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 struct dlm_node_addr {
144 struct list_head list;
147 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
150 static LIST_HEAD(dlm_node_addrs);
151 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
153 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
154 static int dlm_local_count;
155 static int dlm_allow_conn;
158 static struct workqueue_struct *recv_workqueue;
159 static struct workqueue_struct *send_workqueue;
161 static struct hlist_head connection_hash[CONN_HASH_SIZE];
162 static DEFINE_MUTEX(connections_lock);
163 static struct kmem_cache *con_cache;
165 static void process_recv_sockets(struct work_struct *work);
166 static void process_send_sockets(struct work_struct *work);
169 /* This is deliberately very simple because most clusters have simple
170 sequential nodeids, so we should be able to go straight to a connection
171 struct in the array */
172 static inline int nodeid_hash(int nodeid)
174 return nodeid & (CONN_HASH_SIZE-1);
177 static struct connection *__find_con(int nodeid)
180 struct connection *con;
182 r = nodeid_hash(nodeid);
184 hlist_for_each_entry(con, &connection_hash[r], list) {
185 if (con->nodeid == nodeid)
192 * If 'allocation' is zero then we don't attempt to create a new
193 * connection structure for this node.
195 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
197 struct connection *con = NULL;
200 con = __find_con(nodeid);
204 con = kmem_cache_zalloc(con_cache, alloc);
208 r = nodeid_hash(nodeid);
209 hlist_add_head(&con->list, &connection_hash[r]);
211 con->nodeid = nodeid;
212 mutex_init(&con->sock_mutex);
213 INIT_LIST_HEAD(&con->writequeue);
214 spin_lock_init(&con->writequeue_lock);
215 INIT_WORK(&con->swork, process_send_sockets);
216 INIT_WORK(&con->rwork, process_recv_sockets);
218 /* Setup action pointers for child sockets */
220 struct connection *zerocon = __find_con(0);
222 con->connect_action = zerocon->connect_action;
224 con->rx_action = zerocon->rx_action;
230 /* Loop round all connections */
231 static void foreach_conn(void (*conn_func)(struct connection *c))
234 struct hlist_node *n;
235 struct connection *con;
237 for (i = 0; i < CONN_HASH_SIZE; i++) {
238 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
243 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
245 struct connection *con;
247 mutex_lock(&connections_lock);
248 con = __nodeid2con(nodeid, allocation);
249 mutex_unlock(&connections_lock);
254 /* This is a bit drastic, but only called when things go wrong */
255 static struct connection *assoc2con(int assoc_id)
258 struct connection *con;
260 mutex_lock(&connections_lock);
262 for (i = 0 ; i < CONN_HASH_SIZE; i++) {
263 hlist_for_each_entry(con, &connection_hash[i], list) {
264 if (con->sctp_assoc == assoc_id) {
265 mutex_unlock(&connections_lock);
270 mutex_unlock(&connections_lock);
274 static struct dlm_node_addr *find_node_addr(int nodeid)
276 struct dlm_node_addr *na;
278 list_for_each_entry(na, &dlm_node_addrs, list) {
279 if (na->nodeid == nodeid)
285 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
287 switch (x->ss_family) {
289 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
290 struct sockaddr_in *siny = (struct sockaddr_in *)y;
291 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
293 if (sinx->sin_port != siny->sin_port)
298 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
299 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
300 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
302 if (sinx->sin6_port != siny->sin6_port)
312 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
313 struct sockaddr *sa_out)
315 struct sockaddr_storage sas;
316 struct dlm_node_addr *na;
318 if (!dlm_local_count)
321 spin_lock(&dlm_node_addrs_spin);
322 na = find_node_addr(nodeid);
323 if (na && na->addr_count)
324 memcpy(&sas, na->addr[0], sizeof(struct sockaddr_storage));
325 spin_unlock(&dlm_node_addrs_spin);
334 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
339 if (dlm_local_addr[0]->ss_family == AF_INET) {
340 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas;
341 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
342 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
344 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas;
345 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
346 ret6->sin6_addr = in6->sin6_addr;
352 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
354 struct dlm_node_addr *na;
357 spin_lock(&dlm_node_addrs_spin);
358 list_for_each_entry(na, &dlm_node_addrs, list) {
362 if (!addr_compare(na->addr[0], addr))
365 *nodeid = na->nodeid;
369 spin_unlock(&dlm_node_addrs_spin);
