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