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         int buflen;
 486         void (*orig_report)(struct sock *) = NULL;
 487
 488         read_lock_bh(&sk->sk_callback_lock);
 489         con = sock2con(sk);
 490         if (con == NULL)
 491                 goto out;
 492
 493         orig_report = listen_sock.sk_error_report;
 494         if (con->sock == NULL ||
 495             kernel_getpeername(con->sock, (struct sockaddr *)&saddr, &buflen)) {
 496                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 497                                    "sending to node %d, port %d, "
 498                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 499                                    con->nodeid, dlm_config.ci_tcp_port,
 500                                    sk->sk_err, sk->sk_err_soft);
 501         } else if (saddr.ss_family == AF_INET) {
 502                 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
 503
 504                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 505                                    "sending to node %d at %pI4, port %d, "
 506                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 507                                    con->nodeid, &sin4->sin_addr.s_addr,
 508                                    dlm_config.ci_tcp_port, sk->sk_err,
 509                                    sk->sk_err_soft);
 510         } else {
 511                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
 512
 513                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 514                                    "sending to node %d at %u.%u.%u.%u, "
 515                                    "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 516                                    con->nodeid, sin6->sin6_addr.s6_addr32[0],
 517                                    sin6->sin6_addr.s6_addr32[1],
 518                                    sin6->sin6_addr.s6_addr32[2],
 519                                    sin6->sin6_addr.s6_addr32[3],
 520                                    dlm_config.ci_tcp_port, sk->sk_err,
 521                                    sk->sk_err_soft);
 522         }
 523 out:
 524         read_unlock_bh(&sk->sk_callback_lock);
 525         if (orig_report)
 526                 orig_report(sk);
 527 }
 528
 529 /* Note: sk_callback_lock must be locked before calling this function. */
 530 static void save_listen_callbacks(struct socket *sock)
 531 {
 532         struct sock *sk = sock->sk;
 533
 534         listen_sock.sk_data_ready = sk->sk_data_ready;
 535         listen_sock.sk_state_change = sk->sk_state_change;
 536         listen_sock.sk_write_space = sk->sk_write_space;
 537         listen_sock.sk_error_report = sk->sk_error_report;
 538 }
 539
 540 static void restore_callbacks(struct socket *sock)
 541 {
 542         struct sock *sk = sock->sk;
 543
 544         write_lock_bh(&sk->sk_callback_lock);
 545         sk->sk_user_data = NULL;
 546         sk->sk_data_ready = listen_sock.sk_data_ready;
 547         sk->sk_state_change = listen_sock.sk_state_change;
 548         sk->sk_write_space = listen_sock.sk_write_space;
 549         sk->sk_error_report = listen_sock.sk_error_report;
 550         write_unlock_bh(&sk->sk_callback_lock);
 551 }
 552
 553 /* Make a socket active */
 554 static void add_sock(struct socket *sock, struct connection *con)
 555 {
 556         struct sock *sk = sock->sk;
 557
 558         write_lock_bh(&sk->sk_callback_lock);
 559         con->sock = sock;
 560
 561         sk->sk_user_data = con;
 562         /* Install a data_ready callback */
 563         sk->sk_data_ready = lowcomms_data_ready;
 564         sk->sk_write_space = lowcomms_write_space;
 565         sk->sk_state_change = lowcomms_state_change;
 566         sk->sk_allocation = GFP_NOFS;
 567         sk->sk_error_report = lowcomms_error_report;
 568         write_unlock_bh(&sk->sk_callback_lock);
 569 }
 570
 571 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 572    length */
 573 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 574                           int *addr_len)
 575 {
 576         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 577         if (saddr->ss_family == AF_INET) {
 578                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 579                 in4_addr->sin_port = cpu_to_be16(port);
 580                 *addr_len = sizeof(struct sockaddr_in);
 581                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 582         } else {
 583                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 584                 in6_addr->sin6_port = cpu_to_be16(port);
 585                 *addr_len = sizeof(struct sockaddr_in6);
 586         }
 587         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 588 }
 589
 590 /* Close a remote connection and tidy up */
 591 static void close_connection(struct connection *con, bool and_other,
 592                              bool tx, bool rx)
 593 {
 594         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 595
 596         if (tx && !closing && cancel_work_sync(&con->swork)) {
 597                 log_print("canceled swork for node %d", con->nodeid);
 598                 clear_bit(CF_WRITE_PENDING, &con->flags);
 599         }
 600         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 601                 log_print("canceled rwork for node %d", con->nodeid);
 602                 clear_bit(CF_READ_PENDING, &con->flags);
 603         }
 604
 605         mutex_lock(&con->sock_mutex);
 606         if (con->sock) {
 607                 restore_callbacks(con->sock);
 608                 sock_release(con->sock);
 609                 con->sock = NULL;
 610         }
 611         if (con->othercon && and_other) {
 612                 /* Will only re-enter once. */
 613                 close_connection(con->othercon, false, true, true);
 614         }
 615         if (con->rx_page) {
 616                 __free_page(con->rx_page);
 617                 con->rx_page = NULL;
 618         }
 619
 620         con->retries = 0;
 621         mutex_unlock(&con->sock_mutex);
 622         clear_bit(CF_CLOSING, &con->flags);
 623 }
 624
 625 /* Data received from remote end */
 626 static int receive_from_sock(struct connection *con)
 627 {
 628         int ret = 0;
 629         struct msghdr msg = {};
 630         struct kvec iov[2];
 631         unsigned len;
 632         int r;
 633         int call_again_soon = 0;
 634         int nvec;
 635
 636         mutex_lock(&con->sock_mutex);
 637
 638         if (con->sock == NULL) {
 639                 ret = -EAGAIN;
 640                 goto out_close;
 641         }
 642         if (con->nodeid == 0) {
 643                 ret = -EINVAL;
 644                 goto out_close;
 645         }
 646
 647         if (con->rx_page == NULL) {
 648                 /*
 649                  * This doesn't need to be atomic, but I think it should
 650                  * improve performance if it is.
