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 {
 519         mutex_lock(&con->sock_mutex);
 520
 521         if (con->sock) {
 522                 sock_release(con->sock);
 523                 con->sock = NULL;
 524         }
 525         if (con->othercon && and_other) {
 526                 /* Will only re-enter once. */
 527                 close_connection(con->othercon, false);
 528         }
 529         if (con->rx_page) {
 530                 __free_page(con->rx_page);
 531                 con->rx_page = NULL;
 532         }
 533
 534         con->retries = 0;
 535         mutex_unlock(&con->sock_mutex);
 536 }
 537
 538 /* We only send shutdown messages to nodes that are not part of the cluster
 539  * or if we get multiple connections from a node.
 540  */
 541 static void sctp_send_shutdown(sctp_assoc_t associd)
 542 {
 543         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 544         struct msghdr outmessage;
 545         struct cmsghdr *cmsg;
 546         struct sctp_sndrcvinfo *sinfo;
 547         int ret;
 548         struct connection *con;
 549
 550         con = nodeid2con(0,0);
 551         BUG_ON(con == NULL);
 552
 553         outmessage.msg_name = NULL;
 554         outmessage.msg_namelen = 0;
 555         outmessage.msg_control = outcmsg;
 556         outmessage.msg_controllen = sizeof(outcmsg);
 557         outmessage.msg_flags = MSG_EOR;
 558
 559         cmsg = CMSG_FIRSTHDR(&outmessage);
 560         cmsg->cmsg_level = IPPROTO_SCTP;
 561         cmsg->cmsg_type = SCTP_SNDRCV;
 562         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 563         outmessage.msg_controllen = cmsg->cmsg_len;
 564         sinfo = CMSG_DATA(cmsg);
 565         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 566
 567         sinfo->sinfo_flags |= MSG_EOF;
 568         sinfo->sinfo_assoc_id = associd;
 569
 570         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 571
 572         if (ret != 0)
 573                 log_print("send EOF to node failed: %d", ret);
 574 }
 575
 576 static void sctp_init_failed_foreach(struct connection *con)
 577 {
 578
 579         /*
 580          * Don't try to recover base con and handle race where the
 581          * other node's assoc init creates a assoc and we get that
 582          * notification, then we get a notification that our attempt
 583          * failed due. This happens when we are still trying the primary
 584          * address, but the other node has already tried secondary addrs
 585          * and found one that worked.
 586          */
 587         if (!con->nodeid || con->sctp_assoc)
 588                 return;
 589
 590         log_print("Retrying SCTP association init for node %d\n", con->nodeid);
 591
 592         con->try_new_addr = true;
 593         con->sctp_assoc = 0;
 594         if (test_and_clear_bit(CF_INIT_PENDING, &con->flags)) {
 595                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 596                         queue_work(send_workqueue, &con->swork);
 597         }
 598 }
 599
 600 /* INIT failed but we don't know which node...
 601    restart INIT on all pending nodes */
 602 static void sctp_init_failed(void)
 603 {
 604         mutex_lock(&connections_lock);
 605
 606         foreach_conn(sctp_init_failed_foreach);
 607
 608         mutex_unlock(&connections_lock);
 609 }
 610
 611 static void retry_failed_sctp_send(struct connection *recv_con,
 612                                    struct sctp_send_failed *sn_send_failed,
 613                                    char *buf)
 614 {
 615         int len = sn_send_failed->ssf_length - sizeof(struct sctp_send_failed);
 616         struct dlm_mhandle *mh;
 617         struct connection *con;
 618         char *retry_buf;
 619         int nodeid = sn_send_failed->ssf_info.sinfo_ppid;
 620
 621         log_print("Retry sending %d bytes to node id %d", len, nodeid);
 622
 623         if (!nodeid) {
 624                 log_print("Shouldn't resend data via listening connection.");
 625                 return;
 626         }
 627
 628         con = nodeid2con(nodeid, 0);
 629         if (!con) {
 630                 log_print("Could not look up con for nodeid %d\n",
 631                           nodeid);
 632                 return;
 633         }
 634
 635         mh = dlm_lowcomms_get_buffer(nodeid, len, GFP_NOFS, &retry_buf);
 636         if (!mh) {
 637                 log_print("Could not allocate buf for retry.");
 638                 return;
 639         }
 640         memcpy(retry_buf, buf + sizeof(struct sctp_send_failed), len);
 641         dlm_lowcomms_commit_buffer(mh);
 642
 643         /*
 644          * If we got a assoc changed event before the send failed event then
 645          * we only need to retry the send.
 646          */
 647         if (con->sctp_assoc) {
 648                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 649                         queue_work(send_workqueue, &con->swork);
 650         } else
 651                 sctp_init_failed_foreach(con);
 652 }
 653
 654 /* Something happened to an association */
 655 static void process_sctp_notification(struct connection *con,
 656                                       struct msghdr *msg, char *buf)
 657 {
 658         union sctp_notification *sn = (union sctp_notification *)buf;
 659         struct linger linger;
 660
 661         switch (sn->sn_header.sn_type) {
 662         case SCTP_SEND_FAILED:
 663                 retry_failed_sctp_send(con, &sn->sn_send_failed, buf);
 664                 break;
 665         case SCTP_ASSOC_CHANGE:
 666                 switch (sn->sn_assoc_change.sac_state) {
 667                 case SCTP_COMM_UP:
 668                 case SCTP_RESTART:
 669                 {
 670                         /* Check that the new node is in the lockspace */
 671                         struct sctp_prim prim;
 672                         int nodeid;
 673                         int prim_len, ret;
 674                         int addr_len;
 675                         struct connection *new_con;
 676
 677                         /*
 678                          * We get this before any data for an association.
