blob: 7c1e5e5cccd8ae97a87a618ad9818d5b35721553 [file] [log] [blame]
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001/******************************************************************************
2*******************************************************************************
3**
4** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
5** Copyright (C) 2004-2007 Red Hat, Inc. All rights reserved.
6**
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.
10**
11*******************************************************************************
12******************************************************************************/
13
14/*
15 * lowcomms.c
16 *
17 * This is the "low-level" comms layer.
18 *
19 * It is responsible for sending/receiving messages
20 * from other nodes in the cluster.
21 *
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 it's
25 * responsibility. It is this layer's
26 * responsibility to resolve these into IP address or
27 * whatever it needs for inter-node communication.
28 *
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.
38 *
39 * lowcomms will choose to use wither 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 shouldbe 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.
44 *
45 */
46
47#include <asm/ioctls.h>
48#include <net/sock.h>
49#include <net/tcp.h>
50#include <linux/pagemap.h>
51#include <linux/idr.h>
52#include <linux/file.h>
53#include <linux/sctp.h>
54#include <net/sctp/user.h>
55
56#include "dlm_internal.h"
57#include "lowcomms.h"
58#include "midcomms.h"
59#include "config.h"
60
61#define NEEDED_RMEM (4*1024*1024)
62
63struct cbuf {
64 unsigned int base;
65 unsigned int len;
66 unsigned int mask;
67};
68
69static void cbuf_add(struct cbuf *cb, int n)
70{
71 cb->len += n;
72}
73
74static int cbuf_data(struct cbuf *cb)
75{
76 return ((cb->base + cb->len) & cb->mask);
77}
78
79static void cbuf_init(struct cbuf *cb, int size)
80{
81 cb->base = cb->len = 0;
82 cb->mask = size-1;
83}
84
85static void cbuf_eat(struct cbuf *cb, int n)
86{
87 cb->len -= n;
88 cb->base += n;
89 cb->base &= cb->mask;
90}
91
92static bool cbuf_empty(struct cbuf *cb)
93{
94 return cb->len == 0;
95}
96
97struct connection {
98 struct socket *sock; /* NULL if not connected */
99 uint32_t nodeid; /* So we know who we are in the list */
100 struct mutex sock_mutex;
101 unsigned long flags;
102#define CF_READ_PENDING 1
103#define CF_WRITE_PENDING 2
104#define CF_CONNECT_PENDING 3
105#define CF_INIT_PENDING 4
106#define CF_IS_OTHERCON 5
107 struct list_head writequeue; /* List of outgoing writequeue_entries */
108 spinlock_t writequeue_lock;
109 int (*rx_action) (struct connection *); /* What to do when active */
110 void (*connect_action) (struct connection *); /* What to do to connect */
111 struct page *rx_page;
112 struct cbuf cb;
113 int retries;
114#define MAX_CONNECT_RETRIES 3
115 int sctp_assoc;
116 struct connection *othercon;
117 struct work_struct rwork; /* Receive workqueue */
118 struct work_struct swork; /* Send workqueue */
119};
120#define sock2con(x) ((struct connection *)(x)->sk_user_data)
121
122/* An entry waiting to be sent */
123struct writequeue_entry {
124 struct list_head list;
125 struct page *page;
126 int offset;
127 int len;
128 int end;
129 int users;
130 struct connection *con;
131};
132
133static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
134static int dlm_local_count;
135
136/* Work queues */
137static struct workqueue_struct *recv_workqueue;
138static struct workqueue_struct *send_workqueue;
139
140static DEFINE_IDR(connections_idr);
141static DECLARE_MUTEX(connections_lock);
142static int max_nodeid;
143static struct kmem_cache *con_cache;
144
145static void process_recv_sockets(struct work_struct *work);
146static void process_send_sockets(struct work_struct *work);
147
148/*
149 * If 'allocation' is zero then we don't attempt to create a new
150 * connection structure for this node.
151 */
152static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
153{
154 struct connection *con = NULL;
155 int r;
156 int n;
157
158 con = idr_find(&connections_idr, nodeid);
159 if (con || !alloc)
160 return con;
161
162 r = idr_pre_get(&connections_idr, alloc);
163 if (!r)
164 return NULL;
165
166 con = kmem_cache_zalloc(con_cache, alloc);
167 if (!con)
168 return NULL;
169
170 r = idr_get_new_above(&connections_idr, con, nodeid, &n);
171 if (r) {
172 kmem_cache_free(con_cache, con);
173 return NULL;
174 }
175
176 if (n != nodeid) {
177 idr_remove(&connections_idr, n);
178 kmem_cache_free(con_cache, con);
179 return NULL;
180 }
181
182 con->nodeid = nodeid;
183 mutex_init(&con->sock_mutex);
184 INIT_LIST_HEAD(&con->writequeue);
185 spin_lock_init(&con->writequeue_lock);
186 INIT_WORK(&con->swork, process_send_sockets);
187 INIT_WORK(&con->rwork, process_recv_sockets);
188
189 /* Setup action pointers for child sockets */
190 if (con->nodeid) {
191 struct connection *zerocon = idr_find(&connections_idr, 0);
192
193 con->connect_action = zerocon->connect_action;
194 if (!con->rx_action)
195 con->rx_action = zerocon->rx_action;
196 }
197
198 if (nodeid > max_nodeid)
199 max_nodeid = nodeid;
200
201 return con;
202}
203
204static struct connection *nodeid2con(int nodeid, gfp_t allocation)
205{
206 struct connection *con;
207
208 down(&connections_lock);
