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 <net/sctp/user.h>
57 #include "dlm_internal.h"
62 #define NEEDED_RMEM (4*1024*1024)
63 #define CONN_HASH_SIZE 32
71 static void cbuf_add(struct cbuf *cb, int n)
76 static int cbuf_data(struct cbuf *cb)
78 return ((cb->base + cb->len) & cb->mask);
81 static void cbuf_init(struct cbuf *cb, int size)
83 cb->base = cb->len = 0;
87 static void cbuf_eat(struct cbuf *cb, int n)
94 static bool cbuf_empty(struct cbuf *cb)
100 struct socket *sock; /* NULL if not connected */
101 uint32_t nodeid; /* So we know who we are in the list */
102 struct mutex sock_mutex;
104 #define CF_READ_PENDING 1
105 #define CF_WRITE_PENDING 2
106 #define CF_CONNECT_PENDING 3
107 #define CF_INIT_PENDING 4
108 #define CF_IS_OTHERCON 5
110 struct list_head writequeue; /* List of outgoing writequeue_entries */
111 spinlock_t writequeue_lock;
112 int (*rx_action) (struct connection *); /* What to do when active */
113 void (*connect_action) (struct connection *); /* What to do to connect */
114 struct page *rx_page;
117 #define MAX_CONNECT_RETRIES 3
119 struct hlist_node list;
120 struct connection *othercon;
121 struct work_struct rwork; /* Receive workqueue */
122 struct work_struct swork; /* Send workqueue */
124 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
126 /* An entry waiting to be sent */
127 struct writequeue_entry {
128 struct list_head list;
134 struct connection *con;
137 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
138 static int dlm_local_count;
141 static struct workqueue_struct *recv_workqueue;
142 static struct workqueue_struct *send_workqueue;
144 static struct hlist_head connection_hash[CONN_HASH_SIZE];
145 static DEFINE_MUTEX(connections_lock);
146 static struct kmem_cache *con_cache;
148 static void process_recv_sockets(struct work_struct *work);
149 static void process_send_sockets(struct work_struct *work);
152 /* This is deliberately very simple because most clusters have simple
153 sequential nodeids, so we should be able to go straight to a connection
154 struct in the array */
155 static inline int nodeid_hash(int nodeid)
157 return nodeid & (CONN_HASH_SIZE-1);
160 static struct connection *__find_con(int nodeid)
163 struct hlist_node *h;
164 struct connection *con;
166 r = nodeid_hash(nodeid);
168 hlist_for_each_entry(con, h, &connection_hash[r], list) {
169 if (con->nodeid == nodeid)
176 * If 'allocation' is zero then we don't attempt to create a new
177 * connection structure for this node.
179 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
181 struct connection *con = NULL;
184 con = __find_con(nodeid);
188 con = kmem_cache_zalloc(con_cache, alloc);
192 r = nodeid_hash(nodeid);
193 hlist_add_head(&con->list, &connection_hash[r]);
195 con->nodeid = nodeid;
196 mutex_init(&con->sock_mutex);
197 INIT_LIST_HEAD(&con->writequeue);
198 spin_lock_init(&con->writequeue_lock);
199 INIT_WORK(&con->swork, process_send_sockets);
200 INIT_WORK(&con->rwork, process_recv_sockets);
202 /* Setup action pointers for child sockets */
204 struct connection *zerocon = __find_con(0);
206 con->connect_action = zerocon->connect_action;
208 con->rx_action = zerocon->rx_action;
214 /* Loop round all connections */
215 static void foreach_conn(void (*conn_func)(struct connection *c))
218 struct hlist_node *h, *n;
219 struct connection *con;
221 for (i = 0; i < CONN_HASH_SIZE; i++) {
222 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
228 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
230 struct connection *con;
232 mutex_lock(&connections_lock);
233 con = __nodeid2con(nodeid, allocation);
234 mutex_unlock(&connections_lock);
239 /* This is a bit drastic, but only called when things go wrong */
240 static struct connection *assoc2con(int assoc_id)
243 struct hlist_node *h;
244 struct connection *con;
246 mutex_lock(&connections_lock);
248 for (i = 0 ; i < CONN_HASH_SIZE; i++) {
249 hlist_for_each_entry(con, h, &connection_hash[i], list) {
250 if (con && con->sctp_assoc == assoc_id) {
251 mutex_unlock(&connections_lock);
256 mutex_unlock(&connections_lock);
260 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
262 struct sockaddr_storage addr;