 679                          * We verify that the node is in the cluster and
 680                          * then peel off a socket for it.
 681                          */
 682                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 683                                 log_print("COMM_UP for invalid assoc ID %d",
 684                                          (int)sn->sn_assoc_change.sac_assoc_id);
 685                                 sctp_init_failed();
 686                                 return;
 687                         }
 688                         memset(&prim, 0, sizeof(struct sctp_prim));
 689                         prim_len = sizeof(struct sctp_prim);
 690                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 691
 692                         ret = kernel_getsockopt(con->sock,
 693                                                 IPPROTO_SCTP,
 694                                                 SCTP_PRIMARY_ADDR,
 695                                                 (char*)&prim,
 696                                                 &prim_len);
 697                         if (ret < 0) {
 698                                 log_print("getsockopt/sctp_primary_addr on "
 699                                           "new assoc %d failed : %d",
 700                                           (int)sn->sn_assoc_change.sac_assoc_id,
 701                                           ret);
 702
 703                                 /* Retry INIT later */
 704                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 705                                 if (new_con)
 706                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 707                                 return;
 708                         }
 709                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 710                         if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 711                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 712                                 log_print("reject connect from unknown addr");
 713                                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 714                                                      b, sizeof(struct sockaddr_storage));
 715                                 sctp_send_shutdown(prim.ssp_assoc_id);
 716                                 return;
 717                         }
 718
 719                         new_con = nodeid2con(nodeid, GFP_NOFS);
 720                         if (!new_con)
 721                                 return;
 722
 723                         if (new_con->sock) {
 724                                 log_print("reject connect from node %d: "
 725                                           "already has a connection.",
 726                                           nodeid);
 727                                 sctp_send_shutdown(prim.ssp_assoc_id);
 728                                 return;
 729                         }
 730
 731                         /* Peel off a new sock */
 732                         lock_sock(con->sock->sk);
 733                         ret = sctp_do_peeloff(con->sock->sk,
 734                                 sn->sn_assoc_change.sac_assoc_id,
 735                                 &new_con->sock);
 736                         release_sock(con->sock->sk);
 737                         if (ret < 0) {
 738                                 log_print("Can't peel off a socket for "
 739                                           "connection %d to node %d: err=%d",
 740                                           (int)sn->sn_assoc_change.sac_assoc_id,
 741                                           nodeid, ret);
 742                                 return;
 743                         }
 744                         add_sock(new_con->sock, new_con);
 745
 746                         linger.l_onoff = 1;
 747                         linger.l_linger = 0;
 748                         ret = kernel_setsockopt(new_con->sock, SOL_SOCKET, SO_LINGER,
 749                                                 (char *)&linger, sizeof(linger));
 750                         if (ret < 0)
 751                                 log_print("set socket option SO_LINGER failed");
 752
 753                         log_print("connecting to %d sctp association %d",
 754                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 755
 756                         new_con->sctp_assoc = sn->sn_assoc_change.sac_assoc_id;
 757                         new_con->try_new_addr = false;
 758                         /* Send any pending writes */
 759                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 760                         clear_bit(CF_INIT_PENDING, &new_con->flags);
 761                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 762                                 queue_work(send_workqueue, &new_con->swork);
 763                         }
 764                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 765                                 queue_work(recv_workqueue, &new_con->rwork);
 766                 }
 767                 break;
 768
 769                 case SCTP_COMM_LOST:
 770                 case SCTP_SHUTDOWN_COMP:
 771                 {
 772                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 773                         if (con) {
 774                                 con->sctp_assoc = 0;
 775                         }
 776                 }
 777                 break;
 778
 779                 case SCTP_CANT_STR_ASSOC:
 780                 {
 781                         /* Will retry init when we get the send failed notification */
 782                         log_print("Can't start SCTP association - retrying");
 783                 }
 784                 break;
 785
 786                 default:
 787                         log_print("unexpected SCTP assoc change id=%d state=%d",
 788                                   (int)sn->sn_assoc_change.sac_assoc_id,
 789                                   sn->sn_assoc_change.sac_state);
 790                 }
 791         default:
 792                 ; /* fall through */
 793         }
 794 }
 795
 796 /* Data received from remote end */
 797 static int receive_from_sock(struct connection *con)
 798 {
 799         int ret = 0;
 800         struct msghdr msg = {};
 801         struct kvec iov[2];
 802         unsigned len;
 803         int r;
 804         int call_again_soon = 0;
 805         int nvec;
 806         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 807
 808         mutex_lock(&con->sock_mutex);
 809
 810         if (con->sock == NULL) {
 811                 ret = -EAGAIN;
 812                 goto out_close;
 813         }
 814
 815         if (con->rx_page == NULL) {
 816                 /*
 817                  * This doesn't need to be atomic, but I think it should
 818                  * improve performance if it is.
