xref: /linux/net/rds/connection.c (revision 69050f8d6d075dc01af7a5f2f550a8067510366f)
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.. */
49 static DEFINE_SPINLOCK(rds_conn_lock);
50 static unsigned long rds_conn_count;
51 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
52 static struct kmem_cache *rds_conn_slab;
53 
54 static 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 	__be32 lhash, fhash;
61 	u32 hash;
62 
63 	net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret));
64 	net_get_random_once(&rds6_hash_secret, sizeof(rds6_hash_secret));
65 
66 	lhash = laddr->s6_addr32[3];
67 #if IS_ENABLED(CONFIG_IPV6)
68 	fhash = (__force __be32)__ipv6_addr_jhash(faddr, rds6_hash_secret);
69 #else
70 	fhash = faddr->s6_addr32[3];
71 #endif
72 	hash = __inet_ehashfn(lhash, 0, fhash, 0, rds_hash_secret);
73 
74 	return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
75 }
76 
77 #define rds_conn_info_set(var, test, suffix) do {		\
78 	if (test)						\
79 		var |= RDS_INFO_CONNECTION_FLAG_##suffix;	\
80 } while (0)
81 
82 /* rcu read lock must be held or the connection spinlock */
83 static struct rds_connection *rds_conn_lookup(struct net *net,
84 					      struct hlist_head *head,
85 					      const struct in6_addr *laddr,
86 					      const struct in6_addr *faddr,
87 					      struct rds_transport *trans,
88 					      u8 tos, int dev_if)
89 {
90 	struct rds_connection *conn, *ret = NULL;
91 
92 	hlist_for_each_entry_rcu(conn, head, c_hash_node) {
93 		if (ipv6_addr_equal(&conn->c_faddr, faddr) &&
94 		    ipv6_addr_equal(&conn->c_laddr, laddr) &&
95 		    conn->c_trans == trans &&
96 		    conn->c_tos == tos &&
97 		    net == rds_conn_net(conn) &&
98 		    conn->c_dev_if == dev_if) {
99 			ret = conn;
100 			break;
101 		}
102 	}
103 	rdsdebug("returning conn %p for %pI6c -> %pI6c\n", ret,
104 		 laddr, faddr);
105 	return ret;
106 }
107 
108 /*
109  * This is called by transports as they're bringing down a connection.
110  * It clears partial message state so that the transport can start sending
111  * and receiving over this connection again in the future.  It is up to
112  * the transport to have serialized this call with its send and recv.
113  */
114 static void rds_conn_path_reset(struct rds_conn_path *cp)
115 {
116 	struct rds_connection *conn = cp->cp_conn;
117 
118 	rdsdebug("connection %pI6c to %pI6c reset\n",
119 		 &conn->c_laddr, &conn->c_faddr);
120 
121 	rds_stats_inc(s_conn_reset);
122 	rds_send_path_reset(cp);
123 	cp->cp_flags = 0;
124 
125 	/* Do not clear next_rx_seq here, else we cannot distinguish
126 	 * retransmitted packets from new packets, and will hand all
127 	 * of them to the application. That is not consistent with the
128 	 * reliability guarantees of RDS. */
129 }
130 
131 static void __rds_conn_path_init(struct rds_connection *conn,
132 				 struct rds_conn_path *cp, bool is_outgoing)
133 {
134 	spin_lock_init(&cp->cp_lock);
135 	cp->cp_next_tx_seq = 1;
136 	init_waitqueue_head(&cp->cp_waitq);
137 	INIT_LIST_HEAD(&cp->cp_send_queue);
138 	INIT_LIST_HEAD(&cp->cp_retrans);
139 
140 	cp->cp_conn = conn;
141 	atomic_set(&cp->cp_state, RDS_CONN_DOWN);
142 	cp->cp_send_gen = 0;
143 	cp->cp_reconnect_jiffies = 0;
144 	cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
145 	INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker);
146 	INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker);
147 	INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker);
148 	INIT_WORK(&cp->cp_down_w, rds_shutdown_worker);
149 	mutex_init(&cp->cp_cm_lock);
150 	cp->cp_flags = 0;
151 }
152 
153 /*
154  * There is only every one 'conn' for a given pair of addresses in the
155  * system at a time.  They contain messages to be retransmitted and so
156  * span the lifetime of the actual underlying transport connections.
