xref: /linux/fs/dlm/midcomms.c (revision 2363088eba2ecccfb643725e4864af73c4226a04)
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) 2004-2021 Red Hat, Inc.  All rights reserved.
7 **
8 **
9 *******************************************************************************
10 ******************************************************************************/
11 
12 /*
13  * midcomms.c
14  *
15  * This is the appallingly named "mid-level" comms layer. It takes care about
16  * deliver an on application layer "reliable" communication above the used
17  * lowcomms transport layer.
18  *
19  * How it works:
20  *
21  * Each nodes keeps track of all send DLM messages in send_queue with a sequence
22  * number. The receive will send an DLM_ACK message back for every DLM message
23  * received at the other side. If a reconnect happens in lowcomms we will send
24  * all unacknowledged dlm messages again. The receiving side might drop any already
25  * received message by comparing sequence numbers.
26  *
27  * How version detection works:
28  *
29  * Due the fact that dlm has pre-configured node addresses on every side
30  * it is in it's nature that every side connects at starts to transmit
31  * dlm messages which ends in a race. However DLM_RCOM_NAMES, DLM_RCOM_STATUS
32  * and their replies are the first messages which are exchanges. Due backwards
33  * compatibility these messages are not covered by the midcomms re-transmission
34  * layer. These messages have their own re-transmission handling in the dlm
35  * application layer. The version field of every node will be set on these RCOM
36  * messages as soon as they arrived and the node isn't yet part of the nodes
37  * hash. There exists also logic to detect version mismatched if something weird
38  * going on or the first messages isn't an expected one.
39  *
40  * Termination:
41  *
42  * The midcomms layer does a 4 way handshake for termination on DLM protocol
43  * like TCP supports it with half-closed socket support. SCTP doesn't support
44  * half-closed socket, so we do it on DLM layer. Also socket shutdown() can be
45  * interrupted by .e.g. tcp reset itself. Additional there exists the othercon
46  * paradigm in lowcomms which cannot be easily without breaking backwards
47  * compatibility. A node cannot send anything to another node when a DLM_FIN
48  * message was send. There exists additional logic to print a warning if
49  * DLM wants to do it. There exists a state handling like RFC 793 but reduced
50  * to termination only. The event "member removal event" describes the cluster
51  * manager removed the node from internal lists, at this point DLM does not
52  * send any message to the other node. There exists two cases:
53  *
54  * 1. The cluster member was removed and we received a FIN
55  * OR
56  * 2. We received a FIN but the member was not removed yet
57  *
58  * One of these cases will do the CLOSE_WAIT to LAST_ACK change.
59  *
60  *
61  *                              +---------+
62  *                              | CLOSED  |
63  *                              +---------+
64  *                                   | add member/receive RCOM version
65  *                                   |            detection msg
66  *                                   V
67  *                              +---------+
68  *                              |  ESTAB  |
69  *                              +---------+
70  *                       CLOSE    |     |    rcv FIN
71  *                      -------   |     |    -------
72  * +---------+          snd FIN  /       \   snd ACK          +---------+
73  * |  FIN    |<-----------------           ------------------>|  CLOSE  |
74  * | WAIT-1  |------------------                              |   WAIT  |
75  * +---------+          rcv FIN  \                            +---------+
76  * | rcv ACK of FIN   -------   |                            CLOSE  | member
77  * | --------------   snd ACK   |                           ------- | removal
78  * V        x                   V                           snd FIN V event
79  * +---------+                  +---------+                   +---------+
80  * |FINWAIT-2|                  | CLOSING |                   | LAST-ACK|
81  * +---------+                  +---------+                   +---------+
82  * |                rcv ACK of FIN |                 rcv ACK of FIN |
83  * |  rcv FIN       -------------- |                 -------------- |
84  * |  -------              x       V                        x       V
85  *  \ snd ACK                 +---------+                   +---------+
86  *   ------------------------>| CLOSED  |                   | CLOSED  |
87  *                            +---------+                   +---------+
88  *
89  * NOTE: any state can interrupted by midcomms_close() and state will be
90  * switched to CLOSED in case of fencing. There exists also some timeout
91  * handling when we receive the version detection RCOM messages which is
92  * made by observation.
93  *
94  * Future improvements:
95  *
96  * There exists some known issues/improvements of the dlm handling. Some
97  * of them should be done in a next major dlm version bump which makes
98  * it incompatible with previous versions.
99  *
100  * Unaligned memory access:
101  *
102  * There exists cases when the dlm message buffer length is not aligned
103  * to 8 byte. However seems nobody detected any problem with it. This
104  * can be fixed in the next major version bump of dlm.
105  *
106  * Version detection:
107  *
108  * The version detection and how it's done is related to backwards
109  * compatibility. There exists better ways to make a better handling.
110  * However this should be changed in the next major version bump of dlm.
111  *
112  * Tail Size checking:
113  *
114  * There exists a message tail payload in e.g. DLM_MSG however we don't
115  * check it against the message length yet regarding to the receive buffer
116  * length. That need to be validated.
117  *
118  * Fencing bad nodes:
119  *
120  * At timeout places or weird sequence number behaviours we should send
121  * a fencing request to the cluster manager.
122  */
123 
124 /* Debug switch to enable a 5 seconds sleep waiting of a termination.
125  * This can be useful to test fencing while termination is running.
126  * This requires a setup with only gfs2 as dlm user, so that the
127  * last umount will terminate the connection.
128  *
129  * However it became useful to test, while the 5 seconds block in umount
130  * just press the reset button. In a lot of dropping the termination
131  * process can could take several seconds.
132  */
133 #define DLM_DEBUG_FENCE_TERMINATION	0
134 
135 #include <trace/events/dlm.h>
136 #include <net/tcp.h>
137 
138 #include "dlm_internal.h"
139 #include "lowcomms.h"
140 #include "config.h"
141 #include "memory.h"
142 #include "lock.h"
143 #include "util.h"
144 #include "midcomms.h"
145 
146 /* init value for sequence numbers for testing purpose only e.g. overflows */
147 #define DLM_SEQ_INIT		0
148 /* 5 seconds wait to sync ending of dlm */
149 #define DLM_SHUTDOWN_TIMEOUT	msecs_to_jiffies(5000)
150 #define DLM_VERSION_NOT_SET	0
151 #define DLM_SEND_ACK_BACK_MSG_THRESHOLD 32
152 #define DLM_RECV_ACK_BACK_MSG_THRESHOLD (DLM_SEND_ACK_BACK_MSG_THRESHOLD * 8)
153 
154 struct midcomms_node {
155 	int nodeid;
156 	uint32_t version;
157 	atomic_t seq_send;
158 	atomic_t seq_next;
159 	/* These queues are unbound because we cannot drop any message in dlm.