 819                  */
 820                 con->rx_page = alloc_page(GFP_ATOMIC);
 821                 if (con->rx_page == NULL)
 822                         goto out_resched;
 823                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 824         }
 825
 826         /* Only SCTP needs these really */
 827         memset(&incmsg, 0, sizeof(incmsg));
 828         msg.msg_control = incmsg;
 829         msg.msg_controllen = sizeof(incmsg);
 830
 831         /*
 832          * iov[0] is the bit of the circular buffer between the current end
 833          * point (cb.base + cb.len) and the end of the buffer.
 834          */
 835         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 836         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 837         iov[1].iov_len = 0;
 838         nvec = 1;
 839
 840         /*
 841          * iov[1] is the bit of the circular buffer between the start of the
 842          * buffer and the start of the currently used section (cb.base)
 843          */
 844         if (cbuf_data(&con->cb) >= con->cb.base) {
 845                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 846                 iov[1].iov_len = con->cb.base;
 847                 iov[1].iov_base = page_address(con->rx_page);
 848                 nvec = 2;
 849         }
 850         len = iov[0].iov_len + iov[1].iov_len;
 851
 852         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 853                                MSG_DONTWAIT | MSG_NOSIGNAL);
 854         if (ret <= 0)
 855                 goto out_close;
 856
 857         /* Process SCTP notifications */
 858         if (msg.msg_flags & MSG_NOTIFICATION) {
 859                 msg.msg_control = incmsg;
 860                 msg.msg_controllen = sizeof(incmsg);
 861
 862                 process_sctp_notification(con, &msg,
 863                                 page_address(con->rx_page) + con->cb.base);
 864                 mutex_unlock(&con->sock_mutex);
 865                 return 0;
 866         }
 867         BUG_ON(con->nodeid == 0);
 868
 869         if (ret == len)
 870                 call_again_soon = 1;
 871         cbuf_add(&con->cb, ret);
 872         ret = dlm_process_incoming_buffer(con->nodeid,
 873                                           page_address(con->rx_page),
 874                                           con->cb.base, con->cb.len,
 875                                           PAGE_CACHE_SIZE);
 876         if (ret == -EBADMSG) {
 877                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 878                           "iov_len=%u, iov_base[0]=%p, read=%d",
 879                           page_address(con->rx_page), con->cb.base, con->cb.len,
 880                           len, iov[0].iov_base, r);
 881         }
 882         if (ret < 0)
 883                 goto out_close;
 884         cbuf_eat(&con->cb, ret);
 885
 886         if (cbuf_empty(&con->cb) && !call_again_soon) {
 887                 __free_page(con->rx_page);
 888                 con->rx_page = NULL;
 889         }
 890
 891         if (call_again_soon)
 892                 goto out_resched;
 893         mutex_unlock(&con->sock_mutex);
 894         return 0;
 895
 896 out_resched:
 897         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 898                 queue_work(recv_workqueue, &con->rwork);
 899         mutex_unlock(&con->sock_mutex);
 900         return -EAGAIN;
 901
 902 out_close:
 903         mutex_unlock(&con->sock_mutex);
 904         if (ret != -EAGAIN) {
 905                 close_connection(con, false);
 906                 /* Reconnect when there is something to send */
 907         }
 908         /* Don't return success if we really got EOF */
 909         if (ret == 0)
 910                 ret = -EAGAIN;
 911
 912         return ret;
 913 }
 914
 915 /* Listening socket is busy, accept a connection */
 916 static int tcp_accept_from_sock(struct connection *con)
 917 {
 918         int result;
 919         struct sockaddr_storage peeraddr;
 920         struct socket *newsock;
 921         int len;
 922         int nodeid;
 923         struct connection *newcon;
 924         struct connection *addcon;
 925
 926         mutex_lock(&connections_lock);
 927         if (!dlm_allow_conn) {
 928                 mutex_unlock(&connections_lock);
 929                 return -1;
 930         }
 931         mutex_unlock(&connections_lock);
 932
 933         memset(&peeraddr, 0, sizeof(peeraddr));
 934         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
 935                                   SOCK_STREAM, IPPROTO_TCP, &newsock);
 936         if (result < 0)
 937                 return -ENOMEM;
 938
 939         mutex_lock_nested(&con->sock_mutex, 0);
 940
 941         result = -ENOTCONN;
 942         if (con->sock == NULL)
 943                 goto accept_err;
 944
 945         newsock->type = con->sock->type;
 946         newsock->ops = con->sock->ops;
 947
 948         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 949         if (result < 0)
 950                 goto accept_err;
 951
 952         /* Get the connected socket's peer */
 953         memset(&peeraddr, 0, sizeof(peeraddr));
 954         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 955                                   &len, 2)) {
 956                 result = -ECONNABORTED;
 957                 goto accept_err;
 958         }
 959
 960         /* Get the new node's NODEID */
 961         make_sockaddr(&peeraddr, 0, &len);
 962         if (addr_to_nodeid(&peeraddr, &nodeid)) {
 963                 unsigned char *b=(unsigned char *)&peeraddr;
 964                 log_print("connect from non cluster node");
 965                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 966                                      b, sizeof(struct sockaddr_storage));
 967                 sock_release(newsock);
 968                 mutex_unlock(&con->sock_mutex);
 969                 return -1;
 970         }
 971
 972         log_print("got connection from %d", nodeid);
 973
 974         /*  Check to see if we already have a connection to this node. This
 975          *  could happen if the two nodes initiate a connection at roughly
 976          *  the same time and the connections cross on the wire.