157  *
158  * For now they are not garbage collected once they're created.  They
159  * are torn down as the module is removed, if ever.
160  */
161 static struct rds_connection *__rds_conn_create(struct net *net,
162 						const struct in6_addr *laddr,
163 						const struct in6_addr *faddr,
164 						struct rds_transport *trans,
165 						gfp_t gfp, u8 tos,
166 						int is_outgoing,
167 						int dev_if)
168 {
169 	struct rds_connection *conn, *parent = NULL;
170 	struct hlist_head *head = rds_conn_bucket(laddr, faddr);
171 	struct rds_transport *loop_trans;
172 	struct rds_conn_path *free_cp = NULL;
173 	unsigned long flags;
174 	int ret, i;
175 	int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
176 
177 	rcu_read_lock();
178 	conn = rds_conn_lookup(net, head, laddr, faddr, trans, tos, dev_if);
179 	if (conn &&
180 	    conn->c_loopback &&
181 	    conn->c_trans != &rds_loop_transport &&
182 	    ipv6_addr_equal(laddr, faddr) &&
183 	    !is_outgoing) {
184 		/* This is a looped back IB connection, and we're
185 		 * called by the code handling the incoming connect.
186 		 * We need a second connection object into which we
187 		 * can stick the other QP. */
188 		parent = conn;
189 		conn = parent->c_passive;
190 	}
191 	rcu_read_unlock();
192 	if (conn)
193 		goto out;
194 
195 	conn = kmem_cache_zalloc(rds_conn_slab, gfp);
196 	if (!conn) {
197 		conn = ERR_PTR(-ENOMEM);
198 		goto out;
199 	}
200 	conn->c_path = kzalloc_objs(struct rds_conn_path, npaths, gfp);
201 	if (!conn->c_path) {
202 		kmem_cache_free(rds_conn_slab, conn);
203 		conn = ERR_PTR(-ENOMEM);
204 		goto out;
205 	}
206 
207 	INIT_HLIST_NODE(&conn->c_hash_node);
208 	conn->c_laddr = *laddr;
209 	conn->c_isv6 = !ipv6_addr_v4mapped(laddr);
210 	conn->c_faddr = *faddr;
211 	conn->c_dev_if = dev_if;
212 	conn->c_tos = tos;
213 
214 #if IS_ENABLED(CONFIG_IPV6)
215 	/* If the local address is link local, set c_bound_if to be the
216 	 * index used for this connection.  Otherwise, set it to 0 as
217 	 * the socket is not bound to an interface.  c_bound_if is used
218 	 * to look up a socket when a packet is received
219 	 */
220 	if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL)
221 		conn->c_bound_if = dev_if;
222 	else
223 #endif
224 		conn->c_bound_if = 0;
225 
226 	rds_conn_net_set(conn, net);
227 
228 	ret = rds_cong_get_maps(conn);
229 	if (ret) {
230 		kfree(conn->c_path);
231 		kmem_cache_free(rds_conn_slab, conn);
232 		conn = ERR_PTR(ret);
233 		goto out;
234 	}
235 
236 	/*
237 	 * This is where a connection becomes loopback.  If *any* RDS sockets
238 	 * can bind to the destination address then we'd rather the messages
239 	 * flow through loopback rather than either transport.
240 	 */
241 	loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if);
242 	if (loop_trans) {
243 		rds_trans_put(loop_trans);
244 		conn->c_loopback = 1;
245 		if (trans->t_prefer_loopback) {
246 			if (likely(is_outgoing)) {
247 				/* "outgoing" connection to local address.
248 				 * Protocol says it wants the connection
249 				 * handled by the loopback transport.
250 				 * This is what TCP does.
251 				 */
252 				trans = &rds_loop_transport;
253 			} else {
254 				/* No transport currently in use
255 				 * should end up here, but if it
256 				 * does, reset/destroy the connection.
