Loading...
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 <linux/export.h>
37#include <net/inet_hashtables.h>
38
39#include "rds.h"
40#include "loop.h"
41
42#define RDS_CONNECTION_HASH_BITS 12
43#define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
44#define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
45
46/* converting this to RCU is a chore for another day.. */
47static DEFINE_SPINLOCK(rds_conn_lock);
48static unsigned long rds_conn_count;
49static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
50static struct kmem_cache *rds_conn_slab;
51
52static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
53{
54 static u32 rds_hash_secret __read_mostly;
55
56 unsigned long hash;
57
58 net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
59
60 /* Pass NULL, don't need struct net for hash */
61 hash = __inet_ehashfn(be32_to_cpu(laddr), 0,
62 be32_to_cpu(faddr), 0,
63 rds_hash_secret);
64 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
65}
66
67#define rds_conn_info_set(var, test, suffix) do { \
68 if (test) \
69 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
70} while (0)
71
72/* rcu read lock must be held or the connection spinlock */
73static struct rds_connection *rds_conn_lookup(struct net *net,
74 struct hlist_head *head,
75 __be32 laddr, __be32 faddr,
76 struct rds_transport *trans)
77{
78 struct rds_connection *conn, *ret = NULL;
79
80 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
81 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
82 conn->c_trans == trans && net == rds_conn_net(conn)) {
83 ret = conn;
84 break;
85 }
86 }
87 rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
88 &laddr, &faddr);
89 return ret;
90}
91
92/*
93 * This is called by transports as they're bringing down a connection.
94 * It clears partial message state so that the transport can start sending
95 * and receiving over this connection again in the future. It is up to
96 * the transport to have serialized this call with its send and recv.
97 */
98static void rds_conn_path_reset(struct rds_conn_path *cp)
99{
100 struct rds_connection *conn = cp->cp_conn;
101
102 rdsdebug("connection %pI4 to %pI4 reset\n",
103 &conn->c_laddr, &conn->c_faddr);
104
105 rds_stats_inc(s_conn_reset);
106 rds_send_path_reset(cp);
107 cp->cp_flags = 0;
108
109 /* Do not clear next_rx_seq here, else we cannot distinguish
110 * retransmitted packets from new packets, and will hand all
111 * of them to the application. That is not consistent with the
112 * reliability guarantees of RDS. */
113}
114
115static void __rds_conn_path_init(struct rds_connection *conn,
116 struct rds_conn_path *cp, bool is_outgoing)
117{
118 spin_lock_init(&cp->cp_lock);
119 cp->cp_next_tx_seq = 1;
120 init_waitqueue_head(&cp->cp_waitq);
121 INIT_LIST_HEAD(&cp->cp_send_queue);
122 INIT_LIST_HEAD(&cp->cp_retrans);
123
124 cp->cp_conn = conn;
125 atomic_set(&cp->cp_state, RDS_CONN_DOWN);
126 cp->cp_send_gen = 0;
127 cp->cp_reconnect_jiffies = 0;
128 INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
129 INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
130 INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
131 INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
132 mutex_init(&cp->cp_cm_lock);
133 cp->cp_flags = 0;
134}
135
136/*
137 * There is only every one 'conn' for a given pair of addresses in the
138 * system at a time. They contain messages to be retransmitted and so
139 * span the lifetime of the actual underlying transport connections.
140 *
141 * For now they are not garbage collected once they're created. They
142 * are torn down as the module is removed, if ever.
143 */
144static struct rds_connection *__rds_conn_create(struct net *net,
145 __be32 laddr, __be32 faddr,
146 struct rds_transport *trans, gfp_t gfp,
147 int is_outgoing)
148{
149 struct rds_connection *conn, *parent = NULL;
150 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
151 struct rds_transport *loop_trans;
152 unsigned long flags;
153 int ret, i;
154 int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
155
156 rcu_read_lock();
157 conn = rds_conn_lookup(net, head, laddr, faddr, trans);
158 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport &&
159 laddr == faddr && !is_outgoing) {
160 /* This is a looped back IB connection, and we're
161 * called by the code handling the incoming connect.
162 * We need a second connection object into which we
163 * can stick the other QP. */
164 parent = conn;
165 conn = parent->c_passive;
166 }
167 rcu_read_unlock();
168 if (conn)
169 goto out;
170
171 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
172 if (!conn) {
173 conn = ERR_PTR(-ENOMEM);
174 goto out;
175 }
176 conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp);
177 if (!conn->c_path) {
178 kmem_cache_free(rds_conn_slab, conn);
179 conn = ERR_PTR(-ENOMEM);
180 goto out;
181 }
182
183 INIT_HLIST_NODE(&conn->c_hash_node);
184 conn->c_laddr = laddr;
185 conn->c_faddr = faddr;
186
187 rds_conn_net_set(conn, net);
188
189 ret = rds_cong_get_maps(conn);
190 if (ret) {
191 kfree(conn->c_path);
192 kmem_cache_free(rds_conn_slab, conn);
193 conn = ERR_PTR(ret);
194 goto out;
195 }
196
197 /*
198 * This is where a connection becomes loopback. If *any* RDS sockets
199 * can bind to the destination address then we'd rather the messages
200 * flow through loopback rather than either transport.
201 */
202 loop_trans = rds_trans_get_preferred(net, faddr);
203 if (loop_trans) {
204 rds_trans_put(loop_trans);
205 conn->c_loopback = 1;
206 if (is_outgoing && trans->t_prefer_loopback) {
207 /* "outgoing" connection - and the transport
208 * says it wants the connection handled by the
209 * loopback transport. This is what TCP does.
210 */
211 trans = &rds_loop_transport;
212 }
213 }
214
215 conn->c_trans = trans;
216
217 init_waitqueue_head(&conn->c_hs_waitq);
218 for (i = 0; i < npaths; i++) {
219 __rds_conn_path_init(conn, &conn->c_path[i],
220 is_outgoing);
221 conn->c_path[i].cp_index = i;
222 }
223 rcu_read_lock();
224 if (rds_destroy_pending(conn))
225 ret = -ENETDOWN;
226 else
227 ret = trans->conn_alloc(conn, GFP_ATOMIC);
228 if (ret) {
229 rcu_read_unlock();
230 kfree(conn->c_path);
231 kmem_cache_free(rds_conn_slab, conn);
232 conn = ERR_PTR(ret);
233 goto out;
234 }
235
236 rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
237 conn, &laddr, &faddr,
238 strnlen(trans->t_name, sizeof(trans->t_name)) ? trans->t_name :
239 "[unknown]", is_outgoing ? "(outgoing)" : "");
240
241 /*
242 * Since we ran without holding the conn lock, someone could
243 * have created the same conn (either normal or passive) in the
244 * interim. We check while holding the lock. If we won, we complete
245 * init and return our conn. If we lost, we rollback and return the
246 * other one.