160 	 * We could send a fence signal for a specific node to the cluster
161 	 * manager if queues hits some maximum value, however this handling
162 	 * not supported yet.
163 	 */
164 	struct list_head send_queue;
165 	spinlock_t send_queue_lock;
166 	atomic_t send_queue_cnt;
167 #define DLM_NODE_FLAG_CLOSE	1
168 #define DLM_NODE_FLAG_STOP_TX	2
169 #define DLM_NODE_FLAG_STOP_RX	3
170 	atomic_t ulp_delivered;
171 	unsigned long flags;
172 	wait_queue_head_t shutdown_wait;
173 
174 	/* dlm tcp termination state */
175 #define DLM_CLOSED	1
176 #define DLM_ESTABLISHED	2
177 #define DLM_FIN_WAIT1	3
178 #define DLM_FIN_WAIT2	4
179 #define DLM_CLOSE_WAIT	5
180 #define DLM_LAST_ACK	6
181 #define DLM_CLOSING	7
182 	int state;
183 	spinlock_t state_lock;
184 
185 	/* counts how many lockspaces are using this node
186 	 * this refcount is necessary to determine if the
187 	 * node wants to disconnect.
188 	 */
189 	int users;
190 
191 	/* not protected by srcu, node_hash lifetime */
192 	void *debugfs;
193 
194 	struct hlist_node hlist;
195 	struct rcu_head rcu;
196 };
197 
198 struct dlm_mhandle {
199 	const union dlm_packet *inner_p;
200 	struct midcomms_node *node;
201 	struct dlm_opts *opts;
202 	struct dlm_msg *msg;
203 	bool committed;
204 	uint32_t seq;
205 
206 	void (*ack_rcv)(struct midcomms_node *node);
207 
208 	/* get_mhandle/commit srcu idx exchange */
209 	int idx;
210 
211 	struct list_head list;
212 	struct rcu_head rcu;
213 };
214 
215 static struct hlist_head node_hash[CONN_HASH_SIZE];
216 static DEFINE_SPINLOCK(nodes_lock);
217 DEFINE_STATIC_SRCU(nodes_srcu);
218 
219 /* This mutex prevents that midcomms_close() is running while
220  * stop() or remove(). As I experienced invalid memory access
221  * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
222  * resetting machines. I will end in some double deletion in nodes
223  * datastructure.
224  */
225 static DEFINE_MUTEX(close_lock);
226 
227 struct kmem_cache *dlm_midcomms_cache_create(void)
228 {
229 	return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
230 				 0, 0, NULL);
231 }
232 
233 static inline const char *dlm_state_str(int state)
234 {
235 	switch (state) {
236 	case DLM_CLOSED:
237 		return "CLOSED";
238 	case DLM_ESTABLISHED:
239 		return "ESTABLISHED";
240 	case DLM_FIN_WAIT1:
241 		return "FIN_WAIT1";
242 	case DLM_FIN_WAIT2:
243 		return "FIN_WAIT2";
244 	case DLM_CLOSE_WAIT:
245 		return "CLOSE_WAIT";
246 	case DLM_LAST_ACK:
247 		return "LAST_ACK";
248 	case DLM_CLOSING:
249 		return "CLOSING";
250 	default:
251 		return "UNKNOWN";
252 	}
253 }
254 
255 const char *dlm_midcomms_state(struct midcomms_node *node)
256 {
257 	return dlm_state_str(node->state);
258 }
259 
260 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
261 {
262 	return node->flags;
263 }
264 
265 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
266 {
267 	return atomic_read(&node->send_queue_cnt);
268 }
269 
270 uint32_t dlm_midcomms_version(struct midcomms_node *node)
271 {
272 	return node->version;
273 }
274 
275 static struct midcomms_node *__find_node(int nodeid, int r)
276 {
277 	struct midcomms_node *node;
278 
279 	hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
280 		if (node->nodeid == nodeid)
281 			return node;
282 	}
283 
284 	return NULL;
285 }
286 
287 static void dlm_mhandle_release(struct rcu_head *rcu)
288 {
289 	struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
290 
291 	dlm_lowcomms_put_msg(mh->msg);
292 	dlm_free_mhandle(mh);
293 }
294 
295 static void dlm_mhandle_delete(struct midcomms_node *node,
296 			       struct dlm_mhandle *mh)
297 {
298 	list_del_rcu(&mh->list);
299 	atomic_dec(&node->send_queue_cnt);
300 	call_rcu(&mh->rcu, dlm_mhandle_release);
301 }
302 
303 static void dlm_send_queue_flush(struct midcomms_node *node)
304 {
305 	struct dlm_mhandle *mh;
306 
307 	pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
308 
309 	rcu_read_lock();
310 	spin_lock_bh(&node->send_queue_lock);
311 	list_for_each_entry_rcu(mh, &node->send_queue, list) {
312 		dlm_mhandle_delete(node, mh);
313 	}
314 	spin_unlock_bh(&node->send_queue_lock);
315 	rcu_read_unlock();
316 }
317 
318 static void midcomms_node_reset(struct midcomms_node *node)
319 {
320 	pr_debug("reset node %d\n", node->nodeid);
321 
322 	atomic_set(&node->seq_next, DLM_SEQ_INIT);
323 	atomic_set(&node->seq_send, DLM_SEQ_INIT);
324 	atomic_set(&node->ulp_delivered, 0);
325 	node->version = DLM_VERSION_NOT_SET;
326 	node->flags = 0;
327 
328 	dlm_send_queue_flush(node);
329 	node->state = DLM_CLOSED;
330 	wake_up(&node->shutdown_wait);
331 }
332 
333 static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
334 {
335 	struct midcomms_node *node, *tmp;
336 	int r = nodeid_hash(nodeid);
337 
338 	node = __find_node(nodeid, r);
339 	if (node || !alloc)
340 		return node;
341 
342 	node = kmalloc(sizeof(*node), alloc);
343 	if (!node)
344 		return NULL;
345 
346 	node->nodeid = nodeid;
347 	spin_lock_init(&node->state_lock);
348 	spin_lock_init(&node->send_queue_lock);
349 	atomic_set(&node->send_queue_cnt, 0);
350 	INIT_LIST_HEAD(&node->send_queue);
351 	init_waitqueue_head(&node->shutdown_wait);
352 	node->users = 0;
353 	midcomms_node_reset(node);
354 
355 	spin_lock(&nodes_lock);
