier = NULL;
502 }
454 503
455 ro = rm->m_rdma_op; 504 ao = &rm->atomic;
456 if (ro && ro->r_notify && ro->r_notifier) { 505 if (ao->op_active && ao->op_notify && ao->op_notifier) {
457 ro->r_notifier->n_status = status; 506 ao->op_notifier->n_status = status;
458 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue); 507 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
459 ro->r_notifier = NULL; 508 ao->op_notifier = NULL;
460 } 509 }
461 510
462 /* No need to wake the app - caller does this */ 511 /* 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
468 * So speed is not an issue here. 517 * So speed is not an issue here.
469 */ 518 */
470struct rds_message *rds_send_get_message(struct rds_connection *conn, 519struct rds_message *rds_send_get_message(struct rds_connection *conn,
471 struct rds_rdma_op *op) 520 struct rm_rdma_op *op)
472{ 521{
473 struct rds_message *rm, *tmp, *found = NULL; 522 struct rds_message *rm, *tmp, *found = NULL;
474 unsigned long flags; 523 unsigned long flags;
@@ -476,7 +525,7 @@ struct rds_message *rds_send_get_message(struct rds_connection *conn,
476 spin_lock_irqsave(&conn->c_lock, flags); 525 spin_lock_irqsave(&conn->c_lock, flags);
477 526
478 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) { 527 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
479 if (rm->m_rdma_op == op) { 528 if (&rm->rdma == op) {
480 atomic_inc(&rm->m_refcount); 529 atomic_inc(&rm->m_refcount);
481 found = rm; 530 found = rm;
482 goto out; 531 goto out;
@@ -484,7 +533,7 @@ struct rds_message *rds_send_get_message(struct rds_connection *conn,
484 } 533 }
485 534
486 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) { 535 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
487 if (rm->m_rdma_op == op) { 536 if (&rm->rdma == op) {
488 atomic_inc(&rm->m_refcount); 537 atomic_inc(&rm->m_refcount);
489 found = rm; 538 found = rm;
490 break; 539 break;
@@ -544,19 +593,20 @@ void rds_send_remove_from_sock(struct list_head *messages, int status)
544 spin_lock(&rs->rs_lock); 593 spin_lock(&rs->rs_lock);
545 594
546 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { 595 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
547 struct rds_rdma_op *ro = rm->m_rdma_op; 596 struct rm_rdma_op *ro = &rm->rdma;
548 struct rds_notifier *notifier; 597 struct rds_notifier *notifier;
549 598
550 list_del_init(&rm->m_sock_item); 599 list_del_init(&rm->m_sock_item);
551 rds_send_sndbuf_remove(rs, rm); 600 rds_send_sndbuf_remove(rs, rm);
552 601
553 if (ro && ro->r_notifier && (status || ro->r_notify)) { 602 if (ro->op_active && ro->op_notifier &&
554 notifier = ro->r_notifier; 603 (ro->op_notify || (ro->op_recverr && status))) {
604 notifier = ro->op_notifier;
555 list_add_tail(&notifier->n_list, 605 list_add_tail(&notifier->n_list,
556 &rs->rs_notify_queue); 606 &rs->rs_notify_queue);
557 if (!notifier->n_status) 607 if (!notifier->n_status)
558 notifier->n_status = status; 608 notifier->n_status = status;
559 rm->m_rdma_op->r_notifier = NULL; 609 rm->rdma.op_notifier = NULL;
560 } 610 }
561 was_on_sock = 1; 611 was_on_sock = 1;
562 rm->m_rs = NULL; 612 rm->m_rs = NULL;
@@ -619,9 +669,8 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
619{ 669{
620 struct rds_message *rm, *tmp; 670 struct rds_message *rm, *tmp;
621 struct rds_connection *conn; 671 struct rds_connection *conn;
622 unsigned long flags, flags2; 672 unsigned long flags;
623 LIST_HEAD(list); 673 LIST_HEAD(list);
624 int wake = 0;
625 674
626 /* get all the messages we're dropping under the rs lock */ 675 /* get all the messages we're dropping under the rs lock */
627 spin_lock_irqsave(&rs->rs_lock, flags); 676 spin_lock_irqsave(&rs->rs_lock, flags);
@@ -631,59 +680,54 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
631 dest->sin_port != rm->m_inc.i_hdr.h_dport)) 680 dest->sin_port != rm->m_inc.i_hdr.h_dport))
632 continue; 681 continue;
633 682
634 wake = 1;
635 list_move(&rm->m_sock_item, &list); 683 list_move(&rm->m_sock_item, &list);
636 rds_send_sndbuf_remove(rs, rm); 684 rds_send_sndbuf_remove(rs, rm);
637 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); 685 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
638 } 686 }
639 687
640 /* order flag updates with the rs lock */ 688 /* order flag updates with the rs lock */
641 if (wake) 689 smp_mb__after_clear_bit();
642 smp_mb__after_clear_bit();
643 690
644 spin_unlock_irqrestore(&rs->rs_lock, flags); 691 spin_unlock_irqrestore(&rs->rs_lock, flags);
645 692
646 conn = NULL; 693 if (list_empty(&list))
694 return;
647 695
648 /* now remove the messages from the conn list as needed */ 696 /* Remove the messages from the conn */
649 list_for_each_entry(rm, &list, m_sock_item) { 697 list_for_each_entry(rm, &list, m_sock_item) {
650 /* We do this here rather than in the loop above, so that
651 * we don't have to nest m_rs_lock under rs->rs_lock */
652 spin_lock_irqsave(&rm->m_rs_lock, flags2);
653 /* If this is a RDMA operation, notify the app. */
654 spin_lock(&rs->rs_lock);
655 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
656 spin_unlock(&rs->rs_lock);
657 rm->m_rs = NULL;
658 spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
659 698
699 conn = rm->m_inc.i_conn;
700
701 spin_lock_irqsave(&conn->c_lock, flags);
660 /* 702 /*
661 * If we see this flag cleared then we're *sure* that someone 703 * Maybe someone else beat us to removing rm from the conn.