247 */
248 spin_lock_irqsave(&rds_conn_lock, flags);
249 if (parent) {
250 /* Creating passive conn */
251 if (parent->c_passive) {
252 trans->conn_free(conn->c_path[0].cp_transport_data);
253 kfree(conn->c_path);
254 kmem_cache_free(rds_conn_slab, conn);
255 conn = parent->c_passive;
256 } else {
257 parent->c_passive = conn;
258 rds_cong_add_conn(conn);
259 rds_conn_count++;
260 }
261 } else {
262 /* Creating normal conn */
263 struct rds_connection *found;
264
265 found = rds_conn_lookup(net, head, laddr, faddr, trans);
266 if (found) {
267 struct rds_conn_path *cp;
268 int i;
269
270 for (i = 0; i < npaths; i++) {
271 cp = &conn->c_path[i];
272 /* The ->conn_alloc invocation may have
273 * allocated resource for all paths, so all
274 * of them may have to be freed here.
275 */
276 if (cp->cp_transport_data)
277 trans->conn_free(cp->cp_transport_data);
278 }
279 kfree(conn->c_path);
280 kmem_cache_free(rds_conn_slab, conn);
281 conn = found;
282 } else {
283 conn->c_my_gen_num = rds_gen_num;
284 conn->c_peer_gen_num = 0;
285 hlist_add_head_rcu(&conn->c_hash_node, head);
286 rds_cong_add_conn(conn);
287 rds_conn_count++;
288 }
289 }
290 spin_unlock_irqrestore(&rds_conn_lock, flags);
291 rcu_read_unlock();
292
293out:
294 return conn;
295}
296
297struct rds_connection *rds_conn_create(struct net *net,
298 __be32 laddr, __be32 faddr,
299 struct rds_transport *trans, gfp_t gfp)
300{
301 return __rds_conn_create(net, laddr, faddr, trans, gfp, 0);
302}
303EXPORT_SYMBOL_GPL(rds_conn_create);
304
305struct rds_connection *rds_conn_create_outgoing(struct net *net,
306 __be32 laddr, __be32 faddr,
307 struct rds_transport *trans, gfp_t gfp)
308{
309 return __rds_conn_create(net, laddr, faddr, trans, gfp, 1);
310}
311EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
312
313void rds_conn_shutdown(struct rds_conn_path *cp)
314{
315 struct rds_connection *conn = cp->cp_conn;
316
317 /* shut it down unless it's down already */
318 if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
319 /*
320 * Quiesce the connection mgmt handlers before we start tearing
321 * things down. We don't hold the mutex for the entire
322 * duration of the shutdown operation, else we may be
323 * deadlocking with the CM handler. Instead, the CM event
324 * handler is supposed to check for state DISCONNECTING
325 */
326 mutex_lock(&cp->cp_cm_lock);
327 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
328 RDS_CONN_DISCONNECTING) &&
329 !rds_conn_path_transition(cp, RDS_CONN_ERROR,
330 RDS_CONN_DISCONNECTING)) {
331 rds_conn_path_error(cp,
332 "shutdown called in state %d\n",
333 atomic_read(&cp->cp_state));
334 mutex_unlock(&cp->cp_cm_lock);
335 return;
336 }
337 mutex_unlock(&cp->cp_cm_lock);
338
339 wait_event(cp->cp_waitq,
340 !test_bit(RDS_IN_XMIT, &cp->cp_flags));
341 wait_event(cp->cp_waitq,
342 !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
343
344 conn->c_trans->conn_path_shutdown(cp);
345 rds_conn_path_reset(cp);
346
347 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
348 RDS_CONN_DOWN) &&
349 !rds_conn_path_transition(cp, RDS_CONN_ERROR,
350 RDS_CONN_DOWN)) {
351 /* This can happen - eg when we're in the middle of tearing
352 * down the connection, and someone unloads the rds module.
353 * Quite reproducible with loopback connections.
354 * Mostly harmless.
355 *
356 * Note that this also happens with rds-tcp because
357 * we could have triggered rds_conn_path_drop in irq
358 * mode from rds_tcp_state change on the receipt of
359 * a FIN, thus we need to recheck for RDS_CONN_ERROR
360 * here.
361 */
362 rds_conn_path_error(cp, "%s: failed to transition "
363 "to state DOWN, current state "
364 "is %d\n", __func__,
365 atomic_read(&cp->cp_state));
366 return;
367 }
368 }
369
370 /* Then reconnect if it's still live.
371 * The passive side of an IB loopback connection is never added
372 * to the conn hash, so we never trigger a reconnect on this
373 * conn - the reconnect is always triggered by the active peer. */
374 cancel_delayed_work_sync(&cp->cp_conn_w);
375 rcu_read_lock();
376 if (!hlist_unhashed(&conn->c_hash_node)) {
377 rcu_read_unlock();
378 rds_queue_reconnect(cp);
379 } else {
380 rcu_read_unlock();
381 }
382}
383
384/* destroy a single rds_conn_path. rds_conn_destroy() iterates over
385 * all paths using rds_conn_path_destroy()
386 */
387static void rds_conn_path_destroy(struct rds_conn_path *cp)
388{
389 struct rds_message *rm, *rtmp;
390
391 if (!cp->cp_transport_data)
392 return;
393
394 /* make sure lingering queued work won't try to ref the conn */
395 cancel_delayed_work_sync(&cp->cp_send_w);
396 cancel_delayed_work_sync(&cp->cp_recv_w);
397
398 rds_conn_path_drop(cp, true);
399 flush_work(&cp->cp_down_w);
400
401 /* tear down queued messages */
402 list_for_each_entry_safe(rm, rtmp,
403 &cp->cp_send_queue,
404 m_conn_item) {
405 list_del_init(&rm->m_conn_item);
406 BUG_ON(!list_empty(&rm->m_sock_item));
407 rds_message_put(rm);
408 }
409 if (cp->cp_xmit_rm)
410 rds_message_put(cp->cp_xmit_rm);
411
412 WARN_ON(delayed_work_pending(&cp->cp_send_w));
413 WARN_ON(delayed_work_pending(&cp->cp_recv_w));
414 WARN_ON(delayed_work_pending(&cp->cp_conn_w));
415 WARN_ON(work_pending(&cp->cp_down_w));
416
417 cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
418}
419
420/*
421 * Stop and free a connection.
