RDS: fold rdma.h into rds.h
[linux-3.10.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 inline int rds_conn_is_sending(struct rds_connection *conn)
66 {
67         int ret = 0;
68
69         if (!mutex_trylock(&conn->c_send_lock))
70                 ret = 1;
71         else
72                 mutex_unlock(&conn->c_send_lock);
73
74         return ret;
75 }
76
77 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
78                                               __be32 laddr, __be32 faddr,
79                                               struct rds_transport *trans)
80 {
81         struct rds_connection *conn, *ret = NULL;
82         struct hlist_node *pos;
83
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) {
87                         ret = conn;
88                         break;
89                 }
90         }
91         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
92                  &laddr, &faddr);
93         return ret;
94 }
95
96 /*
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.
101  */
102 void rds_conn_reset(struct rds_connection *conn)
103 {
104         rdsdebug("connection %pI4 to %pI4 reset\n",
105           &conn->c_laddr, &conn->c_faddr);
106
107         rds_stats_inc(s_conn_reset);
108         rds_send_reset(conn);
109         conn->c_flags = 0;
110
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. */
115 }
116
117 /*
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.
121  *
122  * For now they are not garbage collected once they're created.  They
123  * are torn down as the module is removed, if ever.
124  */
125 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
126                                        struct rds_transport *trans, gfp_t gfp,
127                                        int is_outgoing)
128 {
129         struct rds_connection *conn, *parent = NULL;
130         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
131         unsigned long flags;
132         int ret;
133
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 &&
137             !is_outgoing) {
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. */
142                 parent = conn;
143                 conn = parent->c_passive;
144         }
145         spin_unlock_irqrestore(&rds_conn_lock, flags);
146         if (conn)
147                 goto out;
148
149         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
150         if (!conn) {
151                 conn = ERR_PTR(-ENOMEM);
152                 goto out;
153         }
154
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;
160
161         mutex_init(&conn->c_send_lock);
162         INIT_LIST_HEAD(&conn->c_send_queue);
163         INIT_LIST_HEAD(&conn->c_retrans);
164
165         ret = rds_cong_get_maps(conn);
166         if (ret) {
167                 kmem_cache_free(rds_conn_slab, conn);
168                 conn = ERR_PTR(ret);
169                 goto out;
170         }
171
172         /*
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.
176          */
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.
183                          */
184                         trans = &rds_loop_transport;
185                 }
186         }
187
188         conn->c_trans = trans;
189
190         ret = trans->conn_alloc(conn, gfp);
191         if (ret) {
192                 kmem_cache_free(rds_conn_slab, conn);
193                 conn = ERR_PTR(ret);
194                 goto out;
195         }
196
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);
204         conn->c_flags = 0;
205
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)" : "");
210
211         /*
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
216          * other one.
217          */
218         spin_lock_irqsave(&rds_conn_lock, flags);
219         if (parent) {
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;
225                 } else {
226                         parent->c_passive = conn;
227                         rds_cong_add_conn(conn);
228                         rds_conn_count++;
229                 }
230         } else {
231                 /* Creating normal conn */
232                 struct rds_connection *found;
233
234                 found = rds_conn_lookup(head, laddr, faddr, trans);
235                 if (found) {
236                         trans->conn_free(conn->c_transport_data);
237                         kmem_cache_free(rds_conn_slab, conn);
238                         conn = found;
239                 } else {
240                         hlist_add_head(&conn->c_hash_node, head);
241                         rds_cong_add_conn(conn);
242                         rds_conn_count++;
243                 }
244         }
245         spin_unlock_irqrestore(&rds_conn_lock, flags);
246
247 out:
248         return conn;
249 }
250
251 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
252                                        struct rds_transport *trans, gfp_t gfp)
253 {
254         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
255 }
256 EXPORT_SYMBOL_GPL(rds_conn_create);
257
258 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
259                                        struct rds_transport *trans, gfp_t gfp)
260 {
261         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
262 }
263 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
264
265 void rds_conn_shutdown(struct rds_connection *conn)
266 {
267         /* shut it down unless it's down already */
268         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
269                 /*
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
275                  */
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);
282                         return;
283                 }
284                 mutex_unlock(&conn->c_cm_lock);
285
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);
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         if (!hlist_unhashed(&conn->c_hash_node))
312                 rds_queue_reconnect(conn);
313 }
314
315 /*
316  * Stop and free a connection.