 651                  */
 652                 con->rx_page = alloc_page(GFP_ATOMIC);
 653                 if (con->rx_page == NULL)
 654                         goto out_resched;
 655                 cbuf_init(&con->cb, PAGE_SIZE);
 656         }
 657
 658         /*
 659          * iov[0] is the bit of the circular buffer between the current end
 660          * point (cb.base + cb.len) and the end of the buffer.
 661          */
 662         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 663         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 664         iov[1].iov_len = 0;
 665         nvec = 1;
 666
 667         /*
 668          * iov[1] is the bit of the circular buffer between the start of the
 669          * buffer and the start of the currently used section (cb.base)
 670          */
 671         if (cbuf_data(&con->cb) >= con->cb.base) {
 672                 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb);
 673                 iov[1].iov_len = con->cb.base;
 674                 iov[1].iov_base = page_address(con->rx_page);
 675                 nvec = 2;
 676         }
 677         len = iov[0].iov_len + iov[1].iov_len;
 678
 679         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 680                                MSG_DONTWAIT | MSG_NOSIGNAL);
 681         if (ret <= 0)
 682                 goto out_close;
 683         else if (ret == len)
 684                 call_again_soon = 1;
 685
 686         cbuf_add(&con->cb, ret);
 687         ret = dlm_process_incoming_buffer(con->nodeid,
 688                                           page_address(con->rx_page),
 689                                           con->cb.base, con->cb.len,
 690                                           PAGE_SIZE);
 691         if (ret == -EBADMSG) {
 692                 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d",
 693                           page_address(con->rx_page), con->cb.base,
 694                           con->cb.len, r);
 695         }
 696         if (ret < 0)
 697                 goto out_close;
 698         cbuf_eat(&con->cb, ret);
 699
 700         if (cbuf_empty(&con->cb) && !call_again_soon) {
 701                 __free_page(con->rx_page);
 702                 con->rx_page = NULL;
 703         }
 704
 705         if (call_again_soon)
 706                 goto out_resched;
 707         mutex_unlock(&con->sock_mutex);
 708         return 0;
 709
 710 out_resched:
 711         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 712                 queue_work(recv_workqueue, &con->rwork);
 713         mutex_unlock(&con->sock_mutex);
 714         return -EAGAIN;
 715
 716 out_close:
 717         mutex_unlock(&con->sock_mutex);
 718         if (ret != -EAGAIN) {
 719                 close_connection(con, true, true, false);
 720                 /* Reconnect when there is something to send */
 721         }
 722         /* Don't return success if we really got EOF */
 723         if (ret == 0)
 724                 ret = -EAGAIN;
 725
 726         return ret;
 727 }
 728
 729 /* Listening socket is busy, accept a connection */
 730 static int tcp_accept_from_sock(struct connection *con)
 731 {
 732         int result;
 733         struct sockaddr_storage peeraddr;
 734         struct socket *newsock;
 735         int len;
 736         int nodeid;
 737         struct connection *newcon;
 738         struct connection *addcon;
 739
 740         mutex_lock(&connections_lock);
 741         if (!dlm_allow_conn) {
 742                 mutex_unlock(&connections_lock);
 743                 return -1;
 744         }
 745         mutex_unlock(&connections_lock);
 746
 747         mutex_lock_nested(&con->sock_mutex, 0);
 748
 749         if (!con->sock) {
 750                 mutex_unlock(&con->sock_mutex);
 751                 return -ENOTCONN;
 752         }
 753
 754         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 755         if (result < 0)
 756                 goto accept_err;
 757
 758         /* Get the connected socket's peer */
 759         memset(&peeraddr, 0, sizeof(peeraddr));
 760         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 761                                   &len, 2)) {
 762                 result = -ECONNABORTED;
 763                 goto accept_err;
 764         }
 765
 766         /* Get the new node's NODEID */
 767         make_sockaddr(&peeraddr, 0, &len);
 768         if (addr_to_nodeid(&peeraddr, &nodeid)) {
 769                 unsigned char *b=(unsigned char *)&peeraddr;
 770                 log_print("connect from non cluster node");
 771                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 772                                      b, sizeof(struct sockaddr_storage));
 773                 sock_release(newsock);
 774                 mutex_unlock(&con->sock_mutex);
 775                 return -1;
 776         }
 777
 778         log_print("got connection from %d", nodeid);
 779
 780         /*  Check to see if we already have a connection to this node. This
 781          *  could happen if the two nodes initiate a connection at roughly
 782          *  the same time and the connections cross on the wire.