209 con = __nodeid2con(nodeid, allocation);
210 up(&connections_lock);
211
212 return con;
213}
214
215/* This is a bit drastic, but only called when things go wrong */
216static struct connection *assoc2con(int assoc_id)
217{
218 int i;
219 struct connection *con;
220
221 down(&connections_lock);
Patrick Caulfield30d3a232007-04-23 16:26:21 +0100222 for (i=0; i<=max_nodeid; i++) {
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100223 con = __nodeid2con(i, 0);
224 if (con && con->sctp_assoc == assoc_id) {
225 up(&connections_lock);
226 return con;
227 }
228 }
229 up(&connections_lock);
230 return NULL;
231}
232
233static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
234{
235 struct sockaddr_storage addr;
236 int error;
237
238 if (!dlm_local_count)
239 return -1;
240
241 error = dlm_nodeid_to_addr(nodeid, &addr);
242 if (error)
243 return error;
244
245 if (dlm_local_addr[0]->ss_family == AF_INET) {
246 struct sockaddr_in *in4 = (struct sockaddr_in *) &addr;
247 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
248 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
249 } else {
250 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &addr;
251 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
252 memcpy(&ret6->sin6_addr, &in6->sin6_addr,
253 sizeof(in6->sin6_addr));
254 }
255
256 return 0;
257}
258
259/* Data available on socket or listen socket received a connect */
260static void lowcomms_data_ready(struct sock *sk, int count_unused)
261{
262 struct connection *con = sock2con(sk);
Patrick Caulfieldafb853f2007-06-01 10:07:26 -0500263 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100264 queue_work(recv_workqueue, &con->rwork);
265}
266
267static void lowcomms_write_space(struct sock *sk)
268{
269 struct connection *con = sock2con(sk);
270
Patrick Caulfieldafb853f2007-06-01 10:07:26 -0500271 if (con && !test_and_set_bit(CF_WRITE_PENDING, &con->flags))
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100272 queue_work(send_workqueue, &con->swork);
273}
274
275static inline void lowcomms_connect_sock(struct connection *con)
276{
277 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
278 queue_work(send_workqueue, &con->swork);
279}
280
281static void lowcomms_state_change(struct sock *sk)
282{
283 if (sk->sk_state == TCP_ESTABLISHED)
284 lowcomms_write_space(sk);
285}
286
287/* Make a socket active */
288static int add_sock(struct socket *sock, struct connection *con)
289{
290 con->sock = sock;
291
292 /* Install a data_ready callback */
293 con->sock->sk->sk_data_ready = lowcomms_data_ready;
294 con->sock->sk->sk_write_space = lowcomms_write_space;
295 con->sock->sk->sk_state_change = lowcomms_state_change;
296 con->sock->sk->sk_user_data = con;
297 return 0;
298}
299
300/* Add the port number to an IPv6 or 4 sockaddr and return the address
301 length */
302static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
303 int *addr_len)
304{
305 saddr->ss_family = dlm_local_addr[0]->ss_family;
306 if (saddr->ss_family == AF_INET) {
307 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
308 in4_addr->sin_port = cpu_to_be16(port);
309 *addr_len = sizeof(struct sockaddr_in);
310 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
311 } else {
312 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
313 in6_addr->sin6_port = cpu_to_be16(port);
314 *addr_len = sizeof(struct sockaddr_in6);
315 }
Patrick Caulfield01c8cab2007-07-17 16:53:15 +0100316 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100317}
318
319/* Close a remote connection and tidy up */
320static void close_connection(struct connection *con, bool and_other)
321{
322 mutex_lock(&con->sock_mutex);
323
324 if (con->sock) {
325 sock_release(con->sock);
326 con->sock = NULL;
327 }
328 if (con->othercon && and_other) {
329 /* Will only re-enter once. */
330 close_connection(con->othercon, false);
331 }
332 if (con->rx_page) {
333 __free_page(con->rx_page);
334 con->rx_page = NULL;
335 }
Patrick Caulfield9e5f2822007-08-02 14:58:14 +0100336
Patrick Caulfield61d96be2007-08-20 15:13:38 +0100337 con->retries = 0;
338 mutex_unlock(&con->sock_mutex);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100339}
340
341/* We only send shutdown messages to nodes that are not part of the cluster */
342static void sctp_send_shutdown(sctp_assoc_t associd)
343{
344 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
345 struct msghdr outmessage;
346 struct cmsghdr *cmsg;
347 struct sctp_sndrcvinfo *sinfo;
348 int ret;
349 struct connection *con;
350
351 con = nodeid2con(0,0);
352 BUG_ON(con == NULL);
353
354 outmessage.msg_name = NULL;
355 outmessage.msg_namelen = 0;
356 outmessage.msg_control = outcmsg;
357 outmessage.msg_controllen = sizeof(outcmsg);
358 outmessage.msg_flags = MSG_EOR;
359
360 cmsg = CMSG_FIRSTHDR(&outmessage);
361 cmsg->cmsg_level = IPPROTO_SCTP;
362 cmsg->cmsg_type = SCTP_SNDRCV;
363 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
364 outmessage.msg_controllen = cmsg->cmsg_len;
365 sinfo = CMSG_DATA(cmsg);
366 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
367
368 sinfo->sinfo_flags |= MSG_EOF;
369 sinfo->sinfo_assoc_id = associd;
370
371 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
372
373 if (ret != 0)
374 log_print("send EOF to node failed: %d", ret);
375}
376
377/* INIT failed but we don't know which node...