317  */
318 void rds_conn_destroy(struct rds_connection *conn)
319 {
320         struct rds_message *rm, *rtmp;
321
322         rdsdebug("freeing conn %p for %pI4 -> "
323                  "%pI4\n", conn, &conn->c_laddr,
324                  &conn->c_faddr);
325
326         hlist_del_init(&conn->c_hash_node);
327
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);
333
334         /* tear down queued messages */
335         list_for_each_entry_safe(rm, rtmp,
336                                  &conn->c_send_queue,
337                                  m_conn_item) {
338                 list_del_init(&rm->m_conn_item);
339                 BUG_ON(!list_empty(&rm->m_sock_item));
340                 rds_message_put(rm);
341         }
342         if (conn->c_xmit_rm)
343                 rds_message_put(conn->c_xmit_rm);
344
345         conn->c_trans->conn_free(conn->c_transport_data);
346
347         /*
348          * The congestion maps aren't freed up here.  They're
349          * freed by rds_cong_exit() after all the connections
350          * have been freed.
351          */
352         rds_cong_remove_conn(conn);
353
354         BUG_ON(!list_empty(&conn->c_retrans));
355         kmem_cache_free(rds_conn_slab, conn);
356
357         rds_conn_count--;
358 }
359 EXPORT_SYMBOL_GPL(rds_conn_destroy);
360
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,
364                                   int want_send)
365 {
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;
371         unsigned long flags;
372         unsigned int total = 0;
373         size_t i;
374
375         len /= sizeof(struct rds_info_message);
376
377         spin_lock_irqsave(&rds_conn_lock, flags);
378
379         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
380              i++, head++) {
381                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
382                         if (want_send)
383                                 list = &conn->c_send_queue;
384                         else
385                                 list = &conn->c_retrans;
386
387                         spin_lock(&conn->c_lock);
388
389                         /* XXX too lazy to maintain counts.. */
390                         list_for_each_entry(rm, list, m_conn_item) {
391                                 total++;
392                                 if (total <= len)
393                                         rds_inc_info_copy(&rm->m_inc, iter,
394                                                           conn->c_laddr,
395                                                           conn->c_faddr, 0);
396                         }
397
398                         spin_unlock(&conn->c_lock);
399                 }
400         }
401
402         spin_unlock_irqrestore(&rds_conn_lock, flags);
403
404         lens->nr = total;
405         lens->each = sizeof(struct rds_info_message);
406 }
407
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)
411 {
412         rds_conn_message_info(sock, len, iter, lens, 1);
413 }
414
415 static void rds_conn_message_info_retrans(struct socket *sock,
416                                           unsigned int len,
417                                           struct rds_info_iterator *iter,
418                                           struct rds_info_lengths *lens)
419 {
420         rds_conn_message_info(sock, len, iter, lens, 0);
421 }
422
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 *),
427                           size_t item_len)
428 {
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;
434         unsigned long flags;
435         size_t i;
436
437         spin_lock_irqsave(&rds_conn_lock, flags);
438
439         lens->nr = 0;
440         lens->each = item_len;
441
442         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
443              i++, head++) {
444                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
445
446                         /* XXX no c_lock usage.. */
447                         if (!visitor(conn, buffer))
448                                 continue;
449
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);
455                                 len -= item_len;
456                         }
457                         lens->nr++;
458                 }
459         }
460
461         spin_unlock_irqrestore(&rds_conn_lock, flags);
462 }
463 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
464
465 static int rds_conn_info_visitor(struct rds_connection *conn,
466                                   void *buffer)
467 {
468         struct rds_info_connection *cinfo = buffer;
469
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));
476         cinfo->flags = 0;
477
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,
483                           CONNECTING);
484         rds_conn_info_set(cinfo->flags,
485                           atomic_read(&conn->c_state) == RDS_CONN_UP,
486                           CONNECTED);
487         return 1;
488 }
489
490 static void rds_conn_info(struct socket *sock, unsigned int len,
491                           struct rds_info_iterator *iter,
492                           struct rds_info_lengths *lens)
493 {
494         rds_for_each_conn_info(sock, len, iter, lens,
495                                 rds_conn_info_visitor,
496                                 sizeof(struct rds_info_connection));
497 }
498
499 int __init rds_conn_init(void)
500 {
501         rds_conn_slab = kmem_cache_create("rds_connection",
502                                           sizeof(struct rds_connection),
503                                           0, 0, NULL);
504         if (!rds_conn_slab)
505                 return -ENOMEM;
506
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);
512
513         return 0;
514 }
515
516 void rds_conn_exit(void)
517 {
518         rds_loop_exit();
519
520         WARN_ON(!hlist_empty(rds_conn_hash));
521
522         kmem_cache_destroy(rds_conn_slab);
523
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);
529 }
530
531 /*
532  * Force a disconnect
533  */
534 void rds_conn_drop(struct rds_connection *conn)
535 {
536         atomic_set(&conn->c_state, RDS_CONN_ERROR);
537         queue_work(rds_wq, &conn->c_down_w);
538 }
539 EXPORT_SYMBOL_GPL(rds_conn_drop);
540
541 /*
542  * An error occurred on the connection
543  */
544 void
545 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
546 {
547         va_list ap;
548
549         va_start(ap, fmt);
550         vprintk(fmt, ap);
551         va_end(ap);
552
553         rds_conn_drop(conn);
554 }