projects/linux·blob master · refs

1515 lines · 39.2 KiB

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
154struct 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
198struct 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
215static struct hlist_head node_hash[CONN_HASH_SIZE];
216static DEFINE_SPINLOCK(nodes_lock);
217DEFINE_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 */
225static DEFINE_MUTEX(close_lock);
226
227struct kmem_cache *dlm_midcomms_cache_create(void)
228{
229 return KMEM_CACHE(dlm_mhandle, 0);
230}
231
232static inline const char *dlm_state_str(int state)
233{
234 switch (state) {
235 case DLM_CLOSED:
236 return "CLOSED";
237 case DLM_ESTABLISHED:
238 return "ESTABLISHED";
239 case DLM_FIN_WAIT1:
240 return "FIN_WAIT1";
241 case DLM_FIN_WAIT2:
242 return "FIN_WAIT2";
243 case DLM_CLOSE_WAIT:
244 return "CLOSE_WAIT";
245 case DLM_LAST_ACK:
246 return "LAST_ACK";
247 case DLM_CLOSING:
248 return "CLOSING";
249 default:
250 return "UNKNOWN";
251 }
252}
253
254const char *dlm_midcomms_state(struct midcomms_node *node)
255{
256 return dlm_state_str(node->state);
257}
258
259unsigned long dlm_midcomms_flags(struct midcomms_node *node)
260{
261 return node->flags;
262}
263
264int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
265{
266 return atomic_read(&node->send_queue_cnt);
267}
268
269uint32_t dlm_midcomms_version(struct midcomms_node *node)
270{
271 return node->version;
272}
273
274static struct midcomms_node *__find_node(int nodeid, int r)
275{
276 struct midcomms_node *node;
277
278 hlist_for_each_entry_srcu(node, &node_hash[r], hlist,
279 srcu_read_lock_held(&nodes_srcu)) {
280 if (node->nodeid == nodeid)
281 return node;
282 }
283
284 return NULL;
285}
286
287static 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
295static 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
303static 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
318static 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
333static struct midcomms_node *nodeid2node(int nodeid)
334{
335 return __find_node(nodeid, nodeid_hash(nodeid));
336}
337
338int dlm_midcomms_addr(int nodeid, struct sockaddr_storage *addr)
339{
340 int ret, idx, r = nodeid_hash(nodeid);
341 struct midcomms_node *node;
342
343 ret = dlm_lowcomms_addr(nodeid, addr);
344 if (ret)
345 return ret;
346
347 idx = srcu_read_lock(&nodes_srcu);
348 node = __find_node(nodeid, r);
349 if (node) {
350 srcu_read_unlock(&nodes_srcu, idx);
351 return 0;
352 }
353 srcu_read_unlock(&nodes_srcu, idx);
354
355 node = kmalloc_obj(*node, GFP_NOFS);
356 if (!node)
357 return -ENOMEM;
358
359 node->debugfs = dlm_create_debug_comms_file(nodeid, node);
360 node->nodeid = nodeid;
361 spin_lock_init(&node->state_lock);
362 spin_lock_init(&node->send_queue_lock);
363 atomic_set(&node->send_queue_cnt, 0);
364 INIT_LIST_HEAD(&node->send_queue);
365 init_waitqueue_head(&node->shutdown_wait);
366 node->users = 0;
367 midcomms_node_reset(node);
368
369 spin_lock_bh(&nodes_lock);
370 hlist_add_head_rcu(&node->hlist, &node_hash[r]);
371 spin_unlock_bh(&nodes_lock);
372
373 return 0;
374}
375
376static int dlm_send_ack(int nodeid, uint32_t seq)
377{
378 int mb_len = sizeof(struct dlm_header);
379 struct dlm_header *m_header;
380 struct dlm_msg *msg;
381 char *ppc;
382
383 msg = dlm_lowcomms_new_msg(nodeid, mb_len, &ppc, NULL, NULL);
384 if (!msg)
385 return -ENOMEM;
386
387 m_header = (struct dlm_header *)ppc;
388
389 m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
390 m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
391 m_header->h_length = cpu_to_le16(mb_len);
392 m_header->h_cmd = DLM_ACK;
393 m_header->u.h_seq = cpu_to_le32(seq);
394
395 dlm_lowcomms_commit_msg(msg);
396 dlm_lowcomms_put_msg(msg);
397
398 return 0;
399}
400
401static void dlm_send_ack_threshold(struct midcomms_node *node,
402 uint32_t threshold)
403{
404 uint32_t oval, nval;
405 bool send_ack;
406
407 /* let only send one user trigger threshold to send ack back */
408 do {
409 oval = atomic_read(&node->ulp_delivered);
410 send_ack = (oval > threshold);
411 /* abort if threshold is not reached */
412 if (!send_ack)
413 break;
414
415 nval = 0;
416 /* try to reset ulp_delivered counter */
417 } while (atomic_cmpxchg(&node->ulp_delivered, oval, nval) != oval);
418
419 if (send_ack)
420 dlm_send_ack(node->nodeid, atomic_read(&node->seq_next));
421}
422
423static int dlm_send_fin(struct midcomms_node *node,
424 void (*ack_rcv)(struct midcomms_node *node))
425{
426 int mb_len = sizeof(struct dlm_header);
427 struct dlm_header *m_header;
428 struct dlm_mhandle *mh;
429 char *ppc;
430
431 mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, &ppc);
