1 /******************************************************************************
2 *******************************************************************************
4 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
5 ** Copyright (C) 2004-2009 Red Hat, Inc. All rights reserved.
7 ** This copyrighted material is made available to anyone wishing to use,
8 ** modify, copy, or redistribute it subject to the terms and conditions
9 ** of the GNU General Public License v.2.
11 *******************************************************************************
12 ******************************************************************************/
17 * This is the "low-level" comms layer.
19 * It is responsible for sending/receiving messages
20 * from other nodes in the cluster.
22 * Cluster nodes are referred to by their nodeids. nodeids are
23 * simply 32 bit numbers to the locking module - if they need to
24 * be expanded for the cluster infrastructure then that is its
25 * responsibility. It is this layer's
26 * responsibility to resolve these into IP address or
27 * whatever it needs for inter-node communication.
29 * The comms level is two kernel threads that deal mainly with
30 * the receiving of messages from other nodes and passing them
31 * up to the mid-level comms layer (which understands the
32 * message format) for execution by the locking core, and
33 * a send thread which does all the setting up of connections
34 * to remote nodes and the sending of data. Threads are not allowed
35 * to send their own data because it may cause them to wait in times
36 * of high load. Also, this way, the sending thread can collect together
37 * messages bound for one node and send them in one block.
39 * lowcomms will choose to use either TCP or SCTP as its transport layer
40 * depending on the configuration variable 'protocol'. This should be set
41 * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
42 * cluster-wide mechanism as it must be the same on all nodes of the cluster
43 * for the DLM to function.
47 #include <asm/ioctls.h>
50 #include <linux/pagemap.h>
51 #include <linux/file.h>
52 #include <linux/mutex.h>
53 #include <linux/sctp.h>
54 #include <linux/slab.h>
55 #include <net/sctp/sctp.h>
56 #include <net/sctp/user.h>
59 #include "dlm_internal.h"
64 #define NEEDED_RMEM (4*1024*1024)
65 #define CONN_HASH_SIZE 32
67 /* Number of messages to send before rescheduling */
68 #define MAX_SEND_MSG_COUNT 25
76 static void cbuf_add(struct cbuf *cb, int n)
81 static int cbuf_data(struct cbuf *cb)
83 return ((cb->base + cb->len) & cb->mask);
86 static void cbuf_init(struct cbuf *cb, int size)
88 cb->base = cb->len = 0;
92 static void cbuf_eat(struct cbuf *cb, int n)
99 static bool cbuf_empty(struct cbuf *cb)
105 struct socket *sock; /* NULL if not connected */
106 uint32_t nodeid; /* So we know who we are in the list */
107 struct mutex sock_mutex;
109 #define CF_READ_PENDING 1
110 #define CF_WRITE_PENDING 2
111 #define CF_CONNECT_PENDING 3
112 #define CF_INIT_PENDING 4
113 #define CF_IS_OTHERCON 5
115 #define CF_APP_LIMITED 7
116 struct list_head writequeue; /* List of outgoing writequeue_entries */
117 spinlock_t writequeue_lock;
118 int (*rx_action) (struct connection *); /* What to do when active */
119 void (*connect_action) (struct connection *); /* What to do to connect */
120 struct page *rx_page;
123 #define MAX_CONNECT_RETRIES 3
125 struct hlist_node list;
126 struct connection *othercon;
127 struct work_struct rwork; /* Receive workqueue */
128 struct work_struct swork; /* Send workqueue */
130 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
132 /* An entry waiting to be sent */
133 struct writequeue_entry {
134 struct list_head list;
140 struct connection *con;
143 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
144 static int dlm_local_count;
147 static struct workqueue_struct *recv_workqueue;
148 static struct workqueue_struct *send_workqueue;
150 static struct hlist_head connection_hash[CONN_HASH_SIZE];
151 static DEFINE_MUTEX(connections_lock);
152 static struct kmem_cache *con_cache;
154 static void process_recv_sockets(struct work_struct *work);
155 static void process_send_sockets(struct work_struct *work);
158 /* This is deliberately very simple because most clusters have simple
159 sequential nodeids, so we should be able to go straight to a connection
160 struct in the array */
161 static inline int nodeid_hash(int nodeid)
163 return nodeid & (CONN_HASH_SIZE-1);
166 static struct connection *__find_con(int nodeid)
169 struct hlist_node *h;
170 struct connection *con;
172 r = nodeid_hash(nodeid);
174 hlist_for_each_entry(con, h, &connection_hash[r], list) {
175 if (con->nodeid == nodeid)
182 * If 'allocation' is zero then we don't attempt to create a new
183 * connection structure for this node.
