RDS: introduce rds_conn_connect_if_down()
[linux-2.6.git] / net / rds / connection.c
1 /*
2  * Copyright (c) 2006 Oracle.  All rights reserved.
3  *
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:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
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.
22  *
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
30  * SOFTWARE.
31  *
32  */
33 #include <linux/kernel.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36 #include <net/inet_hashtables.h>
37
38 #include "rds.h"
39 #include "loop.h"
40
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)
44
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;
50
51 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
52 {
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];
58 }
59
60 #define rds_conn_info_set(var, test, suffix) do {               \
61         if (test)                                               \
62                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
63 } while (0)
64
65 /* rcu read lock must be held or the connection spinlock */
66 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
67                                               __be32 laddr, __be32 faddr,
68                                               struct rds_transport *trans)
69 {
70         struct rds_connection *conn, *ret = NULL;
71         struct hlist_node *pos;
72
73         hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
74                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
75                                 conn->c_trans == trans) {
76                         ret = conn;
77                         break;
78                 }
79         }
80         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
81                  &laddr, &faddr);
82         return ret;
83 }
84
85 /*
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.
90  */
91 void rds_conn_reset(struct rds_connection *conn)
92 {
93         rdsdebug("connection %pI4 to %pI4 reset\n",
94           &conn->c_laddr, &conn->c_faddr);
95
96         rds_stats_inc(s_conn_reset);
97         rds_send_reset(conn);
98         conn->c_flags = 0;
99
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. */
104 }
105
106 /*
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.
110  *
111  * For now they are not garbage collected once they're created.  They
112  * are torn down as the module is removed, if ever.
113  */
114 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
115                                        struct rds_transport *trans, gfp_t gfp,
116                                        int is_outgoing)
117 {
118         struct rds_connection *conn, *parent = NULL;
119         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
120         unsigned long flags;
121         int ret;
122
123
124         rcu_read_lock();
125         conn = rds_conn_lookup(head, laddr, faddr, trans);
126         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
127             !is_outgoing) {
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. */
132                 parent = conn;
133                 conn = parent->c_passive;
134         }
135         rcu_read_unlock();
136         if (conn)
137                 goto out;
138
139         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
140         if (!conn) {
141                 conn = ERR_PTR(-ENOMEM);
142                 goto out;
143         }
144
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;
150
151         spin_lock_init(&conn->c_send_lock);
152         atomic_set(&conn->c_send_generation, 1);
153         atomic_set(&conn->c_senders, 0);
154         INIT_LIST_HEAD(&conn->c_send_queue);
155         INIT_LIST_HEAD(&conn->c_retrans);
156
157         ret = rds_cong_get_maps(conn);
158         if (ret) {
159                 kmem_cache_free(rds_conn_slab, conn);
160                 conn = ERR_PTR(ret);
161                 goto out;
162         }
163
164         /*
165          * This is where a connection becomes loopback.  If *any* RDS sockets
166          * can bind to the destination address then we'd rather the messages
167          * flow through loopback rather than either transport.
168          */
169         if (rds_trans_get_preferred(faddr)) {
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.
175                          */
176                         trans = &rds_loop_transport;
177                 }
178         }
179
180         conn->c_trans = trans;
181
182         ret = trans->conn_alloc(conn, gfp);
183         if (ret) {
184                 kmem_cache_free(rds_conn_slab, conn);
185                 conn = ERR_PTR(ret);
186                 goto out;
187         }
188
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);
196         conn->c_flags = 0;
197
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)" : "");
202
203         /*
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
208          * other one.
209          */
210         spin_lock_irqsave(&rds_conn_lock, flags);
211         if (parent) {
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;
217                 } else {
218                         parent->c_passive = conn;
219                         rds_cong_add_conn(conn);
220                         rds_conn_count++;
221                 }
222         } else {
223                 /* Creating normal conn */
224                 struct rds_connection *found;
225
226                 found = rds_conn_lookup(head, laddr, faddr, trans);
227                 if (found) {
228                         trans->conn_free(conn->c_transport_data);
229                         kmem_cache_free(rds_conn_slab, conn);
230                         conn = found;
231                 } else {
232                         hlist_add_head_rcu(&conn->c_hash_node, head);
233                         rds_cong_add_conn(conn);
234                         rds_conn_count++;
235                 }
236         }
237         spin_unlock_irqrestore(&rds_conn_lock, flags);
238
239 out:
240         return conn;
241 }
242
243 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
244                                        struct rds_transport *trans, gfp_t gfp)
245 {
246         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
247 }
248 EXPORT_SYMBOL_GPL(rds_conn_create);
249
250 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
251                                        struct rds_transport *trans, gfp_t gfp)
252 {
253         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
254 }
255 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
256
257 void 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                 /* verify everybody's out of rds_send_xmit() */
279                 spin_lock_irq(&conn->c_send_lock);
280                 spin_unlock_irq(&conn->c_send_lock);
281
282                 while(atomic_read(&conn->c_senders)) {
283                         schedule_timeout(1);
284                         spin_lock_irq(&conn->c_send_lock);
285                         spin_unlock_irq(&conn->c_send_lock);
286                 }
287
288                 conn->c_trans->conn_shutdown(conn);
289                 rds_conn_reset(conn);
290
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.
