rds: fix rds_send_xmit() serialization
[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         init_waitqueue_head(&conn->c_waitq);
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_rcu(&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                 wait_event(conn->c_waitq,
277                            !test_bit(RDS_IN_XMIT, &conn->c_flags));
278
279                 conn->c_trans->conn_shutdown(conn);
280                 rds_conn_reset(conn);
281
282                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
283                         /* This can happen - eg when we're in the middle of tearing
284                          * down the connection, and someone unloads the rds module.
285                          * Quite reproduceable with loopback connections.
286                          * Mostly harmless.
287                          */
288                         rds_conn_error(conn,
289                                 "%s: failed to transition to state DOWN, "
290                                 "current state is %d\n",
291                                 __func__,
292                                 atomic_read(&conn->c_state));
293                         return;
294                 }
295         }
296
297         /* Then reconnect if it's still live.
298          * The passive side of an IB loopback connection is never added
299          * to the conn hash, so we never trigger a reconnect on this
300          * conn - the reconnect is always triggered by the active peer. */
301         cancel_delayed_work_sync(&conn->c_conn_w);
302         rcu_read_lock();
303         if (!hlist_unhashed(&conn->c_hash_node)) {
304                 rcu_read_unlock();
305                 rds_queue_reconnect(conn);
306         } else {
307                 rcu_read_unlock();
308         }
309 }
310
311 /*
312  * Stop and free a connection.
313  */
314 void rds_conn_destroy(struct rds_connection *conn)
315 {
316         struct rds_message *rm, *rtmp;
317
318         rdsdebug("freeing conn %p for %pI4 -> "
319                  "%pI4\n", conn, &conn->c_laddr,
320                  &conn->c_faddr);
321
322         /* Ensure conn will not be scheduled for reconnect */
323         spin_lock_irq(&rds_conn_lock);
324         hlist_del_init_rcu(&conn->c_hash_node);
325         spin_unlock_irq(&rds_conn_lock);
326
327         synchronize_rcu();
328
329         rds_conn_shutdown(conn);
330
331         /* tear down queued messages */
332         list_for_each_entry_safe(rm, rtmp,
333                                  &conn->c_send_queue,
334                                  m_conn_item) {
335                 list_del_init(&rm->m_conn_item);
336                 BUG_ON(!list_empty(&rm->m_sock_item));
337                 rds_message_put(rm);
338         }
339         if (conn->c_xmit_rm)
340                 rds_message_put(conn->c_xmit_rm);
341
342         conn->c_trans->conn_free(conn->c_transport_data);
343
344         /*
345          * The congestion maps aren't freed up here.  They're
346          * freed by rds_cong_exit() after all the connections
347          * have been freed.
348          */
349         rds_cong_remove_conn(conn);
350
351         BUG_ON(!list_empty(&conn->c_retrans));
352         kmem_cache_free(rds_conn_slab, conn);
353
354         rds_conn_count--;
355 }
356 EXPORT_SYMBOL_GPL(rds_conn_destroy);
357
358 static void rds_conn_message_info(struct socket *sock, unsigned int len,
359                                   struct rds_info_iterator *iter,
360                                   struct rds_info_lengths *lens,
361                                   int want_send)
362 {
363         struct hlist_head *head;
364         struct hlist_node *pos;
365         struct list_head *list;
366         struct rds_connection *conn;
367         struct rds_message *rm;
368         unsigned int total = 0;
369         unsigned long flags;
370         size_t i;
371
372         len /= sizeof(struct rds_info_message);
373
374         rcu_read_lock();
375
376         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
377              i++, head++) {
378                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
379                         if (want_send)
380                                 list = &conn->c_send_queue;
381                         else
382                                 list = &conn->c_retrans;
383
384                         spin_lock_irqsave(&conn->c_lock, flags);
385
386                         /* XXX too lazy to maintain counts.. */
387                         list_for_each_entry(rm, list, m_conn_item) {
388                                 total++;
389                                 if (total <= len)
390                                         rds_inc_info_copy(&rm->m_inc, iter,
391                                                           conn->c_laddr,
392                                                           conn->c_faddr, 0);
393                         }
394
395                         spin_unlock_irqrestore(&conn->c_lock, flags);
396                 }
397         }
398         rcu_read_unlock();
399
400         lens->nr = total;
401         lens->each = sizeof(struct rds_info_message);
402 }
403
404 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
405                                        struct rds_info_iterator *iter,
406                                        struct rds_info_lengths *lens)
407 {