185 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
187 struct connection *con = NULL;
190 con = __find_con(nodeid);
194 con = kmem_cache_zalloc(con_cache, alloc);
198 r = nodeid_hash(nodeid);
199 hlist_add_head(&con->list, &connection_hash[r]);
201 con->nodeid = nodeid;
202 mutex_init(&con->sock_mutex);
203 INIT_LIST_HEAD(&con->writequeue);
204 spin_lock_init(&con->writequeue_lock);
205 INIT_WORK(&con->swork, process_send_sockets);
206 INIT_WORK(&con->rwork, process_recv_sockets);
208 /* Setup action pointers for child sockets */
210 struct connection *zerocon = __find_con(0);
212 con->connect_action = zerocon->connect_action;
214 con->rx_action = zerocon->rx_action;
220 /* Loop round all connections */
221 static void foreach_conn(void (*conn_func)(struct connection *c))
224 struct hlist_node *h, *n;
225 struct connection *con;
227 for (i = 0; i < CONN_HASH_SIZE; i++) {
228 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
234 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
236 struct connection *con;
238 mutex_lock(&connections_lock);
239 con = __nodeid2con(nodeid, allocation);
240 mutex_unlock(&connections_lock);
245 /* This is a bit drastic, but only called when things go wrong */
246 static struct connection *assoc2con(int assoc_id)
249 struct hlist_node *h;
250 struct connection *con;
252 mutex_lock(&connections_lock);
254 for (i = 0 ; i < CONN_HASH_SIZE; i++) {
255 hlist_for_each_entry(con, h, &connection_hash[i], list) {
256 if (con->sctp_assoc == assoc_id) {
257 mutex_unlock(&connections_lock);
262 mutex_unlock(&connections_lock);
266 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
268 struct sockaddr_storage addr;
271 if (!dlm_local_count)
274 error = dlm_nodeid_to_addr(nodeid, &addr);
278 if (dlm_local_addr[0]->ss_family == AF_INET) {
279 struct sockaddr_in *in4 = (struct sockaddr_in *) &addr;
280 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
281 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
283 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &addr;
284 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
285 ret6->sin6_addr = in6->sin6_addr;
291 /* Data available on socket or listen socket received a connect */
292 static void lowcomms_data_ready(struct sock *sk, int count_unused)
294 struct connection *con = sock2con(sk);
295 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
296 queue_work(recv_workqueue, &con->rwork);
299 static void lowcomms_write_space(struct sock *sk)
301 struct connection *con = sock2con(sk);
306 clear_bit(SOCK_NOSPACE, &con->sock->flags);
308 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
309 con->sock->sk->sk_write_pending--;
310 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
313 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
314 queue_work(send_workqueue, &con->swork);
317 static inline void lowcomms_connect_sock(struct connection *con)
319 if (test_bit(CF_CLOSE, &con->flags))
321 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
322 queue_work(send_workqueue, &con->swork);
325 static void lowcomms_state_change(struct sock *sk)
327 if (sk->sk_state == TCP_ESTABLISHED)
328 lowcomms_write_space(sk);
331 int dlm_lowcomms_connect_node(int nodeid)
333 struct connection *con;
335 /* with sctp there's no connecting without sending */
336 if (dlm_config.ci_protocol != 0)
339 if (nodeid == dlm_our_nodeid())
342 con = nodeid2con(nodeid, GFP_NOFS);
345 lowcomms_connect_sock(con);
349 /* Make a socket active */
350 static int add_sock(struct socket *sock, struct connection *con)
354 /* Install a data_ready callback */
355 con->sock->sk->sk_data_ready = lowcomms_data_ready;
356 con->sock->sk->sk_write_space = lowcomms_write_space;
357 con->sock->sk->sk_state_change = lowcomms_state_change;
358 con->sock->sk->sk_user_data = con;
359 con->sock->sk->sk_allocation = GFP_NOFS;
363 /* Add the port number to an IPv6 or 4 sockaddr and return the address
365 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
368 saddr->ss_family = dlm_local_addr[0]->ss_family;
369 if (saddr->ss_family == AF_INET) {
370 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
371 in4_addr->sin_port = cpu_to_be16(port);
372 *addr_len = sizeof(struct sockaddr_in);
373 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
375 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
376 in6_addr->sin6_port = cpu_to_be16(port);
