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
124 struct hlist_node list;
125 struct connection *othercon;
126 struct work_struct rwork; /* Receive workqueue */
127 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;
148 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
151 static LIST_HEAD(dlm_node_addrs);
152 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
154 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
155 static int dlm_local_count;
156 static int dlm_allow_conn;
159 static struct workqueue_struct *recv_workqueue;
160 static struct workqueue_struct *send_workqueue;
162 static struct hlist_head connection_hash[CONN_HASH_SIZE];
163 static DEFINE_MUTEX(connections_lock);
164 static struct kmem_cache *con_cache;
166 static void process_recv_sockets(struct work_struct *work);
167 static void process_send_sockets(struct work_struct *work);
170 /* This is deliberately very simple because most clusters have simple
171 sequential nodeids, so we should be able to go straight to a connection
172 struct in the array */
173 static inline int nodeid_hash(int nodeid)
175 return nodeid & (CONN_HASH_SIZE-1);
178 static struct connection *__find_con(int nodeid)
181 struct connection *con;
183 r = nodeid_hash(nodeid);
185 hlist_for_each_entry(con, &connection_hash[r], list) {
186 if (con->nodeid == nodeid)
193 * If 'allocation' is zero then we don't attempt to create a new
194 * connection structure for this node.
196 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
198 struct connection *con = NULL;
201 con = __find_con(nodeid);
205 con = kmem_cache_zalloc(con_cache, alloc);
209 r = nodeid_hash(nodeid);
210 hlist_add_head(&con->list, &connection_hash[r]);
212 con->nodeid = nodeid;
213 mutex_init(&con->sock_mutex);
214 INIT_LIST_HEAD(&con->writequeue);
215 spin_lock_init(&con->writequeue_lock);
216 INIT_WORK(&con->swork, process_send_sockets);
217 INIT_WORK(&con->rwork, process_recv_sockets);
219 /* Setup action pointers for child sockets */
221 struct connection *zerocon = __find_con(0);
223 con->connect_action = zerocon->connect_action;
225 con->rx_action = zerocon->rx_action;
231 /* Loop round all connections */
232 static void foreach_conn(void (*conn_func)(struct connection *c))
235 struct hlist_node *n;
236 struct connection *con;
238 for (i = 0; i < CONN_HASH_SIZE; i++) {
239 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
244 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
246 struct connection *con;
248 mutex_lock(&connections_lock);
249 con = __nodeid2con(nodeid, allocation);
250 mutex_unlock(&connections_lock);
255 /* This is a bit drastic, but only called when things go wrong */
256 static struct connection *assoc2con(int assoc_id)
259 struct connection *con;
261 mutex_lock(&connections_lock);
263 for (i = 0 ; i < CONN_HASH_SIZE; i++) {
264 hlist_for_each_entry(con, &connection_hash[i], list) {
265 if (con->sctp_assoc == assoc_id) {
266 mutex_unlock(&connections_lock);
271 mutex_unlock(&connections_lock);
275 static struct dlm_node_addr *find_node_addr(int nodeid)
277 struct dlm_node_addr *na;
279 list_for_each_entry(na, &dlm_node_addrs, list) {
280 if (na->nodeid == nodeid)
286 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
288 switch (x->ss_family) {
290 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
291 struct sockaddr_in *siny = (struct sockaddr_in *)y;
292 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
294 if (sinx->sin_port != siny->sin_port)
299 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
300 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
301 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
303 if (sinx->sin6_port != siny->sin6_port)
313 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
314 struct sockaddr *sa_out, bool try_new_addr)
316 struct sockaddr_storage sas;
317 struct dlm_node_addr *na;
319 if (!dlm_local_count)
322 spin_lock(&dlm_node_addrs_spin);
323 na = find_node_addr(nodeid);
324 if (na && na->addr_count) {
326 na->curr_addr_index++;
327 if (na->curr_addr_index == na->addr_count)
328 na->curr_addr_index = 0;
331 memcpy(&sas, na->addr[na->curr_addr_index ],
332 sizeof(struct sockaddr_storage));
334 spin_unlock(&dlm_node_addrs_spin);
343 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
348 if (dlm_local_addr[0]->ss_family == AF_INET) {
349 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas;
350 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
351 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
353 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas;
354 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
355 ret6->sin6_addr = in6->sin6_addr;
361 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
363 struct dlm_node_addr *na;
367 spin_lock(&dlm_node_addrs_spin);
368 list_for_each_entry(na, &dlm_node_addrs, list) {
372 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
373 if (addr_compare(na->addr[addr_i], addr)) {
374 *nodeid = na->nodeid;
381 spin_unlock(&dlm_node_addrs_spin);
385 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
387 struct sockaddr_storage *new_addr;
