xref: /openbmc/linux/fs/dlm/lowcomms.c (revision 5626af8f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
4 **
5 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
6 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
7 **
8 **
9 *******************************************************************************
10 ******************************************************************************/
11 
12 /*
13  * lowcomms.c
14  *
15  * This is the "low-level" comms layer.
16  *
17  * It is responsible for sending/receiving messages
18  * from other nodes in the cluster.
19  *
20  * Cluster nodes are referred to by their nodeids. nodeids are
21  * simply 32 bit numbers to the locking module - if they need to
22  * be expanded for the cluster infrastructure then that is its
23  * responsibility. It is this layer's
24  * responsibility to resolve these into IP address or
25  * whatever it needs for inter-node communication.
26  *
27  * The comms level is two kernel threads that deal mainly with
28  * the receiving of messages from other nodes and passing them
29  * up to the mid-level comms layer (which understands the
30  * message format) for execution by the locking core, and
31  * a send thread which does all the setting up of connections
32  * to remote nodes and the sending of data. Threads are not allowed
33  * to send their own data because it may cause them to wait in times
34  * of high load. Also, this way, the sending thread can collect together
35  * messages bound for one node and send them in one block.
36  *
37  * lowcomms will choose to use either TCP or SCTP as its transport layer
38  * depending on the configuration variable 'protocol'. This should be set
39  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
40  * cluster-wide mechanism as it must be the same on all nodes of the cluster
41  * for the DLM to function.
42  *
43  */
44 
45 #include <asm/ioctls.h>
46 #include <net/sock.h>
47 #include <net/tcp.h>
48 #include <linux/pagemap.h>
49 #include <linux/file.h>
50 #include <linux/mutex.h>
51 #include <linux/sctp.h>
52 #include <linux/slab.h>
53 #include <net/sctp/sctp.h>
54 #include <net/ipv6.h>
55 
56 #include <trace/events/dlm.h>
57 
58 #include "dlm_internal.h"
59 #include "lowcomms.h"
60 #include "midcomms.h"
61 #include "memory.h"
62 #include "config.h"
63 
64 #define NEEDED_RMEM (4*1024*1024)
65 
66 struct connection {
67 	struct socket *sock;	/* NULL if not connected */
68 	uint32_t nodeid;	/* So we know who we are in the list */
69 	/* this semaphore is used to allow parallel recv/send in read
70 	 * lock mode. When we release a sock we need to held the write lock.
71 	 *
72 	 * However this is locking code and not nice. When we remove the
73 	 * othercon handling we can look into other mechanism to synchronize
74 	 * io handling to call sock_release() at the right time.
75 	 */
76 	struct rw_semaphore sock_lock;
77 	unsigned long flags;
78 #define CF_APP_LIMITED 0
79 #define CF_RECV_PENDING 1
80 #define CF_SEND_PENDING 2
81 #define CF_RECV_INTR 3
82 #define CF_IO_STOP 4
83 #define CF_IS_OTHERCON 5
84 	struct list_head writequeue;  /* List of outgoing writequeue_entries */
85 	spinlock_t writequeue_lock;
86 	int retries;
87 	struct hlist_node list;
88 	/* due some connect()/accept() races we currently have this cross over
89 	 * connection attempt second connection for one node.
90 	 *
91 	 * There is a solution to avoid the race by introducing a connect
92 	 * rule as e.g. our_nodeid > nodeid_to_connect who is allowed to
93 	 * connect. Otherside can connect but will only be considered that
94 	 * the other side wants to have a reconnect.
95 	 *
96 	 * However changing to this behaviour will break backwards compatible.
97 	 * In a DLM protocol major version upgrade we should remove this!
98 	 */
99 	struct connection *othercon;
100 	struct work_struct rwork; /* receive worker */
101 	struct work_struct swork; /* send worker */
102 	unsigned char rx_leftover_buf[DLM_MAX_SOCKET_BUFSIZE];
103 	int rx_leftover;
104 	int mark;
105 	int addr_count;
106 	int curr_addr_index;
107 	struct sockaddr_storage addr[DLM_MAX_ADDR_COUNT];
108 	spinlock_t addrs_lock;
109 	struct rcu_head rcu;
110 };
111 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
112 
113 struct listen_connection {
114 	struct socket *sock;
115 	struct work_struct rwork;
116 };
117 
118 #define DLM_WQ_REMAIN_BYTES(e) (PAGE_SIZE - e->end)
119 #define DLM_WQ_LENGTH_BYTES(e) (e->end - e->offset)
120 
121 /* An entry waiting to be sent */
122 struct writequeue_entry {
123 	struct list_head list;
124 	struct page *page;
125 	int offset;
126 	int len;
127 	int end;
128 	int users;
129 	bool dirty;
130 	struct connection *con;
131 	struct list_head msgs;
132 	struct kref ref;
133 };
134 
135 struct dlm_msg {
136 	struct writequeue_entry *entry;
137 	struct dlm_msg *orig_msg;
138 	bool retransmit;
139 	void *ppc;
140 	int len;
141 	int idx; /* new()/commit() idx exchange */
142 
143 	struct list_head list;
144 	struct kref ref;
145 };
146 
147 struct processqueue_entry {
148 	unsigned char *buf;
149 	int nodeid;
150 	int buflen;
151 
152 	struct list_head list;
153 };
154 
155 struct dlm_proto_ops {
156 	bool try_new_addr;
157 	const char *name;
158 	int proto;
159 
160 	int (*connect)(struct connection *con, struct socket *sock,
161 		       struct sockaddr *addr, int addr_len);
162 	void (*sockopts)(struct socket *sock);
163 	int (*bind)(struct socket *sock);
164 	int (*listen_validate)(void);
165 	void (*listen_sockopts)(struct socket *sock);
166 	int (*listen_bind)(struct socket *sock);
167 };
168 
169 static struct listen_sock_callbacks {
170 	void (*sk_error_report)(struct sock *);
171 	void (*sk_data_ready)(struct sock *);
172 	void (*sk_state_change)(struct sock *);
173 	void (*sk_write_space)(struct sock *);
174 } listen_sock;
175 
176 static struct listen_connection listen_con;
177 static struct sockaddr_storage dlm_local_addr[DLM_MAX_ADDR_COUNT];
178 static int dlm_local_count;
179 
180 /* Work queues */
181 static struct workqueue_struct *io_workqueue;
182 static struct workqueue_struct *process_workqueue;
183 
184 static struct hlist_head connection_hash[CONN_HASH_SIZE];
185 static DEFINE_SPINLOCK(connections_lock);
186 DEFINE_STATIC_SRCU(connections_srcu);
187 
188 static const struct dlm_proto_ops *dlm_proto_ops;
189 
190 #define DLM_IO_SUCCESS 0
191 #define DLM_IO_END 1
192 #define DLM_IO_EOF 2
193 #define DLM_IO_RESCHED 3
194 
195 static void process_recv_sockets(struct work_struct *work);
196 static void process_send_sockets(struct work_struct *work);
197 static void process_dlm_messages(struct work_struct *work);
198 
199 static DECLARE_WORK(process_work, process_dlm_messages);
200 static DEFINE_SPINLOCK(processqueue_lock);
201 static bool process_dlm_messages_pending;
202 static LIST_HEAD(processqueue);
203 
204 bool dlm_lowcomms_is_running(void)
205 {
206 	return !!listen_con.sock;
207 }
208 
209 static void lowcomms_queue_swork(struct connection *con)
210 {
211 	assert_spin_locked(&con->writequeue_lock);
212 
213 	if (!test_bit(CF_IO_STOP, &con->flags) &&
214 	    !test_bit(CF_APP_LIMITED, &con->flags) &&
215 	    !test_and_set_bit(CF_SEND_PENDING, &con->flags))
216 		queue_work(io_workqueue, &con->swork);
217 }
218 
219 static void lowcomms_queue_rwork(struct connection *con)
220 {
221 #ifdef CONFIG_LOCKDEP
222 	WARN_ON_ONCE(!lockdep_sock_is_held(con->sock->sk));
223 #endif
224 
225 	if (!test_bit(CF_IO_STOP, &con->flags) &&
226 	    !test_and_set_bit(CF_RECV_PENDING, &con->flags))
227 		queue_work(io_workqueue, &con->rwork);
228 }
229 
230 static void writequeue_entry_ctor(void *data)
231 {
232 	struct writequeue_entry *entry = data;
233 
234 	INIT_LIST_HEAD(&entry->msgs);
235 }
236 
237 struct kmem_cache *dlm_lowcomms_writequeue_cache_create(void)
238 {
239 	return kmem_cache_create("dlm_writequeue", sizeof(struct writequeue_entry),
240 				 0, 0, writequeue_entry_ctor);
241 }
242 
243 struct kmem_cache *dlm_lowcomms_msg_cache_create(void)
244 {
245 	return kmem_cache_create("dlm_msg", sizeof(struct dlm_msg), 0, 0, NULL);
246 }
247 
248 /* need to held writequeue_lock */
249 static struct writequeue_entry *con_next_wq(struct connection *con)
250 {
251 	struct writequeue_entry *e;
252 
253 	e = list_first_entry_or_null(&con->writequeue, struct writequeue_entry,
254 				     list);
255 	/* if len is zero nothing is to send, if there are users filling
256 	 * buffers we wait until the users are done so we can send more.