257 				 */
258 				kfree(conn->c_path);
259 				kmem_cache_free(rds_conn_slab, conn);
260 				conn = ERR_PTR(-EOPNOTSUPP);
261 				goto out;
262 			}
263 		}
264 	}
265 
266 	conn->c_trans = trans;
267 
268 	init_waitqueue_head(&conn->c_hs_waitq);
269 	for (i = 0; i < npaths; i++) {
270 		__rds_conn_path_init(conn, &conn->c_path[i],
271 				     is_outgoing);
272 		conn->c_path[i].cp_index = i;
273 		conn->c_path[i].cp_wq =
274 			alloc_ordered_workqueue("krds_cp_wq#%lu/%d", 0,
275 						rds_conn_count, i);
276 		if (!conn->c_path[i].cp_wq)
277 			conn->c_path[i].cp_wq = rds_wq;
278 	}
279 	rcu_read_lock();
280 	if (rds_destroy_pending(conn))
281 		ret = -ENETDOWN;
282 	else
283 		ret = trans->conn_alloc(conn, GFP_ATOMIC);
284 	if (ret) {
285 		rcu_read_unlock();
286 		free_cp = conn->c_path;
287 		kmem_cache_free(rds_conn_slab, conn);
288 		conn = ERR_PTR(ret);
289 		goto out;
290 	}
291 
292 	rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n",
293 		 conn, laddr, faddr,
294 		 strnlen(trans->t_name, sizeof(trans->t_name)) ?
295 		 trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : "");
296 
297 	/*
298 	 * Since we ran without holding the conn lock, someone could
299 	 * have created the same conn (either normal or passive) in the
300 	 * interim. We check while holding the lock. If we won, we complete
301 	 * init and return our conn. If we lost, we rollback and return the
302 	 * other one.
303 	 */
304 	spin_lock_irqsave(&rds_conn_lock, flags);
305 	if (parent) {
306 		/* Creating passive conn */
307 		if (parent->c_passive) {
308 			trans->conn_free(conn->c_path[0].cp_transport_data);
309 			free_cp = conn->c_path;
310 			kmem_cache_free(rds_conn_slab, conn);
311 			conn = parent->c_passive;
312 		} else {
313 			parent->c_passive = conn;
314 			rds_cong_add_conn(conn);
315 			rds_conn_count++;
316 		}
317 	} else {
318 		/* Creating normal conn */
319 		struct rds_connection *found;
320 
321 		found = rds_conn_lookup(net, head, laddr, faddr, trans,
322 					tos, dev_if);
323 		if (found) {
324 			struct rds_conn_path *cp;
325 			int i;
326 
327 			for (i = 0; i < npaths; i++) {
328 				cp = &conn->c_path[i];
329 				/* The ->conn_alloc invocation may have
330 				 * allocated resource for all paths, so all
331 				 * of them may have to be freed here.
332 				 */
333 				if (cp->cp_transport_data)
334 					trans->conn_free(cp->cp_transport_data);
335 			}
336 			free_cp = conn->c_path;
337 			kmem_cache_free(rds_conn_slab, conn);
338 			conn = found;
339 		} else {
340 			conn->c_my_gen_num = rds_gen_num;
341 			conn->c_peer_gen_num = 0;
342 			hlist_add_head_rcu(&conn->c_hash_node, head);
343 			rds_cong_add_conn(conn);
344 			rds_conn_count++;
345 		}
346 	}
347 	spin_unlock_irqrestore(&rds_conn_lock, flags);
348 	rcu_read_unlock();
349 
350 out:
351 	if (free_cp) {
352 		for (i = 0; i < npaths; i++)
353 			if (free_cp[i].cp_wq != rds_wq)
354 				destroy_workqueue(free_cp[i].cp_wq);
355 		kfree(free_cp);
356 	}
357 
358 	return conn;
359 }
360 
361 struct rds_connection *rds_conn_create(struct net *net,
362 				       const struct in6_addr *laddr,
363 				       const struct in6_addr *faddr,
364 				       struct rds_transport *trans, u8 tos,
365 				       gfp_t gfp, int dev_if)
366 {
367 	return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 0, dev_if);
368 }
369 EXPORT_SYMBOL_GPL(rds_conn_create);
370 
371 struct rds_connection *rds_conn_create_outgoing(struct net *net,
372 						const struct in6_addr *laddr,
373 						const struct in6_addr *faddr,
374 						struct rds_transport *trans,
375 						u8 tos, gfp_t gfp, int dev_if)
376 {
377 	return __rds_conn_create(net, laddr, faddr, trans, gfp, tos, 1, dev_if);
378 }
379 EXPORT_SYMBOL_GPL(rds_conn_create_outgoing);
380 
381 void rds_conn_shutdown(struct rds_conn_path *cp)
382 {
383 	struct rds_connection *conn = cp->cp_conn;
384 
385 	/* shut it down unless it's down already */
386 	if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
387 		/*