388 struct dlm_node_addr *new_node, *na;
390 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
394 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
400 memcpy(new_addr, addr, len);
402 spin_lock(&dlm_node_addrs_spin);
403 na = find_node_addr(nodeid);
405 new_node->nodeid = nodeid;
406 new_node->addr[0] = new_addr;
407 new_node->addr_count = 1;
408 list_add(&new_node->list, &dlm_node_addrs);
409 spin_unlock(&dlm_node_addrs_spin);
413 if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
414 spin_unlock(&dlm_node_addrs_spin);
420 na->addr[na->addr_count++] = new_addr;
421 spin_unlock(&dlm_node_addrs_spin);
426 /* Data available on socket or listen socket received a connect */
427 static void lowcomms_data_ready(struct sock *sk)
429 struct connection *con = sock2con(sk);
430 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
431 queue_work(recv_workqueue, &con->rwork);
434 static void lowcomms_write_space(struct sock *sk)
436 struct connection *con = sock2con(sk);
441 clear_bit(SOCK_NOSPACE, &con->sock->flags);
443 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
444 con->sock->sk->sk_write_pending--;
445 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
448 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
449 queue_work(send_workqueue, &con->swork);
452 static inline void lowcomms_connect_sock(struct connection *con)
454 if (test_bit(CF_CLOSE, &con->flags))
456 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
457 queue_work(send_workqueue, &con->swork);
460 static void lowcomms_state_change(struct sock *sk)
462 if (sk->sk_state == TCP_ESTABLISHED)
463 lowcomms_write_space(sk);
466 int dlm_lowcomms_connect_node(int nodeid)
468 struct connection *con;
470 /* with sctp there's no connecting without sending */
471 if (dlm_config.ci_protocol != 0)
474 if (nodeid == dlm_our_nodeid())
477 con = nodeid2con(nodeid, GFP_NOFS);
480 lowcomms_connect_sock(con);
484 /* Make a socket active */
485 static void add_sock(struct socket *sock, struct connection *con)
489 /* Install a data_ready callback */
490 con->sock->sk->sk_data_ready = lowcomms_data_ready;
491 con->sock->sk->sk_write_space = lowcomms_write_space;
492 con->sock->sk->sk_state_change = lowcomms_state_change;
493 con->sock->sk->sk_user_data = con;
494 con->sock->sk->sk_allocation = GFP_NOFS;
497 /* Add the port number to an IPv6 or 4 sockaddr and return the address
499 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
502 saddr->ss_family = dlm_local_addr[0]->ss_family;
503 if (saddr->ss_family == AF_INET) {
504 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
505 in4_addr->sin_port = cpu_to_be16(port);
506 *addr_len = sizeof(struct sockaddr_in);
507 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
509 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
510 in6_addr->sin6_port = cpu_to_be16(port);
511 *addr_len = sizeof(struct sockaddr_in6);
513 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
516 /* Close a remote connection and tidy up */
517 static void close_connection(struct connection *con, bool and_other,
520 clear_bit(CF_CONNECT_PENDING, &con->flags);
521 clear_bit(CF_WRITE_PENDING, &con->flags);
522 if (tx && cancel_work_sync(&con->swork))
523 log_print("canceled swork for node %d", con->nodeid);
524 if (rx && cancel_work_sync(&con->rwork))
525 log_print("canceled rwork for node %d", con->nodeid);
527 mutex_lock(&con->sock_mutex);
529 sock_release(con->sock);
532 if (con->othercon && and_other) {
533 /* Will only re-enter once. */
534 close_connection(con->othercon, false, true, true);
537 __free_page(con->rx_page);
542 mutex_unlock(&con->sock_mutex);
545 /* We only send shutdown messages to nodes that are not part of the cluster
546 * or if we get multiple connections from a node.
548 static void sctp_send_shutdown(sctp_assoc_t associd)
550 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
551 struct msghdr outmessage;
552 struct cmsghdr *cmsg;
553 struct sctp_sndrcvinfo *sinfo;
555 struct connection *con;
557 con = nodeid2con(0,0);
560 outmessage.msg_name = NULL;
561 outmessage.msg_namelen = 0;
562 outmessage.msg_control = outcmsg;
563 outmessage.msg_controllen = sizeof(outcmsg);
564 outmessage.msg_flags = MSG_EOR;
566 cmsg = CMSG_FIRSTHDR(&outmessage);
567 cmsg->cmsg_level = IPPROTO_SCTP;
568 cmsg->cmsg_type = SCTP_SNDRCV;
569 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
570 outmessage.msg_controllen = cmsg->cmsg_len;
571 sinfo = CMSG_DATA(cmsg);
572 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
574 sinfo->sinfo_flags |= MSG_EOF;
575 sinfo->sinfo_assoc_id = associd;
577 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
580 log_print("send EOF to node failed: %d", ret);
583 static void sctp_init_failed_foreach(struct connection *con)
587 * Don't try to recover base con and handle race where the
588 * other node's assoc init creates a assoc and we get that
589 * notification, then we get a notification that our attempt
590 * failed due. This happens when we are still trying the primary
591 * address, but the other node has already tried secondary addrs
592 * and found one that worked.
