87df15b9f8e4b1072120f252520371f5e8c66146
[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 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
66                                               __be32 laddr, __be32 faddr,
67                                               struct rds_transport *trans)
68 {
69         struct rds_connection *conn, *ret = NULL;
70         struct hlist_node *pos;
71
72         hlist_for_each_entry(conn, pos, head, c_hash_node) {
73                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
74                                 conn->c_trans == trans) {
75                         ret = conn;
76                         break;
77                 }
78         }
79         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
80                  &laddr, &faddr);
81         return ret;
82 }
83
84 /*
85  * This is called by transports as they're bringing down a connection.
86  * It clears partial message state so that the transport can start sending
87  * and receiving over this connection again in the future.  It is up to
88  * the transport to have serialized this call with its send and recv.
89  */
90 void rds_conn_reset(struct rds_connection *conn)
91 {
92         rdsdebug("connection %pI4 to %pI4 reset\n",
93           &conn->c_laddr, &conn->c_faddr);
94
95         rds_stats_inc(s_conn_reset);
96         rds_send_reset(conn);
97         conn->c_flags = 0;
98
99         /* Do not clear next_rx_seq here, else we cannot distinguish
100          * retransmitted packets from new packets, and will hand all
101          * of them to the application. That is not consistent with the
102          * reliability guarantees of RDS. */
103 }
104
105 /*
106  * There is only every one 'conn' for a given pair of addresses in the
107  * system at a time.  They contain messages to be retransmitted and so
108  * span the lifetime of the actual underlying transport connections.
109  *
110  * For now they are not garbage collected once they're created.  They
111  * are torn down as the module is removed, if ever.
112  */
113 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
114                                        struct rds_transport *trans, gfp_t gfp,
115                                        int is_outgoing)
116 {
117         struct rds_connection *conn, *parent = NULL;
118         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
119         unsigned long flags;
120         int ret;
121
122         spin_lock_irqsave(&rds_conn_lock, flags);
123         conn = rds_conn_lookup(head, laddr, faddr, trans);
124         if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
125             !is_outgoing) {
126                 /* This is a looped back IB connection, and we're
127                  * called by the code handling the incoming connect.
128                  * We need a second connection object into which we
129                  * can stick the other QP. */
130                 parent = conn;
131                 conn = parent->c_passive;
132         }
133         spin_unlock_irqrestore(&rds_conn_lock, flags);
134         if (conn)
135                 goto out;
136
137         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
138         if (!conn) {
139                 conn = ERR_PTR(-ENOMEM);
140                 goto out;
141         }
142
143         INIT_HLIST_NODE(&conn->c_hash_node);
144         conn->c_laddr = laddr;
145         conn->c_faddr = faddr;
146         spin_lock_init(&conn->c_lock);
147         conn->c_next_tx_seq = 1;
148
149         spin_lock_init(&conn->c_send_lock);
150         atomic_set(&conn->c_send_generation, 1);
151         atomic_set(&conn->c_senders, 0);
152         INIT_LIST_HEAD(&conn->c_send_queue);
153         INIT_LIST_HEAD(&conn->c_retrans);
154
155         ret = rds_cong_get_maps(conn);
156         if (ret) {
157                 kmem_cache_free(rds_conn_slab, conn);
158                 conn = ERR_PTR(ret);
159                 goto out;
160         }
161
162         /*
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.
166          */
167         if (rds_trans_get_preferred(faddr)) {
168                 conn->c_loopback = 1;
169                 if (is_outgoing && trans->t_prefer_loopback) {
170                         /* "outgoing" connection - and the transport
171                          * says it wants the connection handled by the
172                          * loopback transport. This is what TCP does.
173                          */
174                         trans = &rds_loop_transport;
175                 }
176         }
177
178         conn->c_trans = trans;
179
180         ret = trans->conn_alloc(conn, gfp);
181         if (ret) {
182                 kmem_cache_free(rds_conn_slab, conn);
183                 conn = ERR_PTR(ret);
184                 goto out;
185         }
186
187         atomic_set(&conn->c_state, RDS_CONN_DOWN);
188         conn->c_reconnect_jiffies = 0;
189         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
190         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
191         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
192         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
193         mutex_init(&conn->c_cm_lock);
194         conn->c_flags = 0;
195
196         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
197           conn, &laddr, &faddr,
198           trans->t_name ? trans->t_name : "[unknown]",
199           is_outgoing ? "(outgoing)" : "");
200
201         /*
202          * Since we ran without holding the conn lock, someone could
203          * have created the same conn (either normal or passive) in the
204          * interim. We check while holding the lock. If we won, we complete
205          * init and return our conn. If we lost, we rollback and return the
206          * other one.
