1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
6 ** Copyright (C) 2004-2021 Red Hat, Inc. All rights reserved.
9 *******************************************************************************
10 ******************************************************************************/
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.
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.
27 * How version detection works:
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.
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:
54 * 1. The cluster member was removed and we received a FIN
56 * 2. We received a FIN but the member was not removed yet
58 * One of these cases will do the CLOSE_WAIT to LAST_ACK change.
64 * | add member/receive RCOM version
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 -------------- | -------------- |
85 * \ snd ACK +---------+ +---------+
86 * ------------------------>| CLOSED | | CLOSED |
87 * +---------+ +---------+
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.
94 * Future improvements:
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.
100 * Unaligned memory access:
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.
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.
112 * Tail Size checking:
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.
120 * At timeout places or weird sequence number behaviours we should send
121 * a fencing request to the cluster manager.
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.
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.
133 #define DLM_DEBUG_FENCE_TERMINATION 0
135 #include <trace/events/dlm.h>
138 #include "dlm_internal.h"
139 #include "lockspace.h"
140 #include "lowcomms.h"
145 #include "midcomms.h"
147 /* init value for sequence numbers for testing purpose only e.g. overflows */
148 #define DLM_SEQ_INIT 0
149 /* 5 seconds wait to sync ending of dlm */
150 #define DLM_SHUTDOWN_TIMEOUT msecs_to_jiffies(5000)
151 #define DLM_VERSION_NOT_SET 0
153 struct midcomms_node {
158 /* These queues are unbound because we cannot drop any message in dlm.
159 * We could send a fence signal for a specific node to the cluster
160 * manager if queues hits some maximum value, however this handling
163 struct list_head send_queue;
164 spinlock_t send_queue_lock;
165 atomic_t send_queue_cnt;
166 #define DLM_NODE_FLAG_CLOSE 1
167 #define DLM_NODE_FLAG_STOP_TX 2
168 #define DLM_NODE_FLAG_STOP_RX 3
169 #define DLM_NODE_ULP_DELIVERED 4
171 wait_queue_head_t shutdown_wait;
173 /* dlm tcp termination state */
175 #define DLM_ESTABLISHED 2
176 #define DLM_FIN_WAIT1 3
177 #define DLM_FIN_WAIT2 4
178 #define DLM_CLOSE_WAIT 5
179 #define DLM_LAST_ACK 6
180 #define DLM_CLOSING 7
182 spinlock_t state_lock;
184 /* counts how many lockspaces are using this node
185 * this refcount is necessary to determine if the
186 * node wants to disconnect.
190 /* not protected by srcu, node_hash lifetime */
193 struct hlist_node hlist;
198 const union dlm_packet *inner_p;
199 struct midcomms_node *node;
200 struct dlm_opts *opts;
205 void (*ack_rcv)(struct midcomms_node *node);
207 /* get_mhandle/commit srcu idx exchange */
210 struct list_head list;
214 static struct hlist_head node_hash[CONN_HASH_SIZE];
215 static DEFINE_SPINLOCK(nodes_lock);
216 DEFINE_STATIC_SRCU(nodes_srcu);
218 /* This mutex prevents that midcomms_close() is running while
219 * stop() or remove(). As I experienced invalid memory access
220 * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
221 * resetting machines. I will end in some double deletion in nodes
224 static DEFINE_MUTEX(close_lock);
226 struct kmem_cache *dlm_midcomms_cache_create(void)
228 return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
232 static inline const char *dlm_state_str(int state)
237 case DLM_ESTABLISHED:
238 return "ESTABLISHED";
254 const char *dlm_midcomms_state(struct midcomms_node *node)
256 return dlm_state_str(node->state);
259 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
264 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
266 return atomic_read(&node->send_queue_cnt);
269 uint32_t dlm_midcomms_version(struct midcomms_node *node)
271 return node->version;
274 static struct midcomms_node *__find_node(int nodeid, int r)
276 struct midcomms_node *node;
278 hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
279 if (node->nodeid == nodeid)
286 static void dlm_mhandle_release(struct rcu_head *rcu)
288 struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
290 dlm_lowcomms_put_msg(mh->msg);
291 dlm_free_mhandle(mh);
