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_INIT_PENDING 4
110 #define CF_IS_OTHERCON 5
112 #define CF_APP_LIMITED 7
113 struct list_head writequeue; /* List of outgoing writequeue_entries */
114 spinlock_t writequeue_lock;
115 int (*rx_action) (struct connection *); /* What to do when active */
116 void (*connect_action) (struct connection *); /* What to do to connect */
117 struct page *rx_page;
120 #define MAX_CONNECT_RETRIES 3
121 struct hlist_node list;
122 struct connection *othercon;
123 struct work_struct rwork; /* Receive workqueue */
124 struct work_struct swork; /* Send workqueue */
126 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
128 /* An entry waiting to be sent */
129 struct writequeue_entry {
130 struct list_head list;
136 struct connection *con;
139 struct dlm_node_addr {
140 struct list_head list;
144 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
147 static struct listen_sock_callbacks {
148 void (*sk_error_report)(struct sock *);
149 void (*sk_data_ready)(struct sock *);
150 void (*sk_state_change)(struct sock *);
151 void (*sk_write_space)(struct sock *);
154 static LIST_HEAD(dlm_node_addrs);
155 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
157 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
158 static int dlm_local_count;
159 static int dlm_allow_conn;
162 static struct workqueue_struct *recv_workqueue;
163 static struct workqueue_struct *send_workqueue;
165 static struct hlist_head connection_hash[CONN_HASH_SIZE];
166 static DEFINE_MUTEX(connections_lock);
167 static struct kmem_cache *con_cache;
169 static void process_recv_sockets(struct work_struct *work);
170 static void process_send_sockets(struct work_struct *work);
173 /* This is deliberately very simple because most clusters have simple
174 sequential nodeids, so we should be able to go straight to a connection
175 struct in the array */
176 static inline int nodeid_hash(int nodeid)
178 return nodeid & (CONN_HASH_SIZE-1);
181 static struct connection *__find_con(int nodeid)
184 struct connection *con;
186 r = nodeid_hash(nodeid);
188 hlist_for_each_entry(con, &connection_hash[r], list) {
189 if (con->nodeid == nodeid)
196 * If 'allocation' is zero then we don't attempt to create a new
197 * connection structure for this node.
199 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
201 struct connection *con = NULL;
204 con = __find_con(nodeid);
208 con = kmem_cache_zalloc(con_cache, alloc);
212 r = nodeid_hash(nodeid);
213 hlist_add_head(&con->list, &connection_hash[r]);
215 con->nodeid = nodeid;
216 mutex_init(&con->sock_mutex);
217 INIT_LIST_HEAD(&con->writequeue);
218 spin_lock_init(&con->writequeue_lock);
219 INIT_WORK(&con->swork, process_send_sockets);
220 INIT_WORK(&con->rwork, process_recv_sockets);
222 /* Setup action pointers for child sockets */
224 struct connection *zerocon = __find_con(0);
226 con->connect_action = zerocon->connect_action;
228 con->rx_action = zerocon->rx_action;
234 /* Loop round all connections */
235 static void foreach_conn(void (*conn_func)(struct connection *c))
238 struct hlist_node *n;
239 struct connection *con;
241 for (i = 0; i < CONN_HASH_SIZE; i++) {
242 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
247 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
249 struct connection *con;
251 mutex_lock(&connections_lock);
252 con = __nodeid2con(nodeid, allocation);
253 mutex_unlock(&connections_lock);
258 static struct dlm_node_addr *find_node_addr(int nodeid)
260 struct dlm_node_addr *na;
262 list_for_each_entry(na, &dlm_node_addrs, list) {
263 if (na->nodeid == nodeid)
269 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
271 switch (x->ss_family) {
273 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
274 struct sockaddr_in *siny = (struct sockaddr_in *)y;
275 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
277 if (sinx->sin_port != siny->sin_port)
282 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
283 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