207          */
208         spin_lock_irqsave(&rds_conn_lock, flags);
209         if (parent) {
210                 /* Creating passive conn */
211                 if (parent->c_passive) {
212                         trans->conn_free(conn->c_transport_data);
213                         kmem_cache_free(rds_conn_slab, conn);
214                         conn = parent->c_passive;
215                 } else {
216                         parent->c_passive = conn;
217                         rds_cong_add_conn(conn);
218                         rds_conn_count++;
219                 }
220         } else {
221                 /* Creating normal conn */
222                 struct rds_connection *found;
223
224                 found = rds_conn_lookup(head, laddr, faddr, trans);
225                 if (found) {
226                         trans->conn_free(conn->c_transport_data);
227                         kmem_cache_free(rds_conn_slab, conn);
228                         conn = found;
229                 } else {
230                         hlist_add_head(&conn->c_hash_node, head);
231                         rds_cong_add_conn(conn);
232                         rds_conn_count++;
233                 }
234         }
235         spin_unlock_irqrestore(&rds_conn_lock, flags);
236
237 out:
238         return conn;
239 }
240
241 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
242                                        struct rds_transport *trans, gfp_t gfp)
243 {
244         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
245 }
246 EXPORT_SYMBOL_GPL(rds_conn_create);
247
248 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
249                                        struct rds_transport *trans, gfp_t gfp)
250 {
251         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
252 }
253 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
254
255 void rds_conn_shutdown(struct rds_connection *conn)
256 {
257         /* shut it down unless it's down already */
258         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
259                 /*
260                  * Quiesce the connection mgmt handlers before we start tearing
261                  * things down. We don't hold the mutex for the entire
262                  * duration of the shutdown operation, else we may be
263                  * deadlocking with the CM handler. Instead, the CM event
264                  * handler is supposed to check for state DISCONNECTING
265                  */
266                 mutex_lock(&conn->c_cm_lock);
267                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
268                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
269                         rds_conn_error(conn, "shutdown called in state %d\n",
270                                         atomic_read(&conn->c_state));
271                         mutex_unlock(&conn->c_cm_lock);
272                         return;
273                 }
274                 mutex_unlock(&conn->c_cm_lock);
275
276                 /* verify everybody's out of rds_send_xmit() */
277                 spin_lock_irq(&conn->c_send_lock);
278                 spin_unlock_irq(&conn->c_send_lock);
279
280                 while(atomic_read(&conn->c_senders)) {
281                         schedule_timeout(1);
282                         spin_lock_irq(&conn->c_send_lock);
283                         spin_unlock_irq(&conn->c_send_lock);
284                 }
285
286                 conn->c_trans->conn_shutdown(conn);
287                 rds_conn_reset(conn);
288
289                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
290                         /* This can happen - eg when we're in the middle of tearing
291                          * down the connection, and someone unloads the rds module.
292                          * Quite reproduceable with loopback connections.
293                          * Mostly harmless.
294                          */
295                         rds_conn_error(conn,
296                                 "%s: failed to transition to state DOWN, "
297                                 "current state is %d\n",
298                                 __func__,
299                                 atomic_read(&conn->c_state));
300                         return;
301                 }
302         }
303
304         /* Then reconnect if it's still live.
305          * The passive side of an IB loopback connection is never added
306          * to the conn hash, so we never trigger a reconnect on this
307          * conn - the reconnect is always triggered by the active peer. */
308         cancel_delayed_work_sync(&conn->c_conn_w);
309         if (!hlist_unhashed(&conn->c_hash_node))
310                 rds_queue_reconnect(conn);
311 }
312
313 /*
314  * Stop and free a connection.