265 if (!dlm_local_count)
268 error = dlm_nodeid_to_addr(nodeid, &addr);
272 if (dlm_local_addr[0]->ss_family == AF_INET) {
273 struct sockaddr_in *in4 = (struct sockaddr_in *) &addr;
274 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
275 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
277 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &addr;
278 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
279 ipv6_addr_copy(&ret6->sin6_addr, &in6->sin6_addr);
285 /* Data available on socket or listen socket received a connect */
286 static void lowcomms_data_ready(struct sock *sk, int count_unused)
288 struct connection *con = sock2con(sk);
289 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
290 queue_work(recv_workqueue, &con->rwork);
293 static void lowcomms_write_space(struct sock *sk)
295 struct connection *con = sock2con(sk);
297 if (con && !test_and_set_bit(CF_WRITE_PENDING, &con->flags))
298 queue_work(send_workqueue, &con->swork);
301 static inline void lowcomms_connect_sock(struct connection *con)
303 if (test_bit(CF_CLOSE, &con->flags))
305 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
306 queue_work(send_workqueue, &con->swork);
309 static void lowcomms_state_change(struct sock *sk)
311 if (sk->sk_state == TCP_ESTABLISHED)
312 lowcomms_write_space(sk);
315 int dlm_lowcomms_connect_node(int nodeid)
317 struct connection *con;
319 if (nodeid == dlm_our_nodeid())
322 con = nodeid2con(nodeid, GFP_NOFS);
325 lowcomms_connect_sock(con);
329 /* Make a socket active */
330 static int add_sock(struct socket *sock, struct connection *con)
334 /* Install a data_ready callback */
335 con->sock->sk->sk_data_ready = lowcomms_data_ready;
336 con->sock->sk->sk_write_space = lowcomms_write_space;
337 con->sock->sk->sk_state_change = lowcomms_state_change;
338 con->sock->sk->sk_user_data = con;
339 con->sock->sk->sk_allocation = GFP_NOFS;
343 /* Add the port number to an IPv6 or 4 sockaddr and return the address
345 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
348 saddr->ss_family = dlm_local_addr[0]->ss_family;
349 if (saddr->ss_family == AF_INET) {
350 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
351 in4_addr->sin_port = cpu_to_be16(port);
352 *addr_len = sizeof(struct sockaddr_in);
353 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
355 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
356 in6_addr->sin6_port = cpu_to_be16(port);
357 *addr_len = sizeof(struct sockaddr_in6);
359 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
362 /* Close a remote connection and tidy up */
363 static void close_connection(struct connection *con, bool and_other)
365 mutex_lock(&con->sock_mutex);
368 sock_release(con->sock);
371 if (con->othercon && and_other) {
372 /* Will only re-enter once. */
373 close_connection(con->othercon, false);
376 __free_page(con->rx_page);
381 mutex_unlock(&con->sock_mutex);
384 /* We only send shutdown messages to nodes that are not part of the cluster */
385 static void sctp_send_shutdown(sctp_assoc_t associd)
387 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
388 struct msghdr outmessage;
389 struct cmsghdr *cmsg;
390 struct sctp_sndrcvinfo *sinfo;
392 struct connection *con;
394 con = nodeid2con(0,0);
397 outmessage.msg_name = NULL;
398 outmessage.msg_namelen = 0;
399 outmessage.msg_control = outcmsg;
400 outmessage.msg_controllen = sizeof(outcmsg);
401 outmessage.msg_flags = MSG_EOR;
403 cmsg = CMSG_FIRSTHDR(&outmessage);
404 cmsg->cmsg_level = IPPROTO_SCTP;
405 cmsg->cmsg_type = SCTP_SNDRCV;
406 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
407 outmessage.msg_controllen = cmsg->cmsg_len;
408 sinfo = CMSG_DATA(cmsg);
409 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
411 sinfo->sinfo_flags |= MSG_EOF;
412 sinfo->sinfo_assoc_id = associd;
414 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
417 log_print("send EOF to node failed: %d", ret);
420 static void sctp_init_failed_foreach(struct connection *con)
423 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
424 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
425 queue_work(send_workqueue, &con->swork);
429 /* INIT failed but we don't know which node...