378 restart INIT on all pending nodes */
379static void sctp_init_failed(void)
380{
381 int i;
382 struct connection *con;
383
384 down(&connections_lock);
385 for (i=1; i<=max_nodeid; i++) {
386 con = __nodeid2con(i, 0);
387 if (!con)
388 continue;
389 con->sctp_assoc = 0;
390 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
391 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
392 queue_work(send_workqueue, &con->swork);
393 }
394 }
395 }
396 up(&connections_lock);
397}
398
399/* Something happened to an association */
David Teigland617e82e2007-04-26 13:46:49 -0500400static void process_sctp_notification(struct connection *con,
401 struct msghdr *msg, char *buf)
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100402{
403 union sctp_notification *sn = (union sctp_notification *)buf;
404
405 if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
406 switch (sn->sn_assoc_change.sac_state) {
407
408 case SCTP_COMM_UP:
409 case SCTP_RESTART:
410 {
411 /* Check that the new node is in the lockspace */
412 struct sctp_prim prim;
413 int nodeid;
414 int prim_len, ret;
415 int addr_len;
416 struct connection *new_con;
417 struct file *file;
418 sctp_peeloff_arg_t parg;
419 int parglen = sizeof(parg);
420
421 /*
422 * We get this before any data for an association.
423 * We verify that the node is in the cluster and
424 * then peel off a socket for it.
425 */
426 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
427 log_print("COMM_UP for invalid assoc ID %d",
David Teigland617e82e2007-04-26 13:46:49 -0500428 (int)sn->sn_assoc_change.sac_assoc_id);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100429 sctp_init_failed();
430 return;
431 }
432 memset(&prim, 0, sizeof(struct sctp_prim));
433 prim_len = sizeof(struct sctp_prim);
434 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
435
436 ret = kernel_getsockopt(con->sock,
437 IPPROTO_SCTP,
438 SCTP_PRIMARY_ADDR,
439 (char*)&prim,
440 &prim_len);
441 if (ret < 0) {
442 log_print("getsockopt/sctp_primary_addr on "
443 "new assoc %d failed : %d",
444 (int)sn->sn_assoc_change.sac_assoc_id,
445 ret);
446
447 /* Retry INIT later */
448 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
449 if (new_con)
450 clear_bit(CF_CONNECT_PENDING, &con->flags);
451 return;
452 }
453 make_sockaddr(&prim.ssp_addr, 0, &addr_len);
454 if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
455 int i;
456 unsigned char *b=(unsigned char *)&prim.ssp_addr;
457 log_print("reject connect from unknown addr");
458 for (i=0; i<sizeof(struct sockaddr_storage);i++)
459 printk("%02x ", b[i]);
460 printk("\n");
461 sctp_send_shutdown(prim.ssp_assoc_id);
462 return;
463 }
464
465 new_con = nodeid2con(nodeid, GFP_KERNEL);
466 if (!new_con)
467 return;
468
469 /* Peel off a new sock */
470 parg.associd = sn->sn_assoc_change.sac_assoc_id;
David Teigland617e82e2007-04-26 13:46:49 -0500471 ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
472 SCTP_SOCKOPT_PEELOFF,
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100473 (void *)&parg, &parglen);
Patrick Caulfield30d3a232007-04-23 16:26:21 +0100474 if (ret) {
David Teigland617e82e2007-04-26 13:46:49 -0500475 log_print("Can't peel off a socket for "
476 "connection %d to node %d: err=%d\n",
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100477 parg.associd, nodeid, ret);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100478 }
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100479 file = fget(parg.sd);
480 new_con->sock = SOCKET_I(file->f_dentry->d_inode);
481 add_sock(new_con->sock, new_con);
482 fput(file);
483 put_unused_fd(parg.sd);
484
485 log_print("got new/restarted association %d nodeid %d",
David Teigland617e82e2007-04-26 13:46:49 -0500486 (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100487
488 /* Send any pending writes */
489 clear_bit(CF_CONNECT_PENDING, &new_con->flags);
490 clear_bit(CF_INIT_PENDING, &con->flags);
491 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
492 queue_work(send_workqueue, &new_con->swork);
493 }
494 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
495 queue_work(recv_workqueue, &new_con->rwork);
496 }
497 break;
498
499 case SCTP_COMM_LOST:
500 case SCTP_SHUTDOWN_COMP:
501 {
502 con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
503 if (con) {
504 con->sctp_assoc = 0;
505 }
506 }
507 break;
508
509 /* We don't know which INIT failed, so clear the PENDING flags
510 * on them all. if assoc_id is zero then it will then try
511 * again */
512
513 case SCTP_CANT_STR_ASSOC:
514 {
515 log_print("Can't start SCTP association - retrying");
516 sctp_init_failed();
517 }
518 break;
519
520 default:
521 log_print("unexpected SCTP assoc change id=%d state=%d",
522 (int)sn->sn_assoc_change.sac_assoc_id,
523 sn->sn_assoc_change.sac_state);
524 }
525 }
526}
527
528/* Data received from remote end */
529static int receive_from_sock(struct connection *con)
530{
531 int ret = 0;
532 struct msghdr msg = {};
533 struct kvec iov[2];
534 unsigned len;
535 int r;
536 int call_again_soon = 0;
537 int nvec;
538 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
539
540 mutex_lock(&con->sock_mutex);
541
542 if (con->sock == NULL) {
543 ret = -EAGAIN;
544 goto out_close;
545 }
546
547 if (con->rx_page == NULL) {
548 /*
549 * This doesn't need to be atomic, but I think it should
550 * improve performance if it is.
551 */
552 con->rx_page = alloc_page(GFP_ATOMIC);
553 if (con->rx_page == NULL)
554 goto out_resched;
555 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
556 }
557
558 /* Only SCTP needs these really */
559 memset(&incmsg, 0, sizeof(incmsg));
560 msg.msg_control = incmsg;
561 msg.msg_controllen = sizeof(incmsg);
562
563 /*
564 * iov[0] is the bit of the circular buffer between the current end
565 * point (cb.base + cb.len) and the end of the buffer.