295                          * Mostly harmless.
296                          */
297                         rds_conn_error(conn,
298                                 "%s: failed to transition to state DOWN, "
299                                 "current state is %d\n",
300                                 __func__,
301                                 atomic_read(&conn->c_state));
302                         return;
303                 }
304         }
305
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         rcu_read_lock();
312         if (!hlist_unhashed(&conn->c_hash_node)) {
313                 rcu_read_unlock();
314                 rds_queue_reconnect(conn);
315         } else {
316                 rcu_read_unlock();
317         }
318 }
319
320 /*
321  * Stop and free a connection.
322  */
323 void rds_conn_destroy(struct rds_connection *conn)
324 {
325         struct rds_message *rm, *rtmp;
326
327         rdsdebug("freeing conn %p for %pI4 -> "
328                  "%pI4\n", conn, &conn->c_laddr,
329                  &conn->c_faddr);
330
331         /* Ensure conn will not be scheduled for reconnect */
332         spin_lock_irq(&rds_conn_lock);
333         hlist_del_init_rcu(&conn->c_hash_node);
334         spin_unlock_irq(&rds_conn_lock);
335
336         synchronize_rcu();
337
338         rds_conn_shutdown(conn);
339
340         /* tear down queued messages */
341         list_for_each_entry_safe(rm, rtmp,
342                                  &conn->c_send_queue,
343                                  m_conn_item) {
344                 list_del_init(&rm->m_conn_item);
345                 BUG_ON(!list_empty(&rm->m_sock_item));
346                 rds_message_put(rm);
347         }
348         if (conn->c_xmit_rm)
349                 rds_message_put(conn->c_xmit_rm);
350
351         conn->c_trans->conn_free(conn->c_transport_data);
352
353         /*
354          * The congestion maps aren't freed up here.  They're
355          * freed by rds_cong_exit() after all the connections
356          * have been freed.
357          */
358         rds_cong_remove_conn(conn);
359
360         BUG_ON(!list_empty(&conn->c_retrans));
361         kmem_cache_free(rds_conn_slab, conn);
362
363         rds_conn_count--;
364 }
365 EXPORT_SYMBOL_GPL(rds_conn_destroy);
366
367 static void rds_conn_message_info(struct socket *sock, unsigned int len,
368                                   struct rds_info_iterator *iter,
369                                   struct rds_info_lengths *lens,
370                                   int want_send)
371 {
372         struct hlist_head *head;
373         struct hlist_node *pos;
374         struct list_head *list;
375         struct rds_connection *conn;
376         struct rds_message *rm;
377         unsigned int total = 0;
378         size_t i;
379
380         len /= sizeof(struct rds_info_message);
381
382         rcu_read_lock();
383
384         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
385              i++, head++) {
386                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
387                         if (want_send)
388                                 list = &conn->c_send_queue;
389                         else
390                                 list = &conn->c_retrans;
391
392                         spin_lock(&conn->c_lock);
393
394                         /* XXX too lazy to maintain counts.. */
395                         list_for_each_entry(rm, list, m_conn_item) {
396                                 total++;
397                                 if (total <= len)
398                                         rds_inc_info_copy(&rm->m_inc, iter,
399                                                           conn->c_laddr,
400                                                           conn->c_faddr, 0);
401                         }
402
403                         spin_unlock(&conn->c_lock);
404                 }
405         }
406         rcu_read_unlock();
407
408         lens->nr = total;
409         lens->each = sizeof(struct rds_info_message);
410 }
411
412 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
413                                        struct rds_info_iterator *iter,
414                                        struct rds_info_lengths *lens)
415 {
416         rds_conn_message_info(sock, len, iter, lens, 1);
417 }
418
419 static void rds_conn_message_info_retrans(struct socket *sock,
420                                           unsigned int len,
421                                           struct rds_info_iterator *iter,
422                                           struct rds_info_lengths *lens)
423 {
424         rds_conn_message_info(sock, len, iter, lens, 0);
425 }
426
427 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
428                           struct rds_info_iterator *iter,
429                           struct rds_info_lengths *lens,
430                           int (*visitor)(struct rds_connection *, void *),
