fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2009 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 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.
  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 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.
  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/file.h>
  52 #include <linux/mutex.h>
  53 #include <linux/sctp.h>
  54 #include <linux/slab.h>
  55 #include <net/sctp/sctp.h>
  56 #include <net/ipv6.h>
  57
  58 #include "dlm_internal.h"
  59 #include "lowcomms.h"
  60 #include "midcomms.h"
  61 #include "config.h"
  62
  63 #define NEEDED_RMEM (4*1024*1024)
  64 #define CONN_HASH_SIZE 32
  65
  66 /* Number of messages to send before rescheduling */
  67 #define MAX_SEND_MSG_COUNT 25
  68
  69 struct cbuf {
  70         unsigned int base;
  71         unsigned int len;
  72         unsigned int mask;
  73 };
  74
  75 static void cbuf_add(struct cbuf *cb, int n)
  76 {
  77         cb->len += n;
  78 }
  79
  80 static int cbuf_data(struct cbuf *cb)
  81 {
  82         return ((cb->base + cb->len) & cb->mask);
  83 }
  84
  85 static void cbuf_init(struct cbuf *cb, int size)
  86 {
  87         cb->base = cb->len = 0;
  88         cb->mask = size-1;
  89 }
  90
  91 static void cbuf_eat(struct cbuf *cb, int n)
  92 {
  93         cb->len  -= n;
  94         cb->base += n;
  95         cb->base &= cb->mask;
  96 }
  97
  98 static bool cbuf_empty(struct cbuf *cb)
  99 {
 100         return cb->len == 0;
 101 }
 102
 103 struct connection {
 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;
 107         unsigned long flags;
 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
 113 #define CF_CLOSE 6
 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;
 120         struct cbuf cb;
 121         int retries;
 122 #define MAX_CONNECT_RETRIES 3
 123         int sctp_assoc;
 124         struct hlist_node list;
 125         struct connection *othercon;
 126         struct work_struct rwork; /* Receive workqueue */
 127         struct work_struct swork; /* Send workqueue */
 128         bool try_new_addr;
 129 };
 130 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 131
 132 /* An entry waiting to be sent */
 133 struct writequeue_entry {
 134         struct list_head list;
 135         struct page *page;
 136         int offset;
 137         int len;
 138         int end;
 139         int users;
 140         struct connection *con;
 141 };
 142
 143 struct dlm_node_addr {
 144         struct list_head list;
 145         int nodeid;
 146         int addr_count;
 147         int curr_addr_index;
 148         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 149 };
 150
 151 static LIST_HEAD(dlm_node_addrs);
 152 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 153
 154 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 155 static int dlm_local_count;
 156 static int dlm_allow_conn;
 157
 158 /* Work queues */
 159 static struct workqueue_struct *recv_workqueue;
 160 static struct workqueue_struct *send_workqueue;
 161
 162 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 163 static DEFINE_MUTEX(connections_lock);
 164 static struct kmem_cache *con_cache;
 165
 166 static void process_recv_sockets(struct work_struct *work);
 167 static void process_send_sockets(struct work_struct *work);
 168
 169
 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)
 174 {
 175         return nodeid & (CONN_HASH_SIZE-1);
 176 }
 177
 178 static struct connection *__find_con(int nodeid)
 179 {
 180         int r;
 181         struct connection *con;
 182
 183         r = nodeid_hash(nodeid);
 184
 185         hlist_for_each_entry(con, &connection_hash[r], list) {
 186                 if (con->nodeid == nodeid)
 187                         return con;
 188         }
 189         return NULL;
 190 }
 191
 192 /*
 193  * If 'allocation' is zero then we don't attempt to create a new
 194  * connection structure for this node.
 195  */
 196 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 197 {
 198         struct connection *con = NULL;
 199         int r;
 200
 201         con = __find_con(nodeid);
 202         if (con || !alloc)
 203                 return con;
 204
 205         con = kmem_cache_zalloc(con_cache, alloc);
 206         if (!con)
 207                 return NULL;
 208
 209         r = nodeid_hash(nodeid);
 210         hlist_add_head(&con->list, &connection_hash[r]);
 211
 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);
 218
 219         /* Setup action pointers for child sockets */
 220         if (con->nodeid) {
 221                 struct connection *zerocon = __find_con(0);
 222
 223                 con->connect_action = zerocon->connect_action;
 224                 if (!con->rx_action)
 225                         con->rx_action = zerocon->rx_action;
 226         }
 227
 228         return con;
 229 }
 230
 231 /* Loop round all connections */
 232 static void foreach_conn(void (*conn_func)(struct connection *c))
 233 {
 234         int i;
 235         struct hlist_node *n;
 236         struct connection *con;
 237
 238         for (i = 0; i < CONN_HASH_SIZE; i++) {
 239                 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
 240                         conn_func(con);
 241         }
 242 }
 243
 244 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 245 {
 246         struct connection *con;
 247
 248         mutex_lock(&connections_lock);
 249         con = __nodeid2con(nodeid, allocation);
 250         mutex_unlock(&connections_lock);
 251
 252         return con;
 253 }
 254
 255 /* This is a bit drastic, but only called when things go wrong */
 256 static struct connection *assoc2con(int assoc_id)
 257 {
 258         int i;
 259         struct connection *con;
 260
 261         mutex_lock(&connections_lock);
 262
 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);
 267                                 return con;
 268                         }
 269                 }
 270         }
 271         mutex_unlock(&connections_lock);
 272         return NULL;
 273 }
 274
 275 static struct dlm_node_addr *find_node_addr(int nodeid)
 276 {
 277         struct dlm_node_addr *na;
 278
 279         list_for_each_entry(na, &dlm_node_addrs, list) {
 280                 if (na->nodeid == nodeid)
 281                         return na;
 282         }
 283         return NULL;
 284 }
 285
 286 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
 287 {
 288         switch (x->ss_family) {
 289         case AF_INET: {
 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)
 293                         return 0;
 294                 if (sinx->sin_port != siny->sin_port)
 295                         return 0;
 296                 break;
