fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is its
  25  * responsibility. It is this layer's
  26  * responsibility to resolve these into IP address or
  27  * whatever it needs for inter-node communication.
  28  *
  29  * The comms level is two kernel threads that deal mainly with
  30  * the receiving of messages from other nodes and passing them
  31  * up to the mid-level comms layer (which understands the
  32  * message format) for execution by the locking core, and
  33  * a send thread which does all the setting up of connections
  34  * to remote nodes and the sending of data. Threads are not allowed
  35  * to send their own data because it may cause them to wait in times
  36  * of high load. Also, this way, the sending thread can collect together
  37  * messages bound for one node and send them in one block.
  38  *
  39  * lowcomms will choose to use either TCP or SCTP as its transport layer
  40  * depending on the configuration variable 'protocol'. This should be set
  41  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  42  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  43  * for the DLM to function.
  44  *
  45  */
  46
  47 #include <asm/ioctls.h>
  48 #include <net/sock.h>
  49 #include <net/tcp.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/file.h>
  52 #include <linux/mutex.h>
  53 #include <linux/sctp.h>
  54 #include <linux/slab.h>
  55 #include <net/sctp/sctp.h>
  56 #include <net/ipv6.h>
  57
  58 #include "dlm_internal.h"
  59 #include "lowcomms.h"
  60 #include "midcomms.h"
  61 #include "config.h"
  62
  63 #define NEEDED_RMEM (4*1024*1024)
  64 #define CONN_HASH_SIZE 32
  65
  66 /* Number of messages to send before rescheduling */
  67 #define MAX_SEND_MSG_COUNT 25
  68
  69 struct cbuf {
  70         unsigned int base;
  71         unsigned int len;
  72         unsigned int mask;
  73 };
  74
  75 static void cbuf_add(struct cbuf *cb, int n)
  76 {
  77         cb->len += n;
  78 }
  79
  80 static int cbuf_data(struct cbuf *cb)
  81 {
  82         return ((cb->base + cb->len) & cb->mask);
  83 }
  84
  85 static void cbuf_init(struct cbuf *cb, int size)
  86 {
  87         cb->base = cb->len = 0;
  88         cb->mask = size-1;
  89 }
  90
  91 static void cbuf_eat(struct cbuf *cb, int n)
  92 {
  93         cb->len  -= n;
  94         cb->base += n;
  95         cb->base &= cb->mask;
  96 }
  97
  98 static bool cbuf_empty(struct cbuf *cb)
  99 {
 100         return cb->len == 0;
 101 }
 102
 103 struct connection {
 104         struct socket *sock;    /* NULL if not connected */
 105         uint32_t nodeid;        /* So we know who we are in the list */
 106         struct mutex sock_mutex;
 107         unsigned long flags;
 108 #define CF_READ_PENDING 1
 109 #define CF_WRITE_PENDING 2
 110 #define CF_INIT_PENDING 4
 111 #define CF_IS_OTHERCON 5
 112 #define CF_CLOSE 6
 113 #define CF_APP_LIMITED 7
 114 #define CF_CLOSING 8
 115         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 116         spinlock_t writequeue_lock;
 117         int (*rx_action) (struct connection *); /* What to do when active */
 118         void (*connect_action) (struct connection *);   /* What to do to connect */
 119         struct page *rx_page;
 120         struct cbuf cb;
 121         int retries;
 122 #define MAX_CONNECT_RETRIES 3
 123         struct hlist_node list;
 124         struct connection *othercon;
 125         struct work_struct rwork; /* Receive workqueue */
 126         struct work_struct swork; /* Send workqueue */
 127 };
 128 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 129
 130 /* An entry waiting to be sent */
 131 struct writequeue_entry {
 132         struct list_head list;
 133         struct page *page;
 134         int offset;
 135         int len;
 136         int end;
 137         int users;
 138         struct connection *con;
 139 };
 140
 141 struct dlm_node_addr {
 142         struct list_head list;
 143         int nodeid;
 144         int addr_count;
 145         int curr_addr_index;
 146         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 147 };
 148
 149 static struct listen_sock_callbacks {
 150         void (*sk_error_report)(struct sock *);
 151         void (*sk_data_ready)(struct sock *);
 152         void (*sk_state_change)(struct sock *);
 153         void (*sk_write_space)(struct sock *);
 154 } listen_sock;
 155
 156 static LIST_HEAD(dlm_node_addrs);
 157 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 158
 159 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 160 static int dlm_local_count;
 161 static int dlm_allow_conn;
 162
 163 /* Work queues */
 164 static struct workqueue_struct *recv_workqueue;
 165 static struct workqueue_struct *send_workqueue;
 166
 167 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 168 static DEFINE_MUTEX(connections_lock);
 169 static struct kmem_cache *con_cache;
 170
 171 static void process_recv_sockets(struct work_struct *work);
 172 static void process_send_sockets(struct work_struct *work);
 173
 174
 175 /* This is deliberately very simple because most clusters have simple
 176    sequential nodeids, so we should be able to go straight to a connection
 177    struct in the array */
 178 static inline int nodeid_hash(int nodeid)
 179 {
 180         return nodeid & (CONN_HASH_SIZE-1);
 181 }
 182
 183 static struct connection *__find_con(int nodeid)
 184 {
 185         int r;
 186         struct connection *con;
 187
 188         r = nodeid_hash(nodeid);
 189
 190         hlist_for_each_entry(con, &connection_hash[r], list) {
 191                 if (con->nodeid == nodeid)
 192                         return con;
 193         }
 194         return NULL;
 195 }
 196
 197 /*
 198  * If 'allocation' is zero then we don't attempt to create a new
 199  * connection structure for this node.