356 	/* check again if there was somebody else
357 	 * earlier here to add the node
358 	 */
359 	tmp = __find_node(nodeid, r);
360 	if (tmp) {
361 		spin_unlock(&nodes_lock);
362 		kfree(node);
363 		return tmp;
364 	}
365 
366 	hlist_add_head_rcu(&node->hlist, &node_hash[r]);
367 	spin_unlock(&nodes_lock);
368 
369 	node->debugfs = dlm_create_debug_comms_file(nodeid, node);
370 	return node;
371 }
372 
373 static int dlm_send_ack(int nodeid, uint32_t seq)
374 {
375 	int mb_len = sizeof(struct dlm_header);
376 	struct dlm_header *m_header;
377 	struct dlm_msg *msg;
378 	char *ppc;
379 
380 	msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
381 				   NULL, NULL);
382 	if (!msg)
383 		return -ENOMEM;
384 
385 	m_header = (struct dlm_header *)ppc;
386 
387 	m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
388 	m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
389 	m_header->h_length = cpu_to_le16(mb_len);
390 	m_header->h_cmd = DLM_ACK;
391 	m_header->u.h_seq = cpu_to_le32(seq);
392 
393 	dlm_lowcomms_commit_msg(msg);
394 	dlm_lowcomms_put_msg(msg);
395 
396 	return 0;
397 }
398 
399 static void dlm_send_ack_threshold(struct midcomms_node *node,
400 				   uint32_t threshold)
401 {
402 	uint32_t oval, nval;
403 	bool send_ack;
404 
405 	/* let only send one user trigger threshold to send ack back */
406 	do {
407 		oval = atomic_read(&node->ulp_delivered);
408 		send_ack = (oval > threshold);
409 		/* abort if threshold is not reached */
410 		if (!send_ack)
411 			break;
412 
413 		nval = 0;
414 		/* try to reset ulp_delivered counter */
415 	} while (atomic_cmpxchg(&node->ulp_delivered, oval, nval) != oval);
416 
417 	if (send_ack)
418 		dlm_send_ack(node->nodeid, atomic_read(&node->seq_next));
419 }
420 
421 static int dlm_send_fin(struct midcomms_node *node,
422 			void (*ack_rcv)(struct midcomms_node *node))
423 {
424 	int mb_len = sizeof(struct dlm_header);
425 	struct dlm_header *m_header;
426 	struct dlm_mhandle *mh;
427 	char *ppc;
428 
429 	mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
430 	if (!mh)
431 		return -ENOMEM;
432 
433 	set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
434 	mh->ack_rcv = ack_rcv;
435 
436 	m_header = (struct dlm_header *)ppc;
437 
438 	m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
439 	m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
440 	m_header->h_length = cpu_to_le16(mb_len);
441 	m_header->h_cmd = DLM_FIN;
442 
443 	pr_debug("sending fin msg to node %d\n", node->nodeid);
444 	dlm_midcomms_commit_mhandle(mh, NULL, 0);
445 
446 	return 0;
447 }
448 
449 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
450 {
451 	struct dlm_mhandle *mh;
452 
453 	rcu_read_lock();
454 	list_for_each_entry_rcu(mh, &node->send_queue, list) {
455 		if (before(mh->seq, seq)) {
456 			if (mh->ack_rcv)
457 				mh->ack_rcv(node);
458 		} else {
459 			/* send queue should be ordered */
460 			break;
461 		}
462 	}
463 
464 	spin_lock_bh(&node->send_queue_lock);
465 	list_for_each_entry_rcu(mh, &node->send_queue, list) {
466 		if (before(mh->seq, seq)) {
467 			dlm_mhandle_delete(node, mh);
468 		} else {
469 			/* send queue should be ordered */
470 			break;
471 		}
472 	}
473 	spin_unlock_bh(&node->send_queue_lock);
474 	rcu_read_unlock();
475 }
476 
477 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
478 {
479 	spin_lock(&node->state_lock);
480 	pr_debug("receive passive fin ack from node %d with state %s\n",
481 		 node->nodeid, dlm_state_str(node->state));
482 
483 	switch (node->state) {
484 	case DLM_LAST_ACK:
485 		/* DLM_CLOSED */
486 		midcomms_node_reset(node);
487 		break;
488 	case DLM_CLOSED:
489 		/* not valid but somehow we got what we want */
490 		wake_up(&node->shutdown_wait);
491 		break;
492 	default:
493 		spin_unlock(&node->state_lock);
494 		log_print("%s: unexpected state: %d",
495 			  __func__, node->state);
496 		WARN_ON_ONCE(1);
497 		return;
498 	}
499 	spin_unlock(&node->state_lock);
500 }
501 
502 static void dlm_receive_buffer_3_2_trace(uint32_t seq, union dlm_packet *p)
503 {
504 	switch (p->header.h_cmd) {
505 	case DLM_MSG:
506 		trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
507 		break;
508 	case DLM_RCOM:
509 		trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
510 		break;
511 	default:
512 		break;
513 	}
514 }
515 
516 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
517 					struct midcomms_node *node,
518 					uint32_t seq)
519 {
520 	bool is_expected_seq;
521 	uint32_t oval, nval;
522 
523 	do {
524 		oval = atomic_read(&node->seq_next);
525 		is_expected_seq = (oval == seq);
526 		if (!is_expected_seq)
527 			break;
528 
529 		nval = oval + 1;
530 	} while (atomic_cmpxchg(&node->seq_next, oval, nval) != oval);
531 
532 	if (is_expected_seq) {
533 		switch (p->header.h_cmd) {
534 		case DLM_FIN:
535 			spin_lock(&node->state_lock);
536 			pr_debug("receive fin msg from node %d with state %s\n",
537 				 node->nodeid, dlm_state_str(node->state));
538 
539 			switch (node->state) {
540 			case DLM_ESTABLISHED:
541 				dlm_send_ack(node->nodeid, nval);
542 
543 				/* passive shutdown DLM_LAST_ACK case 1
544 				 * additional we check if the node is used by
545 				 * cluster manager events at all.