662 * else beat us to removing it from the conn. If we race 704 * If we race with their flag update we'll get the lock and
663 * with their flag update we'll get the lock and then really 705 * then really see that the flag has been cleared.
664 * see that the flag has been cleared.
665 */ 706 */
666 if (!test_bit(RDS_MSG_ON_CONN, &rm->m_flags)) 707 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
708 spin_unlock_irqrestore(&conn->c_lock, flags);
667 continue; 709 continue;
668
669 if (conn != rm->m_inc.i_conn) {
670 if (conn)
671 spin_unlock_irqrestore(&conn->c_lock, flags);
672 conn = rm->m_inc.i_conn;
673 spin_lock_irqsave(&conn->c_lock, flags);
674 } 710 }
711 list_del_init(&rm->m_conn_item);
712 spin_unlock_irqrestore(&conn->c_lock, flags);
675 713
676 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { 714 /*
677 list_del_init(&rm->m_conn_item); 715 * Couldn't grab m_rs_lock in top loop (lock ordering),
678 rds_message_put(rm); 716 * but we can now.
679 } 717 */
680 } 718 spin_lock_irqsave(&rm->m_rs_lock, flags);
681 719
682 if (conn) 720 spin_lock(&rs->rs_lock);
683 spin_unlock_irqrestore(&conn->c_lock, flags); 721 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
722 spin_unlock(&rs->rs_lock);
684 723
685 if (wake) 724 rm->m_rs = NULL;
686 rds_wake_sk_sleep(rs); 725 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
726
727 rds_message_put(rm);
728 }
729
730 rds_wake_sk_sleep(rs);
687 731
688 while (!list_empty(&list)) { 732 while (!list_empty(&list)) {
689 rm = list_entry(list.next, struct rds_message, m_sock_item); 733 rm = list_entry(list.next, struct rds_message, m_sock_item);
@@ -763,6 +807,63 @@ out:
763 return *queued; 807 return *queued;
764} 808}
765 809
810/*
811 * rds_message is getting to be quite complicated, and we'd like to allocate
812 * it all in one go. This figures out how big it needs to be up front.
813 */
814static int rds_rm_size(struct msghdr *msg, int data_len)
815{
816 struct cmsghdr *cmsg;
817 int size = 0;
818 int cmsg_groups = 0;
819 int retval;
820
821 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
822 if (!CMSG_OK(msg, cmsg))
823 return -EINVAL;
824
825 if (cmsg->cmsg_level != SOL_RDS)
826 continue;
827
828 switch (cmsg->cmsg_type) {
829 case RDS_CMSG_RDMA_ARGS:
830 cmsg_groups |= 1;
831 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
832 if (retval < 0)
833 return retval;
834 size += retval;
835
836 break;
837
838 case RDS_CMSG_RDMA_DEST:
839 case RDS_CMSG_RDMA_MAP:
840 cmsg_groups |= 2;
841 /* these are valid but do no add any size */
842 break;
843
844 case RDS_CMSG_ATOMIC_CSWP:
845 case RDS_CMSG_ATOMIC_FADD:
846 case RDS_CMSG_MASKED_ATOMIC_CSWP:
847 case RDS_CMSG_MASKED_ATOMIC_FADD:
848 cmsg_groups |= 1;
849 size += sizeof(struct scatterlist);
850 break;
851
852 default:
853 return -EINVAL;
854 }
855
856 }
857
858 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
859
860 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
861 if (cmsg_groups == 3)
862 return -EINVAL;
863
864 return size;
865}
866
766static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, 867static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
767 struct msghdr *msg, int *allocated_mr) 868 struct msghdr *msg, int *allocated_mr)
768{ 869{
@@ -777,7 +878,7 @@ static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
777 continue; 878 continue;
778 879
779 /* As a side effect, RDMA_DEST and RDMA_MAP will set 880 /* As a side effect, RDMA_DEST and RDMA_MAP will set
780 * rm->m_rdma_cookie and rm->m_rdma_mr. 881 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
781 */ 882 */
782 switch (cmsg->cmsg_type) { 883 switch (cmsg->cmsg_type) {
783 case RDS_CMSG_RDMA_ARGS: 884 case RDS_CMSG_RDMA_ARGS:
@@ -793,6 +894,12 @@ static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
