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/user.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_CONNECT_PENDING 3
 111 #define CF_INIT_PENDING 4
 112 #define CF_IS_OTHERCON 5
 113 #define CF_CLOSE 6
 114 #define CF_APP_LIMITED 7
 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         int sctp_assoc;
 124         struct hlist_node list;
 125         struct connection *othercon;
 126         struct work_struct rwork; /* Receive workqueue */
 127         struct work_struct swork; /* Send workqueue */
 128 };
 129 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 130
 131 /* An entry waiting to be sent */
 132 struct writequeue_entry {
 133         struct list_head list;
 134         struct page *page;
 135         int offset;
 136         int len;
 137         int end;
 138         int users;
 139         struct connection *con;
 140 };
 141
 142 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 143 static int dlm_local_count;
 144
 145 /* Work queues */
 146 static struct workqueue_struct *recv_workqueue;
 147 static struct workqueue_struct *send_workqueue;
 148
 149 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 150 static DEFINE_MUTEX(connections_lock);
 151 static struct kmem_cache *con_cache;
 152
 153 static void process_recv_sockets(struct work_struct *work);
 154 static void process_send_sockets(struct work_struct *work);
 155
 156
 157 /* This is deliberately very simple because most clusters have simple
 158    sequential nodeids, so we should be able to go straight to a connection
 159    struct in the array */
 160 static inline int nodeid_hash(int nodeid)
 161 {
 162         return nodeid & (CONN_HASH_SIZE-1);
 163 }
 164
 165 static struct connection *__find_con(int nodeid)
 166 {
 167         int r;
 168         struct hlist_node *h;
 169         struct connection *con;
 170
 171         r = nodeid_hash(nodeid);
 172
 173         hlist_for_each_entry(con, h, &connection_hash[r], list) {
 174                 if (con->nodeid == nodeid)
 175                         return con;
 176         }
 177         return NULL;
 178 }
 179
 180 /*
 181  * If 'allocation' is zero then we don't attempt to create a new
 182  * connection structure for this node.
 183  */
 184 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 185 {
 186         struct connection *con = NULL;
 187         int r;
 188
 189         con = __find_con(nodeid);
 190         if (con || !alloc)
 191                 return con;
 192
 193         con = kmem_cache_zalloc(con_cache, alloc);
 194         if (!con)
 195                 return NULL;
 196
 197         r = nodeid_hash(nodeid);
 198         hlist_add_head(&con->list, &connection_hash[r]);
 199
 200         con->nodeid = nodeid;
 201         mutex_init(&con->sock_mutex);
 202         INIT_LIST_HEAD(&con->writequeue);
 203         spin_lock_init(&con->writequeue_lock);
 204         INIT_WORK(&con->swork, process_send_sockets);
 205         INIT_WORK(&con->rwork, process_recv_sockets);
 206
 207         /* Setup action pointers for child sockets */
 208         if (con->nodeid) {
 209                 struct connection *zerocon = __find_con(0);
 210
 211                 con->connect_action = zerocon->connect_action;
 212                 if (!con->rx_action)
 213                         con->rx_action = zerocon->rx_action;
 214         }
 215
 216         return con;
 217 }
 218
 219 /* Loop round all connections */
 220 static void foreach_conn(void (*conn_func)(struct connection *c))
 221 {
 222         int i;
 223         struct hlist_node *h, *n;
 224         struct connection *con;
 225
 226         for (i = 0; i < CONN_HASH_SIZE; i++) {
 227                 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
 228                         conn_func(con);
 229                 }
 230         }
 231 }
 232
 233 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 234 {
 235         struct connection *con;
 236
 237         mutex_lock(&connections_lock);
 238         con = __nodeid2con(nodeid, allocation);
 239         mutex_unlock(&connections_lock);
 240
 241         return con;
 242 }
 243
 244 /* This is a bit drastic, but only called when things go wrong */
 245 static struct connection *assoc2con(int assoc_id)
 246 {
 247         int i;
 248         struct hlist_node *h;
 249         struct connection *con;
 250
 251         mutex_lock(&connections_lock);
 252
 253         for (i = 0 ; i < CONN_HASH_SIZE; i++) {
 254                 hlist_for_each_entry(con, h, &connection_hash[i], list) {
 255                         if (con->sctp_assoc == assoc_id) {
 256                                 mutex_unlock(&connections_lock);
 257                                 return con;
 258                         }
 259                 }
 260         }
 261         mutex_unlock(&connections_lock);
 262         return NULL;
 263 }
 264
 265 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 266 {
 267         struct sockaddr_storage addr;
 268         int error;
 269
 270         if (!dlm_local_count)
 271                 return -1;
 272
 273         error = dlm_nodeid_to_addr(nodeid, &addr);
 274         if (error)
 275                 return error;
 276
 277         if (dlm_local_addr[0]->ss_family == AF_INET) {
 278                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 279                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 280                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 281         } else {
 282                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 283                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 284                 ret6->sin6_addr = in6->sin6_addr;
 285         }
 286
 287         return 0;
 288 }
 289
 290 /* Data available on socket or listen socket received a connect */
 291 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 292 {
 293         struct connection *con = sock2con(sk);
 294         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 295                 queue_work(recv_workqueue, &con->rwork);
 296 }
 297
 298 static void lowcomms_write_space(struct sock *sk)
 299 {
 300         struct connection *con = sock2con(sk);
 301
 302         if (!con)
 303                 return;
 304
 305         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 306
 307         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 308                 con->sock->sk->sk_write_pending--;
 309                 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
