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 /* Pass NULL, don't need struct net for hash */
55 unsigned long hash = inet_ehashfn(NULL,
56 be32_to_cpu(laddr), 0,
57 be32_to_cpu(faddr), 0);
58 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
61 #define rds_conn_info_set(var, test, suffix) do { \
63 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
66 /* rcu read lock must be held or the connection spinlock */
67 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
68 __be32 laddr, __be32 faddr,
69 struct rds_transport *trans)
71 struct rds_connection *conn, *ret = NULL;
73 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
74 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
75 conn->c_trans == trans) {
80 rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
86 * This is called by transports as they're bringing down a connection.
87 * It clears partial message state so that the transport can start sending
88 * and receiving over this connection again in the future. It is up to
89 * the transport to have serialized this call with its send and recv.
91 static void rds_conn_reset(struct rds_connection *conn)
93 rdsdebug("connection %pI4 to %pI4 reset\n",
94 &conn->c_laddr, &conn->c_faddr);
96 rds_stats_inc(s_conn_reset);
100 /* Do not clear next_rx_seq here, else we cannot distinguish
101 * retransmitted packets from new packets, and will hand all
102 * of them to the application. That is not consistent with the
103 * reliability guarantees of RDS. */
107 * There is only every one 'conn' for a given pair of addresses in the
108 * system at a time. They contain messages to be retransmitted and so
109 * span the lifetime of the actual underlying transport connections.
111 * For now they are not garbage collected once they're created. They
112 * are torn down as the module is removed, if ever.
114 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
115 struct rds_transport *trans, gfp_t gfp,
118 struct rds_connection *conn, *parent = NULL;
119 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
120 struct rds_transport *loop_trans;
125 conn = rds_conn_lookup(head, laddr, faddr, trans);
126 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
128 /* This is a looped back IB connection, and we're
129 * called by the code handling the incoming connect.
130 * We need a second connection object into which we
131 * can stick the other QP. */
133 conn = parent->c_passive;
139 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
141 conn = ERR_PTR(-ENOMEM);
145 INIT_HLIST_NODE(&conn->c_hash_node);
146 conn->c_laddr = laddr;
147 conn->c_faddr = faddr;
148 spin_lock_init(&conn->c_lock);
149 conn->c_next_tx_seq = 1;
151 init_waitqueue_head(&conn->c_waitq);
152 INIT_LIST_HEAD(&conn->c_send_queue);
153 INIT_LIST_HEAD(&conn->c_retrans);
155 ret = rds_cong_get_maps(conn);
157 kmem_cache_free(rds_conn_slab, conn);
163 * This is where a connection becomes loopback. If *any* RDS sockets
164 * can bind to the destination address then we'd rather the messages
165 * flow through loopback rather than either transport.
167 loop_trans = rds_trans_get_preferred(faddr);
169 rds_trans_put(loop_trans);
170 conn->c_loopback = 1;
171 if (is_outgoing && trans->t_prefer_loopback) {
172 /* "outgoing" connection - and the transport
173 * says it wants the connection handled by the
174 * loopback transport. This is what TCP does.
176 trans = &rds_loop_transport;
180 conn->c_trans = trans;
182 ret = trans->conn_alloc(conn, gfp);
184 kmem_cache_free(rds_conn_slab, conn);
189 atomic_set(&conn->c_state, RDS_CONN_DOWN);
190 conn->c_reconnect_jiffies = 0;
191 INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
192 INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
193 INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
194 INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
195 mutex_init(&conn->c_cm_lock);
198 rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
199 conn, &laddr, &faddr,
200 trans->t_name ? trans->t_name : "[unknown]",
201 is_outgoing ? "(outgoing)" : "");
204 * Since we ran without holding the conn lock, someone could
205 * have created the same conn (either normal or passive) in the
206 * interim. We check while holding the lock. If we won, we complete
207 * init and return our conn. If we lost, we rollback and return the
210 spin_lock_irqsave(&rds_conn_lock, flags);
212 /* Creating passive conn */
213 if (parent->c_passive) {
214 trans->conn_free(conn->c_transport_data);
215 kmem_cache_free(rds_conn_slab, conn);
216 conn = parent->c_passive;
218 parent->c_passive = conn;
219 rds_cong_add_conn(conn);
223 /* Creating normal conn */
224 struct rds_connection *found;
226 found = rds_conn_lookup(head, laddr, faddr, trans);
228 trans->conn_free(conn->c_transport_data);
229 kmem_cache_free(rds_conn_slab, conn);
232 hlist_add_head_rcu(&conn->c_hash_node, head);
233 rds_cong_add_conn(conn);
237 spin_unlock_irqrestore(&rds_conn_lock, flags);
243 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
244 struct rds_transport *trans, gfp_t gfp)
246 return __rds_conn_create(laddr, faddr, trans, gfp, 0);
248 EXPORT_SYMBOL_GPL(rds_conn_create);
250 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
251 struct rds_transport *trans, gfp_t gfp)
253 return __rds_conn_create(laddr, faddr, trans, gfp, 1);
255 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
257 void rds_conn_shutdown(struct rds_connection *conn)
259 /* shut it down unless it's down already */
260 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
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
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);
276 mutex_unlock(&conn->c_cm_lock);
278 wait_event(conn->c_waitq,
279 !test_bit(RDS_IN_XMIT, &conn->c_flags));
281 conn->c_trans->conn_shutdown(conn);
282 rds_conn_reset(conn);
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.