566 */
567 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
568 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
569 iov[1].iov_len = 0;
570 nvec = 1;
571
572 /*
573 * iov[1] is the bit of the circular buffer between the start of the
574 * buffer and the start of the currently used section (cb.base)
575 */
576 if (cbuf_data(&con->cb) >= con->cb.base) {
577 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
578 iov[1].iov_len = con->cb.base;
579 iov[1].iov_base = page_address(con->rx_page);
580 nvec = 2;
581 }
582 len = iov[0].iov_len + iov[1].iov_len;
583
584 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
585 MSG_DONTWAIT | MSG_NOSIGNAL);
586 if (ret <= 0)
587 goto out_close;
588
589 /* Process SCTP notifications */
590 if (msg.msg_flags & MSG_NOTIFICATION) {
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100591 msg.msg_control = incmsg;
592 msg.msg_controllen = sizeof(incmsg);
593
594 process_sctp_notification(con, &msg,
David Teigland617e82e2007-04-26 13:46:49 -0500595 page_address(con->rx_page) + con->cb.base);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100596 mutex_unlock(&con->sock_mutex);
597 return 0;
598 }
599 BUG_ON(con->nodeid == 0);
600
601 if (ret == len)
602 call_again_soon = 1;
603 cbuf_add(&con->cb, ret);
604 ret = dlm_process_incoming_buffer(con->nodeid,
605 page_address(con->rx_page),
606 con->cb.base, con->cb.len,
607 PAGE_CACHE_SIZE);
608 if (ret == -EBADMSG) {
David Teigland617e82e2007-04-26 13:46:49 -0500609 log_print("lowcomms: addr=%p, base=%u, len=%u, "
610 "iov_len=%u, iov_base[0]=%p, read=%d",
611 page_address(con->rx_page), con->cb.base, con->cb.len,
612 len, iov[0].iov_base, r);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100613 }
614 if (ret < 0)
615 goto out_close;
616 cbuf_eat(&con->cb, ret);
617
618 if (cbuf_empty(&con->cb) && !call_again_soon) {
619 __free_page(con->rx_page);
620 con->rx_page = NULL;
621 }
622
623 if (call_again_soon)
624 goto out_resched;
625 mutex_unlock(&con->sock_mutex);
626 return 0;
627
628out_resched:
629 if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
630 queue_work(recv_workqueue, &con->rwork);
631 mutex_unlock(&con->sock_mutex);
632 return -EAGAIN;
633
634out_close:
635 mutex_unlock(&con->sock_mutex);
Patrick Caulfield9e5f2822007-08-02 14:58:14 +0100636 if (ret != -EAGAIN) {
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100637 close_connection(con, false);
638 /* Reconnect when there is something to send */
639 }
640 /* Don't return success if we really got EOF */
641 if (ret == 0)
642 ret = -EAGAIN;
643
644 return ret;
645}
646
647/* Listening socket is busy, accept a connection */
648static int tcp_accept_from_sock(struct connection *con)
649{
650 int result;
651 struct sockaddr_storage peeraddr;
652 struct socket *newsock;
653 int len;
654 int nodeid;
655 struct connection *newcon;
656 struct connection *addcon;
657
658 memset(&peeraddr, 0, sizeof(peeraddr));
659 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
660 IPPROTO_TCP, &newsock);
661 if (result < 0)
662 return -ENOMEM;
663
664 mutex_lock_nested(&con->sock_mutex, 0);
665
666 result = -ENOTCONN;
667 if (con->sock == NULL)
668 goto accept_err;
669
670 newsock->type = con->sock->type;
671 newsock->ops = con->sock->ops;
672
673 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
674 if (result < 0)
675 goto accept_err;
676
677 /* Get the connected socket's peer */
678 memset(&peeraddr, 0, sizeof(peeraddr));
679 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
680 &len, 2)) {
681 result = -ECONNABORTED;
682 goto accept_err;
683 }
684
685 /* Get the new node's NODEID */
686 make_sockaddr(&peeraddr, 0, &len);
687 if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
David Teigland617e82e2007-04-26 13:46:49 -0500688 log_print("connect from non cluster node");
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100689 sock_release(newsock);
690 mutex_unlock(&con->sock_mutex);
691 return -1;
692 }
693
694 log_print("got connection from %d", nodeid);
695
696 /* Check to see if we already have a connection to this node. This
697 * could happen if the two nodes initiate a connection at roughly
698 * the same time and the connections cross on the wire.
699 * In this case we store the incoming one in "othercon"
700 */
701 newcon = nodeid2con(nodeid, GFP_KERNEL);
702 if (!newcon) {
703 result = -ENOMEM;
704 goto accept_err;
705 }
706 mutex_lock_nested(&newcon->sock_mutex, 1);
707 if (newcon->sock) {
708 struct connection *othercon = newcon->othercon;
709
710 if (!othercon) {
711 othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
712 if (!othercon) {
David Teigland617e82e2007-04-26 13:46:49 -0500713 log_print("failed to allocate incoming socket");
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100714 mutex_unlock(&newcon->sock_mutex);
715 result = -ENOMEM;
716 goto accept_err;
717 }
718 othercon->nodeid = nodeid;
719 othercon->rx_action = receive_from_sock;
720 mutex_init(&othercon->sock_mutex);
721 INIT_WORK(&othercon->swork, process_send_sockets);
722 INIT_WORK(&othercon->rwork, process_recv_sockets);
723 set_bit(CF_IS_OTHERCON, &othercon->flags);
Patrick Caulfield61d96be2007-08-20 15:13:38 +0100724 }
725 if (!othercon->sock) {
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100726 newcon->othercon = othercon;
Patrick Caulfield97d84832007-06-27 11:36:23 +0100727 othercon->sock = newsock;
728 newsock->sk->sk_user_data = othercon;
729 add_sock(newsock, othercon);
730 addcon = othercon;
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100731 }
Patrick Caulfield97d84832007-06-27 11:36:23 +0100732 else {
733 printk("Extra connection from node %d attempted\n", nodeid);
734 result = -EAGAIN;
akpm@linux-foundation.orgf4fadb23c2007-06-27 14:43:37 -0700735 mutex_unlock(&newcon->sock_mutex);
Patrick Caulfield97d84832007-06-27 11:36:23 +0100736 goto accept_err;
737 }
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100738 }
739 else {
740 newsock->sk->sk_user_data = newcon;
741 newcon->rx_action = receive_from_sock;
742 add_sock(newsock, newcon);
743 addcon = newcon;
744 }
745
746 mutex_unlock(&newcon->sock_mutex);
747
748 /*
749 * Add it to the active queue in case we got data
750 * beween processing the accept adding the socket
751 * to the read_sockets list
752 */
753 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
754 queue_work(recv_workqueue, &addcon->rwork);
755 mutex_unlock(&con->sock_mutex);
756
757 return 0;
758
759accept_err:
760 mutex_unlock(&con->sock_mutex);
761 sock_release(newsock);
762
763 if (result != -EAGAIN)
David Teigland617e82e2007-04-26 13:46:49 -0500764 log_print("error accepting connection from node: %d", result);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100765 return result;
766}
767
768static void free_entry(struct writequeue_entry *e)
769{
770 __free_page(e->page);
771 kfree(e);
772}
773
774/* Initiate an SCTP association.
