Pileus Git - ~andy/linux/blob - 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/sctp/user.h>
  57 #include <net/ipv6.h>
  58
  59 #include "dlm_internal.h"
  60 #include "lowcomms.h"
  61 #include "midcomms.h"
  62 #include "config.h"
  63
  64 #define NEEDED_RMEM (4*1024*1024)
  65 #define CONN_HASH_SIZE 32
  66
  67 /* Number of messages to send before rescheduling */
  68 #define MAX_SEND_MSG_COUNT 25
  69
  70 struct cbuf {
  71         unsigned int base;
  72         unsigned int len;
  73         unsigned int mask;
  74 };
  75
  76 static void cbuf_add(struct cbuf *cb, int n)
  77 {
  78         cb->len += n;
  79 }
  80
  81 static int cbuf_data(struct cbuf *cb)
  82 {
  83         return ((cb->base + cb->len) & cb->mask);
  84 }
  85
  86 static void cbuf_init(struct cbuf *cb, int size)
  87 {
  88         cb->base = cb->len = 0;
  89         cb->mask = size-1;
  90 }
  91
  92 static void cbuf_eat(struct cbuf *cb, int n)
  93 {
  94         cb->len  -= n;
  95         cb->base += n;
  96         cb->base &= cb->mask;
  97 }
  98
  99 static bool cbuf_empty(struct cbuf *cb)
 100 {
 101         return cb->len == 0;
 102 }
 103
 104 struct connection {
 105         struct socket *sock;    /* NULL if not connected */
 106         uint32_t nodeid;        /* So we know who we are in the list */
 107         struct mutex sock_mutex;
 108         unsigned long flags;
 109 #define CF_READ_PENDING 1
 110 #define CF_WRITE_PENDING 2
 111 #define CF_CONNECT_PENDING 3
 112 #define CF_INIT_PENDING 4
 113 #define CF_IS_OTHERCON 5
 114 #define CF_CLOSE 6
 115 #define CF_APP_LIMITED 7
 116         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 117         spinlock_t writequeue_lock;
 118         int (*rx_action) (struct connection *); /* What to do when active */
 119         void (*connect_action) (struct connection *);   /* What to do to connect */
 120         struct page *rx_page;
 121         struct cbuf cb;
 122         int retries;
 123 #define MAX_CONNECT_RETRIES 3
 124         int sctp_assoc;
 125         struct hlist_node list;
 126         struct connection *othercon;
 127         struct work_struct rwork; /* Receive workqueue */
 128         struct work_struct swork; /* Send workqueue */
 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         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 148 };
 149
 150 static LIST_HEAD(dlm_node_addrs);
 151 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 152
 153 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 154 static int dlm_local_count;
 155 static int dlm_allow_conn;
 156
 157 /* Work queues */
 158 static struct workqueue_struct *recv_workqueue;
 159 static struct workqueue_struct *send_workqueue;
 160
 161 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 162 static DEFINE_MUTEX(connections_lock);
 163 static struct kmem_cache *con_cache;
 164
 165 static void process_recv_sockets(struct work_struct *work);
 166 static void process_send_sockets(struct work_struct *work);
 167
 168
 169 /* This is deliberately very simple because most clusters have simple
 170    sequential nodeids, so we should be able to go straight to a connection
 171    struct in the array */
 172 static inline int nodeid_hash(int nodeid)
 173 {
 174         return nodeid & (CONN_HASH_SIZE-1);
 175 }
 176
 177 static struct connection *__find_con(int nodeid)
 178 {
 179         int r;
 180         struct connection *con;
 181
 182         r = nodeid_hash(nodeid);
 183
 184         hlist_for_each_entry(con, &connection_hash[r], list) {
 185                 if (con->nodeid == nodeid)
 186                         return con;
 187         }
 188         return NULL;
 189 }
 190
 191 /*
 192  * If 'allocation' is zero then we don't attempt to create a new
 193  * connection structure for this node.
 194  */
 195 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 196 {
 197         struct connection *con = NULL;
 198         int r;
 199
 200         con = __find_con(nodeid);
 201         if (con || !alloc)
 202                 return con;
 203
 204         con = kmem_cache_zalloc(con_cache, alloc);
 205         if (!con)
 206                 return NULL;
 207
 208         r = nodeid_hash(nodeid);
 209         hlist_add_head(&con->list, &connection_hash[r]);
 210
 211         con->nodeid = nodeid;
 212         mutex_init(&con->sock_mutex);
 213         INIT_LIST_HEAD(&con->writequeue);
 214         spin_lock_init(&con->writequeue_lock);
 215         INIT_WORK(&con->swork, process_send_sockets);
 216         INIT_WORK(&con->rwork, process_recv_sockets);
 217
 218         /* Setup action pointers for child sockets */
 219         if (con->nodeid) {
 220                 struct connection *zerocon = __find_con(0);
 221
 222                 con->connect_action = zerocon->connect_action;
 223                 if (!con->rx_action)
 224                         con->rx_action = zerocon->rx_action;
 225         }
 226
 227         return con;
 228 }
 229
 230 /* Loop round all connections */
 231 static void foreach_conn(void (*conn_func)(struct connection *c))
 232 {
 233         int i;
 234         struct hlist_node *n;
 235         struct connection *con;
 236
 237         for (i = 0; i < CONN_HASH_SIZE; i++) {
 238                 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
 239                         conn_func(con);
 240         }
 241 }
 242
 243 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 244 {
 245         struct connection *con;
 246
 247         mutex_lock(&connections_lock);
 248         con = __nodeid2con(nodeid, allocation);
 249         mutex_unlock(&connections_lock);
 250
 251         return con;
 252 }
 253
 254 /* This is a bit drastic, but only called when things go wrong */
 255 static struct connection *assoc2con(int assoc_id)
 256 {
 257         int i;
 258         struct connection *con;
 259
 260         mutex_lock(&connections_lock);
 261
 262         for (i = 0 ; i < CONN_HASH_SIZE; i++) {
 263                 hlist_for_each_entry(con, &connection_hash[i], list) {
 264                         if (con->sctp_assoc == assoc_id) {
 265                                 mutex_unlock(&connections_lock);
 266                                 return con;
 267                         }
 268                 }
 269         }
 270         mutex_unlock(&connections_lock);
 271         return NULL;
 272 }
 273
 274 static struct dlm_node_addr *find_node_addr(int nodeid)
 275 {
 276         struct dlm_node_addr *na;
 277
 278         list_for_each_entry(na, &dlm_node_addrs, list) {
 279                 if (na->nodeid == nodeid)
 280                         return na;
 281         }
 282         return NULL;
 283 }
 284
 285 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
 286 {
 287         switch (x->ss_family) {
 288         case AF_INET: {
 289                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 290                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 291                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 292                         return 0;
 293                 if (sinx->sin_port != siny->sin_port)
 294                         return 0;
 295                 break;
 296         }
 297         case AF_INET6: {
 298                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 299                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 300                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 301                         return 0;
 302                 if (sinx->sin6_port != siny->sin6_port)
 303                         return 0;
 304                 break;
 305         }
 306         default:
 307                 return 0;
 308         }
 309         return 1;
 310 }
 311
 312 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 313                           struct sockaddr *sa_out)
 314 {
 315         struct sockaddr_storage sas;
 316         struct dlm_node_addr *na;
 317
 318         if (!dlm_local_count)
 319                 return -1;
 320
 321         spin_lock(&dlm_node_addrs_spin);