373 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
375 struct sockaddr_storage *new_addr;
376 struct dlm_node_addr *new_node, *na;
378 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
382 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
388 memcpy(new_addr, addr, len);
390 spin_lock(&dlm_node_addrs_spin);
391 na = find_node_addr(nodeid);
393 new_node->nodeid = nodeid;
394 new_node->addr[0] = new_addr;
395 new_node->addr_count = 1;
396 list_add(&new_node->list, &dlm_node_addrs);
397 spin_unlock(&dlm_node_addrs_spin);
401 if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
402 spin_unlock(&dlm_node_addrs_spin);
408 na->addr[na->addr_count++] = new_addr;
409 spin_unlock(&dlm_node_addrs_spin);
414 /* Data available on socket or listen socket received a connect */
415 static void lowcomms_data_ready(struct sock *sk, int count_unused)
417 struct connection *con = sock2con(sk);
418 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
419 queue_work(recv_workqueue, &con->rwork);
422 static void lowcomms_write_space(struct sock *sk)
424 struct connection *con = sock2con(sk);
429 clear_bit(SOCK_NOSPACE, &con->sock->flags);
431 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
432 con->sock->sk->sk_write_pending--;
433 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
436 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
437 queue_work(send_workqueue, &con->swork);
440 static inline void lowcomms_connect_sock(struct connection *con)
442 if (test_bit(CF_CLOSE, &con->flags))
444 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
445 queue_work(send_workqueue, &con->swork);
448 static void lowcomms_state_change(struct sock *sk)
450 if (sk->sk_state == TCP_ESTABLISHED)
451 lowcomms_write_space(sk);
454 int dlm_lowcomms_connect_node(int nodeid)
456 struct connection *con;
458 /* with sctp there's no connecting without sending */
459 if (dlm_config.ci_protocol != 0)
462 if (nodeid == dlm_our_nodeid())
465 con = nodeid2con(nodeid, GFP_NOFS);
468 lowcomms_connect_sock(con);
472 /* Make a socket active */
473 static void add_sock(struct socket *sock, struct connection *con)
477 /* Install a data_ready callback */
478 con->sock->sk->sk_data_ready = lowcomms_data_ready;
479 con->sock->sk->sk_write_space = lowcomms_write_space;
480 con->sock->sk->sk_state_change = lowcomms_state_change;
481 con->sock->sk->sk_user_data = con;
482 con->sock->sk->sk_allocation = GFP_NOFS;
485 /* Add the port number to an IPv6 or 4 sockaddr and return the address
487 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
490 saddr->ss_family = dlm_local_addr[0]->ss_family;
491 if (saddr->ss_family == AF_INET) {
492 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
493 in4_addr->sin_port = cpu_to_be16(port);
494 *addr_len = sizeof(struct sockaddr_in);
495 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
497 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
498 in6_addr->sin6_port = cpu_to_be16(port);
499 *addr_len = sizeof(struct sockaddr_in6);
501 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
504 /* Close a remote connection and tidy up */
505 static void close_connection(struct connection *con, bool and_other)
507 mutex_lock(&con->sock_mutex);
510 sock_release(con->sock);
513 if (con->othercon && and_other) {
514 /* Will only re-enter once. */
515 close_connection(con->othercon, false);
518 __free_page(con->rx_page);
523 mutex_unlock(&con->sock_mutex);
526 /* We only send shutdown messages to nodes that are not part of the cluster */
527 static void sctp_send_shutdown(sctp_assoc_t associd)
529 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
530 struct msghdr outmessage;
531 struct cmsghdr *cmsg;
532 struct sctp_sndrcvinfo *sinfo;
534 struct connection *con;
536 con = nodeid2con(0,0);
539 outmessage.msg_name = NULL;
540 outmessage.msg_namelen = 0;
541 outmessage.msg_control = outcmsg;
542 outmessage.msg_controllen = sizeof(outcmsg);
543 outmessage.msg_flags = MSG_EOR;
545 cmsg = CMSG_FIRSTHDR(&outmessage);
546 cmsg->cmsg_level = IPPROTO_SCTP;
547 cmsg->cmsg_type = SCTP_SNDRCV;
548 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
549 outmessage.msg_controllen = cmsg->cmsg_len;
550 sinfo = CMSG_DATA(cmsg);
551 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
553 sinfo->sinfo_flags |= MSG_EOF;
554 sinfo->sinfo_assoc_id = associd;
556 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
559 log_print("send EOF to node failed: %d", ret);
562 static void sctp_init_failed_foreach(struct connection *con)
565 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
566 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
567 queue_work(send_workqueue, &con->swork);
571 /* INIT failed but we don't know which node...