775 This is a special case of send_to_sock() in that we don't yet have a
776 peeled-off socket for this association, so we use the listening socket
777 and add the primary IP address of the remote node.
778 */
779static void sctp_init_assoc(struct connection *con)
780{
781 struct sockaddr_storage rem_addr;
782 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
783 struct msghdr outmessage;
784 struct cmsghdr *cmsg;
785 struct sctp_sndrcvinfo *sinfo;
786 struct connection *base_con;
787 struct writequeue_entry *e;
788 int len, offset;
789 int ret;
790 int addrlen;
791 struct kvec iov[1];
792
793 if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
794 return;
795
796 if (con->retries++ > MAX_CONNECT_RETRIES)
797 return;
798
799 log_print("Initiating association with node %d", con->nodeid);
800
801 if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
802 log_print("no address for nodeid %d", con->nodeid);
803 return;
804 }
805 base_con = nodeid2con(0, 0);
806 BUG_ON(base_con == NULL);
807
808 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
809
810 outmessage.msg_name = &rem_addr;
811 outmessage.msg_namelen = addrlen;
812 outmessage.msg_control = outcmsg;
813 outmessage.msg_controllen = sizeof(outcmsg);
814 outmessage.msg_flags = MSG_EOR;
815
816 spin_lock(&con->writequeue_lock);
817 e = list_entry(con->writequeue.next, struct writequeue_entry,
818 list);
819
820 BUG_ON((struct list_head *) e == &con->writequeue);
821
822 len = e->len;
823 offset = e->offset;
824 spin_unlock(&con->writequeue_lock);
825 kmap(e->page);
826
827 /* Send the first block off the write queue */
828 iov[0].iov_base = page_address(e->page)+offset;
829 iov[0].iov_len = len;
830
831 cmsg = CMSG_FIRSTHDR(&outmessage);
832 cmsg->cmsg_level = IPPROTO_SCTP;
833 cmsg->cmsg_type = SCTP_SNDRCV;
834 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
835 sinfo = CMSG_DATA(cmsg);
836 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
837 sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
838 outmessage.msg_controllen = cmsg->cmsg_len;
839
840 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
841 if (ret < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500842 log_print("Send first packet to node %d failed: %d",
843 con->nodeid, ret);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100844
845 /* Try again later */
846 clear_bit(CF_CONNECT_PENDING, &con->flags);
847 clear_bit(CF_INIT_PENDING, &con->flags);
848 }
849 else {
850 spin_lock(&con->writequeue_lock);
851 e->offset += ret;
852 e->len -= ret;
853
854 if (e->len == 0 && e->users == 0) {
855 list_del(&e->list);
856 kunmap(e->page);
857 free_entry(e);
858 }
859 spin_unlock(&con->writequeue_lock);
860 }
861}
862
863/* Connect a new socket to its peer */
864static void tcp_connect_to_sock(struct connection *con)
865{
866 int result = -EHOSTUNREACH;
Lon Hohberger6bd8fed2007-10-25 18:51:54 -0400867 struct sockaddr_storage saddr, src_addr;
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100868 int addr_len;
869 struct socket *sock;
870
871 if (con->nodeid == 0) {
872 log_print("attempt to connect sock 0 foiled");
873 return;
874 }
875
876 mutex_lock(&con->sock_mutex);
877 if (con->retries++ > MAX_CONNECT_RETRIES)
878 goto out;
879
880 /* Some odd races can cause double-connects, ignore them */
881 if (con->sock) {
882 result = 0;
883 goto out;
884 }
885
886 /* Create a socket to communicate with */
887 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
888 IPPROTO_TCP, &sock);
889 if (result < 0)
890 goto out_err;
891
892 memset(&saddr, 0, sizeof(saddr));
893 if (dlm_nodeid_to_addr(con->nodeid, &saddr))
894 goto out_err;
895
896 sock->sk->sk_user_data = con;
897 con->rx_action = receive_from_sock;
898 con->connect_action = tcp_connect_to_sock;
899 add_sock(sock, con);
900
Lon Hohberger6bd8fed2007-10-25 18:51:54 -0400901 /* Bind to our cluster-known address connecting to avoid
902 routing problems */
903 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
904 make_sockaddr(&src_addr, 0, &addr_len);
905 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
906 addr_len);
907 if (result < 0) {
908 log_print("could not bind for connect: %d", result);
909 /* This *may* not indicate a critical error */
910 }
911
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100912 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
913
914 log_print("connecting to %d", con->nodeid);
915 result =
916 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
917 O_NONBLOCK);
918 if (result == -EINPROGRESS)
919 result = 0;
920 if (result == 0)
921 goto out;
922
923out_err:
924 if (con->sock) {
925 sock_release(con->sock);
926 con->sock = NULL;
927 }
928 /*
929 * Some errors are fatal and this list might need adjusting. For other
930 * errors we try again until the max number of retries is reached.