 322         na = find_node_addr(nodeid);
 323         if (na && na->addr_count)
 324                 memcpy(&sas, na->addr[0], sizeof(struct sockaddr_storage));
 325         spin_unlock(&dlm_node_addrs_spin);
 326
 327         if (!na)
 328                 return -EEXIST;
 329
 330         if (!na->addr_count)
 331                 return -ENOENT;
 332
 333         if (sas_out)
 334                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 335
 336         if (!sa_out)
 337                 return 0;
 338
 339         if (dlm_local_addr[0]->ss_family == AF_INET) {
 340                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 341                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 342                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 343         } else {
 344                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 345                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 346                 ret6->sin6_addr = in6->sin6_addr;
 347         }
 348
 349         return 0;
 350 }
 351
 352 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
 353 {
 354         struct dlm_node_addr *na;
 355         int rv = -EEXIST;
 356
 357         spin_lock(&dlm_node_addrs_spin);
 358         list_for_each_entry(na, &dlm_node_addrs, list) {
 359                 if (!na->addr_count)
 360                         continue;
 361
 362                 if (!addr_compare(na->addr[0], addr))
 363                         continue;
 364
 365                 *nodeid = na->nodeid;
 366                 rv = 0;
 367                 break;
 368         }
 369         spin_unlock(&dlm_node_addrs_spin);
 370         return rv;
 371 }
 372
 373 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 374 {
 375         struct sockaddr_storage *new_addr;
 376         struct dlm_node_addr *new_node, *na;
 377
 378         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 379         if (!new_node)
 380                 return -ENOMEM;
 381
 382         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 383         if (!new_addr) {
 384                 kfree(new_node);
 385                 return -ENOMEM;
 386         }
 387
 388         memcpy(new_addr, addr, len);
 389
 390         spin_lock(&dlm_node_addrs_spin);
 391         na = find_node_addr(nodeid);
 392         if (!na) {
 393                 new_node->nodeid = nodeid;
 394                 new_node->addr[0] = new_addr;
 395                 new_node->addr_count = 1;
 396                 list_add(&new_node->list, &dlm_node_addrs);
 397                 spin_unlock(&dlm_node_addrs_spin);
 398                 return 0;
 399         }
 400
 401         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 402                 spin_unlock(&dlm_node_addrs_spin);
 403                 kfree(new_addr);
 404                 kfree(new_node);
 405                 return -ENOSPC;
 406         }
 407
 408         na->addr[na->addr_count++] = new_addr;
 409         spin_unlock(&dlm_node_addrs_spin);
 410         kfree(new_node);
 411         return 0;
 412 }
 413
 414 /* Data available on socket or listen socket received a connect */
 415 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 416 {
 417         struct connection *con = sock2con(sk);
 418         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 419                 queue_work(recv_workqueue, &con->rwork);
 420 }
 421
 422 static void lowcomms_write_space(struct sock *sk)
 423 {
 424         struct connection *con = sock2con(sk);
 425
 426         if (!con)
 427                 return;
 428
 429         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 430
 431         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 432                 con->sock->sk->sk_write_pending--;
 433                 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
 434         }
 435
 436         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 437                 queue_work(send_workqueue, &con->swork);
 438 }
 439
 440 static inline void lowcomms_connect_sock(struct connection *con)
 441 {
 442         if (test_bit(CF_CLOSE, &con->flags))
 443                 return;
 444         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 445                 queue_work(send_workqueue, &con->swork);
 446 }
 447
 448 static void lowcomms_state_change(struct sock *sk)
 449 {
 450         if (sk->sk_state == TCP_ESTABLISHED)
 451                 lowcomms_write_space(sk);
 452 }
 453
 454 int dlm_lowcomms_connect_node(int nodeid)
 455 {
 456         struct connection *con;
 457
 458         /* with sctp there's no connecting without sending */
 459         if (dlm_config.ci_protocol != 0)
 460                 return 0;
 461
 462         if (nodeid == dlm_our_nodeid())
 463                 return 0;
 464
 465         con = nodeid2con(nodeid, GFP_NOFS);
 466         if (!con)
 467                 return -ENOMEM;
 468         lowcomms_connect_sock(con);
 469         return 0;
 470 }
 471
 472 /* Make a socket active */
 473 static void add_sock(struct socket *sock, struct connection *con)
 474 {
 475         con->sock = sock;
 476
 477         /* Install a data_ready callback */
 478         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 479         con->sock->sk->sk_write_space = lowcomms_write_space;
 480         con->sock->sk->sk_state_change = lowcomms_state_change;
 481         con->sock->sk->sk_user_data = con;
 482         con->sock->sk->sk_allocation = GFP_NOFS;
 483 }
 484
 485 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 486    length */
 487 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 488                           int *addr_len)
 489 {
 490         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 491         if (saddr->ss_family == AF_INET) {
 492                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 493                 in4_addr->sin_port = cpu_to_be16(port);
 494                 *addr_len = sizeof(struct sockaddr_in);
 495                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 496         } else {
 497                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 498                 in6_addr->sin6_port = cpu_to_be16(port);
 499                 *addr_len = sizeof(struct sockaddr_in6);
 500         }
 501         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 502 }
 503
 504 /* Close a remote connection and tidy up */
 505 static void close_connection(struct connection *con, bool and_other)
 506 {
 507         mutex_lock(&con->sock_mutex);
 508
 509         if (con->sock) {
 510                 sock_release(con->sock);
 511                 con->sock = NULL;
 512         }
 513         if (con->othercon && and_other) {
 514                 /* Will only re-enter once. */
 515                 close_connection(con->othercon, false);
 516         }
 517         if (con->rx_page) {
 518                 __free_page(con->rx_page);
 519                 con->rx_page = NULL;
 520         }
 521
 522         con->retries = 0;
 523         mutex_unlock(&con->sock_mutex);
 524 }
 525
 526 /* We only send shutdown messages to nodes that are not part of the cluster */
 527 static void sctp_send_shutdown(sctp_assoc_t associd)
 528 {
 529         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 530         struct msghdr outmessage;
 531         struct cmsghdr *cmsg;
 532         struct sctp_sndrcvinfo *sinfo;
 533         int ret;
 534         struct connection *con;
 535
 536         con = nodeid2con(0,0);
 537         BUG_ON(con == NULL);
 538
 539         outmessage.msg_name = NULL;
 540         outmessage.msg_namelen = 0;
 541         outmessage.msg_control = outcmsg;
 542         outmessage.msg_controllen = sizeof(outcmsg);
 543         outmessage.msg_flags = MSG_EOR;
 544
 545         cmsg = CMSG_FIRSTHDR(&outmessage);
 546         cmsg->cmsg_level = IPPROTO_SCTP;
 547         cmsg->cmsg_type = SCTP_SNDRCV;
 548         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 549         outmessage.msg_controllen = cmsg->cmsg_len;