422 *
423 * This can only be used in very limited circumstances. It assumes that once
424 * the conn has been shutdown that no one else is referencing the connection.
425 * We can only ensure this in the rmmod path in the current code.
426 */
427void rds_conn_destroy(struct rds_connection *conn)
428{
429 unsigned long flags;
430 int i;
431 struct rds_conn_path *cp;
432 int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
433
434 rdsdebug("freeing conn %p for %pI4 -> "
435 "%pI4\n", conn, &conn->c_laddr,
436 &conn->c_faddr);
437
438 /* Ensure conn will not be scheduled for reconnect */
439 spin_lock_irq(&rds_conn_lock);
440 hlist_del_init_rcu(&conn->c_hash_node);
441 spin_unlock_irq(&rds_conn_lock);
442 synchronize_rcu();
443
444 /* shut the connection down */
445 for (i = 0; i < npaths; i++) {
446 cp = &conn->c_path[i];
447 rds_conn_path_destroy(cp);
448 BUG_ON(!list_empty(&cp->cp_retrans));
449 }
450
451 /*
452 * The congestion maps aren't freed up here. They're
453 * freed by rds_cong_exit() after all the connections
454 * have been freed.
455 */
456 rds_cong_remove_conn(conn);
457
458 kfree(conn->c_path);
459 kmem_cache_free(rds_conn_slab, conn);
460
461 spin_lock_irqsave(&rds_conn_lock, flags);
462 rds_conn_count--;
463 spin_unlock_irqrestore(&rds_conn_lock, flags);
464}
465EXPORT_SYMBOL_GPL(rds_conn_destroy);
466
467static void rds_conn_message_info(struct socket *sock, unsigned int len,
468 struct rds_info_iterator *iter,
469 struct rds_info_lengths *lens,
470 int want_send)
471{
472 struct hlist_head *head;
473 struct list_head *list;
474 struct rds_connection *conn;
475 struct rds_message *rm;
476 unsigned int total = 0;
477 unsigned long flags;
478 size_t i;
479 int j;
480
481 len /= sizeof(struct rds_info_message);
482
483 rcu_read_lock();
484
485 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
486 i++, head++) {
487 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
488 struct rds_conn_path *cp;
489 int npaths;
490
491 npaths = (conn->c_trans->t_mp_capable ?
492 RDS_MPATH_WORKERS : 1);
493
494 for (j = 0; j < npaths; j++) {
495 cp = &conn->c_path[j];
496 if (want_send)
497 list = &cp->cp_send_queue;
498 else
499 list = &cp->cp_retrans;
500
501 spin_lock_irqsave(&cp->cp_lock, flags);
502
503 /* XXX too lazy to maintain counts.. */
504 list_for_each_entry(rm, list, m_conn_item) {
505 total++;
506 if (total <= len)
507 rds_inc_info_copy(&rm->m_inc,
508 iter,
509 conn->c_laddr,
510 conn->c_faddr,
511 0);
512 }
513
514 spin_unlock_irqrestore(&cp->cp_lock, flags);
515 }
516 }
517 }
518 rcu_read_unlock();
519
520 lens->nr = total;
521 lens->each = sizeof(struct rds_info_message);
522}
523
524static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
525 struct rds_info_iterator *iter,
526 struct rds_info_lengths *lens)
527{
528 rds_conn_message_info(sock, len, iter, lens, 1);
529}
530
531static void rds_conn_message_info_retrans(struct socket *sock,
532 unsigned int len,
533 struct rds_info_iterator *iter,
534 struct rds_info_lengths *lens)
535{
536 rds_conn_message_info(sock, len, iter, lens, 0);
537}
538
539void rds_for_each_conn_info(struct socket *sock, unsigned int len,
540 struct rds_info_iterator *iter,
541 struct rds_info_lengths *lens,
542 int (*visitor)(struct rds_connection *, void *),
543 u64 *buffer,
544 size_t item_len)
545{
546 struct hlist_head *head;
547 struct rds_connection *conn;
548 size_t i;
549
550 rcu_read_lock();
551
552 lens->nr = 0;
553 lens->each = item_len;
554
555 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
556 i++, head++) {
557 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
558
559 /* XXX no c_lock usage.. */
560 if (!visitor(conn, buffer))
561 continue;
562
563 /* We copy as much as we can fit in the buffer,
564 * but we count all items so that the caller
565 * can resize the buffer. */
566 if (len >= item_len) {
567 rds_info_copy(iter, buffer, item_len);
568 len -= item_len;
569 }
570 lens->nr++;
571 }
572 }
573 rcu_read_unlock();
574}
575EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
576
577static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
578 struct rds_info_iterator *iter,
579 struct rds_info_lengths *lens,
580 int (*visitor)(struct rds_conn_path *, void *),
581 u64 *buffer,
582 size_t item_len)
583{
584 struct hlist_head *head;
585 struct rds_connection *conn;
586 size_t i;
587 int j;
588
589 rcu_read_lock();
590
591 lens->nr = 0;
592 lens->each = item_len;
593
594 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
595 i++, head++) {
596 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
597 struct rds_conn_path *cp;
598 int npaths;
599
600 npaths = (conn->c_trans->t_mp_capable ?
601 RDS_MPATH_WORKERS : 1);
602 for (j = 0; j < npaths; j++) {
603 cp = &conn->c_path[j];
604
605 /* XXX no cp_lock usage.. */
606 if (!visitor(cp, buffer))
607 continue;
608 }
609
610 /* We copy as much as we can fit in the buffer,
611 * but we count all items so that the caller
612 * can resize the buffer.