 297         }
 298         case AF_INET6: {
 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))
 302                         return 0;
 303                 if (sinx->sin6_port != siny->sin6_port)
 304                         return 0;
 305                 break;
 306         }
 307         default:
 308                 return 0;
 309         }
 310         return 1;
 311 }
 312
 313 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 314                           struct sockaddr *sa_out, bool try_new_addr)
 315 {
 316         struct sockaddr_storage sas;
 317         struct dlm_node_addr *na;
 318
 319         if (!dlm_local_count)
 320                 return -1;
 321
 322         spin_lock(&dlm_node_addrs_spin);
 323         na = find_node_addr(nodeid);
 324         if (na && na->addr_count) {
 325                 if (try_new_addr) {
 326                         na->curr_addr_index++;
 327                         if (na->curr_addr_index == na->addr_count)
 328                                 na->curr_addr_index = 0;
 329                 }
 330
 331                 memcpy(&sas, na->addr[na->curr_addr_index ],
 332                         sizeof(struct sockaddr_storage));
 333         }
 334         spin_unlock(&dlm_node_addrs_spin);
 335
 336         if (!na)
 337                 return -EEXIST;
 338
 339         if (!na->addr_count)
 340                 return -ENOENT;
 341
 342         if (sas_out)
 343                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 344
 345         if (!sa_out)
 346                 return 0;
 347
 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;
 352         } else {
 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;
 356         }
 357
 358         return 0;
 359 }
 360
 361 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
 362 {
 363         struct dlm_node_addr *na;
 364         int rv = -EEXIST;
 365         int addr_i;
 366
 367         spin_lock(&dlm_node_addrs_spin);
 368         list_for_each_entry(na, &dlm_node_addrs, list) {
 369                 if (!na->addr_count)
 370                         continue;
 371
 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;
 375                                 rv = 0;
 376                                 goto unlock;
 377                         }
 378                 }
 379         }
 380 unlock:
 381         spin_unlock(&dlm_node_addrs_spin);
 382         return rv;
 383 }
 384
 385 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 386 {
 387         struct sockaddr_storage *new_addr;
 388         struct dlm_node_addr *new_node, *na;
 389
 390         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 391         if (!new_node)
 392                 return -ENOMEM;
 393
 394         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 395         if (!new_addr) {
 396                 kfree(new_node);
 397                 return -ENOMEM;
 398         }
 399
 400         memcpy(new_addr, addr, len);
 401
 402         spin_lock(&dlm_node_addrs_spin);
 403         na = find_node_addr(nodeid);
 404         if (!na) {
 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);
 410                 return 0;
 411         }
 412
 413         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 414                 spin_unlock(&dlm_node_addrs_spin);
 415                 kfree(new_addr);
 416                 kfree(new_node);
 417                 return -ENOSPC;
 418         }
 419
 420         na->addr[na->addr_count++] = new_addr;
 421         spin_unlock(&dlm_node_addrs_spin);
 422         kfree(new_node);
 423         return 0;
 424 }
 425
 426 /* Data available on socket or listen socket received a connect */
 427 static void lowcomms_data_ready(struct sock *sk)
 428 {
 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);
 432 }
 433
 434 static void lowcomms_write_space(struct sock *sk)
 435 {
 436         struct connection *con = sock2con(sk);
 437
 438         if (!con)
 439                 return;
 440
 441         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 442
 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);
 446         }
 447
 448         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 449                 queue_work(send_workqueue, &con->swork);
 450 }
 451
 452 static inline void lowcomms_connect_sock(struct connection *con)
 453 {
 454         if (test_bit(CF_CLOSE, &con->flags))
 455                 return;
 456         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 457                 queue_work(send_workqueue, &con->swork);
 458 }
 459
 460 static void lowcomms_state_change(struct sock *sk)
 461 {
 462         if (sk->sk_state == TCP_ESTABLISHED)
 463                 lowcomms_write_space(sk);
 464 }
 465
 466 int dlm_lowcomms_connect_node(int nodeid)
 467 {
 468         struct connection *con;
 469
 470         /* with sctp there's no connecting without sending */
 471         if (dlm_config.ci_protocol != 0)
 472                 return 0;
 473
 474         if (nodeid == dlm_our_nodeid())
 475                 return 0;
 476
 477         con = nodeid2con(nodeid, GFP_NOFS);
 478         if (!con)
 479                 return -ENOMEM;
 480         lowcomms_connect_sock(con);
 481         return 0;
 482 }
 483
 484 /* Make a socket active */
 485 static void add_sock(struct socket *sock, struct connection *con)
 486 {
 487         con->sock = sock;
 488
 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;
 495 }
 496
 497 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 498    length */
 499 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 500                           int *addr_len)
 501 {
 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));
 508         } else {
 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);
 512         }
 513         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 514 }
 515
 516 /* Close a remote connection and tidy up */
 517 static void close_connection(struct connection *con, bool and_other,
 518                              bool tx, bool rx)
 519 {
 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);
 526
 527         mutex_lock(&con->sock_mutex);
 528         if (con->sock) {
 529                 sock_release(con->sock);
 530                 con->sock = NULL;
 531         }
 532         if (con->othercon && and_other) {
 533                 /* Will only re-enter once. */
 534                 close_connection(con->othercon, false, true, true);
 535         }
 536         if (con->rx_page) {
 537                 __free_page(con->rx_page);
 538                 con->rx_page = NULL;
 539         }
 540
 541         con->retries = 0;
 542         mutex_unlock(&con->sock_mutex);
 543 }
 544
 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.