257 	 */
258 	if (!e || e->users || e->len == 0)
259 		return NULL;
260 
261 	return e;
262 }
263 
264 static struct connection *__find_con(int nodeid, int r)
265 {
266 	struct connection *con;
267 
268 	hlist_for_each_entry_rcu(con, &connection_hash[r], list) {
269 		if (con->nodeid == nodeid)
270 			return con;
271 	}
272 
273 	return NULL;
274 }
275 
276 static void dlm_con_init(struct connection *con, int nodeid)
277 {
278 	con->nodeid = nodeid;
279 	init_rwsem(&con->sock_lock);
280 	INIT_LIST_HEAD(&con->writequeue);
281 	spin_lock_init(&con->writequeue_lock);
282 	INIT_WORK(&con->swork, process_send_sockets);
283 	INIT_WORK(&con->rwork, process_recv_sockets);
284 	spin_lock_init(&con->addrs_lock);
285 }
286 
287 /*
288  * If 'allocation' is zero then we don't attempt to create a new
289  * connection structure for this node.
290  */
291 static struct connection *nodeid2con(int nodeid, gfp_t alloc)
292 {
293 	struct connection *con, *tmp;
294 	int r;
295 
296 	r = nodeid_hash(nodeid);
297 	con = __find_con(nodeid, r);
298 	if (con || !alloc)
299 		return con;
300 
301 	con = kzalloc(sizeof(*con), alloc);
302 	if (!con)
303 		return NULL;
304 
305 	dlm_con_init(con, nodeid);
306 
307 	spin_lock(&connections_lock);
308 	/* Because multiple workqueues/threads calls this function it can
309 	 * race on multiple cpu's. Instead of locking hot path __find_con()
310 	 * we just check in rare cases of recently added nodes again
311 	 * under protection of connections_lock. If this is the case we
312 	 * abort our connection creation and return the existing connection.
313 	 */
314 	tmp = __find_con(nodeid, r);
315 	if (tmp) {
316 		spin_unlock(&connections_lock);
317 		kfree(con);
318 		return tmp;
319 	}
320 
321 	hlist_add_head_rcu(&con->list, &connection_hash[r]);
322 	spin_unlock(&connections_lock);
323 
324 	return con;
325 }
326 
327 static int addr_compare(const struct sockaddr_storage *x,
328 			const struct sockaddr_storage *y)
329 {
330 	switch (x->ss_family) {
331 	case AF_INET: {
332 		struct sockaddr_in *sinx = (struct sockaddr_in *)x;
333 		struct sockaddr_in *siny = (struct sockaddr_in *)y;
334 		if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
335 			return 0;
336 		if (sinx->sin_port != siny->sin_port)
337 			return 0;
338 		break;
339 	}
340 	case AF_INET6: {
341 		struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
342 		struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
343 		if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
344 			return 0;
345 		if (sinx->sin6_port != siny->sin6_port)
346 			return 0;
347 		break;
348 	}
349 	default:
350 		return 0;
351 	}
352 	return 1;
353 }
354 
355 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
356 			  struct sockaddr *sa_out, bool try_new_addr,
357 			  unsigned int *mark)
358 {
359 	struct sockaddr_storage sas;
360 	struct connection *con;
361 	int idx;
362 
363 	if (!dlm_local_count)
364 		return -1;
365 
366 	idx = srcu_read_lock(&connections_srcu);
367 	con = nodeid2con(nodeid, 0);
368 	if (!con) {
369 		srcu_read_unlock(&connections_srcu, idx);
370 		return -ENOENT;
371 	}
372 
373 	spin_lock(&con->addrs_lock);
374 	if (!con->addr_count) {
375 		spin_unlock(&con->addrs_lock);
376 		srcu_read_unlock(&connections_srcu, idx);
377 		return -ENOENT;
378 	}
379 
380 	memcpy(&sas, &con->addr[con->curr_addr_index],
381 	       sizeof(struct sockaddr_storage));
382 
383 	if (try_new_addr) {
384 		con->curr_addr_index++;
385 		if (con->curr_addr_index == con->addr_count)
386 			con->curr_addr_index = 0;
387 	}
388 
389 	*mark = con->mark;
390 	spin_unlock(&con->addrs_lock);
391 
392 	if (sas_out)
393 		memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
394 
395 	if (!sa_out) {
396 		srcu_read_unlock(&connections_srcu, idx);
397 		return 0;
398 	}
399 
400 	if (dlm_local_addr[0].ss_family == AF_INET) {
401 		struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
402 		struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
403 		ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
404 	} else {
405 		struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
406 		struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
407 		ret6->sin6_addr = in6->sin6_addr;
408 	}
409 
410 	srcu_read_unlock(&connections_srcu, idx);
411 	return 0;
412 }
413 
414 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid,
415 			  unsigned int *mark)
416 {
417 	struct connection *con;
418 	int i, idx, addr_i;
419 
420 	idx = srcu_read_lock(&connections_srcu);
421 	for (i = 0; i < CONN_HASH_SIZE; i++) {
422 		hlist_for_each_entry_rcu(con, &connection_hash[i], list) {
423 			WARN_ON_ONCE(!con->addr_count);
424 
425 			spin_lock(&con->addrs_lock);
426 			for (addr_i = 0; addr_i < con->addr_count; addr_i++) {
427 				if (addr_compare(&con->addr[addr_i], addr)) {
428 					*nodeid = con->nodeid;
429 					*mark = con->mark;
430 					spin_unlock(&con->addrs_lock);
431 					srcu_read_unlock(&connections_srcu, idx);
432 					return 0;
433 				}
434 			}
435 			spin_unlock(&con->addrs_lock);
436 		}
437 	}
438 	srcu_read_unlock(&connections_srcu, idx);
439 
440 	return -ENOENT;
441 }
442 
443 static bool dlm_lowcomms_con_has_addr(const struct connection *con,
444 				      const struct sockaddr_storage *addr)
445 {
446 	int i;
447 
448 	for (i = 0; i < con->addr_count; i++) {
449 		if (addr_compare(&con->addr[i], addr))
450 			return true;
451 	}
452 
453 	return false;
454 }
455 
456 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
457 {
458 	struct connection *con;
459 	bool ret, idx;
460 
461 	idx = srcu_read_lock(&connections_srcu);
462 	con = nodeid2con(nodeid, GFP_NOFS);
463 	if (!con) {
464 		srcu_read_unlock(&connections_srcu, idx);
465 		return -ENOMEM;
466 	}
467 
468 	spin_lock(&con->addrs_lock);
469 	if (!con->addr_count) {
470 		memcpy(&con->addr[0], addr, sizeof(*addr));
471 		con->addr_count = 1;
472 		con->mark = dlm_config.ci_mark;
473 		spin_unlock(&con->addrs_lock);
474 		srcu_read_unlock(&connections_srcu, idx);
475 		return 0;
476 	}
477 
478 	ret = dlm_lowcomms_con_has_addr(con, addr);
479 	if (ret) {
480 		spin_unlock(&con->addrs_lock);
481 		srcu_read_unlock(&connections_srcu, idx);
482 		return -EEXIST;
483 	}
484 
485 	if (con->addr_count >= DLM_MAX_ADDR_COUNT) {
486 		spin_unlock(&con->addrs_lock);
487 		srcu_read_unlock(&connections_srcu, idx);
488 		return -ENOSPC;
489 	}
490 
491 	memcpy(&con->addr[con->addr_count++], addr, sizeof(*addr));
492 	srcu_read_unlock(&connections_srcu, idx);
493 	spin_unlock(&con->addrs_lock);
494 	return 0;
495 }
496 
497 /* Data available on socket or listen socket received a connect */
498 static void lowcomms_data_ready(struct sock *sk)
499 {
500 	struct connection *con = sock2con(sk);
501 
502 	set_bit(CF_RECV_INTR, &con->flags);
503 	lowcomms_queue_rwork(con);
504 }
505 
506 static void lowcomms_write_space(struct sock *sk)
507 {
508 	struct connection *con = sock2con(sk);
509 
510 	clear_bit(SOCK_NOSPACE, &con->sock->flags);
511 
512 	spin_lock_bh(&con->writequeue_lock);
513 	if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
514 		con->sock->sk->sk_write_pending--;
515 		clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
516 	}
517 
518 	lowcomms_queue_swork(con);
519 	spin_unlock_bh(&con->writequeue_lock);
520 }
521 
522 static void lowcomms_state_change(struct sock *sk)
523 {
524 	/* SCTP layer is not calling sk_data_ready when the connection
525 	 * is done, so we catch the signal through here.