294 static void dlm_mhandle_delete(struct midcomms_node *node,
295 struct dlm_mhandle *mh)
297 list_del_rcu(&mh->list);
298 atomic_dec(&node->send_queue_cnt);
299 call_rcu(&mh->rcu, dlm_mhandle_release);
302 static void dlm_send_queue_flush(struct midcomms_node *node)
304 struct dlm_mhandle *mh;
306 pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
309 spin_lock_bh(&node->send_queue_lock);
310 list_for_each_entry_rcu(mh, &node->send_queue, list) {
311 dlm_mhandle_delete(node, mh);
313 spin_unlock_bh(&node->send_queue_lock);
317 static void midcomms_node_reset(struct midcomms_node *node)
319 pr_debug("reset node %d\n", node->nodeid);
321 node->seq_next = DLM_SEQ_INIT;
322 node->seq_send = DLM_SEQ_INIT;
323 node->version = DLM_VERSION_NOT_SET;
326 dlm_send_queue_flush(node);
327 node->state = DLM_CLOSED;
328 wake_up(&node->shutdown_wait);
331 static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
333 struct midcomms_node *node, *tmp;
334 int r = nodeid_hash(nodeid);
336 node = __find_node(nodeid, r);
340 node = kmalloc(sizeof(*node), alloc);
344 node->nodeid = nodeid;
345 spin_lock_init(&node->state_lock);
346 spin_lock_init(&node->send_queue_lock);
347 atomic_set(&node->send_queue_cnt, 0);
348 INIT_LIST_HEAD(&node->send_queue);
349 init_waitqueue_head(&node->shutdown_wait);
351 midcomms_node_reset(node);
353 spin_lock(&nodes_lock);
354 /* check again if there was somebody else
355 * earlier here to add the node
357 tmp = __find_node(nodeid, r);
359 spin_unlock(&nodes_lock);
364 hlist_add_head_rcu(&node->hlist, &node_hash[r]);
365 spin_unlock(&nodes_lock);
367 node->debugfs = dlm_create_debug_comms_file(nodeid, node);
371 static int dlm_send_ack(int nodeid, uint32_t seq)
373 int mb_len = sizeof(struct dlm_header);
374 struct dlm_header *m_header;
378 msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
383 m_header = (struct dlm_header *)ppc;
385 m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
386 m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
387 m_header->h_length = cpu_to_le16(mb_len);
388 m_header->h_cmd = DLM_ACK;
389 m_header->u.h_seq = cpu_to_le32(seq);
391 dlm_lowcomms_commit_msg(msg);
392 dlm_lowcomms_put_msg(msg);
397 static int dlm_send_fin(struct midcomms_node *node,
398 void (*ack_rcv)(struct midcomms_node *node))
400 int mb_len = sizeof(struct dlm_header);
401 struct dlm_header *m_header;
402 struct dlm_mhandle *mh;
405 mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
409 set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
410 mh->ack_rcv = ack_rcv;
412 m_header = (struct dlm_header *)ppc;
414 m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
415 m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
416 m_header->h_length = cpu_to_le16(mb_len);
417 m_header->h_cmd = DLM_FIN;
419 pr_debug("sending fin msg to node %d\n", node->nodeid);
420 dlm_midcomms_commit_mhandle(mh, NULL, 0);
425 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
427 struct dlm_mhandle *mh;
430 list_for_each_entry_rcu(mh, &node->send_queue, list) {
431 if (before(mh->seq, seq)) {
435 /* send queue should be ordered */
440 spin_lock_bh(&node->send_queue_lock);
441 list_for_each_entry_rcu(mh, &node->send_queue, list) {
442 if (before(mh->seq, seq)) {
443 dlm_mhandle_delete(node, mh);
445 /* send queue should be ordered */
449 spin_unlock_bh(&node->send_queue_lock);
453 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
455 spin_lock(&node->state_lock);
456 pr_debug("receive passive fin ack from node %d with state %s\n",
457 node->nodeid, dlm_state_str(node->state));
459 switch (node->state) {
462 midcomms_node_reset(node);
465 /* not valid but somehow we got what we want */
466 wake_up(&node->shutdown_wait);
469 spin_unlock(&node->state_lock);
470 log_print("%s: unexpected state: %d",
471 __func__, node->state);
475 spin_unlock(&node->state_lock);
478 static void dlm_receive_buffer_3_2_trace(uint32_t seq, union dlm_packet *p)
480 switch (p->header.h_cmd) {
482 trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
485 trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
492 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
493 struct midcomms_node *node,
496 if (seq == node->seq_next) {
499 switch (p->header.h_cmd) {
501 spin_lock(&node->state_lock);
502 pr_debug("receive fin msg from node %d with state %s\n",
503 node->nodeid, dlm_state_str(node->state));
505 switch (node->state) {
506 case DLM_ESTABLISHED:
507 dlm_send_ack(node->nodeid, node->seq_next);
509 /* passive shutdown DLM_LAST_ACK case 1
510 * additional we check if the node is used by
511 * cluster manager events at all.