 547  */
 548 static void sctp_send_shutdown(sctp_assoc_t associd)
 549 {
 550         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 551         struct msghdr outmessage;
 552         struct cmsghdr *cmsg;
 553         struct sctp_sndrcvinfo *sinfo;
 554         int ret;
 555         struct connection *con;
 556
 557         con = nodeid2con(0,0);
 558         BUG_ON(con == NULL);
 559
 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;
 565
 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));
 573
 574         sinfo->sinfo_flags |= MSG_EOF;
 575         sinfo->sinfo_assoc_id = associd;
 576
 577         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 578
 579         if (ret != 0)
 580                 log_print("send EOF to node failed: %d", ret);
 581 }
 582
 583 static void sctp_init_failed_foreach(struct connection *con)
 584 {
 585
 586         /*
 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.
 593          */
 594         if (!con->nodeid || con->sctp_assoc)
 595                 return;
 596
 597         log_print("Retrying SCTP association init for node %d\n", con->nodeid);
 598
 599         con->try_new_addr = true;
 600         con->sctp_assoc = 0;
 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);
 604         }
 605 }
 606
 607 /* INIT failed but we don't know which node...
 608    restart INIT on all pending nodes */
 609 static void sctp_init_failed(void)
 610 {
 611         mutex_lock(&connections_lock);
 612
 613         foreach_conn(sctp_init_failed_foreach);
 614
 615         mutex_unlock(&connections_lock);
 616 }
 617
 618 static void retry_failed_sctp_send(struct connection *recv_con,
 619                                    struct sctp_send_failed *sn_send_failed,
 620                                    char *buf)
 621 {
 622         int len = sn_send_failed->ssf_length - sizeof(struct sctp_send_failed);
 623         struct dlm_mhandle *mh;
 624         struct connection *con;
 625         char *retry_buf;
 626         int nodeid = sn_send_failed->ssf_info.sinfo_ppid;
 627
 628         log_print("Retry sending %d bytes to node id %d", len, nodeid);
 629
 630         if (!nodeid) {
 631                 log_print("Shouldn't resend data via listening connection.");
 632                 return;
 633         }
 634
 635         con = nodeid2con(nodeid, 0);
 636         if (!con) {
 637                 log_print("Could not look up con for nodeid %d\n",
 638                           nodeid);
 639                 return;
 640         }
 641
 642         mh = dlm_lowcomms_get_buffer(nodeid, len, GFP_NOFS, &retry_buf);
 643         if (!mh) {
 644                 log_print("Could not allocate buf for retry.");
 645                 return;
 646         }
 647         memcpy(retry_buf, buf + sizeof(struct sctp_send_failed), len);
 648         dlm_lowcomms_commit_buffer(mh);
 649
 650         /*
 651          * If we got a assoc changed event before the send failed event then
 652          * we only need to retry the send.
 653          */
 654         if (con->sctp_assoc) {
 655                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 656                         queue_work(send_workqueue, &con->swork);
 657         } else
 658                 sctp_init_failed_foreach(con);
 659 }
 660
 661 /* Something happened to an association */
 662 static void process_sctp_notification(struct connection *con,
 663                                       struct msghdr *msg, char *buf)
 664 {
 665         union sctp_notification *sn = (union sctp_notification *)buf;
 666         struct linger linger;
 667
 668         switch (sn->sn_header.sn_type) {
 669         case SCTP_SEND_FAILED:
 670                 retry_failed_sctp_send(con, &sn->sn_send_failed, buf);
 671                 break;
 672         case SCTP_ASSOC_CHANGE:
 673                 switch (sn->sn_assoc_change.sac_state) {
 674                 case SCTP_COMM_UP:
 675                 case SCTP_RESTART:
 676                 {
 677                         /* Check that the new node is in the lockspace */
 678                         struct sctp_prim prim;
 679                         int nodeid;
 680                         int prim_len, ret;
 681                         int addr_len;
 682                         struct connection *new_con;
 683
 684                         /*
 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.