 200  */
 201 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 202 {
 203         struct connection *con = NULL;
 204         int r;
 205
 206         con = __find_con(nodeid);
 207         if (con || !alloc)
 208                 return con;
 209
 210         con = kmem_cache_zalloc(con_cache, alloc);
 211         if (!con)
 212                 return NULL;
 213
 214         r = nodeid_hash(nodeid);
 215         hlist_add_head(&con->list, &connection_hash[r]);
 216
 217         con->nodeid = nodeid;
 218         mutex_init(&con->sock_mutex);
 219         INIT_LIST_HEAD(&con->writequeue);
 220         spin_lock_init(&con->writequeue_lock);
 221         INIT_WORK(&con->swork, process_send_sockets);
 222         INIT_WORK(&con->rwork, process_recv_sockets);
 223
 224         /* Setup action pointers for child sockets */
 225         if (con->nodeid) {
 226                 struct connection *zerocon = __find_con(0);
 227
 228                 con->connect_action = zerocon->connect_action;
 229                 if (!con->rx_action)
 230                         con->rx_action = zerocon->rx_action;
 231         }
 232
 233         return con;
 234 }
 235
 236 /* Loop round all connections */
 237 static void foreach_conn(void (*conn_func)(struct connection *c))
 238 {
 239         int i;
 240         struct hlist_node *n;
 241         struct connection *con;
 242
 243         for (i = 0; i < CONN_HASH_SIZE; i++) {
 244                 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
 245                         conn_func(con);
 246         }
 247 }
 248
 249 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 250 {
 251         struct connection *con;
 252
 253         mutex_lock(&connections_lock);
 254         con = __nodeid2con(nodeid, allocation);
 255         mutex_unlock(&connections_lock);
 256
 257         return con;
 258 }
 259
 260 static struct dlm_node_addr *find_node_addr(int nodeid)
 261 {
 262         struct dlm_node_addr *na;
 263
 264         list_for_each_entry(na, &dlm_node_addrs, list) {
 265                 if (na->nodeid == nodeid)
 266                         return na;
 267         }
 268         return NULL;
 269 }
 270
 271 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
 272 {
 273         switch (x->ss_family) {
 274         case AF_INET: {
 275                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 276                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 277                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 278                         return 0;
 279                 if (sinx->sin_port != siny->sin_port)
 280                         return 0;
 281                 break;
 282         }
 283         case AF_INET6: {
 284                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 285                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 286                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 287                         return 0;
 288                 if (sinx->sin6_port != siny->sin6_port)
 289                         return 0;
 290                 break;
 291         }
 292         default:
 293                 return 0;
 294         }
 295         return 1;
 296 }
 297
 298 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 299                           struct sockaddr *sa_out, bool try_new_addr)
 300 {
 301         struct sockaddr_storage sas;
 302         struct dlm_node_addr *na;
 303
 304         if (!dlm_local_count)
 305                 return -1;
 306
 307         spin_lock(&dlm_node_addrs_spin);
 308         na = find_node_addr(nodeid);
 309         if (na && na->addr_count) {
 310                 memcpy(&sas, na->addr[na->curr_addr_index],
 311                        sizeof(struct sockaddr_storage));
 312
 313                 if (try_new_addr) {
 314                         na->curr_addr_index++;
 315                         if (na->curr_addr_index == na->addr_count)
 316                                 na->curr_addr_index = 0;
 317                 }
 318         }
 319         spin_unlock(&dlm_node_addrs_spin);
 320
 321         if (!na)
 322                 return -EEXIST;
 323
 324         if (!na->addr_count)
 325                 return -ENOENT;
 326
 327         if (sas_out)
 328                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 329
 330         if (!sa_out)
 331                 return 0;
 332
 333         if (dlm_local_addr[0]->ss_family == AF_INET) {
 334                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 335                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 336                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 337         } else {
 338                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 339                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 340                 ret6->sin6_addr = in6->sin6_addr;
 341         }
 342
 343         return 0;
 344 }
 345
 346 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
 347 {
 348         struct dlm_node_addr *na;
 349         int rv = -EEXIST;
 350         int addr_i;
 351
 352         spin_lock(&dlm_node_addrs_spin);
 353         list_for_each_entry(na, &dlm_node_addrs, list) {
 354                 if (!na->addr_count)
 355                         continue;
 356
 357                 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
 358                         if (addr_compare(na->addr[addr_i], addr)) {
 359                                 *nodeid = na->nodeid;
 360                                 rv = 0;
 361                                 goto unlock;
 362                         }
 363                 }
 364         }
 365 unlock:
 366         spin_unlock(&dlm_node_addrs_spin);
 367         return rv;
 368 }
 369
 370 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 371 {
 372         struct sockaddr_storage *new_addr;
 373         struct dlm_node_addr *new_node, *na;
 374
 375         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 376         if (!new_node)
 377                 return -ENOMEM;
 378
 379         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 380         if (!new_addr) {
 381                 kfree(new_node);
 382                 return -ENOMEM;
 383         }
 384
 385         memcpy(new_addr, addr, len);
 386
 387         spin_lock(&dlm_node_addrs_spin);
 388         na = find_node_addr(nodeid);
 389         if (!na) {
 390                 new_node->nodeid = nodeid;
 391                 new_node->addr[0] = new_addr;
 392                 new_node->addr_count = 1;
 393                 list_add(&new_node->list, &dlm_node_addrs);
 394                 spin_unlock(&dlm_node_addrs_spin);
 395                 return 0;
 396         }
 397
 398         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 399                 spin_unlock(&dlm_node_addrs_spin);
 400                 kfree(new_addr);
 401                 kfree(new_node);
 402                 return -ENOSPC;
 403         }
 404
 405         na->addr[na->addr_count++] = new_addr;
 406         spin_unlock(&dlm_node_addrs_spin);
 407         kfree(new_node);
 408         return 0;
 409 }
 410
 411 /* Data available on socket or listen socket received a connect */
 412 static void lowcomms_data_ready(struct sock *sk)
 413 {
 414         struct connection *con;
 415
 416         read_lock_bh(&sk->sk_callback_lock);
 417         con = sock2con(sk);
 418         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 419                 queue_work(recv_workqueue, &con->rwork);
 420         read_unlock_bh(&sk->sk_callback_lock);
 421 }
 422
 423 static void lowcomms_write_space(struct sock *sk)
 424 {
 425         struct connection *con;
 426
 427         read_lock_bh(&sk->sk_callback_lock);
 428         con = sock2con(sk);
 429         if (!con)
 430                 goto out;
 431
 432         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 433
 434         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 435                 con->sock->sk->sk_write_pending--;
 436                 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
 437         }
 438
 439         queue_work(send_workqueue, &con->swork);
 440 out:
 441         read_unlock_bh(&sk->sk_callback_lock);
 442 }
 443
 444 static inline void lowcomms_connect_sock(struct connection *con)
 445 {
 446         if (test_bit(CF_CLOSE, &con->flags))
 447                 return;
 448         queue_work(send_workqueue, &con->swork);
 449         cond_resched();
 450 }
 451
 452 static void lowcomms_state_change(struct sock *sk)
 453 {
 454         /* SCTP layer is not calling sk_data_ready when the connection
 455          * is done, so we catch the signal through here. Also, it
 456          * doesn't switch socket state when entering shutdown, so we
 457          * skip the write in that case.