408         rds_conn_message_info(sock, len, iter, lens, 1);
409 }
410
411 static void rds_conn_message_info_retrans(struct socket *sock,
412                                           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, 0);
417 }
418
419 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
420                           struct rds_info_iterator *iter,
421                           struct rds_info_lengths *lens,
422                           int (*visitor)(struct rds_connection *, void *),
423                           size_t item_len)
424 {
425         uint64_t buffer[(item_len + 7) / 8];
426         struct hlist_head *head;
427         struct hlist_node *pos;
428         struct rds_connection *conn;
429         size_t i;
430
431         rcu_read_lock();
432
433         lens->nr = 0;
434         lens->each = item_len;
435
436         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
437              i++, head++) {
438                 hlist_for_each_entry_rcu(conn, pos, head, c_hash_node) {
439
440                         /* XXX no c_lock usage.. */
441                         if (!visitor(conn, buffer))
442                                 continue;
443
444                         /* We copy as much as we can fit in the buffer,
445                          * but we count all items so that the caller
446                          * can resize the buffer. */
447                         if (len >= item_len) {
448                                 rds_info_copy(iter, buffer, item_len);
449                                 len -= item_len;
450                         }
451                         lens->nr++;
452                 }
453         }
454         rcu_read_unlock();
455 }
456 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
457
458 static int rds_conn_info_visitor(struct rds_connection *conn,
459                                   void *buffer)
460 {
461         struct rds_info_connection *cinfo = buffer;
462
463         cinfo->next_tx_seq = conn->c_next_tx_seq;
464         cinfo->next_rx_seq = conn->c_next_rx_seq;
465         cinfo->laddr = conn->c_laddr;
466         cinfo->faddr = conn->c_faddr;
467         strncpy(cinfo->transport, conn->c_trans->t_name,
468                 sizeof(cinfo->transport));
469         cinfo->flags = 0;
470
471         rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags),
472                           SENDING);
473         /* XXX Future: return the state rather than these funky bits */
474         rds_conn_info_set(cinfo->flags,
475                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
476                           CONNECTING);
477         rds_conn_info_set(cinfo->flags,
478                           atomic_read(&conn->c_state) == RDS_CONN_UP,
479                           CONNECTED);
480         return 1;
481 }
482
483 static void rds_conn_info(struct socket *sock, unsigned int len,
484                           struct rds_info_iterator *iter,
485                           struct rds_info_lengths *lens)
486 {
487         rds_for_each_conn_info(sock, len, iter, lens,
488                                 rds_conn_info_visitor,
489                                 sizeof(struct rds_info_connection));
490 }
491
492 int __init rds_conn_init(void)
493 {
494         rds_conn_slab = kmem_cache_create("rds_connection",
495                                           sizeof(struct rds_connection),
496                                           0, 0, NULL);
497         if (!rds_conn_slab)
498                 return -ENOMEM;
499
500         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
501         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
502                                rds_conn_message_info_send);
503         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
504                                rds_conn_message_info_retrans);
505
506         return 0;
507 }
508
509 void rds_conn_exit(void)
510 {
511         rds_loop_exit();
512
513         WARN_ON(!hlist_empty(rds_conn_hash));
514
515         kmem_cache_destroy(rds_conn_slab);
516
517         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
518         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
519                                  rds_conn_message_info_send);
520         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
521                                  rds_conn_message_info_retrans);
522 }
523
524 /*
525  * Force a disconnect
526  */
527 void rds_conn_drop(struct rds_connection *conn)
528 {
529         atomic_set(&conn->c_state, RDS_CONN_ERROR);
530         queue_work(rds_wq, &conn->c_down_w);
531 }
532 EXPORT_SYMBOL_GPL(rds_conn_drop);
533
534 /*
535  * If the connection is down, trigger a connect. We may have scheduled a
536  * delayed reconnect however - in this case we should not interfere.
537  */
538 void rds_conn_connect_if_down(struct rds_connection *conn)
539 {
540         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
541             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
542                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
543 }
544 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
545
546 /*
547  * An error occurred on the connection
548  */
549 void
550 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
551 {
552         va_list ap;
553
554         va_start(ap, fmt);
555         vprintk(fmt, ap);
556         va_end(ap);
557
558         rds_conn_drop(conn);
559 }