793 if (!ret) 894 if (!ret)
794 *allocated_mr = 1; 895 *allocated_mr = 1;
795 break; 896 break;
897 case RDS_CMSG_ATOMIC_CSWP:
898 case RDS_CMSG_ATOMIC_FADD:
899 case RDS_CMSG_MASKED_ATOMIC_CSWP:
900 case RDS_CMSG_MASKED_ATOMIC_FADD:
901 ret = rds_cmsg_atomic(rs, rm, cmsg);
902 break;
796 903
797 default: 904 default:
798 return -EINVAL; 905 return -EINVAL;
@@ -850,13 +957,26 @@ int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
850 goto out; 957 goto out;
851 } 958 }
852 959
853 rm = rds_message_copy_from_user(msg->msg_iov, payload_len); 960 /* size of rm including all sgs */
854 if (IS_ERR(rm)) { 961 ret = rds_rm_size(msg, payload_len);
855 ret = PTR_ERR(rm); 962 if (ret < 0)
856 rm = NULL; 963 goto out;
964
965 rm = rds_message_alloc(ret, GFP_KERNEL);
966 if (!rm) {
967 ret = -ENOMEM;
857 goto out; 968 goto out;
858 } 969 }
859 970
971 /* Attach data to the rm */
972 if (payload_len) {
973 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
974 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
975 if (ret)
976 goto out;
977 }
978 rm->data.op_active = 1;
979
860 rm->m_daddr = daddr; 980 rm->m_daddr = daddr;
861 981
862 /* rds_conn_create has a spinlock that runs with IRQ off. 982 /* 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,
879 if (ret) 999 if (ret)
880 goto out; 1000 goto out;
881 1001
882 if ((rm->m_rdma_cookie || rm->m_rdma_op) && 1002 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
883 conn->c_trans->xmit_rdma == NULL) {
884 if (printk_ratelimit()) 1003 if (printk_ratelimit())
885 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", 1004 printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
886 rm->m_rdma_op, conn->c_trans->xmit_rdma); 1005 &rm->rdma, conn->c_trans->xmit_rdma);
887 ret = -EOPNOTSUPP; 1006 ret = -EOPNOTSUPP;
888 goto out; 1007 goto out;
889 } 1008 }
890 1009
891 /* If the connection is down, trigger a connect. We may 1010 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
892 * have scheduled a delayed reconnect however - in this case 1011 if (printk_ratelimit())
893 * we should not interfere. 1012 printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
894 */ 1013 &rm->atomic, conn->c_trans->xmit_atomic);
895 if (rds_conn_state(conn) == RDS_CONN_DOWN && 1014 ret = -EOPNOTSUPP;
896 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 1015 goto out;
897 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 1016 }
1017
1018 rds_conn_connect_if_down(conn);
898 1019
899 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); 1020 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
900 if (ret) { 1021 if (ret) {
@@ -938,7 +1059,7 @@ int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
938 rds_stats_inc(s_send_queued); 1059 rds_stats_inc(s_send_queued);
939 1060
940 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags)) 1061 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
941 rds_send_worker(&conn->c_send_w.work); 1062 rds_send_xmit(conn);
942 1063
943 rds_message_put(rm); 1064 rds_message_put(rm);
944 return payload_len; 1065 return payload_len;
@@ -966,20 +1087,15 @@ rds_send_pong(struct rds_connection *conn, __be16 dport)
966 int ret = 0; 1087 int ret = 0;
967 1088
968 rm = rds_message_alloc(0, GFP_ATOMIC); 1089 rm = rds_message_alloc(0, GFP_ATOMIC);
969 if (rm == NULL) { 1090 if (!rm) {
970 ret = -ENOMEM; 1091 ret = -ENOMEM;
971 goto out; 1092 goto out;
972 } 1093 }
973 1094
974 rm->m_daddr = conn->c_faddr; 1095 rm->m_daddr = conn->c_faddr;
1096 rm->data.op_active = 1;
975 1097
976 /* If the connection is down, trigger a connect. We may 1098 rds_conn_connect_if_down(conn);
977 * have scheduled a delayed reconnect however - in this case
978 * we should not interfere.