931 */
932 if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
933 result != -ENETDOWN && result != EINVAL
934 && result != -EPROTONOSUPPORT) {
935 lowcomms_connect_sock(con);
936 result = 0;
937 }
938out:
939 mutex_unlock(&con->sock_mutex);
940 return;
941}
942
943static struct socket *tcp_create_listen_sock(struct connection *con,
944 struct sockaddr_storage *saddr)
945{
946 struct socket *sock = NULL;
947 int result = 0;
948 int one = 1;
949 int addr_len;
950
951 if (dlm_local_addr[0]->ss_family == AF_INET)
952 addr_len = sizeof(struct sockaddr_in);
953 else
954 addr_len = sizeof(struct sockaddr_in6);
955
956 /* Create a socket to communicate with */
David Teigland617e82e2007-04-26 13:46:49 -0500957 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
958 IPPROTO_TCP, &sock);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100959 if (result < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500960 log_print("Can't create listening comms socket");
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100961 goto create_out;
962 }
963
964 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
965 (char *)&one, sizeof(one));
966
967 if (result < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500968 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100969 }
970 sock->sk->sk_user_data = con;
971 con->rx_action = tcp_accept_from_sock;
972 con->connect_action = tcp_connect_to_sock;
973 con->sock = sock;
974
975 /* Bind to our port */
976 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
977 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
978 if (result < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500979 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100980 sock_release(sock);
981 sock = NULL;
982 con->sock = NULL;
983 goto create_out;
984 }
985 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
986 (char *)&one, sizeof(one));
987 if (result < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500988 log_print("Set keepalive failed: %d", result);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100989 }
990
991 result = sock->ops->listen(sock, 5);
992 if (result < 0) {
David Teigland617e82e2007-04-26 13:46:49 -0500993 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +0100994 sock_release(sock);
995 sock = NULL;
996 goto create_out;
997 }
998
999create_out:
1000 return sock;
1001}
1002
1003/* Get local addresses */
1004static void init_local(void)
1005{
1006 struct sockaddr_storage sas, *addr;
1007 int i;
1008
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001009 dlm_local_count = 0;
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001010 for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1011 if (dlm_our_addr(&sas, i))
1012 break;
1013
1014 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1015 if (!addr)
1016 break;
1017 memcpy(addr, &sas, sizeof(*addr));
1018 dlm_local_addr[dlm_local_count++] = addr;
1019 }
1020}
1021
David Teigland617e82e2007-04-26 13:46:49 -05001022/* Bind to an IP address. SCTP allows multiple address so it can do
1023 multi-homing */
1024static int add_sctp_bind_addr(struct connection *sctp_con,
1025 struct sockaddr_storage *addr,
1026 int addr_len, int num)
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001027{
1028 int result = 0;
1029
1030 if (num == 1)
1031 result = kernel_bind(sctp_con->sock,
1032 (struct sockaddr *) addr,
1033 addr_len);
1034 else
1035 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1036 SCTP_SOCKOPT_BINDX_ADD,
1037 (char *)addr, addr_len);
1038
1039 if (result < 0)
1040 log_print("Can't bind to port %d addr number %d",
1041 dlm_config.ci_tcp_port, num);
1042
1043 return result;
1044}
1045
1046/* Initialise SCTP socket and bind to all interfaces */
1047static int sctp_listen_for_all(void)
1048{
1049 struct socket *sock = NULL;
1050 struct sockaddr_storage localaddr;
1051 struct sctp_event_subscribe subscribe;
1052 int result = -EINVAL, num = 1, i, addr_len;
1053 struct connection *con = nodeid2con(0, GFP_KERNEL);
1054 int bufsize = NEEDED_RMEM;
1055
1056 if (!con)
1057 return -ENOMEM;
1058
1059 log_print("Using SCTP for communications");
1060
1061 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1062 IPPROTO_SCTP, &sock);
1063 if (result < 0) {
1064 log_print("Can't create comms socket, check SCTP is loaded");
1065 goto out;
1066 }
1067
1068 /* Listen for events */
1069 memset(&subscribe, 0, sizeof(subscribe));
1070 subscribe.sctp_data_io_event = 1;
1071 subscribe.sctp_association_event = 1;
1072 subscribe.sctp_send_failure_event = 1;
1073 subscribe.sctp_shutdown_event = 1;
1074 subscribe.sctp_partial_delivery_event = 1;
1075
David S. Millerdf61c952007-11-06 23:48:57 -08001076 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001077 (char *)&bufsize, sizeof(bufsize));
1078 if (result)
David Teigland617e82e2007-04-26 13:46:49 -05001079 log_print("Error increasing buffer space on socket %d", result);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001080
1081 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
David Teigland617e82e2007-04-26 13:46:49 -05001082 (char *)&subscribe, sizeof(subscribe));
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001083 if (result < 0) {
1084 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1085 result);
1086 goto create_delsock;
1087 }
1088
1089 /* Init con struct */
1090 sock->sk->sk_user_data = con;
1091 con->sock = sock;
1092 con->sock->sk->sk_data_ready = lowcomms_data_ready;
1093 con->rx_action = receive_from_sock;
