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 <net/inet_hashtables.h>
41 #define RDS_CONNECTION_HASH_BITS 12
42 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
43 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
45 /* converting this to RCU is a chore for another day.. */
46 static DEFINE_SPINLOCK(rds_conn_lock);
47 static unsigned long rds_conn_count;
48 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
49 static struct kmem_cache *rds_conn_slab;
51 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
53 /* Pass NULL, don't need struct net for hash */
54 unsigned long hash = inet_ehashfn(NULL,
55 be32_to_cpu(laddr), 0,
56 be32_to_cpu(faddr), 0);
57 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
60 #define rds_conn_info_set(var, test, suffix) do { \
62 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
65 static inline int rds_conn_is_sending(struct rds_connection *conn)
69 if (!mutex_trylock(&conn->c_send_lock))
72 mutex_unlock(&conn->c_send_lock);
77 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
78 __be32 laddr, __be32 faddr,
79 struct rds_transport *trans)
81 struct rds_connection *conn, *ret = NULL;
82 struct hlist_node *pos;
84 hlist_for_each_entry(conn, pos, head, c_hash_node) {
85 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
86 conn->c_trans == trans) {
91 rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
97 * This is called by transports as they're bringing down a connection.
98 * It clears partial message state so that the transport can start sending
99 * and receiving over this connection again in the future. It is up to
100 * the transport to have serialized this call with its send and recv.
102 void rds_conn_reset(struct rds_connection *conn)
104 rdsdebug("connection %pI4 to %pI4 reset\n",
105 &conn->c_laddr, &conn->c_faddr);
107 rds_stats_inc(s_conn_reset);
108 rds_send_reset(conn);
111 /* Do not clear next_rx_seq here, else we cannot distinguish
112 * retransmitted packets from new packets, and will hand all
113 * of them to the application. That is not consistent with the
114 * reliability guarantees of RDS. */
118 * There is only every one 'conn' for a given pair of addresses in the
119 * system at a time. They contain messages to be retransmitted and so
120 * span the lifetime of the actual underlying transport connections.
122 * For now they are not garbage collected once they're created. They
123 * are torn down as the module is removed, if ever.
125 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
126 struct rds_transport *trans, gfp_t gfp,
129 struct rds_connection *conn, *parent = NULL;
130 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
134 spin_lock_irqsave(&rds_conn_lock, flags);
135 conn = rds_conn_lookup(head, laddr, faddr, trans);
136 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
138 /* This is a looped back IB connection, and we're
139 * called by the code handling the incoming connect.
140 * We need a second connection object into which we
141 * can stick the other QP. */
143 conn = parent->c_passive;
145 spin_unlock_irqrestore(&rds_conn_lock, flags);
149 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
151 conn = ERR_PTR(-ENOMEM);
155 INIT_HLIST_NODE(&conn->c_hash_node);
156 conn->c_laddr = laddr;
157 conn->c_faddr = faddr;
158 spin_lock_init(&conn->c_lock);
159 conn->c_next_tx_seq = 1;
161 mutex_init(&conn->c_send_lock);
162 INIT_LIST_HEAD(&conn->c_send_queue);
163 INIT_LIST_HEAD(&conn->c_retrans);
165 ret = rds_cong_get_maps(conn);
167 kmem_cache_free(rds_conn_slab, conn);
173 * This is where a connection becomes loopback. If *any* RDS sockets
174 * can bind to the destination address then we'd rather the messages
175 * flow through loopback rather than either transport.
177 if (rds_trans_get_preferred(faddr)) {
178 conn->c_loopback = 1;
179 if (is_outgoing && trans->t_prefer_loopback) {
180 /* "outgoing" connection - and the transport
181 * says it wants the connection handled by the
182 * loopback transport. This is what TCP does.