 458          */
 459         if (sk->sk_shutdown) {
 460                 if (sk->sk_shutdown == RCV_SHUTDOWN)
 461                         lowcomms_data_ready(sk);
 462         } else if (sk->sk_state == TCP_ESTABLISHED) {
 463                 lowcomms_write_space(sk);
 464         }
 465 }
 466
 467 int dlm_lowcomms_connect_node(int nodeid)
 468 {
 469         struct connection *con;
 470
 471         if (nodeid == dlm_our_nodeid())
 472                 return 0;
 473
 474         con = nodeid2con(nodeid, GFP_NOFS);
 475         if (!con)
 476                 return -ENOMEM;
 477         lowcomms_connect_sock(con);
 478         return 0;
 479 }
 480
 481 static void lowcomms_error_report(struct sock *sk)
 482 {
 483         struct connection *con;
 484         struct sockaddr_storage saddr;
 485         void (*orig_report)(struct sock *) = NULL;
 486
 487         read_lock_bh(&sk->sk_callback_lock);
 488         con = sock2con(sk);
 489         if (con == NULL)
 490                 goto out;
 491
 492         orig_report = listen_sock.sk_error_report;
 493         if (con->sock == NULL ||
 494             kernel_getpeername(con->sock, (struct sockaddr *)&saddr) < 0) {
 495                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 496                                    "sending to node %d, port %d, "
 497                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 498                                    con->nodeid, dlm_config.ci_tcp_port,
 499                                    sk->sk_err, sk->sk_err_soft);
 500         } else if (saddr.ss_family == AF_INET) {
 501                 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
 502
 503                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 504                                    "sending to node %d at %pI4, port %d, "
 505                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 506                                    con->nodeid, &sin4->sin_addr.s_addr,
 507                                    dlm_config.ci_tcp_port, sk->sk_err,
 508                                    sk->sk_err_soft);
 509         } else {
 510                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
 511
 512                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 513                                    "sending to node %d at %u.%u.%u.%u, "
 514                                    "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 515                                    con->nodeid, sin6->sin6_addr.s6_addr32[0],
 516                                    sin6->sin6_addr.s6_addr32[1],
 517                                    sin6->sin6_addr.s6_addr32[2],
 518                                    sin6->sin6_addr.s6_addr32[3],
 519                                    dlm_config.ci_tcp_port, sk->sk_err,
 520                                    sk->sk_err_soft);
 521         }
 522 out:
 523         read_unlock_bh(&sk->sk_callback_lock);
 524         if (orig_report)
 525                 orig_report(sk);
 526 }
 527
 528 /* Note: sk_callback_lock must be locked before calling this function. */
 529 static void save_listen_callbacks(struct socket *sock)
 530 {
 531         struct sock *sk = sock->sk;
 532
 533         listen_sock.sk_data_ready = sk->sk_data_ready;
 534         listen_sock.sk_state_change = sk->sk_state_change;
 535         listen_sock.sk_write_space = sk->sk_write_space;
 536         listen_sock.sk_error_report = sk->sk_error_report;
 537 }
 538
 539 static void restore_callbacks(struct socket *sock)
 540 {
 541         struct sock *sk = sock->sk;
 542
 543         write_lock_bh(&sk->sk_callback_lock);
 544         sk->sk_user_data = NULL;
 545         sk->sk_data_ready = listen_sock.sk_data_ready;
 546         sk->sk_state_change = listen_sock.sk_state_change;
 547         sk->sk_write_space = listen_sock.sk_write_space;
 548         sk->sk_error_report = listen_sock.sk_error_report;
 549         write_unlock_bh(&sk->sk_callback_lock);
 550 }
 551
 552 /* Make a socket active */
 553 static void add_sock(struct socket *sock, struct connection *con)
 554 {
 555         struct sock *sk = sock->sk;
 556
 557         write_lock_bh(&sk->sk_callback_lock);
 558         con->sock = sock;
 559
 560         sk->sk_user_data = con;
 561         /* Install a data_ready callback */
 562         sk->sk_data_ready = lowcomms_data_ready;
 563         sk->sk_write_space = lowcomms_write_space;
 564         sk->sk_state_change = lowcomms_state_change;
 565         sk->sk_allocation = GFP_NOFS;
 566         sk->sk_error_report = lowcomms_error_report;
 567         write_unlock_bh(&sk->sk_callback_lock);
 568 }
 569
 570 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 571    length */
 572 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 573                           int *addr_len)
 574 {
 575         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 576         if (saddr->ss_family == AF_INET) {
 577                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 578                 in4_addr->sin_port = cpu_to_be16(port);
 579                 *addr_len = sizeof(struct sockaddr_in);
 580                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 581         } else {
 582                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 583                 in6_addr->sin6_port = cpu_to_be16(port);
 584                 *addr_len = sizeof(struct sockaddr_in6);
 585         }
 586         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 587 }
 588
 589 /* Close a remote connection and tidy up */
 590 static void close_connection(struct connection *con, bool and_other,
 591                              bool tx, bool rx)
 592 {
 593         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 594
 595         if (tx && !closing && cancel_work_sync(&con->swork)) {
 596                 log_print("canceled swork for node %d", con->nodeid);
 597                 clear_bit(CF_WRITE_PENDING, &con->flags);
 598         }
 599         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 600                 log_print("canceled rwork for node %d", con->nodeid);
 601                 clear_bit(CF_READ_PENDING, &con->flags);
 602         }
 603
 604         mutex_lock(&con->sock_mutex);
 605         if (con->sock) {
 606                 restore_callbacks(con->sock);
 607                 sock_release(con->sock);
 608                 con->sock = NULL;
 609         }
 610         if (con->othercon && and_other) {
 611                 /* Will only re-enter once. */
 612                 close_connection(con->othercon, false, true, true);
 613         }
 614         if (con->rx_page) {
 615                 __free_page(con->rx_page);
 616                 con->rx_page = NULL;
 617         }
 618
 619         con->retries = 0;
 620         mutex_unlock(&con->sock_mutex);
 621         clear_bit(CF_CLOSING, &con->flags);
 622 }
 623
 624 /* Data received from remote end */
 625 static int receive_from_sock(struct connection *con)
 626 {
 627         int ret = 0;
 628         struct msghdr msg = {};
 629         struct kvec iov[2];
 630         unsigned len;
 631         int r;
 632         int call_again_soon = 0;
 633         int nvec;
 634
 635         mutex_lock(&con->sock_mutex);
 636
 637         if (con->sock == NULL) {
 638                 ret = -EAGAIN;
 639                 goto out_close;
 640         }
 641         if (con->nodeid == 0) {
 642                 ret = -EINVAL;
 643                 goto out_close;
 644         }
 645
 646         if (con->rx_page == NULL) {
 647                 /*
 648                  * This doesn't need to be atomic, but I think it should
 649                  * improve performance if it is.
 650                  */
 651                 con->rx_page = alloc_page(GFP_ATOMIC);
 652                 if (con->rx_page == NULL)
 653                         goto out_resched;
 654                 cbuf_init(&con->cb, PAGE_SIZE);
 655         }
 656
 657         /*
 658          * iov[0] is the bit of the circular buffer between the current end
 659          * point (cb.base + cb.len) and the end of the buffer.