526 	 */
527 	if (sk->sk_shutdown == RCV_SHUTDOWN)
528 		lowcomms_data_ready(sk);
529 }
530 
531 static void lowcomms_listen_data_ready(struct sock *sk)
532 {
533 	queue_work(io_workqueue, &listen_con.rwork);
534 }
535 
536 int dlm_lowcomms_connect_node(int nodeid)
537 {
538 	struct connection *con;
539 	int idx;
540 
541 	if (nodeid == dlm_our_nodeid())
542 		return 0;
543 
544 	idx = srcu_read_lock(&connections_srcu);
545 	con = nodeid2con(nodeid, 0);
546 	if (WARN_ON_ONCE(!con)) {
547 		srcu_read_unlock(&connections_srcu, idx);
548 		return -ENOENT;
549 	}
550 
551 	down_read(&con->sock_lock);
552 	if (!con->sock) {
553 		spin_lock_bh(&con->writequeue_lock);
554 		lowcomms_queue_swork(con);
555 		spin_unlock_bh(&con->writequeue_lock);
556 	}
557 	up_read(&con->sock_lock);
558 	srcu_read_unlock(&connections_srcu, idx);
559 
560 	cond_resched();
561 	return 0;
562 }
563 
564 int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark)
565 {
566 	struct connection *con;
567 	int idx;
568 
569 	idx = srcu_read_lock(&connections_srcu);
570 	con = nodeid2con(nodeid, 0);
571 	if (!con) {
572 		srcu_read_unlock(&connections_srcu, idx);
573 		return -ENOENT;
574 	}
575 
576 	spin_lock(&con->addrs_lock);
577 	con->mark = mark;
578 	spin_unlock(&con->addrs_lock);
579 	srcu_read_unlock(&connections_srcu, idx);
580 	return 0;
581 }
582 
583 static void lowcomms_error_report(struct sock *sk)
584 {
585 	struct connection *con = sock2con(sk);
586 	struct inet_sock *inet;
587 
588 	inet = inet_sk(sk);
589 	switch (sk->sk_family) {
590 	case AF_INET:
591 		printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
592 				   "sending to node %d at %pI4, dport %d, "
593 				   "sk_err=%d/%d\n", dlm_our_nodeid(),
594 				   con->nodeid, &inet->inet_daddr,
595 				   ntohs(inet->inet_dport), sk->sk_err,
596 				   sk->sk_err_soft);
597 		break;
598 #if IS_ENABLED(CONFIG_IPV6)
599 	case AF_INET6:
600 		printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
601 				   "sending to node %d at %pI6c, "
602 				   "dport %d, sk_err=%d/%d\n", dlm_our_nodeid(),
603 				   con->nodeid, &sk->sk_v6_daddr,
604 				   ntohs(inet->inet_dport), sk->sk_err,
605 				   sk->sk_err_soft);
606 		break;
607 #endif
608 	default:
609 		printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
610 				   "invalid socket family %d set, "
611 				   "sk_err=%d/%d\n", dlm_our_nodeid(),
612 				   sk->sk_family, sk->sk_err, sk->sk_err_soft);
613 		break;
614 	}
615 
616 	dlm_midcomms_unack_msg_resend(con->nodeid);
617 
618 	listen_sock.sk_error_report(sk);
619 }
620 
621 static void restore_callbacks(struct sock *sk)
622 {
623 #ifdef CONFIG_LOCKDEP
624 	WARN_ON_ONCE(!lockdep_sock_is_held(sk));
625 #endif
626 
627 	sk->sk_user_data = NULL;
628 	sk->sk_data_ready = listen_sock.sk_data_ready;
629 	sk->sk_state_change = listen_sock.sk_state_change;
630 	sk->sk_write_space = listen_sock.sk_write_space;
631 	sk->sk_error_report = listen_sock.sk_error_report;
632 }
633 
634 /* Make a socket active */
635 static void add_sock(struct socket *sock, struct connection *con)
636 {
637 	struct sock *sk = sock->sk;
638 
639 	lock_sock(sk);
640 	con->sock = sock;
641 
642 	sk->sk_user_data = con;
643 	sk->sk_data_ready = lowcomms_data_ready;
644 	sk->sk_write_space = lowcomms_write_space;
645 	if (dlm_config.ci_protocol == DLM_PROTO_SCTP)
646 		sk->sk_state_change = lowcomms_state_change;
647 	sk->sk_allocation = GFP_NOFS;
648 	sk->sk_use_task_frag = false;
649 	sk->sk_error_report = lowcomms_error_report;
650 	release_sock(sk);
651 }
652 
653 /* Add the port number to an IPv6 or 4 sockaddr and return the address
654    length */
655 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
656 			  int *addr_len)
657 {
658 	saddr->ss_family =  dlm_local_addr[0].ss_family;
659 	if (saddr->ss_family == AF_INET) {
660 		struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
661 		in4_addr->sin_port = cpu_to_be16(port);
662 		*addr_len = sizeof(struct sockaddr_in);
663 		memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
664 	} else {
665 		struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
666 		in6_addr->sin6_port = cpu_to_be16(port);
667 		*addr_len = sizeof(struct sockaddr_in6);
668 	}
669 	memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
670 }
671 
672 static void dlm_page_release(struct kref *kref)
673 {
674 	struct writequeue_entry *e = container_of(kref, struct writequeue_entry,
675 						  ref);
676 
677 	__free_page(e->page);
678 	dlm_free_writequeue(e);
679 }
680 
681 static void dlm_msg_release(struct kref *kref)
682 {
683 	struct dlm_msg *msg = container_of(kref, struct dlm_msg, ref);
684 
685 	kref_put(&msg->entry->ref, dlm_page_release);
686 	dlm_free_msg(msg);
687 }
688 
689 static void free_entry(struct writequeue_entry *e)
690 {
691 	struct dlm_msg *msg, *tmp;
692 
693 	list_for_each_entry_safe(msg, tmp, &e->msgs, list) {
694 		if (msg->orig_msg) {
695 			msg->orig_msg->retransmit = false;
696 			kref_put(&msg->orig_msg->ref, dlm_msg_release);
697 		}
698 
699 		list_del(&msg->list);
700 		kref_put(&msg->ref, dlm_msg_release);
701 	}
702 
703 	list_del(&e->list);
704 	kref_put(&e->ref, dlm_page_release);
705 }
706 
707 static void dlm_close_sock(struct socket **sock)
708 {
709 	lock_sock((*sock)->sk);
710 	restore_callbacks((*sock)->sk);
711 	release_sock((*sock)->sk);
712 
713 	sock_release(*sock);
714 	*sock = NULL;
715 }
716 
717 static void allow_connection_io(struct connection *con)
718 {
719 	if (con->othercon)
720 		clear_bit(CF_IO_STOP, &con->othercon->flags);
721 	clear_bit(CF_IO_STOP, &con->flags);
722 }
723 
724 static void stop_connection_io(struct connection *con)
725 {
726 	if (con->othercon)
727 		stop_connection_io(con->othercon);
728 
729 	down_write(&con->sock_lock);
730 	if (con->sock) {
731 		lock_sock(con->sock->sk);
732 		restore_callbacks(con->sock->sk);
733 
734 		spin_lock_bh(&con->writequeue_lock);
735 		set_bit(CF_IO_STOP, &con->flags);
736 		spin_unlock_bh(&con->writequeue_lock);
737 		release_sock(con->sock->sk);
738 	} else {
739 		spin_lock_bh(&con->writequeue_lock);
740 		set_bit(CF_IO_STOP, &con->flags);
741 		spin_unlock_bh(&con->writequeue_lock);
742 	}
743 	up_write(&con->sock_lock);
744 
745 	cancel_work_sync(&con->swork);
746 	cancel_work_sync(&con->rwork);
747 }
748 
749 /* Close a remote connection and tidy up */
750 static void close_connection(struct connection *con, bool and_other)
751 {
752 	struct writequeue_entry *e;
753 
754 	if (con->othercon && and_other)
755 		close_connection(con->othercon, false);
756 
757 	down_write(&con->sock_lock);
758 	if (!con->sock) {
759 		up_write(&con->sock_lock);
760 		return;
761 	}
762 
763 	dlm_close_sock(&con->sock);
764 
765 	/* if we send a writequeue entry only a half way, we drop the
766 	 * whole entry because reconnection and that we not start of the
767 	 * middle of a msg which will confuse the other end.
