2 * Copyright (c) 2006 Oracle. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36 #include <linux/export.h>
37 #include <net/inet_hashtables.h>
42 #define RDS_CONNECTION_HASH_BITS 12
43 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
44 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
46 /* converting this to RCU is a chore for another day.. */
47 static DEFINE_SPINLOCK(rds_conn_lock);
48 static unsigned long rds_conn_count;
49 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
50 static struct kmem_cache *rds_conn_slab;
52 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
54 static u32 rds_hash_secret __read_mostly;
58 net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
60 /* Pass NULL, don't need struct net for hash */
61 hash = __inet_ehashfn(be32_to_cpu(laddr), 0,
62 be32_to_cpu(faddr), 0,
64 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
67 #define rds_conn_info_set(var, test, suffix) do { \
69 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
72 /* rcu read lock must be held or the connection spinlock */
73 static struct rds_connection *rds_conn_lookup(struct net *net,
74 struct hlist_head *head,
75 __be32 laddr, __be32 faddr,
76 struct rds_transport *trans)
78 struct rds_connection *conn, *ret = NULL;
80 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
81 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
82 conn->c_trans == trans && net == rds_conn_net(conn)) {
87 rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
93 * This is called by transports as they're bringing down a connection.
94 * It clears partial message state so that the transport can start sending
95 * and receiving over this connection again in the future. It is up to
96 * the transport to have serialized this call with its send and recv.
98 static void rds_conn_reset(struct rds_connection *conn)
100 rdsdebug("connection %pI4 to %pI4 reset\n",
101 &conn->c_laddr, &conn->c_faddr);
103 rds_stats_inc(s_conn_reset);
104 rds_send_reset(conn);
107 /* Do not clear next_rx_seq here, else we cannot distinguish
108 * retransmitted packets from new packets, and will hand all
109 * of them to the application. That is not consistent with the
110 * reliability guarantees of RDS. */
114 * There is only every one 'conn' for a given pair of addresses in the
115 * system at a time. They contain messages to be retransmitted and so
116 * span the lifetime of the actual underlying transport connections.
118 * For now they are not garbage collected once they're created. They
119 * are torn down as the module is removed, if ever.
121 static struct rds_connection *__rds_conn_create(struct net *net,
122 __be32 laddr, __be32 faddr,
123 struct rds_transport *trans, gfp_t gfp,
126 struct rds_connection *conn, *parent = NULL;
127 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
128 struct rds_transport *loop_trans;
131 struct rds_transport *otrans = trans;
133 if (!is_outgoing && otrans->t_type == RDS_TRANS_TCP)
136 conn = rds_conn_lookup(net, head, laddr, faddr, trans);
137 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
138 laddr == faddr && !is_outgoing) {
139 /* This is a looped back IB connection, and we're
140 * called by the code handling the incoming connect.
141 * We need a second connection object into which we
142 * can stick the other QP. */
144 conn = parent->c_passive;
151 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
153 conn = ERR_PTR(-ENOMEM);
157 INIT_HLIST_NODE(&conn->c_hash_node);
158 conn->c_laddr = laddr;
159 conn->c_faddr = faddr;
160 spin_lock_init(&conn->c_lock);
161 conn->c_next_tx_seq = 1;
162 rds_conn_net_set(conn, net);
164 init_waitqueue_head(&conn->c_waitq);
165 INIT_LIST_HEAD(&conn->c_send_queue);
166 INIT_LIST_HEAD(&conn->c_retrans);
168 ret = rds_cong_get_maps(conn);
170 kmem_cache_free(rds_conn_slab, conn);
176 * This is where a connection becomes loopback. If *any* RDS sockets
177 * can bind to the destination address then we'd rather the messages
178 * flow through loopback rather than either transport.
180 loop_trans = rds_trans_get_preferred(net, faddr);
182 rds_trans_put(loop_trans);
183 conn->c_loopback = 1;
184 if (is_outgoing && trans->t_prefer_loopback) {
185 /* "outgoing" connection - and the transport
186 * says it wants the connection handled by the
187 * loopback transport. This is what TCP does.