513 if (node->users == 0) {
514 node->state = DLM_LAST_ACK;
515 pr_debug("switch node %d to state %s case 1\n",
516 node->nodeid, dlm_state_str(node->state));
517 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
518 dlm_send_fin(node, dlm_pas_fin_ack_rcv);
520 node->state = DLM_CLOSE_WAIT;
521 pr_debug("switch node %d to state %s\n",
522 node->nodeid, dlm_state_str(node->state));
526 dlm_send_ack(node->nodeid, node->seq_next);
527 node->state = DLM_CLOSING;
528 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
529 pr_debug("switch node %d to state %s\n",
530 node->nodeid, dlm_state_str(node->state));
533 dlm_send_ack(node->nodeid, node->seq_next);
534 midcomms_node_reset(node);
535 pr_debug("switch node %d to state %s\n",
536 node->nodeid, dlm_state_str(node->state));
539 /* probably remove_member caught it, do nothing */
542 spin_unlock(&node->state_lock);
543 log_print("%s: unexpected state: %d",
544 __func__, node->state);
548 spin_unlock(&node->state_lock);
551 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
552 dlm_receive_buffer_3_2_trace(seq, p);
553 dlm_receive_buffer(p, node->nodeid);
554 set_bit(DLM_NODE_ULP_DELIVERED, &node->flags);
558 /* retry to ack message which we already have by sending back
559 * current node->seq_next number as ack.
561 if (seq < node->seq_next)
562 dlm_send_ack(node->nodeid, node->seq_next);
564 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
565 seq, node->seq_next, node->nodeid);
569 static struct midcomms_node *
570 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
571 uint16_t msglen, int (*cb)(struct midcomms_node *node))
573 struct midcomms_node *node = NULL;
574 gfp_t allocation = 0;
577 switch (p->header.h_cmd) {
579 if (msglen < sizeof(struct dlm_rcom)) {
580 log_print("rcom msg too small: %u, will skip this message from node %d",
585 switch (p->rcom.rc_type) {
586 case cpu_to_le32(DLM_RCOM_NAMES):
588 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
590 case cpu_to_le32(DLM_RCOM_STATUS):
592 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
593 node = nodeid2node(nodeid, 0);
595 spin_lock(&node->state_lock);
596 if (node->state != DLM_ESTABLISHED)
597 pr_debug("receive begin RCOM msg from node %d with state %s\n",
598 node->nodeid, dlm_state_str(node->state));
600 switch (node->state) {
602 node->state = DLM_ESTABLISHED;
603 pr_debug("switch node %d to state %s\n",
604 node->nodeid, dlm_state_str(node->state));
606 case DLM_ESTABLISHED:
609 spin_unlock(&node->state_lock);
612 spin_unlock(&node->state_lock);
615 allocation = GFP_NOFS;
626 node = nodeid2node(nodeid, allocation);
628 switch (p->header.h_cmd) {
630 if (msglen < sizeof(struct dlm_opts)) {
631 log_print("opts msg too small: %u, will skip this message from node %d",
636 log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
637 p->opts.o_nextcmd, nodeid);
640 log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
641 p->header.h_cmd, nodeid);
655 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
657 switch (node->version) {
658 case DLM_VERSION_NOT_SET:
659 node->version = DLM_VERSION_3_2;
660 wake_up(&node->shutdown_wait);
661 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
664 case DLM_VERSION_3_2:
667 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
668 DLM_VERSION_3_2, node->nodeid, node->version);
675 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
679 /* we only trust outer header msglen because
680 * it's checked against receive buffer length.