546 				 */
547 				if (node->users == 0) {
548 					node->state = DLM_LAST_ACK;
549 					pr_debug("switch node %d to state %s case 1\n",
550 						 node->nodeid, dlm_state_str(node->state));
551 					set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
552 					dlm_send_fin(node, dlm_pas_fin_ack_rcv);
553 				} else {
554 					node->state = DLM_CLOSE_WAIT;
555 					pr_debug("switch node %d to state %s\n",
556 						 node->nodeid, dlm_state_str(node->state));
557 				}
558 				break;
559 			case DLM_FIN_WAIT1:
560 				dlm_send_ack(node->nodeid, nval);
561 				node->state = DLM_CLOSING;
562 				set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
563 				pr_debug("switch node %d to state %s\n",
564 					 node->nodeid, dlm_state_str(node->state));
565 				break;
566 			case DLM_FIN_WAIT2:
567 				dlm_send_ack(node->nodeid, nval);
568 				midcomms_node_reset(node);
569 				pr_debug("switch node %d to state %s\n",
570 					 node->nodeid, dlm_state_str(node->state));
571 				break;
572 			case DLM_LAST_ACK:
573 				/* probably remove_member caught it, do nothing */
574 				break;
575 			default:
576 				spin_unlock(&node->state_lock);
577 				log_print("%s: unexpected state: %d",
578 					  __func__, node->state);
579 				WARN_ON_ONCE(1);
580 				return;
581 			}
582 			spin_unlock(&node->state_lock);
583 			break;
584 		default:
585 			WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
586 			dlm_receive_buffer_3_2_trace(seq, p);
587 			dlm_receive_buffer(p, node->nodeid);
588 			atomic_inc(&node->ulp_delivered);
589 			/* unlikely case to send ack back when we don't transmit */
590 			dlm_send_ack_threshold(node, DLM_RECV_ACK_BACK_MSG_THRESHOLD);
591 			break;
592 		}
593 	} else {
594 		/* retry to ack message which we already have by sending back
595 		 * current node->seq_next number as ack.
596 		 */
597 		if (seq < oval)
598 			dlm_send_ack(node->nodeid, oval);
599 
600 		log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
601 				      seq, oval, node->nodeid);
602 	}
603 }
604 
605 static struct midcomms_node *
606 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
607 			      uint16_t msglen, int (*cb)(struct midcomms_node *node))
608 {
609 	struct midcomms_node *node = NULL;
610 	gfp_t allocation = 0;
611 	int ret;
612 
613 	switch (p->header.h_cmd) {
614 	case DLM_RCOM:
615 		if (msglen < sizeof(struct dlm_rcom)) {
616 			log_print("rcom msg too small: %u, will skip this message from node %d",
617 				  msglen, nodeid);
618 			return NULL;
619 		}
620 
621 		switch (p->rcom.rc_type) {
622 		case cpu_to_le32(DLM_RCOM_NAMES):
623 			fallthrough;
624 		case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
625 			fallthrough;
626 		case cpu_to_le32(DLM_RCOM_STATUS):
627 			fallthrough;
628 		case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
629 			node = nodeid2node(nodeid, 0);
630 			if (node) {
631 				spin_lock(&node->state_lock);
632 				if (node->state != DLM_ESTABLISHED)
633 					pr_debug("receive begin RCOM msg from node %d with state %s\n",
634 						 node->nodeid, dlm_state_str(node->state));
635 
636 				switch (node->state) {
637 				case DLM_CLOSED:
638 					node->state = DLM_ESTABLISHED;
639 					pr_debug("switch node %d to state %s\n",
640 						 node->nodeid, dlm_state_str(node->state));
641 					break;
642 				case DLM_ESTABLISHED:
643 					break;
644 				default:
645 					spin_unlock(&node->state_lock);
646 					return NULL;
647 				}
648 				spin_unlock(&node->state_lock);
649 			}
650 
651 			allocation = GFP_NOFS;
652 			break;
653 		default:
654 			break;
655 		}
656 
657 		break;
658 	default:
659 		break;
660 	}
661 
662 	node = nodeid2node(nodeid, allocation);
663 	if (!node) {
664 		switch (p->header.h_cmd) {
665 		case DLM_OPTS:
666 			if (msglen < sizeof(struct dlm_opts)) {
667 				log_print("opts msg too small: %u, will skip this message from node %d",
668 					  msglen, nodeid);
669 				return NULL;
670 			}
671 
672 			log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
673 					      p->opts.o_nextcmd, nodeid);
674 			break;
675 		default:
676 			log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
677 					      p->header.h_cmd, nodeid);
678 			break;
679 		}
680 
681 		return NULL;
682 	}
683 
684 	ret = cb(node);
685 	if (ret < 0)
686 		return NULL;
687 
688 	return node;
689 }
690 
691 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
692 {
693 	switch (node->version) {
694 	case DLM_VERSION_NOT_SET:
695 		node->version = DLM_VERSION_3_2;
696 		wake_up(&node->shutdown_wait);
697 		log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
698 			  node->nodeid);
699 		break;
700 	case DLM_VERSION_3_2:
701 		break;
702 	default:
703 		log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
704 				      DLM_VERSION_3_2, node->nodeid, node->version);
705 		return -1;
706 	}
707 
708 	return 0;
709 }
710 
711 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
712 {
713 	int len = msglen;
714 
715 	/* we only trust outer header msglen because
716 	 * it's checked against receive buffer length.
717 	 */
718 	if (len < sizeof(struct dlm_opts))
719 		return -1;
720 	len -= sizeof(struct dlm_opts);
721 
722 	if (len < le16_to_cpu(p->opts.o_optlen))
723 		return -1;
724 	len -= le16_to_cpu(p->opts.o_optlen);
725 
726 	switch (p->opts.o_nextcmd) {
727 	case DLM_FIN:
728 		if (len < sizeof(struct dlm_header)) {
729 			log_print("fin too small: %d, will skip this message from node %d",
730 				  len, nodeid);
731 			return -1;
732 		}
733 
734 		break;
735 	case DLM_MSG:
736 		if (len < sizeof(struct dlm_message)) {
737 			log_print("msg too small: %d, will skip this message from node %d",
738 				  msglen, nodeid);
739 			return -1;
740 		}
741 
742 		break;
743 	case DLM_RCOM:
744 		if (len < sizeof(struct dlm_rcom)) {
745 			log_print("rcom msg too small: %d, will skip this message from node %d",
746 				  len, nodeid);
747 			return -1;
748 		}
749 
750 		break;
751 	default:
752 		log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
753 			  p->opts.o_nextcmd, nodeid);
754 		return -1;
755 	}
756 
757 	return 0;
758 }
759 
760 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
761 {
762 	uint16_t msglen = le16_to_cpu(p->header.h_length);
763 	struct midcomms_node *node;
764 	uint32_t seq;
765 	int ret, idx;
766 
767 	idx = srcu_read_lock(&nodes_srcu);
768 	node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
769 					     dlm_midcomms_version_check_3_2);
770 	if (!node)
771 		goto out;
772 
773 	switch (p->header.h_cmd) {
774 	case DLM_RCOM:
775 		/* these rcom message we use to determine version.