 660          */
 661         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 662         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 663         iov[1].iov_len = 0;
 664         nvec = 1;
 665
 666         /*
 667          * iov[1] is the bit of the circular buffer between the start of the
 668          * buffer and the start of the currently used section (cb.base)
 669          */
 670         if (cbuf_data(&con->cb) >= con->cb.base) {
 671                 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb);
 672                 iov[1].iov_len = con->cb.base;
 673                 iov[1].iov_base = page_address(con->rx_page);
 674                 nvec = 2;
 675         }
 676         len = iov[0].iov_len + iov[1].iov_len;
 677         iov_iter_kvec(&msg.msg_iter, READ, iov, nvec, len);
 678
 679         r = ret = sock_recvmsg(con->sock, &msg, MSG_DONTWAIT | MSG_NOSIGNAL);
 680         if (ret <= 0)
 681                 goto out_close;
 682         else if (ret == len)
 683                 call_again_soon = 1;
 684
 685         cbuf_add(&con->cb, ret);
 686         ret = dlm_process_incoming_buffer(con->nodeid,
 687                                           page_address(con->rx_page),
 688                                           con->cb.base, con->cb.len,
 689                                           PAGE_SIZE);
 690         if (ret == -EBADMSG) {
 691                 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d",
 692                           page_address(con->rx_page), con->cb.base,
 693                           con->cb.len, r);
 694         }
 695         if (ret < 0)
 696                 goto out_close;
 697         cbuf_eat(&con->cb, ret);
 698
 699         if (cbuf_empty(&con->cb) && !call_again_soon) {
 700                 __free_page(con->rx_page);
 701                 con->rx_page = NULL;
 702         }
 703
 704         if (call_again_soon)
 705                 goto out_resched;
 706         mutex_unlock(&con->sock_mutex);
 707         return 0;
 708
 709 out_resched:
 710         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 711                 queue_work(recv_workqueue, &con->rwork);
 712         mutex_unlock(&con->sock_mutex);
 713         return -EAGAIN;
 714
 715 out_close:
 716         mutex_unlock(&con->sock_mutex);
 717         if (ret != -EAGAIN) {
 718                 close_connection(con, true, true, false);
 719                 /* Reconnect when there is something to send */
 720         }
 721         /* Don't return success if we really got EOF */
 722         if (ret == 0)
 723                 ret = -EAGAIN;
 724
 725         return ret;
 726 }
 727
 728 /* Listening socket is busy, accept a connection */
 729 static int tcp_accept_from_sock(struct connection *con)
 730 {
 731         int result;
 732         struct sockaddr_storage peeraddr;
 733         struct socket *newsock;
 734         int len;
 735         int nodeid;
 736         struct connection *newcon;
 737         struct connection *addcon;
 738
 739         mutex_lock(&connections_lock);
 740         if (!dlm_allow_conn) {
 741                 mutex_unlock(&connections_lock);
 742                 return -1;
 743         }
 744         mutex_unlock(&connections_lock);
 745
 746         mutex_lock_nested(&con->sock_mutex, 0);
 747
 748         if (!con->sock) {
 749                 mutex_unlock(&con->sock_mutex);
 750                 return -ENOTCONN;
 751         }
 752
 753         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 754         if (result < 0)
 755                 goto accept_err;
 756
 757         /* Get the connected socket's peer */
 758         memset(&peeraddr, 0, sizeof(peeraddr));
 759         len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
 760         if (len < 0) {
 761                 result = -ECONNABORTED;
 762                 goto accept_err;
 763         }
 764
 765         /* Get the new node's NODEID */
 766         make_sockaddr(&peeraddr, 0, &len);
 767         if (addr_to_nodeid(&peeraddr, &nodeid)) {
 768                 unsigned char *b=(unsigned char *)&peeraddr;
 769                 log_print("connect from non cluster node");
 770                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 771                                      b, sizeof(struct sockaddr_storage));
 772                 sock_release(newsock);
 773                 mutex_unlock(&con->sock_mutex);
 774                 return -1;
 775         }
 776
 777         log_print("got connection from %d", nodeid);
 778
 779         /*  Check to see if we already have a connection to this node. This
 780          *  could happen if the two nodes initiate a connection at roughly
 781          *  the same time and the connections cross on the wire.
 782          *  In this case we store the incoming one in "othercon"
 783          */
 784         newcon = nodeid2con(nodeid, GFP_NOFS);
 785         if (!newcon) {
 786                 result = -ENOMEM;
 787                 goto accept_err;
 788         }
 789         mutex_lock_nested(&newcon->sock_mutex, 1);
 790         if (newcon->sock) {
 791                 struct connection *othercon = newcon->othercon;
 792
 793                 if (!othercon) {
 794                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 795                         if (!othercon) {
 796                                 log_print("failed to allocate incoming socket");
 797                                 mutex_unlock(&newcon->sock_mutex);
 798                                 result = -ENOMEM;
 799                                 goto accept_err;
 800                         }
 801                         othercon->nodeid = nodeid;
 802                         othercon->rx_action = receive_from_sock;
 803                         mutex_init(&othercon->sock_mutex);
 804                         INIT_LIST_HEAD(&othercon->writequeue);
 805                         spin_lock_init(&othercon->writequeue_lock);
 806                         INIT_WORK(&othercon->swork, process_send_sockets);
 807                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 808                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 809                 }
 810                 mutex_lock_nested(&othercon->sock_mutex, 2);
 811                 if (!othercon->sock) {
 812                         newcon->othercon = othercon;
 813                         add_sock(newsock, othercon);
 814                         addcon = othercon;
 815                         mutex_unlock(&othercon->sock_mutex);
 816                 }
 817                 else {
 818                         printk("Extra connection from node %d attempted\n", nodeid);
 819                         result = -EAGAIN;
 820                         mutex_unlock(&othercon->sock_mutex);
 821                         mutex_unlock(&newcon->sock_mutex);
 822                         goto accept_err;
 823                 }
 824         }
 825         else {
 826                 newcon->rx_action = receive_from_sock;
 827                 /* accept copies the sk after we've saved the callbacks, so we
 828                    don't want to save them a second time or comm errors will
 829                    result in calling sk_error_report recursively. */
 830                 add_sock(newsock, newcon);
 831                 addcon = newcon;
 832         }
 833
 834         mutex_unlock(&newcon->sock_mutex);
 835
 836         /*
 837          * Add it to the active queue in case we got data
 838          * between processing the accept adding the socket
 839          * to the read_sockets list
 840          */
 841         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 842                 queue_work(recv_workqueue, &addcon->rwork);
 843         mutex_unlock(&con->sock_mutex);
 844
 845         return 0;
 846
 847 accept_err:
 848         mutex_unlock(&con->sock_mutex);
 849         if (newsock)
 850                 sock_release(newsock);
 851
 852         if (result != -EAGAIN)
 853                 log_print("error accepting connection from node: %d", result);
 854         return result;
 855 }
 856
 857 static int sctp_accept_from_sock(struct connection *con)
 858 {
 859         /* Check that the new node is in the lockspace */
 860         struct sctp_prim prim;
 861         int nodeid;
 862         int prim_len, ret;
 863         int addr_len;
 864         struct connection *newcon;
 865         struct connection *addcon;
 866         struct socket *newsock;
 867
 868         mutex_lock(&connections_lock);
 869         if (!dlm_allow_conn) {
 870                 mutex_unlock(&connections_lock);