1094 con->connect_action = sctp_init_assoc;
1095
1096 /* Bind to all interfaces. */
1097 for (i = 0; i < dlm_local_count; i++) {
1098 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1099 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1100
1101 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1102 if (result)
1103 goto create_delsock;
1104 ++num;
1105 }
1106
1107 result = sock->ops->listen(sock, 5);
1108 if (result < 0) {
1109 log_print("Can't set socket listening");
1110 goto create_delsock;
1111 }
1112
1113 return 0;
1114
1115create_delsock:
1116 sock_release(sock);
1117 con->sock = NULL;
1118out:
1119 return result;
1120}
1121
1122static int tcp_listen_for_all(void)
1123{
1124 struct socket *sock = NULL;
1125 struct connection *con = nodeid2con(0, GFP_KERNEL);
1126 int result = -EINVAL;
1127
1128 if (!con)
1129 return -ENOMEM;
1130
1131 /* We don't support multi-homed hosts */
1132 if (dlm_local_addr[1] != NULL) {
David Teigland617e82e2007-04-26 13:46:49 -05001133 log_print("TCP protocol can't handle multi-homed hosts, "
1134 "try SCTP");
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001135 return -EINVAL;
1136 }
1137
1138 log_print("Using TCP for communications");
1139
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001140 sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1141 if (sock) {
1142 add_sock(sock, con);
1143 result = 0;
1144 }
1145 else {
1146 result = -EADDRINUSE;
1147 }
1148
1149 return result;
1150}
1151
1152
1153
1154static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1155 gfp_t allocation)
1156{
1157 struct writequeue_entry *entry;
1158
1159 entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1160 if (!entry)
1161 return NULL;
1162
1163 entry->page = alloc_page(allocation);
1164 if (!entry->page) {
1165 kfree(entry);
1166 return NULL;
1167 }
1168
1169 entry->offset = 0;
1170 entry->len = 0;
1171 entry->end = 0;
1172 entry->users = 0;
1173 entry->con = con;
1174
1175 return entry;
1176}
1177
David Teigland617e82e2007-04-26 13:46:49 -05001178void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001179{
1180 struct connection *con;
1181 struct writequeue_entry *e;
1182 int offset = 0;
1183 int users = 0;
1184
1185 con = nodeid2con(nodeid, allocation);
1186 if (!con)
1187 return NULL;
1188
1189 spin_lock(&con->writequeue_lock);
1190 e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1191 if ((&e->list == &con->writequeue) ||
1192 (PAGE_CACHE_SIZE - e->end < len)) {
1193 e = NULL;
1194 } else {
1195 offset = e->end;
1196 e->end += len;
1197 users = e->users++;
1198 }
1199 spin_unlock(&con->writequeue_lock);
1200
1201 if (e) {
1202 got_one:
1203 if (users == 0)
1204 kmap(e->page);
1205 *ppc = page_address(e->page) + offset;
1206 return e;
1207 }
1208
1209 e = new_writequeue_entry(con, allocation);
1210 if (e) {
1211 spin_lock(&con->writequeue_lock);
1212 offset = e->end;
1213 e->end += len;
1214 users = e->users++;
1215 list_add_tail(&e->list, &con->writequeue);
1216 spin_unlock(&con->writequeue_lock);
1217 goto got_one;
1218 }
1219 return NULL;
1220}
1221
1222void dlm_lowcomms_commit_buffer(void *mh)
1223{
1224 struct writequeue_entry *e = (struct writequeue_entry *)mh;
1225 struct connection *con = e->con;
1226 int users;
1227
1228 spin_lock(&con->writequeue_lock);
1229 users = --e->users;
1230 if (users)
1231 goto out;
1232 e->len = e->end - e->offset;
1233 kunmap(e->page);
1234 spin_unlock(&con->writequeue_lock);
1235
1236 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1237 queue_work(send_workqueue, &con->swork);
1238 }
1239 return;
1240
1241out:
1242 spin_unlock(&con->writequeue_lock);
1243 return;
1244}
1245
1246/* Send a message */
1247static void send_to_sock(struct connection *con)
1248{
1249 int ret = 0;
1250 ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1251 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1252 struct writequeue_entry *e;
1253 int len, offset;
1254
1255 mutex_lock(&con->sock_mutex);
1256 if (con->sock == NULL)
1257 goto out_connect;
1258
1259 sendpage = con->sock->ops->sendpage;
1260
1261 spin_lock(&con->writequeue_lock);
1262 for (;;) {
1263 e = list_entry(con->writequeue.next, struct writequeue_entry,
1264 list);
1265 if ((struct list_head *) e == &con->writequeue)
1266 break;
1267
1268 len = e->len;
1269 offset = e->offset;
1270 BUG_ON(len == 0 && e->users == 0);
1271 spin_unlock(&con->writequeue_lock);
1272 kmap(e->page);
1273
1274 ret = 0;
1275 if (len) {
1276 ret = sendpage(con->sock, e->page, offset, len,
1277 msg_flags);
Patrick Caulfieldd66f8272007-09-14 08:49:21 +01001278 if (ret == -EAGAIN || ret == 0) {
1279 cond_resched();
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001280 goto out;
Patrick Caulfieldd66f8272007-09-14 08:49:21 +01001281 }
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001282 if (ret <= 0)
1283 goto send_error;
Patrick Caulfieldd66f8272007-09-14 08:49:21 +01001284 }
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001285 /* Don't starve people filling buffers */
1286 cond_resched();
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001287
1288 spin_lock(&con->writequeue_lock);
1289 e->offset += ret;
1290 e->len -= ret;
1291
1292 if (e->len == 0 && e->users == 0) {
1293 list_del(&e->list);
1294 kunmap(e->page);
1295 free_entry(e);
1296 continue;
1297 }
1298 }
1299 spin_unlock(&con->writequeue_lock);
1300out:
1301 mutex_unlock(&con->sock_mutex);
1302 return;
1303
1304send_error:
1305 mutex_unlock(&con->sock_mutex);
1306 close_connection(con, false);
1307 lowcomms_connect_sock(con);