613 */
614 if (len >= item_len) {
615 rds_info_copy(iter, buffer, item_len);
616 len -= item_len;
617 }
618 lens->nr++;
619 }
620 }
621 rcu_read_unlock();
622}
623
624static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
625{
626 struct rds_info_connection *cinfo = buffer;
627
628 cinfo->next_tx_seq = cp->cp_next_tx_seq;
629 cinfo->next_rx_seq = cp->cp_next_rx_seq;
630 cinfo->laddr = cp->cp_conn->c_laddr;
631 cinfo->faddr = cp->cp_conn->c_faddr;
632 strncpy(cinfo->transport, cp->cp_conn->c_trans->t_name,
633 sizeof(cinfo->transport));
634 cinfo->flags = 0;
635
636 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
637 SENDING);
638 /* XXX Future: return the state rather than these funky bits */
639 rds_conn_info_set(cinfo->flags,
640 atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
641 CONNECTING);
642 rds_conn_info_set(cinfo->flags,
643 atomic_read(&cp->cp_state) == RDS_CONN_UP,
644 CONNECTED);
645 return 1;
646}
647
648static void rds_conn_info(struct socket *sock, unsigned int len,
649 struct rds_info_iterator *iter,
650 struct rds_info_lengths *lens)
651{
652 u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
653
654 rds_walk_conn_path_info(sock, len, iter, lens,
655 rds_conn_info_visitor,
656 buffer,
657 sizeof(struct rds_info_connection));
658}
659
660int rds_conn_init(void)
661{
662 rds_conn_slab = kmem_cache_create("rds_connection",
663 sizeof(struct rds_connection),
664 0, 0, NULL);
665 if (!rds_conn_slab)
666 return -ENOMEM;
667
668 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
669 rds_info_register_func(RDS_INFO_SEND_MESSAGES,
670 rds_conn_message_info_send);
671 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
672 rds_conn_message_info_retrans);
673
674 return 0;
675}
676
677void rds_conn_exit(void)
678{
679 rds_loop_exit();
680
681 WARN_ON(!hlist_empty(rds_conn_hash));
682
683 kmem_cache_destroy(rds_conn_slab);
684
685 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
686 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
687 rds_conn_message_info_send);
688 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
689 rds_conn_message_info_retrans);
690}
691
692/*
693 * Force a disconnect
694 */
695void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
696{
697 atomic_set(&cp->cp_state, RDS_CONN_ERROR);
698
699 rcu_read_lock();
700 if (!destroy && rds_destroy_pending(cp->cp_conn)) {
701 rcu_read_unlock();
702 return;
703 }
704 queue_work(rds_wq, &cp->cp_down_w);
705 rcu_read_unlock();
706}
707EXPORT_SYMBOL_GPL(rds_conn_path_drop);
708
709void rds_conn_drop(struct rds_connection *conn)
710{
711 WARN_ON(conn->c_trans->t_mp_capable);
712 rds_conn_path_drop(&conn->c_path[0], false);
713}
714EXPORT_SYMBOL_GPL(rds_conn_drop);
715
716/*
717 * If the connection is down, trigger a connect. We may have scheduled a
718 * delayed reconnect however - in this case we should not interfere.
719 */
720void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
721{
722 rcu_read_lock();
723 if (rds_destroy_pending(cp->cp_conn)) {
724 rcu_read_unlock();
725 return;
726 }
727 if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
728 !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
729 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
730 rcu_read_unlock();
731}
732EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
733
734void rds_conn_connect_if_down(struct rds_connection *conn)
735{
736 WARN_ON(conn->c_trans->t_mp_capable);
737 rds_conn_path_connect_if_down(&conn->c_path[0]);
738}
739EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
740
741void
742__rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
743{
744 va_list ap;
745
746 va_start(ap, fmt);
747 vprintk(fmt, ap);
748 va_end(ap);
749
750 rds_conn_path_drop(cp, false);
751}
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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 <linux/export.h>
37#include <net/ipv6.h>
38#include <net/inet6_hashtables.h>
39#include <net/addrconf.h>
40
41#include "rds.h"
42#include "loop.h"
43
44#define RDS_CONNECTION_HASH_BITS 12
45#define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
46#define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
47
48/* converting this to RCU is a chore for another day.. */
49static DEFINE_SPINLOCK(rds_conn_lock);
50static unsigned long rds_conn_count;
51static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
52static struct kmem_cache *rds_conn_slab;
53
54static struct hlist_head *rds_conn_bucket(const struct in6_addr *laddr,
55 const struct in6_addr *faddr)
56{
57 static u32 rds6_hash_secret __read_mostly;
58 static u32 rds_hash_secret __read_mostly;
59
60 u32 lhash, fhash, hash;
61
62 net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
63 net_get_random_once(&rds6_hash_secret, sizeof(rds6_hash_secret));
64
65 lhash = (__force u32)laddr->s6_addr32[3];
66#if IS_ENABLED(CONFIG_IPV6)
67 fhash = __ipv6_addr_jhash(faddr, rds6_hash_secret);
68#else
69 fhash = (__force u32)faddr->s6_addr32[3];
70#endif
71 hash = __inet_ehashfn(lhash, 0, fhash, 0, rds_hash_secret);
72
73 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
74}
75
76#define rds_conn_info_set(var, test, suffix) do { \
77 if (test) \
78 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \
79} while (0)
80
81/* rcu read lock must be held or the connection spinlock */
82static struct rds_connection *rds_conn_lookup(struct net *net,
83 struct hlist_head *head,
84 const struct in6_addr *laddr,
85 const struct in6_addr *faddr,
86 struct rds_transport *trans,
87 u8 tos, int dev_if)
88{
89 struct rds_connection *conn, *ret = NULL;
90
91 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
92 if (ipv6_addr_equal(&conn->c_faddr, faddr) &&
93 ipv6_addr_equal(&conn->c_laddr, laddr) &&
94 conn->c_trans == trans &&
95 conn->c_tos == tos &&
96 net == rds_conn_net(conn) &&
97 conn->c_dev_if == dev_if) {
98 ret = conn;
99 break;
100 }
101 }
102 rdsdebug("returning conn %p for %pI6c -> %pI6c\n", ret,
103 laddr, faddr);
104 return ret;
105}
106
107/*
108 * This is called by transports as they're bringing down a connection.
109 * It clears partial message state so that the transport can start sending
110 * and receiving over this connection again in the future. It is up to
111 * the transport to have serialized this call with its send and recv.
112 */
113static void rds_conn_path_reset(struct rds_conn_path *cp)
114{
115 struct rds_connection *conn = cp->cp_conn;
116
117 rdsdebug("connection %pI6c to %pI6c reset\n",
118 &conn->c_laddr, &conn->c_faddr);
119
120 rds_stats_inc(s_conn_reset);
121 rds_send_path_reset(cp);
122 cp->cp_flags = 0;
123
124 /* Do not clear next_rx_seq here, else we cannot distinguish
125 * retransmitted packets from new packets, and will hand all
126 * of them to the application. That is not consistent with the
127 * reliability guarantees of RDS. */
128}
129
130static void __rds_conn_path_init(struct rds_connection *conn,
131 struct rds_conn_path *cp, bool is_outgoing)
132{
133 spin_lock_init(&cp->cp_lock);
134 cp->cp_next_tx_seq = 1;
135 init_waitqueue_head(&cp->cp_waitq);
136 INIT_LIST_HEAD(&cp->cp_send_queue);
137 INIT_LIST_HEAD(&cp->cp_retrans);
138
139 cp->cp_conn = conn;
140 atomic_set(&cp->cp_state, RDS_CONN_DOWN);
141 cp->cp_send_gen = 0;
142 cp->cp_reconnect_jiffies = 0;
143 cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
144 INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
145 INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
146 INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
147 INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
148 mutex_init(&cp->cp_cm_lock);
149 cp->cp_flags = 0;
150}
151
152/*
153 * There is only every one 'conn' for a given pair of addresses in the
154 * system at a time. They contain messages to be retransmitted and so
155 * span the lifetime of the actual underlying transport connections.