594 if (!con->nodeid || con->sctp_assoc)
597 log_print("Retrying SCTP association init for node %d\n", con->nodeid);
599 con->try_new_addr = true;
601 if (test_and_clear_bit(CF_INIT_PENDING, &con->flags)) {
602 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
603 queue_work(send_workqueue, &con->swork);
607 /* INIT failed but we don't know which node...
608 restart INIT on all pending nodes */
609 static void sctp_init_failed(void)
611 mutex_lock(&connections_lock);
613 foreach_conn(sctp_init_failed_foreach);
615 mutex_unlock(&connections_lock);
618 static void retry_failed_sctp_send(struct connection *recv_con,
619 struct sctp_send_failed *sn_send_failed,
622 int len = sn_send_failed->ssf_length - sizeof(struct sctp_send_failed);
623 struct dlm_mhandle *mh;
624 struct connection *con;
626 int nodeid = sn_send_failed->ssf_info.sinfo_ppid;
628 log_print("Retry sending %d bytes to node id %d", len, nodeid);
631 log_print("Shouldn't resend data via listening connection.");
635 con = nodeid2con(nodeid, 0);
637 log_print("Could not look up con for nodeid %d\n",
642 mh = dlm_lowcomms_get_buffer(nodeid, len, GFP_NOFS, &retry_buf);
644 log_print("Could not allocate buf for retry.");
647 memcpy(retry_buf, buf + sizeof(struct sctp_send_failed), len);
648 dlm_lowcomms_commit_buffer(mh);
651 * If we got a assoc changed event before the send failed event then
652 * we only need to retry the send.
654 if (con->sctp_assoc) {
655 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
656 queue_work(send_workqueue, &con->swork);
658 sctp_init_failed_foreach(con);
661 /* Something happened to an association */
662 static void process_sctp_notification(struct connection *con,
663 struct msghdr *msg, char *buf)
665 union sctp_notification *sn = (union sctp_notification *)buf;
666 struct linger linger;
668 switch (sn->sn_header.sn_type) {
669 case SCTP_SEND_FAILED:
670 retry_failed_sctp_send(con, &sn->sn_send_failed, buf);
672 case SCTP_ASSOC_CHANGE:
673 switch (sn->sn_assoc_change.sac_state) {
677 /* Check that the new node is in the lockspace */
678 struct sctp_prim prim;
682 struct connection *new_con;
685 * We get this before any data for an association.
686 * We verify that the node is in the cluster and
687 * then peel off a socket for it.
689 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
690 log_print("COMM_UP for invalid assoc ID %d",
691 (int)sn->sn_assoc_change.sac_assoc_id);
695 memset(&prim, 0, sizeof(struct sctp_prim));
696 prim_len = sizeof(struct sctp_prim);
697 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
699 ret = kernel_getsockopt(con->sock,
705 log_print("getsockopt/sctp_primary_addr on "
706 "new assoc %d failed : %d",
707 (int)sn->sn_assoc_change.sac_assoc_id,
710 /* Retry INIT later */
711 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
713 clear_bit(CF_CONNECT_PENDING, &con->flags);
716 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
717 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
718 unsigned char *b=(unsigned char *)&prim.ssp_addr;
719 log_print("reject connect from unknown addr");
720 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
721 b, sizeof(struct sockaddr_storage));
722 sctp_send_shutdown(prim.ssp_assoc_id);
726 new_con = nodeid2con(nodeid, GFP_NOFS);
731 log_print("reject connect from node %d: "
732 "already has a connection.",
734 sctp_send_shutdown(prim.ssp_assoc_id);
738 /* Peel off a new sock */
739 lock_sock(con->sock->sk);
740 ret = sctp_do_peeloff(con->sock->sk,
741 sn->sn_assoc_change.sac_assoc_id,
743 release_sock(con->sock->sk);
745 log_print("Can't peel off a socket for "
746 "connection %d to node %d: err=%d",
747 (int)sn->sn_assoc_change.sac_assoc_id,
751 add_sock(new_con->sock, new_con);
755 ret = kernel_setsockopt(new_con->sock, SOL_SOCKET, SO_LINGER,
756 (char *)&linger, sizeof(linger));
758 log_print("set socket option SO_LINGER failed");
760 log_print("connecting to %d sctp association %d",
761 nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
763 new_con->sctp_assoc = sn->sn_assoc_change.sac_assoc_id;
764 new_con->try_new_addr = false;
765 /* Send any pending writes */
766 clear_bit(CF_CONNECT_PENDING, &new_con->flags);
767 clear_bit(CF_INIT_PENDING, &new_con->flags);
768 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
769 queue_work(send_workqueue, &new_con->swork);
771 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
772 queue_work(recv_workqueue, &new_con->rwork);
777 case SCTP_SHUTDOWN_COMP:
779 con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
786 case SCTP_CANT_STR_ASSOC:
788 /* Will retry init when we get the send failed notification */
789 log_print("Can't start SCTP association - retrying");
794 log_print("unexpected SCTP assoc change id=%d state=%d",
795 (int)sn->sn_assoc_change.sac_assoc_id,
796 sn->sn_assoc_change.sac_state);
803 /* Data received from remote end */
804 static int receive_from_sock(struct connection *con)
807 struct msghdr msg = {};
811 int call_again_soon = 0;
813 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
815 mutex_lock(&con->sock_mutex);
817 if (con->sock == NULL) {
822 if (con->rx_page == NULL) {
824 * This doesn't need to be atomic, but I think it should
825 * improve performance if it is.