 871                 return -1;
 872         }
 873         mutex_unlock(&connections_lock);
 874
 875         mutex_lock_nested(&con->sock_mutex, 0);
 876
 877         ret = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 878         if (ret < 0)
 879                 goto accept_err;
 880
 881         memset(&prim, 0, sizeof(struct sctp_prim));
 882         prim_len = sizeof(struct sctp_prim);
 883
 884         ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR,
 885                                 (char *)&prim, &prim_len);
 886         if (ret < 0) {
 887                 log_print("getsockopt/sctp_primary_addr failed: %d", ret);
 888                 goto accept_err;
 889         }
 890
 891         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 892         ret = addr_to_nodeid(&prim.ssp_addr, &nodeid);
 893         if (ret) {
 894                 unsigned char *b = (unsigned char *)&prim.ssp_addr;
 895
 896                 log_print("reject connect from unknown addr");
 897                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 898                                      b, sizeof(struct sockaddr_storage));
 899                 goto accept_err;
 900         }
 901
 902         newcon = nodeid2con(nodeid, GFP_NOFS);
 903         if (!newcon) {
 904                 ret = -ENOMEM;
 905                 goto accept_err;
 906         }
 907
 908         mutex_lock_nested(&newcon->sock_mutex, 1);
 909
 910         if (newcon->sock) {
 911                 struct connection *othercon = newcon->othercon;
 912
 913                 if (!othercon) {
 914                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 915                         if (!othercon) {
 916                                 log_print("failed to allocate incoming socket");
 917                                 mutex_unlock(&newcon->sock_mutex);
 918                                 ret = -ENOMEM;
 919                                 goto accept_err;
 920                         }
 921                         othercon->nodeid = nodeid;
 922                         othercon->rx_action = receive_from_sock;
 923                         mutex_init(&othercon->sock_mutex);
 924                         INIT_LIST_HEAD(&othercon->writequeue);
 925                         spin_lock_init(&othercon->writequeue_lock);
 926                         INIT_WORK(&othercon->swork, process_send_sockets);
 927                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 928                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 929                 }
 930                 mutex_lock_nested(&othercon->sock_mutex, 2);
 931                 if (!othercon->sock) {
 932                         newcon->othercon = othercon;
 933                         add_sock(newsock, othercon);
 934                         addcon = othercon;
 935                         mutex_unlock(&othercon->sock_mutex);
 936                 } else {
 937                         printk("Extra connection from node %d attempted\n", nodeid);
 938                         ret = -EAGAIN;
 939                         mutex_unlock(&othercon->sock_mutex);
 940                         mutex_unlock(&newcon->sock_mutex);
 941                         goto accept_err;
 942                 }
 943         } else {
 944                 newcon->rx_action = receive_from_sock;
 945                 add_sock(newsock, newcon);
 946                 addcon = newcon;
 947         }
 948
 949         log_print("connected to %d", nodeid);
 950
 951         mutex_unlock(&newcon->sock_mutex);
 952
 953         /*
 954          * Add it to the active queue in case we got data
 955          * between processing the accept adding the socket
 956          * to the read_sockets list
 957          */
 958         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 959                 queue_work(recv_workqueue, &addcon->rwork);
 960         mutex_unlock(&con->sock_mutex);
 961
 962         return 0;
 963
 964 accept_err:
 965         mutex_unlock(&con->sock_mutex);
 966         if (newsock)
 967                 sock_release(newsock);
 968         if (ret != -EAGAIN)
 969                 log_print("error accepting connection from node: %d", ret);
 970
 971         return ret;
 972 }
 973
 974 static void free_entry(struct writequeue_entry *e)
 975 {
 976         __free_page(e->page);
 977         kfree(e);
 978 }
 979
 980 /*
 981  * writequeue_entry_complete - try to delete and free write queue entry
 982  * @e: write queue entry to try to delete
 983  * @completed: bytes completed
 984  *
 985  * writequeue_lock must be held.
 986  */
 987 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 988 {
 989         e->offset += completed;
 990         e->len -= completed;
 991
 992         if (e->len == 0 && e->users == 0) {
 993                 list_del(&e->list);
 994                 free_entry(e);
 995         }
 996 }
 997
 998 /*
 999  * sctp_bind_addrs - bind a SCTP socket to all our addresses
1000  */
1001 static int sctp_bind_addrs(struct connection *con, uint16_t port)
1002 {
1003         struct sockaddr_storage localaddr;
1004         int i, addr_len, result = 0;
1005
1006         for (i = 0; i < dlm_local_count; i++) {
1007                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1008                 make_sockaddr(&localaddr, port, &addr_len);
1009
1010                 if (!i)
1011                         result = kernel_bind(con->sock,
1012                                              (struct sockaddr *)&localaddr,
1013                                              addr_len);
1014                 else
1015                         result = kernel_setsockopt(con->sock, SOL_SCTP,
1016                                                    SCTP_SOCKOPT_BINDX_ADD,
1017                                                    (char *)&localaddr, addr_len);
1018
1019                 if (result < 0) {
1020                         log_print("Can't bind to %d addr number %d, %d.\n",
1021                                   port, i + 1, result);
1022                         break;
1023                 }
1024         }
1025         return result;
1026 }
1027
1028 /* Initiate an SCTP association.
1029    This is a special case of send_to_sock() in that we don't yet have a
1030    peeled-off socket for this association, so we use the listening socket
1031    and add the primary IP address of the remote node.
1032  */
1033 static void sctp_connect_to_sock(struct connection *con)
1034 {
1035         struct sockaddr_storage daddr;
1036         int one = 1;
1037         int result;
1038         int addr_len;
1039         struct socket *sock;
1040         struct timeval tv = { .tv_sec = 5, .tv_usec = 0 };
1041
1042         if (con->nodeid == 0) {
1043                 log_print("attempt to connect sock 0 foiled");
1044                 return;
1045         }
1046
1047         mutex_lock(&con->sock_mutex);
1048
1049         /* Some odd races can cause double-connects, ignore them */
1050         if (con->retries++ > MAX_CONNECT_RETRIES)
1051                 goto out;
1052
1053         if (con->sock) {
1054                 log_print("node %d already connected.", con->nodeid);
1055                 goto out;
1056         }
1057
1058         memset(&daddr, 0, sizeof(daddr));
1059         result = nodeid_to_addr(con->nodeid, &daddr, NULL, true);
1060         if (result < 0) {
1061                 log_print("no address for nodeid %d", con->nodeid);
1062                 goto out;
1063         }
1064
1065         /* Create a socket to communicate with */
1066         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1067                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1068         if (result < 0)
1069                 goto socket_err;
1070
1071         con->rx_action = receive_from_sock;
1072         con->connect_action = sctp_connect_to_sock;
1073         add_sock(sock, con);
1074
1075         /* Bind to all addresses. */
1076         if (sctp_bind_addrs(con, 0))
1077                 goto bind_err;
1078
1079         make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1080
1081         log_print("connecting to %d", con->nodeid);
1082
1083         /* Turn off Nagle's algorithm */
1084         kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1085                           sizeof(one));
1086
1087         /*
1088          * Make sock->ops->connect() function return in specified time,
1089          * since O_NONBLOCK argument in connect() function does not work here,
1090          * then, we should restore the default value of this attribute.