 550         sinfo = CMSG_DATA(cmsg);
 551         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 552
 553         sinfo->sinfo_flags |= MSG_EOF;
 554         sinfo->sinfo_assoc_id = associd;
 555
 556         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 557
 558         if (ret != 0)
 559                 log_print("send EOF to node failed: %d", ret);
 560 }
 561
 562 static void sctp_init_failed_foreach(struct connection *con)
 563 {
 564         con->sctp_assoc = 0;
 565         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 566                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 567                         queue_work(send_workqueue, &con->swork);
 568         }
 569 }
 570
 571 /* INIT failed but we don't know which node...
 572    restart INIT on all pending nodes */
 573 static void sctp_init_failed(void)
 574 {
 575         mutex_lock(&connections_lock);
 576
 577         foreach_conn(sctp_init_failed_foreach);
 578
 579         mutex_unlock(&connections_lock);
 580 }
 581
 582 /* Something happened to an association */
 583 static void process_sctp_notification(struct connection *con,
 584                                       struct msghdr *msg, char *buf)
 585 {
 586         union sctp_notification *sn = (union sctp_notification *)buf;
 587
 588         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 589                 switch (sn->sn_assoc_change.sac_state) {
 590
 591                 case SCTP_COMM_UP:
 592                 case SCTP_RESTART:
 593                 {
 594                         /* Check that the new node is in the lockspace */
 595                         struct sctp_prim prim;
 596                         int nodeid;
 597                         int prim_len, ret;
 598                         int addr_len;
 599                         struct connection *new_con;
 600
 601                         /*
 602                          * We get this before any data for an association.
 603                          * We verify that the node is in the cluster and
 604                          * then peel off a socket for it.
 605                          */
 606                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 607                                 log_print("COMM_UP for invalid assoc ID %d",
 608                                          (int)sn->sn_assoc_change.sac_assoc_id);
 609                                 sctp_init_failed();
 610                                 return;
 611                         }
 612                         memset(&prim, 0, sizeof(struct sctp_prim));
 613                         prim_len = sizeof(struct sctp_prim);
 614                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 615
 616                         ret = kernel_getsockopt(con->sock,
 617                                                 IPPROTO_SCTP,
 618                                                 SCTP_PRIMARY_ADDR,
 619                                                 (char*)&prim,
 620                                                 &prim_len);
 621                         if (ret < 0) {
 622                                 log_print("getsockopt/sctp_primary_addr on "
 623                                           "new assoc %d failed : %d",
 624                                           (int)sn->sn_assoc_change.sac_assoc_id,
 625                                           ret);
 626
 627                                 /* Retry INIT later */
 628                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 629                                 if (new_con)
 630                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 631                                 return;
 632                         }
 633                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 634                         if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 635                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 636                                 log_print("reject connect from unknown addr");
 637                                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 638                                                      b, sizeof(struct sockaddr_storage));
 639                                 sctp_send_shutdown(prim.ssp_assoc_id);
 640                                 return;
 641                         }
 642
 643                         new_con = nodeid2con(nodeid, GFP_NOFS);
 644                         if (!new_con)
 645                                 return;
 646
 647                         /* Peel off a new sock */
 648                         sctp_lock_sock(con->sock->sk);
 649                         ret = sctp_do_peeloff(con->sock->sk,
 650                                 sn->sn_assoc_change.sac_assoc_id,
 651                                 &new_con->sock);
 652                         sctp_release_sock(con->sock->sk);
 653                         if (ret < 0) {
 654                                 log_print("Can't peel off a socket for "
 655                                           "connection %d to node %d: err=%d",
 656                                           (int)sn->sn_assoc_change.sac_assoc_id,
 657                                           nodeid, ret);
 658                                 return;
 659                         }
 660                         add_sock(new_con->sock, new_con);
 661
 662                         log_print("connecting to %d sctp association %d",
 663                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 664
 665                         /* Send any pending writes */
 666                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 667                         clear_bit(CF_INIT_PENDING, &con->flags);
 668                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 669                                 queue_work(send_workqueue, &new_con->swork);
 670                         }
 671                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 672                                 queue_work(recv_workqueue, &new_con->rwork);
 673                 }
 674                 break;
 675
 676                 case SCTP_COMM_LOST:
 677                 case SCTP_SHUTDOWN_COMP:
 678                 {
 679                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 680                         if (con) {
 681                                 con->sctp_assoc = 0;
 682                         }
 683                 }
 684                 break;
 685
 686                 /* We don't know which INIT failed, so clear the PENDING flags
 687                  * on them all.  if assoc_id is zero then it will then try
 688                  * again */
 689
 690                 case SCTP_CANT_STR_ASSOC:
 691                 {
 692                         log_print("Can't start SCTP association - retrying");
 693                         sctp_init_failed();
 694                 }
 695                 break;
 696
 697                 default:
 698                         log_print("unexpected SCTP assoc change id=%d state=%d",
 699                                   (int)sn->sn_assoc_change.sac_assoc_id,
 700                                   sn->sn_assoc_change.sac_state);
 701                 }
 702         }
 703 }
 704
 705 /* Data received from remote end */
 706 static int receive_from_sock(struct connection *con)
 707 {
 708         int ret = 0;
 709         struct msghdr msg = {};
 710         struct kvec iov[2];
 711         unsigned len;
 712         int r;
 713         int call_again_soon = 0;
 714         int nvec;
 715         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 716
 717         mutex_lock(&con->sock_mutex);
 718
 719         if (con->sock == NULL) {
 720                 ret = -EAGAIN;
 721                 goto out_close;
 722         }
 723
 724         if (con->rx_page == NULL) {
 725                 /*
 726                  * This doesn't need to be atomic, but I think it should
 727                  * improve performance if it is.