 977          *  In this case we store the incoming one in "othercon"
 978          */
 979         newcon = nodeid2con(nodeid, GFP_NOFS);
 980         if (!newcon) {
 981                 result = -ENOMEM;
 982                 goto accept_err;
 983         }
 984         mutex_lock_nested(&newcon->sock_mutex, 1);
 985         if (newcon->sock) {
 986                 struct connection *othercon = newcon->othercon;
 987
 988                 if (!othercon) {
 989                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 990                         if (!othercon) {
 991                                 log_print("failed to allocate incoming socket");
 992                                 mutex_unlock(&newcon->sock_mutex);
 993                                 result = -ENOMEM;
 994                                 goto accept_err;
 995                         }
 996                         othercon->nodeid = nodeid;
 997                         othercon->rx_action = receive_from_sock;
 998                         mutex_init(&othercon->sock_mutex);
 999                         INIT_WORK(&othercon->swork, process_send_sockets);
1000                         INIT_WORK(&othercon->rwork, process_recv_sockets);
1001                         set_bit(CF_IS_OTHERCON, &othercon->flags);
1002                 }
1003                 if (!othercon->sock) {
1004                         newcon->othercon = othercon;
1005                         othercon->sock = newsock;
1006                         newsock->sk->sk_user_data = othercon;
1007                         add_sock(newsock, othercon);
1008                         addcon = othercon;
1009                 }
1010                 else {
1011                         printk("Extra connection from node %d attempted\n", nodeid);
1012                         result = -EAGAIN;
1013                         mutex_unlock(&newcon->sock_mutex);
1014                         goto accept_err;
1015                 }
1016         }
1017         else {
1018                 newsock->sk->sk_user_data = newcon;
1019                 newcon->rx_action = receive_from_sock;
1020                 add_sock(newsock, newcon);
1021                 addcon = newcon;
1022         }
1023
1024         mutex_unlock(&newcon->sock_mutex);
1025
1026         /*
1027          * Add it to the active queue in case we got data
1028          * between processing the accept adding the socket
1029          * to the read_sockets list
1030          */
1031         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
1032                 queue_work(recv_workqueue, &addcon->rwork);
1033         mutex_unlock(&con->sock_mutex);
1034
1035         return 0;
1036
1037 accept_err:
1038         mutex_unlock(&con->sock_mutex);
1039         sock_release(newsock);
1040
1041         if (result != -EAGAIN)
1042                 log_print("error accepting connection from node: %d", result);
1043         return result;
1044 }
1045
1046 static void free_entry(struct writequeue_entry *e)
1047 {
1048         __free_page(e->page);
1049         kfree(e);
1050 }
1051
1052 /*
1053  * writequeue_entry_complete - try to delete and free write queue entry
1054  * @e: write queue entry to try to delete
1055  * @completed: bytes completed
1056  *
1057  * writequeue_lock must be held.
1058  */
1059 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1060 {
1061         e->offset += completed;
1062         e->len -= completed;
1063
1064         if (e->len == 0 && e->users == 0) {
1065                 list_del(&e->list);
1066                 free_entry(e);
1067         }
1068 }
1069
1070 /* Initiate an SCTP association.
1071    This is a special case of send_to_sock() in that we don't yet have a
1072    peeled-off socket for this association, so we use the listening socket
1073    and add the primary IP address of the remote node.
1074  */
1075 static void sctp_init_assoc(struct connection *con)
1076 {
1077         struct sockaddr_storage rem_addr;
1078         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
1079         struct msghdr outmessage;
1080         struct cmsghdr *cmsg;
1081         struct sctp_sndrcvinfo *sinfo;
1082         struct connection *base_con;
1083         struct writequeue_entry *e;
1084         int len, offset;
1085         int ret;
1086         int addrlen;
1087         struct kvec iov[1];
1088
1089         mutex_lock(&con->sock_mutex);
1090         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
1091                 goto unlock;
1092
1093         if (nodeid_to_addr(con->nodeid, NULL, (struct sockaddr *)&rem_addr,
1094                            con->try_new_addr)) {
1095                 log_print("no address for nodeid %d", con->nodeid);
1096                 goto unlock;
1097         }
1098         base_con = nodeid2con(0, 0);
1099         BUG_ON(base_con == NULL);
1100
1101         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
1102
1103         outmessage.msg_name = &rem_addr;
1104         outmessage.msg_namelen = addrlen;
1105         outmessage.msg_control = outcmsg;
1106         outmessage.msg_controllen = sizeof(outcmsg);
1107         outmessage.msg_flags = MSG_EOR;
1108
1109         spin_lock(&con->writequeue_lock);
1110
1111         if (list_empty(&con->writequeue)) {
1112                 spin_unlock(&con->writequeue_lock);
1113                 log_print("writequeue empty for nodeid %d", con->nodeid);
1114                 goto unlock;
1115         }
1116
1117         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1118         len = e->len;
1119         offset = e->offset;
1120
1121         /* Send the first block off the write queue */
1122         iov[0].iov_base = page_address(e->page)+offset;
1123         iov[0].iov_len = len;
1124         spin_unlock(&con->writequeue_lock);
1125
1126         if (rem_addr.ss_family == AF_INET) {
1127                 struct sockaddr_in *sin = (struct sockaddr_in *)&rem_addr;
1128                 log_print("Trying to connect to %pI4", &sin->sin_addr.s_addr);
1129         } else {
1130                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&rem_addr;
1131                 log_print("Trying to connect to %pI6", &sin6->sin6_addr);
1132         }
1133
1134         cmsg = CMSG_FIRSTHDR(&outmessage);