432 if (!mh)
433 return -ENOMEM;
434
435 set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
436 mh->ack_rcv = ack_rcv;
437
438 m_header = (struct dlm_header *)ppc;
439
440 m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
441 m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
442 m_header->h_length = cpu_to_le16(mb_len);
443 m_header->h_cmd = DLM_FIN;
444
445 pr_debug("sending fin msg to node %d\n", node->nodeid);
446 dlm_midcomms_commit_mhandle(mh, NULL, 0);
447
448 return 0;
449}
450
451static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
452{
453 struct dlm_mhandle *mh;
454
455 rcu_read_lock();
456 list_for_each_entry_rcu(mh, &node->send_queue, list) {
457 if (before(mh->seq, seq)) {
458 if (mh->ack_rcv)
459 mh->ack_rcv(node);
460 } else {
461 /* send queue should be ordered */
462 break;
463 }
464 }
465
466 spin_lock_bh(&node->send_queue_lock);
467 list_for_each_entry_rcu(mh, &node->send_queue, list) {
468 if (before(mh->seq, seq)) {
469 dlm_mhandle_delete(node, mh);
470 } else {
471 /* send queue should be ordered */
472 break;
473 }
474 }
475 spin_unlock_bh(&node->send_queue_lock);
476 rcu_read_unlock();
477}
478
479static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
480{
481 spin_lock_bh(&node->state_lock);
482 pr_debug("receive passive fin ack from node %d with state %s\n",
483 node->nodeid, dlm_state_str(node->state));
484
485 switch (node->state) {
486 case DLM_LAST_ACK:
487 /* DLM_CLOSED */
488 midcomms_node_reset(node);
489 break;
490 case DLM_CLOSED:
491 /* not valid but somehow we got what we want */
492 wake_up(&node->shutdown_wait);
493 break;
494 default:
495 spin_unlock_bh(&node->state_lock);
496 log_print("%s: unexpected state: %d",
497 __func__, node->state);
498 WARN_ON_ONCE(1);
499 return;
500 }
501 spin_unlock_bh(&node->state_lock);
502}
503
504static void dlm_receive_buffer_3_2_trace(uint32_t seq,
505 const union dlm_packet *p)
506{
507 switch (p->header.h_cmd) {
508 case DLM_MSG:
509 trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
510 break;
511 case DLM_RCOM:
512 trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
513 break;
514 default:
515 break;
516 }
517}
518
519static void dlm_midcomms_receive_buffer(const union dlm_packet *p,
520 struct midcomms_node *node,
521 uint32_t seq)
522{
523 bool is_expected_seq;
524 uint32_t oval, nval;
525
526 do {
527 oval = atomic_read(&node->seq_next);
528 is_expected_seq = (oval == seq);
529 if (!is_expected_seq)
530 break;
531
532 nval = oval + 1;
533 } while (atomic_cmpxchg(&node->seq_next, oval, nval) != oval);
534
535 if (is_expected_seq) {
536 switch (p->header.h_cmd) {
537 case DLM_FIN:
538 spin_lock_bh(&node->state_lock);
539 pr_debug("receive fin msg from node %d with state %s\n",
540 node->nodeid, dlm_state_str(node->state));
541
542 switch (node->state) {
543 case DLM_ESTABLISHED:
544 dlm_send_ack(node->nodeid, nval);
545
546 /* passive shutdown DLM_LAST_ACK case 1
547 * additional we check if the node is used by
548 * cluster manager events at all.
549 */
550 if (node->users == 0) {
551 node->state = DLM_LAST_ACK;
552 pr_debug("switch node %d to state %s case 1\n",
553 node->nodeid, dlm_state_str(node->state));
554 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
555 dlm_send_fin(node, dlm_pas_fin_ack_rcv);
556 } else {
557 node->state = DLM_CLOSE_WAIT;
558 pr_debug("switch node %d to state %s\n",
559 node->nodeid, dlm_state_str(node->state));
560 }
561 break;
562 case DLM_FIN_WAIT1:
563 dlm_send_ack(node->nodeid, nval);
564 node->state = DLM_CLOSING;
565 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
566 pr_debug("switch node %d to state %s\n",
567 node->nodeid, dlm_state_str(node->state));
568 break;
569 case DLM_FIN_WAIT2:
570 dlm_send_ack(node->nodeid, nval);
571 midcomms_node_reset(node);
572 pr_debug("switch node %d to state %s\n",
573 node->nodeid, dlm_state_str(node->state));
574 break;
575 case DLM_LAST_ACK:
576 /* probably remove_member caught it, do nothing */
577 break;
578 default:
579 spin_unlock_bh(&node->state_lock);
580 log_print("%s: unexpected state: %d",
581 __func__, node->state);
582 WARN_ON_ONCE(1);
583 return;
584 }
585 spin_unlock_bh(&node->state_lock);
586 break;
587 default:
588 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
589 dlm_receive_buffer_3_2_trace(seq, p);
590 dlm_receive_buffer(p, node->nodeid);
591 atomic_inc(&node->ulp_delivered);
592 /* unlikely case to send ack back when we don't transmit */
593 dlm_send_ack_threshold(node, DLM_RECV_ACK_BACK_MSG_THRESHOLD);
594 break;
595 }
596 } else {
597 /* retry to ack message which we already have by sending back
598 * current node->seq_next number as ack.