 688                          */
 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);
 692                                 sctp_init_failed();
 693                                 return;
 694                         }
 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;
 698
 699                         ret = kernel_getsockopt(con->sock,
 700                                                 IPPROTO_SCTP,
 701                                                 SCTP_PRIMARY_ADDR,
 702                                                 (char*)&prim,
 703                                                 &prim_len);
 704                         if (ret < 0) {
 705                                 log_print("getsockopt/sctp_primary_addr on "
 706                                           "new assoc %d failed : %d",
 707                                           (int)sn->sn_assoc_change.sac_assoc_id,
 708                                           ret);
 709
 710                                 /* Retry INIT later */
 711                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 712                                 if (new_con)
 713                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 714                                 return;
 715                         }
 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);
 723                                 return;
 724                         }
 725
 726                         new_con = nodeid2con(nodeid, GFP_NOFS);
 727                         if (!new_con)
 728                                 return;
 729
 730                         if (new_con->sock) {
 731                                 log_print("reject connect from node %d: "
 732                                           "already has a connection.",
 733                                           nodeid);
 734                                 sctp_send_shutdown(prim.ssp_assoc_id);
 735                                 return;
 736                         }
 737
 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,
 742                                 &new_con->sock);
 743                         release_sock(con->sock->sk);
 744                         if (ret < 0) {
 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,
 748                                           nodeid, ret);
 749                                 return;
 750                         }
 751                         add_sock(new_con->sock, new_con);
 752
 753                         linger.l_onoff = 1;
 754                         linger.l_linger = 0;
 755                         ret = kernel_setsockopt(new_con->sock, SOL_SOCKET, SO_LINGER,
 756                                                 (char *)&linger, sizeof(linger));
 757                         if (ret < 0)
 758                                 log_print("set socket option SO_LINGER failed");
 759
 760                         log_print("connecting to %d sctp association %d",
 761                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 762
 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);
 770                         }
 771                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 772                                 queue_work(recv_workqueue, &new_con->rwork);
 773                 }
 774                 break;
 775
 776                 case SCTP_COMM_LOST:
 777                 case SCTP_SHUTDOWN_COMP:
 778                 {
 779                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 780                         if (con) {
 781                                 con->sctp_assoc = 0;
 782                         }
 783                 }
 784                 break;
 785
 786                 case SCTP_CANT_STR_ASSOC:
 787                 {
 788                         /* Will retry init when we get the send failed notification */
 789                         log_print("Can't start SCTP association - retrying");
 790                 }
 791                 break;
 792
 793                 default:
 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);
 797                 }
 798         default:
 799                 ; /* fall through */
 800         }
 801 }
 802
 803 /* Data received from remote end */
 804 static int receive_from_sock(struct connection *con)
 805 {
 806         int ret = 0;
 807         struct msghdr msg = {};
 808         struct kvec iov[2];
 809         unsigned len;
 810         int r;
 811         int call_again_soon = 0;
 812         int nvec;
 813         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 814
 815         mutex_lock(&con->sock_mutex);
 816
 817         if (con->sock == NULL) {
 818                 ret = -EAGAIN;
 819                 goto out_close;
 820         }
 821
 822         if (con->rx_page == NULL) {
 823                 /*
 824                  * This doesn't need to be atomic, but I think it should
 825                  * improve performance if it is.
 826                  */
 827                 con->rx_page = alloc_page(GFP_ATOMIC);
 828                 if (con->rx_page == NULL)
 829                         goto out_resched;
 830                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 831         }
 832
 833         /* Only SCTP needs these really */
 834         memset(&incmsg, 0, sizeof(incmsg));
 835         msg.msg_control = incmsg;
 836         msg.msg_controllen = sizeof(incmsg);
 837
 838         /*
 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.
 841          */
 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);
 844         iov[1].iov_len = 0;
 845         nvec = 1;
 846
 847         /*
 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)
 850          */
 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);
 855                 nvec = 2;
 856         }
 857         len = iov[0].iov_len + iov[1].iov_len;
 858
 859         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 860                                MSG_DONTWAIT | MSG_NOSIGNAL);
 861         if (ret <= 0)
 862                 goto out_close;
 863
 864         /* Process SCTP notifications */
 865         if (msg.msg_flags & MSG_NOTIFICATION) {
 866                 msg.msg_control = incmsg;
 867                 msg.msg_controllen = sizeof(incmsg);
 868
 869                 process_sctp_notification(con, &msg,
 870                                 page_address(con->rx_page) + con->cb.base);
 871                 mutex_unlock(&con->sock_mutex);
 872                 return 0;
 873         }
 874         BUG_ON(con->nodeid == 0);
 875
 876         if (ret == len)
 877                 call_again_soon = 1;
 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,
 882                                           PAGE_CACHE_SIZE);
 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);
 888         }
 889         if (ret < 0)
 890                 goto out_close;
 891         cbuf_eat(&con->cb, ret);
 892
 893         if (cbuf_empty(&con->cb) && !call_again_soon) {
 894                 __free_page(con->rx_page);
 895                 con->rx_page = NULL;
 896         }
 897
 898         if (call_again_soon)
 899                 goto out_resched;
 900         mutex_unlock(&con->sock_mutex);
 901         return 0;
 902
 903 out_resched:
 904         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 905                 queue_work(recv_workqueue, &con->rwork);
 906         mutex_unlock(&con->sock_mutex);
 907         return -EAGAIN;
 908
 909 out_close:
 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 */
 914         }
 915         /* Don't return success if we really got EOF */
 916         if (ret == 0)
 917                 ret = -EAGAIN;
 918
 919         return ret;
 920 }
 921
 922 /* Listening socket is busy, accept a connection */
 923 static int tcp_accept_from_sock(struct connection *con)
 924 {
 925         int result;
 926         struct sockaddr_storage peeraddr;
 927         struct socket *newsock;
 928         int len;
 929         int nodeid;
 930         struct connection *newcon;
 931         struct connection *addcon;
 932
 933         mutex_lock(&connections_lock);
 934         if (!dlm_allow_conn) {
 935                 mutex_unlock(&connections_lock);
 936                 return -1;
 937         }
 938         mutex_unlock(&connections_lock);
 939
 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);
 943         if (result < 0)
 944                 return -ENOMEM;
 945
 946         mutex_lock_nested(&con->sock_mutex, 0);