682 if (len < sizeof(struct dlm_opts))
684 len -= sizeof(struct dlm_opts);
686 if (len < le16_to_cpu(p->opts.o_optlen))
688 len -= le16_to_cpu(p->opts.o_optlen);
690 switch (p->opts.o_nextcmd) {
692 if (len < sizeof(struct dlm_header)) {
693 log_print("fin too small: %d, will skip this message from node %d",
700 if (len < sizeof(struct dlm_message)) {
701 log_print("msg too small: %d, will skip this message from node %d",
708 if (len < sizeof(struct dlm_rcom)) {
709 log_print("rcom msg too small: %d, will skip this message from node %d",
716 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
717 p->opts.o_nextcmd, nodeid);
724 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
726 uint16_t msglen = le16_to_cpu(p->header.h_length);
727 struct midcomms_node *node;
731 idx = srcu_read_lock(&nodes_srcu);
732 node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
733 dlm_midcomms_version_check_3_2);
737 switch (p->header.h_cmd) {
739 /* these rcom message we use to determine version.
740 * they have their own retransmission handling and
741 * are the first messages of dlm.
743 * length already checked.
745 switch (p->rcom.rc_type) {
746 case cpu_to_le32(DLM_RCOM_NAMES):
748 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
750 case cpu_to_le32(DLM_RCOM_STATUS):
752 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
755 log_print("unsupported rcom type received: %u, will skip this message from node %d",
756 le32_to_cpu(p->rcom.rc_type), nodeid);
760 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
761 dlm_receive_buffer(p, nodeid);
764 seq = le32_to_cpu(p->header.u.h_seq);
766 ret = dlm_opts_check_msglen(p, msglen, nodeid);
768 log_print("opts msg too small: %u, will skip this message from node %d",
773 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
774 le16_to_cpu(p->opts.o_optlen));
776 /* recheck inner msglen just if it's not garbage */
777 msglen = le16_to_cpu(p->header.h_length);
778 switch (p->header.h_cmd) {
780 if (msglen < sizeof(struct dlm_rcom)) {
781 log_print("inner rcom msg too small: %u, will skip this message from node %d",
788 if (msglen < sizeof(struct dlm_message)) {
789 log_print("inner msg too small: %u, will skip this message from node %d",
796 if (msglen < sizeof(struct dlm_header)) {
797 log_print("inner fin too small: %u, will skip this message from node %d",
804 log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
809 dlm_midcomms_receive_buffer(p, node, seq);
812 seq = le32_to_cpu(p->header.u.h_seq);
813 dlm_receive_ack(node, seq);
816 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
817 p->header.h_cmd, nodeid);
822 srcu_read_unlock(&nodes_srcu, idx);
825 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
827 switch (node->version) {
828 case DLM_VERSION_NOT_SET:
829 node->version = DLM_VERSION_3_1;
830 wake_up(&node->shutdown_wait);
831 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
834 case DLM_VERSION_3_1:
837 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
838 DLM_VERSION_3_1, node->nodeid, node->version);
845 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
847 uint16_t msglen = le16_to_cpu(p->header.h_length);
848 struct midcomms_node *node;
851 idx = srcu_read_lock(&nodes_srcu);
852 node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
853 dlm_midcomms_version_check_3_1);
855 srcu_read_unlock(&nodes_srcu, idx);
858 srcu_read_unlock(&nodes_srcu, idx);
860 switch (p->header.h_cmd) {
862 /* length already checked */
865 if (msglen < sizeof(struct dlm_message)) {
866 log_print("msg too small: %u, will skip this message from node %d",
873 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
874 p->header.h_cmd, nodeid);
878 dlm_receive_buffer(p, nodeid);
881 int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
883 const unsigned char *ptr = buf;
884 const struct dlm_header *hd;
888 while (len >= sizeof(struct dlm_header)) {
889 hd = (struct dlm_header *)ptr;
891 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
892 * less than dlm_header size.