827 con->rx_page = alloc_page(GFP_ATOMIC);
828 if (con->rx_page == NULL)
830 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
833 /* Only SCTP needs these really */
834 memset(&incmsg, 0, sizeof(incmsg));
835 msg.msg_control = incmsg;
836 msg.msg_controllen = sizeof(incmsg);
839 * iov[0] is the bit of the circular buffer between the current end
840 * point (cb.base + cb.len) and the end of the buffer.
842 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
843 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
848 * iov[1] is the bit of the circular buffer between the start of the
849 * buffer and the start of the currently used section (cb.base)
851 if (cbuf_data(&con->cb) >= con->cb.base) {
852 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
853 iov[1].iov_len = con->cb.base;
854 iov[1].iov_base = page_address(con->rx_page);
857 len = iov[0].iov_len + iov[1].iov_len;
859 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
860 MSG_DONTWAIT | MSG_NOSIGNAL);
864 /* Process SCTP notifications */
865 if (msg.msg_flags & MSG_NOTIFICATION) {
866 msg.msg_control = incmsg;
867 msg.msg_controllen = sizeof(incmsg);
869 process_sctp_notification(con, &msg,
870 page_address(con->rx_page) + con->cb.base);
871 mutex_unlock(&con->sock_mutex);
874 BUG_ON(con->nodeid == 0);
878 cbuf_add(&con->cb, ret);
879 ret = dlm_process_incoming_buffer(con->nodeid,
880 page_address(con->rx_page),
881 con->cb.base, con->cb.len,
883 if (ret == -EBADMSG) {
884 log_print("lowcomms: addr=%p, base=%u, len=%u, "
885 "iov_len=%u, iov_base[0]=%p, read=%d",
886 page_address(con->rx_page), con->cb.base, con->cb.len,
887 len, iov[0].iov_base, r);
891 cbuf_eat(&con->cb, ret);
893 if (cbuf_empty(&con->cb) && !call_again_soon) {
894 __free_page(con->rx_page);
900 mutex_unlock(&con->sock_mutex);
904 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
905 queue_work(recv_workqueue, &con->rwork);
906 mutex_unlock(&con->sock_mutex);
910 mutex_unlock(&con->sock_mutex);
911 if (ret != -EAGAIN) {
912 close_connection(con, false, true, false);
913 /* Reconnect when there is something to send */
915 /* Don't return success if we really got EOF */
922 /* Listening socket is busy, accept a connection */
923 static int tcp_accept_from_sock(struct connection *con)
926 struct sockaddr_storage peeraddr;
927 struct socket *newsock;
930 struct connection *newcon;
931 struct connection *addcon;
933 mutex_lock(&connections_lock);
934 if (!dlm_allow_conn) {
935 mutex_unlock(&connections_lock);
938 mutex_unlock(&connections_lock);
940 memset(&peeraddr, 0, sizeof(peeraddr));
941 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
942 SOCK_STREAM, IPPROTO_TCP, &newsock);
946 mutex_lock_nested(&con->sock_mutex, 0);
949 if (con->sock == NULL)
952 newsock->type = con->sock->type;
953 newsock->ops = con->sock->ops;
955 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
959 /* Get the connected socket's peer */
960 memset(&peeraddr, 0, sizeof(peeraddr));
961 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
963 result = -ECONNABORTED;
967 /* Get the new node's NODEID */
968 make_sockaddr(&peeraddr, 0, &len);
969 if (addr_to_nodeid(&peeraddr, &nodeid)) {
970 unsigned char *b=(unsigned char *)&peeraddr;
971 log_print("connect from non cluster node");
972 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
973 b, sizeof(struct sockaddr_storage));
974 sock_release(newsock);
975 mutex_unlock(&con->sock_mutex);
979 log_print("got connection from %d", nodeid);
981 /* Check to see if we already have a connection to this node. This
982 * could happen if the two nodes initiate a connection at roughly
983 * the same time and the connections cross on the wire.