979 */
980 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
981 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
982 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
983 1099
984 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL); 1100 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
985 if (ret) 1101 if (ret)
@@ -999,7 +1115,9 @@ rds_send_pong(struct rds_connection *conn, __be16 dport)
999 rds_stats_inc(s_send_queued); 1115 rds_stats_inc(s_send_queued);
1000 rds_stats_inc(s_send_pong); 1116 rds_stats_inc(s_send_pong);
1001 1117
1002 queue_delayed_work(rds_wq, &conn->c_send_w, 0); 1118 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1119 rds_send_xmit(conn);
1120
1003 rds_message_put(rm); 1121 rds_message_put(rm);
1004 return 0; 1122 return 0;
1005 1123
diff --git a/net/rds/stats.c b/net/rds/stats.c
index 7598eb07cfb1..10c759ccac0c 100644
--- a/net/rds/stats.c
+++ b/net/rds/stats.c
@@ -57,8 +57,8 @@ static const char *const rds_stat_names[] = {
57 "recv_ping", 57 "recv_ping",
58 "send_queue_empty", 58 "send_queue_empty",
59 "send_queue_full", 59 "send_queue_full",
60 "send_sem_contention", 60 "send_lock_contention",
61 "send_sem_queue_raced", 61 "send_lock_queue_raced",
62 "send_immediate_retry", 62 "send_immediate_retry",
63 "send_delayed_retry", 63 "send_delayed_retry",
64 "send_drop_acked", 64 "send_drop_acked",
@@ -143,7 +143,7 @@ void rds_stats_exit(void)
143 rds_info_deregister_func(RDS_INFO_COUNTERS, rds_stats_info); 143 rds_info_deregister_func(RDS_INFO_COUNTERS, rds_stats_info);
144} 144}
145 145
146int __init rds_stats_init(void) 146int rds_stats_init(void)
147{ 147{
148 rds_info_register_func(RDS_INFO_COUNTERS, rds_stats_info); 148 rds_info_register_func(RDS_INFO_COUNTERS, rds_stats_info);
149 return 0; 149 return 0;
diff --git a/net/rds/sysctl.c b/net/rds/sysctl.c
index 7829a20325d3..25ad0c77a26c 100644
--- a/net/rds/sysctl.c
+++ b/net/rds/sysctl.c
@@ -105,13 +105,13 @@ void rds_sysctl_exit(void)
105 unregister_sysctl_table(rds_sysctl_reg_table); 105 unregister_sysctl_table(rds_sysctl_reg_table);
106} 106}
107 107
108int __init rds_sysctl_init(void) 108int rds_sysctl_init(void)
109{ 109{
110 rds_sysctl_reconnect_min = msecs_to_jiffies(1); 110 rds_sysctl_reconnect_min = msecs_to_jiffies(1);
111 rds_sysctl_reconnect_min_jiffies = rds_sysctl_reconnect_min; 111 rds_sysctl_reconnect_min_jiffies = rds_sysctl_reconnect_min;
112 112
113 rds_sysctl_reg_table = register_sysctl_paths(rds_sysctl_path, rds_sysctl_rds_table); 113 rds_sysctl_reg_table = register_sysctl_paths(rds_sysctl_path, rds_sysctl_rds_table);
114 if (rds_sysctl_reg_table == NULL) 114 if (!rds_sysctl_reg_table)
115 return -ENOMEM; 115 return -ENOMEM;
116 return 0; 116 return 0;
117} 117}
diff --git a/net/rds/tcp.c b/net/rds/tcp.c
index babf4577ff7d..eeb08e6ab96b 100644
--- a/net/rds/tcp.c
+++ b/net/rds/tcp.c
@@ -200,7 +200,7 @@ static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
200 struct rds_tcp_connection *tc; 200 struct rds_tcp_connection *tc;
201 201
202 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp); 202 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
203 if (tc == NULL) 203 if (!tc)
204 return -ENOMEM; 204 return -ENOMEM;
205 205
206 tc->t_sock = NULL; 206 tc->t_sock = NULL;
@@ -258,7 +258,6 @@ struct rds_transport rds_tcp_transport = {
258 .laddr_check = rds_tcp_laddr_check, 258 .laddr_check = rds_tcp_laddr_check,
259 .xmit_prepare = rds_tcp_xmit_prepare, 259 .xmit_prepare = rds_tcp_xmit_prepare,
260 .xmit_complete = rds_tcp_xmit_complete, 260 .xmit_complete = rds_tcp_xmit_complete,
261 .xmit_cong_map = rds_tcp_xmit_cong_map,
262 .xmit = rds_tcp_xmit, 261 .xmit = rds_tcp_xmit,
263 .recv = rds_tcp_recv, 262 .recv = rds_tcp_recv,
264 .conn_alloc = rds_tcp_conn_alloc, 263 .conn_alloc = rds_tcp_conn_alloc,
@@ -266,7 +265,6 @@ struct rds_transport rds_tcp_transport = {
266 .conn_connect = rds_tcp_conn_connect, 265 .conn_connect = rds_tcp_conn_connect,
267 .conn_shutdown = rds_tcp_conn_shutdown, 266 .conn_shutdown = rds_tcp_conn_shutdown,
268 .inc_copy_to_user = rds_tcp_inc_copy_to_user, 267 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
269 .inc_purge = rds_tcp_inc_purge,
270 .inc_free = rds_tcp_inc_free, 268 .inc_free = rds_tcp_inc_free,
271 .stats_info_copy = rds_tcp_stats_info_copy, 269 .stats_info_copy = rds_tcp_stats_info_copy,
272 .exit = rds_tcp_exit, 270 .exit = rds_tcp_exit,
@@ -276,14 +274,14 @@ struct rds_transport rds_tcp_transport = {
276 .t_prefer_loopback = 1, 274 .t_prefer_loopback = 1,
277}; 275};
278 276
279int __init rds_tcp_init(void) 277int rds_tcp_init(void)
280{ 278{
281 int ret; 279 int ret;
282 280
283 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection", 281 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
284 sizeof(struct rds_tcp_connection), 282 sizeof(struct rds_tcp_connection),