 310         }
 311
 312         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 313                 queue_work(send_workqueue, &con->swork);
 314 }
 315
 316 static inline void lowcomms_connect_sock(struct connection *con)
 317 {
 318         if (test_bit(CF_CLOSE, &con->flags))
 319                 return;
 320         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 321                 queue_work(send_workqueue, &con->swork);
 322 }
 323
 324 static void lowcomms_state_change(struct sock *sk)
 325 {
 326         if (sk->sk_state == TCP_ESTABLISHED)
 327                 lowcomms_write_space(sk);
 328 }
 329
 330 int dlm_lowcomms_connect_node(int nodeid)
 331 {
 332         struct connection *con;
 333
 334         /* with sctp there's no connecting without sending */
 335         if (dlm_config.ci_protocol != 0)
 336                 return 0;
 337
 338         if (nodeid == dlm_our_nodeid())
 339                 return 0;
 340
 341         con = nodeid2con(nodeid, GFP_NOFS);
 342         if (!con)
 343                 return -ENOMEM;
 344         lowcomms_connect_sock(con);
 345         return 0;
 346 }
 347
 348 /* Make a socket active */
 349 static int add_sock(struct socket *sock, struct connection *con)
 350 {
 351         con->sock = sock;
 352
 353         /* Install a data_ready callback */
 354         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 355         con->sock->sk->sk_write_space = lowcomms_write_space;
 356         con->sock->sk->sk_state_change = lowcomms_state_change;
 357         con->sock->sk->sk_user_data = con;
 358         con->sock->sk->sk_allocation = GFP_NOFS;
 359         return 0;
 360 }
 361
 362 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 363    length */
 364 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 365                           int *addr_len)
 366 {
 367         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 368         if (saddr->ss_family == AF_INET) {
 369                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 370                 in4_addr->sin_port = cpu_to_be16(port);
 371                 *addr_len = sizeof(struct sockaddr_in);
 372                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 373         } else {
 374                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 375                 in6_addr->sin6_port = cpu_to_be16(port);
 376                 *addr_len = sizeof(struct sockaddr_in6);
 377         }
 378         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 379 }
 380
 381 /* Close a remote connection and tidy up */
 382 static void close_connection(struct connection *con, bool and_other)
 383 {
 384         mutex_lock(&con->sock_mutex);
 385
 386         if (con->sock) {
 387                 sock_release(con->sock);
 388                 con->sock = NULL;
 389         }
 390         if (con->othercon && and_other) {
 391                 /* Will only re-enter once. */
 392                 close_connection(con->othercon, false);
 393         }
 394         if (con->rx_page) {
 395                 __free_page(con->rx_page);
 396                 con->rx_page = NULL;
 397         }
 398
 399         con->retries = 0;
 400         mutex_unlock(&con->sock_mutex);
 401 }
 402
 403 /* We only send shutdown messages to nodes that are not part of the cluster */
 404 static void sctp_send_shutdown(sctp_assoc_t associd)
 405 {
 406         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 407         struct msghdr outmessage;
 408         struct cmsghdr *cmsg;
 409         struct sctp_sndrcvinfo *sinfo;
 410         int ret;
 411         struct connection *con;
 412
 413         con = nodeid2con(0,0);
 414         BUG_ON(con == NULL);
 415
 416         outmessage.msg_name = NULL;
 417         outmessage.msg_namelen = 0;
 418         outmessage.msg_control = outcmsg;
 419         outmessage.msg_controllen = sizeof(outcmsg);
 420         outmessage.msg_flags = MSG_EOR;
 421
 422         cmsg = CMSG_FIRSTHDR(&outmessage);
 423         cmsg->cmsg_level = IPPROTO_SCTP;
 424         cmsg->cmsg_type = SCTP_SNDRCV;
 425         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 426         outmessage.msg_controllen = cmsg->cmsg_len;
 427         sinfo = CMSG_DATA(cmsg);
 428         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 429
 430         sinfo->sinfo_flags |= MSG_EOF;
 431         sinfo->sinfo_assoc_id = associd;
 432
 433         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 434
 435         if (ret != 0)
 436                 log_print("send EOF to node failed: %d", ret);
 437 }
 438
 439 static void sctp_init_failed_foreach(struct connection *con)
 440 {
 441         con->sctp_assoc = 0;
 442         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 443                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 444                         queue_work(send_workqueue, &con->swork);
 445         }
 446 }
 447
 448 /* INIT failed but we don't know which node...
 449    restart INIT on all pending nodes */
 450 static void sctp_init_failed(void)
 451 {
 452         mutex_lock(&connections_lock);
 453
 454         foreach_conn(sctp_init_failed_foreach);
 455
 456         mutex_unlock(&connections_lock);
 457 }
 458
 459 /* Something happened to an association */
 460 static void process_sctp_notification(struct connection *con,
 461                                       struct msghdr *msg, char *buf)
 462 {
 463         union sctp_notification *sn = (union sctp_notification *)buf;
 464
 465         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 466                 switch (sn->sn_assoc_change.sac_state) {
 467
 468                 case SCTP_COMM_UP:
 469                 case SCTP_RESTART:
 470                 {
 471                         /* Check that the new node is in the lockspace */
 472                         struct sctp_prim prim;
 473                         int nodeid;
 474                         int prim_len, ret;
 475                         int addr_len;
 476                         struct connection *new_con;
 477                         sctp_peeloff_arg_t parg;
 478                         int parglen = sizeof(parg);
 479                         int err;
 480
 481                         /*
 482                          * We get this before any data for an association.
 483                          * We verify that the node is in the cluster and
 484                          * then peel off a socket for it.