284 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
286 if (sinx->sin6_port != siny->sin6_port)
296 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
297 struct sockaddr *sa_out, bool try_new_addr)
299 struct sockaddr_storage sas;
300 struct dlm_node_addr *na;
302 if (!dlm_local_count)
305 spin_lock(&dlm_node_addrs_spin);
306 na = find_node_addr(nodeid);
307 if (na && na->addr_count) {
308 memcpy(&sas, na->addr[na->curr_addr_index],
309 sizeof(struct sockaddr_storage));
312 na->curr_addr_index++;
313 if (na->curr_addr_index == na->addr_count)
314 na->curr_addr_index = 0;
317 spin_unlock(&dlm_node_addrs_spin);
326 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
331 if (dlm_local_addr[0]->ss_family == AF_INET) {
332 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas;
333 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
334 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
336 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas;
337 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
338 ret6->sin6_addr = in6->sin6_addr;
344 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
346 struct dlm_node_addr *na;
350 spin_lock(&dlm_node_addrs_spin);
351 list_for_each_entry(na, &dlm_node_addrs, list) {
355 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
356 if (addr_compare(na->addr[addr_i], addr)) {
357 *nodeid = na->nodeid;
364 spin_unlock(&dlm_node_addrs_spin);
368 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
370 struct sockaddr_storage *new_addr;
371 struct dlm_node_addr *new_node, *na;
373 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
377 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
383 memcpy(new_addr, addr, len);
385 spin_lock(&dlm_node_addrs_spin);
386 na = find_node_addr(nodeid);
388 new_node->nodeid = nodeid;
389 new_node->addr[0] = new_addr;
390 new_node->addr_count = 1;
391 list_add(&new_node->list, &dlm_node_addrs);
392 spin_unlock(&dlm_node_addrs_spin);
396 if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
397 spin_unlock(&dlm_node_addrs_spin);
403 na->addr[na->addr_count++] = new_addr;
404 spin_unlock(&dlm_node_addrs_spin);
409 /* Data available on socket or listen socket received a connect */
410 static void lowcomms_data_ready(struct sock *sk)
412 struct connection *con = sock2con(sk);
413 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
414 queue_work(recv_workqueue, &con->rwork);
417 static void lowcomms_write_space(struct sock *sk)
419 struct connection *con = sock2con(sk);
424 clear_bit(SOCK_NOSPACE, &con->sock->flags);
426 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
427 con->sock->sk->sk_write_pending--;
428 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
431 queue_work(send_workqueue, &con->swork);
434 static inline void lowcomms_connect_sock(struct connection *con)
436 if (test_bit(CF_CLOSE, &con->flags))
438 queue_work(send_workqueue, &con->swork);
442 static void lowcomms_state_change(struct sock *sk)
444 /* SCTP layer is not calling sk_data_ready when the connection
445 * is done, so we catch the signal through here. Also, it
446 * doesn't switch socket state when entering shutdown, so we
447 * skip the write in that case.
449 if (sk->sk_shutdown) {
450 if (sk->sk_shutdown == RCV_SHUTDOWN)
451 lowcomms_data_ready(sk);
452 } else if (sk->sk_state == TCP_ESTABLISHED) {
453 lowcomms_write_space(sk);
457 int dlm_lowcomms_connect_node(int nodeid)
459 struct connection *con;
461 if (nodeid == dlm_our_nodeid())
464 con = nodeid2con(nodeid, GFP_NOFS);
467 lowcomms_connect_sock(con);
471 static void lowcomms_error_report(struct sock *sk)
473 struct connection *con;
474 struct sockaddr_storage saddr;
476 void (*orig_report)(struct sock *) = NULL;
478 read_lock_bh(&sk->sk_callback_lock);
483 orig_report = listen_sock.sk_error_report;
484 if (con->sock == NULL ||
485 kernel_getpeername(con->sock, (struct sockaddr *)&saddr, &buflen)) {
486 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
487 "sending to node %d, port %d, "
488 "sk_err=%d/%d\n", dlm_our_nodeid(),