 728                  */
 729                 con->rx_page = alloc_page(GFP_ATOMIC);
 730                 if (con->rx_page == NULL)
 731                         goto out_resched;
 732                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 733         }
 734
 735         /* Only SCTP needs these really */
 736         memset(&incmsg, 0, sizeof(incmsg));
 737         msg.msg_control = incmsg;
 738         msg.msg_controllen = sizeof(incmsg);
 739
 740         /*
 741          * iov[0] is the bit of the circular buffer between the current end
 742          * point (cb.base + cb.len) and the end of the buffer.
 743          */
 744         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 745         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 746         iov[1].iov_len = 0;
 747         nvec = 1;
 748
 749         /*
 750          * iov[1] is the bit of the circular buffer between the start of the
 751          * buffer and the start of the currently used section (cb.base)
 752          */
 753         if (cbuf_data(&con->cb) >= con->cb.base) {
 754                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 755                 iov[1].iov_len = con->cb.base;
 756                 iov[1].iov_base = page_address(con->rx_page);
 757                 nvec = 2;
 758         }
 759         len = iov[0].iov_len + iov[1].iov_len;
 760
 761         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 762                                MSG_DONTWAIT | MSG_NOSIGNAL);
 763         if (ret <= 0)
 764                 goto out_close;
 765
 766         /* Process SCTP notifications */
 767         if (msg.msg_flags & MSG_NOTIFICATION) {
 768                 msg.msg_control = incmsg;
 769                 msg.msg_controllen = sizeof(incmsg);
 770
 771                 process_sctp_notification(con, &msg,
 772                                 page_address(con->rx_page) + con->cb.base);
 773                 mutex_unlock(&con->sock_mutex);
 774                 return 0;
 775         }
 776         BUG_ON(con->nodeid == 0);
 777
 778         if (ret == len)
 779                 call_again_soon = 1;
 780         cbuf_add(&con->cb, ret);
 781         ret = dlm_process_incoming_buffer(con->nodeid,
 782                                           page_address(con->rx_page),
 783                                           con->cb.base, con->cb.len,
 784                                           PAGE_CACHE_SIZE);
 785         if (ret == -EBADMSG) {
 786                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 787                           "iov_len=%u, iov_base[0]=%p, read=%d",
 788                           page_address(con->rx_page), con->cb.base, con->cb.len,
 789                           len, iov[0].iov_base, r);
 790         }
 791         if (ret < 0)
 792                 goto out_close;
 793         cbuf_eat(&con->cb, ret);
 794
 795         if (cbuf_empty(&con->cb) && !call_again_soon) {
 796                 __free_page(con->rx_page);
 797                 con->rx_page = NULL;
 798         }
 799
 800         if (call_again_soon)
 801                 goto out_resched;
 802         mutex_unlock(&con->sock_mutex);
 803         return 0;
 804
 805 out_resched:
 806         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 807                 queue_work(recv_workqueue, &con->rwork);
 808         mutex_unlock(&con->sock_mutex);
 809         return -EAGAIN;
 810
 811 out_close:
 812         mutex_unlock(&con->sock_mutex);
 813         if (ret != -EAGAIN) {
 814                 close_connection(con, false);
 815                 /* Reconnect when there is something to send */
 816         }
 817         /* Don't return success if we really got EOF */
 818         if (ret == 0)
 819                 ret = -EAGAIN;
 820
 821         return ret;
 822 }
 823
 824 /* Listening socket is busy, accept a connection */
 825 static int tcp_accept_from_sock(struct connection *con)
 826 {
 827         int result;
 828         struct sockaddr_storage peeraddr;
 829         struct socket *newsock;
 830         int len;
 831         int nodeid;
 832         struct connection *newcon;
 833         struct connection *addcon;
 834
 835         mutex_lock(&connections_lock);
 836         if (!dlm_allow_conn) {
 837                 mutex_unlock(&connections_lock);
 838                 return -1;
 839         }
 840         mutex_unlock(&connections_lock);
 841
 842         memset(&peeraddr, 0, sizeof(peeraddr));
 843         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 844                                   IPPROTO_TCP, &newsock);
 845         if (result < 0)
 846                 return -ENOMEM;
 847
 848         mutex_lock_nested(&con->sock_mutex, 0);
 849
 850         result = -ENOTCONN;
 851         if (con->sock == NULL)
 852                 goto accept_err;
 853
 854         newsock->type = con->sock->type;
 855         newsock->ops = con->sock->ops;
 856
 857         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 858         if (result < 0)
 859                 goto accept_err;
 860
 861         /* Get the connected socket's peer */
 862         memset(&peeraddr, 0, sizeof(peeraddr));
 863         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 864                                   &len, 2)) {
 865                 result = -ECONNABORTED;
 866                 goto accept_err;
 867         }
 868
 869         /* Get the new node's NODEID */
 870         make_sockaddr(&peeraddr, 0, &len);
 871         if (addr_to_nodeid(&peeraddr, &nodeid)) {
 872                 unsigned char *b=(unsigned char *)&peeraddr;
 873                 log_print("connect from non cluster node");
 874                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 875                                      b, sizeof(struct sockaddr_storage));
 876                 sock_release(newsock);
 877                 mutex_unlock(&con->sock_mutex);
 878                 return -1;
 879         }
 880
 881         log_print("got connection from %d", nodeid);
 882
 883         /*  Check to see if we already have a connection to this node. This
 884          *  could happen if the two nodes initiate a connection at roughly
 885          *  the same time and the connections cross on the wire.