285 0, 0, NULL); 283 0, 0, NULL);
286 if (rds_tcp_conn_slab == NULL) { 284 if (!rds_tcp_conn_slab) {
287 ret = -ENOMEM; 285 ret = -ENOMEM;
288 goto out; 286 goto out;
289 } 287 }
diff --git a/net/rds/tcp.h b/net/rds/tcp.h
index 844fa6b9cf5a..f5e6f7bebb50 100644
--- a/net/rds/tcp.h
+++ b/net/rds/tcp.h
@@ -43,7 +43,7 @@ struct rds_tcp_statistics {
43}; 43};
44 44
45/* tcp.c */ 45/* tcp.c */
46int __init rds_tcp_init(void); 46int rds_tcp_init(void);
47void rds_tcp_exit(void); 47void rds_tcp_exit(void);
48void rds_tcp_tune(struct socket *sock); 48void rds_tcp_tune(struct socket *sock);
49void rds_tcp_nonagle(struct socket *sock); 49void rds_tcp_nonagle(struct socket *sock);
@@ -61,16 +61,15 @@ void rds_tcp_conn_shutdown(struct rds_connection *conn);
61void rds_tcp_state_change(struct sock *sk); 61void rds_tcp_state_change(struct sock *sk);
62 62
63/* tcp_listen.c */ 63/* tcp_listen.c */
64int __init rds_tcp_listen_init(void); 64int rds_tcp_listen_init(void);
65void rds_tcp_listen_stop(void); 65void rds_tcp_listen_stop(void);
66void rds_tcp_listen_data_ready(struct sock *sk, int bytes); 66void rds_tcp_listen_data_ready(struct sock *sk, int bytes);
67 67
68/* tcp_recv.c */ 68/* tcp_recv.c */
69int __init rds_tcp_recv_init(void); 69int rds_tcp_recv_init(void);
70void rds_tcp_recv_exit(void); 70void rds_tcp_recv_exit(void);
71void rds_tcp_data_ready(struct sock *sk, int bytes); 71void rds_tcp_data_ready(struct sock *sk, int bytes);
72int rds_tcp_recv(struct rds_connection *conn); 72int rds_tcp_recv(struct rds_connection *conn);
73void rds_tcp_inc_purge(struct rds_incoming *inc);
74void rds_tcp_inc_free(struct rds_incoming *inc); 73void rds_tcp_inc_free(struct rds_incoming *inc);
75int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov, 74int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
76 size_t size); 75 size_t size);
@@ -81,8 +80,6 @@ void rds_tcp_xmit_complete(struct rds_connection *conn);
81int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm, 80int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
82 unsigned int hdr_off, unsigned int sg, unsigned int off); 81 unsigned int hdr_off, unsigned int sg, unsigned int off);
83void rds_tcp_write_space(struct sock *sk); 82void rds_tcp_write_space(struct sock *sk);
84int rds_tcp_xmit_cong_map(struct rds_connection *conn,
85 struct rds_cong_map *map, unsigned long offset);
86 83
87/* tcp_stats.c */ 84/* tcp_stats.c */
88DECLARE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats); 85DECLARE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats);
diff --git a/net/rds/tcp_connect.c b/net/rds/tcp_connect.c
index c397524c039c..a65ee78db0c5 100644
--- a/net/rds/tcp_connect.c
+++ b/net/rds/tcp_connect.c
@@ -45,7 +45,7 @@ void rds_tcp_state_change(struct sock *sk)
45 45
46 read_lock(&sk->sk_callback_lock); 46 read_lock(&sk->sk_callback_lock);
47 conn = sk->sk_user_data; 47 conn = sk->sk_user_data;
48 if (conn == NULL) { 48 if (!conn) {
49 state_change = sk->sk_state_change; 49 state_change = sk->sk_state_change;
50 goto out; 50 goto out;
51 } 51 }
diff --git a/net/rds/tcp_listen.c b/net/rds/tcp_listen.c
index 975183fe6950..ae27869dfc21 100644
--- a/net/rds/tcp_listen.c
+++ b/net/rds/tcp_listen.c
@@ -116,7 +116,7 @@ void rds_tcp_listen_data_ready(struct sock *sk, int bytes)
116 116
117 read_lock(&sk->sk_callback_lock); 117 read_lock(&sk->sk_callback_lock);
118 ready = sk->sk_user_data; 118 ready = sk->sk_user_data;
119 if (ready == NULL) { /* check for teardown race */ 119 if (!ready) { /* check for teardown race */
120 ready = sk->sk_data_ready; 120 ready = sk->sk_data_ready;
121 goto out; 121 goto out;
122 } 122 }
@@ -135,7 +135,7 @@ out:
135 ready(sk, bytes); 135 ready(sk, bytes);
136} 136}
137 137
138int __init rds_tcp_listen_init(void) 138int rds_tcp_listen_init(void)
139{ 139{
140 struct sockaddr_in sin; 140 struct sockaddr_in sin;
141 struct socket *sock = NULL; 141 struct socket *sock = NULL;
@@ -178,7 +178,7 @@ void rds_tcp_listen_stop(void)
178 struct socket *sock = rds_tcp_listen_sock; 178 struct socket *sock = rds_tcp_listen_sock;
179 struct sock *sk; 179 struct sock *sk;
180 180
181 if (sock == NULL) 181 if (!sock)
182 return; 182 return;
183 183
184 sk = sock->sk; 184 sk = sock->sk;
diff --git a/net/rds/tcp_recv.c b/net/rds/tcp_recv.c
index 1aba6878fa5d..7017f3af80b6 100644
--- a/net/rds/tcp_recv.c
+++ b/net/rds/tcp_recv.c
@@ -39,7 +39,7 @@
39 39
40static struct kmem_cache *rds_tcp_incoming_slab; 40static struct kmem_cache *rds_tcp_incoming_slab;
41 41
42void rds_tcp_inc_purge(struct rds_incoming *inc) 42static void rds_tcp_inc_purge(struct rds_incoming *inc)
43{ 43{
44 struct rds_tcp_incoming *tinc; 44 struct rds_tcp_incoming *tinc;
45 tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); 45 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,
190 * processing. 190 * processing.
