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>
58 #include "dlm_internal.h"
63 #define NEEDED_RMEM (4*1024*1024)
64 #define CONN_HASH_SIZE 32
66 /* Number of messages to send before rescheduling */
67 #define MAX_SEND_MSG_COUNT 25
75 static void cbuf_add(struct cbuf *cb, int n)
80 static int cbuf_data(struct cbuf *cb)
82 return ((cb->base + cb->len) & cb->mask);
85 static void cbuf_init(struct cbuf *cb, int size)
87 cb->base = cb->len = 0;
91 static void cbuf_eat(struct cbuf *cb, int n)
98 static bool cbuf_empty(struct cbuf *cb)
104 struct socket *sock; /* NULL if not connected */
105 uint32_t nodeid; /* So we know who we are in the list */
106 struct mutex sock_mutex;
108 #define CF_READ_PENDING 1
109 #define CF_WRITE_PENDING 2
110 #define CF_CONNECT_PENDING 3
111 #define CF_INIT_PENDING 4
112 #define CF_IS_OTHERCON 5
114 #define CF_APP_LIMITED 7
115 struct list_head writequeue; /* List of outgoing writequeue_entries */
116 spinlock_t writequeue_lock;
117 int (*rx_action) (struct connection *); /* What to do when active */
118 void (*connect_action) (struct connection *); /* What to do to connect */
119 struct page *rx_page;
122 #define MAX_CONNECT_RETRIES 3
123 struct hlist_node list;
124 struct connection *othercon;
125 struct work_struct rwork; /* Receive workqueue */
126 struct work_struct swork; /* Send workqueue */
128 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
130 /* An entry waiting to be sent */
131 struct writequeue_entry {
132 struct list_head list;
138 struct connection *con;
141 struct dlm_node_addr {
142 struct list_head list;
146 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
149 static LIST_HEAD(dlm_node_addrs);
150 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
152 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
153 static int dlm_local_count;
154 static int dlm_allow_conn;
157 static struct workqueue_struct *recv_workqueue;
158 static struct workqueue_struct *send_workqueue;
160 static struct hlist_head connection_hash[CONN_HASH_SIZE];
161 static DEFINE_MUTEX(connections_lock);
162 static struct kmem_cache *con_cache;
164 static void process_recv_sockets(struct work_struct *work);
165 static void process_send_sockets(struct work_struct *work);
168 /* This is deliberately very simple because most clusters have simple
169 sequential nodeids, so we should be able to go straight to a connection
170 struct in the array */
171 static inline int nodeid_hash(int nodeid)
173 return nodeid & (CONN_HASH_SIZE-1);
176 static struct connection *__find_con(int nodeid)
179 struct connection *con;
181 r = nodeid_hash(nodeid);
183 hlist_for_each_entry(con, &connection_hash[r], list) {
184 if (con->nodeid == nodeid)
191 * If 'allocation' is zero then we don't attempt to create a new
192 * connection structure for this node.