388 		 * Quiesce the connection mgmt handlers before we start tearing
389 		 * things down. We don't hold the mutex for the entire
390 		 * duration of the shutdown operation, else we may be
391 		 * deadlocking with the CM handler. Instead, the CM event
392 		 * handler is supposed to check for state DISCONNECTING
393 		 */
394 		mutex_lock(&cp->cp_cm_lock);
395 		if (!rds_conn_path_transition(cp, RDS_CONN_UP,
396 					      RDS_CONN_DISCONNECTING) &&
397 		    !rds_conn_path_transition(cp, RDS_CONN_ERROR,
398 					      RDS_CONN_DISCONNECTING) &&
399 		    !rds_conn_path_transition(cp, RDS_CONN_RESETTING,
400 					      RDS_CONN_DISCONNECTING)) {
401 			rds_conn_path_error(cp,
402 					    "shutdown called in state %d\n",
403 					    atomic_read(&cp->cp_state));
404 			mutex_unlock(&cp->cp_cm_lock);
405 			return;
406 		}
407 		mutex_unlock(&cp->cp_cm_lock);
408 
409 		wait_event(cp->cp_waitq,
410 			   !test_bit(RDS_IN_XMIT, &cp->cp_flags));
411 		wait_event(cp->cp_waitq,
412 			   !test_bit(RDS_RECV_REFILL, &cp->cp_flags));
413 
414 		conn->c_trans->conn_path_shutdown(cp);
415 		rds_conn_path_reset(cp);
416 
417 		if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING,
418 					      RDS_CONN_DOWN) &&
419 		    !rds_conn_path_transition(cp, RDS_CONN_ERROR,
420 					      RDS_CONN_DOWN)) {
421 			/* This can happen - eg when we're in the middle of tearing
422 			 * down the connection, and someone unloads the rds module.
423 			 * Quite reproducible with loopback connections.
424 			 * Mostly harmless.
425 			 *
426 			 * Note that this also happens with rds-tcp because
427 			 * we could have triggered rds_conn_path_drop in irq
428 			 * mode from rds_tcp_state change on the receipt of
429 			 * a FIN, thus we need to recheck for RDS_CONN_ERROR
430 			 * here.
431 			 */
432 			rds_conn_path_error(cp, "%s: failed to transition "
433 					    "to state DOWN, current state "
434 					    "is %d\n", __func__,
435 					    atomic_read(&cp->cp_state));
436 			return;
437 		}
438 	}
439 
440 	/* Then reconnect if it's still live.
441 	 * The passive side of an IB loopback connection is never added
442 	 * to the conn hash, so we never trigger a reconnect on this
443 	 * conn - the reconnect is always triggered by the active peer. */
444 	cancel_delayed_work_sync(&cp->cp_conn_w);
445 
446 	clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
447 	rcu_read_lock();
448 	if (!hlist_unhashed(&conn->c_hash_node)) {
449 		rcu_read_unlock();
450 		if (conn->c_trans->t_mp_capable &&
451 		    cp->cp_index == 0)
452 			rds_send_ping(conn, 0);
453 		rds_queue_reconnect(cp);
454 	} else {
455 		rcu_read_unlock();
456 	}
457 
458 	if (conn->c_trans->conn_slots_available)
459 		conn->c_trans->conn_slots_available(conn, false);
460 }
461 
462 /* destroy a single rds_conn_path. rds_conn_destroy() iterates over
463  * all paths using rds_conn_path_destroy()
464  */
465 static void rds_conn_path_destroy(struct rds_conn_path *cp)
466 {
467 	struct rds_message *rm, *rtmp;
468 
469 	if (!cp->cp_transport_data)
470 		return;
471 
472 	/* make sure lingering queued work won't try to ref the conn */
473 	cancel_delayed_work_sync(&cp->cp_send_w);
474 	cancel_delayed_work_sync(&cp->cp_recv_w);
475 
476 	rds_conn_path_drop(cp, true);
477 	flush_work(&cp->cp_down_w);
478 
479 	/* tear down queued messages */
480 	list_for_each_entry_safe(rm, rtmp,
481 				 &cp->cp_send_queue,
482 				 m_conn_item) {
483 		list_del_init(&rm->m_conn_item);
484 		BUG_ON(!list_empty(&rm->m_sock_item));
485 		rds_message_put(rm);
486 	}
487 	if (cp->cp_xmit_rm)
488 		rds_message_put(cp->cp_xmit_rm);
489 
490 	WARN_ON(delayed_work_pending(&cp->cp_send_w));
491 	WARN_ON(delayed_work_pending(&cp->cp_recv_w));
492 	WARN_ON(delayed_work_pending(&cp->cp_conn_w));
493 	WARN_ON(work_pending(&cp->cp_down_w));
494 
495 	if (cp->cp_wq != rds_wq) {
496 		destroy_workqueue(cp->cp_wq);
497 		cp->cp_wq = NULL;
498 	}
499 
500 	cp->cp_conn->c_trans->conn_free(cp->cp_transport_data);
501 }
502 
503 /*
504  * Stop and free a connection.
