RDS: rds_send_xmit() locking/irq fixes
[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         INIT_LIST_HEAD(&conn->c_send_queue);
151         INIT_LIST_HEAD(&conn->c_retrans);
152
153         ret = rds_cong_get_maps(conn);
154         if (ret) {
155                 kmem_cache_free(rds_conn_slab, conn);
156                 conn = ERR_PTR(ret);
157                 goto out;
158         }
159
160         /*
161          * This is where a connection becomes loopback.  If *any* RDS sockets
162          * can bind to the destination address then we'd rather the messages
163          * flow through loopback rather than either transport.
164          */
165         if (rds_trans_get_preferred(faddr)) {
166                 conn->c_loopback = 1;
167                 if (is_outgoing && trans->t_prefer_loopback) {
168                         /* "outgoing" connection - and the transport
169                          * says it wants the connection handled by the
170                          * loopback transport. This is what TCP does.
171                          */
172                         trans = &rds_loop_transport;
173                 }
174         }
175
176         conn->c_trans = trans;
177
178         ret = trans->conn_alloc(conn, gfp);
179         if (ret) {
180                 kmem_cache_free(rds_conn_slab, conn);
181                 conn = ERR_PTR(ret);
182                 goto out;
183         }
184
185         atomic_set(&conn->c_state, RDS_CONN_DOWN);
186         conn->c_reconnect_jiffies = 0;
187         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
188         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
189         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
190         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
191         mutex_init(&conn->c_cm_lock);
192         conn->c_flags = 0;
193
194         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
195           conn, &laddr, &faddr,
196           trans->t_name ? trans->t_name : "[unknown]",
197           is_outgoing ? "(outgoing)" : "");
198
199         /*
200          * Since we ran without holding the conn lock, someone could
201          * have created the same conn (either normal or passive) in the
202          * interim. We check while holding the lock. If we won, we complete
203          * init and return our conn. If we lost, we rollback and return the
204          * other one.
205          */
206         spin_lock_irqsave(&rds_conn_lock, flags);
207         if (parent) {
208                 /* Creating passive conn */
209                 if (parent->c_passive) {
210                         trans->conn_free(conn->c_transport_data);
211                         kmem_cache_free(rds_conn_slab, conn);
212                         conn = parent->c_passive;
213                 } else {
214                         parent->c_passive = conn;
215                         rds_cong_add_conn(conn);
216                         rds_conn_count++;
217                 }
218         } else {
219                 /* Creating normal conn */
220                 struct rds_connection *found;
221
222                 found = rds_conn_lookup(head, laddr, faddr, trans);
223                 if (found) {
224                         trans->conn_free(conn->c_transport_data);
225                         kmem_cache_free(rds_conn_slab, conn);
226                         conn = found;
227                 } else {
228                         hlist_add_head(&conn->c_hash_node, head);
229                         rds_cong_add_conn(conn);
230                         rds_conn_count++;
231                 }
232         }
233         spin_unlock_irqrestore(&rds_conn_lock, flags);
234
235 out:
236         return conn;
237 }
238
239 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
240                                        struct rds_transport *trans, gfp_t gfp)
241 {
242         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
243 }
244 EXPORT_SYMBOL_GPL(rds_conn_create);
245
246 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
247                                        struct rds_transport *trans, gfp_t gfp)
248 {
249         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
250 }
251 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
252
253 void rds_conn_shutdown(struct rds_connection *conn)
254 {
255         /* shut it down unless it's down already */
256         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
257                 /*
258                  * Quiesce the connection mgmt handlers before we start tearing
259                  * things down. We don't hold the mutex for the entire
260                  * duration of the shutdown operation, else we may be
261                  * deadlocking with the CM handler. Instead, the CM event
262                  * handler is supposed to check for state DISCONNECTING
263                  */
264                 mutex_lock(&conn->c_cm_lock);
265                 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
266                  && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
267                         rds_conn_error(conn, "shutdown called in state %d\n",
268                                         atomic_read(&conn->c_state));
269                         mutex_unlock(&conn->c_cm_lock);
270                         return;
271                 }
272                 mutex_unlock(&conn->c_cm_lock);
273
274                 /* verify everybody's out of rds_send_xmit() */
275                 spin_lock_irq(&conn->c_send_lock);
276                 spin_unlock_irq(&conn->c_send_lock);
277
278                 conn->c_trans->conn_shutdown(conn);
279                 rds_conn_reset(conn);
280
281                 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
282                         /* This can happen - eg when we're in the middle of tearing
283                          * down the connection, and someone unloads the rds module.