189 trans = &rds_loop_transport;
194 kmem_cache_free(rds_conn_slab, conn);
195 conn = ERR_PTR(-ENODEV);
199 conn->c_trans = trans;
201 ret = trans->conn_alloc(conn, gfp);
203 kmem_cache_free(rds_conn_slab, conn);
208 atomic_set(&conn->c_state, RDS_CONN_DOWN);
209 conn->c_send_gen = 0;
210 conn->c_reconnect_jiffies = 0;
211 INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
212 INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
213 INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
214 INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
215 mutex_init(&conn->c_cm_lock);
218 rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
219 conn, &laddr, &faddr,
220 trans->t_name ? trans->t_name : "[unknown]",
221 is_outgoing ? "(outgoing)" : "");
224 * Since we ran without holding the conn lock, someone could
225 * have created the same conn (either normal or passive) in the
226 * interim. We check while holding the lock. If we won, we complete
227 * init and return our conn. If we lost, we rollback and return the
230 spin_lock_irqsave(&rds_conn_lock, flags);
232 /* Creating passive conn */
233 if (parent->c_passive) {
234 trans->conn_free(conn->c_transport_data);
235 kmem_cache_free(rds_conn_slab, conn);
236 conn = parent->c_passive;
238 parent->c_passive = conn;
239 rds_cong_add_conn(conn);
243 /* Creating normal conn */
244 struct rds_connection *found;
246 if (!is_outgoing && otrans->t_type == RDS_TRANS_TCP)
249 found = rds_conn_lookup(net, head, laddr, faddr, trans);
251 trans->conn_free(conn->c_transport_data);
252 kmem_cache_free(rds_conn_slab, conn);
255 if ((is_outgoing && otrans->t_type == RDS_TRANS_TCP) ||
256 (otrans->t_type != RDS_TRANS_TCP)) {
257 /* Only the active side should be added to
258 * reconnect list for TCP.
260 hlist_add_head_rcu(&conn->c_hash_node, head);
262 rds_cong_add_conn(conn);
266 spin_unlock_irqrestore(&rds_conn_lock, flags);
272 struct rds_connection *rds_conn_create(struct net *net,
273 __be32 laddr, __be32 faddr,
274 struct rds_transport *trans, gfp_t gfp)
276 return __rds_conn_create(net, laddr, faddr, trans, gfp, 0);
278 EXPORT_SYMBOL_GPL(rds_conn_create);
280 struct rds_connection *rds_conn_create_outgoing(struct net *net,
281 __be32 laddr, __be32 faddr,
282 struct rds_transport *trans, gfp_t gfp)
284 return __rds_conn_create(net, laddr, faddr, trans, gfp, 1);
286 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
288 void rds_conn_shutdown(struct rds_connection *conn)
290 /* shut it down unless it's down already */
291 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
293 * Quiesce the connection mgmt handlers before we start tearing
294 * things down. We don't hold the mutex for the entire
295 * duration of the shutdown operation, else we may be
296 * deadlocking with the CM handler. Instead, the CM event
297 * handler is supposed to check for state DISCONNECTING
299 mutex_lock(&conn->c_cm_lock);
300 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
301 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
302 rds_conn_error(conn, "shutdown called in state %d\n",
303 atomic_read(&conn->c_state));
304 mutex_unlock(&conn->c_cm_lock);
307 mutex_unlock(&conn->c_cm_lock);
309 wait_event(conn->c_waitq,
310 !test_bit(RDS_IN_XMIT, &conn->c_flags));
311 wait_event(conn->c_waitq,
312 !test_bit(RDS_RECV_REFILL, &conn->c_flags));
314 conn->c_trans->conn_shutdown(conn);
315 rds_conn_reset(conn);
317 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
318 /* This can happen - eg when we're in the middle of tearing
319 * down the connection, and someone unloads the rds module.
320 * Quite reproduceable with loopback connections.
324 "%s: failed to transition to state DOWN, "
325 "current state is %d\n",
327 atomic_read(&conn->c_state));
332 /* Then reconnect if it's still live.
333 * The passive side of an IB loopback connection is never added
334 * to the conn hash, so we never trigger a reconnect on this
335 * conn - the reconnect is always triggered by the active peer. */
336 cancel_delayed_work_sync(&conn->c_conn_w);
338 if (!hlist_unhashed(&conn->c_hash_node)) {
340 rds_queue_reconnect(conn);
347 * Stop and free a connection.
349 * This can only be used in very limited circumstances. It assumes that once
350 * the conn has been shutdown that no one else is referencing the connection.
351 * We can only ensure this in the rmmod path in the current code.