572 restart INIT on all pending nodes */
573 static void sctp_init_failed(void)
575 mutex_lock(&connections_lock);
577 foreach_conn(sctp_init_failed_foreach);
579 mutex_unlock(&connections_lock);
582 /* Something happened to an association */
583 static void process_sctp_notification(struct connection *con,
584 struct msghdr *msg, char *buf)
586 union sctp_notification *sn = (union sctp_notification *)buf;
588 if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
589 switch (sn->sn_assoc_change.sac_state) {
594 /* Check that the new node is in the lockspace */
595 struct sctp_prim prim;
599 struct connection *new_con;
602 * We get this before any data for an association.
603 * We verify that the node is in the cluster and
604 * then peel off a socket for it.
606 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
607 log_print("COMM_UP for invalid assoc ID %d",
608 (int)sn->sn_assoc_change.sac_assoc_id);
612 memset(&prim, 0, sizeof(struct sctp_prim));
613 prim_len = sizeof(struct sctp_prim);
614 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
616 ret = kernel_getsockopt(con->sock,
622 log_print("getsockopt/sctp_primary_addr on "
623 "new assoc %d failed : %d",
624 (int)sn->sn_assoc_change.sac_assoc_id,
627 /* Retry INIT later */
628 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
630 clear_bit(CF_CONNECT_PENDING, &con->flags);
633 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
634 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
635 unsigned char *b=(unsigned char *)&prim.ssp_addr;
636 log_print("reject connect from unknown addr");
637 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
638 b, sizeof(struct sockaddr_storage));
639 sctp_send_shutdown(prim.ssp_assoc_id);
643 new_con = nodeid2con(nodeid, GFP_NOFS);
647 /* Peel off a new sock */
648 sctp_lock_sock(con->sock->sk);
649 ret = sctp_do_peeloff(con->sock->sk,
650 sn->sn_assoc_change.sac_assoc_id,
652 sctp_release_sock(con->sock->sk);
654 log_print("Can't peel off a socket for "
655 "connection %d to node %d: err=%d",
656 (int)sn->sn_assoc_change.sac_assoc_id,
660 add_sock(new_con->sock, new_con);
662 log_print("connecting to %d sctp association %d",
663 nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
665 /* Send any pending writes */
666 clear_bit(CF_CONNECT_PENDING, &new_con->flags);
667 clear_bit(CF_INIT_PENDING, &con->flags);
668 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
669 queue_work(send_workqueue, &new_con->swork);
671 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
672 queue_work(recv_workqueue, &new_con->rwork);
677 case SCTP_SHUTDOWN_COMP:
679 con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
686 /* We don't know which INIT failed, so clear the PENDING flags
687 * on them all. if assoc_id is zero then it will then try
690 case SCTP_CANT_STR_ASSOC:
692 log_print("Can't start SCTP association - retrying");
698 log_print("unexpected SCTP assoc change id=%d state=%d",
699 (int)sn->sn_assoc_change.sac_assoc_id,
700 sn->sn_assoc_change.sac_state);
705 /* Data received from remote end */
706 static int receive_from_sock(struct connection *con)
709 struct msghdr msg = {};
713 int call_again_soon = 0;
715 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
717 mutex_lock(&con->sock_mutex);
719 if (con->sock == NULL) {
724 if (con->rx_page == NULL) {
726 * This doesn't need to be atomic, but I think it should
727 * improve performance if it is.
729 con->rx_page = alloc_page(GFP_ATOMIC);
730 if (con->rx_page == NULL)
732 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
735 /* Only SCTP needs these really */
736 memset(&incmsg, 0, sizeof(incmsg));
737 msg.msg_control = incmsg;
738 msg.msg_controllen = sizeof(incmsg);
741 * iov[0] is the bit of the circular buffer between the current end
742 * point (cb.base + cb.len) and the end of the buffer.