599 */
600 if (seq < oval)
601 dlm_send_ack(node->nodeid, oval);
602
603 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
604 seq, oval, node->nodeid);
605 }
606}
607
608static int dlm_opts_check_msglen(const union dlm_packet *p, uint16_t msglen,
609 int nodeid)
610{
611 int len = msglen;
612
613 /* we only trust outer header msglen because
614 * it's checked against receive buffer length.
615 */
616 if (len < sizeof(struct dlm_opts))
617 return -1;
618 len -= sizeof(struct dlm_opts);
619
620 if (len < le16_to_cpu(p->opts.o_optlen))
621 return -1;
622 len -= le16_to_cpu(p->opts.o_optlen);
623
624 switch (p->opts.o_nextcmd) {
625 case DLM_FIN:
626 if (len < sizeof(struct dlm_header)) {
627 log_print("fin too small: %d, will skip this message from node %d",
628 len, nodeid);
629 return -1;
630 }
631
632 break;
633 case DLM_MSG:
634 if (len < sizeof(struct dlm_message)) {
635 log_print("msg too small: %d, will skip this message from node %d",
636 msglen, nodeid);
637 return -1;
638 }
639
640 break;
641 case DLM_RCOM:
642 if (len < sizeof(struct dlm_rcom)) {
643 log_print("rcom msg too small: %d, will skip this message from node %d",
644 len, nodeid);
645 return -1;
646 }
647
648 break;
649 default:
650 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
651 p->opts.o_nextcmd, nodeid);
652 return -1;
653 }
654
655 return 0;
656}
657
658static void dlm_midcomms_receive_buffer_3_2(const union dlm_packet *p, int nodeid)
659{
660 uint16_t msglen = le16_to_cpu(p->header.h_length);
661 struct midcomms_node *node;
662 uint32_t seq;
663 int ret, idx;
664
665 idx = srcu_read_lock(&nodes_srcu);
666 node = nodeid2node(nodeid);
667 if (WARN_ON_ONCE(!node))
668 goto out;
669
670 switch (node->version) {
671 case DLM_VERSION_NOT_SET:
672 node->version = DLM_VERSION_3_2;
673 wake_up(&node->shutdown_wait);
674 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
675 node->nodeid);
676
677 spin_lock(&node->state_lock);
678 switch (node->state) {
679 case DLM_CLOSED:
680 node->state = DLM_ESTABLISHED;
681 pr_debug("switch node %d to state %s\n",
682 node->nodeid, dlm_state_str(node->state));
683 break;
684 default:
685 break;
686 }
687 spin_unlock(&node->state_lock);
688
689 break;
690 case DLM_VERSION_3_2:
691 break;
692 default:
693 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
694 DLM_VERSION_3_2, node->nodeid, node->version);
695 goto out;
696 }
697
698 switch (p->header.h_cmd) {
699 case DLM_RCOM:
700 /* these rcom message we use to determine version.
701 * they have their own retransmission handling and
702 * are the first messages of dlm.
703 *
704 * length already checked.