 783          *  In this case we store the incoming one in "othercon"
 784          */
 785         newcon = nodeid2con(nodeid, GFP_NOFS);
 786         if (!newcon) {
 787                 result = -ENOMEM;
 788                 goto accept_err;
 789         }
 790         mutex_lock_nested(&newcon->sock_mutex, 1);
 791         if (newcon->sock) {
 792                 struct connection *othercon = newcon->othercon;
 793
 794                 if (!othercon) {
 795                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 796                         if (!othercon) {
 797                                 log_print("failed to allocate incoming socket");
 798                                 mutex_unlock(&newcon->sock_mutex);
 799                                 result = -ENOMEM;
 800                                 goto accept_err;
 801                         }
 802                         othercon->nodeid = nodeid;
 803                         othercon->rx_action = receive_from_sock;
 804                         mutex_init(&othercon->sock_mutex);
 805                         INIT_WORK(&othercon->swork, process_send_sockets);
 806                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 807                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 808                 }
 809                 mutex_lock_nested(&othercon->sock_mutex, 2);
 810                 if (!othercon->sock) {
 811                         newcon->othercon = othercon;
 812                         add_sock(newsock, othercon);
 813                         addcon = othercon;
 814                         mutex_unlock(&othercon->sock_mutex);
 815                 }
 816                 else {
 817                         printk("Extra connection from node %d attempted\n", nodeid);
 818                         result = -EAGAIN;
 819                         mutex_unlock(&othercon->sock_mutex);
 820                         mutex_unlock(&newcon->sock_mutex);
 821                         goto accept_err;
 822                 }
 823         }
 824         else {
 825                 newcon->rx_action = receive_from_sock;
 826                 /* accept copies the sk after we've saved the callbacks, so we
 827                    don't want to save them a second time or comm errors will
 828                    result in calling sk_error_report recursively. */
 829                 add_sock(newsock, newcon);
 830                 addcon = newcon;
 831         }
 832
 833         mutex_unlock(&newcon->sock_mutex);
 834
 835         /*
 836          * Add it to the active queue in case we got data
 837          * between processing the accept adding the socket
 838          * to the read_sockets list
 839          */
 840         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 841                 queue_work(recv_workqueue, &addcon->rwork);
 842         mutex_unlock(&con->sock_mutex);
 843
 844         return 0;
 845
 846 accept_err:
 847         mutex_unlock(&con->sock_mutex);
 848         if (newsock)
 849                 sock_release(newsock);
 850
 851         if (result != -EAGAIN)
 852                 log_print("error accepting connection from node: %d", result);
 853         return result;
 854 }
 855
 856 static int sctp_accept_from_sock(struct connection *con)
 857 {
 858         /* Check that the new node is in the lockspace */
 859         struct sctp_prim prim;
 860         int nodeid;
 861         int prim_len, ret;
 862         int addr_len;
 863         struct connection *newcon;
 864         struct connection *addcon;
 865         struct socket *newsock;
 866
 867         mutex_lock(&connections_lock);
 868         if (!dlm_allow_conn) {
 869                 mutex_unlock(&connections_lock);
 870                 return -1;
 871         }
 872         mutex_unlock(&connections_lock);
 873
 874         mutex_lock_nested(&con->sock_mutex, 0);
 875
 876         ret = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 877         if (ret < 0)
 878                 goto accept_err;
 879
 880         memset(&prim, 0, sizeof(struct sctp_prim));
 881         prim_len = sizeof(struct sctp_prim);
 882
 883         ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR,
 884                                 (char *)&prim, &prim_len);
 885         if (ret < 0) {
 886                 log_print("getsockopt/sctp_primary_addr failed: %d", ret);
 887                 goto accept_err;
 888         }
 889
 890         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 891         ret = addr_to_nodeid(&prim.ssp_addr, &nodeid);
 892         if (ret) {
 893                 unsigned char *b = (unsigned char *)&prim.ssp_addr;
 894
 895                 log_print("reject connect from unknown addr");
 896                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 897                                      b, sizeof(struct sockaddr_storage));
 898                 goto accept_err;
 899         }
 900
 901         newcon = nodeid2con(nodeid, GFP_NOFS);
 902         if (!newcon) {
 903                 ret = -ENOMEM;
 904                 goto accept_err;
 905         }
 906
 907         mutex_lock_nested(&newcon->sock_mutex, 1);
 908
 909         if (newcon->sock) {
 910                 struct connection *othercon = newcon->othercon;
 911
 912                 if (!othercon) {
 913                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 914                         if (!othercon) {
 915                                 log_print("failed to allocate incoming socket");
 916                                 mutex_unlock(&newcon->sock_mutex);
 917                                 ret = -ENOMEM;
 918                                 goto accept_err;
 919                         }
 920                         othercon->nodeid = nodeid;
 921                         othercon->rx_action = receive_from_sock;
 922                         mutex_init(&othercon->sock_mutex);
 923                         INIT_WORK(&othercon->swork, process_send_sockets);
 924                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 925                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 926                 }