1135         cmsg->cmsg_level = IPPROTO_SCTP;
1136         cmsg->cmsg_type = SCTP_SNDRCV;
1137         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
1138         sinfo = CMSG_DATA(cmsg);
1139         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
1140         sinfo->sinfo_ppid = cpu_to_le32(con->nodeid);
1141         outmessage.msg_controllen = cmsg->cmsg_len;
1142         sinfo->sinfo_flags |= SCTP_ADDR_OVER;
1143
1144         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
1145         if (ret < 0) {
1146                 log_print("Send first packet to node %d failed: %d",
1147                           con->nodeid, ret);
1148
1149                 /* Try again later */
1150                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1151                 clear_bit(CF_INIT_PENDING, &con->flags);
1152         }
1153         else {
1154                 spin_lock(&con->writequeue_lock);
1155                 writequeue_entry_complete(e, ret);
1156                 spin_unlock(&con->writequeue_lock);
1157         }
1158
1159 unlock:
1160         mutex_unlock(&con->sock_mutex);
1161 }
1162
1163 /* Connect a new socket to its peer */
1164 static void tcp_connect_to_sock(struct connection *con)
1165 {
1166         struct sockaddr_storage saddr, src_addr;
1167         int addr_len;
1168         struct socket *sock = NULL;
1169         int one = 1;
1170         int result;
1171
1172         if (con->nodeid == 0) {
1173                 log_print("attempt to connect sock 0 foiled");
1174                 return;
1175         }
1176
1177         mutex_lock(&con->sock_mutex);
1178         if (con->retries++ > MAX_CONNECT_RETRIES)
1179                 goto out;
1180
1181         /* Some odd races can cause double-connects, ignore them */
1182         if (con->sock)
1183                 goto out;
1184
1185         /* Create a socket to communicate with */
1186         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1187                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1188         if (result < 0)
1189                 goto out_err;
1190
1191         memset(&saddr, 0, sizeof(saddr));
1192         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1193         if (result < 0) {
1194                 log_print("no address for nodeid %d", con->nodeid);
1195                 goto out_err;
1196         }
1197
1198         sock->sk->sk_user_data = con;
1199         con->rx_action = receive_from_sock;
1200         con->connect_action = tcp_connect_to_sock;
1201         add_sock(sock, con);
1202
1203         /* Bind to our cluster-known address connecting to avoid
1204            routing problems */
1205         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1206         make_sockaddr(&src_addr, 0, &addr_len);
1207         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1208                                  addr_len);
1209         if (result < 0) {
1210                 log_print("could not bind for connect: %d", result);
1211                 /* This *may* not indicate a critical error */
1212         }
1213
1214         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1215
1216         log_print("connecting to %d", con->nodeid);
1217
1218         /* Turn off Nagle's algorithm */
1219         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1220                           sizeof(one));
1221
1222         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1223                                    O_NONBLOCK);
1224         if (result == -EINPROGRESS)
1225                 result = 0;
1226         if (result == 0)
1227                 goto out;
1228
1229 out_err:
1230         if (con->sock) {
1231                 sock_release(con->sock);
1232                 con->sock = NULL;
1233         } else if (sock) {
1234                 sock_release(sock);
1235         }
1236         /*
1237          * Some errors are fatal and this list might need adjusting. For other
1238          * errors we try again until the max number of retries is reached.
1239          */
1240         if (result != -EHOSTUNREACH &&
1241             result != -ENETUNREACH &&
1242             result != -ENETDOWN &&
1243             result != -EINVAL &&
1244             result != -EPROTONOSUPPORT) {
1245                 log_print("connect %d try %d error %d", con->nodeid,
1246                           con->retries, result);
1247                 mutex_unlock(&con->sock_mutex);
1248                 msleep(1000);
1249                 lowcomms_connect_sock(con);
1250                 return;
1251         }
1252 out:
1253         mutex_unlock(&con->sock_mutex);
1254         return;
1255 }
1256
1257 static struct socket *tcp_create_listen_sock(struct connection *con,
1258                                              struct sockaddr_storage *saddr)
1259 {
1260         struct socket *sock = NULL;
1261         int result = 0;
1262         int one = 1;
1263         int addr_len;
1264
1265         if (dlm_local_addr[0]->ss_family == AF_INET)
1266                 addr_len = sizeof(struct sockaddr_in);
1267         else
1268                 addr_len = sizeof(struct sockaddr_in6);
1269
1270         /* Create a socket to communicate with */
1271         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1272                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1273         if (result < 0) {
1274                 log_print("Can't create listening comms socket");
1275                 goto create_out;
1276         }
1277
1278         /* Turn off Nagle's algorithm */
1279         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1280                           sizeof(one));
1281
1282         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1283                                    (char *)&one, sizeof(one));
1284
1285         if (result < 0) {
1286                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1287         }
1288         con->rx_action = tcp_accept_from_sock;
1289         con->connect_action = tcp_connect_to_sock;
1290
1291         /* Bind to our port */
1292         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1293         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1294         if (result < 0) {