894 * Some messages does not have a 8 byte length boundary yet
895 * which can occur in a unaligned memory access of some dlm
896 * messages. However this problem need to be fixed at the
897 * sending side, for now it seems nobody run into architecture
898 * related issues yet but it slows down some processing.
899 * Fixing this issue should be scheduled in future by doing
900 * the next major version bump.
902 msglen = le16_to_cpu(hd->h_length);
903 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
904 msglen < sizeof(struct dlm_header)) {
905 log_print("received invalid length header: %u from node %d, will abort message parsing",
910 /* caller will take care that leftover
911 * will be parsed next call with more data
925 * Called from the low-level comms layer to process a buffer of
928 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
930 const unsigned char *ptr = buf;
931 const struct dlm_header *hd;
935 while (len >= sizeof(struct dlm_header)) {
936 hd = (struct dlm_header *)ptr;
938 msglen = le16_to_cpu(hd->h_length);
942 switch (hd->h_version) {
943 case cpu_to_le32(DLM_VERSION_3_1):
944 dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
946 case cpu_to_le32(DLM_VERSION_3_2):
947 dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
950 log_print("received invalid version header: %u from node %d, will skip this message",
951 le32_to_cpu(hd->h_version), nodeid);
963 void dlm_midcomms_receive_done(int nodeid)
965 struct midcomms_node *node;
968 idx = srcu_read_lock(&nodes_srcu);
969 node = nodeid2node(nodeid, 0);
971 srcu_read_unlock(&nodes_srcu, idx);
975 /* old protocol, we do nothing */
976 switch (node->version) {
977 case DLM_VERSION_3_2:
980 srcu_read_unlock(&nodes_srcu, idx);
984 /* do nothing if we didn't delivered stateful to ulp */
985 if (!test_and_clear_bit(DLM_NODE_ULP_DELIVERED,
987 srcu_read_unlock(&nodes_srcu, idx);
991 spin_lock(&node->state_lock);
992 /* we only ack if state is ESTABLISHED */
993 switch (node->state) {
994 case DLM_ESTABLISHED:
995 spin_unlock(&node->state_lock);
996 dlm_send_ack(node->nodeid, node->seq_next);
999 spin_unlock(&node->state_lock);
1000 /* do nothing FIN has it's own ack send */
1003 srcu_read_unlock(&nodes_srcu, idx);
1006 void dlm_midcomms_unack_msg_resend(int nodeid)
1008 struct midcomms_node *node;
1009 struct dlm_mhandle *mh;
1012 idx = srcu_read_lock(&nodes_srcu);
1013 node = nodeid2node(nodeid, 0);
1015 srcu_read_unlock(&nodes_srcu, idx);
1019 /* old protocol, we don't support to retransmit on failure */
1020 switch (node->version) {
1021 case DLM_VERSION_3_2:
1024 srcu_read_unlock(&nodes_srcu, idx);
1029 list_for_each_entry_rcu(mh, &node->send_queue, list) {
1033 ret = dlm_lowcomms_resend_msg(mh->msg);
1035 log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1036 mh->seq, node->nodeid);
1039 srcu_read_unlock(&nodes_srcu, idx);
1042 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1045 opts->o_header.h_cmd = DLM_OPTS;
1046 opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1047 opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1048 opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1049 opts->o_header.u.h_seq = cpu_to_le32(seq);