 886          *  In this case we store the incoming one in "othercon"
 887          */
 888         newcon = nodeid2con(nodeid, GFP_NOFS);
 889         if (!newcon) {
 890                 result = -ENOMEM;
 891                 goto accept_err;
 892         }
 893         mutex_lock_nested(&newcon->sock_mutex, 1);
 894         if (newcon->sock) {
 895                 struct connection *othercon = newcon->othercon;
 896
 897                 if (!othercon) {
 898                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 899                         if (!othercon) {
 900                                 log_print("failed to allocate incoming socket");
 901                                 mutex_unlock(&newcon->sock_mutex);
 902                                 result = -ENOMEM;
 903                                 goto accept_err;
 904                         }
 905                         othercon->nodeid = nodeid;
 906                         othercon->rx_action = receive_from_sock;
 907                         mutex_init(&othercon->sock_mutex);
 908                         INIT_WORK(&othercon->swork, process_send_sockets);
 909                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 910                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 911                 }
 912                 if (!othercon->sock) {
 913                         newcon->othercon = othercon;
 914                         othercon->sock = newsock;
 915                         newsock->sk->sk_user_data = othercon;
 916                         add_sock(newsock, othercon);
 917                         addcon = othercon;
 918                 }
 919                 else {
 920                         printk("Extra connection from node %d attempted\n", nodeid);
 921                         result = -EAGAIN;
 922                         mutex_unlock(&newcon->sock_mutex);
 923                         goto accept_err;
 924                 }
 925         }
 926         else {
 927                 newsock->sk->sk_user_data = newcon;
 928                 newcon->rx_action = receive_from_sock;
 929                 add_sock(newsock, newcon);
 930                 addcon = newcon;
 931         }
 932
 933         mutex_unlock(&newcon->sock_mutex);
 934
 935         /*
 936          * Add it to the active queue in case we got data
 937          * between processing the accept adding the socket
 938          * to the read_sockets list
 939          */
 940         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 941                 queue_work(recv_workqueue, &addcon->rwork);
 942         mutex_unlock(&con->sock_mutex);
 943
 944         return 0;
 945
 946 accept_err:
 947         mutex_unlock(&con->sock_mutex);
 948         sock_release(newsock);
 949
 950         if (result != -EAGAIN)
 951                 log_print("error accepting connection from node: %d", result);
 952         return result;
 953 }
 954
 955 static void free_entry(struct writequeue_entry *e)
 956 {
 957         __free_page(e->page);
 958         kfree(e);
 959 }
 960
 961 /* Initiate an SCTP association.
 962    This is a special case of send_to_sock() in that we don't yet have a
 963    peeled-off socket for this association, so we use the listening socket
 964    and add the primary IP address of the remote node.
 965  */
 966 static void sctp_init_assoc(struct connection *con)
 967 {
 968         struct sockaddr_storage rem_addr;
 969         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 970         struct msghdr outmessage;
 971         struct cmsghdr *cmsg;
 972         struct sctp_sndrcvinfo *sinfo;
 973         struct connection *base_con;
 974         struct writequeue_entry *e;
 975         int len, offset;
 976         int ret;
 977         int addrlen;
 978         struct kvec iov[1];
 979
 980         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 981                 return;
 982
 983         if (con->retries++ > MAX_CONNECT_RETRIES)
 984                 return;
 985
 986         if (nodeid_to_addr(con->nodeid, NULL, (struct sockaddr *)&rem_addr)) {
 987                 log_print("no address for nodeid %d", con->nodeid);
 988                 return;
 989         }
 990         base_con = nodeid2con(0, 0);
 991         BUG_ON(base_con == NULL);
 992
 993         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 994
 995         outmessage.msg_name = &rem_addr;
 996         outmessage.msg_namelen = addrlen;
 997         outmessage.msg_control = outcmsg;
 998         outmessage.msg_controllen = sizeof(outcmsg);
 999         outmessage.msg_flags = MSG_EOR;
1000
1001         spin_lock(&con->writequeue_lock);
1002
1003         if (list_empty(&con->writequeue)) {
1004                 spin_unlock(&con->writequeue_lock);
1005                 log_print("writequeue empty for nodeid %d", con->nodeid);
1006                 return;
1007         }
1008
1009         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1010         len = e->len;
1011         offset = e->offset;
1012         spin_unlock(&con->writequeue_lock);
1013
1014         /* Send the first block off the write queue */
1015         iov[0].iov_base = page_address(e->page)+offset;
1016         iov[0].iov_len = len;
1017
1018         cmsg = CMSG_FIRSTHDR(&outmessage);
1019         cmsg->cmsg_level = IPPROTO_SCTP;
1020         cmsg->cmsg_type = SCTP_SNDRCV;
1021         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
1022         sinfo = CMSG_DATA(cmsg);
1023         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
1024         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
1025         outmessage.msg_controllen = cmsg->cmsg_len;
1026
1027         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
1028         if (ret < 0) {
1029                 log_print("Send first packet to node %d failed: %d",
1030                           con->nodeid, ret);
1031
1032                 /* Try again later */
1033                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1034                 clear_bit(CF_INIT_PENDING, &con->flags);
1035         }
1036         else {
1037                 spin_lock(&con->writequeue_lock);
1038                 e->offset += ret;
1039                 e->len -= ret;
1040
1041                 if (e->len == 0 && e->users == 0) {
1042                         list_del(&e->list);
1043                         free_entry(e);
1044                 }
1045                 spin_unlock(&con->writequeue_lock);
1046         }
1047 }
1048
1049 /* Connect a new socket to its peer */
1050 static void tcp_connect_to_sock(struct connection *con)
1051 {
1052         struct sockaddr_storage saddr, src_addr;
1053         int addr_len;
1054         struct socket *sock = NULL;
1055         int one = 1;
1056         int result;
1057
1058         if (con->nodeid == 0) {
1059                 log_print("attempt to connect sock 0 foiled");
1060                 return;
1061         }
1062
1063         mutex_lock(&con->sock_mutex);
1064         if (con->retries++ > MAX_CONNECT_RETRIES)
1065                 goto out;
1066
1067         /* Some odd races can cause double-connects, ignore them */
1068         if (con->sock)
1069                 goto out;
1070
1071         /* Create a socket to communicate with */
1072         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1073                                   IPPROTO_TCP, &sock);
1074         if (result < 0)
1075                 goto out_err;
1076
1077         memset(&saddr, 0, sizeof(saddr));
1078         result = nodeid_to_addr(con->nodeid, &saddr, NULL);
1079         if (result < 0) {
1080                 log_print("no address for nodeid %d", con->nodeid);
1081                 goto out_err;
1082         }
1083
1084         sock->sk->sk_user_data = con;
1085         con->rx_action = receive_from_sock;
1086         con->connect_action = tcp_connect_to_sock;
1087         add_sock(sock, con);
1088
1089         /* Bind to our cluster-known address connecting to avoid
1090            routing problems */
1091         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1092         make_sockaddr(&src_addr, 0, &addr_len);
1093         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1094                                  addr_len);
1095         if (result < 0) {
1096                 log_print("could not bind for connect: %d", result);
1097                 /* This *may* not indicate a critical error */
1098         }
1099
1100         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1101
1102         log_print("connecting to %d", con->nodeid);
1103
1104         /* Turn off Nagle's algorithm */
1105         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1106                           sizeof(one));
1107
1108         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1109                                    O_NONBLOCK);
1110         if (result == -EINPROGRESS)
1111                 result = 0;
1112         if (result == 0)
1113                 goto out;
1114
1115 out_err:
1116         if (con->sock) {
1117                 sock_release(con->sock);
1118                 con->sock = NULL;
1119         } else if (sock) {
1120                 sock_release(sock);
1121         }
1122         /*
1123          * Some errors are fatal and this list might need adjusting. For other
1124          * errors we try again until the max number of retries is reached.