505  *
506  * This can only be used in very limited circumstances.  It assumes that once
507  * the conn has been shutdown that no one else is referencing the connection.
508  * We can only ensure this in the rmmod path in the current code.
509  */
510 void rds_conn_destroy(struct rds_connection *conn)
511 {
512 	unsigned long flags;
513 	int i;
514 	struct rds_conn_path *cp;
515 	int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1);
516 
517 	rdsdebug("freeing conn %p for %pI4 -> "
518 		 "%pI4\n", conn, &conn->c_laddr,
519 		 &conn->c_faddr);
520 
521 	/* Ensure conn will not be scheduled for reconnect */
522 	spin_lock_irq(&rds_conn_lock);
523 	hlist_del_init_rcu(&conn->c_hash_node);
524 	spin_unlock_irq(&rds_conn_lock);
525 	synchronize_rcu();
526 
527 	/* shut the connection down */
528 	for (i = 0; i < npaths; i++) {
529 		cp = &conn->c_path[i];
530 		rds_conn_path_destroy(cp);
531 		BUG_ON(!list_empty(&cp->cp_retrans));
532 	}
533 
534 	/*
535 	 * The congestion maps aren't freed up here.  They're
536 	 * freed by rds_cong_exit() after all the connections
537 	 * have been freed.
538 	 */
539 	rds_cong_remove_conn(conn);
540 
541 	kfree(conn->c_path);
542 	kmem_cache_free(rds_conn_slab, conn);
543 
544 	spin_lock_irqsave(&rds_conn_lock, flags);
545 	rds_conn_count--;
546 	spin_unlock_irqrestore(&rds_conn_lock, flags);
547 }
548 EXPORT_SYMBOL_GPL(rds_conn_destroy);
549 
550 static void __rds_inc_msg_cp(struct rds_incoming *inc,
551 			     struct rds_info_iterator *iter,
552 			     void *saddr, void *daddr, int flip, bool isv6)
553 {
554 #if IS_ENABLED(CONFIG_IPV6)
555 	if (isv6)
556 		rds6_inc_info_copy(inc, iter, saddr, daddr, flip);
557 	else
558 #endif
559 		rds_inc_info_copy(inc, iter, *(__be32 *)saddr,
560 				  *(__be32 *)daddr, flip);
561 }
562 
563 static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len,
564 				      struct rds_info_iterator *iter,
565 				      struct rds_info_lengths *lens,
566 				      int want_send, bool isv6)
567 {
568 	struct hlist_head *head;
569 	struct list_head *list;
570 	struct rds_connection *conn;
571 	struct rds_message *rm;
572 	unsigned int total = 0;
573 	unsigned long flags;
574 	size_t i;
575 	int j;
576 
577 	if (isv6)
578 		len /= sizeof(struct rds6_info_message);
579 	else
580 		len /= sizeof(struct rds_info_message);
581 
582 	rcu_read_lock();
583 
584 	for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
585 	     i++, head++) {
586 		hlist_for_each_entry_rcu(conn, head, c_hash_node) {
587 			struct rds_conn_path *cp;
588 			int npaths;
589 
590 			if (!isv6 && conn->c_isv6)
591 				continue;
592 
593 			npaths = (conn->c_trans->t_mp_capable ?