1091          */
1092         kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
1093                           sizeof(tv));
1094         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1095                                    0);
1096         memset(&tv, 0, sizeof(tv));
1097         kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
1098                           sizeof(tv));
1099
1100         if (result == -EINPROGRESS)
1101                 result = 0;
1102         if (result == 0)
1103                 goto out;
1104
1105 bind_err:
1106         con->sock = NULL;
1107         sock_release(sock);
1108
1109 socket_err:
1110         /*
1111          * Some errors are fatal and this list might need adjusting. For other
1112          * errors we try again until the max number of retries is reached.
1113          */
1114         if (result != -EHOSTUNREACH &&
1115             result != -ENETUNREACH &&
1116             result != -ENETDOWN &&
1117             result != -EINVAL &&
1118             result != -EPROTONOSUPPORT) {
1119                 log_print("connect %d try %d error %d", con->nodeid,
1120                           con->retries, result);
1121                 mutex_unlock(&con->sock_mutex);
1122                 msleep(1000);
1123                 lowcomms_connect_sock(con);
1124                 return;
1125         }
1126
1127 out:
1128         mutex_unlock(&con->sock_mutex);
1129 }
1130
1131 /* Connect a new socket to its peer */
1132 static void tcp_connect_to_sock(struct connection *con)
1133 {
1134         struct sockaddr_storage saddr, src_addr;
1135         int addr_len;
1136         struct socket *sock = NULL;
1137         int one = 1;
1138         int result;
1139
1140         if (con->nodeid == 0) {
1141                 log_print("attempt to connect sock 0 foiled");
1142                 return;
1143         }
1144
1145         mutex_lock(&con->sock_mutex);
1146         if (con->retries++ > MAX_CONNECT_RETRIES)
1147                 goto out;
1148
1149         /* Some odd races can cause double-connects, ignore them */
1150         if (con->sock)
1151                 goto out;
1152
1153         /* Create a socket to communicate with */
1154         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1155                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1156         if (result < 0)
1157                 goto out_err;
1158
1159         memset(&saddr, 0, sizeof(saddr));
1160         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1161         if (result < 0) {
1162                 log_print("no address for nodeid %d", con->nodeid);
1163                 goto out_err;
1164         }
1165
1166         con->rx_action = receive_from_sock;
1167         con->connect_action = tcp_connect_to_sock;
1168         add_sock(sock, con);
1169
1170         /* Bind to our cluster-known address connecting to avoid
1171            routing problems */
1172         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1173         make_sockaddr(&src_addr, 0, &addr_len);
1174         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1175                                  addr_len);
1176         if (result < 0) {
1177                 log_print("could not bind for connect: %d", result);
1178                 /* This *may* not indicate a critical error */
1179         }
1180
1181         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1182
1183         log_print("connecting to %d", con->nodeid);
1184
1185         /* Turn off Nagle's algorithm */
1186         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1187                           sizeof(one));
1188
1189         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1190                                    O_NONBLOCK);
1191         if (result == -EINPROGRESS)
1192                 result = 0;
1193         if (result == 0)
1194                 goto out;
1195
1196 out_err:
1197         if (con->sock) {
1198                 sock_release(con->sock);
1199                 con->sock = NULL;
1200         } else if (sock) {
1201                 sock_release(sock);
1202         }
1203         /*
1204          * Some errors are fatal and this list might need adjusting. For other
1205          * errors we try again until the max number of retries is reached.
1206          */
1207         if (result != -EHOSTUNREACH &&
1208             result != -ENETUNREACH &&
1209             result != -ENETDOWN &&
1210             result != -EINVAL &&
1211             result != -EPROTONOSUPPORT) {
1212                 log_print("connect %d try %d error %d", con->nodeid,
1213                           con->retries, result);
1214                 mutex_unlock(&con->sock_mutex);
1215                 msleep(1000);
1216                 lowcomms_connect_sock(con);
1217                 return;
1218         }
1219 out:
1220         mutex_unlock(&con->sock_mutex);
1221         return;
1222 }
1223
1224 static struct socket *tcp_create_listen_sock(struct connection *con,
1225                                              struct sockaddr_storage *saddr)
1226 {
1227         struct socket *sock = NULL;
1228         int result = 0;
1229         int one = 1;
1230         int addr_len;
1231
1232         if (dlm_local_addr[0]->ss_family == AF_INET)
1233                 addr_len = sizeof(struct sockaddr_in);
1234         else
1235                 addr_len = sizeof(struct sockaddr_in6);
1236
1237         /* Create a socket to communicate with */
1238         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1239                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1240         if (result < 0) {
1241                 log_print("Can't create listening comms socket");
1242                 goto create_out;
1243         }
1244
1245         /* Turn off Nagle's algorithm */
1246         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1247                           sizeof(one));
1248
1249         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1250                                    (char *)&one, sizeof(one));
1251
1252         if (result < 0) {
1253                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1254         }
1255         write_lock_bh(&sock->sk->sk_callback_lock);
1256         sock->sk->sk_user_data = con;
1257         save_listen_callbacks(sock);
1258         con->rx_action = tcp_accept_from_sock;
1259         con->connect_action = tcp_connect_to_sock;
1260         write_unlock_bh(&sock->sk->sk_callback_lock);
1261
1262         /* Bind to our port */
1263         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1264         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1265         if (result < 0) {
1266                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1267                 sock_release(sock);
1268                 sock = NULL;
1269                 con->sock = NULL;
1270                 goto create_out;
1271         }
1272         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1273                                  (char *)&one, sizeof(one));
1274         if (result < 0) {
1275                 log_print("Set keepalive failed: %d", result);
1276         }
1277
1278         result = sock->ops->listen(sock, 5);
1279         if (result < 0) {
1280                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1281                 sock_release(sock);
1282                 sock = NULL;
1283                 goto create_out;
1284         }
1285
1286 create_out:
1287         return sock;
1288 }
1289
1290 /* Get local addresses */
1291 static void init_local(void)
1292 {
1293         struct sockaddr_storage sas, *addr;
1294         int i;
1295