194 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
196 struct connection *con = NULL;
199 con = __find_con(nodeid);
203 con = kmem_cache_zalloc(con_cache, alloc);
207 r = nodeid_hash(nodeid);
208 hlist_add_head(&con->list, &connection_hash[r]);
210 con->nodeid = nodeid;
211 mutex_init(&con->sock_mutex);
212 INIT_LIST_HEAD(&con->writequeue);
213 spin_lock_init(&con->writequeue_lock);
214 INIT_WORK(&con->swork, process_send_sockets);
215 INIT_WORK(&con->rwork, process_recv_sockets);
217 /* Setup action pointers for child sockets */
219 struct connection *zerocon = __find_con(0);
221 con->connect_action = zerocon->connect_action;
223 con->rx_action = zerocon->rx_action;
229 /* Loop round all connections */
230 static void foreach_conn(void (*conn_func)(struct connection *c))
233 struct hlist_node *n;
234 struct connection *con;
236 for (i = 0; i < CONN_HASH_SIZE; i++) {
237 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
242 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
244 struct connection *con;
246 mutex_lock(&connections_lock);
247 con = __nodeid2con(nodeid, allocation);
248 mutex_unlock(&connections_lock);
253 static struct dlm_node_addr *find_node_addr(int nodeid)
255 struct dlm_node_addr *na;
257 list_for_each_entry(na, &dlm_node_addrs, list) {
258 if (na->nodeid == nodeid)
264 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
266 switch (x->ss_family) {
268 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
269 struct sockaddr_in *siny = (struct sockaddr_in *)y;
270 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
272 if (sinx->sin_port != siny->sin_port)
277 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
278 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
279 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
281 if (sinx->sin6_port != siny->sin6_port)
291 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
292 struct sockaddr *sa_out, bool try_new_addr)
294 struct sockaddr_storage sas;
295 struct dlm_node_addr *na;
297 if (!dlm_local_count)
300 spin_lock(&dlm_node_addrs_spin);
301 na = find_node_addr(nodeid);
302 if (na && na->addr_count) {
303 memcpy(&sas, na->addr[na->curr_addr_index],
304 sizeof(struct sockaddr_storage));
307 na->curr_addr_index++;
308 if (na->curr_addr_index == na->addr_count)
309 na->curr_addr_index = 0;
312 spin_unlock(&dlm_node_addrs_spin);
321 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
326 if (dlm_local_addr[0]->ss_family == AF_INET) {
327 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas;
328 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
329 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
331 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas;
332 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
333 ret6->sin6_addr = in6->sin6_addr;
339 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
341 struct dlm_node_addr *na;
345 spin_lock(&dlm_node_addrs_spin);
346 list_for_each_entry(na, &dlm_node_addrs, list) {
350 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
351 if (addr_compare(na->addr[addr_i], addr)) {
352 *nodeid = na->nodeid;
359 spin_unlock(&dlm_node_addrs_spin);
363 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
365 struct sockaddr_storage *new_addr;
366 struct dlm_node_addr *new_node, *na;
368 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
372 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
378 memcpy(new_addr, addr, len);
380 spin_lock(&dlm_node_addrs_spin);
381 na = find_node_addr(nodeid);
383 new_node->nodeid = nodeid;
384 new_node->addr[0] = new_addr;
385 new_node->addr_count = 1;
386 list_add(&new_node->list, &dlm_node_addrs);
387 spin_unlock(&dlm_node_addrs_spin);
391 if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
392 spin_unlock(&dlm_node_addrs_spin);
398 na->addr[na->addr_count++] = new_addr;
399 spin_unlock(&dlm_node_addrs_spin);
404 /* Data available on socket or listen socket received a connect */
405 static void lowcomms_data_ready(struct sock *sk)
407 struct connection *con = sock2con(sk);
408 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
409 queue_work(recv_workqueue, &con->rwork);
412 static void lowcomms_write_space(struct sock *sk)
414 struct connection *con = sock2con(sk);
419 clear_bit(SOCK_NOSPACE, &con->sock->flags);
421 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
422 con->sock->sk->sk_write_pending--;
423 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
426 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
427 queue_work(send_workqueue, &con->swork);
430 static inline void lowcomms_connect_sock(struct connection *con)
432 if (test_bit(CF_CLOSE, &con->flags))
434 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
435 queue_work(send_workqueue, &con->swork);
438 static void lowcomms_state_change(struct sock *sk)
440 /* SCTP layer is not calling sk_data_ready when the connection
441 * is done, so we catch the signal through here. Also, it
442 * doesn't switch socket state when entering shutdown, so we
443 * skip the write in that case.