191 */ 191 */
192 while (left) { 192 while (left) {
193 if (tinc == NULL) { 193 if (!tinc) {
194 tinc = kmem_cache_alloc(rds_tcp_incoming_slab, 194 tinc = kmem_cache_alloc(rds_tcp_incoming_slab,
195 arg->gfp); 195 arg->gfp);
196 if (tinc == NULL) { 196 if (!tinc) {
197 desc->error = -ENOMEM; 197 desc->error = -ENOMEM;
198 goto out; 198 goto out;
199 } 199 }
@@ -229,7 +229,7 @@ static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb,
229 229
230 if (left && tc->t_tinc_data_rem) { 230 if (left && tc->t_tinc_data_rem) {
231 clone = skb_clone(skb, arg->gfp); 231 clone = skb_clone(skb, arg->gfp);
232 if (clone == NULL) { 232 if (!clone) {
233 desc->error = -ENOMEM; 233 desc->error = -ENOMEM;
234 goto out; 234 goto out;
235 } 235 }
@@ -326,7 +326,7 @@ void rds_tcp_data_ready(struct sock *sk, int bytes)
326 326
327 read_lock(&sk->sk_callback_lock); 327 read_lock(&sk->sk_callback_lock);
328 conn = sk->sk_user_data; 328 conn = sk->sk_user_data;
329 if (conn == NULL) { /* check for teardown race */ 329 if (!conn) { /* check for teardown race */
330 ready = sk->sk_data_ready; 330 ready = sk->sk_data_ready;
331 goto out; 331 goto out;
332 } 332 }
@@ -342,12 +342,12 @@ out:
342 ready(sk, bytes); 342 ready(sk, bytes);
343} 343}
344 344
345int __init rds_tcp_recv_init(void) 345int rds_tcp_recv_init(void)
346{ 346{
347 rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming", 347 rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming",
348 sizeof(struct rds_tcp_incoming), 348 sizeof(struct rds_tcp_incoming),
349 0, 0, NULL); 349 0, 0, NULL);
350 if (rds_tcp_incoming_slab == NULL) 350 if (!rds_tcp_incoming_slab)
351 return -ENOMEM; 351 return -ENOMEM;
352 return 0; 352 return 0;
353} 353}
diff --git a/net/rds/tcp_send.c b/net/rds/tcp_send.c
index a28b895ff0d1..2979fb4a4b9a 100644
--- a/net/rds/tcp_send.c
+++ b/net/rds/tcp_send.c
@@ -77,56 +77,6 @@ int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
77} 77}
78 78
79/* the core send_sem serializes this with other xmit and shutdown */ 79/* the core send_sem serializes this with other xmit and shutdown */
80int rds_tcp_xmit_cong_map(struct rds_connection *conn,
81 struct rds_cong_map *map, unsigned long offset)
82{
83 static struct rds_header rds_tcp_map_header = {
84 .h_flags = RDS_FLAG_CONG_BITMAP,
85 };
86 struct rds_tcp_connection *tc = conn->c_transport_data;
87 unsigned long i;
88 int ret;
89 int copied = 0;
90
91 /* Some problem claims cpu_to_be32(constant) isn't a constant. */
92 rds_tcp_map_header.h_len = cpu_to_be32(RDS_CONG_MAP_BYTES);
93
94 if (offset < sizeof(struct rds_header)) {
95 ret = rds_tcp_sendmsg(tc->t_sock,
96 (void *)&rds_tcp_map_header + offset,
97 sizeof(struct rds_header) - offset);
98 if (ret <= 0)
99 return ret;
100 offset += ret;
101 copied = ret;
102 if (offset < sizeof(struct rds_header))
103 return ret;
104 }
105
106 offset -= sizeof(struct rds_header);
107 i = offset / PAGE_SIZE;
108 offset = offset % PAGE_SIZE;
109 BUG_ON(i >= RDS_CONG_MAP_PAGES);
110
111 do {
112 ret = tc->t_sock->ops->sendpage(tc->t_sock,
113 virt_to_page(map->m_page_addrs[i]),
114 offset, PAGE_SIZE - offset,
115 MSG_DONTWAIT);
116 if (ret <= 0)
117 break;
118 copied += ret;
119 offset += ret;
120 if (offset == PAGE_SIZE) {
121 offset = 0;
122 i++;
123 }
124 } while (i < RDS_CONG_MAP_PAGES);
125
126 return copied ? copied : ret;
127}
128
129/* the core send_sem serializes this with other xmit and shutdown */
130int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm, 80int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
131 unsigned int hdr_off, unsigned int sg, unsigned int off) 81 unsigned int hdr_off, unsigned int sg, unsigned int off)
132{ 82{
@@ -166,21 +116,21 @@ int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
166 goto out; 116 goto out;
167 } 117 }
168 118
169 while (sg < rm->m_nents) { 119 while (sg < rm->data.op_nents) {
170 ret = tc->t_sock->ops->sendpage(tc->t_sock, 120 ret = tc->t_sock->ops->sendpage(tc->t_sock,
171 sg_page(&rm->m_sg[sg]), 121 sg_page(&rm->data.op_sg[sg]),