768 	 *
769 	 * we can always drop messages because retransmits, but what we
770 	 * cannot allow is to transmit half messages which may be processed
771 	 * at the other side.
772 	 *
773 	 * our policy is to start on a clean state when disconnects, we don't
774 	 * know what's send/received on transport layer in this case.
775 	 */
776 	spin_lock_bh(&con->writequeue_lock);
777 	if (!list_empty(&con->writequeue)) {
778 		e = list_first_entry(&con->writequeue, struct writequeue_entry,
779 				     list);
780 		if (e->dirty)
781 			free_entry(e);
782 	}
783 	spin_unlock_bh(&con->writequeue_lock);
784 
785 	con->rx_leftover = 0;
786 	con->retries = 0;
787 	clear_bit(CF_APP_LIMITED, &con->flags);
788 	clear_bit(CF_RECV_PENDING, &con->flags);
789 	clear_bit(CF_SEND_PENDING, &con->flags);
790 	up_write(&con->sock_lock);
791 }
792 
793 static struct processqueue_entry *new_processqueue_entry(int nodeid,
794 							 int buflen)
795 {
796 	struct processqueue_entry *pentry;
797 
798 	pentry = kmalloc(sizeof(*pentry), GFP_NOFS);
799 	if (!pentry)
800 		return NULL;
801 
802 	pentry->buf = kmalloc(buflen, GFP_NOFS);
803 	if (!pentry->buf) {
804 		kfree(pentry);
805 		return NULL;
806 	}
807 
808 	pentry->nodeid = nodeid;
809 	return pentry;
810 }
811 
812 static void free_processqueue_entry(struct processqueue_entry *pentry)
813 {
814 	kfree(pentry->buf);
815 	kfree(pentry);
816 }
817 
818 struct dlm_processed_nodes {
819 	int nodeid;
820 
821 	struct list_head list;
822 };
823 
824 static void add_processed_node(int nodeid, struct list_head *processed_nodes)
825 {
826 	struct dlm_processed_nodes *n;
827 
828 	list_for_each_entry(n, processed_nodes, list) {
829 		/* we already remembered this node */
830 		if (n->nodeid == nodeid)
831 			return;
832 	}
833 
834 	/* if it's fails in worst case we simple don't send an ack back.
835 	 * We try it next time.
836 	 */
837 	n = kmalloc(sizeof(*n), GFP_NOFS);
838 	if (!n)
839 		return;
840 
841 	n->nodeid = nodeid;
842 	list_add(&n->list, processed_nodes);
843 }
844 
845 static void process_dlm_messages(struct work_struct *work)
846 {
847 	struct dlm_processed_nodes *n, *n_tmp;
848 	struct processqueue_entry *pentry;
849 	LIST_HEAD(processed_nodes);
850 
851 	spin_lock(&processqueue_lock);
852 	pentry = list_first_entry_or_null(&processqueue,
853 					  struct processqueue_entry, list);
854 	if (WARN_ON_ONCE(!pentry)) {
855 		spin_unlock(&processqueue_lock);
856 		return;
857 	}
858 
859 	list_del(&pentry->list);
860 	spin_unlock(&processqueue_lock);
861 
862 	for (;;) {
863 		dlm_process_incoming_buffer(pentry->nodeid, pentry->buf,
864 					    pentry->buflen);
865 		add_processed_node(pentry->nodeid, &processed_nodes);
866 		free_processqueue_entry(pentry);
867 
868 		spin_lock(&processqueue_lock);
869 		pentry = list_first_entry_or_null(&processqueue,
870 						  struct processqueue_entry, list);
871 		if (!pentry) {
872 			process_dlm_messages_pending = false;
873 			spin_unlock(&processqueue_lock);
874 			break;
875 		}
876 
877 		list_del(&pentry->list);
878 		spin_unlock(&processqueue_lock);
879 	}
880 
881 	/* send ack back after we processed couple of messages */
882 	list_for_each_entry_safe(n, n_tmp, &processed_nodes, list) {
883 		list_del(&n->list);
884 		dlm_midcomms_receive_done(n->nodeid);
885 		kfree(n);
886 	}
887 }
888 
889 /* Data received from remote end */
890 static int receive_from_sock(struct connection *con, int buflen)
891 {
892 	struct processqueue_entry *pentry;
893 	int ret, buflen_real;
894 	struct msghdr msg;
895 	struct kvec iov;
896 
897 	pentry = new_processqueue_entry(con->nodeid, buflen);
898 	if (!pentry)
899 		return DLM_IO_RESCHED;
900 
901 	memcpy(pentry->buf, con->rx_leftover_buf, con->rx_leftover);
902 
903 	/* calculate new buffer parameter regarding last receive and
904 	 * possible leftover bytes
905 	 */
906 	iov.iov_base = pentry->buf + con->rx_leftover;
907 	iov.iov_len = buflen - con->rx_leftover;
908 
909 	memset(&msg, 0, sizeof(msg));
910 	msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
911 	clear_bit(CF_RECV_INTR, &con->flags);
912 again:
913 	ret = kernel_recvmsg(con->sock, &msg, &iov, 1, iov.iov_len,
914 			     msg.msg_flags);
915 	trace_dlm_recv(con->nodeid, ret);
916 	if (ret == -EAGAIN) {
917 		lock_sock(con->sock->sk);
918 		if (test_and_clear_bit(CF_RECV_INTR, &con->flags)) {
919 			release_sock(con->sock->sk);
920 			goto again;
921 		}
922 
923 		clear_bit(CF_RECV_PENDING, &con->flags);
924 		release_sock(con->sock->sk);
925 		free_processqueue_entry(pentry);
926 		return DLM_IO_END;
927 	} else if (ret == 0) {
928 		/* close will clear CF_RECV_PENDING */
929 		free_processqueue_entry(pentry);
930 		return DLM_IO_EOF;
931 	} else if (ret < 0) {
932 		free_processqueue_entry(pentry);
933 		return ret;
934 	}
935 
936 	/* new buflen according readed bytes and leftover from last receive */
937 	buflen_real = ret + con->rx_leftover;
938 	ret = dlm_validate_incoming_buffer(con->nodeid, pentry->buf,
939 					   buflen_real);
940 	if (ret < 0) {
941 		free_processqueue_entry(pentry);
942 		return ret;
943 	}
944 
945 	pentry->buflen = ret;
946 
947 	/* calculate leftover bytes from process and put it into begin of
948 	 * the receive buffer, so next receive we have the full message
949 	 * at the start address of the receive buffer.
950 	 */
951 	con->rx_leftover = buflen_real - ret;
952 	memmove(con->rx_leftover_buf, pentry->buf + ret,
953 		con->rx_leftover);
954 
955 	spin_lock(&processqueue_lock);
956 	list_add_tail(&pentry->list, &processqueue);
957 	if (!process_dlm_messages_pending) {
958 		process_dlm_messages_pending = true;
959 		queue_work(process_workqueue, &process_work);
960 	}
961 	spin_unlock(&processqueue_lock);
962 
963 	return DLM_IO_SUCCESS;
964 }
965 
966 /* Listening socket is busy, accept a connection */
967 static int accept_from_sock(void)
968 {
969 	struct sockaddr_storage peeraddr;
970 	int len, idx, result, nodeid;
971 	struct connection *newcon;
972 	struct socket *newsock;
973 	unsigned int mark;
974 
975 	result = kernel_accept(listen_con.sock, &newsock, O_NONBLOCK);
976 	if (result == -EAGAIN)
977 		return DLM_IO_END;
978 	else if (result < 0)
979 		goto accept_err;
980 
981 	/* Get the connected socket's peer */
982 	memset(&peeraddr, 0, sizeof(peeraddr));
983 	len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
984 	if (len < 0) {
985 		result = -ECONNABORTED;
986 		goto accept_err;
987 	}
988 
989 	/* Get the new node's NODEID */
990 	make_sockaddr(&peeraddr, 0, &len);
991 	if (addr_to_nodeid(&peeraddr, &nodeid, &mark)) {
992 		switch (peeraddr.ss_family) {
993 		case AF_INET: {
994 			struct sockaddr_in *sin = (struct sockaddr_in *)&peeraddr;
995 
996 			log_print("connect from non cluster IPv4 node %pI4",
997 				  &sin->sin_addr);
998 			break;
999 		}
1000 #if IS_ENABLED(CONFIG_IPV6)
1001 		case AF_INET6: {
1002 			struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&peeraddr;
1003 
1004 			log_print("connect from non cluster IPv6 node %pI6c",
1005 				  &sin6->sin6_addr);
1006 			break;
1007 		}
1008 #endif
1009 		default:
1010 			log_print("invalid family from non cluster node");
1011 			break;
1012 		}
1013 
1014 		sock_release(newsock);
1015 		return -1;
1016 	}
1017 
1018 	log_print("got connection from %d", nodeid);
1019 
1020 	/*  Check to see if we already have a connection to this node. This
1021 	 *  could happen if the two nodes initiate a connection at roughly
1022 	 *  the same time and the connections cross on the wire.