184 trans = &rds_loop_transport;
188 conn->c_trans = trans;
190 ret = trans->conn_alloc(conn, gfp);
192 kmem_cache_free(rds_conn_slab, conn);
197 atomic_set(&conn->c_state, RDS_CONN_DOWN);
198 conn->c_reconnect_jiffies = 0;
199 INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
200 INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
201 INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
202 INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
203 mutex_init(&conn->c_cm_lock);
206 rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
207 conn, &laddr, &faddr,
208 trans->t_name ? trans->t_name : "[unknown]",
209 is_outgoing ? "(outgoing)" : "");
212 * Since we ran without holding the conn lock, someone could
213 * have created the same conn (either normal or passive) in the
214 * interim. We check while holding the lock. If we won, we complete
215 * init and return our conn. If we lost, we rollback and return the
218 spin_lock_irqsave(&rds_conn_lock, flags);
220 /* Creating passive conn */
221 if (parent->c_passive) {
222 trans->conn_free(conn->c_transport_data);
223 kmem_cache_free(rds_conn_slab, conn);
224 conn = parent->c_passive;
226 parent->c_passive = conn;
227 rds_cong_add_conn(conn);
231 /* Creating normal conn */
232 struct rds_connection *found;
234 found = rds_conn_lookup(head, laddr, faddr, trans);
236 trans->conn_free(conn->c_transport_data);
237 kmem_cache_free(rds_conn_slab, conn);
240 hlist_add_head(&conn->c_hash_node, head);
241 rds_cong_add_conn(conn);
245 spin_unlock_irqrestore(&rds_conn_lock, flags);
251 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
252 struct rds_transport *trans, gfp_t gfp)
254 return __rds_conn_create(laddr, faddr, trans, gfp, 0);
256 EXPORT_SYMBOL_GPL(rds_conn_create);
258 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
259 struct rds_transport *trans, gfp_t gfp)
261 return __rds_conn_create(laddr, faddr, trans, gfp, 1);
263 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
265 void rds_conn_shutdown(struct rds_connection *conn)
267 /* shut it down unless it's down already */
268 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
270 * Quiesce the connection mgmt handlers before we start tearing
271 * things down. We don't hold the mutex for the entire
272 * duration of the shutdown operation, else we may be
273 * deadlocking with the CM handler. Instead, the CM event
274 * handler is supposed to check for state DISCONNECTING
276 mutex_lock(&conn->c_cm_lock);
277 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
278 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
279 rds_conn_error(conn, "shutdown called in state %d\n",
280 atomic_read(&conn->c_state));
281 mutex_unlock(&conn->c_cm_lock);
284 mutex_unlock(&conn->c_cm_lock);
286 mutex_lock(&conn->c_send_lock);
287 conn->c_trans->conn_shutdown(conn);
288 rds_conn_reset(conn);
289 mutex_unlock(&conn->c_send_lock);
291 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
292 /* This can happen - eg when we're in the middle of tearing
293 * down the connection, and someone unloads the rds module.
294 * Quite reproduceable with loopback connections.
298 "%s: failed to transition to state DOWN, "
299 "current state is %d\n",
301 atomic_read(&conn->c_state));
306 /* Then reconnect if it's still live.
307 * The passive side of an IB loopback connection is never added
308 * to the conn hash, so we never trigger a reconnect on this
309 * conn - the reconnect is always triggered by the active peer. */
310 cancel_delayed_work_sync(&conn->c_conn_w);
311 if (!hlist_unhashed(&conn->c_hash_node))
312 rds_queue_reconnect(conn);
316 * Stop and free a connection.