776 		 * they have their own retransmission handling and
777 		 * are the first messages of dlm.
778 		 *
779 		 * length already checked.
780 		 */
781 		switch (p->rcom.rc_type) {
782 		case cpu_to_le32(DLM_RCOM_NAMES):
783 			fallthrough;
784 		case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
785 			fallthrough;
786 		case cpu_to_le32(DLM_RCOM_STATUS):
787 			fallthrough;
788 		case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
789 			break;
790 		default:
791 			log_print("unsupported rcom type received: %u, will skip this message from node %d",
792 				  le32_to_cpu(p->rcom.rc_type), nodeid);
793 			goto out;
794 		}
795 
796 		WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
797 		dlm_receive_buffer(p, nodeid);
798 		break;
799 	case DLM_OPTS:
800 		seq = le32_to_cpu(p->header.u.h_seq);
801 
802 		ret = dlm_opts_check_msglen(p, msglen, nodeid);
803 		if (ret < 0) {
804 			log_print("opts msg too small: %u, will skip this message from node %d",
805 				  msglen, nodeid);
806 			goto out;
807 		}
808 
809 		p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
810 					 le16_to_cpu(p->opts.o_optlen));
811 
812 		/* recheck inner msglen just if it's not garbage */
813 		msglen = le16_to_cpu(p->header.h_length);
814 		switch (p->header.h_cmd) {
815 		case DLM_RCOM:
816 			if (msglen < sizeof(struct dlm_rcom)) {
817 				log_print("inner rcom msg too small: %u, will skip this message from node %d",
818 					  msglen, nodeid);
819 				goto out;
820 			}
821 
822 			break;
823 		case DLM_MSG:
824 			if (msglen < sizeof(struct dlm_message)) {
825 				log_print("inner msg too small: %u, will skip this message from node %d",
826 					  msglen, nodeid);
827 				goto out;
828 			}
829 
830 			break;
831 		case DLM_FIN:
832 			if (msglen < sizeof(struct dlm_header)) {
833 				log_print("inner fin too small: %u, will skip this message from node %d",
834 					  msglen, nodeid);
835 				goto out;
836 			}
837 
838 			break;
839 		default:
840 			log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
841 				  msglen, nodeid);
842 			goto out;
843 		}
844 
845 		dlm_midcomms_receive_buffer(p, node, seq);
846 		break;
847 	case DLM_ACK:
848 		seq = le32_to_cpu(p->header.u.h_seq);
849 		dlm_receive_ack(node, seq);
850 		break;
851 	default:
852 		log_print("unsupported h_cmd received: %u, will skip this message from node %d",
853 			  p->header.h_cmd, nodeid);
854 		break;
855 	}
856 
857 out:
858 	srcu_read_unlock(&nodes_srcu, idx);
859 }
860 
861 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
862 {
863 	switch (node->version) {
864 	case DLM_VERSION_NOT_SET:
865 		node->version = DLM_VERSION_3_1;
866 		wake_up(&node->shutdown_wait);
867 		log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
868 			  node->nodeid);
869 		break;
870 	case DLM_VERSION_3_1:
871 		break;
872 	default:
873 		log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
874 				      DLM_VERSION_3_1, node->nodeid, node->version);
875 		return -1;
876 	}
877 
878 	return 0;
879 }
880 
881 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
882 {
883 	uint16_t msglen = le16_to_cpu(p->header.h_length);
884 	struct midcomms_node *node;
885 	int idx;
886 
887 	idx = srcu_read_lock(&nodes_srcu);
888 	node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
889 					     dlm_midcomms_version_check_3_1);
890 	if (!node) {
891 		srcu_read_unlock(&nodes_srcu, idx);
892 		return;
893 	}
894 	srcu_read_unlock(&nodes_srcu, idx);
895 
896 	switch (p->header.h_cmd) {
897 	case DLM_RCOM:
898 		/* length already checked */
899 		break;
900 	case DLM_MSG:
901 		if (msglen < sizeof(struct dlm_message)) {
902 			log_print("msg too small: %u, will skip this message from node %d",
903 				  msglen, nodeid);
904 			return;
905 		}
906 
907 		break;
908 	default:
909 		log_print("unsupported h_cmd received: %u, will skip this message from node %d",
910 			  p->header.h_cmd, nodeid);
911 		return;
912 	}
913 
914 	dlm_receive_buffer(p, nodeid);
915 }
916 
917 int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
918 {
919 	const unsigned char *ptr = buf;
920 	const struct dlm_header *hd;
921 	uint16_t msglen;
922 	int ret = 0;
923 
924 	while (len >= sizeof(struct dlm_header)) {
925 		hd = (struct dlm_header *)ptr;
926 
927 		/* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
928 		 * less than dlm_header size.
929 		 *
930 		 * Some messages does not have a 8 byte length boundary yet
931 		 * which can occur in a unaligned memory access of some dlm
932 		 * messages. However this problem need to be fixed at the
933 		 * sending side, for now it seems nobody run into architecture
934 		 * related issues yet but it slows down some processing.
935 		 * Fixing this issue should be scheduled in future by doing
936 		 * the next major version bump.
937 		 */
938 		msglen = le16_to_cpu(hd->h_length);
939 		if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
940 		    msglen < sizeof(struct dlm_header)) {
941 			log_print("received invalid length header: %u from node %d, will abort message parsing",
942 				  msglen, nodeid);
943 			return -EBADMSG;
944 		}
945 
946 		/* caller will take care that leftover
947 		 * will be parsed next call with more data
948 		 */
949 		if (msglen > len)
950 			break;
951 
952 		ret += msglen;
953 		len -= msglen;
954 		ptr += msglen;
955 	}
956 
957 	return ret;
958 }
959 
960 /*
961  * Called from the low-level comms layer to process a buffer of
962  * commands.