744 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
745 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
750 * iov[1] is the bit of the circular buffer between the start of the
751 * buffer and the start of the currently used section (cb.base)
753 if (cbuf_data(&con->cb) >= con->cb.base) {
754 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
755 iov[1].iov_len = con->cb.base;
756 iov[1].iov_base = page_address(con->rx_page);
759 len = iov[0].iov_len + iov[1].iov_len;
761 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
762 MSG_DONTWAIT | MSG_NOSIGNAL);
766 /* Process SCTP notifications */
767 if (msg.msg_flags & MSG_NOTIFICATION) {
768 msg.msg_control = incmsg;
769 msg.msg_controllen = sizeof(incmsg);
771 process_sctp_notification(con, &msg,
772 page_address(con->rx_page) + con->cb.base);
773 mutex_unlock(&con->sock_mutex);
776 BUG_ON(con->nodeid == 0);
780 cbuf_add(&con->cb, ret);
781 ret = dlm_process_incoming_buffer(con->nodeid,
782 page_address(con->rx_page),
783 con->cb.base, con->cb.len,
785 if (ret == -EBADMSG) {
786 log_print("lowcomms: addr=%p, base=%u, len=%u, "
787 "iov_len=%u, iov_base[0]=%p, read=%d",
788 page_address(con->rx_page), con->cb.base, con->cb.len,
789 len, iov[0].iov_base, r);
793 cbuf_eat(&con->cb, ret);
795 if (cbuf_empty(&con->cb) && !call_again_soon) {
796 __free_page(con->rx_page);
802 mutex_unlock(&con->sock_mutex);
806 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
807 queue_work(recv_workqueue, &con->rwork);
808 mutex_unlock(&con->sock_mutex);
812 mutex_unlock(&con->sock_mutex);
813 if (ret != -EAGAIN) {
814 close_connection(con, false);
815 /* Reconnect when there is something to send */
817 /* Don't return success if we really got EOF */
824 /* Listening socket is busy, accept a connection */
825 static int tcp_accept_from_sock(struct connection *con)
828 struct sockaddr_storage peeraddr;
829 struct socket *newsock;
832 struct connection *newcon;
833 struct connection *addcon;
835 mutex_lock(&connections_lock);
836 if (!dlm_allow_conn) {
837 mutex_unlock(&connections_lock);
840 mutex_unlock(&connections_lock);
842 memset(&peeraddr, 0, sizeof(peeraddr));
843 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
844 IPPROTO_TCP, &newsock);
848 mutex_lock_nested(&con->sock_mutex, 0);
851 if (con->sock == NULL)
854 newsock->type = con->sock->type;
855 newsock->ops = con->sock->ops;
857 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
861 /* Get the connected socket's peer */
862 memset(&peeraddr, 0, sizeof(peeraddr));
863 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
865 result = -ECONNABORTED;
869 /* Get the new node's NODEID */
870 make_sockaddr(&peeraddr, 0, &len);
871 if (addr_to_nodeid(&peeraddr, &nodeid)) {
872 unsigned char *b=(unsigned char *)&peeraddr;
873 log_print("connect from non cluster node");
874 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
875 b, sizeof(struct sockaddr_storage));
876 sock_release(newsock);
877 mutex_unlock(&con->sock_mutex);
881 log_print("got connection from %d", nodeid);
883 /* Check to see if we already have a connection to this node. This
884 * could happen if the two nodes initiate a connection at roughly
885 * the same time and the connections cross on the wire.