156 *
157 * For now they are not garbage collected once they're created. They
158 * are torn down as the module is removed, if ever.
159 */
160static struct rds_connection *__rds_conn_create(struct net *net,
161 const struct in6_addr *laddr,
162 const struct in6_addr *faddr,
163 struct rds_transport *trans,
164 gfp_t gfp, u8 tos,
165 int is_outgoing,
166 int dev_if)
167{
168 struct rds_connection *conn, *parent = NULL;
169 struct hlist_head *head = rds_conn_bucket(laddr, faddr);
170 struct rds_transport *loop_trans;
171 unsigned long flags;
172 int ret, i;
173 int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
174
175 rcu_read_lock();
176 conn = rds_conn_lookup(net, head, laddr, faddr, trans, tos, dev_if);
177 if (conn &&
178 conn->c_loopback &&
179 conn->c_trans != &rds_loop_transport &&
180 ipv6_addr_equal(laddr, faddr) &&
181 !is_outgoing) {
182 /* This is a looped back IB connection, and we're
183 * called by the code handling the incoming connect.
184 * We need a second connection object into which we
185 * can stick the other QP. */
186 parent = conn;
187 conn = parent->c_passive;
188 }
189 rcu_read_unlock();
190 if (conn)
191 goto out;
192
193 conn = kmem_cache_zalloc(rds_conn_slab, gfp);
194 if (!conn) {
195 conn = ERR_PTR(-ENOMEM);
196 goto out;
197 }
198 conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp);
199 if (!conn->c_path) {
200 kmem_cache_free(rds_conn_slab, conn);
201 conn = ERR_PTR(-ENOMEM);
202 goto out;
203 }
204
205 INIT_HLIST_NODE(&conn->c_hash_node);
206 conn->c_laddr = *laddr;
207 conn->c_isv6 = !ipv6_addr_v4mapped(laddr);
208 conn->c_faddr = *faddr;
209 conn->c_dev_if = dev_if;
210 conn->c_tos = tos;
211
212#if IS_ENABLED(CONFIG_IPV6)
213 /* If the local address is link local, set c_bound_if to be the
214 * index used for this connection. Otherwise, set it to 0 as
215 * the socket is not bound to an interface. c_bound_if is used
216 * to look up a socket when a packet is received
217 */
218 if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL)
219 conn->c_bound_if = dev_if;
220 else
221#endif
222 conn->c_bound_if = 0;
223
224 rds_conn_net_set(conn, net);
225
226 ret = rds_cong_get_maps(conn);
227 if (ret) {
228 kfree(conn->c_path);
229 kmem_cache_free(rds_conn_slab, conn);
230 conn = ERR_PTR(ret);
231 goto out;
232 }
233
234 /*
235 * This is where a connection becomes loopback. If *any* RDS sockets
236 * can bind to the destination address then we'd rather the messages
237 * flow through loopback rather than either transport.
238 */
239 loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if);
240 if (loop_trans) {
241 rds_trans_put(loop_trans);
242 conn->c_loopback = 1;
243 if (trans->t_prefer_loopback) {
244 if (likely(is_outgoing)) {
245 /* "outgoing" connection to local address.
246 * Protocol says it wants the connection
247 * handled by the loopback transport.
248 * This is what TCP does.
249 */
250 trans = &rds_loop_transport;
251 } else {
252 /* No transport currently in use
253 * should end up here, but if it
254 * does, reset/destroy the connection.
255 */
256 kfree(conn->c_path);
257 kmem_cache_free(rds_conn_slab, conn);
258 conn = ERR_PTR(-EOPNOTSUPP);
259 goto out;
260 }
261 }
262 }
263
264 conn->c_trans = trans;
265
266 init_waitqueue_head(&conn->c_hs_waitq);
267 for (i = 0; i < npaths; i++) {
268 __rds_conn_path_init(conn, &conn->c_path[i],
269 is_outgoing);
270 conn->c_path[i].cp_index = i;
271 }
272 rcu_read_lock();
273 if (rds_destroy_pending(conn))
274 ret = -ENETDOWN;
275 else
276 ret = trans->conn_alloc(conn, GFP_ATOMIC);
277 if (ret) {
278 rcu_read_unlock();
279 kfree(conn->c_path);
280 kmem_cache_free(rds_conn_slab, conn);
281 conn = ERR_PTR(ret);
282 goto out;
283 }
284
285 rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n",
286 conn, laddr, faddr,
287 strnlen(trans->t_name, sizeof(trans->t_name)) ?
288 trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : "");
289
290 /*
291 * Since we ran without holding the conn lock, someone could
292 * have created the same conn (either normal or passive) in the
293 * interim. We check while holding the lock. If we won, we complete
294 * init and return our conn. If we lost, we rollback and return the
295 * other one.
296 */
297 spin_lock_irqsave(&rds_conn_lock, flags);
298 if (parent) {
299 /* Creating passive conn */
300 if (parent->c_passive) {
301 trans->conn_free(conn->c_path[0].cp_transport_data);
302 kfree(conn->c_path);
303 kmem_cache_free(rds_conn_slab, conn);
304 conn = parent->c_passive;
305 } else {
306 parent->c_passive = conn;
307 rds_cong_add_conn(conn);
308 rds_conn_count++;
309 }
310 } else {
311 /* Creating normal conn */
312 struct rds_connection *found;
313
314 found = rds_conn_lookup(net, head, laddr, faddr, trans,
315 tos, dev_if);
316 if (found) {
317 struct rds_conn_path *cp;
318 int i;
319
320 for (i = 0; i < npaths; i++) {
321 cp = &conn->c_path[i];
322 /* The ->conn_alloc invocation may have
323 * allocated resource for all paths, so all
324 * of them may have to be freed here.