 485                          */
 486                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 487                                 log_print("COMM_UP for invalid assoc ID %d",
 488                                          (int)sn->sn_assoc_change.sac_assoc_id);
 489                                 sctp_init_failed();
 490                                 return;
 491                         }
 492                         memset(&prim, 0, sizeof(struct sctp_prim));
 493                         prim_len = sizeof(struct sctp_prim);
 494                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 495
 496                         ret = kernel_getsockopt(con->sock,
 497                                                 IPPROTO_SCTP,
 498                                                 SCTP_PRIMARY_ADDR,
 499                                                 (char*)&prim,
 500                                                 &prim_len);
 501                         if (ret < 0) {
 502                                 log_print("getsockopt/sctp_primary_addr on "
 503                                           "new assoc %d failed : %d",
 504                                           (int)sn->sn_assoc_change.sac_assoc_id,
 505                                           ret);
 506
 507                                 /* Retry INIT later */
 508                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 509                                 if (new_con)
 510                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 511                                 return;
 512                         }
 513                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 514                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 515                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 516                                 log_print("reject connect from unknown addr");
 517                                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 518                                                      b, sizeof(struct sockaddr_storage));
 519                                 sctp_send_shutdown(prim.ssp_assoc_id);
 520                                 return;
 521                         }
 522
 523                         new_con = nodeid2con(nodeid, GFP_NOFS);
 524                         if (!new_con)
 525                                 return;
 526
 527                         /* Peel off a new sock */
 528                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 529                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 530                                                 SCTP_SOCKOPT_PEELOFF,
 531                                                 (void *)&parg, &parglen);
 532                         if (ret < 0) {
 533                                 log_print("Can't peel off a socket for "
 534                                           "connection %d to node %d: err=%d",
 535                                           parg.associd, nodeid, ret);
 536                                 return;
 537                         }
 538                         new_con->sock = sockfd_lookup(parg.sd, &err);
 539                         if (!new_con->sock) {
 540                                 log_print("sockfd_lookup error %d", err);
 541                                 return;
 542                         }
 543                         add_sock(new_con->sock, new_con);
 544                         sockfd_put(new_con->sock);
 545
 546                         log_print("connecting to %d sctp association %d",
 547                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 548
 549                         /* Send any pending writes */
 550                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 551                         clear_bit(CF_INIT_PENDING, &con->flags);
 552                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 553                                 queue_work(send_workqueue, &new_con->swork);
 554                         }
 555                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 556                                 queue_work(recv_workqueue, &new_con->rwork);
 557                 }
 558                 break;
 559
 560                 case SCTP_COMM_LOST:
 561                 case SCTP_SHUTDOWN_COMP:
 562                 {
 563                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 564                         if (con) {
 565                                 con->sctp_assoc = 0;
 566                         }
 567                 }
 568                 break;
 569
 570                 /* We don't know which INIT failed, so clear the PENDING flags
 571                  * on them all.  if assoc_id is zero then it will then try
 572                  * again */
 573
 574                 case SCTP_CANT_STR_ASSOC:
 575                 {
 576                         log_print("Can't start SCTP association - retrying");
 577                         sctp_init_failed();
 578                 }
 579                 break;
 580
 581                 default:
 582                         log_print("unexpected SCTP assoc change id=%d state=%d",
 583                                   (int)sn->sn_assoc_change.sac_assoc_id,
 584                                   sn->sn_assoc_change.sac_state);
 585                 }
 586         }
 587 }
 588
 589 /* Data received from remote end */
 590 static int receive_from_sock(struct connection *con)
 591 {
 592         int ret = 0;
 593         struct msghdr msg = {};
 594         struct kvec iov[2];
 595         unsigned len;
 596         int r;
 597         int call_again_soon = 0;
 598         int nvec;
 599         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 600
 601         mutex_lock(&con->sock_mutex);
 602
 603         if (con->sock == NULL) {
 604                 ret = -EAGAIN;
 605                 goto out_close;
 606         }
 607
 608         if (con->rx_page == NULL) {
 609                 /*
 610                  * This doesn't need to be atomic, but I think it should
 611                  * improve performance if it is.
 612                  */
 613                 con->rx_page = alloc_page(GFP_ATOMIC);
 614                 if (con->rx_page == NULL)
 615                         goto out_resched;
 616                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 617         }
 618
 619         /* Only SCTP needs these really */
 620         memset(&incmsg, 0, sizeof(incmsg));
 621         msg.msg_control = incmsg;
 622         msg.msg_controllen = sizeof(incmsg);
 623
 624         /*
 625          * iov[0] is the bit of the circular buffer between the current end
 626          * point (cb.base + cb.len) and the end of the buffer.