1125          */
1126         if (result != -EHOSTUNREACH &&
1127             result != -ENETUNREACH &&
1128             result != -ENETDOWN &&
1129             result != -EINVAL &&
1130             result != -EPROTONOSUPPORT) {
1131                 log_print("connect %d try %d error %d", con->nodeid,
1132                           con->retries, result);
1133                 mutex_unlock(&con->sock_mutex);
1134                 msleep(1000);
1135                 lowcomms_connect_sock(con);
1136                 return;
1137         }
1138 out:
1139         mutex_unlock(&con->sock_mutex);
1140         return;
1141 }
1142
1143 static struct socket *tcp_create_listen_sock(struct connection *con,
1144                                              struct sockaddr_storage *saddr)
1145 {
1146         struct socket *sock = NULL;
1147         int result = 0;
1148         int one = 1;
1149         int addr_len;
1150
1151         if (dlm_local_addr[0]->ss_family == AF_INET)
1152                 addr_len = sizeof(struct sockaddr_in);
1153         else
1154                 addr_len = sizeof(struct sockaddr_in6);
1155
1156         /* Create a socket to communicate with */
1157         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1158                                   IPPROTO_TCP, &sock);
1159         if (result < 0) {
1160                 log_print("Can't create listening comms socket");
1161                 goto create_out;
1162         }
1163
1164         /* Turn off Nagle's algorithm */
1165         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1166                           sizeof(one));
1167
1168         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1169                                    (char *)&one, sizeof(one));
1170
1171         if (result < 0) {
1172                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1173         }
1174         con->rx_action = tcp_accept_from_sock;
1175         con->connect_action = tcp_connect_to_sock;
1176
1177         /* Bind to our port */
1178         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1179         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1180         if (result < 0) {
1181                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1182                 sock_release(sock);
1183                 sock = NULL;
1184                 con->sock = NULL;
1185                 goto create_out;
1186         }
1187         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1188                                  (char *)&one, sizeof(one));
1189         if (result < 0) {
1190                 log_print("Set keepalive failed: %d", result);
1191         }
1192
1193         result = sock->ops->listen(sock, 5);
1194         if (result < 0) {
1195                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1196                 sock_release(sock);
1197                 sock = NULL;
1198                 goto create_out;
1199         }
1200
1201 create_out:
1202         return sock;
1203 }
1204
1205 /* Get local addresses */
1206 static void init_local(void)
1207 {
1208         struct sockaddr_storage sas, *addr;
1209         int i;
1210
1211         dlm_local_count = 0;
1212         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1213                 if (dlm_our_addr(&sas, i))
1214                         break;
1215
1216                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1217                 if (!addr)
1218                         break;
1219                 memcpy(addr, &sas, sizeof(*addr));
1220                 dlm_local_addr[dlm_local_count++] = addr;
1221         }
1222 }
1223
1224 /* Bind to an IP address. SCTP allows multiple address so it can do
1225    multi-homing */
1226 static int add_sctp_bind_addr(struct connection *sctp_con,
1227                               struct sockaddr_storage *addr,
1228                               int addr_len, int num)
1229 {
1230         int result = 0;
1231
1232         if (num == 1)
1233                 result = kernel_bind(sctp_con->sock,
1234                                      (struct sockaddr *) addr,
1235                                      addr_len);
1236         else
1237                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1238                                            SCTP_SOCKOPT_BINDX_ADD,
1239                                            (char *)addr, addr_len);
1240
1241         if (result < 0)
1242                 log_print("Can't bind to port %d addr number %d",
1243                           dlm_config.ci_tcp_port, num);
1244
1245         return result;
1246 }
1247
1248 /* Initialise SCTP socket and bind to all interfaces */
1249 static int sctp_listen_for_all(void)
1250 {
1251         struct socket *sock = NULL;
1252         struct sockaddr_storage localaddr;
1253         struct sctp_event_subscribe subscribe;
1254         int result = -EINVAL, num = 1, i, addr_len;
1255         struct connection *con = nodeid2con(0, GFP_NOFS);
1256         int bufsize = NEEDED_RMEM;
1257
1258         if (!con)
1259                 return -ENOMEM;
1260
1261         log_print("Using SCTP for communications");
1262
1263         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1264                                   IPPROTO_SCTP, &sock);
1265         if (result < 0) {
1266                 log_print("Can't create comms socket, check SCTP is loaded");
1267                 goto out;
1268         }
1269
1270         /* Listen for events */
1271         memset(&subscribe, 0, sizeof(subscribe));
1272         subscribe.sctp_data_io_event = 1;
1273         subscribe.sctp_association_event = 1;
1274         subscribe.sctp_send_failure_event = 1;
1275         subscribe.sctp_shutdown_event = 1;
1276         subscribe.sctp_partial_delivery_event = 1;
1277
1278         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1279                                  (char *)&bufsize, sizeof(bufsize));
1280         if (result)
1281                 log_print("Error increasing buffer space on socket %d", result);
1282
1283         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1284                                    (char *)&subscribe, sizeof(subscribe));
1285         if (result < 0) {
1286                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1287                           result);