705 */
706 switch (p->rcom.rc_type) {
707 case cpu_to_le32(DLM_RCOM_NAMES):
708 fallthrough;
709 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
710 fallthrough;
711 case cpu_to_le32(DLM_RCOM_STATUS):
712 fallthrough;
713 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
714 break;
715 default:
716 log_print("unsupported rcom type received: %u, will skip this message from node %d",
717 le32_to_cpu(p->rcom.rc_type), nodeid);
718 goto out;
719 }
720
721 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
722 dlm_receive_buffer(p, nodeid);
723 break;
724 case DLM_OPTS:
725 seq = le32_to_cpu(p->header.u.h_seq);
726
727 ret = dlm_opts_check_msglen(p, msglen, nodeid);
728 if (ret < 0) {
729 log_print("opts msg too small: %u, will skip this message from node %d",
730 msglen, nodeid);
731 goto out;
732 }
733
734 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
735 le16_to_cpu(p->opts.o_optlen));
736
737 /* recheck inner msglen just if it's not garbage */
738 msglen = le16_to_cpu(p->header.h_length);
739 switch (p->header.h_cmd) {
740 case DLM_RCOM:
741 if (msglen < sizeof(struct dlm_rcom)) {
742 log_print("inner rcom msg too small: %u, will skip this message from node %d",
743 msglen, nodeid);
744 goto out;
745 }
746
747 break;
748 case DLM_MSG:
749 if (msglen < sizeof(struct dlm_message)) {
750 log_print("inner msg too small: %u, will skip this message from node %d",
751 msglen, nodeid);
752 goto out;
753 }
754
755 break;
756 case DLM_FIN:
757 if (msglen < sizeof(struct dlm_header)) {
758 log_print("inner fin too small: %u, will skip this message from node %d",
759 msglen, nodeid);
760 goto out;
761 }
762
763 break;
764 default:
765 log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
766 msglen, nodeid);
767 goto out;
768 }
769
770 dlm_midcomms_receive_buffer(p, node, seq);
771 break;
772 case DLM_ACK:
773 seq = le32_to_cpu(p->header.u.h_seq);
774 dlm_receive_ack(node, seq);
775 break;
776 default:
777 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
778 p->header.h_cmd, nodeid);
779 break;
780 }
781
782out:
783 srcu_read_unlock(&nodes_srcu, idx);
784}
785
786static void dlm_midcomms_receive_buffer_3_1(const union dlm_packet *p, int nodeid)
787{
788 uint16_t msglen = le16_to_cpu(p->header.h_length);
789 struct midcomms_node *node;
790 int idx;
791
792 idx = srcu_read_lock(&nodes_srcu);
793 node = nodeid2node(nodeid);
794 if (WARN_ON_ONCE(!node)) {
795 srcu_read_unlock(&nodes_srcu, idx);
796 return;
797 }
798
799 switch (node->version) {
800 case DLM_VERSION_NOT_SET:
801 node->version = DLM_VERSION_3_1;
802 wake_up(&node->shutdown_wait);
803 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
804 node->nodeid);
805 break;
806 case DLM_VERSION_3_1:
807 break;
808 default:
809 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
810 DLM_VERSION_3_1, node->nodeid, node->version);
811 srcu_read_unlock(&nodes_srcu, idx);
812 return;
813 }
814 srcu_read_unlock(&nodes_srcu, idx);
815
816 switch (p->header.h_cmd) {
817 case DLM_RCOM:
818 /* length already checked */
819 break;
820 case DLM_MSG:
821 if (msglen < sizeof(struct dlm_message)) {
822 log_print("msg too small: %u, will skip this message from node %d",
823 msglen, nodeid);
824 return;
825 }
826
827 break;
828 default:
829 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
830 p->header.h_cmd, nodeid);
831 return;
832 }
833
834 dlm_receive_buffer(p, nodeid);
835}
836
837int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
838{
839 const unsigned char *ptr = buf;
840 const struct dlm_header *hd;
841 uint16_t msglen;
842 int ret = 0;
843
844 while (len >= sizeof(struct dlm_header)) {
845 hd = (struct dlm_header *)ptr;
846
847 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
848 * less than dlm_header size.
849 *
850 * Some messages does not have a 8 byte length boundary yet
851 * which can occur in a unaligned memory access of some dlm
852 * messages. However this problem need to be fixed at the
853 * sending side, for now it seems nobody run into architecture
854 * related issues yet but it slows down some processing.
855 * Fixing this issue should be scheduled in future by doing
856 * the next major version bump.
857 */
858 msglen = le16_to_cpu(hd->h_length);
859 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
860 msglen < sizeof(struct dlm_header)) {
861 log_print("received invalid length header: %u from node %d, will abort message parsing",
862 msglen, nodeid);
863 return -EBADMSG;
864 }
865
866 /* caller will take care that leftover
867 * will be parsed next call with more data
868 */
869 if (msglen > len)
870 break;
871
872 ret += msglen;
873 len -= msglen;
874 ptr += msglen;
875 }
876
877 return ret;
878}
879
880/*
881 * Called from the low-level comms layer to process a buffer of
882 * commands.