 627          */
 628         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 629         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 630         iov[1].iov_len = 0;
 631         nvec = 1;
 632
 633         /*
 634          * iov[1] is the bit of the circular buffer between the start of the
 635          * buffer and the start of the currently used section (cb.base)
 636          */
 637         if (cbuf_data(&con->cb) >= con->cb.base) {
 638                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 639                 iov[1].iov_len = con->cb.base;
 640                 iov[1].iov_base = page_address(con->rx_page);
 641                 nvec = 2;
 642         }
 643         len = iov[0].iov_len + iov[1].iov_len;
 644
 645         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 646                                MSG_DONTWAIT | MSG_NOSIGNAL);
 647         if (ret <= 0)
 648                 goto out_close;
 649
 650         /* Process SCTP notifications */
 651         if (msg.msg_flags & MSG_NOTIFICATION) {
 652                 msg.msg_control = incmsg;
 653                 msg.msg_controllen = sizeof(incmsg);
 654
 655                 process_sctp_notification(con, &msg,
 656                                 page_address(con->rx_page) + con->cb.base);
 657                 mutex_unlock(&con->sock_mutex);
 658                 return 0;
 659         }
 660         BUG_ON(con->nodeid == 0);
 661
 662         if (ret == len)
 663                 call_again_soon = 1;
 664         cbuf_add(&con->cb, ret);
 665         ret = dlm_process_incoming_buffer(con->nodeid,
 666                                           page_address(con->rx_page),
 667                                           con->cb.base, con->cb.len,
 668                                           PAGE_CACHE_SIZE);
 669         if (ret == -EBADMSG) {
 670                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 671                           "iov_len=%u, iov_base[0]=%p, read=%d",
 672                           page_address(con->rx_page), con->cb.base, con->cb.len,
 673                           len, iov[0].iov_base, r);
 674         }
 675         if (ret < 0)
 676                 goto out_close;
 677         cbuf_eat(&con->cb, ret);
 678
 679         if (cbuf_empty(&con->cb) && !call_again_soon) {
 680                 __free_page(con->rx_page);
 681                 con->rx_page = NULL;
 682         }
 683
 684         if (call_again_soon)
 685                 goto out_resched;
 686         mutex_unlock(&con->sock_mutex);
 687         return 0;
 688
 689 out_resched:
 690         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 691                 queue_work(recv_workqueue, &con->rwork);
 692         mutex_unlock(&con->sock_mutex);
 693         return -EAGAIN;
 694
 695 out_close:
 696         mutex_unlock(&con->sock_mutex);
 697         if (ret != -EAGAIN) {
 698                 close_connection(con, false);
 699                 /* Reconnect when there is something to send */
 700         }
 701         /* Don't return success if we really got EOF */
 702         if (ret == 0)
 703                 ret = -EAGAIN;
 704
 705         return ret;
 706 }
 707
 708 /* Listening socket is busy, accept a connection */
 709 static int tcp_accept_from_sock(struct connection *con)
 710 {
 711         int result;
 712         struct sockaddr_storage peeraddr;
 713         struct socket *newsock;
 714         int len;
 715         int nodeid;
 716         struct connection *newcon;
 717         struct connection *addcon;
 718
 719         memset(&peeraddr, 0, sizeof(peeraddr));
 720         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 721                                   IPPROTO_TCP, &newsock);
 722         if (result < 0)
 723                 return -ENOMEM;
 724
 725         mutex_lock_nested(&con->sock_mutex, 0);
 726
 727         result = -ENOTCONN;
 728         if (con->sock == NULL)
 729                 goto accept_err;
 730
 731         newsock->type = con->sock->type;
 732         newsock->ops = con->sock->ops;
 733
 734         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 735         if (result < 0)
 736                 goto accept_err;
 737
 738         /* Get the connected socket's peer */
 739         memset(&peeraddr, 0, sizeof(peeraddr));
 740         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 741                                   &len, 2)) {
 742                 result = -ECONNABORTED;
 743                 goto accept_err;
 744         }
 745
 746         /* Get the new node's NODEID */
 747         make_sockaddr(&peeraddr, 0, &len);
 748         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 749                 unsigned char *b=(unsigned char *)&peeraddr;
 750                 log_print("connect from non cluster node");
 751                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 752                                      b, sizeof(struct sockaddr_storage));
 753                 sock_release(newsock);
 754                 mutex_unlock(&con->sock_mutex);
 755                 return -1;
 756         }
 757
 758         log_print("got connection from %d", nodeid);
 759
 760         /*  Check to see if we already have a connection to this node. This
 761          *  could happen if the two nodes initiate a connection at roughly
 762          *  the same time and the connections cross on the wire.
 763          *  In this case we store the incoming one in "othercon"
 764          */
 765         newcon = nodeid2con(nodeid, GFP_NOFS);
 766         if (!newcon) {
 767                 result = -ENOMEM;
 768                 goto accept_err;
 769         }
 770         mutex_lock_nested(&newcon->sock_mutex, 1);
 771         if (newcon->sock) {
 772                 struct connection *othercon = newcon->othercon;
 773
 774                 if (!othercon) {
 775                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 776                         if (!othercon) {
 777                                 log_print("failed to allocate incoming socket");
 778                                 mutex_unlock(&newcon->sock_mutex);
 779                                 result = -ENOMEM;
 780                                 goto accept_err;
 781                         }
 782                         othercon->nodeid = nodeid;
 783                         othercon->rx_action = receive_from_sock;
 784                         mutex_init(&othercon->sock_mutex);
 785                         INIT_WORK(&othercon->swork, process_send_sockets);
 786                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 787                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 788                 }
 789                 if (!othercon->sock) {
 790                         newcon->othercon = othercon;
 791                         othercon->sock = newsock;
 792                         newsock->sk->sk_user_data = othercon;
 793                         add_sock(newsock, othercon);
 794                         addcon = othercon;
 795                 }
 796                 else {
 797                         printk("Extra connection from node %d attempted\n", nodeid);
 798                         result = -EAGAIN;
 799                         mutex_unlock(&newcon->sock_mutex);
 800                         goto accept_err;
 801                 }
 802         }
 803         else {
 804                 newsock->sk->sk_user_data = newcon;
 805                 newcon->rx_action = receive_from_sock;
 806                 add_sock(newsock, newcon);
 807                 addcon = newcon;
 808         }
 809
 810         mutex_unlock(&newcon->sock_mutex);
 811
 812         /*
 813          * Add it to the active queue in case we got data
 814          * between processing the accept adding the socket
 815          * to the read_sockets list
 816          */
 817         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 818                 queue_work(recv_workqueue, &addcon->rwork);
 819         mutex_unlock(&con->sock_mutex);
 820
 821         return 0;
 822
 823 accept_err:
 824         mutex_unlock(&con->sock_mutex);
 825         sock_release(newsock);
 826
 827         if (result != -EAGAIN)
 828                 log_print("error accepting connection from node: %d", result);
 829         return result;
 830 }
 831
 832 static void free_entry(struct writequeue_entry *e)
 833 {
 834         __free_page(e->page);
 835         kfree(e);
 836 }
 837
 838 /* Initiate an SCTP association.