 947
 948         result = -ENOTCONN;
 949         if (con->sock == NULL)
 950                 goto accept_err;
 951
 952         newsock->type = con->sock->type;
 953         newsock->ops = con->sock->ops;
 954
 955         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 956         if (result < 0)
 957                 goto accept_err;
 958
 959         /* Get the connected socket's peer */
 960         memset(&peeraddr, 0, sizeof(peeraddr));
 961         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 962                                   &len, 2)) {
 963                 result = -ECONNABORTED;
 964                 goto accept_err;
 965         }
 966
 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);
 976                 return -1;
 977         }
 978
 979         log_print("got connection from %d", nodeid);
 980
 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"
 985          */
 986         newcon = nodeid2con(nodeid, GFP_NOFS);
 987         if (!newcon) {
 988                 result = -ENOMEM;
 989                 goto accept_err;
 990         }
 991         mutex_lock_nested(&newcon->sock_mutex, 1);
 992         if (newcon->sock) {
 993                 struct connection *othercon = newcon->othercon;
 994
 995                 if (!othercon) {
 996                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 997                         if (!othercon) {
 998                                 log_print("failed to allocate incoming socket");
 999                                 mutex_unlock(&newcon->sock_mutex);
1000                                 result = -ENOMEM;
1001                                 goto accept_err;
1002                         }
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);
1009                 }
1010                 if (!othercon->sock) {
1011                         newcon->othercon = othercon;
1012                         othercon->sock = newsock;
1013                         newsock->sk->sk_user_data = othercon;
1014                         add_sock(newsock, othercon);
1015                         addcon = othercon;
1016                 }
1017                 else {
1018                         printk("Extra connection from node %d attempted\n", nodeid);
1019                         result = -EAGAIN;
1020                         mutex_unlock(&newcon->sock_mutex);
1021                         goto accept_err;
1022                 }
1023         }
1024         else {
1025                 newsock->sk->sk_user_data = newcon;
1026                 newcon->rx_action = receive_from_sock;
1027                 add_sock(newsock, newcon);
1028                 addcon = newcon;
1029         }
1030
1031         mutex_unlock(&newcon->sock_mutex);
1032
1033         /*
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
1037          */
1038         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
1039                 queue_work(recv_workqueue, &addcon->rwork);
1040         mutex_unlock(&con->sock_mutex);
1041
1042         return 0;
1043
1044 accept_err:
1045         mutex_unlock(&con->sock_mutex);
1046         sock_release(newsock);
1047
1048         if (result != -EAGAIN)
1049                 log_print("error accepting connection from node: %d", result);
1050         return result;
1051 }
1052
1053 static void free_entry(struct writequeue_entry *e)
1054 {
1055         __free_page(e->page);
1056         kfree(e);
1057 }
1058
1059 /*
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
1063  *
1064  * writequeue_lock must be held.
1065  */
1066 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1067 {
1068         e->offset += completed;
1069         e->len -= completed;
1070
1071         if (e->len == 0 && e->users == 0) {
1072                 list_del(&e->list);
1073                 free_entry(e);
1074         }
1075 }
1076
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.
1081  */
1082 static void sctp_init_assoc(struct connection *con)
1083 {
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;
1091         int len, offset;
1092         int ret;
1093         int addrlen;
1094         struct kvec iov[1];
1095
1096         mutex_lock(&con->sock_mutex);
1097         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
1098                 goto unlock;
1099
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);
1103                 goto unlock;
1104         }
1105         base_con = nodeid2con(0, 0);
1106         BUG_ON(base_con == NULL);
1107
1108         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
1109
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;
1115
1116         spin_lock(&con->writequeue_lock);
1117
1118         if (list_empty(&con->writequeue)) {
1119                 spin_unlock(&con->writequeue_lock);
1120                 log_print("writequeue empty for nodeid %d", con->nodeid);
1121                 goto unlock;
1122         }
1123
1124         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1125         len = e->len;
1126         offset = e->offset;
1127
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);
1132
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);
1136         } else {
1137                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&rem_addr;
1138                 log_print("Trying to connect to %pI6", &sin6->sin6_addr);
1139         }
1140
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;
1150
1151         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
1152         if (ret < 0) {
1153                 log_print("Send first packet to node %d failed: %d",
1154                           con->nodeid, ret);
1155
1156                 /* Try again later */
1157                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1158                 clear_bit(CF_INIT_PENDING, &con->flags);
1159         }
1160         else {
1161                 spin_lock(&con->writequeue_lock);
1162                 writequeue_entry_complete(e, ret);
1163                 spin_unlock(&con->writequeue_lock);
1164         }
1165
1166 unlock:
1167         mutex_unlock(&con->sock_mutex);
1168 }
1169
1170 /* Connect a new socket to its peer */
1171 static void tcp_connect_to_sock(struct connection *con)
1172 {
1173         struct sockaddr_storage saddr, src_addr;
1174         int addr_len;
1175         struct socket *sock = NULL;
1176         int one = 1;
1177         int result;
1178
1179         if (con->nodeid == 0) {
1180                 log_print("attempt to connect sock 0 foiled");
1181                 return;
1182         }
1183
1184         mutex_lock(&con->sock_mutex);
1185         if (con->retries++ > MAX_CONNECT_RETRIES)
1186                 goto out;
1187
1188         /* Some odd races can cause double-connects, ignore them */
1189         if (con->sock)
1190                 goto out;
1191
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);
1195         if (result < 0)
1196                 goto out_err;
1197
1198         memset(&saddr, 0, sizeof(saddr));
1199         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1200         if (result < 0) {
1201                 log_print("no address for nodeid %d", con->nodeid);
1202                 goto out_err;
1203         }
1204
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);
1209
1210         /* Bind to our cluster-known address connecting to avoid
1211            routing problems */
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,
1215                                  addr_len);
1216         if (result < 0) {
1217                 log_print("could not bind for connect: %d", result);
1218                 /* This *may* not indicate a critical error */
1219         }
1220
1221         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1222
1223         log_print("connecting to %d", con->nodeid);
1224
1225         /* Turn off Nagle's algorithm */
1226         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1227                           sizeof(one));
1228
1229         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1230                                    O_NONBLOCK);
1231         if (result == -EINPROGRESS)
1232                 result = 0;
1233         if (result == 0)
1234                 goto out;
1235
1236 out_err:
1237         if (con->sock) {
1238                 sock_release(con->sock);
1239                 con->sock = NULL;
1240         } else if (sock) {
1241                 sock_release(sock);
1242         }
1243         /*
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.