284                          * Quite reproduceable with loopback connections.
285                          * Mostly harmless.
286                          */
287                         rds_conn_error(conn,
288                                 "%s: failed to transition to state DOWN, "
289                                 "current state is %d\n",
290                                 __func__,
291                                 atomic_read(&conn->c_state));
292                         return;
293                 }
294         }
295
296         /* Then reconnect if it's still live.
297          * The passive side of an IB loopback connection is never added
298          * to the conn hash, so we never trigger a reconnect on this
299          * conn - the reconnect is always triggered by the active peer. */
300         cancel_delayed_work_sync(&conn->c_conn_w);
301         if (!hlist_unhashed(&conn->c_hash_node))
302                 rds_queue_reconnect(conn);
303 }
304
305 /*
306  * Stop and free a connection.
307  */
308 void rds_conn_destroy(struct rds_connection *conn)
309 {
310         struct rds_message *rm, *rtmp;
311
312         rdsdebug("freeing conn %p for %pI4 -> "
313                  "%pI4\n", conn, &conn->c_laddr,
314                  &conn->c_faddr);
315
316         hlist_del_init(&conn->c_hash_node);
317
318         /* wait for the rds thread to shut it down */
319         atomic_set(&conn->c_state, RDS_CONN_ERROR);
320         cancel_delayed_work(&conn->c_conn_w);
321         queue_work(rds_wq, &conn->c_down_w);
322         flush_workqueue(rds_wq);
323
324         /* tear down queued messages */
325         list_for_each_entry_safe(rm, rtmp,
326                                  &conn->c_send_queue,
327                                  m_conn_item) {
328                 list_del_init(&rm->m_conn_item);
329                 BUG_ON(!list_empty(&rm->m_sock_item));
330                 rds_message_put(rm);
331         }
332         if (conn->c_xmit_rm)
333                 rds_message_put(conn->c_xmit_rm);
334
335         conn->c_trans->conn_free(conn->c_transport_data);
336
337         /*
338          * The congestion maps aren't freed up here.  They're
339          * freed by rds_cong_exit() after all the connections
340          * have been freed.
341          */
342         rds_cong_remove_conn(conn);
343
344         BUG_ON(!list_empty(&conn->c_retrans));
345         kmem_cache_free(rds_conn_slab, conn);
346
347         rds_conn_count--;
348 }
349 EXPORT_SYMBOL_GPL(rds_conn_destroy);
350
351 static void rds_conn_message_info(struct socket *sock, unsigned int len,
352                                   struct rds_info_iterator *iter,
353                                   struct rds_info_lengths *lens,
354                                   int want_send)
355 {
356         struct hlist_head *head;
357         struct hlist_node *pos;
358         struct list_head *list;
359         struct rds_connection *conn;
360         struct rds_message *rm;
361         unsigned long flags;
362         unsigned int total = 0;
363         size_t i;
364
365         len /= sizeof(struct rds_info_message);
366
367         spin_lock_irqsave(&rds_conn_lock, flags);
368
369         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
370              i++, head++) {
371                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
372                         if (want_send)
373                                 list = &conn->c_send_queue;
374                         else
375                                 list = &conn->c_retrans;
376
377                         spin_lock(&conn->c_lock);
378
379                         /* XXX too lazy to maintain counts.. */
380                         list_for_each_entry(rm, list, m_conn_item) {
381                                 total++;
382                                 if (total <= len)
383                                         rds_inc_info_copy(&rm->m_inc, iter,
384                                                           conn->c_laddr,
385                                                           conn->c_faddr, 0);
386                         }
387
388                         spin_unlock(&conn->c_lock);
389                 }
390         }
391
392         spin_unlock_irqrestore(&rds_conn_lock, flags);
393
394         lens->nr = total;
395         lens->each = sizeof(struct rds_info_message);
396 }
397
398 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
399                                        struct rds_info_iterator *iter,
400                                        struct rds_info_lengths *lens)
401 {
402         rds_conn_message_info(sock, len, iter, lens, 1);
403 }
404
405 static void rds_conn_message_info_retrans(struct socket *sock,
406                                           unsigned int len,
407                                           struct rds_info_iterator *iter,
408                                           struct rds_info_lengths *lens)
