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