445 if (sk->sk_shutdown) {
446 if (sk->sk_shutdown == RCV_SHUTDOWN)
447 lowcomms_data_ready(sk);
448 } else if (sk->sk_state == TCP_ESTABLISHED) {
449 lowcomms_write_space(sk);
453 int dlm_lowcomms_connect_node(int nodeid)
455 struct connection *con;
457 if (nodeid == dlm_our_nodeid())
460 con = nodeid2con(nodeid, GFP_NOFS);
463 lowcomms_connect_sock(con);
467 /* Make a socket active */
468 static void add_sock(struct socket *sock, struct connection *con)
472 /* Install a data_ready callback */
473 con->sock->sk->sk_data_ready = lowcomms_data_ready;
474 con->sock->sk->sk_write_space = lowcomms_write_space;
475 con->sock->sk->sk_state_change = lowcomms_state_change;
476 con->sock->sk->sk_user_data = con;
477 con->sock->sk->sk_allocation = GFP_NOFS;
480 /* Add the port number to an IPv6 or 4 sockaddr and return the address
482 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
485 saddr->ss_family = dlm_local_addr[0]->ss_family;
486 if (saddr->ss_family == AF_INET) {
487 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
488 in4_addr->sin_port = cpu_to_be16(port);
489 *addr_len = sizeof(struct sockaddr_in);
490 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
492 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
493 in6_addr->sin6_port = cpu_to_be16(port);
494 *addr_len = sizeof(struct sockaddr_in6);
496 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
499 /* Close a remote connection and tidy up */
500 static void close_connection(struct connection *con, bool and_other,
503 clear_bit(CF_CONNECT_PENDING, &con->flags);
504 clear_bit(CF_WRITE_PENDING, &con->flags);
505 if (tx && cancel_work_sync(&con->swork))
506 log_print("canceled swork for node %d", con->nodeid);
507 if (rx && cancel_work_sync(&con->rwork))
508 log_print("canceled rwork for node %d", con->nodeid);
510 mutex_lock(&con->sock_mutex);
512 sock_release(con->sock);
515 if (con->othercon && and_other) {
516 /* Will only re-enter once. */
517 close_connection(con->othercon, false, true, true);
520 __free_page(con->rx_page);
525 mutex_unlock(&con->sock_mutex);
528 /* Data received from remote end */
529 static int receive_from_sock(struct connection *con)
532 struct msghdr msg = {};
536 int call_again_soon = 0;
539 mutex_lock(&con->sock_mutex);
541 if (con->sock == NULL) {
546 if (con->rx_page == NULL) {
548 * This doesn't need to be atomic, but I think it should
549 * improve performance if it is.
551 con->rx_page = alloc_page(GFP_ATOMIC);
552 if (con->rx_page == NULL)
554 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
558 * iov[0] is the bit of the circular buffer between the current end
559 * point (cb.base + cb.len) and the end of the buffer.
561 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
562 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
567 * iov[1] is the bit of the circular buffer between the start of the
568 * buffer and the start of the currently used section (cb.base)
570 if (cbuf_data(&con->cb) >= con->cb.base) {
571 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
572 iov[1].iov_len = con->cb.base;
573 iov[1].iov_base = page_address(con->rx_page);
576 len = iov[0].iov_len + iov[1].iov_len;
578 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
579 MSG_DONTWAIT | MSG_NOSIGNAL);
585 BUG_ON(con->nodeid == 0);
587 cbuf_add(&con->cb, ret);
588 ret = dlm_process_incoming_buffer(con->nodeid,
589 page_address(con->rx_page),
590 con->cb.base, con->cb.len,
592 if (ret == -EBADMSG) {
593 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d",
594 page_address(con->rx_page), con->cb.base,
599 cbuf_eat(&con->cb, ret);
601 if (cbuf_empty(&con->cb) && !call_again_soon) {
602 __free_page(con->rx_page);
608 mutex_unlock(&con->sock_mutex);
612 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
613 queue_work(recv_workqueue, &con->rwork);
614 mutex_unlock(&con->sock_mutex);
618 mutex_unlock(&con->sock_mutex);
619 if (ret != -EAGAIN) {
620 close_connection(con, false, true, false);
621 /* Reconnect when there is something to send */
623 /* Don't return success if we really got EOF */
630 /* Listening socket is busy, accept a connection */
631 static int tcp_accept_from_sock(struct connection *con)
634 struct sockaddr_storage peeraddr;
635 struct socket *newsock;
638 struct connection *newcon;
639 struct connection *addcon;
641 mutex_lock(&connections_lock);
642 if (!dlm_allow_conn) {
643 mutex_unlock(&connections_lock);
646 mutex_unlock(&connections_lock);
648 memset(&peeraddr, 0, sizeof(peeraddr));
649 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
650 SOCK_STREAM, IPPROTO_TCP, &newsock);
654 mutex_lock_nested(&con->sock_mutex, 0);
657 if (con->sock == NULL)
660 newsock->type = con->sock->type;
661 newsock->ops = con->sock->ops;
663 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
667 /* Get the connected socket's peer */
668 memset(&peeraddr, 0, sizeof(peeraddr));
669 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
671 result = -ECONNABORTED;
675 /* Get the new node's NODEID */
676 make_sockaddr(&peeraddr, 0, &len);
677 if (addr_to_nodeid(&peeraddr, &nodeid)) {
678 unsigned char *b=(unsigned char *)&peeraddr;
679 log_print("connect from non cluster node");
680 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
681 b, sizeof(struct sockaddr_storage));
682 sock_release(newsock);
683 mutex_unlock(&con->sock_mutex);
687 log_print("got connection from %d", nodeid);
689 /* Check to see if we already have a connection to this node. This
690 * could happen if the two nodes initiate a connection at roughly
691 * the same time and the connections cross on the wire.