430 restart INIT on all pending nodes */
431 static void sctp_init_failed(void)
433 mutex_lock(&connections_lock);
435 foreach_conn(sctp_init_failed_foreach);
437 mutex_unlock(&connections_lock);
440 /* Something happened to an association */
441 static void process_sctp_notification(struct connection *con,
442 struct msghdr *msg, char *buf)
444 union sctp_notification *sn = (union sctp_notification *)buf;
446 if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
447 switch (sn->sn_assoc_change.sac_state) {
452 /* Check that the new node is in the lockspace */
453 struct sctp_prim prim;
457 struct connection *new_con;
459 sctp_peeloff_arg_t parg;
460 int parglen = sizeof(parg);
463 * We get this before any data for an association.
464 * We verify that the node is in the cluster and
465 * then peel off a socket for it.
467 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
468 log_print("COMM_UP for invalid assoc ID %d",
469 (int)sn->sn_assoc_change.sac_assoc_id);
473 memset(&prim, 0, sizeof(struct sctp_prim));
474 prim_len = sizeof(struct sctp_prim);
475 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
477 ret = kernel_getsockopt(con->sock,
483 log_print("getsockopt/sctp_primary_addr on "
484 "new assoc %d failed : %d",
485 (int)sn->sn_assoc_change.sac_assoc_id,
488 /* Retry INIT later */
489 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
491 clear_bit(CF_CONNECT_PENDING, &con->flags);
494 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
495 if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
497 unsigned char *b=(unsigned char *)&prim.ssp_addr;
498 log_print("reject connect from unknown addr");
499 for (i=0; i<sizeof(struct sockaddr_storage);i++)
500 printk("%02x ", b[i]);
502 sctp_send_shutdown(prim.ssp_assoc_id);
506 new_con = nodeid2con(nodeid, GFP_NOFS);
510 /* Peel off a new sock */
511 parg.associd = sn->sn_assoc_change.sac_assoc_id;
512 ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
513 SCTP_SOCKOPT_PEELOFF,
514 (void *)&parg, &parglen);
516 log_print("Can't peel off a socket for "
517 "connection %d to node %d: err=%d\n",
518 parg.associd, nodeid, ret);
520 file = fget(parg.sd);
521 new_con->sock = SOCKET_I(file->f_dentry->d_inode);
522 add_sock(new_con->sock, new_con);
524 put_unused_fd(parg.sd);
526 log_print("got new/restarted association %d nodeid %d",
527 (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
529 /* Send any pending writes */
530 clear_bit(CF_CONNECT_PENDING, &new_con->flags);
531 clear_bit(CF_INIT_PENDING, &con->flags);
532 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
533 queue_work(send_workqueue, &new_con->swork);
535 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
536 queue_work(recv_workqueue, &new_con->rwork);
541 case SCTP_SHUTDOWN_COMP:
543 con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
550 /* We don't know which INIT failed, so clear the PENDING flags
551 * on them all. if assoc_id is zero then it will then try
554 case SCTP_CANT_STR_ASSOC:
556 log_print("Can't start SCTP association - retrying");
562 log_print("unexpected SCTP assoc change id=%d state=%d",
563 (int)sn->sn_assoc_change.sac_assoc_id,
564 sn->sn_assoc_change.sac_state);
569 /* Data received from remote end */
570 static int receive_from_sock(struct connection *con)
573 struct msghdr msg = {};
577 int call_again_soon = 0;
579 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
581 mutex_lock(&con->sock_mutex);
583 if (con->sock == NULL) {
588 if (con->rx_page == NULL) {
590 * This doesn't need to be atomic, but I think it should
591 * improve performance if it is.