1052 static void midcomms_new_msg_cb(void *data)
1054 struct dlm_mhandle *mh = data;
1056 atomic_inc(&mh->node->send_queue_cnt);
1058 spin_lock_bh(&mh->node->send_queue_lock);
1059 list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1060 spin_unlock_bh(&mh->node->send_queue_lock);
1062 mh->seq = mh->node->seq_send++;
1065 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1066 int len, gfp_t allocation, char **ppc)
1068 struct dlm_opts *opts;
1069 struct dlm_msg *msg;
1071 msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1072 allocation, ppc, midcomms_new_msg_cb, mh);
1076 opts = (struct dlm_opts *)*ppc;
1079 /* add possible options here */
1080 dlm_fill_opts_header(opts, len, mh->seq);
1082 *ppc += sizeof(*opts);
1083 mh->inner_p = (const union dlm_packet *)*ppc;
1087 /* avoid false positive for nodes_srcu, unlock happens in
1088 * dlm_midcomms_commit_mhandle which is a must call if success
1091 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1092 gfp_t allocation, char **ppc)
1094 struct midcomms_node *node;
1095 struct dlm_mhandle *mh;
1096 struct dlm_msg *msg;
1099 idx = srcu_read_lock(&nodes_srcu);
1100 node = nodeid2node(nodeid, 0);
1106 /* this is a bug, however we going on and hope it will be resolved */
1107 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1109 mh = dlm_allocate_mhandle(allocation);
1113 mh->committed = false;
1118 switch (node->version) {
1119 case DLM_VERSION_3_1:
1120 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1123 dlm_free_mhandle(mh);
1128 case DLM_VERSION_3_2:
1129 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1132 dlm_free_mhandle(mh);
1138 dlm_free_mhandle(mh);
1145 /* keep in mind that is a must to call
1146 * dlm_midcomms_commit_msg() which releases
1147 * nodes_srcu using mh->idx which is assumed
1148 * here that the application will call it.
1153 srcu_read_unlock(&nodes_srcu, idx);
1158 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1159 const void *name, int namelen)
1161 switch (mh->inner_p->header.h_cmd) {
1163 trace_dlm_send_message(mh->node->nodeid, mh->seq,
1164 &mh->inner_p->message,
1168 trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1169 &mh->inner_p->rcom);
1172 /* nothing to trace */
1177 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1178 const void *name, int namelen)
1180 /* nexthdr chain for fast lookup */
1181 mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1182 mh->committed = true;
1183 dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1184 dlm_lowcomms_commit_msg(mh->msg);
1187 /* avoid false positive for nodes_srcu, lock was happen in
1188 * dlm_midcomms_get_mhandle
1191 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1192 const void *name, int namelen)
1195 switch (mh->node->version) {
1196 case DLM_VERSION_3_1:
1197 srcu_read_unlock(&nodes_srcu, mh->idx);
1199 dlm_lowcomms_commit_msg(mh->msg);
1200 dlm_lowcomms_put_msg(mh->msg);
1201 /* mh is not part of rcu list in this case */
1202 dlm_free_mhandle(mh);
1204 case DLM_VERSION_3_2:
1205 /* held rcu read lock here, because we sending the
1206 * dlm message out, when we do that we could receive
1207 * an ack back which releases the mhandle and we
1208 * get a use after free.