489 con->nodeid, dlm_config.ci_tcp_port,
490 sk->sk_err, sk->sk_err_soft);
491 } else if (saddr.ss_family == AF_INET) {
492 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
494 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
495 "sending to node %d at %pI4, port %d, "
496 "sk_err=%d/%d\n", dlm_our_nodeid(),
497 con->nodeid, &sin4->sin_addr.s_addr,
498 dlm_config.ci_tcp_port, sk->sk_err,
501 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
503 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
504 "sending to node %d at %u.%u.%u.%u, "
505 "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
506 con->nodeid, sin6->sin6_addr.s6_addr32[0],
507 sin6->sin6_addr.s6_addr32[1],
508 sin6->sin6_addr.s6_addr32[2],
509 sin6->sin6_addr.s6_addr32[3],
510 dlm_config.ci_tcp_port, sk->sk_err,
514 read_unlock_bh(&sk->sk_callback_lock);
519 /* Note: sk_callback_lock must be locked before calling this function. */
520 static void save_listen_callbacks(struct socket *sock)
522 struct sock *sk = sock->sk;
524 listen_sock.sk_data_ready = sk->sk_data_ready;
525 listen_sock.sk_state_change = sk->sk_state_change;
526 listen_sock.sk_write_space = sk->sk_write_space;
527 listen_sock.sk_error_report = sk->sk_error_report;
530 static void restore_callbacks(struct socket *sock)
532 struct sock *sk = sock->sk;
534 write_lock_bh(&sk->sk_callback_lock);
535 sk->sk_user_data = NULL;
536 sk->sk_data_ready = listen_sock.sk_data_ready;
537 sk->sk_state_change = listen_sock.sk_state_change;
538 sk->sk_write_space = listen_sock.sk_write_space;
539 sk->sk_error_report = listen_sock.sk_error_report;
540 write_unlock_bh(&sk->sk_callback_lock);
543 /* Make a socket active */
544 static void add_sock(struct socket *sock, struct connection *con)
546 struct sock *sk = sock->sk;
548 write_lock_bh(&sk->sk_callback_lock);
551 sk->sk_user_data = con;
552 /* Install a data_ready callback */
553 sk->sk_data_ready = lowcomms_data_ready;
554 sk->sk_write_space = lowcomms_write_space;
555 sk->sk_state_change = lowcomms_state_change;
556 sk->sk_allocation = GFP_NOFS;
557 sk->sk_error_report = lowcomms_error_report;
558 write_unlock_bh(&sk->sk_callback_lock);
561 /* Add the port number to an IPv6 or 4 sockaddr and return the address
563 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
566 saddr->ss_family = dlm_local_addr[0]->ss_family;
567 if (saddr->ss_family == AF_INET) {
568 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
569 in4_addr->sin_port = cpu_to_be16(port);
570 *addr_len = sizeof(struct sockaddr_in);
571 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
573 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
574 in6_addr->sin6_port = cpu_to_be16(port);
575 *addr_len = sizeof(struct sockaddr_in6);
577 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
580 /* Close a remote connection and tidy up */
581 static void close_connection(struct connection *con, bool and_other,
584 if (tx && cancel_work_sync(&con->swork))
585 log_print("canceled swork for node %d", con->nodeid);
586 if (rx && cancel_work_sync(&con->rwork))
587 log_print("canceled rwork for node %d", con->nodeid);
589 mutex_lock(&con->sock_mutex);
591 restore_callbacks(con->sock);
592 sock_release(con->sock);
595 if (con->othercon && and_other) {
596 /* Will only re-enter once. */
597 close_connection(con->othercon, false, true, true);
600 __free_page(con->rx_page);
605 mutex_unlock(&con->sock_mutex);
608 /* Data received from remote end */
609 static int receive_from_sock(struct connection *con)
612 struct msghdr msg = {};
616 int call_again_soon = 0;
619 mutex_lock(&con->sock_mutex);
621 if (con->sock == NULL) {
625 if (con->nodeid == 0) {
630 if (con->rx_page == NULL) {
632 * This doesn't need to be atomic, but I think it should
633 * improve performance if it is.
635 con->rx_page = alloc_page(GFP_ATOMIC);
636 if (con->rx_page == NULL)
638 cbuf_init(&con->cb, PAGE_SIZE);
642 * iov[0] is the bit of the circular buffer between the current end
643 * point (cb.base + cb.len) and the end of the buffer.