883 */
884int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
885{
886 const unsigned char *ptr = buf;
887 const struct dlm_header *hd;
888 uint16_t msglen;
889 int ret = 0;
890
891 while (len >= sizeof(struct dlm_header)) {
892 hd = (struct dlm_header *)ptr;
893
894 msglen = le16_to_cpu(hd->h_length);
895 if (msglen > len)
896 break;
897
898 switch (hd->h_version) {
899 case cpu_to_le32(DLM_VERSION_3_1):
900 dlm_midcomms_receive_buffer_3_1((const union dlm_packet *)ptr, nodeid);
901 break;
902 case cpu_to_le32(DLM_VERSION_3_2):
903 dlm_midcomms_receive_buffer_3_2((const union dlm_packet *)ptr, nodeid);
904 break;
905 default:
906 log_print("received invalid version header: %u from node %d, will skip this message",
907 le32_to_cpu(hd->h_version), nodeid);
908 break;
909 }
910
911 ret += msglen;
912 len -= msglen;
913 ptr += msglen;
914 }
915
916 return ret;
917}
918
919void dlm_midcomms_unack_msg_resend(int nodeid)
920{
921 struct midcomms_node *node;
922 struct dlm_mhandle *mh;
923 int idx, ret;
924
925 idx = srcu_read_lock(&nodes_srcu);
926 node = nodeid2node(nodeid);
927 if (WARN_ON_ONCE(!node)) {
928 srcu_read_unlock(&nodes_srcu, idx);
929 return;
930 }
931
932 /* old protocol, we don't support to retransmit on failure */
933 switch (node->version) {
934 case DLM_VERSION_3_2:
935 break;
936 default:
937 srcu_read_unlock(&nodes_srcu, idx);
938 return;
939 }
940
941 rcu_read_lock();
942 list_for_each_entry_rcu(mh, &node->send_queue, list) {
943 if (!mh->committed)
944 continue;
945
946 ret = dlm_lowcomms_resend_msg(mh->msg);
947 if (!ret)
948 log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
949 mh->seq, node->nodeid);
950 }
951 rcu_read_unlock();
952 srcu_read_unlock(&nodes_srcu, idx);
953}
954
955static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
956 uint32_t seq)
957{
958 opts->o_header.h_cmd = DLM_OPTS;
959 opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
960 opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
961 opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
962 opts->o_header.u.h_seq = cpu_to_le32(seq);
963}
964
965static void midcomms_new_msg_cb(void *data)
966{
967 struct dlm_mhandle *mh = data;
968
969 atomic_inc(&mh->node->send_queue_cnt);
970
971 spin_lock_bh(&mh->node->send_queue_lock);
972 /* need to be locked with list_add_tail_rcu() because list is ordered */
973 mh->seq = atomic_fetch_inc(&mh->node->seq_send);
974 list_add_tail_rcu(&mh->list, &mh->node->send_queue);
975 spin_unlock_bh(&mh->node->send_queue_lock);
976}
977
978static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
979 int len, char **ppc)
980{
981 struct dlm_opts *opts;
982 struct dlm_msg *msg;
983
984 msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
985 ppc, midcomms_new_msg_cb, mh);
986 if (!msg)
987 return NULL;
988
989 opts = (struct dlm_opts *)*ppc;
990 mh->opts = opts;
991
992 /* add possible options here */
993 dlm_fill_opts_header(opts, len, mh->seq);
994
995 *ppc += sizeof(*opts);
996 mh->inner_p = (const union dlm_packet *)*ppc;
997 return msg;
998}
999
1000/* avoid false positive for nodes_srcu, unlock happens in
1001 * dlm_midcomms_commit_mhandle which is a must call if success
1002 */
1003#ifndef __CHECKER__
1004struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len, char **ppc)
1005{
1006 struct midcomms_node *node;
1007 struct dlm_mhandle *mh;
1008 struct dlm_msg *msg;
1009 int idx;
1010
1011 idx = srcu_read_lock(&nodes_srcu);
1012 node = nodeid2node(nodeid);
1013 if (WARN_ON_ONCE(!node))
1014 goto err;
1015
1016 /* this is a bug, however we going on and hope it will be resolved */
1017 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1018
1019 mh = dlm_allocate_mhandle();
1020 if (!mh)
1021 goto err;
1022
1023 mh->committed = false;
1024 mh->ack_rcv = NULL;
1025 mh->idx = idx;
1026 mh->node = node;
1027
1028 switch (node->version) {
1029 case DLM_VERSION_3_1:
1030 msg = dlm_lowcomms_new_msg(nodeid, len, ppc, NULL, NULL);
1031 if (!msg) {
1032 dlm_free_mhandle(mh);
1033 goto err;
1034 }
1035
1036 break;
1037 case DLM_VERSION_3_2:
1038 /* send ack back if necessary */
1039 dlm_send_ack_threshold(node, DLM_SEND_ACK_BACK_MSG_THRESHOLD);
1040
1041 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, ppc);
1042 if (!msg) {
1043 dlm_free_mhandle(mh);
1044 goto err;
1045 }
1046 break;
1047 default:
1048 dlm_free_mhandle(mh);
1049 WARN_ON_ONCE(1);
1050 goto err;
1051 }
1052
1053 mh->msg = msg;
1054
1055 /* keep in mind that is a must to call
1056 * dlm_midcomms_commit_msg() which releases
1057 * nodes_srcu using mh->idx which is assumed
1058 * here that the application will call it.