1288                 goto create_delsock;
1289         }
1290
1291         /* Init con struct */
1292         sock->sk->sk_user_data = con;
1293         con->sock = sock;
1294         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1295         con->rx_action = receive_from_sock;
1296         con->connect_action = sctp_init_assoc;
1297
1298         /* Bind to all interfaces. */
1299         for (i = 0; i < dlm_local_count; i++) {
1300                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1301                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1302
1303                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1304                 if (result)
1305                         goto create_delsock;
1306                 ++num;
1307         }
1308
1309         result = sock->ops->listen(sock, 5);
1310         if (result < 0) {
1311                 log_print("Can't set socket listening");
1312                 goto create_delsock;
1313         }
1314
1315         return 0;
1316
1317 create_delsock:
1318         sock_release(sock);
1319         con->sock = NULL;
1320 out:
1321         return result;
1322 }
1323
1324 static int tcp_listen_for_all(void)
1325 {
1326         struct socket *sock = NULL;
1327         struct connection *con = nodeid2con(0, GFP_NOFS);
1328         int result = -EINVAL;
1329
1330         if (!con)
1331                 return -ENOMEM;
1332
1333         /* We don't support multi-homed hosts */
1334         if (dlm_local_addr[1] != NULL) {
1335                 log_print("TCP protocol can't handle multi-homed hosts, "
1336                           "try SCTP");
1337                 return -EINVAL;
1338         }
1339
1340         log_print("Using TCP for communications");
1341
1342         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1343         if (sock) {
1344                 add_sock(sock, con);
1345                 result = 0;
1346         }
1347         else {
1348                 result = -EADDRINUSE;
1349         }
1350
1351         return result;
1352 }
1353
1354
1355
1356 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1357                                                      gfp_t allocation)
1358 {
1359         struct writequeue_entry *entry;
1360
1361         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1362         if (!entry)
1363                 return NULL;
1364
1365         entry->page = alloc_page(allocation);
1366         if (!entry->page) {
1367                 kfree(entry);
1368                 return NULL;
1369         }
1370
1371         entry->offset = 0;
1372         entry->len = 0;
1373         entry->end = 0;
1374         entry->users = 0;
1375         entry->con = con;
1376
1377         return entry;
1378 }
1379
1380 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1381 {
1382         struct connection *con;
1383         struct writequeue_entry *e;
1384         int offset = 0;
1385
1386         con = nodeid2con(nodeid, allocation);
1387         if (!con)
1388                 return NULL;
1389
1390         spin_lock(&con->writequeue_lock);
1391         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1392         if ((&e->list == &con->writequeue) ||
1393             (PAGE_CACHE_SIZE - e->end < len)) {
1394                 e = NULL;
1395         } else {
1396                 offset = e->end;
1397                 e->end += len;
1398                 e->users++;
1399         }
1400         spin_unlock(&con->writequeue_lock);
1401
1402         if (e) {
1403         got_one:
1404                 *ppc = page_address(e->page) + offset;
1405                 return e;
1406         }
1407
1408         e = new_writequeue_entry(con, allocation);
1409         if (e) {
1410                 spin_lock(&con->writequeue_lock);
1411                 offset = e->end;
1412                 e->end += len;
1413                 e->users++;
1414                 list_add_tail(&e->list, &con->writequeue);
1415                 spin_unlock(&con->writequeue_lock);
1416                 goto got_one;
1417         }
1418         return NULL;
1419 }
1420
1421 void dlm_lowcomms_commit_buffer(void *mh)
1422 {
1423         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1424         struct connection *con = e->con;
1425         int users;
1426
1427         spin_lock(&con->writequeue_lock);
1428         users = --e->users;
1429         if (users)
1430                 goto out;
1431         e->len = e->end - e->offset;
1432         spin_unlock(&con->writequeue_lock);
1433
1434         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1435                 queue_work(send_workqueue, &con->swork);
1436         }
1437         return;
1438
1439 out:
1440         spin_unlock(&con->writequeue_lock);
1441         return;
1442 }
1443
1444 /* Send a message */
1445 static void send_to_sock(struct connection *con)
1446 {
1447         int ret = 0;
1448         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1449         struct writequeue_entry *e;
1450         int len, offset;
1451         int count = 0;
1452
1453         mutex_lock(&con->sock_mutex);
1454         if (con->sock == NULL)
1455                 goto out_connect;
1456
1457         spin_lock(&con->writequeue_lock);
1458         for (;;) {
1459                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1460                                list);
1461                 if ((struct list_head *) e == &con->writequeue)
1462                         break;
1463
1464                 len = e->len;
1465                 offset = e->offset;
1466                 BUG_ON(len == 0 && e->users == 0);
1467                 spin_unlock(&con->writequeue_lock);
1468
1469                 ret = 0;
1470                 if (len) {
1471                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1472                                               msg_flags);
1473                         if (ret == -EAGAIN || ret == 0) {
1474                                 if (ret == -EAGAIN &&
1475                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1476                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1477                                         /* Notify TCP that we're limited by the
1478                                          * application window size.