645 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
646 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
651 * iov[1] is the bit of the circular buffer between the start of the
652 * buffer and the start of the currently used section (cb.base)
654 if (cbuf_data(&con->cb) >= con->cb.base) {
655 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb);
656 iov[1].iov_len = con->cb.base;
657 iov[1].iov_base = page_address(con->rx_page);
660 len = iov[0].iov_len + iov[1].iov_len;
662 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
663 MSG_DONTWAIT | MSG_NOSIGNAL);
669 cbuf_add(&con->cb, ret);
670 ret = dlm_process_incoming_buffer(con->nodeid,
671 page_address(con->rx_page),
672 con->cb.base, con->cb.len,
674 if (ret == -EBADMSG) {
675 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d",
676 page_address(con->rx_page), con->cb.base,
681 cbuf_eat(&con->cb, ret);
683 if (cbuf_empty(&con->cb) && !call_again_soon) {
684 __free_page(con->rx_page);
690 mutex_unlock(&con->sock_mutex);
694 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
695 queue_work(recv_workqueue, &con->rwork);
696 mutex_unlock(&con->sock_mutex);
700 mutex_unlock(&con->sock_mutex);
701 if (ret != -EAGAIN) {
702 close_connection(con, false, true, false);
703 /* Reconnect when there is something to send */
705 /* Don't return success if we really got EOF */
712 /* Listening socket is busy, accept a connection */
713 static int tcp_accept_from_sock(struct connection *con)
716 struct sockaddr_storage peeraddr;
717 struct socket *newsock;
720 struct connection *newcon;
721 struct connection *addcon;
723 mutex_lock(&connections_lock);
724 if (!dlm_allow_conn) {
725 mutex_unlock(&connections_lock);
728 mutex_unlock(&connections_lock);
730 memset(&peeraddr, 0, sizeof(peeraddr));
731 result = sock_create_lite(dlm_local_addr[0]->ss_family,
732 SOCK_STREAM, IPPROTO_TCP, &newsock);
736 mutex_lock_nested(&con->sock_mutex, 0);
739 if (con->sock == NULL)
742 newsock->type = con->sock->type;
743 newsock->ops = con->sock->ops;
745 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK, true);
749 /* Get the connected socket's peer */
750 memset(&peeraddr, 0, sizeof(peeraddr));
751 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
753 result = -ECONNABORTED;
757 /* Get the new node's NODEID */
758 make_sockaddr(&peeraddr, 0, &len);
759 if (addr_to_nodeid(&peeraddr, &nodeid)) {
760 unsigned char *b=(unsigned char *)&peeraddr;
761 log_print("connect from non cluster node");
762 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
763 b, sizeof(struct sockaddr_storage));
764 sock_release(newsock);
765 mutex_unlock(&con->sock_mutex);
769 log_print("got connection from %d", nodeid);
771 /* Check to see if we already have a connection to this node. This
772 * could happen if the two nodes initiate a connection at roughly
773 * the same time and the connections cross on the wire.
774 * In this case we store the incoming one in "othercon"
776 newcon = nodeid2con(nodeid, GFP_NOFS);
781 mutex_lock_nested(&newcon->sock_mutex, 1);
783 struct connection *othercon = newcon->othercon;
786 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
788 log_print("failed to allocate incoming socket");
789 mutex_unlock(&newcon->sock_mutex);
793 othercon->nodeid = nodeid;
794 othercon->rx_action = receive_from_sock;
795 mutex_init(&othercon->sock_mutex);
796 INIT_WORK(&othercon->swork, process_send_sockets);
797 INIT_WORK(&othercon->rwork, process_recv_sockets);
798 set_bit(CF_IS_OTHERCON, &othercon->flags);
800 if (!othercon->sock) {
801 newcon->othercon = othercon;
802 othercon->sock = newsock;
803 newsock->sk->sk_user_data = othercon;
804 add_sock(newsock, othercon);
808 printk("Extra connection from node %d attempted\n", nodeid);
810 mutex_unlock(&newcon->sock_mutex);
815 newsock->sk->sk_user_data = newcon;
816 newcon->rx_action = receive_from_sock;
817 /* accept copies the sk after we've saved the callbacks, so we
818 don't want to save them a second time or comm errors will
819 result in calling sk_error_report recursively. */
820 add_sock(newsock, newcon);
824 mutex_unlock(&newcon->sock_mutex);
827 * Add it to the active queue in case we got data
828 * between processing the accept adding the socket
829 * to the read_sockets list
831 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
832 queue_work(recv_workqueue, &addcon->rwork);
833 mutex_unlock(&con->sock_mutex);
838 mutex_unlock(&con->sock_mutex);
839 sock_release(newsock);
841 if (result != -EAGAIN)
842 log_print("error accepting connection from node: %d", result);
846 static int sctp_accept_from_sock(struct connection *con)
848 /* Check that the new node is in the lockspace */
849 struct sctp_prim prim;
853 struct connection *newcon;
854 struct connection *addcon;
855 struct socket *newsock;
857 mutex_lock(&connections_lock);