377 *addr_len = sizeof(struct sockaddr_in6);
379 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
382 /* Close a remote connection and tidy up */
383 static void close_connection(struct connection *con, bool and_other)
385 mutex_lock(&con->sock_mutex);
388 sock_release(con->sock);
391 if (con->othercon && and_other) {
392 /* Will only re-enter once. */
393 close_connection(con->othercon, false);
396 __free_page(con->rx_page);
401 mutex_unlock(&con->sock_mutex);
404 /* We only send shutdown messages to nodes that are not part of the cluster */
405 static void sctp_send_shutdown(sctp_assoc_t associd)
407 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
408 struct msghdr outmessage;
409 struct cmsghdr *cmsg;
410 struct sctp_sndrcvinfo *sinfo;
412 struct connection *con;
414 con = nodeid2con(0,0);
417 outmessage.msg_name = NULL;
418 outmessage.msg_namelen = 0;
419 outmessage.msg_control = outcmsg;
420 outmessage.msg_controllen = sizeof(outcmsg);
421 outmessage.msg_flags = MSG_EOR;
423 cmsg = CMSG_FIRSTHDR(&outmessage);
424 cmsg->cmsg_level = IPPROTO_SCTP;
425 cmsg->cmsg_type = SCTP_SNDRCV;
426 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
427 outmessage.msg_controllen = cmsg->cmsg_len;
428 sinfo = CMSG_DATA(cmsg);
429 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
431 sinfo->sinfo_flags |= MSG_EOF;
432 sinfo->sinfo_assoc_id = associd;
434 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
437 log_print("send EOF to node failed: %d", ret);
440 static void sctp_init_failed_foreach(struct connection *con)
443 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
444 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
445 queue_work(send_workqueue, &con->swork);
449 /* INIT failed but we don't know which node...
450 restart INIT on all pending nodes */
451 static void sctp_init_failed(void)
453 mutex_lock(&connections_lock);
455 foreach_conn(sctp_init_failed_foreach);
457 mutex_unlock(&connections_lock);
460 /* Something happened to an association */
461 static void process_sctp_notification(struct connection *con,
462 struct msghdr *msg, char *buf)
464 union sctp_notification *sn = (union sctp_notification *)buf;
466 if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
467 switch (sn->sn_assoc_change.sac_state) {
472 /* Check that the new node is in the lockspace */
473 struct sctp_prim prim;
477 struct connection *new_con;
480 * We get this before any data for an association.
481 * We verify that the node is in the cluster and
482 * then peel off a socket for it.
484 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
485 log_print("COMM_UP for invalid assoc ID %d",
486 (int)sn->sn_assoc_change.sac_assoc_id);
490 memset(&prim, 0, sizeof(struct sctp_prim));
491 prim_len = sizeof(struct sctp_prim);
492 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
494 ret = kernel_getsockopt(con->sock,
500 log_print("getsockopt/sctp_primary_addr on "
501 "new assoc %d failed : %d",
502 (int)sn->sn_assoc_change.sac_assoc_id,
505 /* Retry INIT later */
506 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
508 clear_bit(CF_CONNECT_PENDING, &con->flags);
511 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
512 if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
513 unsigned char *b=(unsigned char *)&prim.ssp_addr;
514 log_print("reject connect from unknown addr");
515 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
516 b, sizeof(struct sockaddr_storage));
517 sctp_send_shutdown(prim.ssp_assoc_id);
521 new_con = nodeid2con(nodeid, GFP_NOFS);
525 /* Peel off a new sock */
526 sctp_lock_sock(con->sock->sk);
527 ret = sctp_do_peeloff(con->sock->sk,
528 sn->sn_assoc_change.sac_assoc_id,
530 sctp_release_sock(con->sock->sk);
532 log_print("Can't peel off a socket for "
533 "connection %d to node %d: err=%d",
534 (int)sn->sn_assoc_change.sac_assoc_id,
538 add_sock(new_con->sock, new_con);
540 log_print("connecting to %d sctp association %d",
541 nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
543 /* Send any pending writes */
544 clear_bit(CF_CONNECT_PENDING, &new_con->flags);
545 clear_bit(CF_INIT_PENDING, &con->flags);
546 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
547 queue_work(send_workqueue, &new_con->swork);
549 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
550 queue_work(recv_workqueue, &new_con->rwork);