353 void rds_conn_destroy(struct rds_connection *conn)
355 struct rds_message *rm, *rtmp;
358 rdsdebug("freeing conn %p for %pI4 -> "
359 "%pI4\n", conn, &conn->c_laddr,
362 /* Ensure conn will not be scheduled for reconnect */
363 spin_lock_irq(&rds_conn_lock);
364 hlist_del_init_rcu(&conn->c_hash_node);
365 spin_unlock_irq(&rds_conn_lock);
368 /* shut the connection down */
370 flush_work(&conn->c_down_w);
372 /* make sure lingering queued work won't try to ref the conn */
373 cancel_delayed_work_sync(&conn->c_send_w);
374 cancel_delayed_work_sync(&conn->c_recv_w);
376 /* tear down queued messages */
377 list_for_each_entry_safe(rm, rtmp,
380 list_del_init(&rm->m_conn_item);
381 BUG_ON(!list_empty(&rm->m_sock_item));
385 rds_message_put(conn->c_xmit_rm);
387 conn->c_trans->conn_free(conn->c_transport_data);
390 * The congestion maps aren't freed up here. They're
391 * freed by rds_cong_exit() after all the connections
394 rds_cong_remove_conn(conn);
396 BUG_ON(!list_empty(&conn->c_retrans));
397 kmem_cache_free(rds_conn_slab, conn);
399 spin_lock_irqsave(&rds_conn_lock, flags);
401 spin_unlock_irqrestore(&rds_conn_lock, flags);
403 EXPORT_SYMBOL_GPL(rds_conn_destroy);
405 static void rds_conn_message_info(struct socket *sock, unsigned int len,
406 struct rds_info_iterator *iter,
407 struct rds_info_lengths *lens,
410 struct hlist_head *head;
411 struct list_head *list;
412 struct rds_connection *conn;
413 struct rds_message *rm;
414 unsigned int total = 0;
418 len /= sizeof(struct rds_info_message);
422 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
424 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
426 list = &conn->c_send_queue;
428 list = &conn->c_retrans;
430 spin_lock_irqsave(&conn->c_lock, flags);
432 /* XXX too lazy to maintain counts.. */
433 list_for_each_entry(rm, list, m_conn_item) {
436 rds_inc_info_copy(&rm->m_inc, iter,
441 spin_unlock_irqrestore(&conn->c_lock, flags);
447 lens->each = sizeof(struct rds_info_message);
450 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
451 struct rds_info_iterator *iter,
452 struct rds_info_lengths *lens)
454 rds_conn_message_info(sock, len, iter, lens, 1);
457 static void rds_conn_message_info_retrans(struct socket *sock,
459 struct rds_info_iterator *iter,
460 struct rds_info_lengths *lens)
462 rds_conn_message_info(sock, len, iter, lens, 0);
465 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
466 struct rds_info_iterator *iter,
467 struct rds_info_lengths *lens,
468 int (*visitor)(struct rds_connection *, void *),
471 uint64_t buffer[(item_len + 7) / 8];
472 struct hlist_head *head;
473 struct rds_connection *conn;
479 lens->each = item_len;
481 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
483 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
485 /* XXX no c_lock usage.. */
486 if (!visitor(conn, buffer))
489 /* We copy as much as we can fit in the buffer,
490 * but we count all items so that the caller
491 * can resize the buffer. */
492 if (len >= item_len) {
493 rds_info_copy(iter, buffer, item_len);
501 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
503 static int rds_conn_info_visitor(struct rds_connection *conn,
506 struct rds_info_connection *cinfo = buffer;
508 cinfo->next_tx_seq = conn->c_next_tx_seq;
509 cinfo->next_rx_seq = conn->c_next_rx_seq;
510 cinfo->laddr = conn->c_laddr;
511 cinfo->faddr = conn->c_faddr;
512 strncpy(cinfo->transport, conn->c_trans->t_name,
513 sizeof(cinfo->transport));
516 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
518 /* XXX Future: return the state rather than these funky bits */
519 rds_conn_info_set(cinfo->flags,
520 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
522 rds_conn_info_set(cinfo->flags,
523 atomic_read(&conn->c_state) == RDS_CONN_UP,
528 static void rds_conn_info(struct socket *sock, unsigned int len,
529 struct rds_info_iterator *iter,
530 struct rds_info_lengths *lens)
532 rds_for_each_conn_info(sock, len, iter, lens,
533 rds_conn_info_visitor,
534 sizeof(struct rds_info_connection));
537 int rds_conn_init(void)
539 rds_conn_slab = kmem_cache_create("rds_connection",
540 sizeof(struct rds_connection),
545 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
546 rds_info_register_func(RDS_INFO_SEND_MESSAGES,
547 rds_conn_message_info_send);
548 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
549 rds_conn_message_info_retrans);
554 void rds_conn_exit(void)
558 WARN_ON(!hlist_empty(rds_conn_hash));
560 kmem_cache_destroy(rds_conn_slab);
562 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
563 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
564 rds_conn_message_info_send);
565 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
566 rds_conn_message_info_retrans);
572 void rds_conn_drop(struct rds_connection *conn)
574 atomic_set(&conn->c_state, RDS_CONN_ERROR);
575 queue_work(rds_wq, &conn->c_down_w);
577 EXPORT_SYMBOL_GPL(rds_conn_drop);
580 * If the connection is down, trigger a connect. We may have scheduled a
581 * delayed reconnect however - in this case we should not interfere.
583 void rds_conn_connect_if_down(struct rds_connection *conn)
585 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
586 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
587 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
589 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
592 * An error occurred on the connection
595 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)