858 if (!dlm_allow_conn) {
859 mutex_unlock(&connections_lock);
862 mutex_unlock(&connections_lock);
864 mutex_lock_nested(&con->sock_mutex, 0);
866 ret = kernel_accept(con->sock, &newsock, O_NONBLOCK);
870 memset(&prim, 0, sizeof(struct sctp_prim));
871 prim_len = sizeof(struct sctp_prim);
873 ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR,
874 (char *)&prim, &prim_len);
876 log_print("getsockopt/sctp_primary_addr failed: %d", ret);
880 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
881 ret = addr_to_nodeid(&prim.ssp_addr, &nodeid);
883 unsigned char *b = (unsigned char *)&prim.ssp_addr;
885 log_print("reject connect from unknown addr");
886 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
887 b, sizeof(struct sockaddr_storage));
891 newcon = nodeid2con(nodeid, GFP_NOFS);
897 mutex_lock_nested(&newcon->sock_mutex, 1);
900 struct connection *othercon = newcon->othercon;
903 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
905 log_print("failed to allocate incoming socket");
906 mutex_unlock(&newcon->sock_mutex);
910 othercon->nodeid = nodeid;
911 othercon->rx_action = receive_from_sock;
912 mutex_init(&othercon->sock_mutex);
913 INIT_WORK(&othercon->swork, process_send_sockets);
914 INIT_WORK(&othercon->rwork, process_recv_sockets);
915 set_bit(CF_IS_OTHERCON, &othercon->flags);
917 if (!othercon->sock) {
918 newcon->othercon = othercon;
919 othercon->sock = newsock;
920 newsock->sk->sk_user_data = othercon;
921 add_sock(newsock, othercon);
924 printk("Extra connection from node %d attempted\n", nodeid);
926 mutex_unlock(&newcon->sock_mutex);
930 newsock->sk->sk_user_data = newcon;
931 newcon->rx_action = receive_from_sock;
932 add_sock(newsock, newcon);
936 log_print("connected to %d", nodeid);
938 mutex_unlock(&newcon->sock_mutex);
941 * Add it to the active queue in case we got data
942 * between processing the accept adding the socket
943 * to the read_sockets list
945 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
946 queue_work(recv_workqueue, &addcon->rwork);
947 mutex_unlock(&con->sock_mutex);
952 mutex_unlock(&con->sock_mutex);
954 sock_release(newsock);
956 log_print("error accepting connection from node: %d", ret);
961 static void free_entry(struct writequeue_entry *e)
963 __free_page(e->page);
968 * writequeue_entry_complete - try to delete and free write queue entry
969 * @e: write queue entry to try to delete
970 * @completed: bytes completed
972 * writequeue_lock must be held.
974 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
976 e->offset += completed;
979 if (e->len == 0 && e->users == 0) {
986 * sctp_bind_addrs - bind a SCTP socket to all our addresses
988 static int sctp_bind_addrs(struct connection *con, uint16_t port)
990 struct sockaddr_storage localaddr;
991 int i, addr_len, result = 0;
993 for (i = 0; i < dlm_local_count; i++) {
994 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
995 make_sockaddr(&localaddr, port, &addr_len);
998 result = kernel_bind(con->sock,
999 (struct sockaddr *)&localaddr,
1002 result = kernel_setsockopt(con->sock, SOL_SCTP,
1003 SCTP_SOCKOPT_BINDX_ADD,
1004 (char *)&localaddr, addr_len);
1007 log_print("Can't bind to %d addr number %d, %d.\n",
1008 port, i + 1, result);
1015 /* Initiate an SCTP association.
1016 This is a special case of send_to_sock() in that we don't yet have a
1017 peeled-off socket for this association, so we use the listening socket
1018 and add the primary IP address of the remote node.
1020 static void sctp_connect_to_sock(struct connection *con)
1022 struct sockaddr_storage daddr;
1026 struct socket *sock;
1028 if (con->nodeid == 0) {
1029 log_print("attempt to connect sock 0 foiled");
1033 mutex_lock(&con->sock_mutex);
1035 /* Some odd races can cause double-connects, ignore them */
1036 if (con->retries++ > MAX_CONNECT_RETRIES)
1040 log_print("node %d already connected.", con->nodeid);
1044 memset(&daddr, 0, sizeof(daddr));
1045 result = nodeid_to_addr(con->nodeid, &daddr, NULL, true);
1047 log_print("no address for nodeid %d", con->nodeid);
1051 /* Create a socket to communicate with */
1052 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1053 SOCK_STREAM, IPPROTO_SCTP, &sock);
1057 sock->sk->sk_user_data = con;
1058 con->rx_action = receive_from_sock;
1059 con->connect_action = sctp_connect_to_sock;
1060 add_sock(sock, con);
1062 /* Bind to all addresses. */
1063 if (sctp_bind_addrs(con, 0))
1066 make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1068 log_print("connecting to %d", con->nodeid);
1070 /* Turn off Nagle's algorithm */
1071 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1074 result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1076 if (result == -EINPROGRESS)
1087 * Some errors are fatal and this list might need adjusting. For other
1088 * errors we try again until the max number of retries is reached.