 927                 mutex_lock_nested(&othercon->sock_mutex, 2);
 928                 if (!othercon->sock) {
 929                         newcon->othercon = othercon;
 930                         add_sock(newsock, othercon);
 931                         addcon = othercon;
 932                         mutex_unlock(&othercon->sock_mutex);
 933                 } else {
 934                         printk("Extra connection from node %d attempted\n", nodeid);
 935                         ret = -EAGAIN;
 936                         mutex_unlock(&othercon->sock_mutex);
 937                         mutex_unlock(&newcon->sock_mutex);
 938                         goto accept_err;
 939                 }
 940         } else {
 941                 newcon->rx_action = receive_from_sock;
 942                 add_sock(newsock, newcon);
 943                 addcon = newcon;
 944         }
 945
 946         log_print("connected to %d", nodeid);
 947
 948         mutex_unlock(&newcon->sock_mutex);
 949
 950         /*
 951          * Add it to the active queue in case we got data
 952          * between processing the accept adding the socket
 953          * to the read_sockets list
 954          */
 955         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 956                 queue_work(recv_workqueue, &addcon->rwork);
 957         mutex_unlock(&con->sock_mutex);
 958
 959         return 0;
 960
 961 accept_err:
 962         mutex_unlock(&con->sock_mutex);
 963         if (newsock)
 964                 sock_release(newsock);
 965         if (ret != -EAGAIN)
 966                 log_print("error accepting connection from node: %d", ret);
 967
 968         return ret;
 969 }
 970
 971 static void free_entry(struct writequeue_entry *e)
 972 {
 973         __free_page(e->page);
 974         kfree(e);
 975 }
 976
 977 /*
 978  * writequeue_entry_complete - try to delete and free write queue entry
 979  * @e: write queue entry to try to delete
 980  * @completed: bytes completed
 981  *
 982  * writequeue_lock must be held.
 983  */
 984 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 985 {
 986         e->offset += completed;
 987         e->len -= completed;
 988
 989         if (e->len == 0 && e->users == 0) {
 990                 list_del(&e->list);
 991                 free_entry(e);
 992         }
 993 }
 994
 995 /*
 996  * sctp_bind_addrs - bind a SCTP socket to all our addresses
 997  */
 998 static int sctp_bind_addrs(struct connection *con, uint16_t port)
 999 {
1000         struct sockaddr_storage localaddr;
1001         int i, addr_len, result = 0;
1002
1003         for (i = 0; i < dlm_local_count; i++) {
1004                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1005                 make_sockaddr(&localaddr, port, &addr_len);
1006
1007                 if (!i)
1008                         result = kernel_bind(con->sock,
1009                                              (struct sockaddr *)&localaddr,
1010                                              addr_len);
1011                 else
1012                         result = kernel_setsockopt(con->sock, SOL_SCTP,
1013                                                    SCTP_SOCKOPT_BINDX_ADD,
1014                                                    (char *)&localaddr, addr_len);
1015
1016                 if (result < 0) {
1017                         log_print("Can't bind to %d addr number %d, %d.\n",
1018                                   port, i + 1, result);
1019                         break;
1020                 }
1021         }
1022         return result;
1023 }
1024
1025 /* Initiate an SCTP association.
1026    This is a special case of send_to_sock() in that we don't yet have a
1027    peeled-off socket for this association, so we use the listening socket
1028    and add the primary IP address of the remote node.
1029  */
1030 static void sctp_connect_to_sock(struct connection *con)
1031 {
1032         struct sockaddr_storage daddr;
1033         int one = 1;
1034         int result;
1035         int addr_len;
1036         struct socket *sock;
1037
1038         if (con->nodeid == 0) {
1039                 log_print("attempt to connect sock 0 foiled");
1040                 return;
1041         }
1042
1043         mutex_lock(&con->sock_mutex);
1044
1045         /* Some odd races can cause double-connects, ignore them */
1046         if (con->retries++ > MAX_CONNECT_RETRIES)
1047                 goto out;
1048
1049         if (con->sock) {
1050                 log_print("node %d already connected.", con->nodeid);
1051                 goto out;
1052         }
1053
1054         memset(&daddr, 0, sizeof(daddr));
1055         result = nodeid_to_addr(con->nodeid, &daddr, NULL, true);
1056         if (result < 0) {
1057                 log_print("no address for nodeid %d", con->nodeid);
1058                 goto out;
1059         }
1060
1061         /* Create a socket to communicate with */
1062         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1063                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1064         if (result < 0)
1065                 goto socket_err;
1066
1067         con->rx_action = receive_from_sock;
1068         con->connect_action = sctp_connect_to_sock;
1069         add_sock(sock, con);
1070
1071         /* Bind to all addresses. */
1072         if (sctp_bind_addrs(con, 0))
1073                 goto bind_err;
1074
1075         make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1076
1077         log_print("connecting to %d", con->nodeid);
1078
1079         /* Turn off Nagle's algorithm */
1080         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1081                           sizeof(one));
1082
1083         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1084                                    O_NONBLOCK);
1085         if (result == -EINPROGRESS)
1086                 result = 0;
1087         if (result == 0)
1088                 goto out;
1089
1090 bind_err:
1091         con->sock = NULL;
1092         sock_release(sock);
1093
1094 socket_err:
1095         /*
1096          * Some errors are fatal and this list might need adjusting. For other
1097          * errors we try again until the max number of retries is reached.