692 * In this case we store the incoming one in "othercon"
694 newcon = nodeid2con(nodeid, GFP_NOFS);
699 mutex_lock_nested(&newcon->sock_mutex, 1);
701 struct connection *othercon = newcon->othercon;
704 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
706 log_print("failed to allocate incoming socket");
707 mutex_unlock(&newcon->sock_mutex);
711 othercon->nodeid = nodeid;
712 othercon->rx_action = receive_from_sock;
713 mutex_init(&othercon->sock_mutex);
714 INIT_WORK(&othercon->swork, process_send_sockets);
715 INIT_WORK(&othercon->rwork, process_recv_sockets);
716 set_bit(CF_IS_OTHERCON, &othercon->flags);
718 if (!othercon->sock) {
719 newcon->othercon = othercon;
720 othercon->sock = newsock;
721 newsock->sk->sk_user_data = othercon;
722 add_sock(newsock, othercon);
726 printk("Extra connection from node %d attempted\n", nodeid);
728 mutex_unlock(&newcon->sock_mutex);
733 newsock->sk->sk_user_data = newcon;
734 newcon->rx_action = receive_from_sock;
735 add_sock(newsock, newcon);
739 mutex_unlock(&newcon->sock_mutex);
742 * Add it to the active queue in case we got data
743 * between processing the accept adding the socket
744 * to the read_sockets list
746 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
747 queue_work(recv_workqueue, &addcon->rwork);
748 mutex_unlock(&con->sock_mutex);
753 mutex_unlock(&con->sock_mutex);
754 sock_release(newsock);
756 if (result != -EAGAIN)
757 log_print("error accepting connection from node: %d", result);
761 int sctp_accept_from_sock(struct connection *con)
763 /* Check that the new node is in the lockspace */
764 struct sctp_prim prim;
768 struct connection *newcon;
769 struct connection *addcon;
770 struct socket *newsock;
772 mutex_lock(&connections_lock);
773 if (!dlm_allow_conn) {
774 mutex_unlock(&connections_lock);
777 mutex_unlock(&connections_lock);
779 mutex_lock_nested(&con->sock_mutex, 0);
781 ret = kernel_accept(con->sock, &newsock, O_NONBLOCK);
785 memset(&prim, 0, sizeof(struct sctp_prim));
786 prim_len = sizeof(struct sctp_prim);
788 ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR,
789 (char *)&prim, &prim_len);
791 log_print("getsockopt/sctp_primary_addr failed: %d", ret);
795 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
796 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
797 unsigned char *b = (unsigned char *)&prim.ssp_addr;
799 log_print("reject connect from unknown addr");
800 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
801 b, sizeof(struct sockaddr_storage));
805 newcon = nodeid2con(nodeid, GFP_NOFS);
811 mutex_lock_nested(&newcon->sock_mutex, 1);
814 struct connection *othercon = newcon->othercon;
817 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
819 log_print("failed to allocate incoming socket");
820 mutex_unlock(&newcon->sock_mutex);
824 othercon->nodeid = nodeid;
825 othercon->rx_action = receive_from_sock;
826 mutex_init(&othercon->sock_mutex);
827 INIT_WORK(&othercon->swork, process_send_sockets);
828 INIT_WORK(&othercon->rwork, process_recv_sockets);
829 set_bit(CF_IS_OTHERCON, &othercon->flags);
831 if (!othercon->sock) {
832 newcon->othercon = othercon;
833 othercon->sock = newsock;
834 newsock->sk->sk_user_data = othercon;
835 add_sock(newsock, othercon);