1295                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1296                 sock_release(sock);
1297                 sock = NULL;
1298                 con->sock = NULL;
1299                 goto create_out;
1300         }
1301         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1302                                  (char *)&one, sizeof(one));
1303         if (result < 0) {
1304                 log_print("Set keepalive failed: %d", result);
1305         }
1306
1307         result = sock->ops->listen(sock, 5);
1308         if (result < 0) {
1309                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1310                 sock_release(sock);
1311                 sock = NULL;
1312                 goto create_out;
1313         }
1314
1315 create_out:
1316         return sock;
1317 }
1318
1319 /* Get local addresses */
1320 static void init_local(void)
1321 {
1322         struct sockaddr_storage sas, *addr;
1323         int i;
1324
1325         dlm_local_count = 0;
1326         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1327                 if (dlm_our_addr(&sas, i))
1328                         break;
1329
1330                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1331                 if (!addr)
1332                         break;
1333                 memcpy(addr, &sas, sizeof(*addr));
1334                 dlm_local_addr[dlm_local_count++] = addr;
1335         }
1336 }
1337
1338 /* Bind to an IP address. SCTP allows multiple address so it can do
1339    multi-homing */
1340 static int add_sctp_bind_addr(struct connection *sctp_con,
1341                               struct sockaddr_storage *addr,
1342                               int addr_len, int num)
1343 {
1344         int result = 0;
1345
1346         if (num == 1)
1347                 result = kernel_bind(sctp_con->sock,
1348                                      (struct sockaddr *) addr,
1349                                      addr_len);
1350         else
1351                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1352                                            SCTP_SOCKOPT_BINDX_ADD,
1353                                            (char *)addr, addr_len);
1354
1355         if (result < 0)
1356                 log_print("Can't bind to port %d addr number %d",
1357                           dlm_config.ci_tcp_port, num);
1358
1359         return result;
1360 }
1361
1362 /* Initialise SCTP socket and bind to all interfaces */
1363 static int sctp_listen_for_all(void)
1364 {
1365         struct socket *sock = NULL;
1366         struct sockaddr_storage localaddr;
1367         struct sctp_event_subscribe subscribe;
1368         int result = -EINVAL, num = 1, i, addr_len;
1369         struct connection *con = nodeid2con(0, GFP_NOFS);
1370         int bufsize = NEEDED_RMEM;
1371         int one = 1;
1372
1373         if (!con)
1374                 return -ENOMEM;
1375
1376         log_print("Using SCTP for communications");
1377
1378         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1379                                   SOCK_SEQPACKET, IPPROTO_SCTP, &sock);
1380         if (result < 0) {
1381                 log_print("Can't create comms socket, check SCTP is loaded");
1382                 goto out;
1383         }
1384
1385         /* Listen for events */
1386         memset(&subscribe, 0, sizeof(subscribe));
1387         subscribe.sctp_data_io_event = 1;
1388         subscribe.sctp_association_event = 1;
1389         subscribe.sctp_send_failure_event = 1;
1390         subscribe.sctp_shutdown_event = 1;
1391         subscribe.sctp_partial_delivery_event = 1;
1392
1393         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1394                                  (char *)&bufsize, sizeof(bufsize));
1395         if (result)
1396                 log_print("Error increasing buffer space on socket %d", result);
1397
1398         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1399                                    (char *)&subscribe, sizeof(subscribe));
1400         if (result < 0) {
1401                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1402                           result);
1403                 goto create_delsock;
1404         }
1405
1406         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1407                                    sizeof(one));
1408         if (result < 0)
1409                 log_print("Could not set SCTP NODELAY error %d\n", result);
1410
1411         /* Init con struct */
1412         sock->sk->sk_user_data = con;
1413         con->sock = sock;
1414         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1415         con->rx_action = receive_from_sock;
1416         con->connect_action = sctp_init_assoc;
1417
1418         /* Bind to all interfaces. */
1419         for (i = 0; i < dlm_local_count; i++) {
1420                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1421                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1422
1423                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1424                 if (result)
1425                         goto create_delsock;
1426                 ++num;
1427         }
1428
1429         result = sock->ops->listen(sock, 5);
1430         if (result < 0) {
1431                 log_print("Can't set socket listening");
1432                 goto create_delsock;
1433         }
1434
1435         return 0;
1436
1437 create_delsock:
1438         sock_release(sock);
1439         con->sock = NULL;
1440 out:
1441         return result;
1442 }
1443
1444 static int tcp_listen_for_all(void)
1445 {
1446         struct socket *sock = NULL;
1447         struct connection *con = nodeid2con(0, GFP_NOFS);
1448         int result = -EINVAL;
1449
1450         if (!con)
1451                 return -ENOMEM;
1452
1453         /* We don't support multi-homed hosts */
1454         if (dlm_local_addr[1] != NULL) {
1455                 log_print("TCP protocol can't handle multi-homed hosts, "
1456                           "try SCTP");
1457                 return -EINVAL;
1458         }
1459
1460         log_print("Using TCP for communications");
1461