555 case SCTP_SHUTDOWN_COMP:
557 con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
564 /* We don't know which INIT failed, so clear the PENDING flags
565 * on them all. if assoc_id is zero then it will then try
568 case SCTP_CANT_STR_ASSOC:
570 log_print("Can't start SCTP association - retrying");
576 log_print("unexpected SCTP assoc change id=%d state=%d",
577 (int)sn->sn_assoc_change.sac_assoc_id,
578 sn->sn_assoc_change.sac_state);
583 /* Data received from remote end */
584 static int receive_from_sock(struct connection *con)
587 struct msghdr msg = {};
591 int call_again_soon = 0;
593 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
595 mutex_lock(&con->sock_mutex);
597 if (con->sock == NULL) {
602 if (con->rx_page == NULL) {
604 * This doesn't need to be atomic, but I think it should
605 * improve performance if it is.
607 con->rx_page = alloc_page(GFP_ATOMIC);
608 if (con->rx_page == NULL)
610 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
613 /* Only SCTP needs these really */
614 memset(&incmsg, 0, sizeof(incmsg));
615 msg.msg_control = incmsg;
616 msg.msg_controllen = sizeof(incmsg);
619 * iov[0] is the bit of the circular buffer between the current end
620 * point (cb.base + cb.len) and the end of the buffer.
622 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
623 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
628 * iov[1] is the bit of the circular buffer between the start of the
629 * buffer and the start of the currently used section (cb.base)
631 if (cbuf_data(&con->cb) >= con->cb.base) {
632 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
633 iov[1].iov_len = con->cb.base;
634 iov[1].iov_base = page_address(con->rx_page);
637 len = iov[0].iov_len + iov[1].iov_len;
639 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
640 MSG_DONTWAIT | MSG_NOSIGNAL);
644 /* Process SCTP notifications */
645 if (msg.msg_flags & MSG_NOTIFICATION) {
646 msg.msg_control = incmsg;
647 msg.msg_controllen = sizeof(incmsg);
649 process_sctp_notification(con, &msg,
650 page_address(con->rx_page) + con->cb.base);
651 mutex_unlock(&con->sock_mutex);
654 BUG_ON(con->nodeid == 0);
658 cbuf_add(&con->cb, ret);
659 ret = dlm_process_incoming_buffer(con->nodeid,
660 page_address(con->rx_page),
661 con->cb.base, con->cb.len,
663 if (ret == -EBADMSG) {
664 log_print("lowcomms: addr=%p, base=%u, len=%u, "
665 "iov_len=%u, iov_base[0]=%p, read=%d",
666 page_address(con->rx_page), con->cb.base, con->cb.len,
667 len, iov[0].iov_base, r);
671 cbuf_eat(&con->cb, ret);
673 if (cbuf_empty(&con->cb) && !call_again_soon) {
674 __free_page(con->rx_page);
680 mutex_unlock(&con->sock_mutex);
684 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
685 queue_work(recv_workqueue, &con->rwork);
686 mutex_unlock(&con->sock_mutex);
690 mutex_unlock(&con->sock_mutex);
691 if (ret != -EAGAIN) {
692 close_connection(con, false);
693 /* Reconnect when there is something to send */
695 /* Don't return success if we really got EOF */
702 /* Listening socket is busy, accept a connection */
703 static int tcp_accept_from_sock(struct connection *con)
706 struct sockaddr_storage peeraddr;
707 struct socket *newsock;
710 struct connection *newcon;
711 struct connection *addcon;
713 memset(&peeraddr, 0, sizeof(peeraddr));
714 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
715 IPPROTO_TCP, &newsock);
719 mutex_lock_nested(&con->sock_mutex, 0);
722 if (con->sock == NULL)
725 newsock->type = con->sock->type;
726 newsock->ops = con->sock->ops;
728 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
732 /* Get the connected socket's peer */
733 memset(&peeraddr, 0, sizeof(peeraddr));
734 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
736 result = -ECONNABORTED;
740 /* Get the new node's NODEID */
741 make_sockaddr(&peeraddr, 0, &len);
742 if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
743 unsigned char *b=(unsigned char *)&peeraddr;
744 log_print("connect from non cluster node");
745 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
746 b, sizeof(struct sockaddr_storage));
747 sock_release(newsock);
748 mutex_unlock(&con->sock_mutex);
752 log_print("got connection from %d", nodeid);
754 /* Check to see if we already have a connection to this node. This
755 * could happen if the two nodes initiate a connection at roughly
756 * the same time and the connections cross on the wire.