838 printk("Extra connection from node %d attempted\n", nodeid);
840 mutex_unlock(&newcon->sock_mutex);
844 newsock->sk->sk_user_data = newcon;
845 newcon->rx_action = receive_from_sock;
846 add_sock(newsock, newcon);
850 log_print("connected to %d", nodeid);
852 mutex_unlock(&newcon->sock_mutex);
855 * Add it to the active queue in case we got data
856 * between processing the accept adding the socket
857 * to the read_sockets list
859 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
860 queue_work(recv_workqueue, &addcon->rwork);
861 mutex_unlock(&con->sock_mutex);
866 mutex_unlock(&con->sock_mutex);
868 sock_release(newsock);
870 log_print("error accepting connection from node: %d", ret);
875 static void free_entry(struct writequeue_entry *e)
877 __free_page(e->page);
882 * writequeue_entry_complete - try to delete and free write queue entry
883 * @e: write queue entry to try to delete
884 * @completed: bytes completed
886 * writequeue_lock must be held.
888 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
890 e->offset += completed;
893 if (e->len == 0 && e->users == 0) {
900 * sctp_bind_addrs - bind a SCTP socket to all our addresses
902 static int sctp_bind_addrs(struct connection *con, uint16_t port)
904 struct sockaddr_storage localaddr;
905 int i, addr_len, result = 0;
907 for (i = 0; i < dlm_local_count; i++) {
908 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
909 make_sockaddr(&localaddr, port, &addr_len);
912 result = kernel_bind(con->sock,
913 (struct sockaddr *)&localaddr,
916 result = kernel_setsockopt(con->sock, SOL_SCTP,
917 SCTP_SOCKOPT_BINDX_ADD,
918 (char *)&localaddr, addr_len);
921 log_print("Can't bind to %d addr number %d, %d.\n",
922 port, i + 1, result);
929 /* Initiate an SCTP association.
930 This is a special case of send_to_sock() in that we don't yet have a
931 peeled-off socket for this association, so we use the listening socket
932 and add the primary IP address of the remote node.
934 static void sctp_connect_to_sock(struct connection *con)
936 struct sockaddr_storage daddr;
942 if (con->nodeid == 0) {
943 log_print("attempt to connect sock 0 foiled");
947 mutex_lock(&con->sock_mutex);
949 /* Some odd races can cause double-connects, ignore them */
950 if (con->retries++ > MAX_CONNECT_RETRIES)
954 log_print("node %d already connected.", con->nodeid);
958 memset(&daddr, 0, sizeof(daddr));
959 result = nodeid_to_addr(con->nodeid, &daddr, NULL, true);
961 log_print("no address for nodeid %d", con->nodeid);
965 /* Create a socket to communicate with */
966 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
967 SOCK_STREAM, IPPROTO_SCTP, &sock);
971 sock->sk->sk_user_data = con;
972 con->rx_action = receive_from_sock;
973 con->connect_action = sctp_connect_to_sock;
976 /* Bind to all addresses. */
977 if (sctp_bind_addrs(con, 0))
980 make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
982 log_print("connecting to %d", con->nodeid);
984 /* Turn off Nagle's algorithm */
985 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
988 result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
990 if (result == -EINPROGRESS)
1002 * Some errors are fatal and this list might need adjusting. For other
1003 * errors we try again until the max number of retries is reached.