1462         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1463         if (sock) {
1464                 add_sock(sock, con);
1465                 result = 0;
1466         }
1467         else {
1468                 result = -EADDRINUSE;
1469         }
1470
1471         return result;
1472 }
1473
1474
1475
1476 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1477                                                      gfp_t allocation)
1478 {
1479         struct writequeue_entry *entry;
1480
1481         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1482         if (!entry)
1483                 return NULL;
1484
1485         entry->page = alloc_page(allocation);
1486         if (!entry->page) {
1487                 kfree(entry);
1488                 return NULL;
1489         }
1490
1491         entry->offset = 0;
1492         entry->len = 0;
1493         entry->end = 0;
1494         entry->users = 0;
1495         entry->con = con;
1496
1497         return entry;
1498 }
1499
1500 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1501 {
1502         struct connection *con;
1503         struct writequeue_entry *e;
1504         int offset = 0;
1505
1506         con = nodeid2con(nodeid, allocation);
1507         if (!con)
1508                 return NULL;
1509
1510         spin_lock(&con->writequeue_lock);
1511         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1512         if ((&e->list == &con->writequeue) ||
1513             (PAGE_CACHE_SIZE - e->end < len)) {
1514                 e = NULL;
1515         } else {
1516                 offset = e->end;
1517                 e->end += len;
1518                 e->users++;
1519         }
1520         spin_unlock(&con->writequeue_lock);
1521
1522         if (e) {
1523         got_one:
1524                 *ppc = page_address(e->page) + offset;
1525                 return e;
1526         }
1527
1528         e = new_writequeue_entry(con, allocation);
1529         if (e) {
1530                 spin_lock(&con->writequeue_lock);
1531                 offset = e->end;
1532                 e->end += len;
1533                 e->users++;
1534                 list_add_tail(&e->list, &con->writequeue);
1535                 spin_unlock(&con->writequeue_lock);
1536                 goto got_one;
1537         }
1538         return NULL;
1539 }
1540
1541 void dlm_lowcomms_commit_buffer(void *mh)
1542 {
1543         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1544         struct connection *con = e->con;
1545         int users;
1546
1547         spin_lock(&con->writequeue_lock);
1548         users = --e->users;
1549         if (users)
1550                 goto out;
1551         e->len = e->end - e->offset;
1552         spin_unlock(&con->writequeue_lock);
1553
1554         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1555                 queue_work(send_workqueue, &con->swork);
1556         }
1557         return;
1558
1559 out:
1560         spin_unlock(&con->writequeue_lock);
1561         return;
1562 }
1563
1564 /* Send a message */
1565 static void send_to_sock(struct connection *con)
1566 {
1567         int ret = 0;
1568         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1569         struct writequeue_entry *e;
1570         int len, offset;
1571         int count = 0;
1572
1573         mutex_lock(&con->sock_mutex);
1574         if (con->sock == NULL)
1575                 goto out_connect;
1576
1577         spin_lock(&con->writequeue_lock);
1578         for (;;) {
1579                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1580                                list);
1581                 if ((struct list_head *) e == &con->writequeue)
1582                         break;
1583
1584                 len = e->len;
1585                 offset = e->offset;
1586                 BUG_ON(len == 0 && e->users == 0);
1587                 spin_unlock(&con->writequeue_lock);
1588
1589                 ret = 0;
1590                 if (len) {
1591                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1592                                               msg_flags);
1593                         if (ret == -EAGAIN || ret == 0) {
1594                                 if (ret == -EAGAIN &&
1595                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1596                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1597                                         /* Notify TCP that we're limited by the
1598                                          * application window size.
1599                                          */
1600                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1601                                         con->sock->sk->sk_write_pending++;
1602                                 }
1603                                 cond_resched();
1604                                 goto out;
1605                         } else if (ret < 0)
1606                                 goto send_error;
1607                 }
1608
1609                 /* Don't starve people filling buffers */
1610                 if (++count >= MAX_SEND_MSG_COUNT) {
1611                         cond_resched();
1612                         count = 0;
1613                 }
1614
1615                 spin_lock(&con->writequeue_lock);
1616                 writequeue_entry_complete(e, ret);
1617         }
1618         spin_unlock(&con->writequeue_lock);
1619 out:
1620         mutex_unlock(&con->sock_mutex);
1621         return;
1622
1623 send_error:
1624         mutex_unlock(&con->sock_mutex);
1625         close_connection(con, false);
1626         lowcomms_connect_sock(con);
1627         return;
1628
1629 out_connect:
1630         mutex_unlock(&con->sock_mutex);
1631         if (!test_bit(CF_INIT_PENDING, &con->flags))
1632                 lowcomms_connect_sock(con);
1633 }
1634
1635 static void clean_one_writequeue(struct connection *con)
1636 {
1637         struct writequeue_entry *e, *safe;
1638
1639         spin_lock(&con->writequeue_lock);
1640         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1641                 list_del(&e->list);
1642                 free_entry(e);
1643         }
1644         spin_unlock(&con->writequeue_lock);
1645 }
1646