984 * In this case we store the incoming one in "othercon"
986 newcon = nodeid2con(nodeid, GFP_NOFS);
991 mutex_lock_nested(&newcon->sock_mutex, 1);
993 struct connection *othercon = newcon->othercon;
996 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
998 log_print("failed to allocate incoming socket");
999 mutex_unlock(&newcon->sock_mutex);
1003 othercon->nodeid = nodeid;
1004 othercon->rx_action = receive_from_sock;
1005 mutex_init(&othercon->sock_mutex);
1006 INIT_WORK(&othercon->swork, process_send_sockets);
1007 INIT_WORK(&othercon->rwork, process_recv_sockets);
1008 set_bit(CF_IS_OTHERCON, &othercon->flags);
1010 if (!othercon->sock) {
1011 newcon->othercon = othercon;
1012 othercon->sock = newsock;
1013 newsock->sk->sk_user_data = othercon;
1014 add_sock(newsock, othercon);
1018 printk("Extra connection from node %d attempted\n", nodeid);
1020 mutex_unlock(&newcon->sock_mutex);
1025 newsock->sk->sk_user_data = newcon;
1026 newcon->rx_action = receive_from_sock;
1027 add_sock(newsock, newcon);
1031 mutex_unlock(&newcon->sock_mutex);
1034 * Add it to the active queue in case we got data
1035 * between processing the accept adding the socket
1036 * to the read_sockets list
1038 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
1039 queue_work(recv_workqueue, &addcon->rwork);
1040 mutex_unlock(&con->sock_mutex);
1045 mutex_unlock(&con->sock_mutex);
1046 sock_release(newsock);
1048 if (result != -EAGAIN)
1049 log_print("error accepting connection from node: %d", result);
1053 static void free_entry(struct writequeue_entry *e)
1055 __free_page(e->page);
1060 * writequeue_entry_complete - try to delete and free write queue entry
1061 * @e: write queue entry to try to delete
1062 * @completed: bytes completed
1064 * writequeue_lock must be held.
1066 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1068 e->offset += completed;
1069 e->len -= completed;
1071 if (e->len == 0 && e->users == 0) {
1077 /* Initiate an SCTP association.
1078 This is a special case of send_to_sock() in that we don't yet have a
1079 peeled-off socket for this association, so we use the listening socket
1080 and add the primary IP address of the remote node.
1082 static void sctp_init_assoc(struct connection *con)
1084 struct sockaddr_storage rem_addr;
1085 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
1086 struct msghdr outmessage;
1087 struct cmsghdr *cmsg;
1088 struct sctp_sndrcvinfo *sinfo;
1089 struct connection *base_con;
1090 struct writequeue_entry *e;
1096 mutex_lock(&con->sock_mutex);
1097 if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
1100 if (nodeid_to_addr(con->nodeid, NULL, (struct sockaddr *)&rem_addr,
1101 con->try_new_addr)) {
1102 log_print("no address for nodeid %d", con->nodeid);
1105 base_con = nodeid2con(0, 0);
1106 BUG_ON(base_con == NULL);
1108 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
1110 outmessage.msg_name = &rem_addr;
1111 outmessage.msg_namelen = addrlen;
1112 outmessage.msg_control = outcmsg;
1113 outmessage.msg_controllen = sizeof(outcmsg);
1114 outmessage.msg_flags = MSG_EOR;
1116 spin_lock(&con->writequeue_lock);
1118 if (list_empty(&con->writequeue)) {
1119 spin_unlock(&con->writequeue_lock);
1120 log_print("writequeue empty for nodeid %d", con->nodeid);
1124 e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1128 /* Send the first block off the write queue */
1129 iov[0].iov_base = page_address(e->page)+offset;
1130 iov[0].iov_len = len;
1131 spin_unlock(&con->writequeue_lock);
1133 if (rem_addr.ss_family == AF_INET) {
1134 struct sockaddr_in *sin = (struct sockaddr_in *)&rem_addr;
1135 log_print("Trying to connect to %pI4", &sin->sin_addr.s_addr);
1137 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&rem_addr;
1138 log_print("Trying to connect to %pI6", &sin6->sin6_addr);
1141 cmsg = CMSG_FIRSTHDR(&outmessage);
1142 cmsg->cmsg_level = IPPROTO_SCTP;
1143 cmsg->cmsg_type = SCTP_SNDRCV;
1144 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
1145 sinfo = CMSG_DATA(cmsg);
1146 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
1147 sinfo->sinfo_ppid = cpu_to_le32(con->nodeid);
1148 outmessage.msg_controllen = cmsg->cmsg_len;
1149 sinfo->sinfo_flags |= SCTP_ADDR_OVER;
1151 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
1153 log_print("Send first packet to node %d failed: %d",
1156 /* Try again later */
1157 clear_bit(CF_CONNECT_PENDING, &con->flags);