1296         dlm_local_count = 0;
1297         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1298                 if (dlm_our_addr(&sas, i))
1299                         break;
1300
1301                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1302                 if (!addr)
1303                         break;
1304                 dlm_local_addr[dlm_local_count++] = addr;
1305         }
1306 }
1307
1308 /* Initialise SCTP socket and bind to all interfaces */
1309 static int sctp_listen_for_all(void)
1310 {
1311         struct socket *sock = NULL;
1312         int result = -EINVAL;
1313         struct connection *con = nodeid2con(0, GFP_NOFS);
1314         int bufsize = NEEDED_RMEM;
1315         int one = 1;
1316
1317         if (!con)
1318                 return -ENOMEM;
1319
1320         log_print("Using SCTP for communications");
1321
1322         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1323                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1324         if (result < 0) {
1325                 log_print("Can't create comms socket, check SCTP is loaded");
1326                 goto out;
1327         }
1328
1329         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1330                                  (char *)&bufsize, sizeof(bufsize));
1331         if (result)
1332                 log_print("Error increasing buffer space on socket %d", result);
1333
1334         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1335                                    sizeof(one));
1336         if (result < 0)
1337                 log_print("Could not set SCTP NODELAY error %d\n", result);
1338
1339         write_lock_bh(&sock->sk->sk_callback_lock);
1340         /* Init con struct */
1341         sock->sk->sk_user_data = con;
1342         save_listen_callbacks(sock);
1343         con->sock = sock;
1344         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1345         con->rx_action = sctp_accept_from_sock;
1346         con->connect_action = sctp_connect_to_sock;
1347
1348         write_unlock_bh(&sock->sk->sk_callback_lock);
1349
1350         /* Bind to all addresses. */
1351         if (sctp_bind_addrs(con, dlm_config.ci_tcp_port))
1352                 goto create_delsock;
1353
1354         result = sock->ops->listen(sock, 5);
1355         if (result < 0) {
1356                 log_print("Can't set socket listening");
1357                 goto create_delsock;
1358         }
1359
1360         return 0;
1361
1362 create_delsock:
1363         sock_release(sock);
1364         con->sock = NULL;
1365 out:
1366         return result;
1367 }
1368
1369 static int tcp_listen_for_all(void)
1370 {
1371         struct socket *sock = NULL;
1372         struct connection *con = nodeid2con(0, GFP_NOFS);
1373         int result = -EINVAL;
1374
1375         if (!con)
1376                 return -ENOMEM;
1377
1378         /* We don't support multi-homed hosts */
1379         if (dlm_local_addr[1] != NULL) {
1380                 log_print("TCP protocol can't handle multi-homed hosts, "
1381                           "try SCTP");
1382                 return -EINVAL;
1383         }
1384
1385         log_print("Using TCP for communications");
1386
1387         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1388         if (sock) {
1389                 add_sock(sock, con);
1390                 result = 0;
1391         }
1392         else {
1393                 result = -EADDRINUSE;
1394         }
1395
1396         return result;
1397 }
1398
1399
1400
1401 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1402                                                      gfp_t allocation)
1403 {
1404         struct writequeue_entry *entry;
1405
1406         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1407         if (!entry)
1408                 return NULL;
1409
1410         entry->page = alloc_page(allocation);
1411         if (!entry->page) {
1412                 kfree(entry);
1413                 return NULL;
1414         }
1415
1416         entry->offset = 0;
1417         entry->len = 0;
1418         entry->end = 0;
1419         entry->users = 0;
1420         entry->con = con;
1421
1422         return entry;
1423 }
1424
1425 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1426 {
1427         struct connection *con;
1428         struct writequeue_entry *e;
1429         int offset = 0;
1430
1431         con = nodeid2con(nodeid, allocation);
1432         if (!con)
1433                 return NULL;
1434
1435         spin_lock(&con->writequeue_lock);
1436         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1437         if ((&e->list == &con->writequeue) ||
1438             (PAGE_SIZE - e->end < len)) {
1439                 e = NULL;
1440         } else {
1441                 offset = e->end;
1442                 e->end += len;
1443                 e->users++;
1444         }
1445         spin_unlock(&con->writequeue_lock);
1446
1447         if (e) {
1448         got_one:
1449                 *ppc = page_address(e->page) + offset;
1450                 return e;
1451         }
1452
1453         e = new_writequeue_entry(con, allocation);
1454         if (e) {
1455                 spin_lock(&con->writequeue_lock);
1456                 offset = e->end;
1457                 e->end += len;
1458                 e->users++;
1459                 list_add_tail(&e->list, &con->writequeue);
1460                 spin_unlock(&con->writequeue_lock);
1461                 goto got_one;
1462         }
1463         return NULL;
1464 }
1465
1466 void dlm_lowcomms_commit_buffer(void *mh)
1467 {
1468         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1469         struct connection *con = e->con;
1470         int users;
1471
1472         spin_lock(&con->writequeue_lock);
1473         users = --e->users;
1474         if (users)
1475                 goto out;
1476         e->len = e->end - e->offset;
1477         spin_unlock(&con->writequeue_lock);
1478
1479         queue_work(send_workqueue, &con->swork);
1480         return;
1481
1482 out:
1483         spin_unlock(&con->writequeue_lock);
1484         return;
1485 }
1486
1487 /* Send a message */
1488 static void send_to_sock(struct connection *con)
1489 {
1490         int ret = 0;
1491         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1492         struct writequeue_entry *e;
1493         int len, offset;
1494         int count = 0;
1495
1496         mutex_lock(&con->sock_mutex);
1497         if (con->sock == NULL)
1498                 goto out_connect;
1499
1500         spin_lock(&con->writequeue_lock);
1501         for (;;) {
1502                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1503                                list);
1504                 if ((struct list_head *) e == &con->writequeue)
1505                         break;
1506
1507                 len = e->len;
1508                 offset = e->offset;
1509                 BUG_ON(len == 0 && e->users == 0);
1510                 spin_unlock(&con->writequeue_lock);
1511
1512                 ret = 0;
1513                 if (len) {
1514                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1515                                               msg_flags);
1516                         if (ret == -EAGAIN || ret == 0) {
1517                                 if (ret == -EAGAIN &&
1518                                     test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1519                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1520                                         /* Notify TCP that we're limited by the
1521                                          * application window size.