594 				 RDS_MPATH_WORKERS : 1);
595 
596 			for (j = 0; j < npaths; j++) {
597 				cp = &conn->c_path[j];
598 				if (want_send)
599 					list = &cp->cp_send_queue;
600 				else
601 					list = &cp->cp_retrans;
602 
603 				spin_lock_irqsave(&cp->cp_lock, flags);
604 
605 				/* XXX too lazy to maintain counts.. */
606 				list_for_each_entry(rm, list, m_conn_item) {
607 					total++;
608 					if (total <= len)
609 						__rds_inc_msg_cp(&rm->m_inc,
610 								 iter,
611 								 &conn->c_laddr,
612 								 &conn->c_faddr,
613 								 0, isv6);
614 				}
615 
616 				spin_unlock_irqrestore(&cp->cp_lock, flags);
617 			}
618 		}
619 	}
620 	rcu_read_unlock();
621 
622 	lens->nr = total;
623 	if (isv6)
624 		lens->each = sizeof(struct rds6_info_message);
625 	else
626 		lens->each = sizeof(struct rds_info_message);
627 }
628 
629 static void rds_conn_message_info(struct socket *sock, unsigned int len,
630 				  struct rds_info_iterator *iter,
631 				  struct rds_info_lengths *lens,
632 				  int want_send)
633 {
634 	rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false);
635 }
636 
637 #if IS_ENABLED(CONFIG_IPV6)
638 static void rds6_conn_message_info(struct socket *sock, unsigned int len,
639 				   struct rds_info_iterator *iter,
640 				   struct rds_info_lengths *lens,
641 				   int want_send)
642 {
643 	rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true);
644 }
645 #endif
646 
647 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
648 				       struct rds_info_iterator *iter,
649 				       struct rds_info_lengths *lens)
650 {
651 	rds_conn_message_info(sock, len, iter, lens, 1);
652 }
653 
654 #if IS_ENABLED(CONFIG_IPV6)
655 static void rds6_conn_message_info_send(struct socket *sock, unsigned int len,
656 					struct rds_info_iterator *iter,
657 					struct rds_info_lengths *lens)
658 {
659 	rds6_conn_message_info(sock, len, iter, lens, 1);
660 }
661 #endif
662 
663 static void rds_conn_message_info_retrans(struct socket *sock,
664 					  unsigned int len,
665 					  struct rds_info_iterator *iter,
666 					  struct rds_info_lengths *lens)
667 {
668 	rds_conn_message_info(sock, len, iter, lens, 0);
669 }
670 
671 #if IS_ENABLED(CONFIG_IPV6)
672 static void rds6_conn_message_info_retrans(struct socket *sock,
673 					   unsigned int len,
674 					   struct rds_info_iterator *iter,
675 					   struct rds_info_lengths *lens)
676 {
677 	rds6_conn_message_info(sock, len, iter, lens, 0);
678 }
679 #endif
680 
681 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
682 			  struct rds_info_iterator *iter,
683 			  struct rds_info_lengths *lens,
684 			  int (*visitor)(struct rds_connection *, void *),
685 			  u64 *buffer,
686 			  size_t item_len)
687 {
688 	struct hlist_head *head;
689 	struct rds_connection *conn;
690 	size_t i;
691 
692 	rcu_read_lock();
693 
694 	lens->nr = 0;
695 	lens->each = item_len;
696 
697 	for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
698 	     i++, head++) {
699 		hlist_for_each_entry_rcu(conn, head, c_hash_node) {
700 
701 			/* XXX no c_lock usage.. */
702 			if (!visitor(conn, buffer))
703 				continue;
704 
705 			/* We copy as much as we can fit in the buffer,
706 			 * but we count all items so that the caller
707 			 * can resize the buffer. */
708 			if (len >= item_len) {
709 				rds_info_copy(iter, buffer, item_len);
710 				len -= item_len;
711 			}
712 			lens->nr++;
713 		}
714 	}
715 	rcu_read_unlock();
716 }
717 EXPORT_SYMBOL_GPL(rds_for_each_conn_info);
718 
719 static void rds_walk_conn_path_info(struct socket *sock, unsigned int len,
720 				    struct rds_info_iterator *iter,
721 				    struct rds_info_lengths *lens,
722 				    int (*visitor)(struct rds_conn_path *, void *),
723 				    u64 *buffer,
724 				    size_t item_len)
725 {
726 	struct hlist_head *head;
727 	struct rds_connection *conn;
728 	size_t i;
729 
730 	rcu_read_lock();
731 
732 	lens->nr = 0;
733 	lens->each = item_len;
734 
735 	for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
736 	     i++, head++) {
737 		hlist_for_each_entry_rcu(conn, head, c_hash_node) {
738 			struct rds_conn_path *cp;
739 
740 			/* XXX We only copy the information from the first
741 			 * path for now.  The problem is that if there are
742 			 * more than one underlying paths, we cannot report
743 			 * information of all of them using the existing
744 			 * API.  For example, there is only one next_tx_seq,
745 			 * which path's next_tx_seq should we report?  It is
746 			 * a bug in the design of MPRDS.