886 * In this case we store the incoming one in "othercon"
888 newcon = nodeid2con(nodeid, GFP_NOFS);
893 mutex_lock_nested(&newcon->sock_mutex, 1);
895 struct connection *othercon = newcon->othercon;
898 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
900 log_print("failed to allocate incoming socket");
901 mutex_unlock(&newcon->sock_mutex);
905 othercon->nodeid = nodeid;
906 othercon->rx_action = receive_from_sock;
907 mutex_init(&othercon->sock_mutex);
908 INIT_WORK(&othercon->swork, process_send_sockets);
909 INIT_WORK(&othercon->rwork, process_recv_sockets);
910 set_bit(CF_IS_OTHERCON, &othercon->flags);
912 if (!othercon->sock) {
913 newcon->othercon = othercon;
914 othercon->sock = newsock;
915 newsock->sk->sk_user_data = othercon;
916 add_sock(newsock, othercon);
920 printk("Extra connection from node %d attempted\n", nodeid);
922 mutex_unlock(&newcon->sock_mutex);
927 newsock->sk->sk_user_data = newcon;
928 newcon->rx_action = receive_from_sock;
929 add_sock(newsock, newcon);
933 mutex_unlock(&newcon->sock_mutex);
936 * Add it to the active queue in case we got data
937 * between processing the accept adding the socket
938 * to the read_sockets list
940 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
941 queue_work(recv_workqueue, &addcon->rwork);
942 mutex_unlock(&con->sock_mutex);
947 mutex_unlock(&con->sock_mutex);
948 sock_release(newsock);
950 if (result != -EAGAIN)
951 log_print("error accepting connection from node: %d", result);
955 static void free_entry(struct writequeue_entry *e)
957 __free_page(e->page);
961 /* Initiate an SCTP association.
962 This is a special case of send_to_sock() in that we don't yet have a
963 peeled-off socket for this association, so we use the listening socket
964 and add the primary IP address of the remote node.
966 static void sctp_init_assoc(struct connection *con)
968 struct sockaddr_storage rem_addr;
969 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
970 struct msghdr outmessage;
971 struct cmsghdr *cmsg;
972 struct sctp_sndrcvinfo *sinfo;
973 struct connection *base_con;
974 struct writequeue_entry *e;
980 if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
983 if (con->retries++ > MAX_CONNECT_RETRIES)
986 if (nodeid_to_addr(con->nodeid, NULL, (struct sockaddr *)&rem_addr)) {
987 log_print("no address for nodeid %d", con->nodeid);
990 base_con = nodeid2con(0, 0);
991 BUG_ON(base_con == NULL);
993 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
995 outmessage.msg_name = &rem_addr;
996 outmessage.msg_namelen = addrlen;
997 outmessage.msg_control = outcmsg;
998 outmessage.msg_controllen = sizeof(outcmsg);
999 outmessage.msg_flags = MSG_EOR;
1001 spin_lock(&con->writequeue_lock);
1003 if (list_empty(&con->writequeue)) {
1004 spin_unlock(&con->writequeue_lock);
1005 log_print("writequeue empty for nodeid %d", con->nodeid);
1009 e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1012 spin_unlock(&con->writequeue_lock);
1014 /* Send the first block off the write queue */
1015 iov[0].iov_base = page_address(e->page)+offset;
1016 iov[0].iov_len = len;
1018 cmsg = CMSG_FIRSTHDR(&outmessage);
1019 cmsg->cmsg_level = IPPROTO_SCTP;
1020 cmsg->cmsg_type = SCTP_SNDRCV;
1021 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
1022 sinfo = CMSG_DATA(cmsg);
1023 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
1024 sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
1025 outmessage.msg_controllen = cmsg->cmsg_len;
1027 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
1029 log_print("Send first packet to node %d failed: %d",
1032 /* Try again later */
1033 clear_bit(CF_CONNECT_PENDING, &con->flags);
1034 clear_bit(CF_INIT_PENDING, &con->flags);
1037 spin_lock(&con->writequeue_lock);
1041 if (e->len == 0 && e->users == 0) {
1045 spin_unlock(&con->writequeue_lock);
1049 /* Connect a new socket to its peer */
1050 static void tcp_connect_to_sock(struct connection *con)
1052 struct sockaddr_storage saddr, src_addr;
1054 struct socket *sock = NULL;
1058 if (con->nodeid == 0) {
1059 log_print("attempt to connect sock 0 foiled");
1063 mutex_lock(&con->sock_mutex);
1064 if (con->retries++ > MAX_CONNECT_RETRIES)
1067 /* Some odd races can cause double-connects, ignore them */
1071 /* Create a socket to communicate with */
1072 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1073 IPPROTO_TCP, &sock);
1077 memset(&saddr, 0, sizeof(saddr));
1078 result = nodeid_to_addr(con->nodeid, &saddr, NULL);
1080 log_print("no address for nodeid %d", con->nodeid);
1084 sock->sk->sk_user_data = con;
1085 con->rx_action = receive_from_sock;
1086 con->connect_action = tcp_connect_to_sock;
1087 add_sock(sock, con);
1089 /* Bind to our cluster-known address connecting to avoid
1091 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1092 make_sockaddr(&src_addr, 0, &addr_len);
1093 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1096 log_print("could not bind for connect: %d", result);
1097 /* This *may* not indicate a critical error */
1100 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1102 log_print("connecting to %d", con->nodeid);
1104 /* Turn off Nagle's algorithm */
1105 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1108 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1110 if (result == -EINPROGRESS)
1117 sock_release(con->sock);
1123 * Some errors are fatal and this list might need adjusting. For other
1124 * errors we try again until the max number of retries is reached.