 839    This is a special case of send_to_sock() in that we don't yet have a
 840    peeled-off socket for this association, so we use the listening socket
 841    and add the primary IP address of the remote node.
 842  */
 843 static void sctp_init_assoc(struct connection *con)
 844 {
 845         struct sockaddr_storage rem_addr;
 846         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 847         struct msghdr outmessage;
 848         struct cmsghdr *cmsg;
 849         struct sctp_sndrcvinfo *sinfo;
 850         struct connection *base_con;
 851         struct writequeue_entry *e;
 852         int len, offset;
 853         int ret;
 854         int addrlen;
 855         struct kvec iov[1];
 856
 857         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 858                 return;
 859
 860         if (con->retries++ > MAX_CONNECT_RETRIES)
 861                 return;
 862
 863         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 864                 log_print("no address for nodeid %d", con->nodeid);
 865                 return;
 866         }
 867         base_con = nodeid2con(0, 0);
 868         BUG_ON(base_con == NULL);
 869
 870         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 871
 872         outmessage.msg_name = &rem_addr;
 873         outmessage.msg_namelen = addrlen;
 874         outmessage.msg_control = outcmsg;
 875         outmessage.msg_controllen = sizeof(outcmsg);
 876         outmessage.msg_flags = MSG_EOR;
 877
 878         spin_lock(&con->writequeue_lock);
 879
 880         if (list_empty(&con->writequeue)) {
 881                 spin_unlock(&con->writequeue_lock);
 882                 log_print("writequeue empty for nodeid %d", con->nodeid);
 883                 return;
 884         }
 885
 886         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
 887         len = e->len;
 888         offset = e->offset;
 889         spin_unlock(&con->writequeue_lock);
 890
 891         /* Send the first block off the write queue */
 892         iov[0].iov_base = page_address(e->page)+offset;
 893         iov[0].iov_len = len;
 894
 895         cmsg = CMSG_FIRSTHDR(&outmessage);
 896         cmsg->cmsg_level = IPPROTO_SCTP;
 897         cmsg->cmsg_type = SCTP_SNDRCV;
 898         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 899         sinfo = CMSG_DATA(cmsg);
 900         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 901         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 902         outmessage.msg_controllen = cmsg->cmsg_len;
 903
 904         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 905         if (ret < 0) {
 906                 log_print("Send first packet to node %d failed: %d",
 907                           con->nodeid, ret);
 908
 909                 /* Try again later */
 910                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 911                 clear_bit(CF_INIT_PENDING, &con->flags);
 912         }
 913         else {
 914                 spin_lock(&con->writequeue_lock);
 915                 e->offset += ret;
 916                 e->len -= ret;
 917
 918                 if (e->len == 0 && e->users == 0) {
 919                         list_del(&e->list);
 920                         free_entry(e);
 921                 }
 922                 spin_unlock(&con->writequeue_lock);
 923         }
 924 }
 925
 926 /* Connect a new socket to its peer */
 927 static void tcp_connect_to_sock(struct connection *con)
 928 {
 929         int result = -EHOSTUNREACH;
 930         struct sockaddr_storage saddr, src_addr;
 931         int addr_len;
 932         struct socket *sock = NULL;
 933         int one = 1;
 934
 935         if (con->nodeid == 0) {
 936                 log_print("attempt to connect sock 0 foiled");
 937                 return;
 938         }
 939
 940         mutex_lock(&con->sock_mutex);
 941         if (con->retries++ > MAX_CONNECT_RETRIES)
 942                 goto out;
 943
 944         /* Some odd races can cause double-connects, ignore them */
 945         if (con->sock) {
 946                 result = 0;
 947                 goto out;
 948         }
 949
 950         /* Create a socket to communicate with */
 951         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 952                                   IPPROTO_TCP, &sock);
 953         if (result < 0)
 954                 goto out_err;
 955
 956         memset(&saddr, 0, sizeof(saddr));
 957         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 958                 goto out_err;
 959
 960         sock->sk->sk_user_data = con;
 961         con->rx_action = receive_from_sock;
 962         con->connect_action = tcp_connect_to_sock;
 963         add_sock(sock, con);
 964
 965         /* Bind to our cluster-known address connecting to avoid
 966            routing problems */
 967         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 968         make_sockaddr(&src_addr, 0, &addr_len);
 969         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 970                                  addr_len);
 971         if (result < 0) {
 972                 log_print("could not bind for connect: %d", result);
 973                 /* This *may* not indicate a critical error */
 974         }
 975
 976         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 977
 978         log_print("connecting to %d", con->nodeid);
 979
 980         /* Turn off Nagle's algorithm */
 981         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
 982                           sizeof(one));
 983
 984         result =
 985                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 986                                    O_NONBLOCK);
 987         if (result == -EINPROGRESS)
 988                 result = 0;
 989         if (result == 0)
 990                 goto out;
 991
 992 out_err:
 993         if (con->sock) {
 994                 sock_release(con->sock);
 995                 con->sock = NULL;
 996         } else if (sock) {
 997                 sock_release(sock);
 998         }
 999         /*
1000          * Some errors are fatal and this list might need adjusting. For other
1001          * errors we try again until the max number of retries is reached.