747 			 */
748 			cp = conn->c_path;
749 
750 			/* XXX no cp_lock usage.. */
751 			if (!visitor(cp, buffer))
752 				continue;
753 
754 			/* We copy as much as we can fit in the buffer,
755 			 * but we count all items so that the caller
756 			 * can resize the buffer.
757 			 */
758 			if (len >= item_len) {
759 				rds_info_copy(iter, buffer, item_len);
760 				len -= item_len;
761 			}
762 			lens->nr++;
763 		}
764 	}
765 	rcu_read_unlock();
766 }
767 
768 static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
769 {
770 	struct rds_info_connection *cinfo = buffer;
771 	struct rds_connection *conn = cp->cp_conn;
772 
773 	if (conn->c_isv6)
774 		return 0;
775 
776 	cinfo->next_tx_seq = cp->cp_next_tx_seq;
777 	cinfo->next_rx_seq = cp->cp_next_rx_seq;
778 	cinfo->laddr = conn->c_laddr.s6_addr32[3];
779 	cinfo->faddr = conn->c_faddr.s6_addr32[3];
780 	cinfo->tos = conn->c_tos;
781 	strscpy_pad(cinfo->transport, conn->c_trans->t_name);
782 	cinfo->flags = 0;
783 
784 	rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
785 			  SENDING);
786 	/* XXX Future: return the state rather than these funky bits */
787 	rds_conn_info_set(cinfo->flags,
788 			  atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
789 			  CONNECTING);
790 	rds_conn_info_set(cinfo->flags,
791 			  atomic_read(&cp->cp_state) == RDS_CONN_UP,
792 			  CONNECTED);
793 	return 1;
794 }
795 
796 #if IS_ENABLED(CONFIG_IPV6)
797 static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer)
798 {
799 	struct rds6_info_connection *cinfo6 = buffer;
800 	struct rds_connection *conn = cp->cp_conn;
801 
802 	cinfo6->next_tx_seq = cp->cp_next_tx_seq;
803 	cinfo6->next_rx_seq = cp->cp_next_rx_seq;
804 	cinfo6->laddr = conn->c_laddr;
805 	cinfo6->faddr = conn->c_faddr;
806 	strscpy_pad(cinfo6->transport, conn->c_trans->t_name);
807 	cinfo6->flags = 0;
808 
809 	rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags),
810 			  SENDING);
811 	/* XXX Future: return the state rather than these funky bits */
812 	rds_conn_info_set(cinfo6->flags,
813 			  atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING,
814 			  CONNECTING);
815 	rds_conn_info_set(cinfo6->flags,
816 			  atomic_read(&cp->cp_state) == RDS_CONN_UP,
817 			  CONNECTED);
818 	/* Just return 1 as there is no error case. This is a helper function
819 	 * for rds_walk_conn_path_info() and it wants a return value.