1246          */
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);
1255                 msleep(1000);
1256                 lowcomms_connect_sock(con);
1257                 return;
1258         }
1259 out:
1260         mutex_unlock(&con->sock_mutex);
1261         return;
1262 }
1263
1264 static struct socket *tcp_create_listen_sock(struct connection *con,
1265                                              struct sockaddr_storage *saddr)
1266 {
1267         struct socket *sock = NULL;
1268         int result = 0;
1269         int one = 1;
1270         int addr_len;
1271
1272         if (dlm_local_addr[0]->ss_family == AF_INET)
1273                 addr_len = sizeof(struct sockaddr_in);
1274         else
1275                 addr_len = sizeof(struct sockaddr_in6);
1276
1277         /* Create a socket to communicate with */
1278         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1279                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1280         if (result < 0) {
1281                 log_print("Can't create listening comms socket");
1282                 goto create_out;
1283         }
1284
1285         /* Turn off Nagle's algorithm */
1286         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1287                           sizeof(one));
1288
1289         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1290                                    (char *)&one, sizeof(one));
1291
1292         if (result < 0) {
1293                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1294         }
1295         con->rx_action = tcp_accept_from_sock;
1296         con->connect_action = tcp_connect_to_sock;
1297
1298         /* Bind to our port */
1299         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1300         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1301         if (result < 0) {
1302                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1303                 sock_release(sock);
1304                 sock = NULL;
1305                 con->sock = NULL;
1306                 goto create_out;
1307         }
1308         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1309                                  (char *)&one, sizeof(one));
1310         if (result < 0) {
1311                 log_print("Set keepalive failed: %d", result);
1312         }
1313
1314         result = sock->ops->listen(sock, 5);
1315         if (result < 0) {
1316                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1317                 sock_release(sock);
1318                 sock = NULL;
1319                 goto create_out;
1320         }
1321
1322 create_out:
1323         return sock;
1324 }
1325
1326 /* Get local addresses */
1327 static void init_local(void)
1328 {
1329         struct sockaddr_storage sas, *addr;
1330         int i;
1331
1332         dlm_local_count = 0;
1333         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1334                 if (dlm_our_addr(&sas, i))
1335                         break;
1336
1337                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1338                 if (!addr)
1339                         break;
1340                 memcpy(addr, &sas, sizeof(*addr));
1341                 dlm_local_addr[dlm_local_count++] = addr;
1342         }
1343 }
1344
1345 /* Bind to an IP address. SCTP allows multiple address so it can do
1346    multi-homing */
1347 static int add_sctp_bind_addr(struct connection *sctp_con,
1348                               struct sockaddr_storage *addr,
1349                               int addr_len, int num)
1350 {
1351         int result = 0;
1352
1353         if (num == 1)
1354                 result = kernel_bind(sctp_con->sock,
1355                                      (struct sockaddr *) addr,
1356                                      addr_len);
1357         else
1358                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1359                                            SCTP_SOCKOPT_BINDX_ADD,
1360                                            (char *)addr, addr_len);
1361
1362         if (result < 0)
1363                 log_print("Can't bind to port %d addr number %d",
1364                           dlm_config.ci_tcp_port, num);
1365
1366         return result;
1367 }
1368
1369 /* Initialise SCTP socket and bind to all interfaces */
1370 static int sctp_listen_for_all(void)
1371 {
1372         struct socket *sock = NULL;
1373         struct sockaddr_storage localaddr;
1374         struct sctp_event_subscribe subscribe;
1375         int result = -EINVAL, num = 1, i, addr_len;
1376         struct connection *con = nodeid2con(0, GFP_NOFS);
1377         int bufsize = NEEDED_RMEM;
1378         int one = 1;
1379
1380         if (!con)
1381                 return -ENOMEM;
1382
1383         log_print("Using SCTP for communications");
1384
1385         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1386                                   SOCK_SEQPACKET, IPPROTO_SCTP, &sock);
1387         if (result < 0) {
1388                 log_print("Can't create comms socket, check SCTP is loaded");
1389                 goto out;
1390         }
1391
1392         /* Listen for events */
1393         memset(&subscribe, 0, sizeof(subscribe));
1394         subscribe.sctp_data_io_event = 1;
1395         subscribe.sctp_association_event = 1;
1396         subscribe.sctp_send_failure_event = 1;
1397         subscribe.sctp_shutdown_event = 1;
1398         subscribe.sctp_partial_delivery_event = 1;
1399
1400         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1401                                  (char *)&bufsize, sizeof(bufsize));
1402         if (result)
1403                 log_print("Error increasing buffer space on socket %d", result);
1404
1405         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1406                                    (char *)&subscribe, sizeof(subscribe));
1407         if (result < 0) {
1408                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1409                           result);
1410                 goto create_delsock;
1411         }
1412
1413         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1414                                    sizeof(one));
1415         if (result < 0)