1647 /* Called from recovery when it knows that a node has
1648    left the cluster */
1649 int dlm_lowcomms_close(int nodeid)
1650 {
1651         struct connection *con;
1652         struct dlm_node_addr *na;
1653
1654         log_print("closing connection to node %d", nodeid);
1655         con = nodeid2con(nodeid, 0);
1656         if (con) {
1657                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1658                 clear_bit(CF_WRITE_PENDING, &con->flags);
1659                 set_bit(CF_CLOSE, &con->flags);
1660                 if (cancel_work_sync(&con->swork))
1661                         log_print("canceled swork for node %d", nodeid);
1662                 if (cancel_work_sync(&con->rwork))
1663                         log_print("canceled rwork for node %d", nodeid);
1664                 clean_one_writequeue(con);
1665                 close_connection(con, true);
1666         }
1667
1668         spin_lock(&dlm_node_addrs_spin);
1669         na = find_node_addr(nodeid);
1670         if (na) {
1671                 list_del(&na->list);
1672                 while (na->addr_count--)
1673                         kfree(na->addr[na->addr_count]);
1674                 kfree(na);
1675         }
1676         spin_unlock(&dlm_node_addrs_spin);
1677
1678         return 0;
1679 }
1680
1681 /* Receive workqueue function */
1682 static void process_recv_sockets(struct work_struct *work)
1683 {
1684         struct connection *con = container_of(work, struct connection, rwork);
1685         int err;
1686
1687         clear_bit(CF_READ_PENDING, &con->flags);
1688         do {
1689                 err = con->rx_action(con);
1690         } while (!err);
1691 }
1692
1693 /* Send workqueue function */
1694 static void process_send_sockets(struct work_struct *work)
1695 {
1696         struct connection *con = container_of(work, struct connection, swork);
1697
1698         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1699                 con->connect_action(con);
1700                 set_bit(CF_WRITE_PENDING, &con->flags);
1701         }
1702         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1703                 send_to_sock(con);
1704 }
1705
1706
1707 /* Discard all entries on the write queues */
1708 static void clean_writequeues(void)
1709 {
1710         foreach_conn(clean_one_writequeue);
1711 }
1712
1713 static void work_stop(void)
1714 {
1715         destroy_workqueue(recv_workqueue);
1716         destroy_workqueue(send_workqueue);
1717 }
1718
1719 static int work_start(void)
1720 {
1721         recv_workqueue = alloc_workqueue("dlm_recv",
1722                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1723         if (!recv_workqueue) {
1724                 log_print("can't start dlm_recv");
1725                 return -ENOMEM;
1726         }
1727
1728         send_workqueue = alloc_workqueue("dlm_send",
1729                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1730         if (!send_workqueue) {
1731                 log_print("can't start dlm_send");
1732                 destroy_workqueue(recv_workqueue);
1733                 return -ENOMEM;
1734         }
1735
1736         return 0;
1737 }
1738
1739 static void stop_conn(struct connection *con)
1740 {
1741         con->flags |= 0x0F;
1742         if (con->sock && con->sock->sk)
1743                 con->sock->sk->sk_user_data = NULL;
1744 }
1745
1746 static void free_conn(struct connection *con)
1747 {
1748         close_connection(con, true);
1749         if (con->othercon)
1750                 kmem_cache_free(con_cache, con->othercon);
1751         hlist_del(&con->list);
1752         kmem_cache_free(con_cache, con);
1753 }
1754
1755 void dlm_lowcomms_stop(void)
1756 {
1757         /* Set all the flags to prevent any
1758            socket activity.
1759         */
1760         mutex_lock(&connections_lock);
1761         dlm_allow_conn = 0;
1762         foreach_conn(stop_conn);
1763         mutex_unlock(&connections_lock);
1764
1765         work_stop();
1766
1767         mutex_lock(&connections_lock);
1768         clean_writequeues();
1769
1770         foreach_conn(free_conn);
1771
1772         mutex_unlock(&connections_lock);
1773         kmem_cache_destroy(con_cache);
1774 }
1775
1776 int dlm_lowcomms_start(void)
1777 {
1778         int error = -EINVAL;
1779         struct connection *con;
1780         int i;
1781
1782         for (i = 0; i < CONN_HASH_SIZE; i++)
1783                 INIT_HLIST_HEAD(&connection_hash[i]);
1784
1785         init_local();
1786         if (!dlm_local_count) {
1787                 error = -ENOTCONN;
1788                 log_print("no local IP address has been set");
1789                 goto fail;
1790         }
1791
1792         error = -ENOMEM;
1793         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1794                                       __alignof__(struct connection), 0,
1795                                       NULL);
1796         if (!con_cache)
1797                 goto fail;
1798
1799         error = work_start();
1800         if (error)
1801                 goto fail_destroy;
1802
1803         dlm_allow_conn = 1;
1804
1805         /* Start listening */
1806         if (dlm_config.ci_protocol == 0)
1807                 error = tcp_listen_for_all();
1808         else
1809                 error = sctp_listen_for_all();
1810         if (error)
1811                 goto fail_unlisten;
1812
1813         return 0;
1814
1815 fail_unlisten:
1816         dlm_allow_conn = 0;
1817         con = nodeid2con(0,0);
1818         if (con) {
1819                 close_connection(con, false);
1820                 kmem_cache_free(con_cache, con);
1821         }
1822 fail_destroy:
1823         kmem_cache_destroy(con_cache);
1824 fail:
1825         return error;
1826 }
1827
1828 void dlm_lowcomms_exit(void)
1829 {
1830         struct dlm_node_addr *na, *safe;
1831
1832         spin_lock(&dlm_node_addrs_spin);
1833         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1834                 list_del(&na->list);
1835                 while (na->addr_count--)
1836                         kfree(na->addr[na->addr_count]);
1837                 kfree(na);
1838         }
1839         spin_unlock(&dlm_node_addrs_spin);
1840 }