820 	 */
821 	return 1;
822 }
823 #endif
824 
825 static void rds_conn_info(struct socket *sock, unsigned int len,
826 			  struct rds_info_iterator *iter,
827 			  struct rds_info_lengths *lens)
828 {
829 	u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8];
830 
831 	rds_walk_conn_path_info(sock, len, iter, lens,
832 				rds_conn_info_visitor,
833 				buffer,
834 				sizeof(struct rds_info_connection));
835 }
836 
837 #if IS_ENABLED(CONFIG_IPV6)
838 static void rds6_conn_info(struct socket *sock, unsigned int len,
839 			   struct rds_info_iterator *iter,
840 			   struct rds_info_lengths *lens)
841 {
842 	u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8];
843 
844 	rds_walk_conn_path_info(sock, len, iter, lens,
845 				rds6_conn_info_visitor,
846 				buffer,
847 				sizeof(struct rds6_info_connection));
848 }
849 #endif
850 
851 int rds_conn_init(void)
852 {
853 	int ret;
854 
855 	ret = rds_loop_net_init(); /* register pernet callback */
856 	if (ret)
857 		return ret;
858 
859 	rds_conn_slab = KMEM_CACHE(rds_connection, 0);
860 	if (!rds_conn_slab) {
861 		rds_loop_net_exit();
862 		return -ENOMEM;
863 	}
864 
865 	rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
866 	rds_info_register_func(RDS_INFO_SEND_MESSAGES,
867 			       rds_conn_message_info_send);
868 	rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
869 			       rds_conn_message_info_retrans);
870 #if IS_ENABLED(CONFIG_IPV6)
871 	rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
872 	rds_info_register_func(RDS6_INFO_SEND_MESSAGES,
873 			       rds6_conn_message_info_send);
874 	rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES,
875 			       rds6_conn_message_info_retrans);
876 #endif
877 	return 0;
878 }
879 
880 void rds_conn_exit(void)
881 {
882 	rds_loop_net_exit(); /* unregister pernet callback */
883 	rds_loop_exit();
884 
885 	WARN_ON(!hlist_empty(rds_conn_hash));
886 
887 	kmem_cache_destroy(rds_conn_slab);
888 
889 	rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
890 	rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
891 				 rds_conn_message_info_send);
892 	rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
893 				 rds_conn_message_info_retrans);
894 #if IS_ENABLED(CONFIG_IPV6)
895 	rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info);
896 	rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES,
897 				 rds6_conn_message_info_send);
898 	rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES,
899 				 rds6_conn_message_info_retrans);
900 #endif
901 }
902 
903 /*
904  * Force a disconnect
905  */
906 void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy)
907 {
908 	atomic_set(&cp->cp_state, RDS_CONN_ERROR);
909 
910 	rcu_read_lock();
911 	if (!destroy && rds_destroy_pending(cp->cp_conn)) {
912 		rcu_read_unlock();
913 		return;
914 	}
915 	queue_work(cp->cp_wq, &cp->cp_down_w);
916 	rcu_read_unlock();
917 }
918 EXPORT_SYMBOL_GPL(rds_conn_path_drop);
919 
920 void rds_conn_drop(struct rds_connection *conn)
921 {
922 	WARN_ON(conn->c_trans->t_mp_capable);
923 	rds_conn_path_drop(&conn->c_path[0], false);
924 }
925 EXPORT_SYMBOL_GPL(rds_conn_drop);
926 
927 /*
928  * If the connection is down, trigger a connect. We may have scheduled a
929  * delayed reconnect however - in this case we should not interfere.
930  */
931 void rds_conn_path_connect_if_down(struct rds_conn_path *cp)
932 {
933 	rcu_read_lock();
934 	if (rds_destroy_pending(cp->cp_conn)) {
935 		rcu_read_unlock();
936 		return;
937 	}
938 	if (rds_conn_path_state(cp) == RDS_CONN_DOWN &&
939 	    !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags))
940 		queue_delayed_work(cp->cp_wq, &cp->cp_conn_w, 0);
941 	rcu_read_unlock();
942 }
943 EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down);
944 
945 /* Check connectivity of all paths
946  */
947 void rds_check_all_paths(struct rds_connection *conn)
948 {
949 	int i = 0;
950 
951 	do {
952 		rds_conn_path_connect_if_down(&conn->c_path[i]);
953 	} while (++i < conn->c_npaths);
954 }
955 
956 void rds_conn_connect_if_down(struct rds_connection *conn)
957 {
958 	WARN_ON(conn->c_trans->t_mp_capable);
959 	rds_conn_path_connect_if_down(&conn->c_path[0]);
960 }
961 EXPORT_SYMBOL_GPL(rds_conn_connect_if_down);
962 
963 void
964 __rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...)
965 {
966 	va_list ap;
967 
968 	va_start(ap, fmt);
969 	vprintk(fmt, ap);
970 	va_end(ap);
971 
972 	rds_conn_path_drop(cp, false);
973 }
974