757 * In this case we store the incoming one in "othercon"
759 newcon = nodeid2con(nodeid, GFP_NOFS);
764 mutex_lock_nested(&newcon->sock_mutex, 1);
766 struct connection *othercon = newcon->othercon;
769 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
771 log_print("failed to allocate incoming socket");
772 mutex_unlock(&newcon->sock_mutex);
776 othercon->nodeid = nodeid;
777 othercon->rx_action = receive_from_sock;
778 mutex_init(&othercon->sock_mutex);
779 INIT_WORK(&othercon->swork, process_send_sockets);
780 INIT_WORK(&othercon->rwork, process_recv_sockets);
781 set_bit(CF_IS_OTHERCON, &othercon->flags);
783 if (!othercon->sock) {
784 newcon->othercon = othercon;
785 othercon->sock = newsock;
786 newsock->sk->sk_user_data = othercon;
787 add_sock(newsock, othercon);
791 printk("Extra connection from node %d attempted\n", nodeid);
793 mutex_unlock(&newcon->sock_mutex);
798 newsock->sk->sk_user_data = newcon;
799 newcon->rx_action = receive_from_sock;
800 add_sock(newsock, newcon);
804 mutex_unlock(&newcon->sock_mutex);
807 * Add it to the active queue in case we got data
808 * between processing the accept adding the socket
809 * to the read_sockets list
811 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
812 queue_work(recv_workqueue, &addcon->rwork);
813 mutex_unlock(&con->sock_mutex);
818 mutex_unlock(&con->sock_mutex);
819 sock_release(newsock);
821 if (result != -EAGAIN)
822 log_print("error accepting connection from node: %d", result);
826 static void free_entry(struct writequeue_entry *e)
828 __free_page(e->page);
832 /* Initiate an SCTP association.
833 This is a special case of send_to_sock() in that we don't yet have a
834 peeled-off socket for this association, so we use the listening socket
835 and add the primary IP address of the remote node.
837 static void sctp_init_assoc(struct connection *con)
839 struct sockaddr_storage rem_addr;
840 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
841 struct msghdr outmessage;
842 struct cmsghdr *cmsg;
843 struct sctp_sndrcvinfo *sinfo;
844 struct connection *base_con;
845 struct writequeue_entry *e;
851 if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
854 if (con->retries++ > MAX_CONNECT_RETRIES)
857 if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
858 log_print("no address for nodeid %d", con->nodeid);
861 base_con = nodeid2con(0, 0);
862 BUG_ON(base_con == NULL);
864 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
866 outmessage.msg_name = &rem_addr;
867 outmessage.msg_namelen = addrlen;
868 outmessage.msg_control = outcmsg;
869 outmessage.msg_controllen = sizeof(outcmsg);
870 outmessage.msg_flags = MSG_EOR;
872 spin_lock(&con->writequeue_lock);
874 if (list_empty(&con->writequeue)) {
875 spin_unlock(&con->writequeue_lock);
876 log_print("writequeue empty for nodeid %d", con->nodeid);
880 e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
883 spin_unlock(&con->writequeue_lock);
885 /* Send the first block off the write queue */
886 iov[0].iov_base = page_address(e->page)+offset;
887 iov[0].iov_len = len;
889 cmsg = CMSG_FIRSTHDR(&outmessage);
890 cmsg->cmsg_level = IPPROTO_SCTP;
891 cmsg->cmsg_type = SCTP_SNDRCV;
892 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
893 sinfo = CMSG_DATA(cmsg);
894 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
895 sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
896 outmessage.msg_controllen = cmsg->cmsg_len;
898 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
900 log_print("Send first packet to node %d failed: %d",
903 /* Try again later */
904 clear_bit(CF_CONNECT_PENDING, &con->flags);
905 clear_bit(CF_INIT_PENDING, &con->flags);
908 spin_lock(&con->writequeue_lock);
912 if (e->len == 0 && e->users == 0) {
916 spin_unlock(&con->writequeue_lock);
920 /* Connect a new socket to its peer */
921 static void tcp_connect_to_sock(struct connection *con)
923 int result = -EHOSTUNREACH;
924 struct sockaddr_storage saddr, src_addr;
926 struct socket *sock = NULL;