963  */
964 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
965 {
966 	const unsigned char *ptr = buf;
967 	const struct dlm_header *hd;
968 	uint16_t msglen;
969 	int ret = 0;
970 
971 	while (len >= sizeof(struct dlm_header)) {
972 		hd = (struct dlm_header *)ptr;
973 
974 		msglen = le16_to_cpu(hd->h_length);
975 		if (msglen > len)
976 			break;
977 
978 		switch (hd->h_version) {
979 		case cpu_to_le32(DLM_VERSION_3_1):
980 			dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
981 			break;
982 		case cpu_to_le32(DLM_VERSION_3_2):
983 			dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
984 			break;
985 		default:
986 			log_print("received invalid version header: %u from node %d, will skip this message",
987 				  le32_to_cpu(hd->h_version), nodeid);
988 			break;
989 		}
990 
991 		ret += msglen;
992 		len -= msglen;
993 		ptr += msglen;
994 	}
995 
996 	return ret;
997 }
998 
999 void dlm_midcomms_unack_msg_resend(int nodeid)
1000 {
1001 	struct midcomms_node *node;
1002 	struct dlm_mhandle *mh;
1003 	int idx, ret;
1004 
1005 	idx = srcu_read_lock(&nodes_srcu);
1006 	node = nodeid2node(nodeid, 0);
1007 	if (!node) {
1008 		srcu_read_unlock(&nodes_srcu, idx);
1009 		return;
1010 	}
1011 
1012 	/* old protocol, we don't support to retransmit on failure */
1013 	switch (node->version) {
1014 	case DLM_VERSION_3_2:
1015 		break;
1016 	default:
1017 		srcu_read_unlock(&nodes_srcu, idx);
1018 		return;
1019 	}
1020 
1021 	rcu_read_lock();
1022 	list_for_each_entry_rcu(mh, &node->send_queue, list) {
1023 		if (!mh->committed)
1024 			continue;
1025 
1026 		ret = dlm_lowcomms_resend_msg(mh->msg);
1027 		if (!ret)
1028 			log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1029 					      mh->seq, node->nodeid);
1030 	}
1031 	rcu_read_unlock();
1032 	srcu_read_unlock(&nodes_srcu, idx);
1033 }
1034 
1035 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1036 				 uint32_t seq)
1037 {
1038 	opts->o_header.h_cmd = DLM_OPTS;
1039 	opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1040 	opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1041 	opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1042 	opts->o_header.u.h_seq = cpu_to_le32(seq);
1043 }
1044 
1045 static void midcomms_new_msg_cb(void *data)
1046 {
1047 	struct dlm_mhandle *mh = data;
1048 
1049 	atomic_inc(&mh->node->send_queue_cnt);
1050 
1051 	spin_lock_bh(&mh->node->send_queue_lock);
1052 	list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1053 	spin_unlock_bh(&mh->node->send_queue_lock);
1054 
1055 	mh->seq = atomic_fetch_inc(&mh->node->seq_send);
1056 }
1057 
1058 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1059 						int len, gfp_t allocation, char **ppc)
1060 {
1061 	struct dlm_opts *opts;
1062 	struct dlm_msg *msg;
1063 
1064 	msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1065 				   allocation, ppc, midcomms_new_msg_cb, mh);
1066 	if (!msg)
1067 		return NULL;
1068 
1069 	opts = (struct dlm_opts *)*ppc;
1070 	mh->opts = opts;
1071 
1072 	/* add possible options here */
1073 	dlm_fill_opts_header(opts, len, mh->seq);
1074 
1075 	*ppc += sizeof(*opts);
1076 	mh->inner_p = (const union dlm_packet *)*ppc;
1077 	return msg;
1078 }
1079 
1080 /* avoid false positive for nodes_srcu, unlock happens in
1081  * dlm_midcomms_commit_mhandle which is a must call if success
1082  */
1083 #ifndef __CHECKER__
1084 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1085 					     gfp_t allocation, char **ppc)
1086 {
1087 	struct midcomms_node *node;
1088 	struct dlm_mhandle *mh;
1089 	struct dlm_msg *msg;
1090 	int idx;
1091 
1092 	idx = srcu_read_lock(&nodes_srcu);
1093 	node = nodeid2node(nodeid, 0);
1094 	if (!node) {
1095 		WARN_ON_ONCE(1);
1096 		goto err;
1097 	}
1098 
1099 	/* this is a bug, however we going on and hope it will be resolved */
1100 	WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1101 
1102 	mh = dlm_allocate_mhandle(allocation);
1103 	if (!mh)
1104 		goto err;
1105 
1106 	mh->committed = false;
1107 	mh->ack_rcv = NULL;
1108 	mh->idx = idx;
1109 	mh->node = node;
1110 
1111 	switch (node->version) {
1112 	case DLM_VERSION_3_1:
1113 		msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1114 					   NULL, NULL);
1115 		if (!msg) {
1116 			dlm_free_mhandle(mh);
1117 			goto err;
1118 		}
1119 
1120 		break;
1121 	case DLM_VERSION_3_2:
1122 		msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1123 					       ppc);
1124 		if (!msg) {
1125 			dlm_free_mhandle(mh);
1126 			goto err;
1127 		}
1128 
1129 		/* send ack back if necessary */
1130 		dlm_send_ack_threshold(node, DLM_SEND_ACK_BACK_MSG_THRESHOLD);
1131 		break;
1132 	default:
1133 		dlm_free_mhandle(mh);
1134 		WARN_ON_ONCE(1);
1135 		goto err;
1136 	}
1137 
1138 	mh->msg = msg;
1139 
1140 	/* keep in mind that is a must to call
1141 	 * dlm_midcomms_commit_msg() which releases
1142 	 * nodes_srcu using mh->idx which is assumed
1143 	 * here that the application will call it.
1144 	 */
1145 	return mh;
1146 
1147 err:
1148 	srcu_read_unlock(&nodes_srcu, idx);
1149 	return NULL;
1150 }
1151 #endif
1152 
1153 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1154 					      const void *name, int namelen)
1155 {
1156 	switch (mh->inner_p->header.h_cmd) {
1157 	case DLM_MSG:
1158 		trace_dlm_send_message(mh->node->nodeid, mh->seq,
1159 				       &mh->inner_p->message,
1160 				       name, namelen);
1161 		break;
1162 	case DLM_RCOM:
1163 		trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1164 				    &mh->inner_p->rcom);
1165 		break;
1166 	default:
1167 		/* nothing to trace */
1168 		break;
1169 	}
1170 }
1171 
1172 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1173 					const void *name, int namelen)
1174 {
1175 	/* nexthdr chain for fast lookup */
1176 	mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1177 	mh->committed = true;
1178 	dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1179 	dlm_lowcomms_commit_msg(mh->msg);
1180 }
1181 
1182 /* avoid false positive for nodes_srcu, lock was happen in
1183  * dlm_midcomms_get_mhandle
1184  */
1185 #ifndef __CHECKER__
1186 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1187 				 const void *name, int namelen)
1188 {
1189 
1190 	switch (mh->node->version) {
1191 	case DLM_VERSION_3_1:
1192 		srcu_read_unlock(&nodes_srcu, mh->idx);
1193 
1194 		dlm_lowcomms_commit_msg(mh->msg);
1195 		dlm_lowcomms_put_msg(mh->msg);
1196 		/* mh is not part of rcu list in this case */
1197 		dlm_free_mhandle(mh);
1198 		break;
1199 	case DLM_VERSION_3_2:
1200 		/* held rcu read lock here, because we sending the
1201 		 * dlm message out, when we do that we could receive
1202 		 * an ack back which releases the mhandle and we
1203 		 * get a use after free.