1002          */
1003         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
1004             result != -ENETDOWN && result != -EINVAL
1005             && result != -EPROTONOSUPPORT) {
1006                 lowcomms_connect_sock(con);
1007                 result = 0;
1008         }
1009 out:
1010         mutex_unlock(&con->sock_mutex);
1011         return;
1012 }
1013
1014 static struct socket *tcp_create_listen_sock(struct connection *con,
1015                                              struct sockaddr_storage *saddr)
1016 {
1017         struct socket *sock = NULL;
1018         int result = 0;
1019         int one = 1;
1020         int addr_len;
1021
1022         if (dlm_local_addr[0]->ss_family == AF_INET)
1023                 addr_len = sizeof(struct sockaddr_in);
1024         else
1025                 addr_len = sizeof(struct sockaddr_in6);
1026
1027         /* Create a socket to communicate with */
1028         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1029                                   IPPROTO_TCP, &sock);
1030         if (result < 0) {
1031                 log_print("Can't create listening comms socket");
1032                 goto create_out;
1033         }
1034
1035         /* Turn off Nagle's algorithm */
1036         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1037                           sizeof(one));
1038
1039         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1040                                    (char *)&one, sizeof(one));
1041
1042         if (result < 0) {
1043                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1044         }
1045         sock->sk->sk_user_data = con;
1046         con->rx_action = tcp_accept_from_sock;
1047         con->connect_action = tcp_connect_to_sock;
1048         con->sock = sock;
1049
1050         /* Bind to our port */
1051         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1052         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1053         if (result < 0) {
1054                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1055                 sock_release(sock);
1056                 sock = NULL;
1057                 con->sock = NULL;
1058                 goto create_out;
1059         }
1060         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1061                                  (char *)&one, sizeof(one));
1062         if (result < 0) {
1063                 log_print("Set keepalive failed: %d", result);
1064         }
1065
1066         result = sock->ops->listen(sock, 5);
1067         if (result < 0) {
1068                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1069                 sock_release(sock);
1070                 sock = NULL;
1071                 goto create_out;
1072         }
1073
1074 create_out:
1075         return sock;
1076 }
1077
1078 /* Get local addresses */
1079 static void init_local(void)
1080 {
1081         struct sockaddr_storage sas, *addr;
1082         int i;
1083
1084         dlm_local_count = 0;
1085         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1086                 if (dlm_our_addr(&sas, i))
1087                         break;
1088
1089                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1090                 if (!addr)
1091                         break;
1092                 memcpy(addr, &sas, sizeof(*addr));
1093                 dlm_local_addr[dlm_local_count++] = addr;
1094         }
1095 }
1096
1097 /* Bind to an IP address. SCTP allows multiple address so it can do
1098    multi-homing */
1099 static int add_sctp_bind_addr(struct connection *sctp_con,
1100                               struct sockaddr_storage *addr,
1101                               int addr_len, int num)
1102 {
1103         int result = 0;
1104
1105         if (num == 1)
1106                 result = kernel_bind(sctp_con->sock,
1107                                      (struct sockaddr *) addr,
1108                                      addr_len);
1109         else
1110                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1111                                            SCTP_SOCKOPT_BINDX_ADD,
1112                                            (char *)addr, addr_len);
1113
1114         if (result < 0)
1115                 log_print("Can't bind to port %d addr number %d",
1116                           dlm_config.ci_tcp_port, num);
1117
1118         return result;
1119 }
1120
1121 /* Initialise SCTP socket and bind to all interfaces */
1122 static int sctp_listen_for_all(void)
1123 {
1124         struct socket *sock = NULL;
1125         struct sockaddr_storage localaddr;
1126         struct sctp_event_subscribe subscribe;
1127         int result = -EINVAL, num = 1, i, addr_len;
1128         struct connection *con = nodeid2con(0, GFP_NOFS);
1129         int bufsize = NEEDED_RMEM;
1130
1131         if (!con)
1132                 return -ENOMEM;
1133
1134         log_print("Using SCTP for communications");
1135
1136         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1137                                   IPPROTO_SCTP, &sock);
1138         if (result < 0) {
1139                 log_print("Can't create comms socket, check SCTP is loaded");
1140                 goto out;
1141         }
1142
1143         /* Listen for events */
1144         memset(&subscribe, 0, sizeof(subscribe));
1145         subscribe.sctp_data_io_event = 1;
1146         subscribe.sctp_association_event = 1;
1147         subscribe.sctp_send_failure_event = 1;
1148         subscribe.sctp_shutdown_event = 1;
1149         subscribe.sctp_partial_delivery_event = 1;
1150
1151         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1152                                  (char *)&bufsize, sizeof(bufsize));
1153         if (result)
1154                 log_print("Error increasing buffer space on socket %d", result);
1155
1156         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1157                                    (char *)&subscribe, sizeof(subscribe));
1158         if (result < 0) {
1159                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1160                           result);
1161                 goto create_delsock;
1162         }
1163
1164         /* Init con struct */