1005 if (result != -EHOSTUNREACH &&
1006 result != -ENETUNREACH &&
1007 result != -ENETDOWN &&
1008 result != -EINVAL &&
1009 result != -EPROTONOSUPPORT) {
1010 log_print("connect %d try %d error %d", con->nodeid,
1011 con->retries, result);
1012 mutex_unlock(&con->sock_mutex);
1014 clear_bit(CF_CONNECT_PENDING, &con->flags);
1015 lowcomms_connect_sock(con);
1020 mutex_unlock(&con->sock_mutex);
1023 /* Connect a new socket to its peer */
1024 static void tcp_connect_to_sock(struct connection *con)
1026 struct sockaddr_storage saddr, src_addr;
1028 struct socket *sock = NULL;
1032 if (con->nodeid == 0) {
1033 log_print("attempt to connect sock 0 foiled");
1037 mutex_lock(&con->sock_mutex);
1038 if (con->retries++ > MAX_CONNECT_RETRIES)
1041 /* Some odd races can cause double-connects, ignore them */
1045 /* Create a socket to communicate with */
1046 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1047 SOCK_STREAM, IPPROTO_TCP, &sock);
1051 memset(&saddr, 0, sizeof(saddr));
1052 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1054 log_print("no address for nodeid %d", con->nodeid);
1058 sock->sk->sk_user_data = con;
1059 con->rx_action = receive_from_sock;
1060 con->connect_action = tcp_connect_to_sock;
1061 add_sock(sock, con);
1063 /* Bind to our cluster-known address connecting to avoid
1065 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1066 make_sockaddr(&src_addr, 0, &addr_len);
1067 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1070 log_print("could not bind for connect: %d", result);
1071 /* This *may* not indicate a critical error */
1074 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1076 log_print("connecting to %d", con->nodeid);
1078 /* Turn off Nagle's algorithm */
1079 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1082 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1084 if (result == -EINPROGRESS)
1091 sock_release(con->sock);
1097 * Some errors are fatal and this list might need adjusting. For other
1098 * errors we try again until the max number of retries is reached.
1100 if (result != -EHOSTUNREACH &&
1101 result != -ENETUNREACH &&
1102 result != -ENETDOWN &&
1103 result != -EINVAL &&
1104 result != -EPROTONOSUPPORT) {
1105 log_print("connect %d try %d error %d", con->nodeid,
1106 con->retries, result);
1107 mutex_unlock(&con->sock_mutex);
1109 clear_bit(CF_CONNECT_PENDING, &con->flags);
1110 lowcomms_connect_sock(con);
1114 mutex_unlock(&con->sock_mutex);
1118 static struct socket *tcp_create_listen_sock(struct connection *con,
1119 struct sockaddr_storage *saddr)
1121 struct socket *sock = NULL;
1126 if (dlm_local_addr[0]->ss_family == AF_INET)
1127 addr_len = sizeof(struct sockaddr_in);
1129 addr_len = sizeof(struct sockaddr_in6);
1131 /* Create a socket to communicate with */
1132 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1133 SOCK_STREAM, IPPROTO_TCP, &sock);
1135 log_print("Can't create listening comms socket");
1139 /* Turn off Nagle's algorithm */
1140 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1143 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1144 (char *)&one, sizeof(one));
1147 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1149 con->rx_action = tcp_accept_from_sock;
1150 con->connect_action = tcp_connect_to_sock;
1152 /* Bind to our port */
1153 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1154 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1156 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1162 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1163 (char *)&one, sizeof(one));
1165 log_print("Set keepalive failed: %d", result);
1168 result = sock->ops->listen(sock, 5);
1170 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1180 /* Get local addresses */
1181 static void init_local(void)
1183 struct sockaddr_storage sas, *addr;
1186 dlm_local_count = 0;