1059 */
1060 return mh;
1061
1062err:
1063 srcu_read_unlock(&nodes_srcu, idx);
1064 return NULL;
1065}
1066#endif
1067
1068static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1069 const void *name, int namelen)
1070{
1071 switch (mh->inner_p->header.h_cmd) {
1072 case DLM_MSG:
1073 trace_dlm_send_message(mh->node->nodeid, mh->seq,
1074 &mh->inner_p->message,
1075 name, namelen);
1076 break;
1077 case DLM_RCOM:
1078 trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1079 &mh->inner_p->rcom);
1080 break;
1081 default:
1082 /* nothing to trace */
1083 break;
1084 }
1085}
1086
1087static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1088 const void *name, int namelen)
1089{
1090 /* nexthdr chain for fast lookup */
1091 mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1092 mh->committed = true;
1093 dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1094 dlm_lowcomms_commit_msg(mh->msg);
1095}
1096
1097/* avoid false positive for nodes_srcu, lock was happen in
1098 * dlm_midcomms_get_mhandle
1099 */
1100#ifndef __CHECKER__
1101void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1102 const void *name, int namelen)
1103{
1104
1105 switch (mh->node->version) {
1106 case DLM_VERSION_3_1:
1107 srcu_read_unlock(&nodes_srcu, mh->idx);
1108
1109 dlm_lowcomms_commit_msg(mh->msg);
1110 dlm_lowcomms_put_msg(mh->msg);
1111 /* mh is not part of rcu list in this case */
1112 dlm_free_mhandle(mh);
1113 break;
1114 case DLM_VERSION_3_2:
1115 /* held rcu read lock here, because we sending the
1116 * dlm message out, when we do that we could receive
1117 * an ack back which releases the mhandle and we
1118 * get a use after free.
1119 */
1120 rcu_read_lock();
1121 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1122 srcu_read_unlock(&nodes_srcu, mh->idx);
1123 rcu_read_unlock();
1124 break;
1125 default:
1126 srcu_read_unlock(&nodes_srcu, mh->idx);
1127 WARN_ON_ONCE(1);
1128 break;
1129 }
1130}
1131#endif
1132
1133int dlm_midcomms_start(void)
1134{
1135 return dlm_lowcomms_start();
1136}
1137
1138void dlm_midcomms_stop(void)
1139{
1140 dlm_lowcomms_stop();
1141}
1142
1143void dlm_midcomms_init(void)
1144{
1145 int i;
1146
1147 for (i = 0; i < CONN_HASH_SIZE; i++)
1148 INIT_HLIST_HEAD(&node_hash[i]);
1149
1150 dlm_lowcomms_init();
1151}
1152
1153static void midcomms_node_release(struct rcu_head *rcu)
1154{
1155 struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1156
1157 WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1158 dlm_send_queue_flush(node);
1159 kfree(node);
1160}
1161
1162void dlm_midcomms_exit(void)
1163{
1164 struct midcomms_node *node;
1165 int i, idx;
1166
1167 idx = srcu_read_lock(&nodes_srcu);
1168 for (i = 0; i < CONN_HASH_SIZE; i++) {
1169 hlist_for_each_entry_srcu(node, &node_hash[i], hlist,
1170 srcu_read_lock_held(&nodes_srcu)) {
1171 dlm_delete_debug_comms_file(node->debugfs);
1172
1173 spin_lock(&nodes_lock);
1174 hlist_del_rcu(&node->hlist);
1175 spin_unlock(&nodes_lock);
1176
1177 call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1178 }
1179 }
1180 srcu_read_unlock(&nodes_srcu, idx);
1181 srcu_barrier(&nodes_srcu);
1182
1183 dlm_lowcomms_exit();
1184}
1185
1186static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1187{
1188 spin_lock_bh(&node->state_lock);
1189 pr_debug("receive active fin ack from node %d with state %s\n",
1190 node->nodeid, dlm_state_str(node->state));
1191
1192 switch (node->state) {
1193 case DLM_FIN_WAIT1:
1194 node->state = DLM_FIN_WAIT2;
1195 pr_debug("switch node %d to state %s\n",
1196 node->nodeid, dlm_state_str(node->state));
1197 break;
1198 case DLM_CLOSING:
1199 midcomms_node_reset(node);
1200 pr_debug("switch node %d to state %s\n",
1201 node->nodeid, dlm_state_str(node->state));
1202 break;
1203 case DLM_CLOSED:
1204 /* not valid but somehow we got what we want */
1205 wake_up(&node->shutdown_wait);
1206 break;
1207 default:
1208 spin_unlock_bh(&node->state_lock);
1209 log_print("%s: unexpected state: %d",
1210 __func__, node->state);
1211 WARN_ON_ONCE(1);
1212 return;
1213 }
1214 spin_unlock_bh(&node->state_lock);
1215}
1216
1217void dlm_midcomms_add_member(int nodeid)
1218{
1219 struct midcomms_node *node;
1220 int idx;
1221
1222 idx = srcu_read_lock(&nodes_srcu);
1223 node = nodeid2node(nodeid);
1224 if (WARN_ON_ONCE(!node)) {
1225 srcu_read_unlock(&nodes_srcu, idx);
1226 return;
1227 }
1228
1229 spin_lock_bh(&node->state_lock);
1230 if (!node->users) {
1231 pr_debug("receive add member from node %d with state %s\n",
1232 node->nodeid, dlm_state_str(node->state));
1233 switch (node->state) {
1234 case DLM_ESTABLISHED:
1235 break;
1236 case DLM_CLOSED:
1237 node->state = DLM_ESTABLISHED;
1238 pr_debug("switch node %d to state %s\n",
1239 node->nodeid, dlm_state_str(node->state));
1240 break;
1241 default:
1242 /* some invalid state passive shutdown
1243 * was failed, we try to reset and
1244 * hope it will go on.