1023 	 *  In this case we store the incoming one in "othercon"
1024 	 */
1025 	idx = srcu_read_lock(&connections_srcu);
1026 	newcon = nodeid2con(nodeid, 0);
1027 	if (WARN_ON_ONCE(!newcon)) {
1028 		srcu_read_unlock(&connections_srcu, idx);
1029 		result = -ENOENT;
1030 		goto accept_err;
1031 	}
1032 
1033 	sock_set_mark(newsock->sk, mark);
1034 
1035 	down_write(&newcon->sock_lock);
1036 	if (newcon->sock) {
1037 		struct connection *othercon = newcon->othercon;
1038 
1039 		if (!othercon) {
1040 			othercon = kzalloc(sizeof(*othercon), GFP_NOFS);
1041 			if (!othercon) {
1042 				log_print("failed to allocate incoming socket");
1043 				up_write(&newcon->sock_lock);
1044 				srcu_read_unlock(&connections_srcu, idx);
1045 				result = -ENOMEM;
1046 				goto accept_err;
1047 			}
1048 
1049 			dlm_con_init(othercon, nodeid);
1050 			lockdep_set_subclass(&othercon->sock_lock, 1);
1051 			newcon->othercon = othercon;
1052 			set_bit(CF_IS_OTHERCON, &othercon->flags);
1053 		} else {
1054 			/* close other sock con if we have something new */
1055 			close_connection(othercon, false);
1056 		}
1057 
1058 		down_write(&othercon->sock_lock);
1059 		add_sock(newsock, othercon);
1060 
1061 		/* check if we receved something while adding */
1062 		lock_sock(othercon->sock->sk);
1063 		lowcomms_queue_rwork(othercon);
1064 		release_sock(othercon->sock->sk);
1065 		up_write(&othercon->sock_lock);
1066 	}
1067 	else {
1068 		/* accept copies the sk after we've saved the callbacks, so we
1069 		   don't want to save them a second time or comm errors will
1070 		   result in calling sk_error_report recursively. */
1071 		add_sock(newsock, newcon);
1072 
1073 		/* check if we receved something while adding */
1074 		lock_sock(newcon->sock->sk);
1075 		lowcomms_queue_rwork(newcon);
1076 		release_sock(newcon->sock->sk);
1077 	}
1078 	up_write(&newcon->sock_lock);
1079 	srcu_read_unlock(&connections_srcu, idx);
1080 
1081 	return DLM_IO_SUCCESS;
1082 
1083 accept_err:
1084 	if (newsock)
1085 		sock_release(newsock);
1086 
1087 	return result;
1088 }
1089 
1090 /*
1091  * writequeue_entry_complete - try to delete and free write queue entry
1092  * @e: write queue entry to try to delete
1093  * @completed: bytes completed
1094  *
1095  * writequeue_lock must be held.
1096  */
1097 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1098 {
1099 	e->offset += completed;
1100 	e->len -= completed;
1101 	/* signal that page was half way transmitted */
1102 	e->dirty = true;
1103 
1104 	if (e->len == 0 && e->users == 0)
1105 		free_entry(e);
1106 }
1107 
1108 /*
1109  * sctp_bind_addrs - bind a SCTP socket to all our addresses
1110  */
1111 static int sctp_bind_addrs(struct socket *sock, uint16_t port)
1112 {
1113 	struct sockaddr_storage localaddr;
1114 	struct sockaddr *addr = (struct sockaddr *)&localaddr;
1115 	int i, addr_len, result = 0;
1116 
1117 	for (i = 0; i < dlm_local_count; i++) {
1118 		memcpy(&localaddr, &dlm_local_addr[i], sizeof(localaddr));
1119 		make_sockaddr(&localaddr, port, &addr_len);
1120 
1121 		if (!i)
1122 			result = kernel_bind(sock, addr, addr_len);
1123 		else
1124 			result = sock_bind_add(sock->sk, addr, addr_len);
1125 
1126 		if (result < 0) {
1127 			log_print("Can't bind to %d addr number %d, %d.\n",
1128 				  port, i + 1, result);
1129 			break;
1130 		}
1131 	}
1132 	return result;
1133 }
1134 
1135 /* Get local addresses */
1136 static void init_local(void)
1137 {
1138 	struct sockaddr_storage sas;
1139 	int i;
1140 
1141 	dlm_local_count = 0;
1142 	for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1143 		if (dlm_our_addr(&sas, i))
1144 			break;
1145 
1146 		memcpy(&dlm_local_addr[dlm_local_count++], &sas, sizeof(sas));
1147 	}
1148 }
1149 
1150 static struct writequeue_entry *new_writequeue_entry(struct connection *con)
1151 {
1152 	struct writequeue_entry *entry;
1153 
1154 	entry = dlm_allocate_writequeue();
1155 	if (!entry)
1156 		return NULL;
1157 
1158 	entry->page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
1159 	if (!entry->page) {
1160 		dlm_free_writequeue(entry);
1161 		return NULL;
1162 	}
1163 
1164 	entry->offset = 0;
1165 	entry->len = 0;
1166 	entry->end = 0;
1167 	entry->dirty = false;
1168 	entry->con = con;
1169 	entry->users = 1;
1170 	kref_init(&entry->ref);
1171 	return entry;
1172 }
1173 
1174 static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
1175 					     char **ppc, void (*cb)(void *data),
1176 					     void *data)
1177 {
1178 	struct writequeue_entry *e;
1179 
1180 	spin_lock_bh(&con->writequeue_lock);
1181 	if (!list_empty(&con->writequeue)) {
1182 		e = list_last_entry(&con->writequeue, struct writequeue_entry, list);
1183 		if (DLM_WQ_REMAIN_BYTES(e) >= len) {
1184 			kref_get(&e->ref);
1185 
1186 			*ppc = page_address(e->page) + e->end;
1187 			if (cb)
1188 				cb(data);
1189 
1190 			e->end += len;
1191 			e->users++;
1192 			goto out;
1193 		}
1194 	}
1195 
1196 	e = new_writequeue_entry(con);
1197 	if (!e)
1198 		goto out;
1199 
1200 	kref_get(&e->ref);
1201 	*ppc = page_address(e->page);
1202 	e->end += len;
1203 	if (cb)
1204 		cb(data);
1205 
1206 	list_add_tail(&e->list, &con->writequeue);
1207 
1208 out:
1209 	spin_unlock_bh(&con->writequeue_lock);
1210 	return e;
1211 };
1212 
1213 static struct dlm_msg *dlm_lowcomms_new_msg_con(struct connection *con, int len,
1214 						gfp_t allocation, char **ppc,
1215 						void (*cb)(void *data),
1216 						void *data)
1217 {
1218 	struct writequeue_entry *e;
1219 	struct dlm_msg *msg;
1220 
1221 	msg = dlm_allocate_msg(allocation);
1222 	if (!msg)
1223 		return NULL;
1224 
1225 	kref_init(&msg->ref);
1226 
1227 	e = new_wq_entry(con, len, ppc, cb, data);
1228 	if (!e) {
1229 		dlm_free_msg(msg);
1230 		return NULL;
1231 	}
1232 
1233 	msg->retransmit = false;
1234 	msg->orig_msg = NULL;
1235 	msg->ppc = *ppc;
1236 	msg->len = len;
1237 	msg->entry = e;
1238 
1239 	return msg;
1240 }
1241 
1242 /* avoid false positive for nodes_srcu, unlock happens in
1243  * dlm_lowcomms_commit_msg which is a must call if success
1244  */
1245 #ifndef __CHECKER__
1246 struct dlm_msg *dlm_lowcomms_new_msg(int nodeid, int len, gfp_t allocation,
1247 				     char **ppc, void (*cb)(void *data),
1248 				     void *data)
1249 {
1250 	struct connection *con;
1251 	struct dlm_msg *msg;
1252 	int idx;
1253 
1254 	if (len > DLM_MAX_SOCKET_BUFSIZE ||
1255 	    len < sizeof(struct dlm_header)) {
1256 		BUILD_BUG_ON(PAGE_SIZE < DLM_MAX_SOCKET_BUFSIZE);
1257 		log_print("failed to allocate a buffer of size %d", len);
1258 		WARN_ON_ONCE(1);
1259 		return NULL;
1260 	}
1261 
1262 	idx = srcu_read_lock(&connections_srcu);
1263 	con = nodeid2con(nodeid, 0);
1264 	if (WARN_ON_ONCE(!con)) {
1265 		srcu_read_unlock(&connections_srcu, idx);
1266 		return NULL;
1267 	}
1268 
1269 	msg = dlm_lowcomms_new_msg_con(con, len, allocation, ppc, cb, data);
1270 	if (!msg) {
1271 		srcu_read_unlock(&connections_srcu, idx);
1272 		return NULL;
1273 	}
1274 
1275 	/* for dlm_lowcomms_commit_msg() */
1276 	kref_get(&msg->ref);
1277 	/* we assume if successful commit must called */
1278 	msg->idx = idx;
1279 	return msg;