1187 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1188 if (dlm_our_addr(&sas, i))
1191 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1194 memcpy(addr, &sas, sizeof(*addr));
1195 dlm_local_addr[dlm_local_count++] = addr;
1199 /* Initialise SCTP socket and bind to all interfaces */
1200 static int sctp_listen_for_all(void)
1202 struct socket *sock = NULL;
1203 int result = -EINVAL;
1204 struct connection *con = nodeid2con(0, GFP_NOFS);
1205 int bufsize = NEEDED_RMEM;
1211 log_print("Using SCTP for communications");
1213 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1214 SOCK_STREAM, IPPROTO_SCTP, &sock);
1216 log_print("Can't create comms socket, check SCTP is loaded");
1220 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1221 (char *)&bufsize, sizeof(bufsize));
1223 log_print("Error increasing buffer space on socket %d", result);
1225 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1228 log_print("Could not set SCTP NODELAY error %d\n", result);
1230 /* Init con struct */
1231 sock->sk->sk_user_data = con;
1233 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1234 con->rx_action = sctp_accept_from_sock;
1235 con->connect_action = sctp_connect_to_sock;
1237 /* Bind to all addresses. */
1238 if (sctp_bind_addrs(con, dlm_config.ci_tcp_port))
1239 goto create_delsock;
1241 result = sock->ops->listen(sock, 5);
1243 log_print("Can't set socket listening");
1244 goto create_delsock;
1256 static int tcp_listen_for_all(void)
1258 struct socket *sock = NULL;
1259 struct connection *con = nodeid2con(0, GFP_NOFS);
1260 int result = -EINVAL;
1265 /* We don't support multi-homed hosts */
1266 if (dlm_local_addr[1] != NULL) {
1267 log_print("TCP protocol can't handle multi-homed hosts, "
1272 log_print("Using TCP for communications");
1274 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1276 add_sock(sock, con);
1280 result = -EADDRINUSE;
1288 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1291 struct writequeue_entry *entry;
1293 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1297 entry->page = alloc_page(allocation);
1312 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1314 struct connection *con;
1315 struct writequeue_entry *e;
1318 con = nodeid2con(nodeid, allocation);
1322 spin_lock(&con->writequeue_lock);
1323 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1324 if ((&e->list == &con->writequeue) ||
1325 (PAGE_CACHE_SIZE - e->end < len)) {
1332 spin_unlock(&con->writequeue_lock);
1336 *ppc = page_address(e->page) + offset;
1340 e = new_writequeue_entry(con, allocation);
1342 spin_lock(&con->writequeue_lock);
1346 list_add_tail(&e->list, &con->writequeue);
1347 spin_unlock(&con->writequeue_lock);
1353 void dlm_lowcomms_commit_buffer(void *mh)
1355 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1356 struct connection *con = e->con;
1359 spin_lock(&con->writequeue_lock);
1363 e->len = e->end - e->offset;
1364 spin_unlock(&con->writequeue_lock);
1366 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1367 queue_work(send_workqueue, &con->swork);
1372 spin_unlock(&con->writequeue_lock);
1376 /* Send a message */
1377 static void send_to_sock(struct connection *con)
1380 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1381 struct writequeue_entry *e;
1385 mutex_lock(&con->sock_mutex);
1386 if (con->sock == NULL)
1389 spin_lock(&con->writequeue_lock);
1391 e = list_entry(con->writequeue.next, struct writequeue_entry,
1393 if ((struct list_head *) e == &con->writequeue)
1398 BUG_ON(len == 0 && e->users == 0);
1399 spin_unlock(&con->writequeue_lock);
1403 ret = kernel_sendpage(con->sock, e->page, offset, len,
1405 if (ret == -EAGAIN || ret == 0) {
1406 if (ret == -EAGAIN &&
1407 test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1408 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1409 /* Notify TCP that we're limited by the
1410 * application window size.