1211 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1212 srcu_read_unlock(&nodes_srcu, mh->idx);
1216 srcu_read_unlock(&nodes_srcu, mh->idx);
1223 int dlm_midcomms_start(void)
1225 return dlm_lowcomms_start();
1228 void dlm_midcomms_stop(void)
1230 dlm_lowcomms_stop();
1233 void dlm_midcomms_init(void)
1237 for (i = 0; i < CONN_HASH_SIZE; i++)
1238 INIT_HLIST_HEAD(&node_hash[i]);
1240 dlm_lowcomms_init();
1243 void dlm_midcomms_exit(void)
1245 dlm_lowcomms_exit();
1248 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1250 spin_lock(&node->state_lock);
1251 pr_debug("receive active fin ack from node %d with state %s\n",
1252 node->nodeid, dlm_state_str(node->state));
1254 switch (node->state) {
1256 node->state = DLM_FIN_WAIT2;
1257 pr_debug("switch node %d to state %s\n",
1258 node->nodeid, dlm_state_str(node->state));
1261 midcomms_node_reset(node);
1262 pr_debug("switch node %d to state %s\n",
1263 node->nodeid, dlm_state_str(node->state));
1266 /* not valid but somehow we got what we want */
1267 wake_up(&node->shutdown_wait);
1270 spin_unlock(&node->state_lock);
1271 log_print("%s: unexpected state: %d",
1272 __func__, node->state);
1276 spin_unlock(&node->state_lock);
1279 void dlm_midcomms_add_member(int nodeid)
1281 struct midcomms_node *node;
1284 if (nodeid == dlm_our_nodeid())
1287 idx = srcu_read_lock(&nodes_srcu);
1288 node = nodeid2node(nodeid, GFP_NOFS);
1290 srcu_read_unlock(&nodes_srcu, idx);
1294 spin_lock(&node->state_lock);
1296 pr_debug("receive add member from node %d with state %s\n",
1297 node->nodeid, dlm_state_str(node->state));
1298 switch (node->state) {
1299 case DLM_ESTABLISHED:
1302 node->state = DLM_ESTABLISHED;
1303 pr_debug("switch node %d to state %s\n",
1304 node->nodeid, dlm_state_str(node->state));
1307 /* some invalid state passive shutdown
1308 * was failed, we try to reset and
1309 * hope it will go on.
1311 log_print("reset node %d because shutdown stuck",
1314 midcomms_node_reset(node);
1315 node->state = DLM_ESTABLISHED;
1321 pr_debug("node %d users inc count %d\n", nodeid, node->users);
1322 spin_unlock(&node->state_lock);
1324 srcu_read_unlock(&nodes_srcu, idx);
1327 void dlm_midcomms_remove_member(int nodeid)
1329 struct midcomms_node *node;
1332 if (nodeid == dlm_our_nodeid())
1335 idx = srcu_read_lock(&nodes_srcu);
1336 node = nodeid2node(nodeid, 0);
1338 srcu_read_unlock(&nodes_srcu, idx);
1342 spin_lock(&node->state_lock);
1344 pr_debug("node %d users dec count %d\n", nodeid, node->users);
1346 /* hitting users count to zero means the
1347 * other side is running dlm_midcomms_stop()
1348 * we meet us to have a clean disconnect.
1350 if (node->users == 0) {
1351 pr_debug("receive remove member from node %d with state %s\n",
1352 node->nodeid, dlm_state_str(node->state));
1353 switch (node->state) {
1354 case DLM_ESTABLISHED:
1356 case DLM_CLOSE_WAIT:
1357 /* passive shutdown DLM_LAST_ACK case 2 */
1358 node->state = DLM_LAST_ACK;
1359 pr_debug("switch node %d to state %s case 2\n",
1360 node->nodeid, dlm_state_str(node->state));
1361 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1362 dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1365 /* probably receive fin caught it, do nothing */
1368 /* already gone, do nothing */
1371 log_print("%s: unexpected state: %d",
1372 __func__, node->state);
1376 spin_unlock(&node->state_lock);
1378 srcu_read_unlock(&nodes_srcu, idx);
1381 static void midcomms_node_release(struct rcu_head *rcu)
1383 struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1385 WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1386 dlm_send_queue_flush(node);
1390 void dlm_midcomms_version_wait(void)
1392 struct midcomms_node *node;
1395 idx = srcu_read_lock(&nodes_srcu);
1396 for (i = 0; i < CONN_HASH_SIZE; i++) {
1397 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1398 ret = wait_event_timeout(node->shutdown_wait,
1399 node->version != DLM_VERSION_NOT_SET ||
1400 node->state == DLM_CLOSED ||
1401 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1402 DLM_SHUTDOWN_TIMEOUT);
1403 if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1404 pr_debug("version wait timed out for node %d with state %s\n",
1405 node->nodeid, dlm_state_str(node->state));
1408 srcu_read_unlock(&nodes_srcu, idx);
1411 static void midcomms_shutdown(struct midcomms_node *node)
1415 /* old protocol, we don't wait for pending operations */
1416 switch (node->version) {
1417 case DLM_VERSION_3_2:
1423 spin_lock(&node->state_lock);
1424 pr_debug("receive active shutdown for node %d with state %s\n",
1425 node->nodeid, dlm_state_str(node->state));
1426 switch (node->state) {
1427 case DLM_ESTABLISHED:
1428 node->state = DLM_FIN_WAIT1;
1429 pr_debug("switch node %d to state %s case 2\n",
1430 node->nodeid, dlm_state_str(node->state));
1431 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1434 /* we have what we want */
1437 /* busy to enter DLM_FIN_WAIT1, wait until passive
1438 * done in shutdown_wait to enter DLM_CLOSED.