1280 }
1281 #endif
1282 
1283 static void _dlm_lowcomms_commit_msg(struct dlm_msg *msg)
1284 {
1285 	struct writequeue_entry *e = msg->entry;
1286 	struct connection *con = e->con;
1287 	int users;
1288 
1289 	spin_lock_bh(&con->writequeue_lock);
1290 	kref_get(&msg->ref);
1291 	list_add(&msg->list, &e->msgs);
1292 
1293 	users = --e->users;
1294 	if (users)
1295 		goto out;
1296 
1297 	e->len = DLM_WQ_LENGTH_BYTES(e);
1298 
1299 	lowcomms_queue_swork(con);
1300 
1301 out:
1302 	spin_unlock_bh(&con->writequeue_lock);
1303 	return;
1304 }
1305 
1306 /* avoid false positive for nodes_srcu, lock was happen in
1307  * dlm_lowcomms_new_msg
1308  */
1309 #ifndef __CHECKER__
1310 void dlm_lowcomms_commit_msg(struct dlm_msg *msg)
1311 {
1312 	_dlm_lowcomms_commit_msg(msg);
1313 	srcu_read_unlock(&connections_srcu, msg->idx);
1314 	/* because dlm_lowcomms_new_msg() */
1315 	kref_put(&msg->ref, dlm_msg_release);
1316 }
1317 #endif
1318 
1319 void dlm_lowcomms_put_msg(struct dlm_msg *msg)
1320 {
1321 	kref_put(&msg->ref, dlm_msg_release);
1322 }
1323 
1324 /* does not held connections_srcu, usage lowcomms_error_report only */
1325 int dlm_lowcomms_resend_msg(struct dlm_msg *msg)
1326 {
1327 	struct dlm_msg *msg_resend;
1328 	char *ppc;
1329 
1330 	if (msg->retransmit)
1331 		return 1;
1332 
1333 	msg_resend = dlm_lowcomms_new_msg_con(msg->entry->con, msg->len,
1334 					      GFP_ATOMIC, &ppc, NULL, NULL);
1335 	if (!msg_resend)
1336 		return -ENOMEM;
1337 
1338 	msg->retransmit = true;
1339 	kref_get(&msg->ref);
1340 	msg_resend->orig_msg = msg;
1341 
1342 	memcpy(ppc, msg->ppc, msg->len);
1343 	_dlm_lowcomms_commit_msg(msg_resend);
1344 	dlm_lowcomms_put_msg(msg_resend);
1345 
1346 	return 0;
1347 }
1348 
1349 /* Send a message */
1350 static int send_to_sock(struct connection *con)
1351 {
1352 	const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1353 	struct writequeue_entry *e;
1354 	int len, offset, ret;
1355 
1356 	spin_lock_bh(&con->writequeue_lock);
1357 	e = con_next_wq(con);
1358 	if (!e) {
1359 		clear_bit(CF_SEND_PENDING, &con->flags);
1360 		spin_unlock_bh(&con->writequeue_lock);
1361 		return DLM_IO_END;
1362 	}
1363 
1364 	len = e->len;
1365 	offset = e->offset;
1366 	WARN_ON_ONCE(len == 0 && e->users == 0);
1367 	spin_unlock_bh(&con->writequeue_lock);
1368 
1369 	ret = kernel_sendpage(con->sock, e->page, offset, len,
1370 			      msg_flags);
1371 	trace_dlm_send(con->nodeid, ret);
1372 	if (ret == -EAGAIN || ret == 0) {
1373 		lock_sock(con->sock->sk);
1374 		spin_lock_bh(&con->writequeue_lock);
1375 		if (test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1376 		    !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1377 			/* Notify TCP that we're limited by the
1378 			 * application window size.
1379 			 */
1380 			set_bit(SOCK_NOSPACE, &con->sock->sk->sk_socket->flags);
1381 			con->sock->sk->sk_write_pending++;
1382 
1383 			clear_bit(CF_SEND_PENDING, &con->flags);
1384 			spin_unlock_bh(&con->writequeue_lock);
1385 			release_sock(con->sock->sk);
1386 
1387 			/* wait for write_space() event */
1388 			return DLM_IO_END;
1389 		}
1390 		spin_unlock_bh(&con->writequeue_lock);
1391 		release_sock(con->sock->sk);
1392 
1393 		return DLM_IO_RESCHED;
1394 	} else if (ret < 0) {
1395 		return ret;
1396 	}
1397 
1398 	spin_lock_bh(&con->writequeue_lock);
1399 	writequeue_entry_complete(e, ret);
1400 	spin_unlock_bh(&con->writequeue_lock);
1401 
1402 	return DLM_IO_SUCCESS;
1403 }
1404 
1405 static void clean_one_writequeue(struct connection *con)
1406 {
1407 	struct writequeue_entry *e, *safe;
1408 
1409 	spin_lock_bh(&con->writequeue_lock);
1410 	list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1411 		free_entry(e);
1412 	}
1413 	spin_unlock_bh(&con->writequeue_lock);
1414 }
1415 
1416 static void connection_release(struct rcu_head *rcu)
1417 {
1418 	struct connection *con = container_of(rcu, struct connection, rcu);
1419 
1420 	WARN_ON_ONCE(!list_empty(&con->writequeue));
1421 	WARN_ON_ONCE(con->sock);
1422 	kfree(con);
1423 }
1424 
1425 /* Called from recovery when it knows that a node has
1426    left the cluster */
1427 int dlm_lowcomms_close(int nodeid)
1428 {
1429 	struct connection *con;
1430 	int idx;
1431 
1432 	log_print("closing connection to node %d", nodeid);
1433 
1434 	idx = srcu_read_lock(&connections_srcu);
1435 	con = nodeid2con(nodeid, 0);
1436 	if (WARN_ON_ONCE(!con)) {
1437 		srcu_read_unlock(&connections_srcu, idx);
1438 		return -ENOENT;
1439 	}
1440 
1441 	stop_connection_io(con);
1442 	log_print("io handling for node: %d stopped", nodeid);
1443 	close_connection(con, true);
1444 
1445 	spin_lock(&connections_lock);
1446 	hlist_del_rcu(&con->list);
1447 	spin_unlock(&connections_lock);
1448 
1449 	clean_one_writequeue(con);
1450 	call_srcu(&connections_srcu, &con->rcu, connection_release);
1451 	if (con->othercon) {
1452 		clean_one_writequeue(con->othercon);
1453 		if (con->othercon)
1454 			call_srcu(&connections_srcu, &con->othercon->rcu, connection_release);
1455 	}
1456 	srcu_read_unlock(&connections_srcu, idx);
1457 
1458 	/* for debugging we print when we are done to compare with other
1459 	 * messages in between. This function need to be correctly synchronized
1460 	 * with io handling
1461 	 */
1462 	log_print("closing connection to node %d done", nodeid);
1463 
1464 	return 0;
1465 }
1466 
1467 /* Receive worker function */
1468 static void process_recv_sockets(struct work_struct *work)
1469 {
1470 	struct connection *con = container_of(work, struct connection, rwork);
1471 	int ret, buflen;
1472 
1473 	down_read(&con->sock_lock);
1474 	if (!con->sock) {
1475 		up_read(&con->sock_lock);
1476 		return;
1477 	}
1478 
1479 	buflen = READ_ONCE(dlm_config.ci_buffer_size);
1480 	do {
1481 		ret = receive_from_sock(con, buflen);
1482 	} while (ret == DLM_IO_SUCCESS);
1483 	up_read(&con->sock_lock);
1484 
1485 	switch (ret) {
1486 	case DLM_IO_END:
1487 		/* CF_RECV_PENDING cleared */
1488 		break;
1489 	case DLM_IO_EOF:
1490 		close_connection(con, false);
1491 		/* CF_RECV_PENDING cleared */
1492 		break;
1493 	case DLM_IO_RESCHED:
1494 		cond_resched();
1495 		queue_work(io_workqueue, &con->rwork);
1496 		/* CF_RECV_PENDING not cleared */
1497 		break;
1498 	default:
1499 		if (ret < 0) {
1500 			if (test_bit(CF_IS_OTHERCON, &con->flags)) {
1501 				close_connection(con, false);
1502 			} else {
1503 				spin_lock_bh(&con->writequeue_lock);
1504 				lowcomms_queue_swork(con);
1505 				spin_unlock_bh(&con->writequeue_lock);
1506 			}
1507 
1508 			/* CF_RECV_PENDING cleared for othercon
1509 			 * we trigger send queue if not already done
1510 			 * and process_send_sockets will handle it
1511 			 */
1512 			break;
1513 		}
1514 
1515 		WARN_ON_ONCE(1);
1516 		break;
1517 	}
1518 }
1519 
1520 static void process_listen_recv_socket(struct work_struct *work)
1521 {
1522 	int ret;
1523 
1524 	if (WARN_ON_ONCE(!listen_con.sock))
1525 		return;
1526 
1527 	do {
1528 		ret = accept_from_sock();
1529 	} while (ret == DLM_IO_SUCCESS);
1530 
1531 	if (ret < 0)
1532 		log_print("critical error accepting connection: %d", ret);
1533 }
1534 