1245 */
1246 log_print("reset node %d because shutdown stuck",
1247 node->nodeid);
1248
1249 midcomms_node_reset(node);
1250 node->state = DLM_ESTABLISHED;
1251 break;
1252 }
1253 }
1254
1255 node->users++;
1256 pr_debug("node %d users inc count %d\n", nodeid, node->users);
1257 spin_unlock_bh(&node->state_lock);
1258
1259 srcu_read_unlock(&nodes_srcu, idx);
1260}
1261
1262void dlm_midcomms_remove_member(int nodeid)
1263{
1264 struct midcomms_node *node;
1265 int idx;
1266
1267 idx = srcu_read_lock(&nodes_srcu);
1268 node = nodeid2node(nodeid);
1269 /* in case of dlm_midcomms_close() removes node */
1270 if (!node) {
1271 srcu_read_unlock(&nodes_srcu, idx);
1272 return;
1273 }
1274
1275 spin_lock_bh(&node->state_lock);
1276 /* case of dlm_midcomms_addr() created node but
1277 * was not added before because dlm_midcomms_close()
1278 * removed the node
1279 */
1280 if (!node->users) {
1281 spin_unlock_bh(&node->state_lock);
1282 srcu_read_unlock(&nodes_srcu, idx);
1283 return;
1284 }
1285
1286 node->users--;
1287 pr_debug("node %d users dec count %d\n", nodeid, node->users);
1288
1289 /* hitting users count to zero means the
1290 * other side is running dlm_midcomms_stop()
1291 * we meet us to have a clean disconnect.
1292 */
1293 if (node->users == 0) {
1294 pr_debug("receive remove member from node %d with state %s\n",
1295 node->nodeid, dlm_state_str(node->state));
1296 switch (node->state) {
1297 case DLM_ESTABLISHED:
1298 break;
1299 case DLM_CLOSE_WAIT:
1300 /* passive shutdown DLM_LAST_ACK case 2 */
1301 node->state = DLM_LAST_ACK;
1302 pr_debug("switch node %d to state %s case 2\n",
1303 node->nodeid, dlm_state_str(node->state));
1304 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1305 dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1306 break;
1307 case DLM_LAST_ACK:
1308 /* probably receive fin caught it, do nothing */
1309 break;
1310 case DLM_CLOSED:
1311 /* already gone, do nothing */
1312 break;
1313 default:
1314 log_print("%s: unexpected state: %d",
1315 __func__, node->state);
1316 break;
1317 }
1318 }
1319 spin_unlock_bh(&node->state_lock);
1320
1321 srcu_read_unlock(&nodes_srcu, idx);
1322}
1323
1324void dlm_midcomms_version_wait(void)
1325{
1326 struct midcomms_node *node;
1327 int i, idx, ret;
1328
1329 idx = srcu_read_lock(&nodes_srcu);
1330 for (i = 0; i < CONN_HASH_SIZE; i++) {
1331 hlist_for_each_entry_srcu(node, &node_hash[i], hlist,
1332 srcu_read_lock_held(&nodes_srcu)) {
1333 ret = wait_event_timeout(node->shutdown_wait,
1334 node->version != DLM_VERSION_NOT_SET ||
1335 node->state == DLM_CLOSED ||
1336 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1337 DLM_SHUTDOWN_TIMEOUT);
1338 if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1339 pr_debug("version wait timed out for node %d with state %s\n",
1340 node->nodeid, dlm_state_str(node->state));
1341 }
1342 }
1343 srcu_read_unlock(&nodes_srcu, idx);
1344}
1345
1346static void midcomms_shutdown(struct midcomms_node *node)
1347{
1348 int ret;
1349
1350 /* old protocol, we don't wait for pending operations */
1351 switch (node->version) {
1352 case DLM_VERSION_3_2:
1353 break;
1354 default:
1355 return;
1356 }
1357
1358 spin_lock_bh(&node->state_lock);
1359 pr_debug("receive active shutdown for node %d with state %s\n",
1360 node->nodeid, dlm_state_str(node->state));
1361 switch (node->state) {
1362 case DLM_ESTABLISHED:
1363 node->state = DLM_FIN_WAIT1;
1364 pr_debug("switch node %d to state %s case 2\n",
1365 node->nodeid, dlm_state_str(node->state));
1366 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1367 break;
1368 case DLM_CLOSED:
1369 /* we have what we want */
1370 break;
1371 default:
1372 /* busy to enter DLM_FIN_WAIT1, wait until passive
1373 * done in shutdown_wait to enter DLM_CLOSED.