325 */
326 if (cp->cp_transport_data)
327 trans->conn_free(cp->cp_transport_data);
328 }
329 kfree(conn->c_path);
330 kmem_cache_free(rds_conn_slab, conn);
331 conn = found;
332 } else {
333 conn->c_my_gen_num = rds_gen_num;
334 conn->c_peer_gen_num = 0;
335 hlist_add_head_rcu(&conn->c_hash_node, head);
336 rds_cong_add_conn(conn);
337 rds_conn_count++;
338 }
339 }
340 spin_unlock_irqrestore(&rds_conn_lock, flags);
341 rcu_read_unlock();
342
343out:
344 return conn;
345}
346
347struct rds_connection *rds_conn_create(struct net *net,
348 const struct in6_addr *laddr,
349 const struct in6_addr *faddr,
350 struct rds_transport *trans, u8 tos,
351 gfp_t gfp, int dev_if)
352{
353 return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 0, dev_if);
354}
355EXPORT_SYMBOL_GPL(rds_conn_create);
356
357struct rds_connection *rds_conn_create_outgoing(struct net *net,
358 const struct in6_addr *laddr,
359 const struct in6_addr *faddr,
360 struct rds_transport *trans,
361 u8 tos, gfp_t gfp, int dev_if)
362{
363 return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 1, dev_if);
364}
365EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
366
367void rds_conn_shutdown(struct rds_conn_path *cp)
368{
369 struct rds_connection *conn = cp->cp_conn;
370
371 /* shut it down unless it's down already */
372 if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
373 /*
374 * Quiesce the connection mgmt handlers before we start tearing
375 * things down. We don't hold the mutex for the entire
376 * duration of the shutdown operation, else we may be
377 * deadlocking with the CM handler. Instead, the CM event
378 * handler is supposed to check for state DISCONNECTING
379 */
380 mutex_lock(&cp->cp_cm_lock);
381 if (!rds_conn_path_transition(cp, RDS_CONN_UP,
382 RDS_CONN_DISCONNECTING) &&
383 !rds_conn_path_transition(cp, RDS_CONN_ERROR,
384 RDS_CONN_DISCONNECTING)) {
385 rds_conn_path_error(cp,
386 "shutdown called in state %d\n",
387 atomic_read(&cp->cp_state));
388 mutex_unlock(&cp->cp_cm_lock);
389 return;
390 }
391 mutex_unlock(&cp->cp_cm_lock);
392
393 wait_event(cp->cp_waitq,
394 !test_bit(RDS_IN_XMIT, &cp->cp_flags));
395 wait_event(cp->cp_waitq,
396 !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
397
398 conn->c_trans->conn_path_shutdown(cp);
399 rds_conn_path_reset(cp);
400
401 if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
402 RDS_CONN_DOWN) &&
403 !rds_conn_path_transition(cp, RDS_CONN_ERROR,
404 RDS_CONN_DOWN)) {
405 /* This can happen - eg when we're in the middle of tearing
406 * down the connection, and someone unloads the rds module.
407 * Quite reproducible with loopback connections.
408 * Mostly harmless.
409 *
410 * Note that this also happens with rds-tcp because
411 * we could have triggered rds_conn_path_drop in irq
412 * mode from rds_tcp_state change on the receipt of
413 * a FIN, thus we need to recheck for RDS_CONN_ERROR
414 * here.
415 */
416 rds_conn_path_error(cp, "%s: failed to transition "
417 "to state DOWN, current state "
418 "is %d\n", __func__,
419 atomic_read(&cp->cp_state));
420 return;
421 }
422 }
423
424 /* Then reconnect if it's still live.
425 * The passive side of an IB loopback connection is never added
426 * to the conn hash, so we never trigger a reconnect on this
427 * conn - the reconnect is always triggered by the active peer. */
428 cancel_delayed_work_sync(&cp->cp_conn_w);
429 rcu_read_lock();
430 if (!hlist_unhashed(&conn->c_hash_node)) {
431 rcu_read_unlock();
432 rds_queue_reconnect(cp);
433 } else {
434 rcu_read_unlock();
435 }
436}
437
438/* destroy a single rds_conn_path. rds_conn_destroy() iterates over
439 * all paths using rds_conn_path_destroy()
440 */
441static void rds_conn_path_destroy(struct rds_conn_path *cp)
442{
443 struct rds_message *rm, *rtmp;
444
445 if (!cp->cp_transport_data)
446 return;
447
448 /* make sure lingering queued work won't try to ref the conn */
449 cancel_delayed_work_sync(&cp->cp_send_w);
450 cancel_delayed_work_sync(&cp->cp_recv_w);
451
452 rds_conn_path_drop(cp, true);
453 flush_work(&cp->cp_down_w);
454
455 /* tear down queued messages */
456 list_for_each_entry_safe(rm, rtmp,
457 &cp->cp_send_queue,
458 m_conn_item) {
459 list_del_init(&rm->m_conn_item);
460 BUG_ON(!list_empty(&rm->m_sock_item));
461 rds_message_put(rm);
462 }
463 if (cp->cp_xmit_rm)
464 rds_message_put(cp->cp_xmit_rm);
465
466 WARN_ON(delayed_work_pending(&cp->cp_send_w));
467 WARN_ON(delayed_work_pending(&cp->cp_recv_w));
468 WARN_ON(delayed_work_pending(&cp->cp_conn_w));
469 WARN_ON(work_pending(&cp->cp_down_w));
470
471 cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
472}
473
474/*
475 * Stop and free a connection.
476 *
477 * This can only be used in very limited circumstances. It assumes that once
478 * the conn has been shutdown that no one else is referencing the connection.
479 * We can only ensure this in the rmmod path in the current code.
480 */
481void rds_conn_destroy(struct rds_connection *conn)
482{
483 unsigned long flags;
484 int i;
485 struct rds_conn_path *cp;
486 int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
487
488 rdsdebug("freeing conn %p for %pI4 -> "
489 "%pI4\n", conn, &conn->c_laddr,
490 &conn->c_faddr);
491
492 /* Ensure conn will not be scheduled for reconnect */
493 spin_lock_irq(&rds_conn_lock);
494 hlist_del_init_rcu(&conn->c_hash_node);
495 spin_unlock_irq(&rds_conn_lock);
496 synchronize_rcu();
497
498 /* shut the connection down */
499 for (i = 0; i < npaths; i++) {
500 cp = &conn->c_path[i];
501 rds_conn_path_destroy(cp);
502 BUG_ON(!list_empty(&cp->cp_retrans));
503 }
504
505 /*
506 * The congestion maps aren't freed up here. They're
507 * freed by rds_cong_exit() after all the connections
508 * have been freed.