431                           size_t item_len)
432 {
433         uint64_t buffer[(item_len + 7) / 8];
434         struct hlist_head *head;
435         struct hlist_node *pos;
436         struct rds_connection *conn;
437         size_t i;
438
439         rcu_read_lock();
440
441         lens->nr = 0;
442         lens->each = item_len;
443
444         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
445              i++, head++) {
446                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
447
448                         /* XXX no c_lock usage.. */
449                         if (!visitor(conn, buffer))
450                                 continue;
451
452                         /* We copy as much as we can fit in the buffer,
453                          * but we count all items so that the caller
454                          * can resize the buffer. */
455                         if (len >= item_len) {
456                                 rds_info_copy(iter, buffer, item_len);
457                                 len -= item_len;
458                         }
459                         lens->nr++;
460                 }
461         }
462         rcu_read_unlock();
463 }
464 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
465
466 static int rds_conn_info_visitor(struct rds_connection *conn,
467                                   void *buffer)
468 {
469         struct rds_info_connection *cinfo = buffer;
470
471         cinfo->next_tx_seq = conn->c_next_tx_seq;
472         cinfo->next_rx_seq = conn->c_next_rx_seq;
473         cinfo->laddr = conn->c_laddr;
474         cinfo->faddr = conn->c_faddr;
475         strncpy(cinfo->transport, conn->c_trans->t_name,
476                 sizeof(cinfo->transport));
477         cinfo->flags = 0;
478
479         rds_conn_info_set(cinfo->flags,
480                           spin_is_locked(&conn->c_send_lock), SENDING);
481         /* XXX Future: return the state rather than these funky bits */
482         rds_conn_info_set(cinfo->flags,
483                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
484                           CONNECTING);
485         rds_conn_info_set(cinfo->flags,
486                           atomic_read(&conn->c_state) == RDS_CONN_UP,
487                           CONNECTED);
488         return 1;
489 }
490
491 static void rds_conn_info(struct socket *sock, unsigned int len,
492                           struct rds_info_iterator *iter,
493                           struct rds_info_lengths *lens)
494 {
495         rds_for_each_conn_info(sock, len, iter, lens,
496                                 rds_conn_info_visitor,
497                                 sizeof(struct rds_info_connection));
498 }
499
500 int __init rds_conn_init(void)
501 {
502         rds_conn_slab = kmem_cache_create("rds_connection",
503                                           sizeof(struct rds_connection),
504                                           0, 0, NULL);
505         if (!rds_conn_slab)
506                 return -ENOMEM;
507
508         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
509         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
510                                rds_conn_message_info_send);
511         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
512                                rds_conn_message_info_retrans);
513
514         return 0;
515 }
516
517 void rds_conn_exit(void)
518 {
519         rds_loop_exit();
520
521         WARN_ON(!hlist_empty(rds_conn_hash));
522
523         kmem_cache_destroy(rds_conn_slab);
524
525         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
526         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
527                                  rds_conn_message_info_send);
528         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
529                                  rds_conn_message_info_retrans);
530 }
531
532 /*
533  * Force a disconnect
534  */
535 void rds_conn_drop(struct rds_connection *conn)
536 {
537         atomic_set(&conn->c_state, RDS_CONN_ERROR);
538         queue_work(rds_wq, &conn->c_down_w);
539 }
540 EXPORT_SYMBOL_GPL(rds_conn_drop);
541
542 /*
543  * If the connection is down, trigger a connect. We may have scheduled a
544  * delayed reconnect however - in this case we should not interfere.
545  */
546 void rds_conn_connect_if_down(struct rds_connection *conn)
547 {
548         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
549             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
550                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
551 }
552 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
553
554 /*
555  * An error occurred on the connection
556  */
557 void
558 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
559 {
560         va_list ap;
561
562         va_start(ap, fmt);
563         vprintk(fmt, ap);
564         va_end(ap);
565
566         rds_conn_drop(conn);
567 }