929 if (con->nodeid == 0) {
930 log_print("attempt to connect sock 0 foiled");
934 mutex_lock(&con->sock_mutex);
935 if (con->retries++ > MAX_CONNECT_RETRIES)
938 /* Some odd races can cause double-connects, ignore them */
944 /* Create a socket to communicate with */
945 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
950 memset(&saddr, 0, sizeof(saddr));
951 if (dlm_nodeid_to_addr(con->nodeid, &saddr))
954 sock->sk->sk_user_data = con;
955 con->rx_action = receive_from_sock;
956 con->connect_action = tcp_connect_to_sock;
959 /* Bind to our cluster-known address connecting to avoid
961 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
962 make_sockaddr(&src_addr, 0, &addr_len);
963 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
966 log_print("could not bind for connect: %d", result);
967 /* This *may* not indicate a critical error */
970 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
972 log_print("connecting to %d", con->nodeid);
974 /* Turn off Nagle's algorithm */
975 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
979 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
981 if (result == -EINPROGRESS)
988 sock_release(con->sock);
994 * Some errors are fatal and this list might need adjusting. For other
995 * errors we try again until the max number of retries is reached.
997 if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
998 result != -ENETDOWN && result != -EINVAL
999 && result != -EPROTONOSUPPORT) {
1000 lowcomms_connect_sock(con);
1004 mutex_unlock(&con->sock_mutex);
1008 static struct socket *tcp_create_listen_sock(struct connection *con,
1009 struct sockaddr_storage *saddr)
1011 struct socket *sock = NULL;
1016 if (dlm_local_addr[0]->ss_family == AF_INET)
1017 addr_len = sizeof(struct sockaddr_in);
1019 addr_len = sizeof(struct sockaddr_in6);
1021 /* Create a socket to communicate with */
1022 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1023 IPPROTO_TCP, &sock);
1025 log_print("Can't create listening comms socket");
1029 /* Turn off Nagle's algorithm */
1030 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1033 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1034 (char *)&one, sizeof(one));
1037 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1039 sock->sk->sk_user_data = con;
1040 con->rx_action = tcp_accept_from_sock;
1041 con->connect_action = tcp_connect_to_sock;
1044 /* Bind to our port */
1045 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1046 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1048 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1054 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1055 (char *)&one, sizeof(one));
1057 log_print("Set keepalive failed: %d", result);
1060 result = sock->ops->listen(sock, 5);
1062 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1072 /* Get local addresses */
1073 static void init_local(void)
1075 struct sockaddr_storage sas, *addr;
1078 dlm_local_count = 0;
1079 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1080 if (dlm_our_addr(&sas, i))
1083 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1086 memcpy(addr, &sas, sizeof(*addr));
1087 dlm_local_addr[dlm_local_count++] = addr;
1091 /* Bind to an IP address. SCTP allows multiple address so it can do
1093 static int add_sctp_bind_addr(struct connection *sctp_con,
1094 struct sockaddr_storage *addr,
1095 int addr_len, int num)
1100 result = kernel_bind(sctp_con->sock,
1101 (struct sockaddr *) addr,
1104 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1105 SCTP_SOCKOPT_BINDX_ADD,
1106 (char *)addr, addr_len);
1109 log_print("Can't bind to port %d addr number %d",
1110 dlm_config.ci_tcp_port, num);
1115 /* Initialise SCTP socket and bind to all interfaces */
1116 static int sctp_listen_for_all(void)
1118 struct socket *sock = NULL;
1119 struct sockaddr_storage localaddr;
1120 struct sctp_event_subscribe subscribe;
1121 int result = -EINVAL, num = 1, i, addr_len;