1126 if (result != -EHOSTUNREACH &&
1127 result != -ENETUNREACH &&
1128 result != -ENETDOWN &&
1129 result != -EINVAL &&
1130 result != -EPROTONOSUPPORT) {
1131 log_print("connect %d try %d error %d", con->nodeid,
1132 con->retries, result);
1133 mutex_unlock(&con->sock_mutex);
1135 lowcomms_connect_sock(con);
1139 mutex_unlock(&con->sock_mutex);
1143 static struct socket *tcp_create_listen_sock(struct connection *con,
1144 struct sockaddr_storage *saddr)
1146 struct socket *sock = NULL;
1151 if (dlm_local_addr[0]->ss_family == AF_INET)
1152 addr_len = sizeof(struct sockaddr_in);
1154 addr_len = sizeof(struct sockaddr_in6);
1156 /* Create a socket to communicate with */
1157 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1158 IPPROTO_TCP, &sock);
1160 log_print("Can't create listening comms socket");
1164 /* Turn off Nagle's algorithm */
1165 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1168 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1169 (char *)&one, sizeof(one));
1172 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1174 con->rx_action = tcp_accept_from_sock;
1175 con->connect_action = tcp_connect_to_sock;
1177 /* Bind to our port */
1178 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1179 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1181 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1187 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1188 (char *)&one, sizeof(one));
1190 log_print("Set keepalive failed: %d", result);
1193 result = sock->ops->listen(sock, 5);
1195 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1205 /* Get local addresses */
1206 static void init_local(void)
1208 struct sockaddr_storage sas, *addr;
1211 dlm_local_count = 0;
1212 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1213 if (dlm_our_addr(&sas, i))
1216 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1219 memcpy(addr, &sas, sizeof(*addr));
1220 dlm_local_addr[dlm_local_count++] = addr;
1224 /* Bind to an IP address. SCTP allows multiple address so it can do
1226 static int add_sctp_bind_addr(struct connection *sctp_con,
1227 struct sockaddr_storage *addr,
1228 int addr_len, int num)
1233 result = kernel_bind(sctp_con->sock,
1234 (struct sockaddr *) addr,
1237 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1238 SCTP_SOCKOPT_BINDX_ADD,
1239 (char *)addr, addr_len);
1242 log_print("Can't bind to port %d addr number %d",
1243 dlm_config.ci_tcp_port, num);
1248 /* Initialise SCTP socket and bind to all interfaces */
1249 static int sctp_listen_for_all(void)
1251 struct socket *sock = NULL;
1252 struct sockaddr_storage localaddr;
1253 struct sctp_event_subscribe subscribe;
1254 int result = -EINVAL, num = 1, i, addr_len;
1255 struct connection *con = nodeid2con(0, GFP_NOFS);
1256 int bufsize = NEEDED_RMEM;
1261 log_print("Using SCTP for communications");
1263 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1264 IPPROTO_SCTP, &sock);
1266 log_print("Can't create comms socket, check SCTP is loaded");
1270 /* Listen for events */
1271 memset(&subscribe, 0, sizeof(subscribe));
1272 subscribe.sctp_data_io_event = 1;
1273 subscribe.sctp_association_event = 1;
1274 subscribe.sctp_send_failure_event = 1;
1275 subscribe.sctp_shutdown_event = 1;
1276 subscribe.sctp_partial_delivery_event = 1;
1278 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1279 (char *)&bufsize, sizeof(bufsize));