291 "%s: failed to transition to state DOWN, "
292 "current state is %d\n",
294 atomic_read(&conn->c_state));
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);
305 if (!hlist_unhashed(&conn->c_hash_node)) {
307 rds_queue_reconnect(conn);
314 * Stop and free a connection.
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.
320 void rds_conn_destroy(struct rds_connection *conn)
322 struct rds_message *rm, *rtmp;
325 rdsdebug("freeing conn %p for %pI4 -> "
326 "%pI4\n", conn, &conn->c_laddr,
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);
335 /* shut the connection down */
337 flush_work(&conn->c_down_w);
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);
343 /* tear down queued messages */
344 list_for_each_entry_safe(rm, rtmp,
347 list_del_init(&rm->m_conn_item);
348 BUG_ON(!list_empty(&rm->m_sock_item));
352 rds_message_put(conn->c_xmit_rm);
354 conn->c_trans->conn_free(conn->c_transport_data);
357 * The congestion maps aren't freed up here. They're
358 * freed by rds_cong_exit() after all the connections
361 rds_cong_remove_conn(conn);
363 BUG_ON(!list_empty(&conn->c_retrans));
364 kmem_cache_free(rds_conn_slab, conn);
366 spin_lock_irqsave(&rds_conn_lock, flags);
368 spin_unlock_irqrestore(&rds_conn_lock, flags);
370 EXPORT_SYMBOL_GPL(rds_conn_destroy);
372 static void rds_conn_message_info(struct socket *sock, unsigned int len,
373 struct rds_info_iterator *iter,
374 struct rds_info_lengths *lens,
377 struct hlist_head *head;
378 struct list_head *list;
379 struct rds_connection *conn;
380 struct rds_message *rm;
381 unsigned int total = 0;
385 len /= sizeof(struct rds_info_message);
389 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
391 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
393 list = &conn->c_send_queue;
395 list = &conn->c_retrans;
397 spin_lock_irqsave(&conn->c_lock, flags);
399 /* XXX too lazy to maintain counts.. */
400 list_for_each_entry(rm, list, m_conn_item) {
403 rds_inc_info_copy(&rm->m_inc, iter,
408 spin_unlock_irqrestore(&conn->c_lock, flags);
414 lens->each = sizeof(struct rds_info_message);
417 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
418 struct rds_info_iterator *iter,
419 struct rds_info_lengths *lens)
421 rds_conn_message_info(sock, len, iter, lens, 1);
424 static void rds_conn_message_info_retrans(struct socket *sock,
426 struct rds_info_iterator *iter,
427 struct rds_info_lengths *lens)
429 rds_conn_message_info(sock, len, iter, lens, 0);
432 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
433 struct rds_info_iterator *iter,
434 struct rds_info_lengths *lens,
435 int (*visitor)(struct rds_connection *, void *),
438 uint64_t buffer[(item_len + 7) / 8];
439 struct hlist_head *head;
440 struct rds_connection *conn;
446 lens->each = item_len;
448 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
450 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
452 /* XXX no c_lock usage.. */
453 if (!visitor(conn, buffer))
456 /* We copy as much as we can fit in the buffer,
457 * but we count all items so that the caller
458 * can resize the buffer. */
459 if (len >= item_len) {
460 rds_info_copy(iter, buffer, item_len);
468 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
470 static int rds_conn_info_visitor(struct rds_connection *conn,
473 struct rds_info_connection *cinfo = buffer;
475 cinfo->next_tx_seq = conn->c_next_tx_seq;
476 cinfo->next_rx_seq = conn->c_next_rx_seq;
477 cinfo->laddr = conn->c_laddr;
478 cinfo->faddr = conn->c_faddr;
479 strncpy(cinfo->transport, conn->c_trans->t_name,
480 sizeof(cinfo->transport));
483 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
485 /* XXX Future: return the state rather than these funky bits */
486 rds_conn_info_set(cinfo->flags,
487 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
489 rds_conn_info_set(cinfo->flags,
490 atomic_read(&conn->c_state) == RDS_CONN_UP,
495 static void rds_conn_info(struct socket *sock, unsigned int len,
496 struct rds_info_iterator *iter,
497 struct rds_info_lengths *lens)
499 rds_for_each_conn_info(sock, len, iter, lens,
500 rds_conn_info_visitor,
501 sizeof(struct rds_info_connection));
504 int rds_conn_init(void)
506 rds_conn_slab = kmem_cache_create("rds_connection",
507 sizeof(struct rds_connection),
512 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
513 rds_info_register_func(RDS_INFO_SEND_MESSAGES,
514 rds_conn_message_info_send);
515 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
516 rds_conn_message_info_retrans);
521 void rds_conn_exit(void)
525 WARN_ON(!hlist_empty(rds_conn_hash));
527 kmem_cache_destroy(rds_conn_slab);
529 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
530 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
531 rds_conn_message_info_send);
532 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
533 rds_conn_message_info_retrans);
539 void rds_conn_drop(struct rds_connection *conn)
541 atomic_set(&conn->c_state, RDS_CONN_ERROR);
542 queue_work(rds_wq, &conn->c_down_w);
544 EXPORT_SYMBOL_GPL(rds_conn_drop);
547 * If the connection is down, trigger a connect. We may have scheduled a
548 * delayed reconnect however - in this case we should not interfere.
550 void rds_conn_connect_if_down(struct rds_connection *conn)
552 if (rds_conn_state(conn) == RDS_CONN_DOWN &&
553 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
554 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
556 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
559 * An error occurred on the connection
562 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)