593 con->rx_page = alloc_page(GFP_ATOMIC);
594 if (con->rx_page == NULL)
596 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
599 /* Only SCTP needs these really */
600 memset(&incmsg, 0, sizeof(incmsg));
601 msg.msg_control = incmsg;
602 msg.msg_controllen = sizeof(incmsg);
605 * iov[0] is the bit of the circular buffer between the current end
606 * point (cb.base + cb.len) and the end of the buffer.
608 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
609 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
614 * iov[1] is the bit of the circular buffer between the start of the
615 * buffer and the start of the currently used section (cb.base)
617 if (cbuf_data(&con->cb) >= con->cb.base) {
618 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
619 iov[1].iov_len = con->cb.base;
620 iov[1].iov_base = page_address(con->rx_page);
623 len = iov[0].iov_len + iov[1].iov_len;
625 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
626 MSG_DONTWAIT | MSG_NOSIGNAL);
630 /* Process SCTP notifications */
631 if (msg.msg_flags & MSG_NOTIFICATION) {
632 msg.msg_control = incmsg;
633 msg.msg_controllen = sizeof(incmsg);
635 process_sctp_notification(con, &msg,
636 page_address(con->rx_page) + con->cb.base);
637 mutex_unlock(&con->sock_mutex);
640 BUG_ON(con->nodeid == 0);
644 cbuf_add(&con->cb, ret);
645 ret = dlm_process_incoming_buffer(con->nodeid,
646 page_address(con->rx_page),
647 con->cb.base, con->cb.len,
649 if (ret == -EBADMSG) {
650 log_print("lowcomms: addr=%p, base=%u, len=%u, "
651 "iov_len=%u, iov_base[0]=%p, read=%d",
652 page_address(con->rx_page), con->cb.base, con->cb.len,
653 len, iov[0].iov_base, r);
657 cbuf_eat(&con->cb, ret);
659 if (cbuf_empty(&con->cb) && !call_again_soon) {
660 __free_page(con->rx_page);
666 mutex_unlock(&con->sock_mutex);
670 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
671 queue_work(recv_workqueue, &con->rwork);
672 mutex_unlock(&con->sock_mutex);
676 mutex_unlock(&con->sock_mutex);
677 if (ret != -EAGAIN) {
678 close_connection(con, false);
679 /* Reconnect when there is something to send */
681 /* Don't return success if we really got EOF */
688 /* Listening socket is busy, accept a connection */
689 static int tcp_accept_from_sock(struct connection *con)
692 struct sockaddr_storage peeraddr;
693 struct socket *newsock;
696 struct connection *newcon;
697 struct connection *addcon;
699 memset(&peeraddr, 0, sizeof(peeraddr));
700 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
701 IPPROTO_TCP, &newsock);
705 mutex_lock_nested(&con->sock_mutex, 0);
708 if (con->sock == NULL)
711 newsock->type = con->sock->type;
712 newsock->ops = con->sock->ops;
714 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
718 /* Get the connected socket's peer */
719 memset(&peeraddr, 0, sizeof(peeraddr));
720 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
722 result = -ECONNABORTED;
726 /* Get the new node's NODEID */
727 make_sockaddr(&peeraddr, 0, &len);
728 if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
729 log_print("connect from non cluster node");
730 sock_release(newsock);
731 mutex_unlock(&con->sock_mutex);
735 log_print("got connection from %d", nodeid);
737 /* Check to see if we already have a connection to this node. This
738 * could happen if the two nodes initiate a connection at roughly
739 * the same time and the connections cross on the wire.