1281 log_print("Error increasing buffer space on socket %d", result);
1283 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1284 (char *)&subscribe, sizeof(subscribe));
1286 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1288 goto create_delsock;
1291 /* Init con struct */
1292 sock->sk->sk_user_data = con;
1294 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1295 con->rx_action = receive_from_sock;
1296 con->connect_action = sctp_init_assoc;
1298 /* Bind to all interfaces. */
1299 for (i = 0; i < dlm_local_count; i++) {
1300 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1301 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1303 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1305 goto create_delsock;
1309 result = sock->ops->listen(sock, 5);
1311 log_print("Can't set socket listening");
1312 goto create_delsock;
1324 static int tcp_listen_for_all(void)
1326 struct socket *sock = NULL;
1327 struct connection *con = nodeid2con(0, GFP_NOFS);
1328 int result = -EINVAL;
1333 /* We don't support multi-homed hosts */
1334 if (dlm_local_addr[1] != NULL) {
1335 log_print("TCP protocol can't handle multi-homed hosts, "
1340 log_print("Using TCP for communications");
1342 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1344 add_sock(sock, con);
1348 result = -EADDRINUSE;
1356 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1359 struct writequeue_entry *entry;
1361 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1365 entry->page = alloc_page(allocation);
1380 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1382 struct connection *con;
1383 struct writequeue_entry *e;
1386 con = nodeid2con(nodeid, allocation);
1390 spin_lock(&con->writequeue_lock);
1391 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1392 if ((&e->list == &con->writequeue) ||
1393 (PAGE_CACHE_SIZE - e->end < len)) {
1400 spin_unlock(&con->writequeue_lock);
1404 *ppc = page_address(e->page) + offset;
1408 e = new_writequeue_entry(con, allocation);
1410 spin_lock(&con->writequeue_lock);
1414 list_add_tail(&e->list, &con->writequeue);
1415 spin_unlock(&con->writequeue_lock);
1421 void dlm_lowcomms_commit_buffer(void *mh)
1423 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1424 struct connection *con = e->con;
1427 spin_lock(&con->writequeue_lock);
1431 e->len = e->end - e->offset;
1432 spin_unlock(&con->writequeue_lock);
1434 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1435 queue_work(send_workqueue, &con->swork);
1440 spin_unlock(&con->writequeue_lock);
1444 /* Send a message */
1445 static void send_to_sock(struct connection *con)
1448 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1449 struct writequeue_entry *e;
1453 mutex_lock(&con->sock_mutex);
1454 if (con->sock == NULL)
1457 spin_lock(&con->writequeue_lock);
1459 e = list_entry(con->writequeue.next, struct writequeue_entry,
1461 if ((struct list_head *) e == &con->writequeue)
1466 BUG_ON(len == 0 && e->users == 0);
1467 spin_unlock(&con->writequeue_lock);
1471 ret = kernel_sendpage(con->sock, e->page, offset, len,
1473 if (ret == -EAGAIN || ret == 0) {
1474 if (ret == -EAGAIN &&
1475 test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1476 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1477 /* Notify TCP that we're limited by the
1478 * application window size.