1122 struct connection *con = nodeid2con(0, GFP_NOFS);
1123 int bufsize = NEEDED_RMEM;
1128 log_print("Using SCTP for communications");
1130 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1131 IPPROTO_SCTP, &sock);
1133 log_print("Can't create comms socket, check SCTP is loaded");
1137 /* Listen for events */
1138 memset(&subscribe, 0, sizeof(subscribe));
1139 subscribe.sctp_data_io_event = 1;
1140 subscribe.sctp_association_event = 1;
1141 subscribe.sctp_send_failure_event = 1;
1142 subscribe.sctp_shutdown_event = 1;
1143 subscribe.sctp_partial_delivery_event = 1;
1145 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1146 (char *)&bufsize, sizeof(bufsize));
1148 log_print("Error increasing buffer space on socket %d", result);
1150 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1151 (char *)&subscribe, sizeof(subscribe));
1153 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1155 goto create_delsock;
1158 /* Init con struct */
1159 sock->sk->sk_user_data = con;
1161 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1162 con->rx_action = receive_from_sock;
1163 con->connect_action = sctp_init_assoc;
1165 /* Bind to all interfaces. */
1166 for (i = 0; i < dlm_local_count; i++) {
1167 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1168 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1170 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1172 goto create_delsock;
1176 result = sock->ops->listen(sock, 5);
1178 log_print("Can't set socket listening");
1179 goto create_delsock;
1191 static int tcp_listen_for_all(void)
1193 struct socket *sock = NULL;
1194 struct connection *con = nodeid2con(0, GFP_NOFS);
1195 int result = -EINVAL;
1200 /* We don't support multi-homed hosts */
1201 if (dlm_local_addr[1] != NULL) {
1202 log_print("TCP protocol can't handle multi-homed hosts, "
1207 log_print("Using TCP for communications");
1209 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1211 add_sock(sock, con);
1215 result = -EADDRINUSE;
1223 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1226 struct writequeue_entry *entry;
1228 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1232 entry->page = alloc_page(allocation);
1247 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1249 struct connection *con;
1250 struct writequeue_entry *e;
1254 con = nodeid2con(nodeid, allocation);
1258 spin_lock(&con->writequeue_lock);
1259 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1260 if ((&e->list == &con->writequeue) ||
1261 (PAGE_CACHE_SIZE - e->end < len)) {
1268 spin_unlock(&con->writequeue_lock);
1272 *ppc = page_address(e->page) + offset;
1276 e = new_writequeue_entry(con, allocation);
1278 spin_lock(&con->writequeue_lock);
1282 list_add_tail(&e->list, &con->writequeue);
1283 spin_unlock(&con->writequeue_lock);
1289 void dlm_lowcomms_commit_buffer(void *mh)
1291 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1292 struct connection *con = e->con;
1295 spin_lock(&con->writequeue_lock);
1299 e->len = e->end - e->offset;
1300 spin_unlock(&con->writequeue_lock);
1302 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1303 queue_work(send_workqueue, &con->swork);
1308 spin_unlock(&con->writequeue_lock);
1312 /* Send a message */
1313 static void send_to_sock(struct connection *con)
1316 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1317 struct writequeue_entry *e;
1321 mutex_lock(&con->sock_mutex);
1322 if (con->sock == NULL)
1325 spin_lock(&con->writequeue_lock);
1327 e = list_entry(con->writequeue.next, struct writequeue_entry,
1329 if ((struct list_head *) e == &con->writequeue)
1334 BUG_ON(len == 0 && e->users == 0);
1335 spin_unlock(&con->writequeue_lock);
1339 ret = kernel_sendpage(con->sock, e->page, offset, len,
1341 if (ret == -EAGAIN || ret == 0) {
1342 if (ret == -EAGAIN &&
1343 test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1344 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1345 /* Notify TCP that we're limited by the
1346 * application window size.