740 * In this case we store the incoming one in "othercon"
742 newcon = nodeid2con(nodeid, GFP_NOFS);
747 mutex_lock_nested(&newcon->sock_mutex, 1);
749 struct connection *othercon = newcon->othercon;
752 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
754 log_print("failed to allocate incoming socket");
755 mutex_unlock(&newcon->sock_mutex);
759 othercon->nodeid = nodeid;
760 othercon->rx_action = receive_from_sock;
761 mutex_init(&othercon->sock_mutex);
762 INIT_WORK(&othercon->swork, process_send_sockets);
763 INIT_WORK(&othercon->rwork, process_recv_sockets);
764 set_bit(CF_IS_OTHERCON, &othercon->flags);
766 if (!othercon->sock) {
767 newcon->othercon = othercon;
768 othercon->sock = newsock;
769 newsock->sk->sk_user_data = othercon;
770 add_sock(newsock, othercon);
774 printk("Extra connection from node %d attempted\n", nodeid);
776 mutex_unlock(&newcon->sock_mutex);
781 newsock->sk->sk_user_data = newcon;
782 newcon->rx_action = receive_from_sock;
783 add_sock(newsock, newcon);
787 mutex_unlock(&newcon->sock_mutex);
790 * Add it to the active queue in case we got data
791 * beween processing the accept adding the socket
792 * to the read_sockets list
794 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
795 queue_work(recv_workqueue, &addcon->rwork);
796 mutex_unlock(&con->sock_mutex);
801 mutex_unlock(&con->sock_mutex);
802 sock_release(newsock);
804 if (result != -EAGAIN)
805 log_print("error accepting connection from node: %d", result);
809 static void free_entry(struct writequeue_entry *e)
811 __free_page(e->page);
815 /* Initiate an SCTP association.
816 This is a special case of send_to_sock() in that we don't yet have a
817 peeled-off socket for this association, so we use the listening socket
818 and add the primary IP address of the remote node.
820 static void sctp_init_assoc(struct connection *con)
822 struct sockaddr_storage rem_addr;
823 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
824 struct msghdr outmessage;
825 struct cmsghdr *cmsg;
826 struct sctp_sndrcvinfo *sinfo;
827 struct connection *base_con;
828 struct writequeue_entry *e;
834 if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
837 if (con->retries++ > MAX_CONNECT_RETRIES)
840 log_print("Initiating association with node %d", con->nodeid);
842 if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
843 log_print("no address for nodeid %d", con->nodeid);
846 base_con = nodeid2con(0, 0);
847 BUG_ON(base_con == NULL);
849 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
851 outmessage.msg_name = &rem_addr;
852 outmessage.msg_namelen = addrlen;
853 outmessage.msg_control = outcmsg;
854 outmessage.msg_controllen = sizeof(outcmsg);
855 outmessage.msg_flags = MSG_EOR;
857 spin_lock(&con->writequeue_lock);
858 e = list_entry(con->writequeue.next, struct writequeue_entry,
861 BUG_ON((struct list_head *) e == &con->writequeue);
865 spin_unlock(&con->writequeue_lock);
867 /* Send the first block off the write queue */
868 iov[0].iov_base = page_address(e->page)+offset;
869 iov[0].iov_len = len;
871 cmsg = CMSG_FIRSTHDR(&outmessage);
872 cmsg->cmsg_level = IPPROTO_SCTP;
873 cmsg->cmsg_type = SCTP_SNDRCV;
874 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
875 sinfo = CMSG_DATA(cmsg);
876 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
877 sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
878 outmessage.msg_controllen = cmsg->cmsg_len;
880 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
882 log_print("Send first packet to node %d failed: %d",
885 /* Try again later */
886 clear_bit(CF_CONNECT_PENDING, &con->flags);
887 clear_bit(CF_INIT_PENDING, &con->flags);
890 spin_lock(&con->writequeue_lock);
894 if (e->len == 0 && e->users == 0) {
898 spin_unlock(&con->writequeue_lock);
902 /* Connect a new socket to its peer */
903 static void tcp_connect_to_sock(struct connection *con)