1374 */
1375 break;
1376 }
1377 spin_unlock_bh(&node->state_lock);
1378
1379 if (DLM_DEBUG_FENCE_TERMINATION)
1380 msleep(5000);
1381
1382 /* wait for other side dlm + fin */
1383 ret = wait_event_timeout(node->shutdown_wait,
1384 node->state == DLM_CLOSED ||
1385 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1386 DLM_SHUTDOWN_TIMEOUT);
1387 if (!ret)
1388 pr_debug("active shutdown timed out for node %d with state %s\n",
1389 node->nodeid, dlm_state_str(node->state));
1390 else
1391 pr_debug("active shutdown done for node %d with state %s\n",
1392 node->nodeid, dlm_state_str(node->state));
1393}
1394
1395void dlm_midcomms_shutdown(void)
1396{
1397 struct midcomms_node *node;
1398 int i, idx;
1399
1400 mutex_lock(&close_lock);
1401 idx = srcu_read_lock(&nodes_srcu);
1402 for (i = 0; i < CONN_HASH_SIZE; i++) {
1403 hlist_for_each_entry_srcu(node, &node_hash[i], hlist,
1404 srcu_read_lock_held(&nodes_srcu)) {
1405 midcomms_shutdown(node);
1406 }
1407 }
1408
1409 dlm_lowcomms_shutdown();
1410
1411 for (i = 0; i < CONN_HASH_SIZE; i++) {
1412 hlist_for_each_entry_srcu(node, &node_hash[i], hlist,
1413 srcu_read_lock_held(&nodes_srcu)) {
1414 midcomms_node_reset(node);
1415 }
1416 }
1417 srcu_read_unlock(&nodes_srcu, idx);
1418 mutex_unlock(&close_lock);
1419}
1420
1421int dlm_midcomms_close(int nodeid)
1422{
1423 struct midcomms_node *node;
1424 int idx, ret;
1425
1426 idx = srcu_read_lock(&nodes_srcu);
1427 /* Abort pending close/remove operation */
1428 node = nodeid2node(nodeid);
1429 if (node) {
1430 /* let shutdown waiters leave */
1431 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1432 wake_up(&node->shutdown_wait);
1433 }
1434 srcu_read_unlock(&nodes_srcu, idx);
1435
1436 synchronize_srcu(&nodes_srcu);
1437
1438 mutex_lock(&close_lock);
1439 idx = srcu_read_lock(&nodes_srcu);
1440 node = nodeid2node(nodeid);
1441 if (!node) {
1442 srcu_read_unlock(&nodes_srcu, idx);
1443 mutex_unlock(&close_lock);
1444 return dlm_lowcomms_close(nodeid);
1445 }
1446
1447 ret = dlm_lowcomms_close(nodeid);
1448 dlm_delete_debug_comms_file(node->debugfs);
1449
1450 spin_lock_bh(&nodes_lock);
1451 hlist_del_rcu(&node->hlist);
1452 spin_unlock_bh(&nodes_lock);
1453 srcu_read_unlock(&nodes_srcu, idx);
1454
1455 /* wait that all readers left until flush send queue */
1456 synchronize_srcu(&nodes_srcu);
1457
1458 /* drop all pending dlm messages, this is fine as
1459 * this function get called when the node is fenced
1460 */
1461 dlm_send_queue_flush(node);
1462
1463 call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1464 mutex_unlock(&close_lock);
1465
1466 return ret;
1467}
1468
1469/* debug functionality to send raw dlm msg from user space */
1470struct dlm_rawmsg_data {
1471 struct midcomms_node *node;
1472 void *buf;
1473};
1474
1475static void midcomms_new_rawmsg_cb(void *data)
1476{
1477 struct dlm_rawmsg_data *rd = data;
1478 struct dlm_header *h = rd->buf;
1479
1480 switch (h->h_version) {
1481 case cpu_to_le32(DLM_VERSION_3_1):
1482 break;
1483 default:
1484 switch (h->h_cmd) {
1485 case DLM_OPTS:
1486 if (!h->u.h_seq)
1487 h->u.h_seq = cpu_to_le32(atomic_fetch_inc(&rd->node->seq_send));
1488 break;
1489 default:
1490 break;
1491 }
1492 break;
1493 }
1494}
1495
1496int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1497 int buflen)
1498{
1499 struct dlm_rawmsg_data rd;
1500 struct dlm_msg *msg;
1501 char *msgbuf;
1502
1503 rd.node = node;
1504 rd.buf = buf;
1505
1506 msg = dlm_lowcomms_new_msg(node->nodeid, buflen, &msgbuf,
1507 midcomms_new_rawmsg_cb, &rd);
1508 if (!msg)
1509 return -ENOMEM;
1510
1511 memcpy(msgbuf, buf, buflen);
1512 dlm_lowcomms_commit_msg(msg);
1513 return 0;
1514}
1515