1480 set_bit(SOCK_NOSPACE, &con->sock->flags);
1481 con->sock->sk->sk_write_pending++;
1489 /* Don't starve people filling buffers */
1490 if (++count >= MAX_SEND_MSG_COUNT) {
1495 spin_lock(&con->writequeue_lock);
1499 if (e->len == 0 && e->users == 0) {
1504 spin_unlock(&con->writequeue_lock);
1506 mutex_unlock(&con->sock_mutex);
1510 mutex_unlock(&con->sock_mutex);
1511 close_connection(con, false);
1512 lowcomms_connect_sock(con);
1516 mutex_unlock(&con->sock_mutex);
1517 if (!test_bit(CF_INIT_PENDING, &con->flags))
1518 lowcomms_connect_sock(con);
1521 static void clean_one_writequeue(struct connection *con)
1523 struct writequeue_entry *e, *safe;
1525 spin_lock(&con->writequeue_lock);
1526 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1530 spin_unlock(&con->writequeue_lock);
1533 /* Called from recovery when it knows that a node has
1535 int dlm_lowcomms_close(int nodeid)
1537 struct connection *con;
1538 struct dlm_node_addr *na;
1540 log_print("closing connection to node %d", nodeid);
1541 con = nodeid2con(nodeid, 0);
1543 clear_bit(CF_CONNECT_PENDING, &con->flags);
1544 clear_bit(CF_WRITE_PENDING, &con->flags);
1545 set_bit(CF_CLOSE, &con->flags);
1546 if (cancel_work_sync(&con->swork))
1547 log_print("canceled swork for node %d", nodeid);
1548 if (cancel_work_sync(&con->rwork))
1549 log_print("canceled rwork for node %d", nodeid);
1550 clean_one_writequeue(con);
1551 close_connection(con, true);
1554 spin_lock(&dlm_node_addrs_spin);
1555 na = find_node_addr(nodeid);
1557 list_del(&na->list);
1558 while (na->addr_count--)
1559 kfree(na->addr[na->addr_count]);
1562 spin_unlock(&dlm_node_addrs_spin);
1567 /* Receive workqueue function */
1568 static void process_recv_sockets(struct work_struct *work)
1570 struct connection *con = container_of(work, struct connection, rwork);
1573 clear_bit(CF_READ_PENDING, &con->flags);
1575 err = con->rx_action(con);
1579 /* Send workqueue function */
1580 static void process_send_sockets(struct work_struct *work)
1582 struct connection *con = container_of(work, struct connection, swork);
1584 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1585 con->connect_action(con);
1586 set_bit(CF_WRITE_PENDING, &con->flags);
1588 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1593 /* Discard all entries on the write queues */
1594 static void clean_writequeues(void)
1596 foreach_conn(clean_one_writequeue);
1599 static void work_stop(void)
1601 destroy_workqueue(recv_workqueue);
1602 destroy_workqueue(send_workqueue);
1605 static int work_start(void)
1607 recv_workqueue = alloc_workqueue("dlm_recv",
1608 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1609 if (!recv_workqueue) {
1610 log_print("can't start dlm_recv");
1614 send_workqueue = alloc_workqueue("dlm_send",
1615 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1616 if (!send_workqueue) {
1617 log_print("can't start dlm_send");
1618 destroy_workqueue(recv_workqueue);
1625 static void stop_conn(struct connection *con)
1628 if (con->sock && con->sock->sk)
1629 con->sock->sk->sk_user_data = NULL;
1632 static void free_conn(struct connection *con)
1634 close_connection(con, true);
1636 kmem_cache_free(con_cache, con->othercon);
1637 hlist_del(&con->list);
1638 kmem_cache_free(con_cache, con);
1641 void dlm_lowcomms_stop(void)
1643 /* Set all the flags to prevent any
1646 mutex_lock(&connections_lock);
1648 foreach_conn(stop_conn);
1649 mutex_unlock(&connections_lock);
1653 mutex_lock(&connections_lock);
1654 clean_writequeues();
1656 foreach_conn(free_conn);
1658 mutex_unlock(&connections_lock);
1659 kmem_cache_destroy(con_cache);
1662 int dlm_lowcomms_start(void)
1664 int error = -EINVAL;
1665 struct connection *con;
1668 for (i = 0; i < CONN_HASH_SIZE; i++)
1669 INIT_HLIST_HEAD(&connection_hash[i]);
1672 if (!dlm_local_count) {
1674 log_print("no local IP address has been set");
1679 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1680 __alignof__(struct connection), 0,
1685 error = work_start();
1691 /* Start listening */
1692 if (dlm_config.ci_protocol == 0)
1693 error = tcp_listen_for_all();
1695 error = sctp_listen_for_all();
1703 con = nodeid2con(0,0);
1705 close_connection(con, false);
1706 kmem_cache_free(con_cache, con);
1709 kmem_cache_destroy(con_cache);
1714 void dlm_lowcomms_exit(void)
1716 struct dlm_node_addr *na, *safe;
1718 spin_lock(&dlm_node_addrs_spin);
1719 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1720 list_del(&na->list);
1721 while (na->addr_count--)
1722 kfree(na->addr[na->addr_count]);
1725 spin_unlock(&dlm_node_addrs_spin);