1098          */
1099         if (result != -EHOSTUNREACH &&
1100             result != -ENETUNREACH &&
1101             result != -ENETDOWN &&
1102             result != -EINVAL &&
1103             result != -EPROTONOSUPPORT) {
1104                 log_print("connect %d try %d error %d", con->nodeid,
1105                           con->retries, result);
1106                 mutex_unlock(&con->sock_mutex);
1107                 msleep(1000);
1108                 lowcomms_connect_sock(con);
1109                 return;
1110         }
1111
1112 out:
1113         mutex_unlock(&con->sock_mutex);
1114 }
1115
1116 /* Connect a new socket to its peer */
1117 static void tcp_connect_to_sock(struct connection *con)
1118 {
1119         struct sockaddr_storage saddr, src_addr;
1120         int addr_len;
1121         struct socket *sock = NULL;
1122         int one = 1;
1123         int result;
1124
1125         if (con->nodeid == 0) {
1126                 log_print("attempt to connect sock 0 foiled");
1127                 return;
1128         }
1129
1130         mutex_lock(&con->sock_mutex);
1131         if (con->retries++ > MAX_CONNECT_RETRIES)
1132                 goto out;
1133
1134         /* Some odd races can cause double-connects, ignore them */
1135         if (con->sock)
1136                 goto out;
1137
1138         /* Create a socket to communicate with */
1139         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1140                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1141         if (result < 0)
1142                 goto out_err;
1143
1144         memset(&saddr, 0, sizeof(saddr));
1145         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1146         if (result < 0) {
1147                 log_print("no address for nodeid %d", con->nodeid);
1148                 goto out_err;
1149         }
1150
1151         con->rx_action = receive_from_sock;
1152         con->connect_action = tcp_connect_to_sock;
1153         add_sock(sock, con);
1154
1155         /* Bind to our cluster-known address connecting to avoid
1156            routing problems */
1157         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1158         make_sockaddr(&src_addr, 0, &addr_len);
1159         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1160                                  addr_len);
1161         if (result < 0) {
1162                 log_print("could not bind for connect: %d", result);
1163                 /* This *may* not indicate a critical error */
1164         }
1165
1166         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1167
1168         log_print("connecting to %d", con->nodeid);
1169
1170         /* Turn off Nagle's algorithm */
1171         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1172                           sizeof(one));
1173
1174         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1175                                    O_NONBLOCK);
1176         if (result == -EINPROGRESS)
1177                 result = 0;
1178         if (result == 0)
1179                 goto out;
1180
1181 out_err:
1182         if (con->sock) {
1183                 sock_release(con->sock);
1184                 con->sock = NULL;
1185         } else if (sock) {
1186                 sock_release(sock);
1187         }
1188         /*
1189          * Some errors are fatal and this list might need adjusting. For other
1190          * errors we try again until the max number of retries is reached.
1191          */
1192         if (result != -EHOSTUNREACH &&
1193             result != -ENETUNREACH &&
1194             result != -ENETDOWN &&
1195             result != -EINVAL &&
1196             result != -EPROTONOSUPPORT) {
1197                 log_print("connect %d try %d error %d", con->nodeid,
1198                           con->retries, result);
1199                 mutex_unlock(&con->sock_mutex);
1200                 msleep(1000);
1201                 lowcomms_connect_sock(con);
1202                 return;
1203         }
1204 out:
1205         mutex_unlock(&con->sock_mutex);
1206         return;
1207 }
1208
1209 static struct socket *tcp_create_listen_sock(struct connection *con,
1210                                              struct sockaddr_storage *saddr)
1211 {
1212         struct socket *sock = NULL;
1213         int result = 0;
1214         int one = 1;
1215         int addr_len;
1216
1217         if (dlm_local_addr[0]->ss_family == AF_INET)
1218                 addr_len = sizeof(struct sockaddr_in);
1219         else
1220                 addr_len = sizeof(struct sockaddr_in6);
1221
1222         /* Create a socket to communicate with */
1223         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1224                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1225         if (result < 0) {
1226                 log_print("Can't create listening comms socket");
1227                 goto create_out;
1228         }
1229
1230         /* Turn off Nagle's algorithm */
1231         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1232                           sizeof(one));
1233
1234         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1235                                    (char *)&one, sizeof(one));
1236
1237         if (result < 0) {
1238                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1239         }
1240         write_lock_bh(&sock->sk->sk_callback_lock);
1241         sock->sk->sk_user_data = con;
1242         save_listen_callbacks(sock);
1243         con->rx_action = tcp_accept_from_sock;
1244         con->connect_action = tcp_connect_to_sock;
1245         write_unlock_bh(&sock->sk->sk_callback_lock);
1246
1247         /* Bind to our port */
1248         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1249         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1250         if (result < 0) {
1251                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1252                 sock_release(sock);
1253                 sock = NULL;