1090 if (result != -EHOSTUNREACH &&
1091 result != -ENETUNREACH &&
1092 result != -ENETDOWN &&
1093 result != -EINVAL &&
1094 result != -EPROTONOSUPPORT) {
1095 log_print("connect %d try %d error %d", con->nodeid,
1096 con->retries, result);
1097 mutex_unlock(&con->sock_mutex);
1099 lowcomms_connect_sock(con);
1104 mutex_unlock(&con->sock_mutex);
1107 /* Connect a new socket to its peer */
1108 static void tcp_connect_to_sock(struct connection *con)
1110 struct sockaddr_storage saddr, src_addr;
1112 struct socket *sock = NULL;
1116 if (con->nodeid == 0) {
1117 log_print("attempt to connect sock 0 foiled");
1121 mutex_lock(&con->sock_mutex);
1122 if (con->retries++ > MAX_CONNECT_RETRIES)
1125 /* Some odd races can cause double-connects, ignore them */
1129 /* Create a socket to communicate with */
1130 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1131 SOCK_STREAM, IPPROTO_TCP, &sock);
1135 memset(&saddr, 0, sizeof(saddr));
1136 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1138 log_print("no address for nodeid %d", con->nodeid);
1142 sock->sk->sk_user_data = con;
1143 con->rx_action = receive_from_sock;
1144 con->connect_action = tcp_connect_to_sock;
1145 add_sock(sock, con);
1147 /* Bind to our cluster-known address connecting to avoid
1149 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1150 make_sockaddr(&src_addr, 0, &addr_len);
1151 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1154 log_print("could not bind for connect: %d", result);
1155 /* This *may* not indicate a critical error */
1158 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1160 log_print("connecting to %d", con->nodeid);
1162 /* Turn off Nagle's algorithm */
1163 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1166 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1168 if (result == -EINPROGRESS)
1175 sock_release(con->sock);
1181 * Some errors are fatal and this list might need adjusting. For other
1182 * errors we try again until the max number of retries is reached.
1184 if (result != -EHOSTUNREACH &&
1185 result != -ENETUNREACH &&
1186 result != -ENETDOWN &&
1187 result != -EINVAL &&
1188 result != -EPROTONOSUPPORT) {
1189 log_print("connect %d try %d error %d", con->nodeid,
1190 con->retries, result);
1191 mutex_unlock(&con->sock_mutex);
1193 lowcomms_connect_sock(con);
1197 mutex_unlock(&con->sock_mutex);
1201 static struct socket *tcp_create_listen_sock(struct connection *con,
1202 struct sockaddr_storage *saddr)
1204 struct socket *sock = NULL;
1209 if (dlm_local_addr[0]->ss_family == AF_INET)
1210 addr_len = sizeof(struct sockaddr_in);
1212 addr_len = sizeof(struct sockaddr_in6);
1214 /* Create a socket to communicate with */
1215 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1216 SOCK_STREAM, IPPROTO_TCP, &sock);
1218 log_print("Can't create listening comms socket");
1222 /* Turn off Nagle's algorithm */
1223 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1226 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1227 (char *)&one, sizeof(one));
1230 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1232 sock->sk->sk_user_data = con;
1233 save_listen_callbacks(sock);
1234 con->rx_action = tcp_accept_from_sock;
1235 con->connect_action = tcp_connect_to_sock;
1237 /* Bind to our port */
1238 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1239 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1241 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1247 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1248 (char *)&one, sizeof(one));
1250 log_print("Set keepalive failed: %d", result);
1253 result = sock->ops->listen(sock, 5);
1255 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1265 /* Get local addresses */
1266 static void init_local(void)
1268 struct sockaddr_storage sas, *addr;
1271 dlm_local_count = 0;
1272 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1273 if (dlm_our_addr(&sas, i))