1308 return;
1309
1310out_connect:
1311 mutex_unlock(&con->sock_mutex);
1312 if (!test_bit(CF_INIT_PENDING, &con->flags))
1313 lowcomms_connect_sock(con);
1314 return;
1315}
1316
1317static void clean_one_writequeue(struct connection *con)
1318{
1319 struct list_head *list;
1320 struct list_head *temp;
1321
1322 spin_lock(&con->writequeue_lock);
1323 list_for_each_safe(list, temp, &con->writequeue) {
1324 struct writequeue_entry *e =
1325 list_entry(list, struct writequeue_entry, list);
1326 list_del(&e->list);
1327 free_entry(e);
1328 }
1329 spin_unlock(&con->writequeue_lock);
1330}
1331
1332/* Called from recovery when it knows that a node has
1333 left the cluster */
1334int dlm_lowcomms_close(int nodeid)
1335{
1336 struct connection *con;
1337
1338 log_print("closing connection to node %d", nodeid);
1339 con = nodeid2con(nodeid, 0);
1340 if (con) {
1341 clean_one_writequeue(con);
1342 close_connection(con, true);
1343 }
1344 return 0;
1345}
1346
1347/* Receive workqueue function */
1348static void process_recv_sockets(struct work_struct *work)
1349{
1350 struct connection *con = container_of(work, struct connection, rwork);
1351 int err;
1352
1353 clear_bit(CF_READ_PENDING, &con->flags);
1354 do {
1355 err = con->rx_action(con);
1356 } while (!err);
1357}
1358
1359/* Send workqueue function */
1360static void process_send_sockets(struct work_struct *work)
1361{
1362 struct connection *con = container_of(work, struct connection, swork);
1363
1364 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1365 con->connect_action(con);
1366 }
1367 clear_bit(CF_WRITE_PENDING, &con->flags);
1368 send_to_sock(con);
1369}
1370
1371
1372/* Discard all entries on the write queues */
1373static void clean_writequeues(void)
1374{
1375 int nodeid;
1376
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001377 for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1378 struct connection *con = __nodeid2con(nodeid, 0);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001379
1380 if (con)
1381 clean_one_writequeue(con);
1382 }
1383}
1384
1385static void work_stop(void)
1386{
1387 destroy_workqueue(recv_workqueue);
1388 destroy_workqueue(send_workqueue);
1389}
1390
1391static int work_start(void)
1392{
1393 int error;
1394 recv_workqueue = create_workqueue("dlm_recv");
1395 error = IS_ERR(recv_workqueue);
1396 if (error) {
1397 log_print("can't start dlm_recv %d", error);
1398 return error;
1399 }
1400
1401 send_workqueue = create_singlethread_workqueue("dlm_send");
1402 error = IS_ERR(send_workqueue);
1403 if (error) {
1404 log_print("can't start dlm_send %d", error);
1405 destroy_workqueue(recv_workqueue);
1406 return error;
1407 }
1408
1409 return 0;
1410}
1411
1412void dlm_lowcomms_stop(void)
1413{
1414 int i;
1415 struct connection *con;
1416
1417 /* Set all the flags to prevent any
1418 socket activity.
1419 */
1420 down(&connections_lock);
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001421 for (i = 0; i <= max_nodeid; i++) {
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001422 con = __nodeid2con(i, 0);
Patrick Caulfieldafb853f2007-06-01 10:07:26 -05001423 if (con) {
Patrick Caulfield9e5f2822007-08-02 14:58:14 +01001424 con->flags |= 0x0F;
Patrick Caulfieldafb853f2007-06-01 10:07:26 -05001425 if (con->sock)
1426 con->sock->sk->sk_user_data = NULL;
1427 }
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001428 }
1429 up(&connections_lock);
1430
1431 work_stop();
1432
1433 down(&connections_lock);
1434 clean_writequeues();
1435
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001436 for (i = 0; i <= max_nodeid; i++) {
Josef Bacik2439fe502007-04-19 17:59:05 -04001437 con = __nodeid2con(i, 0);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001438 if (con) {
1439 close_connection(con, true);
Patrick Caulfeld39bd4172008-01-09 15:06:27 +00001440 if (con->othercon)
1441 kmem_cache_free(con_cache, con->othercon);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001442 kmem_cache_free(con_cache, con);
1443 }
1444 }
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001445 max_nodeid = 0;
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001446 up(&connections_lock);
1447 kmem_cache_destroy(con_cache);
Patrick Caulfield30d3a232007-04-23 16:26:21 +01001448 idr_init(&connections_idr);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001449}
1450
1451int dlm_lowcomms_start(void)
1452{
1453 int error = -EINVAL;
1454 struct connection *con;
1455
1456 init_local();
1457 if (!dlm_local_count) {
David Teigland617e82e2007-04-26 13:46:49 -05001458 error = -ENOTCONN;
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001459 log_print("no local IP address has been set");
1460 goto out;
1461 }
1462
1463 error = -ENOMEM;
1464 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1465 __alignof__(struct connection), 0,
Paul Mundt20c2df82007-07-20 10:11:58 +09001466 NULL);
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001467 if (!con_cache)
1468 goto out;
1469
Patrick Caulfield6ed7257b2007-04-17 15:39:57 +01001470 /* Start listening */
1471 if (dlm_config.ci_protocol == 0)
1472 error = tcp_listen_for_all();
1473 else
1474 error = sctp_listen_for_all();
1475 if (error)
1476 goto fail_unlisten;
1477
1478 error = work_start();
1479 if (error)
1480 goto fail_unlisten;
1481
1482 return 0;
1483
1484fail_unlisten:
1485 con = nodeid2con(0,0);
1486 if (con) {
1487 close_connection(con, false);
1488 kmem_cache_free(con_cache, con);
1489 }
1490 kmem_cache_destroy(con_cache);
1491
1492out:
1493 return error;
1494}