1165         sock->sk->sk_user_data = con;
1166         con->sock = sock;
1167         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1168         con->rx_action = receive_from_sock;
1169         con->connect_action = sctp_init_assoc;
1170
1171         /* Bind to all interfaces. */
1172         for (i = 0; i < dlm_local_count; i++) {
1173                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1174                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1175
1176                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1177                 if (result)
1178                         goto create_delsock;
1179                 ++num;
1180         }
1181
1182         result = sock->ops->listen(sock, 5);
1183         if (result < 0) {
1184                 log_print("Can't set socket listening");
1185                 goto create_delsock;
1186         }
1187
1188         return 0;
1189
1190 create_delsock:
1191         sock_release(sock);
1192         con->sock = NULL;
1193 out:
1194         return result;
1195 }
1196
1197 static int tcp_listen_for_all(void)
1198 {
1199         struct socket *sock = NULL;
1200         struct connection *con = nodeid2con(0, GFP_NOFS);
1201         int result = -EINVAL;
1202
1203         if (!con)
1204                 return -ENOMEM;
1205
1206         /* We don't support multi-homed hosts */
1207         if (dlm_local_addr[1] != NULL) {
1208                 log_print("TCP protocol can't handle multi-homed hosts, "
1209                           "try SCTP");
1210                 return -EINVAL;
1211         }
1212
1213         log_print("Using TCP for communications");
1214
1215         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1216         if (sock) {
1217                 add_sock(sock, con);
1218                 result = 0;
1219         }
1220         else {
1221                 result = -EADDRINUSE;
1222         }
1223
1224         return result;
1225 }
1226
1227
1228
1229 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1230                                                      gfp_t allocation)
1231 {
1232         struct writequeue_entry *entry;
1233
1234         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1235         if (!entry)
1236                 return NULL;
1237
1238         entry->page = alloc_page(allocation);
1239         if (!entry->page) {
1240                 kfree(entry);
1241                 return NULL;
1242         }
1243
1244         entry->offset = 0;
1245         entry->len = 0;
1246         entry->end = 0;
1247         entry->users = 0;
1248         entry->con = con;
1249
1250         return entry;
1251 }
1252
1253 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1254 {
1255         struct connection *con;
1256         struct writequeue_entry *e;
1257         int offset = 0;
1258         int users = 0;
1259
1260         con = nodeid2con(nodeid, allocation);
1261         if (!con)
1262                 return NULL;
1263
1264         spin_lock(&con->writequeue_lock);
1265         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1266         if ((&e->list == &con->writequeue) ||
1267             (PAGE_CACHE_SIZE - e->end < len)) {
1268                 e = NULL;
1269         } else {
1270                 offset = e->end;
1271                 e->end += len;
1272                 users = e->users++;
1273         }
1274         spin_unlock(&con->writequeue_lock);
1275
1276         if (e) {
1277         got_one:
1278                 *ppc = page_address(e->page) + offset;
1279                 return e;
1280         }
1281
1282         e = new_writequeue_entry(con, allocation);
1283         if (e) {
1284                 spin_lock(&con->writequeue_lock);
1285                 offset = e->end;
1286                 e->end += len;
1287                 users = e->users++;
1288                 list_add_tail(&e->list, &con->writequeue);
1289                 spin_unlock(&con->writequeue_lock);
1290                 goto got_one;
1291         }
1292         return NULL;
1293 }
1294
1295 void dlm_lowcomms_commit_buffer(void *mh)
1296 {
1297         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1298         struct connection *con = e->con;
1299         int users;
1300
1301         spin_lock(&con->writequeue_lock);
1302         users = --e->users;
1303         if (users)
1304                 goto out;
1305         e->len = e->end - e->offset;
1306         spin_unlock(&con->writequeue_lock);
1307
1308         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1309                 queue_work(send_workqueue, &con->swork);
1310         }
1311         return;
1312
1313 out:
1314         spin_unlock(&con->writequeue_lock);
1315         return;
1316 }
1317
1318 /* Send a message */
1319 static void send_to_sock(struct connection *con)
1320 {
1321         int ret = 0;
1322         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1323         struct writequeue_entry *e;
1324         int len, offset;
1325         int count = 0;
1326
1327         mutex_lock(&con->sock_mutex);
1328         if (con->sock == NULL)
1329                 goto out_connect;
1330
1331         spin_lock(&con->writequeue_lock);
1332         for (;;) {
1333                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1334                                list);
1335                 if ((struct list_head *) e == &con->writequeue)
1336                         break;
1337
1338                 len = e->len;
1339                 offset = e->offset;
1340                 BUG_ON(len == 0 && e->users == 0);
1341                 spin_unlock(&con->writequeue_lock);
1342
1343                 ret = 0;
1344                 if (len) {
1345                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1346                                               msg_flags);
1347                         if (ret == -EAGAIN || ret == 0) {
1348                                 if (ret == -EAGAIN &&
1349                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1350                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1351                                         /* Notify TCP that we're limited by the
1352                                          * application window size.