1254                 con->sock = NULL;
1255                 goto create_out;
1256         }
1257         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1258                                  (char *)&one, sizeof(one));
1259         if (result < 0) {
1260                 log_print("Set keepalive failed: %d", result);
1261         }
1262
1263         result = sock->ops->listen(sock, 5);
1264         if (result < 0) {
1265                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1266                 sock_release(sock);
1267                 sock = NULL;
1268                 goto create_out;
1269         }
1270
1271 create_out:
1272         return sock;
1273 }
1274
1275 /* Get local addresses */
1276 static void init_local(void)
1277 {
1278         struct sockaddr_storage sas, *addr;
1279         int i;
1280
1281         dlm_local_count = 0;
1282         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1283                 if (dlm_our_addr(&sas, i))
1284                         break;
1285
1286                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1287                 if (!addr)
1288                         break;
1289                 dlm_local_addr[dlm_local_count++] = addr;
1290         }
1291 }
1292
1293 /* Initialise SCTP socket and bind to all interfaces */
1294 static int sctp_listen_for_all(void)
1295 {
1296         struct socket *sock = NULL;
1297         int result = -EINVAL;
1298         struct connection *con = nodeid2con(0, GFP_NOFS);
1299         int bufsize = NEEDED_RMEM;
1300         int one = 1;
1301
1302         if (!con)
1303                 return -ENOMEM;
1304
1305         log_print("Using SCTP for communications");
1306
1307         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1308                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1309         if (result < 0) {
1310                 log_print("Can't create comms socket, check SCTP is loaded");
1311                 goto out;
1312         }
1313
1314         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1315                                  (char *)&bufsize, sizeof(bufsize));
1316         if (result)
1317                 log_print("Error increasing buffer space on socket %d", result);
1318
1319         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1320                                    sizeof(one));
1321         if (result < 0)
1322                 log_print("Could not set SCTP NODELAY error %d\n", result);
1323
1324         write_lock_bh(&sock->sk->sk_callback_lock);
1325         /* Init con struct */
1326         sock->sk->sk_user_data = con;
1327         save_listen_callbacks(sock);
1328         con->sock = sock;
1329         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1330         con->rx_action = sctp_accept_from_sock;
1331         con->connect_action = sctp_connect_to_sock;
1332
1333         write_unlock_bh(&sock->sk->sk_callback_lock);
1334
1335         /* Bind to all addresses. */
1336         if (sctp_bind_addrs(con, dlm_config.ci_tcp_port))
1337                 goto create_delsock;
1338
1339         result = sock->ops->listen(sock, 5);
1340         if (result < 0) {
1341                 log_print("Can't set socket listening");
1342                 goto create_delsock;
1343         }
1344
1345         return 0;
1346
1347 create_delsock:
1348         sock_release(sock);
1349         con->sock = NULL;
1350 out:
1351         return result;
1352 }
1353
1354 static int tcp_listen_for_all(void)
1355 {
1356         struct socket *sock = NULL;
1357         struct connection *con = nodeid2con(0, GFP_NOFS);
1358         int result = -EINVAL;
1359
1360         if (!con)
1361                 return -ENOMEM;
1362
1363         /* We don't support multi-homed hosts */
1364         if (dlm_local_addr[1] != NULL) {
1365                 log_print("TCP protocol can't handle multi-homed hosts, "
1366                           "try SCTP");
1367                 return -EINVAL;
1368         }
1369
1370         log_print("Using TCP for communications");
1371
1372         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1373         if (sock) {
1374                 add_sock(sock, con);
1375                 result = 0;
1376         }
1377         else {
1378                 result = -EADDRINUSE;
1379         }
1380
1381         return result;
1382 }
1383
1384
1385
1386 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1387                                                      gfp_t allocation)
1388 {
1389         struct writequeue_entry *entry;
1390
1391         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1392         if (!entry)
1393                 return NULL;
1394
1395         entry->page = alloc_page(allocation);
1396         if (!entry->page) {
1397                 kfree(entry);
1398                 return NULL;
1399         }
1400
1401         entry->offset = 0;
1402         entry->len = 0;
1403         entry->end = 0;
1404         entry->users = 0;
1405         entry->con = con;
1406
1407         return entry;
1408 }
1409
1410 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1411 {
1412         struct connection *con;
1413         struct writequeue_entry *e;
1414         int offset = 0;
1415
1416         con = nodeid2con(nodeid, allocation);
1417         if (!con)
1418                 return NULL;
1419
1420         spin_lock(&con->writequeue_lock);
1421         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1422         if ((&e->list == &con->writequeue) ||
1423             (PAGE_SIZE - e->end < len)) {
1424                 e = NULL;
1425         } else {
1426                 offset = e->end;
1427                 e->end += len;
1428                 e->users++;
1429         }
1430         spin_unlock(&con->writequeue_lock);
1431
1432         if (e) {
1433         got_one:
1434                 *ppc = page_address(e->page) + offset;
1435                 return e;
1436         }
1437
1438         e = new_writequeue_entry(con, allocation);
1439         if (e) {