1416                 log_print("Could not set SCTP NODELAY error %d\n", result);
1417
1418         /* Init con struct */
1419         sock->sk->sk_user_data = con;
1420         con->sock = sock;
1421         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1422         con->rx_action = receive_from_sock;
1423         con->connect_action = sctp_init_assoc;
1424
1425         /* Bind to all interfaces. */
1426         for (i = 0; i < dlm_local_count; i++) {
1427                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1428                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1429
1430                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1431                 if (result)
1432                         goto create_delsock;
1433                 ++num;
1434         }
1435
1436         result = sock->ops->listen(sock, 5);
1437         if (result < 0) {
1438                 log_print("Can't set socket listening");
1439                 goto create_delsock;
1440         }
1441
1442         return 0;
1443
1444 create_delsock:
1445         sock_release(sock);
1446         con->sock = NULL;
1447 out:
1448         return result;
1449 }
1450
1451 static int tcp_listen_for_all(void)
1452 {
1453         struct socket *sock = NULL;
1454         struct connection *con = nodeid2con(0, GFP_NOFS);
1455         int result = -EINVAL;
1456
1457         if (!con)
1458                 return -ENOMEM;
1459
1460         /* We don't support multi-homed hosts */
1461         if (dlm_local_addr[1] != NULL) {
1462                 log_print("TCP protocol can't handle multi-homed hosts, "
1463                           "try SCTP");
1464                 return -EINVAL;
1465         }
1466
1467         log_print("Using TCP for communications");
1468
1469         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1470         if (sock) {
1471                 add_sock(sock, con);
1472                 result = 0;
1473         }
1474         else {
1475                 result = -EADDRINUSE;
1476         }
1477
1478         return result;
1479 }
1480
1481
1482
1483 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1484                                                      gfp_t allocation)
1485 {
1486         struct writequeue_entry *entry;
1487
1488         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1489         if (!entry)
1490                 return NULL;
1491
1492         entry->page = alloc_page(allocation);
1493         if (!entry->page) {
1494                 kfree(entry);
1495                 return NULL;
1496         }
1497
1498         entry->offset = 0;
1499         entry->len = 0;
1500         entry->end = 0;
1501         entry->users = 0;
1502         entry->con = con;
1503
1504         return entry;
1505 }
1506
1507 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1508 {
1509         struct connection *con;
1510         struct writequeue_entry *e;
1511         int offset = 0;
1512
1513         con = nodeid2con(nodeid, allocation);
1514         if (!con)
1515                 return NULL;
1516
1517         spin_lock(&con->writequeue_lock);
1518         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1519         if ((&e->list == &con->writequeue) ||
1520             (PAGE_CACHE_SIZE - e->end < len)) {
1521                 e = NULL;
1522         } else {
1523                 offset = e->end;
1524                 e->end += len;
1525                 e->users++;
1526         }
1527         spin_unlock(&con->writequeue_lock);
1528
1529         if (e) {
1530         got_one:
1531                 *ppc = page_address(e->page) + offset;
1532                 return e;
1533         }
1534
1535         e = new_writequeue_entry(con, allocation);
1536         if (e) {
1537                 spin_lock(&con->writequeue_lock);
1538                 offset = e->end;
1539                 e->end += len;
1540                 e->users++;
1541                 list_add_tail(&e->list, &con->writequeue);
1542                 spin_unlock(&con->writequeue_lock);
1543                 goto got_one;
1544         }
1545         return NULL;
1546 }
1547
1548 void dlm_lowcomms_commit_buffer(void *mh)
1549 {
1550         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1551         struct connection *con = e->con;
1552         int users;
1553
1554         spin_lock(&con->writequeue_lock);
1555         users = --e->users;
1556         if (users)
1557                 goto out;
1558         e->len = e->end - e->offset;
1559         spin_unlock(&con->writequeue_lock);
1560
1561         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1562                 queue_work(send_workqueue, &con->swork);
1563         }
1564         return;
1565
1566 out:
1567         spin_unlock(&con->writequeue_lock);
1568         return;
1569 }
1570
1571 /* Send a message */
1572 static void send_to_sock(struct connection *con)
1573 {
1574         int ret = 0;
1575         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1576         struct writequeue_entry *e;
1577         int len, offset;
1578         int count = 0;
1579
1580         mutex_lock(&con->sock_mutex);
1581         if (con->sock == NULL)
1582                 goto out_connect;
1583
1584         spin_lock(&con->writequeue_lock);
1585         for (;;) {
1586                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1587                                list);
1588                 if ((struct list_head *) e == &con->writequeue)
1589                         break;
1590
1591                 len = e->len;
1592                 offset = e->offset;
1593                 BUG_ON(len == 0 && e->users == 0);
1594                 spin_unlock(&con->writequeue_lock);
1595
1596                 ret = 0;
1597                 if (len) {
1598                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1599                                               msg_flags);
1600                         if (ret == -EAGAIN || ret == 0) {
1601                                 if (ret == -EAGAIN &&
1602                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1603                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1604                                         /* Notify TCP that we're limited by the
1605                                          * application window size.