509 */
510 rds_cong_remove_conn(conn);
511
512 kfree(conn->c_path);
513 kmem_cache_free(rds_conn_slab, conn);
514
515 spin_lock_irqsave(&rds_conn_lock, flags);
516 rds_conn_count--;
517 spin_unlock_irqrestore(&rds_conn_lock, flags);
518}
519EXPORT_SYMBOL_GPL(rds_conn_destroy);
520
521static void __rds_inc_msg_cp(struct rds_incoming *inc,
522 struct rds_info_iterator *iter,
523 void *saddr, void *daddr, int flip, bool isv6)
524{
525#if IS_ENABLED(CONFIG_IPV6)
526 if (isv6)
527 rds6_inc_info_copy(inc, iter, saddr, daddr, flip);
528 else
529#endif
530 rds_inc_info_copy(inc, iter, *(__be32 *)saddr,
531 *(__be32 *)daddr, flip);
532}
533
534static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len,
535 struct rds_info_iterator *iter,
536 struct rds_info_lengths *lens,
537 int want_send, bool isv6)
538{
539 struct hlist_head *head;
540 struct list_head *list;
541 struct rds_connection *conn;
542 struct rds_message *rm;
543 unsigned int total = 0;
544 unsigned long flags;
545 size_t i;
546 int j;
547
548 if (isv6)
549 len /= sizeof(struct rds6_info_message);
550 else
551 len /= sizeof(struct rds_info_message);
552
553 rcu_read_lock();
554
555 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
556 i++, head++) {
557 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
558 struct rds_conn_path *cp;
559 int npaths;
560
561 if (!isv6 && conn->c_isv6)
562 continue;
563
564 npaths = (conn->c_trans->t_mp_capable ?
565 RDS_MPATH_WORKERS : 1);
566
567 for (j = 0; j < npaths; j++) {
568 cp = &conn->c_path[j];
569 if (want_send)
570 list = &cp->cp_send_queue;
571 else
572 list = &cp->cp_retrans;
573
574 spin_lock_irqsave(&cp->cp_lock, flags);
575
576 /* XXX too lazy to maintain counts.. */
577 list_for_each_entry(rm, list, m_conn_item) {
578 total++;
579 if (total <= len)
580 __rds_inc_msg_cp(&rm->m_inc,
581 iter,
582 &conn->c_laddr,
583 &conn->c_faddr,
584 0, isv6);
585 }
586
587 spin_unlock_irqrestore(&cp->cp_lock, flags);
588 }
589 }
590 }
591 rcu_read_unlock();
592
593 lens->nr = total;
594 if (isv6)
595 lens->each = sizeof(struct rds6_info_message);
596 else
597 lens->each = sizeof(struct rds_info_message);
598}
599
600static void rds_conn_message_info(struct socket *sock, unsigned int len,
601 struct rds_info_iterator *iter,
602 struct rds_info_lengths *lens,
603 int want_send)
604{
605 rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false);
606}
607
608#if IS_ENABLED(CONFIG_IPV6)
609static void rds6_conn_message_info(struct socket *sock, unsigned int len,
610 struct rds_info_iterator *iter,
611 struct rds_info_lengths *lens,
612 int want_send)
613{
614 rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true);
615}
616#endif
617
618static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
619 struct rds_info_iterator *iter,
620 struct rds_info_lengths *lens)
621{
622 rds_conn_message_info(sock, len, iter, lens, 1);
623}
624
625#if IS_ENABLED(CONFIG_IPV6)
626static void rds6_conn_message_info_send(struct socket *sock, unsigned int len,
627 struct rds_info_iterator *iter,
628 struct rds_info_lengths *lens)
629{
630 rds6_conn_message_info(sock, len, iter, lens, 1);
631}
632#endif
633
634static void rds_conn_message_info_retrans(struct socket *sock,
635 unsigned int len,
636 struct rds_info_iterator *iter,
637 struct rds_info_lengths *lens)
638{
639 rds_conn_message_info(sock, len, iter, lens, 0);
640}
641
642#if IS_ENABLED(CONFIG_IPV6)
643static void rds6_conn_message_info_retrans(struct socket *sock,
644 unsigned int len,
645 struct rds_info_iterator *iter,
646 struct rds_info_lengths *lens)
647{
648 rds6_conn_message_info(sock, len, iter, lens, 0);
649}
650#endif
651
652void rds_for_each_conn_info(struct socket *sock, unsigned int len,
653 struct rds_info_iterator *iter,
654 struct rds_info_lengths *lens,
655 int (*visitor)(struct rds_connection *, void *),
656 u64 *buffer,
657 size_t item_len)
658{
659 struct hlist_head *head;
660 struct rds_connection *conn;
661 size_t i;
662
663 rcu_read_lock();
664
665 lens->nr = 0;
666 lens->each = item_len;
667
668 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
669 i++, head++) {
670 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
671
672 /* XXX no c_lock usage.. */
673 if (!visitor(conn, buffer))
674 continue;
675
676 /* We copy as much as we can fit in the buffer,
677 * but we count all items so that the caller
678 * can resize the buffer. */
679 if (len >= item_len) {
680 rds_info_copy(iter, buffer, item_len);
681 len -= item_len;
682 }
683 lens->nr++;
684 }
685 }
686 rcu_read_unlock();
687}
688EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
689
690static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
691 struct rds_info_iterator *iter,
692 struct rds_info_lengths *lens,
693 int (*visitor)(struct rds_conn_path *, void *),
694 u64 *buffer,
695 size_t item_len)
696{
697 struct hlist_head *head;
698 struct rds_connection *conn;
699 size_t i;
700
701 rcu_read_lock();
702
703 lens->nr = 0;
704 lens->each = item_len;
705
706 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
707 i++, head++) {
708 hlist_for_each_entry_rcu(conn, head, c_hash_node) {
709 struct rds_conn_path *cp;
710
711 /* XXX We only copy the information from the first
712 * path for now. The problem is that if there are
713 * more than one underlying paths, we cannot report
714 * information of all of them using the existing
715 * API. For example, there is only one next_tx_seq,
716 * which path's next_tx_seq should we report? It is
717 * a bug in the design of MPRDS.
718 */
719 cp = conn->c_path;
720
721 /* XXX no cp_lock usage.. */
722 if (!visitor(cp, buffer))
723 continue;
724
725 /* We copy as much as we can fit in the buffer,
726 * but we count all items so that the caller
727 * can resize the buffer.