1204 		 */
1205 		rcu_read_lock();
1206 		dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1207 		srcu_read_unlock(&nodes_srcu, mh->idx);
1208 		rcu_read_unlock();
1209 		break;
1210 	default:
1211 		srcu_read_unlock(&nodes_srcu, mh->idx);
1212 		WARN_ON_ONCE(1);
1213 		break;
1214 	}
1215 }
1216 #endif
1217 
1218 int dlm_midcomms_start(void)
1219 {
1220 	return dlm_lowcomms_start();
1221 }
1222 
1223 void dlm_midcomms_stop(void)
1224 {
1225 	dlm_lowcomms_stop();
1226 }
1227 
1228 void dlm_midcomms_init(void)
1229 {
1230 	int i;
1231 
1232 	for (i = 0; i < CONN_HASH_SIZE; i++)
1233 		INIT_HLIST_HEAD(&node_hash[i]);
1234 
1235 	dlm_lowcomms_init();
1236 }
1237 
1238 void dlm_midcomms_exit(void)
1239 {
1240 	dlm_lowcomms_exit();
1241 }
1242 
1243 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1244 {
1245 	spin_lock(&node->state_lock);
1246 	pr_debug("receive active fin ack from node %d with state %s\n",
1247 		 node->nodeid, dlm_state_str(node->state));
1248 
1249 	switch (node->state) {
1250 	case DLM_FIN_WAIT1:
1251 		node->state = DLM_FIN_WAIT2;
1252 		pr_debug("switch node %d to state %s\n",
1253 			 node->nodeid, dlm_state_str(node->state));
1254 		break;
1255 	case DLM_CLOSING:
1256 		midcomms_node_reset(node);
1257 		pr_debug("switch node %d to state %s\n",
1258 			 node->nodeid, dlm_state_str(node->state));
1259 		break;
1260 	case DLM_CLOSED:
1261 		/* not valid but somehow we got what we want */
1262 		wake_up(&node->shutdown_wait);
1263 		break;
1264 	default:
1265 		spin_unlock(&node->state_lock);
1266 		log_print("%s: unexpected state: %d",
1267 			  __func__, node->state);
1268 		WARN_ON_ONCE(1);
1269 		return;
1270 	}
1271 	spin_unlock(&node->state_lock);
1272 }
1273 
1274 void dlm_midcomms_add_member(int nodeid)
1275 {
1276 	struct midcomms_node *node;
1277 	int idx;
1278 
1279 	idx = srcu_read_lock(&nodes_srcu);
1280 	node = nodeid2node(nodeid, GFP_NOFS);
1281 	if (!node) {
1282 		srcu_read_unlock(&nodes_srcu, idx);
1283 		return;
1284 	}
1285 
1286 	spin_lock(&node->state_lock);
1287 	if (!node->users) {
1288 		pr_debug("receive add member from node %d with state %s\n",
1289 			 node->nodeid, dlm_state_str(node->state));
1290 		switch (node->state) {
1291 		case DLM_ESTABLISHED:
1292 			break;
1293 		case DLM_CLOSED:
1294 			node->state = DLM_ESTABLISHED;
1295 			pr_debug("switch node %d to state %s\n",
1296 				 node->nodeid, dlm_state_str(node->state));
1297 			break;
1298 		default:
1299 			/* some invalid state passive shutdown
1300 			 * was failed, we try to reset and
1301 			 * hope it will go on.
1302 			 */
1303 			log_print("reset node %d because shutdown stuck",
1304 				  node->nodeid);
1305 
1306 			midcomms_node_reset(node);
1307 			node->state = DLM_ESTABLISHED;
1308 			break;
1309 		}
1310 	}
1311 
1312 	node->users++;
1313 	pr_debug("node %d users inc count %d\n", nodeid, node->users);
1314 	spin_unlock(&node->state_lock);
1315 
1316 	srcu_read_unlock(&nodes_srcu, idx);
1317 }
1318 
1319 void dlm_midcomms_remove_member(int nodeid)
1320 {
1321 	struct midcomms_node *node;
1322 	int idx;
1323 
1324 	idx = srcu_read_lock(&nodes_srcu);
1325 	node = nodeid2node(nodeid, 0);
1326 	if (!node) {
1327 		srcu_read_unlock(&nodes_srcu, idx);
1328 		return;
1329 	}
1330 
1331 	spin_lock(&node->state_lock);
1332 	node->users--;
1333 	pr_debug("node %d users dec count %d\n", nodeid, node->users);
1334 
1335 	/* hitting users count to zero means the
1336 	 * other side is running dlm_midcomms_stop()
1337 	 * we meet us to have a clean disconnect.