1348 set_bit(SOCK_NOSPACE, &con->sock->flags);
1349 con->sock->sk->sk_write_pending++;
1358 /* Don't starve people filling buffers */
1359 if (++count >= MAX_SEND_MSG_COUNT) {
1364 spin_lock(&con->writequeue_lock);
1368 if (e->len == 0 && e->users == 0) {
1374 spin_unlock(&con->writequeue_lock);
1376 mutex_unlock(&con->sock_mutex);
1380 mutex_unlock(&con->sock_mutex);
1381 close_connection(con, false);
1382 lowcomms_connect_sock(con);
1386 mutex_unlock(&con->sock_mutex);
1387 if (!test_bit(CF_INIT_PENDING, &con->flags))
1388 lowcomms_connect_sock(con);
1392 static void clean_one_writequeue(struct connection *con)
1394 struct writequeue_entry *e, *safe;
1396 spin_lock(&con->writequeue_lock);
1397 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1401 spin_unlock(&con->writequeue_lock);
1404 /* Called from recovery when it knows that a node has
1406 int dlm_lowcomms_close(int nodeid)
1408 struct connection *con;
1410 log_print("closing connection to node %d", nodeid);
1411 con = nodeid2con(nodeid, 0);
1413 clear_bit(CF_CONNECT_PENDING, &con->flags);
1414 clear_bit(CF_WRITE_PENDING, &con->flags);
1415 set_bit(CF_CLOSE, &con->flags);
1416 if (cancel_work_sync(&con->swork))
1417 log_print("canceled swork for node %d", nodeid);
1418 if (cancel_work_sync(&con->rwork))
1419 log_print("canceled rwork for node %d", nodeid);
1420 clean_one_writequeue(con);
1421 close_connection(con, true);
1426 /* Receive workqueue function */
1427 static void process_recv_sockets(struct work_struct *work)
1429 struct connection *con = container_of(work, struct connection, rwork);
1432 clear_bit(CF_READ_PENDING, &con->flags);
1434 err = con->rx_action(con);
1438 /* Send workqueue function */
1439 static void process_send_sockets(struct work_struct *work)
1441 struct connection *con = container_of(work, struct connection, swork);
1443 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1444 con->connect_action(con);
1445 set_bit(CF_WRITE_PENDING, &con->flags);
1447 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1452 /* Discard all entries on the write queues */
1453 static void clean_writequeues(void)
1455 foreach_conn(clean_one_writequeue);
1458 static void work_stop(void)
1460 destroy_workqueue(recv_workqueue);
1461 destroy_workqueue(send_workqueue);
1464 static int work_start(void)
1466 recv_workqueue = alloc_workqueue("dlm_recv",
1467 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1468 if (!recv_workqueue) {
1469 log_print("can't start dlm_recv");
1473 send_workqueue = alloc_workqueue("dlm_send",
1474 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1475 if (!send_workqueue) {
1476 log_print("can't start dlm_send");
1477 destroy_workqueue(recv_workqueue);
1484 static void stop_conn(struct connection *con)
1487 if (con->sock && con->sock->sk)
1488 con->sock->sk->sk_user_data = NULL;
1491 static void free_conn(struct connection *con)
1493 close_connection(con, true);
1495 kmem_cache_free(con_cache, con->othercon);
1496 hlist_del(&con->list);
1497 kmem_cache_free(con_cache, con);
1500 void dlm_lowcomms_stop(void)
1502 /* Set all the flags to prevent any
1505 mutex_lock(&connections_lock);
1506 foreach_conn(stop_conn);
1507 mutex_unlock(&connections_lock);
1511 mutex_lock(&connections_lock);
1512 clean_writequeues();
1514 foreach_conn(free_conn);
1516 mutex_unlock(&connections_lock);
1517 kmem_cache_destroy(con_cache);
1520 int dlm_lowcomms_start(void)
1522 int error = -EINVAL;
1523 struct connection *con;
1526 for (i = 0; i < CONN_HASH_SIZE; i++)
1527 INIT_HLIST_HEAD(&connection_hash[i]);
1530 if (!dlm_local_count) {
1532 log_print("no local IP address has been set");
1537 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1538 __alignof__(struct connection), 0,
1543 /* Start listening */
1544 if (dlm_config.ci_protocol == 0)
1545 error = tcp_listen_for_all();
1547 error = sctp_listen_for_all();
1551 error = work_start();
1558 con = nodeid2con(0,0);
1560 close_connection(con, false);
1561 kmem_cache_free(con_cache, con);
1563 kmem_cache_destroy(con_cache);