728 */
729 if (len >= item_len) {
730 rds_info_copy(iter, buffer, item_len);
731 len -= item_len;
732 }
733 lens->nr++;
734 }
735 }
736 rcu_read_unlock();
737}
738
739static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
740{
741 struct rds_info_connection *cinfo = buffer;
742 struct rds_connection *conn = cp->cp_conn;
743
744 if (conn->c_isv6)
745 return 0;
746
747 cinfo->next_tx_seq = cp->cp_next_tx_seq;
748 cinfo->next_rx_seq = cp->cp_next_rx_seq;
749 cinfo->laddr = conn->c_laddr.s6_addr32[3];
750 cinfo->faddr = conn->c_faddr.s6_addr32[3];
751 cinfo->tos = conn->c_tos;
752 strncpy(cinfo->transport, conn->c_trans->t_name,
753 sizeof(cinfo->transport));
754 cinfo->flags = 0;
755
756 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
757 SENDING);
758 /* XXX Future: return the state rather than these funky bits */
759 rds_conn_info_set(cinfo->flags,
760 atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
761 CONNECTING);
762 rds_conn_info_set(cinfo->flags,
763 atomic_read(&cp->cp_state) == RDS_CONN_UP,
764 CONNECTED);
765 return 1;
766}
767
768#if IS_ENABLED(CONFIG_IPV6)
769static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
770{
771 struct rds6_info_connection *cinfo6 = buffer;
772 struct rds_connection *conn = cp->cp_conn;
773
774 cinfo6->next_tx_seq = cp->cp_next_tx_seq;
775 cinfo6->next_rx_seq = cp->cp_next_rx_seq;
776 cinfo6->laddr = conn->c_laddr;
777 cinfo6->faddr = conn->c_faddr;
778 strncpy(cinfo6->transport, conn->c_trans->t_name,
779 sizeof(cinfo6->transport));
780 cinfo6->flags = 0;
781
782 rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
783 SENDING);
784 /* XXX Future: return the state rather than these funky bits */
785 rds_conn_info_set(cinfo6->flags,
786 atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
787 CONNECTING);
788 rds_conn_info_set(cinfo6->flags,
789 atomic_read(&cp->cp_state) == RDS_CONN_UP,
790 CONNECTED);
791 /* Just return 1 as there is no error case. This is a helper function
792 * for rds_walk_conn_path_info() and it wants a return value.
793 */
794 return 1;
795}
796#endif
797
798static void rds_conn_info(struct socket *sock, unsigned int len,
799 struct rds_info_iterator *iter,
800 struct rds_info_lengths *lens)
801{
802 u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
803
804 rds_walk_conn_path_info(sock, len, iter, lens,
805 rds_conn_info_visitor,
806 buffer,
807 sizeof(struct rds_info_connection));
808}
809
810#if IS_ENABLED(CONFIG_IPV6)
811static void rds6_conn_info(struct socket *sock, unsigned int len,
812 struct rds_info_iterator *iter,
813 struct rds_info_lengths *lens)
814{
815 u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8];
816
817 rds_walk_conn_path_info(sock, len, iter, lens,
818 rds6_conn_info_visitor,
819 buffer,
820 sizeof(struct rds6_info_connection));
821}
822#endif
823
824int rds_conn_init(void)
825{
826 int ret;
827
828 ret = rds_loop_net_init(); /* register pernet callback */
829 if (ret)
830 return ret;
831
832 rds_conn_slab = KMEM_CACHE(rds_connection, 0);
833 if (!rds_conn_slab) {
834 rds_loop_net_exit();
835 return -ENOMEM;
836 }
837
838 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
839 rds_info_register_func(RDS_INFO_SEND_MESSAGES,
840 rds_conn_message_info_send);
841 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
842 rds_conn_message_info_retrans);
843#if IS_ENABLED(CONFIG_IPV6)
844 rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
845 rds_info_register_func(RDS6_INFO_SEND_MESSAGES,
846 rds6_conn_message_info_send);
847 rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES,
848 rds6_conn_message_info_retrans);
849#endif
850 return 0;
851}
852
853void rds_conn_exit(void)
854{
855 rds_loop_net_exit(); /* unregister pernet callback */
856 rds_loop_exit();
857
858 WARN_ON(!hlist_empty(rds_conn_hash));
859
860 kmem_cache_destroy(rds_conn_slab);
861
862 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
863 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
864 rds_conn_message_info_send);
865 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
866 rds_conn_message_info_retrans);
867#if IS_ENABLED(CONFIG_IPV6)
868 rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
869 rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES,
870 rds6_conn_message_info_send);
871 rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES,
872 rds6_conn_message_info_retrans);
873#endif
874}
875
876/*
877 * Force a disconnect
878 */
879void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
880{
881 atomic_set(&cp->cp_state, RDS_CONN_ERROR);
882
883 rcu_read_lock();
884 if (!destroy && rds_destroy_pending(cp->cp_conn)) {
885 rcu_read_unlock();
886 return;
887 }
888 queue_work(rds_wq, &cp->cp_down_w);
889 rcu_read_unlock();
890}
891EXPORT_SYMBOL_GPL(rds_conn_path_drop);
892
893void rds_conn_drop(struct rds_connection *conn)
894{
895 WARN_ON(conn->c_trans->t_mp_capable);
896 rds_conn_path_drop(&conn->c_path[0], false);
897}
898EXPORT_SYMBOL_GPL(rds_conn_drop);
899
900/*
901 * If the connection is down, trigger a connect. We may have scheduled a
902 * delayed reconnect however - in this case we should not interfere.
903 */
904void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
905{
906 rcu_read_lock();
907 if (rds_destroy_pending(cp->cp_conn)) {
908 rcu_read_unlock();
909 return;
910 }
911 if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
912 !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
913 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
914 rcu_read_unlock();
915}
916EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
917
918/* Check connectivity of all paths
919 */
920void rds_check_all_paths(struct rds_connection *conn)
921{
922 int i = 0;
923
924 do {
925 rds_conn_path_connect_if_down(&conn->c_path[i]);
926 } while (++i < conn->c_npaths);
927}
928
929void rds_conn_connect_if_down(struct rds_connection *conn)
930{
931 WARN_ON(conn->c_trans->t_mp_capable);
932 rds_conn_path_connect_if_down(&conn->c_path[0]);
933}
934EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
935
936void
937__rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
938{
939 va_list ap;
940
941 va_start(ap, fmt);
942 vprintk(fmt, ap);
943 va_end(ap);
944
945 rds_conn_path_drop(cp, false);
946}