315  */
316 void rds_conn_destroy(struct rds_connection *conn)
317 {
318         struct rds_message *rm, *rtmp;
319
320         rdsdebug("freeing conn %p for %pI4 -> "
321                  "%pI4\n", conn, &conn->c_laddr,
322                  &conn->c_faddr);
323
324         /* Ensure conn will not be scheduled for reconnect */
325         spin_lock_irq(&rds_conn_lock);
326         hlist_del_init(&conn->c_hash_node);
327         spin_unlock_irq(&rds_conn_lock);
328
329         /* wait for the rds thread to shut it down */
330         atomic_set(&conn->c_state, RDS_CONN_ERROR);
331         cancel_delayed_work(&conn->c_conn_w);
332         queue_work(rds_wq, &conn->c_down_w);
333         flush_workqueue(rds_wq);
334
335         /* tear down queued messages */
336         list_for_each_entry_safe(rm, rtmp,
337                                  &conn->c_send_queue,
338                                  m_conn_item) {
339                 list_del_init(&rm->m_conn_item);
340                 BUG_ON(!list_empty(&rm->m_sock_item));
341                 rds_message_put(rm);
342         }
343         if (conn->c_xmit_rm)
344                 rds_message_put(conn->c_xmit_rm);
345
346         conn->c_trans->conn_free(conn->c_transport_data);
347
348         /*
349          * The congestion maps aren't freed up here.  They're
350          * freed by rds_cong_exit() after all the connections
351          * have been freed.
352          */
353         rds_cong_remove_conn(conn);
354
355         BUG_ON(!list_empty(&conn->c_retrans));
356         kmem_cache_free(rds_conn_slab, conn);
357
358         rds_conn_count--;
359 }
360 EXPORT_SYMBOL_GPL(rds_conn_destroy);
361
362 static void rds_conn_message_info(struct socket *sock, unsigned int len,
363                                   struct rds_info_iterator *iter,
364                                   struct rds_info_lengths *lens,
365                                   int want_send)
366 {
367         struct hlist_head *head;
368         struct hlist_node *pos;
369         struct list_head *list;
370         struct rds_connection *conn;
371         struct rds_message *rm;
372         unsigned long flags;
373         unsigned int total = 0;
374         size_t i;
375
376         len /= sizeof(struct rds_info_message);
377
378         spin_lock_irqsave(&rds_conn_lock, flags);
379
380         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
381              i++, head++) {
382                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
383                         if (want_send)
384                                 list = &conn->c_send_queue;
385                         else
386                                 list = &conn->c_retrans;
387
388                         spin_lock(&conn->c_lock);
389
390                         /* XXX too lazy to maintain counts.. */
391                         list_for_each_entry(rm, list, m_conn_item) {
392                                 total++;
393                                 if (total <= len)
394                                         rds_inc_info_copy(&rm->m_inc, iter,
395                                                           conn->c_laddr,
396                                                           conn->c_faddr, 0);
397                         }
398
399                         spin_unlock(&conn->c_lock);
400                 }
401         }
402
403         spin_unlock_irqrestore(&rds_conn_lock, flags);
404
405         lens->nr = total;
406         lens->each = sizeof(struct rds_info_message);
407 }
408
409 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
410                                        struct rds_info_iterator *iter,
411                                        struct rds_info_lengths *lens)
412 {
413         rds_conn_message_info(sock, len, iter, lens, 1);
414 }
415
416 static void rds_conn_message_info_retrans(struct socket *sock,
417                                           unsigned int len,
418                                           struct rds_info_iterator *iter,
419                                           struct rds_info_lengths *lens)
420 {
421         rds_conn_message_info(sock, len, iter, lens, 0);
422 }
423
424 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
425                           struct rds_info_iterator *iter,
426                           struct rds_info_lengths *lens,
427                           int (*visitor)(struct rds_connection *, void *),
428                           size_t item_len)
429 {
430         uint64_t buffer[(item_len + 7) / 8];
431         struct hlist_head *head;
432         struct hlist_node *pos;
433         struct hlist_node *tmp;
434         struct rds_connection *conn;
435         unsigned long flags;
436         size_t i;
437
438         spin_lock_irqsave(&rds_conn_lock, flags);
439
440         lens->nr = 0;
441         lens->each = item_len;
442
443         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
444              i++, head++) {
445                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
446
447                         /* XXX no c_lock usage.. */
448                         if (!visitor(conn, buffer))
449                                 continue;
450
451                         /* We copy as much as we can fit in the buffer,
452                          * but we count all items so that the caller
453                          * can resize the buffer. */
454                         if (len >= item_len) {
455                                 rds_info_copy(iter, buffer, item_len);
456                                 len -= item_len;
457                         }
458                         lens->nr++;
459                 }
460         }
461
462         spin_unlock_irqrestore(&rds_conn_lock, flags);
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  * An error occurred on the connection
544  */
545 void
546 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
547 {
548         va_list ap;
549
550         va_start(ap, fmt);
551         vprintk(fmt, ap);
552         va_end(ap);
553
554         rds_conn_drop(conn);
555 }