1158 clear_bit(CF_INIT_PENDING, &con->flags);
1161 spin_lock(&con->writequeue_lock);
1162 writequeue_entry_complete(e, ret);
1163 spin_unlock(&con->writequeue_lock);
1167 mutex_unlock(&con->sock_mutex);
1170 /* Connect a new socket to its peer */
1171 static void tcp_connect_to_sock(struct connection *con)
1173 struct sockaddr_storage saddr, src_addr;
1175 struct socket *sock = NULL;
1179 if (con->nodeid == 0) {
1180 log_print("attempt to connect sock 0 foiled");
1184 mutex_lock(&con->sock_mutex);
1185 if (con->retries++ > MAX_CONNECT_RETRIES)
1188 /* Some odd races can cause double-connects, ignore them */
1192 /* Create a socket to communicate with */
1193 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1194 SOCK_STREAM, IPPROTO_TCP, &sock);
1198 memset(&saddr, 0, sizeof(saddr));
1199 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1201 log_print("no address for nodeid %d", con->nodeid);
1205 sock->sk->sk_user_data = con;
1206 con->rx_action = receive_from_sock;
1207 con->connect_action = tcp_connect_to_sock;
1208 add_sock(sock, con);
1210 /* Bind to our cluster-known address connecting to avoid
1212 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1213 make_sockaddr(&src_addr, 0, &addr_len);
1214 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1217 log_print("could not bind for connect: %d", result);
1218 /* This *may* not indicate a critical error */
1221 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1223 log_print("connecting to %d", con->nodeid);
1225 /* Turn off Nagle's algorithm */
1226 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1229 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1231 if (result == -EINPROGRESS)
1238 sock_release(con->sock);
1244 * Some errors are fatal and this list might need adjusting. For other
1245 * errors we try again until the max number of retries is reached.
1247 if (result != -EHOSTUNREACH &&
1248 result != -ENETUNREACH &&
1249 result != -ENETDOWN &&
1250 result != -EINVAL &&
1251 result != -EPROTONOSUPPORT) {
1252 log_print("connect %d try %d error %d", con->nodeid,
1253 con->retries, result);
1254 mutex_unlock(&con->sock_mutex);
1256 clear_bit(CF_CONNECT_PENDING, &con->flags);
1257 lowcomms_connect_sock(con);
1261 mutex_unlock(&con->sock_mutex);
1265 static struct socket *tcp_create_listen_sock(struct connection *con,
1266 struct sockaddr_storage *saddr)
1268 struct socket *sock = NULL;
1273 if (dlm_local_addr[0]->ss_family == AF_INET)
1274 addr_len = sizeof(struct sockaddr_in);
1276 addr_len = sizeof(struct sockaddr_in6);
1278 /* Create a socket to communicate with */
1279 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1280 SOCK_STREAM, IPPROTO_TCP, &sock);
1282 log_print("Can't create listening comms socket");
1286 /* Turn off Nagle's algorithm */
1287 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1290 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1291 (char *)&one, sizeof(one));
1294 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1296 con->rx_action = tcp_accept_from_sock;
1297 con->connect_action = tcp_connect_to_sock;
1299 /* Bind to our port */
1300 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1301 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1303 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1309 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1310 (char *)&one, sizeof(one));
1312 log_print("Set keepalive failed: %d", result);
1315 result = sock->ops->listen(sock, 5);
1317 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1327 /* Get local addresses */
1328 static void init_local(void)
1330 struct sockaddr_storage sas, *addr;
1333 dlm_local_count = 0;
1334 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1335 if (dlm_our_addr(&sas, i))
1338 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1341 memcpy(addr, &sas, sizeof(*addr));
1342 dlm_local_addr[dlm_local_count++] = addr;
1346 /* Bind to an IP address. SCTP allows multiple address so it can do
1348 static int add_sctp_bind_addr(struct connection *sctp_con,
1349 struct sockaddr_storage *addr,
1350 int addr_len, int num)
1355 result = kernel_bind(sctp_con->sock,
1356 (struct sockaddr *) addr,
1359 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1360 SCTP_SOCKOPT_BINDX_ADD,
1361 (char *)addr, addr_len);
1364 log_print("Can't bind to port %d addr number %d",
1365 dlm_config.ci_tcp_port, num);