172 rm->m_sg[sg].offset + off, 122 rm->data.op_sg[sg].offset + off,
173 rm->m_sg[sg].length - off, 123 rm->data.op_sg[sg].length - off,
174 MSG_DONTWAIT|MSG_NOSIGNAL); 124 MSG_DONTWAIT|MSG_NOSIGNAL);
175 rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->m_sg[sg]), 125 rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
176 rm->m_sg[sg].offset + off, rm->m_sg[sg].length - off, 126 rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
177 ret); 127 ret);
178 if (ret <= 0) 128 if (ret <= 0)
179 break; 129 break;
180 130
181 off += ret; 131 off += ret;
182 done += ret; 132 done += ret;
183 if (off == rm->m_sg[sg].length) { 133 if (off == rm->data.op_sg[sg].length) {
184 off = 0; 134 off = 0;
185 sg++; 135 sg++;
186 } 136 }
@@ -226,7 +176,7 @@ void rds_tcp_write_space(struct sock *sk)
226 176
227 read_lock(&sk->sk_callback_lock); 177 read_lock(&sk->sk_callback_lock);
228 conn = sk->sk_user_data; 178 conn = sk->sk_user_data;
229 if (conn == NULL) { 179 if (!conn) {
230 write_space = sk->sk_write_space; 180 write_space = sk->sk_write_space;
231 goto out; 181 goto out;
232 } 182 }
diff --git a/net/rds/threads.c b/net/rds/threads.c
index 786c20eaaf5e..0fd90f8c5f59 100644
--- a/net/rds/threads.c
+++ b/net/rds/threads.c
@@ -61,7 +61,7 @@
61 * 61 *
62 * Transition to state DISCONNECTING/DOWN: 62 * Transition to state DISCONNECTING/DOWN:
63 * - Inside the shutdown worker; synchronizes with xmit path 63 * - Inside the shutdown worker; synchronizes with xmit path
64 * through c_send_lock, and with connection management callbacks 64 * through RDS_IN_XMIT, and with connection management callbacks
65 * via c_cm_lock. 65 * via c_cm_lock.
66 * 66 *
67 * For receive callbacks, we rely on the underlying transport 67 * For receive callbacks, we rely on the underlying transport
@@ -110,7 +110,7 @@ EXPORT_SYMBOL_GPL(rds_connect_complete);
110 * We should *always* start with a random backoff; otherwise a broken connection 110 * We should *always* start with a random backoff; otherwise a broken connection
111 * will always take several iterations to be re-established. 111 * will always take several iterations to be re-established.
112 */ 112 */
113static void rds_queue_reconnect(struct rds_connection *conn) 113void rds_queue_reconnect(struct rds_connection *conn)
114{ 114{
115 unsigned long rand; 115 unsigned long rand;
116 116
@@ -156,58 +156,6 @@ void rds_connect_worker(struct work_struct *work)
156 } 156 }
157} 157}
158 158
159void rds_shutdown_worker(struct work_struct *work)
160{
161 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
162
163 /* shut it down unless it's down already */
164 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
165 /*
166 * Quiesce the connection mgmt handlers before we start tearing
167 * things down. We don't hold the mutex for the entire
168 * duration of the shutdown operation, else we may be
169 * deadlocking with the CM handler. Instead, the CM event
170 * handler is supposed to check for state DISCONNECTING
171 */
172 mutex_lock(&conn->c_cm_lock);
173 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING) &&
174 !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
175 rds_conn_error(conn, "shutdown called in state %d\n",
176 atomic_read(&conn->c_state));
177 mutex_unlock(&conn->c_cm_lock);
178 return;
179 }
180 mutex_unlock(&conn->c_cm_lock);
181
182 mutex_lock(&conn->c_send_lock);
183 conn->c_trans->conn_shutdown(conn);
184 rds_conn_reset(conn);
185 mutex_unlock(&conn->c_send_lock);
186
187 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
188 /* This can happen - eg when we're in the middle of tearing
189 * down the connection, and someone unloads the rds module.
190 * Quite reproduceable with loopback connections.
191 * Mostly harmless.
192 */
193 rds_conn_error(conn,
194 "%s: failed to transition to state DOWN, "
195 "current state is %d\n",
196 __func__,
197 atomic_read(&conn->c_state));
198 return;
199 }
200 }
201
202 /* Then reconnect if it's still live.