1535 static int dlm_connect(struct connection *con)
1536 {
1537 	struct sockaddr_storage addr;
1538 	int result, addr_len;
1539 	struct socket *sock;
1540 	unsigned int mark;
1541 
1542 	memset(&addr, 0, sizeof(addr));
1543 	result = nodeid_to_addr(con->nodeid, &addr, NULL,
1544 				dlm_proto_ops->try_new_addr, &mark);
1545 	if (result < 0) {
1546 		log_print("no address for nodeid %d", con->nodeid);
1547 		return result;
1548 	}
1549 
1550 	/* Create a socket to communicate with */
1551 	result = sock_create_kern(&init_net, dlm_local_addr[0].ss_family,
1552 				  SOCK_STREAM, dlm_proto_ops->proto, &sock);
1553 	if (result < 0)
1554 		return result;
1555 
1556 	sock_set_mark(sock->sk, mark);
1557 	dlm_proto_ops->sockopts(sock);
1558 
1559 	result = dlm_proto_ops->bind(sock);
1560 	if (result < 0) {
1561 		sock_release(sock);
1562 		return result;
1563 	}
1564 
1565 	add_sock(sock, con);
1566 
1567 	log_print_ratelimited("connecting to %d", con->nodeid);
1568 	make_sockaddr(&addr, dlm_config.ci_tcp_port, &addr_len);
1569 	result = dlm_proto_ops->connect(con, sock, (struct sockaddr *)&addr,
1570 					addr_len);
1571 	switch (result) {
1572 	case -EINPROGRESS:
1573 		/* not an error */
1574 		fallthrough;
1575 	case 0:
1576 		break;
1577 	default:
1578 		if (result < 0)
1579 			dlm_close_sock(&con->sock);
1580 
1581 		break;
1582 	}
1583 
1584 	return result;
1585 }
1586 
1587 /* Send worker function */
1588 static void process_send_sockets(struct work_struct *work)
1589 {
1590 	struct connection *con = container_of(work, struct connection, swork);
1591 	int ret;
1592 
1593 	WARN_ON_ONCE(test_bit(CF_IS_OTHERCON, &con->flags));
1594 
1595 	down_read(&con->sock_lock);
1596 	if (!con->sock) {
1597 		up_read(&con->sock_lock);
1598 		down_write(&con->sock_lock);
1599 		if (!con->sock) {
1600 			ret = dlm_connect(con);
1601 			switch (ret) {
1602 			case 0:
1603 				break;
1604 			case -EINPROGRESS:
1605 				/* avoid spamming resched on connection
1606 				 * we might can switch to a state_change
1607 				 * event based mechanism if established
1608 				 */
1609 				msleep(100);
1610 				break;
1611 			default:
1612 				/* CF_SEND_PENDING not cleared */
1613 				up_write(&con->sock_lock);
1614 				log_print("connect to node %d try %d error %d",
1615 					  con->nodeid, con->retries++, ret);
1616 				msleep(1000);
1617 				/* For now we try forever to reconnect. In
1618 				 * future we should send a event to cluster
1619 				 * manager to fence itself after certain amount
1620 				 * of retries.
1621 				 */
1622 				queue_work(io_workqueue, &con->swork);
1623 				return;
1624 			}
1625 		}
1626 		downgrade_write(&con->sock_lock);
1627 	}
1628 
1629 	do {
1630 		ret = send_to_sock(con);
1631 	} while (ret == DLM_IO_SUCCESS);
1632 	up_read(&con->sock_lock);
1633 
1634 	switch (ret) {
1635 	case DLM_IO_END:
1636 		/* CF_SEND_PENDING cleared */
1637 		break;
1638 	case DLM_IO_RESCHED:
1639 		/* CF_SEND_PENDING not cleared */
1640 		cond_resched();
1641 		queue_work(io_workqueue, &con->swork);
1642 		break;
1643 	default:
1644 		if (ret < 0) {
1645 			close_connection(con, false);
1646 
1647 			/* CF_SEND_PENDING cleared */
1648 			spin_lock_bh(&con->writequeue_lock);
1649 			lowcomms_queue_swork(con);
1650 			spin_unlock_bh(&con->writequeue_lock);
1651 			break;
1652 		}
1653 
1654 		WARN_ON_ONCE(1);
1655 		break;
1656 	}
1657 }
1658 
1659 static void work_stop(void)
1660 {
1661 	if (io_workqueue) {
1662 		destroy_workqueue(io_workqueue);
1663 		io_workqueue = NULL;
1664 	}
1665 
1666 	if (process_workqueue) {
1667 		destroy_workqueue(process_workqueue);
1668 		process_workqueue = NULL;
1669 	}
1670 }
1671 
1672 static int work_start(void)
1673 {
1674 	io_workqueue = alloc_workqueue("dlm_io", WQ_HIGHPRI | WQ_MEM_RECLAIM,
1675 				       0);
1676 	if (!io_workqueue) {
1677 		log_print("can't start dlm_io");
1678 		return -ENOMEM;
1679 	}
1680 
1681 	/* ordered dlm message process queue,
1682 	 * should be converted to a tasklet
1683 	 */
1684 	process_workqueue = alloc_ordered_workqueue("dlm_process",
1685 						    WQ_HIGHPRI | WQ_MEM_RECLAIM);
1686 	if (!process_workqueue) {
1687 		log_print("can't start dlm_process");
1688 		destroy_workqueue(io_workqueue);
1689 		io_workqueue = NULL;
1690 		return -ENOMEM;
1691 	}
1692 
1693 	return 0;
1694 }
1695 
1696 void dlm_lowcomms_shutdown(void)
1697 {
1698 	/* stop lowcomms_listen_data_ready calls */
1699 	lock_sock(listen_con.sock->sk);
1700 	listen_con.sock->sk->sk_data_ready = listen_sock.sk_data_ready;
1701 	release_sock(listen_con.sock->sk);
1702 
1703 	cancel_work_sync(&listen_con.rwork);
1704 	dlm_close_sock(&listen_con.sock);
1705 
1706 	flush_workqueue(process_workqueue);
1707 }
1708 
1709 void dlm_lowcomms_shutdown_node(int nodeid, bool force)
1710 {
1711 	struct connection *con;
1712 	int idx;
1713 
1714 	idx = srcu_read_lock(&connections_srcu);
1715 	con = nodeid2con(nodeid, 0);
1716 	if (WARN_ON_ONCE(!con)) {
1717 		srcu_read_unlock(&connections_srcu, idx);
1718 		return;
1719 	}
1720 
1721 	flush_work(&con->swork);
1722 	stop_connection_io(con);
1723 	WARN_ON_ONCE(!force && !list_empty(&con->writequeue));
1724 	close_connection(con, true);
1725 	clean_one_writequeue(con);
1726 	if (con->othercon)
1727 		clean_one_writequeue(con->othercon);
1728 	allow_connection_io(con);
1729 	srcu_read_unlock(&connections_srcu, idx);
1730 }
1731 
1732 void dlm_lowcomms_stop(void)
1733 {
1734 	work_stop();
1735 	dlm_proto_ops = NULL;
1736 }
1737 
1738 static int dlm_listen_for_all(void)
1739 {
1740 	struct socket *sock;
1741 	int result;
1742 
1743 	log_print("Using %s for communications",
1744 		  dlm_proto_ops->name);
1745 
1746 	result = dlm_proto_ops->listen_validate();
1747 	if (result < 0)
1748 		return result;
1749 
1750 	result = sock_create_kern(&init_net, dlm_local_addr[0].ss_family,
1751 				  SOCK_STREAM, dlm_proto_ops->proto, &sock);
1752 	if (result < 0) {
1753 		log_print("Can't create comms socket: %d", result);
1754 		return result;
1755 	}
1756 
1757 	sock_set_mark(sock->sk, dlm_config.ci_mark);
1758 	dlm_proto_ops->listen_sockopts(sock);
1759 
1760 	result = dlm_proto_ops->listen_bind(sock);
1761 	if (result < 0)
1762 		goto out;
1763 
1764 	lock_sock(sock->sk);
1765 	listen_sock.sk_data_ready = sock->sk->sk_data_ready;
1766 	listen_sock.sk_write_space = sock->sk->sk_write_space;
1767 	listen_sock.sk_error_report = sock->sk->sk_error_report;
1768 	listen_sock.sk_state_change = sock->sk->sk_state_change;
1769 
1770 	listen_con.sock = sock;
1771 
1772 	sock->sk->sk_allocation = GFP_NOFS;
1773 	sock->sk->sk_use_task_frag = false;
1774 	sock->sk->sk_data_ready = lowcomms_listen_data_ready;
1775 	release_sock(sock->sk);
1776 
1777 	result = sock->ops->listen(sock, 5);
1778 	if (result < 0) {
1779 		dlm_close_sock(&listen_con.sock);
1780 		return result;
1781 	}
1782 
1783 	return 0;
1784 
1785 out:
1786 	sock_release(sock);
1787 	return result;
1788 }
1789 
1790 static int dlm_tcp_bind(struct socket *sock)
1791 {
1792 	struct sockaddr_storage src_addr;
1793 	int result, addr_len;
1794 
1795 	/* Bind to our cluster-known address connecting to avoid
1796 	 * routing problems.