1606                                          */
1607                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1608                                         con->sock->sk->sk_write_pending++;
1609                                 }
1610                                 cond_resched();
1611                                 goto out;
1612                         } else if (ret < 0)
1613                                 goto send_error;
1614                 }
1615
1616                 /* Don't starve people filling buffers */
1617                 if (++count >= MAX_SEND_MSG_COUNT) {
1618                         cond_resched();
1619                         count = 0;
1620                 }
1621
1622                 spin_lock(&con->writequeue_lock);
1623                 writequeue_entry_complete(e, ret);
1624         }
1625         spin_unlock(&con->writequeue_lock);
1626 out:
1627         mutex_unlock(&con->sock_mutex);
1628         return;
1629
1630 send_error:
1631         mutex_unlock(&con->sock_mutex);
1632         close_connection(con, false, false, true);
1633         lowcomms_connect_sock(con);
1634         return;
1635
1636 out_connect:
1637         mutex_unlock(&con->sock_mutex);
1638         if (!test_bit(CF_INIT_PENDING, &con->flags))
1639                 lowcomms_connect_sock(con);
1640 }
1641
1642 static void clean_one_writequeue(struct connection *con)
1643 {
1644         struct writequeue_entry *e, *safe;
1645
1646         spin_lock(&con->writequeue_lock);
1647         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1648                 list_del(&e->list);
1649                 free_entry(e);
1650         }
1651         spin_unlock(&con->writequeue_lock);
1652 }
1653
1654 /* Called from recovery when it knows that a node has
1655    left the cluster */
1656 int dlm_lowcomms_close(int nodeid)
1657 {
1658         struct connection *con;
1659         struct dlm_node_addr *na;
1660
1661         log_print("closing connection to node %d", nodeid);
1662         con = nodeid2con(nodeid, 0);
1663         if (con) {
1664                 set_bit(CF_CLOSE, &con->flags);
1665                 close_connection(con, true, true, true);
1666                 clean_one_writequeue(con);
1667         }
1668
1669         spin_lock(&dlm_node_addrs_spin);
1670         na = find_node_addr(nodeid);
1671         if (na) {
1672                 list_del(&na->list);
1673                 while (na->addr_count--)
1674                         kfree(na->addr[na->addr_count]);
1675                 kfree(na);
1676         }
1677         spin_unlock(&dlm_node_addrs_spin);
1678
1679         return 0;
1680 }
1681
1682 /* Receive workqueue function */
1683 static void process_recv_sockets(struct work_struct *work)
1684 {
1685         struct connection *con = container_of(work, struct connection, rwork);
1686         int err;
1687
1688         clear_bit(CF_READ_PENDING, &con->flags);
1689         do {
1690                 err = con->rx_action(con);
1691         } while (!err);
1692 }
1693
1694 /* Send workqueue function */
1695 static void process_send_sockets(struct work_struct *work)
1696 {
1697         struct connection *con = container_of(work, struct connection, swork);
1698
1699         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1700                 con->connect_action(con);
1701                 set_bit(CF_WRITE_PENDING, &con->flags);
1702         }
1703         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1704                 send_to_sock(con);
1705 }
1706
1707
1708 /* Discard all entries on the write queues */
1709 static void clean_writequeues(void)
1710 {
1711         foreach_conn(clean_one_writequeue);
1712 }
1713
1714 static void work_stop(void)
1715 {
1716         destroy_workqueue(recv_workqueue);
1717         destroy_workqueue(send_workqueue);
1718 }
1719
1720 static int work_start(void)
1721 {
1722         recv_workqueue = alloc_workqueue("dlm_recv",
1723                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1724         if (!recv_workqueue) {
1725                 log_print("can't start dlm_recv");
1726                 return -ENOMEM;
1727         }
1728
1729         send_workqueue = alloc_workqueue("dlm_send",
1730                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1731         if (!send_workqueue) {
1732                 log_print("can't start dlm_send");
1733                 destroy_workqueue(recv_workqueue);
1734                 return -ENOMEM;
1735         }
1736
1737         return 0;
1738 }
1739
1740 static void stop_conn(struct connection *con)
1741 {
1742         con->flags |= 0x0F;
1743         if (con->sock && con->sock->sk)
1744                 con->sock->sk->sk_user_data = NULL;
1745 }
1746
1747 static void free_conn(struct connection *con)
1748 {
1749         close_connection(con, true, true, true);
1750         if (con->othercon)
1751                 kmem_cache_free(con_cache, con->othercon);
1752         hlist_del(&con->list);
1753         kmem_cache_free(con_cache, con);
1754 }
1755
1756 void dlm_lowcomms_stop(void)
1757 {
1758         /* Set all the flags to prevent any
1759            socket activity.
1760         */
1761         mutex_lock(&connections_lock);
1762         dlm_allow_conn = 0;
1763         foreach_conn(stop_conn);
1764         mutex_unlock(&connections_lock);
1765
1766         work_stop();
1767
1768         mutex_lock(&connections_lock);
1769         clean_writequeues();
1770
1771         foreach_conn(free_conn);
1772
1773         mutex_unlock(&connections_lock);
1774         kmem_cache_destroy(con_cache);
1775 }
1776
1777 int dlm_lowcomms_start(void)
1778 {
1779         int error = -EINVAL;
1780         struct connection *con;
1781         int i;
1782
1783         for (i = 0; i < CONN_HASH_SIZE; i++)
1784                 INIT_HLIST_HEAD(&connection_hash[i]);
1785
1786         init_local();
1787         if (!dlm_local_count) {
1788                 error = -ENOTCONN;
1789                 log_print("no local IP address has been set");
1790                 goto fail;
1791         }
1792
1793         error = -ENOMEM;
1794         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1795                                       __alignof__(struct connection), 0,
1796                                       NULL);
1797         if (!con_cache)
1798                 goto fail;
1799
1800         error = work_start();
1801         if (error)
1802                 goto fail_destroy;
1803
1804         dlm_allow_conn = 1;
1805
1806         /* Start listening */
1807         if (dlm_config.ci_protocol == 0)
1808                 error = tcp_listen_for_all();
1809         else
1810                 error = sctp_listen_for_all();
1811         if (error)
1812                 goto fail_unlisten;
1813
1814         return 0;
1815
1816 fail_unlisten:
1817         dlm_allow_conn = 0;
1818         con = nodeid2con(0,0);
1819         if (con) {
1820                 close_connection(con, false, true, true);
1821                 kmem_cache_free(con_cache, con);
1822         }
1823 fail_destroy:
1824         kmem_cache_destroy(con_cache);
1825 fail:
1826         return error;
1827 }
1828
1829 void dlm_lowcomms_exit(void)
1830 {
1831         struct dlm_node_addr *na, *safe;
1832
1833         spin_lock(&dlm_node_addrs_spin);
1834         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1835                 list_del(&na->list);
1836                 while (na->addr_count--)
1837                         kfree(na->addr[na->addr_count]);
1838                 kfree(na);
1839         }
1840         spin_unlock(&dlm_node_addrs_spin);
1841 }