1412 set_bit(SOCK_NOSPACE, &con->sock->flags);
1413 con->sock->sk->sk_write_pending++;
1421 /* Don't starve people filling buffers */
1422 if (++count >= MAX_SEND_MSG_COUNT) {
1427 spin_lock(&con->writequeue_lock);
1428 writequeue_entry_complete(e, ret);
1430 spin_unlock(&con->writequeue_lock);
1432 mutex_unlock(&con->sock_mutex);
1436 mutex_unlock(&con->sock_mutex);
1437 close_connection(con, false, false, true);
1438 lowcomms_connect_sock(con);
1442 mutex_unlock(&con->sock_mutex);
1443 lowcomms_connect_sock(con);
1446 static void clean_one_writequeue(struct connection *con)
1448 struct writequeue_entry *e, *safe;
1450 spin_lock(&con->writequeue_lock);
1451 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1455 spin_unlock(&con->writequeue_lock);
1458 /* Called from recovery when it knows that a node has
1460 int dlm_lowcomms_close(int nodeid)
1462 struct connection *con;
1463 struct dlm_node_addr *na;
1465 log_print("closing connection to node %d", nodeid);
1466 con = nodeid2con(nodeid, 0);
1468 set_bit(CF_CLOSE, &con->flags);
1469 close_connection(con, true, true, true);
1470 clean_one_writequeue(con);
1473 spin_lock(&dlm_node_addrs_spin);
1474 na = find_node_addr(nodeid);
1476 list_del(&na->list);
1477 while (na->addr_count--)
1478 kfree(na->addr[na->addr_count]);
1481 spin_unlock(&dlm_node_addrs_spin);
1486 /* Receive workqueue function */
1487 static void process_recv_sockets(struct work_struct *work)
1489 struct connection *con = container_of(work, struct connection, rwork);
1492 clear_bit(CF_READ_PENDING, &con->flags);
1494 err = con->rx_action(con);
1498 /* Send workqueue function */
1499 static void process_send_sockets(struct work_struct *work)
1501 struct connection *con = container_of(work, struct connection, swork);
1503 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1504 con->connect_action(con);
1505 set_bit(CF_WRITE_PENDING, &con->flags);
1507 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1512 /* Discard all entries on the write queues */
1513 static void clean_writequeues(void)
1515 foreach_conn(clean_one_writequeue);
1518 static void work_stop(void)
1520 destroy_workqueue(recv_workqueue);
1521 destroy_workqueue(send_workqueue);
1524 static int work_start(void)
1526 recv_workqueue = alloc_workqueue("dlm_recv",
1527 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1528 if (!recv_workqueue) {
1529 log_print("can't start dlm_recv");
1533 send_workqueue = alloc_workqueue("dlm_send",
1534 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1535 if (!send_workqueue) {
1536 log_print("can't start dlm_send");
1537 destroy_workqueue(recv_workqueue);
1544 static void stop_conn(struct connection *con)
1547 if (con->sock && con->sock->sk)
1548 con->sock->sk->sk_user_data = NULL;
1551 static void free_conn(struct connection *con)
1553 close_connection(con, true, true, true);
1555 kmem_cache_free(con_cache, con->othercon);
1556 hlist_del(&con->list);
1557 kmem_cache_free(con_cache, con);
1560 void dlm_lowcomms_stop(void)
1562 /* Set all the flags to prevent any
1565 mutex_lock(&connections_lock);
1567 foreach_conn(stop_conn);
1568 mutex_unlock(&connections_lock);
1572 mutex_lock(&connections_lock);
1573 clean_writequeues();
1575 foreach_conn(free_conn);
1577 mutex_unlock(&connections_lock);
1578 kmem_cache_destroy(con_cache);
1581 int dlm_lowcomms_start(void)
1583 int error = -EINVAL;
1584 struct connection *con;
1587 for (i = 0; i < CONN_HASH_SIZE; i++)
1588 INIT_HLIST_HEAD(&connection_hash[i]);
1591 if (!dlm_local_count) {
1593 log_print("no local IP address has been set");
1598 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1599 __alignof__(struct connection), 0,
1604 error = work_start();
1610 /* Start listening */
1611 if (dlm_config.ci_protocol == 0)
1612 error = tcp_listen_for_all();
1614 error = sctp_listen_for_all();
1622 con = nodeid2con(0,0);
1624 close_connection(con, false, true, true);
1625 kmem_cache_free(con_cache, con);
1628 kmem_cache_destroy(con_cache);
1633 void dlm_lowcomms_exit(void)
1635 struct dlm_node_addr *na, *safe;
1637 spin_lock(&dlm_node_addrs_spin);
1638 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1639 list_del(&na->list);
1640 while (na->addr_count--)
1641 kfree(na->addr[na->addr_count]);
1644 spin_unlock(&dlm_node_addrs_spin);