1440                 spin_lock(&con->writequeue_lock);
1441                 offset = e->end;
1442                 e->end += len;
1443                 e->users++;
1444                 list_add_tail(&e->list, &con->writequeue);
1445                 spin_unlock(&con->writequeue_lock);
1446                 goto got_one;
1447         }
1448         return NULL;
1449 }
1450
1451 void dlm_lowcomms_commit_buffer(void *mh)
1452 {
1453         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1454         struct connection *con = e->con;
1455         int users;
1456
1457         spin_lock(&con->writequeue_lock);
1458         users = --e->users;
1459         if (users)
1460                 goto out;
1461         e->len = e->end - e->offset;
1462         spin_unlock(&con->writequeue_lock);
1463
1464         queue_work(send_workqueue, &con->swork);
1465         return;
1466
1467 out:
1468         spin_unlock(&con->writequeue_lock);
1469         return;
1470 }
1471
1472 /* Send a message */
1473 static void send_to_sock(struct connection *con)
1474 {
1475         int ret = 0;
1476         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1477         struct writequeue_entry *e;
1478         int len, offset;
1479         int count = 0;
1480
1481         mutex_lock(&con->sock_mutex);
1482         if (con->sock == NULL)
1483                 goto out_connect;
1484
1485         spin_lock(&con->writequeue_lock);
1486         for (;;) {
1487                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1488                                list);
1489                 if ((struct list_head *) e == &con->writequeue)
1490                         break;
1491
1492                 len = e->len;
1493                 offset = e->offset;
1494                 BUG_ON(len == 0 && e->users == 0);
1495                 spin_unlock(&con->writequeue_lock);
1496
1497                 ret = 0;
1498                 if (len) {
1499                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1500                                               msg_flags);
1501                         if (ret == -EAGAIN || ret == 0) {
1502                                 if (ret == -EAGAIN &&
1503                                     test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1504                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1505                                         /* Notify TCP that we're limited by the
1506                                          * application window size.
1507                                          */
1508                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1509                                         con->sock->sk->sk_write_pending++;
1510                                 }
1511                                 cond_resched();
1512                                 goto out;
1513                         } else if (ret < 0)
1514                                 goto send_error;
1515                 }
1516
1517                 /* Don't starve people filling buffers */
1518                 if (++count >= MAX_SEND_MSG_COUNT) {
1519                         cond_resched();
1520                         count = 0;
1521                 }
1522
1523                 spin_lock(&con->writequeue_lock);
1524                 writequeue_entry_complete(e, ret);
1525         }
1526         spin_unlock(&con->writequeue_lock);
1527 out:
1528         mutex_unlock(&con->sock_mutex);
1529         return;
1530
1531 send_error:
1532         mutex_unlock(&con->sock_mutex);
1533         close_connection(con, true, false, true);
1534         /* Requeue the send work. When the work daemon runs again, it will try
1535            a new connection, then call this function again. */
1536         queue_work(send_workqueue, &con->swork);
1537         return;
1538
1539 out_connect:
1540         mutex_unlock(&con->sock_mutex);
1541         queue_work(send_workqueue, &con->swork);
1542         cond_resched();
1543 }
1544
1545 static void clean_one_writequeue(struct connection *con)
1546 {
1547         struct writequeue_entry *e, *safe;
1548
1549         spin_lock(&con->writequeue_lock);
1550         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1551                 list_del(&e->list);
1552                 free_entry(e);
1553         }
1554         spin_unlock(&con->writequeue_lock);
1555 }
1556
1557 /* Called from recovery when it knows that a node has
1558    left the cluster */
1559 int dlm_lowcomms_close(int nodeid)
1560 {
1561         struct connection *con;
1562         struct dlm_node_addr *na;
1563
1564         log_print("closing connection to node %d", nodeid);
1565         con = nodeid2con(nodeid, 0);
1566         if (con) {
1567                 set_bit(CF_CLOSE, &con->flags);
1568                 close_connection(con, true, true, true);
1569                 clean_one_writequeue(con);
1570         }
1571
1572         spin_lock(&dlm_node_addrs_spin);
1573         na = find_node_addr(nodeid);
1574         if (na) {
1575                 list_del(&na->list);
1576                 while (na->addr_count--)
1577                         kfree(na->addr[na->addr_count]);
1578                 kfree(na);
1579         }
1580         spin_unlock(&dlm_node_addrs_spin);
1581
1582         return 0;
1583 }
1584
1585 /* Receive workqueue function */
1586 static void process_recv_sockets(struct work_struct *work)
1587 {
1588         struct connection *con = container_of(work, struct connection, rwork);
1589         int err;
1590
1591         clear_bit(CF_READ_PENDING, &con->flags);
1592         do {
1593                 err = con->rx_action(con);
1594         } while (!err);
1595 }
1596
1597 /* Send workqueue function */
1598 static void process_send_sockets(struct work_struct *work)
1599 {
1600         struct connection *con = container_of(work, struct connection, swork);
1601
1602         clear_bit(CF_WRITE_PENDING, &con->flags);
1603         if (con->sock == NULL) /* not mutex protected so check it inside too */