1338 	 */
1339 	if (node->users == 0) {
1340 		pr_debug("receive remove member from node %d with state %s\n",
1341 			 node->nodeid, dlm_state_str(node->state));
1342 		switch (node->state) {
1343 		case DLM_ESTABLISHED:
1344 			break;
1345 		case DLM_CLOSE_WAIT:
1346 			/* passive shutdown DLM_LAST_ACK case 2 */
1347 			node->state = DLM_LAST_ACK;
1348 			pr_debug("switch node %d to state %s case 2\n",
1349 				 node->nodeid, dlm_state_str(node->state));
1350 			set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1351 			dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1352 			break;
1353 		case DLM_LAST_ACK:
1354 			/* probably receive fin caught it, do nothing */
1355 			break;
1356 		case DLM_CLOSED:
1357 			/* already gone, do nothing */
1358 			break;
1359 		default:
1360 			log_print("%s: unexpected state: %d",
1361 				  __func__, node->state);
1362 			break;
1363 		}
1364 	}
1365 	spin_unlock(&node->state_lock);
1366 
1367 	srcu_read_unlock(&nodes_srcu, idx);
1368 }
1369 
1370 static void midcomms_node_release(struct rcu_head *rcu)
1371 {
1372 	struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1373 
1374 	WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1375 	dlm_send_queue_flush(node);
1376 	kfree(node);
1377 }
1378 
1379 void dlm_midcomms_version_wait(void)
1380 {
1381 	struct midcomms_node *node;
1382 	int i, idx, ret;
1383 
1384 	idx = srcu_read_lock(&nodes_srcu);
1385 	for (i = 0; i < CONN_HASH_SIZE; i++) {
1386 		hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1387 			ret = wait_event_timeout(node->shutdown_wait,
1388 						 node->version != DLM_VERSION_NOT_SET ||
1389 						 node->state == DLM_CLOSED ||
1390 						 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1391 						 DLM_SHUTDOWN_TIMEOUT);
1392 			if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1393 				pr_debug("version wait timed out for node %d with state %s\n",
1394 					 node->nodeid, dlm_state_str(node->state));
1395 		}
1396 	}
1397 	srcu_read_unlock(&nodes_srcu, idx);
1398 }
1399 
1400 static void midcomms_shutdown(struct midcomms_node *node)
1401 {
1402 	int ret;
1403 
1404 	/* old protocol, we don't wait for pending operations */
1405 	switch (node->version) {
1406 	case DLM_VERSION_3_2:
1407 		break;
1408 	default:
1409 		return;
1410 	}
1411 
1412 	spin_lock(&node->state_lock);
1413 	pr_debug("receive active shutdown for node %d with state %s\n",
1414 		 node->nodeid, dlm_state_str(node->state));
1415 	switch (node->state) {
1416 	case DLM_ESTABLISHED:
1417 		node->state = DLM_FIN_WAIT1;
1418 		pr_debug("switch node %d to state %s case 2\n",
1419 			 node->nodeid, dlm_state_str(node->state));
1420 		dlm_send_fin(node, dlm_act_fin_ack_rcv);
1421 		break;
1422 	case DLM_CLOSED:
1423 		/* we have what we want */
1424 		break;
1425 	default:
1426 		/* busy to enter DLM_FIN_WAIT1, wait until passive
1427 		 * done in shutdown_wait to enter DLM_CLOSED.
1428 		 */
1429 		break;
1430 	}
1431 	spin_unlock(&node->state_lock);
1432 
1433 	if (DLM_DEBUG_FENCE_TERMINATION)
1434 		msleep(5000);
1435 
1436 	/* wait for other side dlm + fin */
1437 	ret = wait_event_timeout(node->shutdown_wait,
1438 				 node->state == DLM_CLOSED ||
1439 				 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1440 				 DLM_SHUTDOWN_TIMEOUT);
1441 	if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1442 		pr_debug("active shutdown timed out for node %d with state %s\n",
1443 			 node->nodeid, dlm_state_str(node->state));
1444 	else
1445 		pr_debug("active shutdown done for node %d with state %s\n",
1446 			 node->nodeid, dlm_state_str(node->state));
1447 }
1448 
1449 void dlm_midcomms_shutdown(void)
1450 {
1451 	struct midcomms_node *node;
1452 	int i, idx;
1453 
1454 	mutex_lock(&close_lock);
1455 	idx = srcu_read_lock(&nodes_srcu);
1456 	for (i = 0; i < CONN_HASH_SIZE; i++) {
1457 		hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1458 			midcomms_shutdown(node);
1459 
1460 			dlm_delete_debug_comms_file(node->debugfs);
1461 
1462 			spin_lock(&nodes_lock);
1463 			hlist_del_rcu(&node->hlist);
1464 			spin_unlock(&nodes_lock);
1465 
1466 			call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1467 		}
1468 	}
1469 	srcu_read_unlock(&nodes_srcu, idx);
1470 	mutex_unlock(&close_lock);
1471 
1472 	dlm_lowcomms_shutdown();
1473 }
1474 
1475 int dlm_midcomms_close(int nodeid)
1476 {
1477 	struct midcomms_node *node;
1478 	int idx, ret;
1479 
1480 	idx = srcu_read_lock(&nodes_srcu);
1481 	/* Abort pending close/remove operation */
1482 	node = nodeid2node(nodeid, 0);
1483 	if (node) {
1484 		/* let shutdown waiters leave */
1485 		set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1486 		wake_up(&node->shutdown_wait);
1487 	}
1488 	srcu_read_unlock(&nodes_srcu, idx);
1489 
1490 	synchronize_srcu(&nodes_srcu);
1491 
1492 	idx = srcu_read_lock(&nodes_srcu);
1493 	mutex_lock(&close_lock);
1494 	node = nodeid2node(nodeid, 0);
1495 	if (!node) {
1496 		mutex_unlock(&close_lock);
1497 		srcu_read_unlock(&nodes_srcu, idx);
1498 		return dlm_lowcomms_close(nodeid);
1499 	}
1500 
1501 	ret = dlm_lowcomms_close(nodeid);
1502 	spin_lock(&node->state_lock);
1503 	midcomms_node_reset(node);
1504 	spin_unlock(&node->state_lock);
1505 	srcu_read_unlock(&nodes_srcu, idx);
1506 	mutex_unlock(&close_lock);
1507 
1508 	return ret;
1509 }
1510 
1511 /* debug functionality to send raw dlm msg from user space */
1512 struct dlm_rawmsg_data {
1513 	struct midcomms_node *node;
1514 	void *buf;
1515 };
1516 
1517 static void midcomms_new_rawmsg_cb(void *data)
1518 {
1519 	struct dlm_rawmsg_data *rd = data;
1520 	struct dlm_header *h = rd->buf;
1521 
1522 	switch (h->h_version) {
1523 	case cpu_to_le32(DLM_VERSION_3_1):
1524 		break;
1525 	default:
1526 		switch (h->h_cmd) {
1527 		case DLM_OPTS:
1528 			if (!h->u.h_seq)
1529 				h->u.h_seq = cpu_to_le32(atomic_fetch_inc(&rd->node->seq_send));
1530 			break;
1531 		default:
1532 			break;
1533 		}
1534 		break;
1535 	}
1536 }
1537 
1538 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1539 			     int buflen)
1540 {
1541 	struct dlm_rawmsg_data rd;
1542 	struct dlm_msg *msg;
1543 	char *msgbuf;
1544 
1545 	rd.node = node;
1546 	rd.buf = buf;
1547 
1548 	msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1549 				   &msgbuf, midcomms_new_rawmsg_cb, &rd);
1550 	if (!msg)
1551 		return -ENOMEM;
1552 
1553 	memcpy(msgbuf, buf, buflen);
1554 	dlm_lowcomms_commit_msg(msg);
1555 	return 0;
1556 }
1557 
1558