1797 	 */
1798 	memcpy(&src_addr, &dlm_local_addr[0], sizeof(src_addr));
1799 	make_sockaddr(&src_addr, 0, &addr_len);
1800 
1801 	result = sock->ops->bind(sock, (struct sockaddr *)&src_addr,
1802 				 addr_len);
1803 	if (result < 0) {
1804 		/* This *may* not indicate a critical error */
1805 		log_print("could not bind for connect: %d", result);
1806 	}
1807 
1808 	return 0;
1809 }
1810 
1811 static int dlm_tcp_connect(struct connection *con, struct socket *sock,
1812 			   struct sockaddr *addr, int addr_len)
1813 {
1814 	return sock->ops->connect(sock, addr, addr_len, O_NONBLOCK);
1815 }
1816 
1817 static int dlm_tcp_listen_validate(void)
1818 {
1819 	/* We don't support multi-homed hosts */
1820 	if (dlm_local_count > 1) {
1821 		log_print("TCP protocol can't handle multi-homed hosts, try SCTP");
1822 		return -EINVAL;
1823 	}
1824 
1825 	return 0;
1826 }
1827 
1828 static void dlm_tcp_sockopts(struct socket *sock)
1829 {
1830 	/* Turn off Nagle's algorithm */
1831 	tcp_sock_set_nodelay(sock->sk);
1832 }
1833 
1834 static void dlm_tcp_listen_sockopts(struct socket *sock)
1835 {
1836 	dlm_tcp_sockopts(sock);
1837 	sock_set_reuseaddr(sock->sk);
1838 }
1839 
1840 static int dlm_tcp_listen_bind(struct socket *sock)
1841 {
1842 	int addr_len;
1843 
1844 	/* Bind to our port */
1845 	make_sockaddr(&dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len);
1846 	return sock->ops->bind(sock, (struct sockaddr *)&dlm_local_addr[0],
1847 			       addr_len);
1848 }
1849 
1850 static const struct dlm_proto_ops dlm_tcp_ops = {
1851 	.name = "TCP",
1852 	.proto = IPPROTO_TCP,
1853 	.connect = dlm_tcp_connect,
1854 	.sockopts = dlm_tcp_sockopts,
1855 	.bind = dlm_tcp_bind,
1856 	.listen_validate = dlm_tcp_listen_validate,
1857 	.listen_sockopts = dlm_tcp_listen_sockopts,
1858 	.listen_bind = dlm_tcp_listen_bind,
1859 };
1860 
1861 static int dlm_sctp_bind(struct socket *sock)
1862 {
1863 	return sctp_bind_addrs(sock, 0);
1864 }
1865 
1866 static int dlm_sctp_connect(struct connection *con, struct socket *sock,
1867 			    struct sockaddr *addr, int addr_len)
1868 {
1869 	int ret;
1870 
1871 	/*
1872 	 * Make sock->ops->connect() function return in specified time,
1873 	 * since O_NONBLOCK argument in connect() function does not work here,
1874 	 * then, we should restore the default value of this attribute.
1875 	 */
1876 	sock_set_sndtimeo(sock->sk, 5);
1877 	ret = sock->ops->connect(sock, addr, addr_len, 0);
1878 	sock_set_sndtimeo(sock->sk, 0);
1879 	return ret;
1880 }
1881 
1882 static int dlm_sctp_listen_validate(void)
1883 {
1884 	if (!IS_ENABLED(CONFIG_IP_SCTP)) {
1885 		log_print("SCTP is not enabled by this kernel");
1886 		return -EOPNOTSUPP;
1887 	}
1888 
1889 	request_module("sctp");
1890 	return 0;
1891 }
1892 
1893 static int dlm_sctp_bind_listen(struct socket *sock)
1894 {
1895 	return sctp_bind_addrs(sock, dlm_config.ci_tcp_port);
1896 }
1897 
1898 static void dlm_sctp_sockopts(struct socket *sock)
1899 {
1900 	/* Turn off Nagle's algorithm */
1901 	sctp_sock_set_nodelay(sock->sk);
1902 	sock_set_rcvbuf(sock->sk, NEEDED_RMEM);
1903 }
1904 
1905 static const struct dlm_proto_ops dlm_sctp_ops = {
1906 	.name = "SCTP",
1907 	.proto = IPPROTO_SCTP,
1908 	.try_new_addr = true,
1909 	.connect = dlm_sctp_connect,
1910 	.sockopts = dlm_sctp_sockopts,
1911 	.bind = dlm_sctp_bind,
1912 	.listen_validate = dlm_sctp_listen_validate,
1913 	.listen_sockopts = dlm_sctp_sockopts,
1914 	.listen_bind = dlm_sctp_bind_listen,
1915 };
1916 
1917 int dlm_lowcomms_start(void)
1918 {
1919 	int error;
1920 
1921 	init_local();
1922 	if (!dlm_local_count) {
1923 		error = -ENOTCONN;
1924 		log_print("no local IP address has been set");
1925 		goto fail;
1926 	}
1927 
1928 	error = work_start();
1929 	if (error)
1930 		goto fail;
1931 
1932 	/* Start listening */
1933 	switch (dlm_config.ci_protocol) {
1934 	case DLM_PROTO_TCP:
1935 		dlm_proto_ops = &dlm_tcp_ops;
1936 		break;
1937 	case DLM_PROTO_SCTP:
1938 		dlm_proto_ops = &dlm_sctp_ops;
1939 		break;
1940 	default:
1941 		log_print("Invalid protocol identifier %d set",
1942 			  dlm_config.ci_protocol);
1943 		error = -EINVAL;
1944 		goto fail_proto_ops;
1945 	}
1946 
1947 	error = dlm_listen_for_all();
1948 	if (error)
1949 		goto fail_listen;
1950 
1951 	return 0;
1952 
1953 fail_listen:
1954 	dlm_proto_ops = NULL;
1955 fail_proto_ops:
1956 	work_stop();
1957 fail:
1958 	return error;
1959 }
1960 
1961 void dlm_lowcomms_init(void)
1962 {
1963 	int i;
1964 
1965 	for (i = 0; i < CONN_HASH_SIZE; i++)
1966 		INIT_HLIST_HEAD(&connection_hash[i]);
1967 
1968 	INIT_WORK(&listen_con.rwork, process_listen_recv_socket);
1969 }
1970 
1971 void dlm_lowcomms_exit(void)
1972 {
1973 	struct connection *con;
1974 	int i, idx;
1975 
1976 	idx = srcu_read_lock(&connections_srcu);
1977 	for (i = 0; i < CONN_HASH_SIZE; i++) {
1978 		hlist_for_each_entry_rcu(con, &connection_hash[i], list) {
1979 			spin_lock(&connections_lock);
1980 			hlist_del_rcu(&con->list);
1981 			spin_unlock(&connections_lock);
1982 
1983 			if (con->othercon)
1984 				call_srcu(&connections_srcu, &con->othercon->rcu,
1985 					  connection_release);
1986 			call_srcu(&connections_srcu, &con->rcu, connection_release);
1987 		}
1988 	}
1989 	srcu_read_unlock(&connections_srcu, idx);
1990 }
1991