1276 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1279 dlm_local_addr[dlm_local_count++] = addr;
1283 /* Initialise SCTP socket and bind to all interfaces */
1284 static int sctp_listen_for_all(void)
1286 struct socket *sock = NULL;
1287 int result = -EINVAL;
1288 struct connection *con = nodeid2con(0, GFP_NOFS);
1289 int bufsize = NEEDED_RMEM;
1295 log_print("Using SCTP for communications");
1297 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1298 SOCK_STREAM, IPPROTO_SCTP, &sock);
1300 log_print("Can't create comms socket, check SCTP is loaded");
1304 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1305 (char *)&bufsize, sizeof(bufsize));
1307 log_print("Error increasing buffer space on socket %d", result);
1309 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1312 log_print("Could not set SCTP NODELAY error %d\n", result);
1314 write_lock_bh(&sock->sk->sk_callback_lock);
1315 /* Init con struct */
1316 sock->sk->sk_user_data = con;
1317 save_listen_callbacks(sock);
1319 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1320 con->rx_action = sctp_accept_from_sock;
1321 con->connect_action = sctp_connect_to_sock;
1323 write_unlock_bh(&sock->sk->sk_callback_lock);
1325 /* Bind to all addresses. */
1326 if (sctp_bind_addrs(con, dlm_config.ci_tcp_port))
1327 goto create_delsock;
1329 result = sock->ops->listen(sock, 5);
1331 log_print("Can't set socket listening");
1332 goto create_delsock;
1344 static int tcp_listen_for_all(void)
1346 struct socket *sock = NULL;
1347 struct connection *con = nodeid2con(0, GFP_NOFS);
1348 int result = -EINVAL;
1353 /* We don't support multi-homed hosts */
1354 if (dlm_local_addr[1] != NULL) {
1355 log_print("TCP protocol can't handle multi-homed hosts, "
1360 log_print("Using TCP for communications");
1362 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1364 add_sock(sock, con);
1368 result = -EADDRINUSE;
1376 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1379 struct writequeue_entry *entry;
1381 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1385 entry->page = alloc_page(allocation);
1400 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1402 struct connection *con;
1403 struct writequeue_entry *e;
1406 con = nodeid2con(nodeid, allocation);
1410 spin_lock(&con->writequeue_lock);
1411 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1412 if ((&e->list == &con->writequeue) ||
1413 (PAGE_SIZE - e->end < len)) {
1420 spin_unlock(&con->writequeue_lock);
1424 *ppc = page_address(e->page) + offset;
1428 e = new_writequeue_entry(con, allocation);
1430 spin_lock(&con->writequeue_lock);
1434 list_add_tail(&e->list, &con->writequeue);
1435 spin_unlock(&con->writequeue_lock);
1441 void dlm_lowcomms_commit_buffer(void *mh)
1443 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1444 struct connection *con = e->con;
1447 spin_lock(&con->writequeue_lock);
1451 e->len = e->end - e->offset;
1452 spin_unlock(&con->writequeue_lock);
1454 queue_work(send_workqueue, &con->swork);
1458 spin_unlock(&con->writequeue_lock);
1462 /* Send a message */
1463 static void send_to_sock(struct connection *con)
1466 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1467 struct writequeue_entry *e;
1471 mutex_lock(&con->sock_mutex);
1472 if (con->sock == NULL)
1475 spin_lock(&con->writequeue_lock);
1477 e = list_entry(con->writequeue.next, struct writequeue_entry,
1479 if ((struct list_head *) e == &con->writequeue)
1484 BUG_ON(len == 0 && e->users == 0);
1485 spin_unlock(&con->writequeue_lock);
1489 ret = kernel_sendpage(con->sock, e->page, offset, len,
1491 if (ret == -EAGAIN || ret == 0) {
1492 if (ret == -EAGAIN &&
1493 test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1494 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1495 /* Notify TCP that we're limited by the
1496 * application window size.