409 {
410         rds_conn_message_info(sock, len, iter, lens, 0);
411 }
412
413 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
414                           struct rds_info_iterator *iter,
415                           struct rds_info_lengths *lens,
416                           int (*visitor)(struct rds_connection *, void *),
417                           size_t item_len)
418 {
419         uint64_t buffer[(item_len + 7) / 8];
420         struct hlist_head *head;
421         struct hlist_node *pos;
422         struct hlist_node *tmp;
423         struct rds_connection *conn;
424         unsigned long flags;
425         size_t i;
426
427         spin_lock_irqsave(&rds_conn_lock, flags);
428
429         lens->nr = 0;
430         lens->each = item_len;
431
432         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
433              i++, head++) {
434                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
435
436                         /* XXX no c_lock usage.. */
437                         if (!visitor(conn, buffer))
438                                 continue;
439
440                         /* We copy as much as we can fit in the buffer,
441                          * but we count all items so that the caller
442                          * can resize the buffer. */
443                         if (len >= item_len) {
444                                 rds_info_copy(iter, buffer, item_len);
445                                 len -= item_len;
446                         }
447                         lens->nr++;
448                 }
449         }
450
451         spin_unlock_irqrestore(&rds_conn_lock, flags);
452 }
453 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
454
455 static int rds_conn_info_visitor(struct rds_connection *conn,
456                                   void *buffer)
457 {
458         struct rds_info_connection *cinfo = buffer;
459
460         cinfo->next_tx_seq = conn->c_next_tx_seq;
461         cinfo->next_rx_seq = conn->c_next_rx_seq;
462         cinfo->laddr = conn->c_laddr;
463         cinfo->faddr = conn->c_faddr;
464         strncpy(cinfo->transport, conn->c_trans->t_name,
465                 sizeof(cinfo->transport));
466         cinfo->flags = 0;
467
468         rds_conn_info_set(cinfo->flags,
469                           spin_is_locked(&conn->c_send_lock), SENDING);
470         /* XXX Future: return the state rather than these funky bits */
471         rds_conn_info_set(cinfo->flags,
472                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
473                           CONNECTING);
474         rds_conn_info_set(cinfo->flags,
475                           atomic_read(&conn->c_state) == RDS_CONN_UP,
476                           CONNECTED);
477         return 1;
478 }
479
480 static void rds_conn_info(struct socket *sock, unsigned int len,
481                           struct rds_info_iterator *iter,
482                           struct rds_info_lengths *lens)
483 {
484         rds_for_each_conn_info(sock, len, iter, lens,
485                                 rds_conn_info_visitor,
486                                 sizeof(struct rds_info_connection));
487 }
488
489 int __init rds_conn_init(void)
490 {
491         rds_conn_slab = kmem_cache_create("rds_connection",
492                                           sizeof(struct rds_connection),
493                                           0, 0, NULL);
494         if (!rds_conn_slab)
495                 return -ENOMEM;
496
497         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
498         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
499                                rds_conn_message_info_send);
500         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
501                                rds_conn_message_info_retrans);
502
503         return 0;
504 }
505
506 void rds_conn_exit(void)
507 {
508         rds_loop_exit();
509
510         WARN_ON(!hlist_empty(rds_conn_hash));
511
512         kmem_cache_destroy(rds_conn_slab);
513
514         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
515         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
516                                  rds_conn_message_info_send);
517         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
518                                  rds_conn_message_info_retrans);
519 }
520
521 /*
522  * Force a disconnect
523  */
524 void rds_conn_drop(struct rds_connection *conn)
525 {
526         atomic_set(&conn->c_state, RDS_CONN_ERROR);
527         queue_work(rds_wq, &conn->c_down_w);
528 }
529 EXPORT_SYMBOL_GPL(rds_conn_drop);
530
531 /*
532  * An error occurred on the connection
533  */
534 void
535 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
536 {
537         va_list ap;
538
539         va_start(ap, fmt);
540         vprintk(fmt, ap);
541         va_end(ap);
542
543         rds_conn_drop(conn);
544 }