1353                                          */
1354                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1355                                         con->sock->sk->sk_write_pending++;
1356                                 }
1357                                 cond_resched();
1358                                 goto out;
1359                         }
1360                         if (ret <= 0)
1361                                 goto send_error;
1362                 }
1363
1364                 /* Don't starve people filling buffers */
1365                 if (++count >= MAX_SEND_MSG_COUNT) {
1366                         cond_resched();
1367                         count = 0;
1368                 }
1369
1370                 spin_lock(&con->writequeue_lock);
1371                 e->offset += ret;
1372                 e->len -= ret;
1373
1374                 if (e->len == 0 && e->users == 0) {
1375                         list_del(&e->list);
1376                         free_entry(e);
1377                         continue;
1378                 }
1379         }
1380         spin_unlock(&con->writequeue_lock);
1381 out:
1382         mutex_unlock(&con->sock_mutex);
1383         return;
1384
1385 send_error:
1386         mutex_unlock(&con->sock_mutex);
1387         close_connection(con, false);
1388         lowcomms_connect_sock(con);
1389         return;
1390
1391 out_connect:
1392         mutex_unlock(&con->sock_mutex);
1393         if (!test_bit(CF_INIT_PENDING, &con->flags))
1394                 lowcomms_connect_sock(con);
1395         return;
1396 }
1397
1398 static void clean_one_writequeue(struct connection *con)
1399 {
1400         struct writequeue_entry *e, *safe;
1401
1402         spin_lock(&con->writequeue_lock);
1403         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1404                 list_del(&e->list);
1405                 free_entry(e);
1406         }
1407         spin_unlock(&con->writequeue_lock);
1408 }
1409
1410 /* Called from recovery when it knows that a node has
1411    left the cluster */
1412 int dlm_lowcomms_close(int nodeid)
1413 {
1414         struct connection *con;
1415
1416         log_print("closing connection to node %d", nodeid);
1417         con = nodeid2con(nodeid, 0);
1418         if (con) {
1419                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1420                 clear_bit(CF_WRITE_PENDING, &con->flags);
1421                 set_bit(CF_CLOSE, &con->flags);
1422                 if (cancel_work_sync(&con->swork))
1423                         log_print("canceled swork for node %d", nodeid);
1424                 if (cancel_work_sync(&con->rwork))
1425                         log_print("canceled rwork for node %d", nodeid);
1426                 clean_one_writequeue(con);
1427                 close_connection(con, true);
1428         }
1429         return 0;
1430 }
1431
1432 /* Receive workqueue function */
1433 static void process_recv_sockets(struct work_struct *work)
1434 {
1435         struct connection *con = container_of(work, struct connection, rwork);
1436         int err;
1437
1438         clear_bit(CF_READ_PENDING, &con->flags);
1439         do {
1440                 err = con->rx_action(con);
1441         } while (!err);
1442 }
1443
1444 /* Send workqueue function */
1445 static void process_send_sockets(struct work_struct *work)
1446 {
1447         struct connection *con = container_of(work, struct connection, swork);
1448
1449         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1450                 con->connect_action(con);
1451                 set_bit(CF_WRITE_PENDING, &con->flags);
1452         }
1453         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1454                 send_to_sock(con);
1455 }
1456
1457
1458 /* Discard all entries on the write queues */
1459 static void clean_writequeues(void)
1460 {
1461         foreach_conn(clean_one_writequeue);
1462 }
1463
1464 static void work_stop(void)
1465 {
1466         destroy_workqueue(recv_workqueue);
1467         destroy_workqueue(send_workqueue);
1468 }
1469
1470 static int work_start(void)
1471 {
1472         recv_workqueue = alloc_workqueue("dlm_recv",
1473                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1474         if (!recv_workqueue) {
1475                 log_print("can't start dlm_recv");
1476                 return -ENOMEM;
1477         }
1478
1479         send_workqueue = alloc_workqueue("dlm_send",
1480                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1481         if (!send_workqueue) {
1482                 log_print("can't start dlm_send");
1483                 destroy_workqueue(recv_workqueue);
1484                 return -ENOMEM;
1485         }
1486
1487         return 0;
1488 }
1489
1490 static void stop_conn(struct connection *con)
1491 {
1492         con->flags |= 0x0F;
1493         if (con->sock && con->sock->sk)
1494                 con->sock->sk->sk_user_data = NULL;
1495 }
1496
1497 static void free_conn(struct connection *con)
1498 {
1499         close_connection(con, true);
1500         if (con->othercon)
1501                 kmem_cache_free(con_cache, con->othercon);
1502         hlist_del(&con->list);
1503         kmem_cache_free(con_cache, con);
1504 }
1505
1506 void dlm_lowcomms_stop(void)
1507 {
1508         /* Set all the flags to prevent any
1509            socket activity.
1510         */
1511         mutex_lock(&connections_lock);
1512         foreach_conn(stop_conn);
1513         mutex_unlock(&connections_lock);
1514
1515         work_stop();
1516
1517         mutex_lock(&connections_lock);
1518         clean_writequeues();
1519
1520         foreach_conn(free_conn);
1521
1522         mutex_unlock(&connections_lock);
1523         kmem_cache_destroy(con_cache);
1524 }
1525
1526 int dlm_lowcomms_start(void)
1527 {
1528         int error = -EINVAL;
1529         struct connection *con;
1530         int i;
1531
1532         for (i = 0; i < CONN_HASH_SIZE; i++)
1533                 INIT_HLIST_HEAD(&connection_hash[i]);
1534
1535         init_local();
1536         if (!dlm_local_count) {
1537                 error = -ENOTCONN;
1538                 log_print("no local IP address has been set");
1539                 goto out;
1540         }
1541
1542         error = -ENOMEM;
1543         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1544                                       __alignof__(struct connection), 0,
1545                                       NULL);
1546         if (!con_cache)
1547                 goto out;
1548
1549         /* Start listening */
1550         if (dlm_config.ci_protocol == 0)
1551                 error = tcp_listen_for_all();
1552         else
1553                 error = sctp_listen_for_all();
1554         if (error)
1555                 goto fail_unlisten;
1556
1557         error = work_start();
1558         if (error)
1559                 goto fail_unlisten;
1560
1561         return 0;
1562
1563 fail_unlisten:
1564         con = nodeid2con(0,0);
1565         if (con) {
1566                 close_connection(con, false);
1567                 kmem_cache_free(con_cache, con);
1568         }
1569         kmem_cache_destroy(con_cache);
1570
1571 out:
1572         return error;
1573 }