1604                 con->connect_action(con);
1605         if (!list_empty(&con->writequeue))
1606                 send_to_sock(con);
1607 }
1608
1609
1610 /* Discard all entries on the write queues */
1611 static void clean_writequeues(void)
1612 {
1613         foreach_conn(clean_one_writequeue);
1614 }
1615
1616 static void work_stop(void)
1617 {
1618         destroy_workqueue(recv_workqueue);
1619         destroy_workqueue(send_workqueue);
1620 }
1621
1622 static int work_start(void)
1623 {
1624         recv_workqueue = alloc_workqueue("dlm_recv",
1625                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1626         if (!recv_workqueue) {
1627                 log_print("can't start dlm_recv");
1628                 return -ENOMEM;
1629         }
1630
1631         send_workqueue = alloc_workqueue("dlm_send",
1632                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1633         if (!send_workqueue) {
1634                 log_print("can't start dlm_send");
1635                 destroy_workqueue(recv_workqueue);
1636                 return -ENOMEM;
1637         }
1638
1639         return 0;
1640 }
1641
1642 static void _stop_conn(struct connection *con, bool and_other)
1643 {
1644         mutex_lock(&con->sock_mutex);
1645         set_bit(CF_CLOSE, &con->flags);
1646         set_bit(CF_READ_PENDING, &con->flags);
1647         set_bit(CF_WRITE_PENDING, &con->flags);
1648         if (con->sock && con->sock->sk) {
1649                 write_lock_bh(&con->sock->sk->sk_callback_lock);
1650                 con->sock->sk->sk_user_data = NULL;
1651                 write_unlock_bh(&con->sock->sk->sk_callback_lock);
1652         }
1653         if (con->othercon && and_other)
1654                 _stop_conn(con->othercon, false);
1655         mutex_unlock(&con->sock_mutex);
1656 }
1657
1658 static void stop_conn(struct connection *con)
1659 {
1660         _stop_conn(con, true);
1661 }
1662
1663 static void free_conn(struct connection *con)
1664 {
1665         close_connection(con, true, true, true);
1666         if (con->othercon)
1667                 kmem_cache_free(con_cache, con->othercon);
1668         hlist_del(&con->list);
1669         kmem_cache_free(con_cache, con);
1670 }
1671
1672 static void work_flush(void)
1673 {
1674         int ok;
1675         int i;
1676         struct hlist_node *n;
1677         struct connection *con;
1678
1679         flush_workqueue(recv_workqueue);
1680         flush_workqueue(send_workqueue);
1681         do {
1682                 ok = 1;
1683                 foreach_conn(stop_conn);
1684                 flush_workqueue(recv_workqueue);
1685                 flush_workqueue(send_workqueue);
1686                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1687                         hlist_for_each_entry_safe(con, n,
1688                                                   &connection_hash[i], list) {
1689                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1690                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1691                                 if (con->othercon) {
1692                                         ok &= test_bit(CF_READ_PENDING,
1693                                                        &con->othercon->flags);
1694                                         ok &= test_bit(CF_WRITE_PENDING,
1695                                                        &con->othercon->flags);
1696                                 }
1697                         }
1698                 }
1699         } while (!ok);
1700 }
1701
1702 void dlm_lowcomms_stop(void)
1703 {
1704         /* Set all the flags to prevent any
1705            socket activity.
1706         */
1707         mutex_lock(&connections_lock);
1708         dlm_allow_conn = 0;
1709         mutex_unlock(&connections_lock);
1710         work_flush();
1711         clean_writequeues();
1712         foreach_conn(free_conn);
1713         work_stop();
1714
1715         kmem_cache_destroy(con_cache);
1716 }
1717
1718 int dlm_lowcomms_start(void)
1719 {
1720         int error = -EINVAL;
1721         struct connection *con;
1722         int i;
1723
1724         for (i = 0; i < CONN_HASH_SIZE; i++)
1725                 INIT_HLIST_HEAD(&connection_hash[i]);
1726
1727         init_local();
1728         if (!dlm_local_count) {
1729                 error = -ENOTCONN;
1730                 log_print("no local IP address has been set");
1731                 goto fail;
1732         }
1733
1734         error = -ENOMEM;
1735         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1736                                       __alignof__(struct connection), 0,
1737                                       NULL);
1738         if (!con_cache)
1739                 goto fail;
1740
1741         error = work_start();
1742         if (error)
1743                 goto fail_destroy;
1744
1745         dlm_allow_conn = 1;
1746
1747         /* Start listening */
1748         if (dlm_config.ci_protocol == 0)
1749                 error = tcp_listen_for_all();
1750         else
1751                 error = sctp_listen_for_all();
1752         if (error)
1753                 goto fail_unlisten;
1754
1755         return 0;
1756
1757 fail_unlisten:
1758         dlm_allow_conn = 0;
1759         con = nodeid2con(0,0);
1760         if (con) {
1761                 close_connection(con, false, true, true);
1762                 kmem_cache_free(con_cache, con);
1763         }
1764 fail_destroy:
1765         kmem_cache_destroy(con_cache);
1766 fail:
1767         return error;
1768 }
1769
1770 void dlm_lowcomms_exit(void)
1771 {
1772         struct dlm_node_addr *na, *safe;
1773
1774         spin_lock(&dlm_node_addrs_spin);
1775         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1776                 list_del(&na->list);
1777                 while (na->addr_count--)
1778                         kfree(na->addr[na->addr_count]);
1779                 kfree(na);
1780         }
1781         spin_unlock(&dlm_node_addrs_spin);
1782 }