1522                                          */
1523                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1524                                         con->sock->sk->sk_write_pending++;
1525                                 }
1526                                 cond_resched();
1527                                 goto out;
1528                         } else if (ret < 0)
1529                                 goto send_error;
1530                 }
1531
1532                 /* Don't starve people filling buffers */
1533                 if (++count >= MAX_SEND_MSG_COUNT) {
1534                         cond_resched();
1535                         count = 0;
1536                 }
1537
1538                 spin_lock(&con->writequeue_lock);
1539                 writequeue_entry_complete(e, ret);
1540         }
1541         spin_unlock(&con->writequeue_lock);
1542 out:
1543         mutex_unlock(&con->sock_mutex);
1544         return;
1545
1546 send_error:
1547         mutex_unlock(&con->sock_mutex);
1548         close_connection(con, true, false, true);
1549         /* Requeue the send work. When the work daemon runs again, it will try
1550            a new connection, then call this function again. */
1551         queue_work(send_workqueue, &con->swork);
1552         return;
1553
1554 out_connect:
1555         mutex_unlock(&con->sock_mutex);
1556         queue_work(send_workqueue, &con->swork);
1557         cond_resched();
1558 }
1559
1560 static void clean_one_writequeue(struct connection *con)
1561 {
1562         struct writequeue_entry *e, *safe;
1563
1564         spin_lock(&con->writequeue_lock);
1565         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1566                 list_del(&e->list);
1567                 free_entry(e);
1568         }
1569         spin_unlock(&con->writequeue_lock);
1570 }
1571
1572 /* Called from recovery when it knows that a node has
1573    left the cluster */
1574 int dlm_lowcomms_close(int nodeid)
1575 {
1576         struct connection *con;
1577         struct dlm_node_addr *na;
1578
1579         log_print("closing connection to node %d", nodeid);
1580         con = nodeid2con(nodeid, 0);
1581         if (con) {
1582                 set_bit(CF_CLOSE, &con->flags);
1583                 close_connection(con, true, true, true);
1584                 clean_one_writequeue(con);
1585         }
1586
1587         spin_lock(&dlm_node_addrs_spin);
1588         na = find_node_addr(nodeid);
1589         if (na) {
1590                 list_del(&na->list);
1591                 while (na->addr_count--)
1592                         kfree(na->addr[na->addr_count]);
1593                 kfree(na);
1594         }
1595         spin_unlock(&dlm_node_addrs_spin);
1596
1597         return 0;
1598 }
1599
1600 /* Receive workqueue function */
1601 static void process_recv_sockets(struct work_struct *work)
1602 {
1603         struct connection *con = container_of(work, struct connection, rwork);
1604         int err;
1605
1606         clear_bit(CF_READ_PENDING, &con->flags);
1607         do {
1608                 err = con->rx_action(con);
1609         } while (!err);
1610 }
1611
1612 /* Send workqueue function */
1613 static void process_send_sockets(struct work_struct *work)
1614 {
1615         struct connection *con = container_of(work, struct connection, swork);
1616
1617         clear_bit(CF_WRITE_PENDING, &con->flags);
1618         if (con->sock == NULL) /* not mutex protected so check it inside too */
1619                 con->connect_action(con);
1620         if (!list_empty(&con->writequeue))
1621                 send_to_sock(con);
1622 }
1623
1624
1625 /* Discard all entries on the write queues */
1626 static void clean_writequeues(void)
1627 {
1628         foreach_conn(clean_one_writequeue);
1629 }
1630
1631 static void work_stop(void)
1632 {
1633         destroy_workqueue(recv_workqueue);
1634         destroy_workqueue(send_workqueue);
1635 }
1636
1637 static int work_start(void)
1638 {
1639         recv_workqueue = alloc_workqueue("dlm_recv",
1640                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1641         if (!recv_workqueue) {
1642                 log_print("can't start dlm_recv");
1643                 return -ENOMEM;
1644         }
1645
1646         send_workqueue = alloc_workqueue("dlm_send",
1647                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1648         if (!send_workqueue) {
1649                 log_print("can't start dlm_send");
1650                 destroy_workqueue(recv_workqueue);
1651                 return -ENOMEM;
1652         }
1653
1654         return 0;
1655 }
1656
1657 static void _stop_conn(struct connection *con, bool and_other)
1658 {
1659         mutex_lock(&con->sock_mutex);
1660         set_bit(CF_CLOSE, &con->flags);
1661         set_bit(CF_READ_PENDING, &con->flags);
1662         set_bit(CF_WRITE_PENDING, &con->flags);
1663         if (con->sock && con->sock->sk) {
1664                 write_lock_bh(&con->sock->sk->sk_callback_lock);
1665                 con->sock->sk->sk_user_data = NULL;
1666                 write_unlock_bh(&con->sock->sk->sk_callback_lock);
1667         }
1668         if (con->othercon && and_other)
1669                 _stop_conn(con->othercon, false);
1670         mutex_unlock(&con->sock_mutex);
1671 }
1672
1673 static void stop_conn(struct connection *con)
1674 {
1675         _stop_conn(con, true);
1676 }
1677
1678 static void free_conn(struct connection *con)
1679 {
1680         close_connection(con, true, true, true);
1681         if (con->othercon)
1682                 kmem_cache_free(con_cache, con->othercon);
1683         hlist_del(&con->list);
1684         kmem_cache_free(con_cache, con);
1685 }
1686
1687 static void work_flush(void)
1688 {
1689         int ok;
1690         int i;
1691         struct hlist_node *n;
1692         struct connection *con;
1693
1694         flush_workqueue(recv_workqueue);
1695         flush_workqueue(send_workqueue);
1696         do {
1697                 ok = 1;
1698                 foreach_conn(stop_conn);
1699                 flush_workqueue(recv_workqueue);
1700                 flush_workqueue(send_workqueue);
1701                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1702                         hlist_for_each_entry_safe(con, n,
1703                                                   &connection_hash[i], list) {
1704                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1705                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1706                                 if (con->othercon) {
1707                                         ok &= test_bit(CF_READ_PENDING,
1708                                                        &con->othercon->flags);
1709                                         ok &= test_bit(CF_WRITE_PENDING,
1710                                                        &con->othercon->flags);
1711                                 }
1712                         }
1713                 }
1714         } while (!ok);
1715 }
1716
1717 void dlm_lowcomms_stop(void)
1718 {
1719         /* Set all the flags to prevent any
1720            socket activity.
1721         */
1722         mutex_lock(&connections_lock);
1723         dlm_allow_conn = 0;
1724         mutex_unlock(&connections_lock);
1725         work_flush();
1726         clean_writequeues();
1727         foreach_conn(free_conn);
1728         work_stop();
1729
1730         kmem_cache_destroy(con_cache);
1731 }
1732
1733 int dlm_lowcomms_start(void)
1734 {
1735         int error = -EINVAL;
1736         struct connection *con;
1737         int i;
1738
1739         for (i = 0; i < CONN_HASH_SIZE; i++)
1740                 INIT_HLIST_HEAD(&connection_hash[i]);
1741
1742         init_local();
1743         if (!dlm_local_count) {
1744                 error = -ENOTCONN;
1745                 log_print("no local IP address has been set");
1746                 goto fail;
1747         }
1748
1749         error = -ENOMEM;
1750         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1751                                       __alignof__(struct connection), 0,
1752                                       NULL);
1753         if (!con_cache)
1754                 goto fail;
1755
1756         error = work_start();
1757         if (error)
1758                 goto fail_destroy;
1759
1760         dlm_allow_conn = 1;
1761
1762         /* Start listening */
1763         if (dlm_config.ci_protocol == 0)
1764                 error = tcp_listen_for_all();
1765         else
1766                 error = sctp_listen_for_all();
1767         if (error)
1768                 goto fail_unlisten;
1769
1770         return 0;
1771
1772 fail_unlisten:
1773         dlm_allow_conn = 0;
1774         con = nodeid2con(0,0);
1775         if (con) {
1776                 close_connection(con, false, true, true);
1777                 kmem_cache_free(con_cache, con);
1778         }
1779 fail_destroy:
1780         kmem_cache_destroy(con_cache);
1781 fail:
1782         return error;
1783 }
1784
1785 void dlm_lowcomms_exit(void)
1786 {
1787         struct dlm_node_addr *na, *safe;
1788
1789         spin_lock(&dlm_node_addrs_spin);
1790         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1791                 list_del(&na->list);
1792                 while (na->addr_count--)
1793                         kfree(na->addr[na->addr_count]);
1794                 kfree(na);
1795         }
1796         spin_unlock(&dlm_node_addrs_spin);
1797 }