203 * The passive side of an IB loopback connection is never added
204 * to the conn hash, so we never trigger a reconnect on this
205 * conn - the reconnect is always triggered by the active peer. */
206 cancel_delayed_work(&conn->c_conn_w);
207 if (!hlist_unhashed(&conn->c_hash_node))
208 rds_queue_reconnect(conn);
209}
210
211void rds_send_worker(struct work_struct *work) 159void rds_send_worker(struct work_struct *work)
212{ 160{
213 struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work); 161 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)
252 } 200 }
253} 201}
254 202
203void rds_shutdown_worker(struct work_struct *work)
204{
205 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
206
207 rds_conn_shutdown(conn);
208}
209
255void rds_threads_exit(void) 210void rds_threads_exit(void)
256{ 211{
257 destroy_workqueue(rds_wq); 212 destroy_workqueue(rds_wq);
258} 213}
259 214
260int __init rds_threads_init(void) 215int rds_threads_init(void)
261{ 216{
262 rds_wq = create_workqueue("krdsd"); 217 rds_wq = create_singlethread_workqueue("krdsd");
263 if (rds_wq == NULL) 218 if (!rds_wq)
264 return -ENOMEM; 219 return -ENOMEM;
265 220
266 return 0; 221 return 0;
diff --git a/net/rds/transport.c b/net/rds/transport.c
index 7e1067901353..7f2ac4fec367 100644
--- a/net/rds/transport.c
+++ b/net/rds/transport.c
@@ -71,19 +71,28 @@ void rds_trans_unregister(struct rds_transport *trans)
71} 71}
72EXPORT_SYMBOL_GPL(rds_trans_unregister); 72EXPORT_SYMBOL_GPL(rds_trans_unregister);
73 73
74void rds_trans_put(struct rds_transport *trans)
75{
76 if (trans && trans->t_owner)
77 module_put(trans->t_owner);
78}
79
74struct rds_transport *rds_trans_get_preferred(__be32 addr) 80struct rds_transport *rds_trans_get_preferred(__be32 addr)
75{ 81{
76 struct rds_transport *ret = NULL; 82 struct rds_transport *ret = NULL;
77 int i; 83 struct rds_transport *trans;
84 unsigned int i;
78 85
79 if (IN_LOOPBACK(ntohl(addr))) 86 if (IN_LOOPBACK(ntohl(addr)))
80 return &rds_loop_transport; 87 return &rds_loop_transport;
81 88
82 down_read(&rds_trans_sem); 89 down_read(&rds_trans_sem);
83 for (i = 0; i < RDS_TRANS_COUNT; i++) 90 for (i = 0; i < RDS_TRANS_COUNT; i++) {
84 { 91 trans = transports[i];
85 if (transports[i] && (transports[i]->laddr_check(addr) == 0)) { 92
86 ret = transports[i]; 93 if (trans && (trans->laddr_check(addr) == 0) &&
94 (!trans->t_owner || try_module_get(trans->t_owner))) {
95 ret = trans;
87 break; 96 break;
88 } 97 }
89 } 98 }
diff --git a/net/rds/xlist.h b/net/rds/xlist.h
new file mode 100644
index 000000000000..e6b5190daddd
--- /dev/null
+++ b/net/rds/xlist.h
@@ -0,0 +1,80 @@
1#ifndef _LINUX_XLIST_H
2#define _LINUX_XLIST_H
3
4#include <linux/stddef.h>
5#include <linux/poison.h>
6#include <linux/prefetch.h>
7#include <asm/system.h>
8
9struct xlist_head {
10 struct xlist_head *next;
11};
12
13static inline void INIT_XLIST_HEAD(struct xlist_head *list)
14{
15 list->next = NULL;
16}
17
18static inline int xlist_empty(struct xlist_head *head)
19{
20 return head->next == NULL;
21}
22
23static inline void xlist_add(struct xlist_head *new, struct xlist_head *tail,
24 struct xlist_head *head)
25{
26 struct xlist_head *cur;
27 struct xlist_head *check;
28
29 while (1) {
30 cur = head->next;
31 tail->next = cur;
32 check = cmpxchg(&head->next, cur, new);
33 if (check == cur)
34 break;
35 }
36}
37
38static inline struct xlist_head *xlist_del_head(struct xlist_head *head)
39{
40 struct xlist_head *cur;
41 struct xlist_head *check;
42 struct xlist_head *next;
43
44 while (1) {
45 cur = head->next;
46 if (!cur)
47 goto out;
48
49 next = cur->next;
50 check = cmpxchg(&head->next, cur, next);
51 if (check == cur)
52 goto out;
53 }
54out:
55 return cur;
56}
57
58static inline struct xlist_head *xlist_del_head_fast(struct xlist_head *head)
59{
60 struct xlist_head *cur;
61
62 cur = head->next;
63 if (!cur)
64 return NULL;
65
66 head->next = cur->next;
67 return cur;
68}
69
70static inline void xlist_splice(struct xlist_head *list,
71 struct xlist_head *head)
72{
73 struct xlist_head *cur;
74
75 WARN_ON(head->next);
76 cur = xchg(&list->next, NULL);
77 head->next = cur;
78}
79
80#endif
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-25 23:40:49 -0500