1370 /* Initialise SCTP socket and bind to all interfaces */
1371 static int sctp_listen_for_all(void)
1373 struct socket *sock = NULL;
1374 struct sockaddr_storage localaddr;
1375 struct sctp_event_subscribe subscribe;
1376 int result = -EINVAL, num = 1, i, addr_len;
1377 struct connection *con = nodeid2con(0, GFP_NOFS);
1378 int bufsize = NEEDED_RMEM;
1384 log_print("Using SCTP for communications");
1386 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1387 SOCK_SEQPACKET, IPPROTO_SCTP, &sock);
1389 log_print("Can't create comms socket, check SCTP is loaded");
1393 /* Listen for events */
1394 memset(&subscribe, 0, sizeof(subscribe));
1395 subscribe.sctp_data_io_event = 1;
1396 subscribe.sctp_association_event = 1;
1397 subscribe.sctp_send_failure_event = 1;
1398 subscribe.sctp_shutdown_event = 1;
1399 subscribe.sctp_partial_delivery_event = 1;
1401 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1402 (char *)&bufsize, sizeof(bufsize));
1404 log_print("Error increasing buffer space on socket %d", result);
1406 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1407 (char *)&subscribe, sizeof(subscribe));
1409 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1411 goto create_delsock;
1414 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1417 log_print("Could not set SCTP NODELAY error %d\n", result);
1419 /* Init con struct */
1420 sock->sk->sk_user_data = con;
1422 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1423 con->rx_action = receive_from_sock;
1424 con->connect_action = sctp_init_assoc;
1426 /* Bind to all interfaces. */
1427 for (i = 0; i < dlm_local_count; i++) {
1428 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1429 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1431 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1433 goto create_delsock;
1437 result = sock->ops->listen(sock, 5);
1439 log_print("Can't set socket listening");
1440 goto create_delsock;
1452 static int tcp_listen_for_all(void)
1454 struct socket *sock = NULL;
1455 struct connection *con = nodeid2con(0, GFP_NOFS);
1456 int result = -EINVAL;
1461 /* We don't support multi-homed hosts */
1462 if (dlm_local_addr[1] != NULL) {
1463 log_print("TCP protocol can't handle multi-homed hosts, "
1468 log_print("Using TCP for communications");
1470 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1472 add_sock(sock, con);
1476 result = -EADDRINUSE;
1484 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1487 struct writequeue_entry *entry;
1489 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1493 entry->page = alloc_page(allocation);
1508 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1510 struct connection *con;
1511 struct writequeue_entry *e;
1514 con = nodeid2con(nodeid, allocation);
1518 spin_lock(&con->writequeue_lock);
1519 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1520 if ((&e->list == &con->writequeue) ||
1521 (PAGE_CACHE_SIZE - e->end < len)) {
1528 spin_unlock(&con->writequeue_lock);
1532 *ppc = page_address(e->page) + offset;
1536 e = new_writequeue_entry(con, allocation);
1538 spin_lock(&con->writequeue_lock);
1542 list_add_tail(&e->list, &con->writequeue);
1543 spin_unlock(&con->writequeue_lock);
1549 void dlm_lowcomms_commit_buffer(void *mh)
1551 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1552 struct connection *con = e->con;
1555 spin_lock(&con->writequeue_lock);
1559 e->len = e->end - e->offset;
1560 spin_unlock(&con->writequeue_lock);
1562 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1563 queue_work(send_workqueue, &con->swork);
1568 spin_unlock(&con->writequeue_lock);
1572 /* Send a message */
1573 static void send_to_sock(struct connection *con)
1576 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1577 struct writequeue_entry *e;
1581 mutex_lock(&con->sock_mutex);
1582 if (con->sock == NULL)
1585 spin_lock(&con->writequeue_lock);
1587 e = list_entry(con->writequeue.next, struct writequeue_entry,
1589 if ((struct list_head *) e == &con->writequeue)
1594 BUG_ON(len == 0 && e->users == 0);
1595 spin_unlock(&con->writequeue_lock);
1599 ret = kernel_sendpage(con->sock, e->page, offset, len,
1601 if (ret == -EAGAIN || ret == 0) {
1602 if (ret == -EAGAIN &&
1603 test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1604 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1605 /* Notify TCP that we're limited by the
1606 * application window size.