318 void rds_conn_destroy(struct rds_connection *conn)
320 struct rds_message *rm, *rtmp;
322 rdsdebug("freeing conn %p for %pI4 -> "
323 "%pI4\n", conn, &conn->c_laddr,
326 hlist_del_init(&conn->c_hash_node);
328 /* wait for the rds thread to shut it down */
329 atomic_set(&conn->c_state, RDS_CONN_ERROR);
330 cancel_delayed_work(&conn->c_conn_w);
331 queue_work(rds_wq, &conn->c_down_w);
332 flush_workqueue(rds_wq);
334 /* tear down queued messages */
335 list_for_each_entry_safe(rm, rtmp,
338 list_del_init(&rm->m_conn_item);
339 BUG_ON(!list_empty(&rm->m_sock_item));
343 rds_message_put(conn->c_xmit_rm);
345 conn->c_trans->conn_free(conn->c_transport_data);
348 * The congestion maps aren't freed up here. They're
349 * freed by rds_cong_exit() after all the connections
352 rds_cong_remove_conn(conn);
354 BUG_ON(!list_empty(&conn->c_retrans));
355 kmem_cache_free(rds_conn_slab, conn);
359 EXPORT_SYMBOL_GPL(rds_conn_destroy);
361 static void rds_conn_message_info(struct socket *sock, unsigned int len,
362 struct rds_info_iterator *iter,
363 struct rds_info_lengths *lens,
366 struct hlist_head *head;
367 struct hlist_node *pos;
368 struct list_head *list;
369 struct rds_connection *conn;
370 struct rds_message *rm;
372 unsigned int total = 0;
375 len /= sizeof(struct rds_info_message);
377 spin_lock_irqsave(&rds_conn_lock, flags);
379 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
381 hlist_for_each_entry(conn, pos, head, c_hash_node) {
383 list = &conn->c_send_queue;
385 list = &conn->c_retrans;
387 spin_lock(&conn->c_lock);
389 /* XXX too lazy to maintain counts.. */
390 list_for_each_entry(rm, list, m_conn_item) {
393 rds_inc_info_copy(&rm->m_inc, iter,
398 spin_unlock(&conn->c_lock);
402 spin_unlock_irqrestore(&rds_conn_lock, flags);
405 lens->each = sizeof(struct rds_info_message);
408 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
409 struct rds_info_iterator *iter,
410 struct rds_info_lengths *lens)
412 rds_conn_message_info(sock, len, iter, lens, 1);
415 static void rds_conn_message_info_retrans(struct socket *sock,
417 struct rds_info_iterator *iter,
418 struct rds_info_lengths *lens)
420 rds_conn_message_info(sock, len, iter, lens, 0);
423 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
424 struct rds_info_iterator *iter,
425 struct rds_info_lengths *lens,
426 int (*visitor)(struct rds_connection *, void *),
429 uint64_t buffer[(item_len + 7) / 8];
430 struct hlist_head *head;
431 struct hlist_node *pos;
432 struct hlist_node *tmp;
433 struct rds_connection *conn;
437 spin_lock_irqsave(&rds_conn_lock, flags);
440 lens->each = item_len;
442 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
444 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
446 /* XXX no c_lock usage.. */
447 if (!visitor(conn, buffer))
450 /* We copy as much as we can fit in the buffer,
451 * but we count all items so that the caller
452 * can resize the buffer. */
453 if (len >= item_len) {
454 rds_info_copy(iter, buffer, item_len);
461 spin_unlock_irqrestore(&rds_conn_lock, flags);
463 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
465 static int rds_conn_info_visitor(struct rds_connection *conn,
468 struct rds_info_connection *cinfo = buffer;
470 cinfo->next_tx_seq = conn->c_next_tx_seq;
471 cinfo->next_rx_seq = conn->c_next_rx_seq;
472 cinfo->laddr = conn->c_laddr;
473 cinfo->faddr = conn->c_faddr;
474 strncpy(cinfo->transport, conn->c_trans->t_name,
475 sizeof(cinfo->transport));
478 rds_conn_info_set(cinfo->flags,
479 rds_conn_is_sending(conn), SENDING);
480 /* XXX Future: return the state rather than these funky bits */
481 rds_conn_info_set(cinfo->flags,
482 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
484 rds_conn_info_set(cinfo->flags,
485 atomic_read(&conn->c_state) == RDS_CONN_UP,
490 static void rds_conn_info(struct socket *sock, unsigned int len,
491 struct rds_info_iterator *iter,
492 struct rds_info_lengths *lens)
494 rds_for_each_conn_info(sock, len, iter, lens,
495 rds_conn_info_visitor,
496 sizeof(struct rds_info_connection));
499 int __init rds_conn_init(void)
501 rds_conn_slab = kmem_cache_create("rds_connection",
502 sizeof(struct rds_connection),
507 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
508 rds_info_register_func(RDS_INFO_SEND_MESSAGES,
509 rds_conn_message_info_send);
510 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
511 rds_conn_message_info_retrans);
516 void rds_conn_exit(void)
520 WARN_ON(!hlist_empty(rds_conn_hash));
522 kmem_cache_destroy(rds_conn_slab);
524 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
525 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
526 rds_conn_message_info_send);
527 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
528 rds_conn_message_info_retrans);
534 void rds_conn_drop(struct rds_connection *conn)
536 atomic_set(&conn->c_state, RDS_CONN_ERROR);
537 queue_work(rds_wq, &conn->c_down_w);
539 EXPORT_SYMBOL_GPL(rds_conn_drop);
542 * An error occurred on the connection
545 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)