1442 spin_unlock(&node->state_lock);
1444 if (DLM_DEBUG_FENCE_TERMINATION)
1447 /* wait for other side dlm + fin */
1448 ret = wait_event_timeout(node->shutdown_wait,
1449 node->state == DLM_CLOSED ||
1450 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1451 DLM_SHUTDOWN_TIMEOUT);
1452 if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1453 pr_debug("active shutdown timed out for node %d with state %s\n",
1454 node->nodeid, dlm_state_str(node->state));
1456 pr_debug("active shutdown done for node %d with state %s\n",
1457 node->nodeid, dlm_state_str(node->state));
1460 void dlm_midcomms_shutdown(void)
1462 struct midcomms_node *node;
1465 mutex_lock(&close_lock);
1466 idx = srcu_read_lock(&nodes_srcu);
1467 for (i = 0; i < CONN_HASH_SIZE; i++) {
1468 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1469 midcomms_shutdown(node);
1471 dlm_delete_debug_comms_file(node->debugfs);
1473 spin_lock(&nodes_lock);
1474 hlist_del_rcu(&node->hlist);
1475 spin_unlock(&nodes_lock);
1477 call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1480 srcu_read_unlock(&nodes_srcu, idx);
1481 mutex_unlock(&close_lock);
1483 dlm_lowcomms_shutdown();
1486 int dlm_midcomms_close(int nodeid)
1488 struct midcomms_node *node;
1491 if (nodeid == dlm_our_nodeid())
1494 dlm_stop_lockspaces_check();
1496 idx = srcu_read_lock(&nodes_srcu);
1497 /* Abort pending close/remove operation */
1498 node = nodeid2node(nodeid, 0);
1500 /* let shutdown waiters leave */
1501 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1502 wake_up(&node->shutdown_wait);
1504 srcu_read_unlock(&nodes_srcu, idx);
1506 synchronize_srcu(&nodes_srcu);
1508 idx = srcu_read_lock(&nodes_srcu);
1509 mutex_lock(&close_lock);
1510 node = nodeid2node(nodeid, 0);
1512 mutex_unlock(&close_lock);
1513 srcu_read_unlock(&nodes_srcu, idx);
1514 return dlm_lowcomms_close(nodeid);
1517 ret = dlm_lowcomms_close(nodeid);
1518 spin_lock(&node->state_lock);
1519 midcomms_node_reset(node);
1520 spin_unlock(&node->state_lock);
1521 srcu_read_unlock(&nodes_srcu, idx);
1522 mutex_unlock(&close_lock);
1527 /* debug functionality to send raw dlm msg from user space */
1528 struct dlm_rawmsg_data {
1529 struct midcomms_node *node;
1533 static void midcomms_new_rawmsg_cb(void *data)
1535 struct dlm_rawmsg_data *rd = data;
1536 struct dlm_header *h = rd->buf;
1538 switch (h->h_version) {
1539 case cpu_to_le32(DLM_VERSION_3_1):
1545 h->u.h_seq = cpu_to_le32(rd->node->seq_send++);
1554 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1557 struct dlm_rawmsg_data rd;
1558 struct dlm_msg *msg;
1564 msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1565 &msgbuf, midcomms_new_rawmsg_cb, &rd);
1569 memcpy(msgbuf, buf, buflen);
1570 dlm_lowcomms_commit_msg(msg);