905 int result = -EHOSTUNREACH;
906 struct sockaddr_storage saddr, src_addr;
908 struct socket *sock = NULL;
910 if (con->nodeid == 0) {
911 log_print("attempt to connect sock 0 foiled");
915 mutex_lock(&con->sock_mutex);
916 if (con->retries++ > MAX_CONNECT_RETRIES)
919 /* Some odd races can cause double-connects, ignore them */
925 /* Create a socket to communicate with */
926 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
931 memset(&saddr, 0, sizeof(saddr));
932 if (dlm_nodeid_to_addr(con->nodeid, &saddr))
935 sock->sk->sk_user_data = con;
936 con->rx_action = receive_from_sock;
937 con->connect_action = tcp_connect_to_sock;
940 /* Bind to our cluster-known address connecting to avoid
942 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
943 make_sockaddr(&src_addr, 0, &addr_len);
944 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
947 log_print("could not bind for connect: %d", result);
948 /* This *may* not indicate a critical error */
951 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
953 log_print("connecting to %d", con->nodeid);
955 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
957 if (result == -EINPROGRESS)
964 sock_release(con->sock);
970 * Some errors are fatal and this list might need adjusting. For other
971 * errors we try again until the max number of retries is reached.
973 if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
974 result != -ENETDOWN && result != -EINVAL
975 && result != -EPROTONOSUPPORT) {
976 lowcomms_connect_sock(con);
980 mutex_unlock(&con->sock_mutex);
984 static struct socket *tcp_create_listen_sock(struct connection *con,
985 struct sockaddr_storage *saddr)
987 struct socket *sock = NULL;
992 if (dlm_local_addr[0]->ss_family == AF_INET)
993 addr_len = sizeof(struct sockaddr_in);
995 addr_len = sizeof(struct sockaddr_in6);
997 /* Create a socket to communicate with */
998 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1001 log_print("Can't create listening comms socket");
1005 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1006 (char *)&one, sizeof(one));
1009 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1011 sock->sk->sk_user_data = con;
1012 con->rx_action = tcp_accept_from_sock;
1013 con->connect_action = tcp_connect_to_sock;
1016 /* Bind to our port */
1017 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1018 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1020 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1026 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1027 (char *)&one, sizeof(one));
1029 log_print("Set keepalive failed: %d", result);
1032 result = sock->ops->listen(sock, 5);
1034 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1044 /* Get local addresses */
1045 static void init_local(void)
1047 struct sockaddr_storage sas, *addr;
1050 dlm_local_count = 0;
1051 for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1052 if (dlm_our_addr(&sas, i))
1055 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1058 memcpy(addr, &sas, sizeof(*addr));
1059 dlm_local_addr[dlm_local_count++] = addr;
1063 /* Bind to an IP address. SCTP allows multiple address so it can do
1065 static int add_sctp_bind_addr(struct connection *sctp_con,
1066 struct sockaddr_storage *addr,
1067 int addr_len, int num)
1072 result = kernel_bind(sctp_con->sock,
1073 (struct sockaddr *) addr,
1076 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1077 SCTP_SOCKOPT_BINDX_ADD,
1078 (char *)addr, addr_len);
1081 log_print("Can't bind to port %d addr number %d",
1082 dlm_config.ci_tcp_port, num);
1087 /* Initialise SCTP socket and bind to all interfaces */
1088 static int sctp_listen_for_all(void)
1090 struct socket *sock = NULL;
1091 struct sockaddr_storage localaddr;
1092 struct sctp_event_subscribe subscribe;
1093 int result = -EINVAL, num = 1, i, addr_len;