1479                                          */
1480                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1481                                         con->sock->sk->sk_write_pending++;
1482                                 }
1483                                 cond_resched();
1484                                 goto out;
1485                         } else if (ret < 0)
1486                                 goto send_error;
1487                 }
1488
1489                 /* Don't starve people filling buffers */
1490                 if (++count >= MAX_SEND_MSG_COUNT) {
1491                         cond_resched();
1492                         count = 0;
1493                 }
1494
1495                 spin_lock(&con->writequeue_lock);
1496                 e->offset += ret;
1497                 e->len -= ret;
1498
1499                 if (e->len == 0 && e->users == 0) {
1500                         list_del(&e->list);
1501                         free_entry(e);
1502                 }
1503         }
1504         spin_unlock(&con->writequeue_lock);
1505 out:
1506         mutex_unlock(&con->sock_mutex);
1507         return;
1508
1509 send_error:
1510         mutex_unlock(&con->sock_mutex);
1511         close_connection(con, false);
1512         lowcomms_connect_sock(con);
1513         return;
1514
1515 out_connect:
1516         mutex_unlock(&con->sock_mutex);
1517         if (!test_bit(CF_INIT_PENDING, &con->flags))
1518                 lowcomms_connect_sock(con);
1519 }
1520
1521 static void clean_one_writequeue(struct connection *con)
1522 {
1523         struct writequeue_entry *e, *safe;
1524
1525         spin_lock(&con->writequeue_lock);
1526         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1527                 list_del(&e->list);
1528                 free_entry(e);
1529         }
1530         spin_unlock(&con->writequeue_lock);
1531 }
1532
1533 /* Called from recovery when it knows that a node has
1534    left the cluster */
1535 int dlm_lowcomms_close(int nodeid)
1536 {
1537         struct connection *con;
1538         struct dlm_node_addr *na;
1539
1540         log_print("closing connection to node %d", nodeid);
1541         con = nodeid2con(nodeid, 0);
1542         if (con) {
1543                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1544                 clear_bit(CF_WRITE_PENDING, &con->flags);
1545                 set_bit(CF_CLOSE, &con->flags);
1546                 if (cancel_work_sync(&con->swork))
1547                         log_print("canceled swork for node %d", nodeid);
1548                 if (cancel_work_sync(&con->rwork))
1549                         log_print("canceled rwork for node %d", nodeid);
1550                 clean_one_writequeue(con);
1551                 close_connection(con, true);
1552         }
1553
1554         spin_lock(&dlm_node_addrs_spin);
1555         na = find_node_addr(nodeid);
1556         if (na) {
1557                 list_del(&na->list);
1558                 while (na->addr_count--)
1559                         kfree(na->addr[na->addr_count]);
1560                 kfree(na);
1561         }
1562         spin_unlock(&dlm_node_addrs_spin);
1563
1564         return 0;
1565 }
1566
1567 /* Receive workqueue function */
1568 static void process_recv_sockets(struct work_struct *work)
1569 {
1570         struct connection *con = container_of(work, struct connection, rwork);
1571         int err;
1572
1573         clear_bit(CF_READ_PENDING, &con->flags);
1574         do {
1575                 err = con->rx_action(con);
1576         } while (!err);
1577 }
1578
1579 /* Send workqueue function */
1580 static void process_send_sockets(struct work_struct *work)
1581 {
1582         struct connection *con = container_of(work, struct connection, swork);
1583
1584         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1585                 con->connect_action(con);
1586                 set_bit(CF_WRITE_PENDING, &con->flags);
1587         }
1588         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1589                 send_to_sock(con);
1590 }
1591
1592
1593 /* Discard all entries on the write queues */
1594 static void clean_writequeues(void)
1595 {
1596         foreach_conn(clean_one_writequeue);
1597 }
1598
1599 static void work_stop(void)
1600 {
1601         destroy_workqueue(recv_workqueue);
1602         destroy_workqueue(send_workqueue);
1603 }
1604
1605 static int work_start(void)
1606 {
1607         recv_workqueue = alloc_workqueue("dlm_recv",
1608                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1609         if (!recv_workqueue) {
1610                 log_print("can't start dlm_recv");
1611                 return -ENOMEM;
1612         }
1613
1614         send_workqueue = alloc_workqueue("dlm_send",
1615                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1616         if (!send_workqueue) {
1617                 log_print("can't start dlm_send");
1618                 destroy_workqueue(recv_workqueue);
1619                 return -ENOMEM;
1620         }
1621
1622         return 0;
1623 }
1624
1625 static void stop_conn(struct connection *con)
1626 {
1627         con->flags |= 0x0F;
1628         if (con->sock && con->sock->sk)
1629                 con->sock->sk->sk_user_data = NULL;
1630 }
1631
1632 static void free_conn(struct connection *con)
1633 {
1634         close_connection(con, true);
1635         if (con->othercon)
1636                 kmem_cache_free(con_cache, con->othercon);
1637         hlist_del(&con->list);
1638         kmem_cache_free(con_cache, con);
1639 }
1640
1641 void dlm_lowcomms_stop(void)
1642 {
1643         /* Set all the flags to prevent any
1644            socket activity.
1645         */
1646         mutex_lock(&connections_lock);
1647         dlm_allow_conn = 0;
1648         foreach_conn(stop_conn);
1649         mutex_unlock(&connections_lock);
1650
1651         work_stop();
1652
1653         mutex_lock(&connections_lock);
1654         clean_writequeues();
1655
1656         foreach_conn(free_conn);
1657
1658         mutex_unlock(&connections_lock);
1659         kmem_cache_destroy(con_cache);
1660 }
1661
1662 int dlm_lowcomms_start(void)
1663 {
1664         int error = -EINVAL;
1665         struct connection *con;
1666         int i;
1667
1668         for (i = 0; i < CONN_HASH_SIZE; i++)
1669                 INIT_HLIST_HEAD(&connection_hash[i]);
1670
1671         init_local();
1672         if (!dlm_local_count) {
1673                 error = -ENOTCONN;
1674                 log_print("no local IP address has been set");
1675                 goto fail;
1676         }
1677
1678         error = -ENOMEM;
1679         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1680                                       __alignof__(struct connection), 0,
1681                                       NULL);
1682         if (!con_cache)
1683                 goto fail;
1684
1685         error = work_start();
1686         if (error)
1687                 goto fail_destroy;
1688
1689         dlm_allow_conn = 1;
1690
1691         /* Start listening */
1692         if (dlm_config.ci_protocol == 0)
1693                 error = tcp_listen_for_all();
1694         else
1695                 error = sctp_listen_for_all();
1696         if (error)
1697                 goto fail_unlisten;
1698
1699         return 0;
1700
1701 fail_unlisten:
1702         dlm_allow_conn = 0;
1703         con = nodeid2con(0,0);
1704         if (con) {
1705                 close_connection(con, false);
1706                 kmem_cache_free(con_cache, con);
1707         }
1708 fail_destroy:
1709         kmem_cache_destroy(con_cache);
1710 fail:
1711         return error;
1712 }
1713
1714 void dlm_lowcomms_exit(void)
1715 {
1716         struct dlm_node_addr *na, *safe;
1717
1718         spin_lock(&dlm_node_addrs_spin);
1719         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1720                 list_del(&na->list);
1721                 while (na->addr_count--)
1722                         kfree(na->addr[na->addr_count]);
1723                 kfree(na);
1724         }
1725         spin_unlock(&dlm_node_addrs_spin);
1726 }