1498 set_bit(SOCK_NOSPACE, &con->sock->flags);
1499 con->sock->sk->sk_write_pending++;
1507 /* Don't starve people filling buffers */
1508 if (++count >= MAX_SEND_MSG_COUNT) {
1513 spin_lock(&con->writequeue_lock);
1514 writequeue_entry_complete(e, ret);
1516 spin_unlock(&con->writequeue_lock);
1518 mutex_unlock(&con->sock_mutex);
1522 mutex_unlock(&con->sock_mutex);
1523 close_connection(con, false, false, true);
1524 /* Requeue the send work. When the work daemon runs again, it will try
1525 a new connection, then call this function again. */
1526 queue_work(send_workqueue, &con->swork);
1530 mutex_unlock(&con->sock_mutex);
1531 queue_work(send_workqueue, &con->swork);
1535 static void clean_one_writequeue(struct connection *con)
1537 struct writequeue_entry *e, *safe;
1539 spin_lock(&con->writequeue_lock);
1540 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1544 spin_unlock(&con->writequeue_lock);
1547 /* Called from recovery when it knows that a node has
1549 int dlm_lowcomms_close(int nodeid)
1551 struct connection *con;
1552 struct dlm_node_addr *na;
1554 log_print("closing connection to node %d", nodeid);
1555 con = nodeid2con(nodeid, 0);
1557 set_bit(CF_CLOSE, &con->flags);
1558 close_connection(con, true, true, true);
1559 clean_one_writequeue(con);
1562 spin_lock(&dlm_node_addrs_spin);
1563 na = find_node_addr(nodeid);
1565 list_del(&na->list);
1566 while (na->addr_count--)
1567 kfree(na->addr[na->addr_count]);
1570 spin_unlock(&dlm_node_addrs_spin);
1575 /* Receive workqueue function */
1576 static void process_recv_sockets(struct work_struct *work)
1578 struct connection *con = container_of(work, struct connection, rwork);
1581 clear_bit(CF_READ_PENDING, &con->flags);
1583 err = con->rx_action(con);
1587 /* Send workqueue function */
1588 static void process_send_sockets(struct work_struct *work)
1590 struct connection *con = container_of(work, struct connection, swork);
1592 if (con->sock == NULL) /* not mutex protected so check it inside too */
1593 con->connect_action(con);
1594 if (!list_empty(&con->writequeue))
1599 /* Discard all entries on the write queues */
1600 static void clean_writequeues(void)
1602 foreach_conn(clean_one_writequeue);
1605 static void work_stop(void)
1607 destroy_workqueue(recv_workqueue);
1608 destroy_workqueue(send_workqueue);
1611 static int work_start(void)
1613 recv_workqueue = alloc_workqueue("dlm_recv",
1614 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1615 if (!recv_workqueue) {
1616 log_print("can't start dlm_recv");
1620 send_workqueue = alloc_workqueue("dlm_send",
1621 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1622 if (!send_workqueue) {
1623 log_print("can't start dlm_send");
1624 destroy_workqueue(recv_workqueue);
1631 static void _stop_conn(struct connection *con, bool and_other)
1633 mutex_lock(&con->sock_mutex);
1634 set_bit(CF_READ_PENDING, &con->flags);
1635 if (con->sock && con->sock->sk)
1636 con->sock->sk->sk_user_data = NULL;
1637 if (con->othercon && and_other)
1638 _stop_conn(con->othercon, false);
1639 mutex_unlock(&con->sock_mutex);
1642 static void stop_conn(struct connection *con)
1644 _stop_conn(con, true);
1647 static void free_conn(struct connection *con)
1649 close_connection(con, true, true, true);
1651 kmem_cache_free(con_cache, con->othercon);
1652 hlist_del(&con->list);
1653 kmem_cache_free(con_cache, con);
1656 static void work_flush(void)
1660 struct hlist_node *n;
1661 struct connection *con;
1663 flush_workqueue(recv_workqueue);
1664 flush_workqueue(send_workqueue);
1667 foreach_conn(stop_conn);
1668 flush_workqueue(recv_workqueue);
1669 flush_workqueue(send_workqueue);
1670 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1671 hlist_for_each_entry_safe(con, n,
1672 &connection_hash[i], list) {
1673 ok &= test_bit(CF_READ_PENDING, &con->flags);
1675 ok &= test_bit(CF_READ_PENDING,
1676 &con->othercon->flags);
1682 void dlm_lowcomms_stop(void)
1684 /* Set all the flags to prevent any
1687 mutex_lock(&connections_lock);
1689 mutex_unlock(&connections_lock);
1691 clean_writequeues();
1692 foreach_conn(free_conn);
1695 kmem_cache_destroy(con_cache);
1698 int dlm_lowcomms_start(void)
1700 int error = -EINVAL;
1701 struct connection *con;
1704 for (i = 0; i < CONN_HASH_SIZE; i++)
1705 INIT_HLIST_HEAD(&connection_hash[i]);
1708 if (!dlm_local_count) {
1710 log_print("no local IP address has been set");
1715 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1716 __alignof__(struct connection), 0,
1721 error = work_start();
1727 /* Start listening */
1728 if (dlm_config.ci_protocol == 0)
1729 error = tcp_listen_for_all();
1731 error = sctp_listen_for_all();
1739 con = nodeid2con(0,0);
1741 close_connection(con, false, true, true);
1742 kmem_cache_free(con_cache, con);
1745 kmem_cache_destroy(con_cache);
1750 void dlm_lowcomms_exit(void)
1752 struct dlm_node_addr *na, *safe;
1754 spin_lock(&dlm_node_addrs_spin);
1755 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1756 list_del(&na->list);
1757 while (na->addr_count--)
1758 kfree(na->addr[na->addr_count]);
1761 spin_unlock(&dlm_node_addrs_spin);