1608 set_bit(SOCK_NOSPACE, &con->sock->flags);
1609 con->sock->sk->sk_write_pending++;
1617 /* Don't starve people filling buffers */
1618 if (++count >= MAX_SEND_MSG_COUNT) {
1623 spin_lock(&con->writequeue_lock);
1624 writequeue_entry_complete(e, ret);
1626 spin_unlock(&con->writequeue_lock);
1628 mutex_unlock(&con->sock_mutex);
1632 mutex_unlock(&con->sock_mutex);
1633 close_connection(con, false, false, true);
1634 lowcomms_connect_sock(con);
1638 mutex_unlock(&con->sock_mutex);
1639 if (!test_bit(CF_INIT_PENDING, &con->flags))
1640 lowcomms_connect_sock(con);
1643 static void clean_one_writequeue(struct connection *con)
1645 struct writequeue_entry *e, *safe;
1647 spin_lock(&con->writequeue_lock);
1648 list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1652 spin_unlock(&con->writequeue_lock);
1655 /* Called from recovery when it knows that a node has
1657 int dlm_lowcomms_close(int nodeid)
1659 struct connection *con;
1660 struct dlm_node_addr *na;
1662 log_print("closing connection to node %d", nodeid);
1663 con = nodeid2con(nodeid, 0);
1665 set_bit(CF_CLOSE, &con->flags);
1666 close_connection(con, true, true, true);
1667 clean_one_writequeue(con);
1670 spin_lock(&dlm_node_addrs_spin);
1671 na = find_node_addr(nodeid);
1673 list_del(&na->list);
1674 while (na->addr_count--)
1675 kfree(na->addr[na->addr_count]);
1678 spin_unlock(&dlm_node_addrs_spin);
1683 /* Receive workqueue function */
1684 static void process_recv_sockets(struct work_struct *work)
1686 struct connection *con = container_of(work, struct connection, rwork);
1689 clear_bit(CF_READ_PENDING, &con->flags);
1691 err = con->rx_action(con);
1695 /* Send workqueue function */
1696 static void process_send_sockets(struct work_struct *work)
1698 struct connection *con = container_of(work, struct connection, swork);
1700 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1701 con->connect_action(con);
1702 set_bit(CF_WRITE_PENDING, &con->flags);
1704 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1709 /* Discard all entries on the write queues */
1710 static void clean_writequeues(void)
1712 foreach_conn(clean_one_writequeue);
1715 static void work_stop(void)
1717 destroy_workqueue(recv_workqueue);
1718 destroy_workqueue(send_workqueue);
1721 static int work_start(void)
1723 recv_workqueue = alloc_workqueue("dlm_recv",
1724 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1725 if (!recv_workqueue) {
1726 log_print("can't start dlm_recv");
1730 send_workqueue = alloc_workqueue("dlm_send",
1731 WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1732 if (!send_workqueue) {
1733 log_print("can't start dlm_send");
1734 destroy_workqueue(recv_workqueue);
1741 static void stop_conn(struct connection *con)
1744 if (con->sock && con->sock->sk)
1745 con->sock->sk->sk_user_data = NULL;
1748 static void free_conn(struct connection *con)
1750 close_connection(con, true, true, true);
1752 kmem_cache_free(con_cache, con->othercon);
1753 hlist_del(&con->list);
1754 kmem_cache_free(con_cache, con);
1757 void dlm_lowcomms_stop(void)
1759 /* Set all the flags to prevent any
1762 mutex_lock(&connections_lock);
1764 foreach_conn(stop_conn);
1765 mutex_unlock(&connections_lock);
1769 mutex_lock(&connections_lock);
1770 clean_writequeues();
1772 foreach_conn(free_conn);
1774 mutex_unlock(&connections_lock);
1775 kmem_cache_destroy(con_cache);
1778 int dlm_lowcomms_start(void)
1780 int error = -EINVAL;
1781 struct connection *con;
1784 for (i = 0; i < CONN_HASH_SIZE; i++)
1785 INIT_HLIST_HEAD(&connection_hash[i]);
1788 if (!dlm_local_count) {
1790 log_print("no local IP address has been set");
1795 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1796 __alignof__(struct connection), 0,
1801 error = work_start();
1807 /* Start listening */
1808 if (dlm_config.ci_protocol == 0)
1809 error = tcp_listen_for_all();
1811 error = sctp_listen_for_all();
1819 con = nodeid2con(0,0);
1821 close_connection(con, false, true, true);
1822 kmem_cache_free(con_cache, con);
1825 kmem_cache_destroy(con_cache);
1830 void dlm_lowcomms_exit(void)
1832 struct dlm_node_addr *na, *safe;
1834 spin_lock(&dlm_node_addrs_spin);
1835 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1836 list_del(&na->list);
1837 while (na->addr_count--)
1838 kfree(na->addr[na->addr_count]);
1841 spin_unlock(&dlm_node_addrs_spin);