1094 struct connection *con = nodeid2con(0, GFP_KERNEL);
1095 int bufsize = NEEDED_RMEM;
1100 log_print("Using SCTP for communications");
1102 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1103 IPPROTO_SCTP, &sock);
1105 log_print("Can't create comms socket, check SCTP is loaded");
1109 /* Listen for events */
1110 memset(&subscribe, 0, sizeof(subscribe));
1111 subscribe.sctp_data_io_event = 1;
1112 subscribe.sctp_association_event = 1;
1113 subscribe.sctp_send_failure_event = 1;
1114 subscribe.sctp_shutdown_event = 1;
1115 subscribe.sctp_partial_delivery_event = 1;
1117 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1118 (char *)&bufsize, sizeof(bufsize));
1120 log_print("Error increasing buffer space on socket %d", result);
1122 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1123 (char *)&subscribe, sizeof(subscribe));
1125 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1127 goto create_delsock;
1130 /* Init con struct */
1131 sock->sk->sk_user_data = con;
1133 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1134 con->rx_action = receive_from_sock;
1135 con->connect_action = sctp_init_assoc;
1137 /* Bind to all interfaces. */
1138 for (i = 0; i < dlm_local_count; i++) {
1139 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1140 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1142 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1144 goto create_delsock;
1148 result = sock->ops->listen(sock, 5);
1150 log_print("Can't set socket listening");
1151 goto create_delsock;
1163 static int tcp_listen_for_all(void)
1165 struct socket *sock = NULL;
1166 struct connection *con = nodeid2con(0, GFP_KERNEL);
1167 int result = -EINVAL;
1172 /* We don't support multi-homed hosts */
1173 if (dlm_local_addr[1] != NULL) {
1174 log_print("TCP protocol can't handle multi-homed hosts, "
1179 log_print("Using TCP for communications");
1181 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1183 add_sock(sock, con);
1187 result = -EADDRINUSE;
1195 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1198 struct writequeue_entry *entry;
1200 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1204 entry->page = alloc_page(allocation);
1219 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1221 struct connection *con;
1222 struct writequeue_entry *e;
1226 con = nodeid2con(nodeid, allocation);
1230 spin_lock(&con->writequeue_lock);
1231 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1232 if ((&e->list == &con->writequeue) ||
1233 (PAGE_CACHE_SIZE - e->end < len)) {
1240 spin_unlock(&con->writequeue_lock);
1244 *ppc = page_address(e->page) + offset;
1248 e = new_writequeue_entry(con, allocation);
1250 spin_lock(&con->writequeue_lock);
1254 list_add_tail(&e->list, &con->writequeue);
1255 spin_unlock(&con->writequeue_lock);
1261 void dlm_lowcomms_commit_buffer(void *mh)
1263 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1264 struct connection *con = e->con;
1267 spin_lock(&con->writequeue_lock);
1271 e->len = e->end - e->offset;
1272 spin_unlock(&con->writequeue_lock);
1274 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1275 queue_work(send_workqueue, &con->swork);
1280 spin_unlock(&con->writequeue_lock);
1284 /* Send a message */
1285 static void send_to_sock(struct connection *con)
1288 ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1289 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1290 struct writequeue_entry *e;
1293 mutex_lock(&con->sock_mutex);
1294 if (con->sock == NULL)
1297 sendpage = con->sock->ops->sendpage;
1299 spin_lock(&con->writequeue_lock);
1301 e = list_entry(con->writequeue.next, struct writequeue_entry,
1303 if ((struct list_head *) e == &con->writequeue)
1308 BUG_ON(len == 0 && e->users == 0);
1309 spin_unlock(&con->writequeue_lock);
1313 ret = sendpage(con->sock, e->page, offset, len,
1315 if (ret == -EAGAIN || ret == 0) {
1322 /* Don't starve people filling buffers */
1325 spin_lock(&con->writequeue_lock);
1329 if (e->len == 0 && e->users == 0) {
1335 spin_unlock(&con->writequeue_lock);
1337 mutex_unlock(&con->sock_mutex);
1341 mutex_unlock(&con->sock_mutex);
1342 close_connection(con, false);
1343 lowcomms_connect_sock(con);
1347 mutex_unlock(&con->sock_mutex);
1348 if (!test_bit(CF_INIT_PENDING, &con->flags))
1349 lowcomms_connect_sock(con);
1353 static void clean_one_writequeue(struct connection *con)
1355 struct writequeue_entry *e, *safe;
1357 spin_lock(&con->writequeue_lock);
1358 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1362 spin_unlock(&con->writequeue_lock);
1365 /* Called from recovery when it knows that a node has
1367 int dlm_lowcomms_close(int nodeid)
1369 struct connection *con;
1371 log_print("closing connection to node %d", nodeid);
1372 con = nodeid2con(nodeid, 0);
1374 clear_bit(CF_CONNECT_PENDING, &con->flags);
1375 clear_bit(CF_WRITE_PENDING, &con->flags);
1376 set_bit(CF_CLOSE, &con->flags);
1377 if (cancel_work_sync(&con->swork))
1378 log_print("canceled swork for node %d", nodeid);
1379 if (cancel_work_sync(&con->rwork))
1380 log_print("canceled rwork for node %d", nodeid);
1381 clean_one_writequeue(con);
1382 close_connection(con, true);
1387 /* Receive workqueue function */
1388 static void process_recv_sockets(struct work_struct *work)
1390 struct connection *con = container_of(work, struct connection, rwork);
1393 clear_bit(CF_READ_PENDING, &con->flags);
1395 err = con->rx_action(con);
1399 /* Send workqueue function */
1400 static void process_send_sockets(struct work_struct *work)
1402 struct connection *con = container_of(work, struct connection, swork);
1404 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1405 con->connect_action(con);
1406 set_bit(CF_WRITE_PENDING, &con->flags);
1408 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1413 /* Discard all entries on the write queues */
1414 static void clean_writequeues(void)
1416 foreach_conn(clean_one_writequeue);
1419 static void work_stop(void)
1421 destroy_workqueue(recv_workqueue);
1422 destroy_workqueue(send_workqueue);
1425 static int work_start(void)
1428 recv_workqueue = create_workqueue("dlm_recv");
1429 error = IS_ERR(recv_workqueue);
1431 log_print("can't start dlm_recv %d", error);
1435 send_workqueue = create_singlethread_workqueue("dlm_send");
1436 error = IS_ERR(send_workqueue);
1438 log_print("can't start dlm_send %d", error);
1439 destroy_workqueue(recv_workqueue);
1446 static void stop_conn(struct connection *con)
1449 if (con->sock && con->sock->sk)
1450 con->sock->sk->sk_user_data = NULL;
1453 static void free_conn(struct connection *con)
1455 close_connection(con, true);
1457 kmem_cache_free(con_cache, con->othercon);
1458 hlist_del(&con->list);
1459 kmem_cache_free(con_cache, con);
1462 void dlm_lowcomms_stop(void)
1464 /* Set all the flags to prevent any
1467 mutex_lock(&connections_lock);
1468 foreach_conn(stop_conn);
1469 mutex_unlock(&connections_lock);
1473 mutex_lock(&connections_lock);
1474 clean_writequeues();
1476 foreach_conn(free_conn);
1478 mutex_unlock(&connections_lock);
1479 kmem_cache_destroy(con_cache);
1482 int dlm_lowcomms_start(void)
1484 int error = -EINVAL;
1485 struct connection *con;
1488 for (i = 0; i < CONN_HASH_SIZE; i++)
1489 INIT_HLIST_HEAD(&connection_hash[i]);
1492 if (!dlm_local_count) {
1494 log_print("no local IP address has been set");
1499 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1500 __alignof__(struct connection), 0,
1505 /* Start listening */
1506 if (dlm_config.ci_protocol == 0)
1507 error = tcp_listen_for_all();
1509 error = sctp_listen_for_all();
1513 error = work_start();
1520 con = nodeid2con(0,0);
1522 close_connection(con, false);
1523 kmem_cache_free(con_cache, con);
1525 kmem_cache_destroy(con_cache);