fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2007 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 it's
  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 wither 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 shouldbe 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/idr.h>
  52 #include <linux/file.h>
  53 #include <linux/sctp.h>
  54 #include <net/sctp/user.h>
  55
  56 #include "dlm_internal.h"
  57 #include "lowcomms.h"
  58 #include "midcomms.h"
  59 #include "config.h"
  60
  61 #define NEEDED_RMEM (4*1024*1024)
  62
  63 struct cbuf {
  64         unsigned int base;
  65         unsigned int len;
  66         unsigned int mask;
  67 };
  68
  69 static void cbuf_add(struct cbuf *cb, int n)
  70 {
  71         cb->len += n;
  72 }
  73
  74 static int cbuf_data(struct cbuf *cb)
  75 {
  76         return ((cb->base + cb->len) & cb->mask);
  77 }
  78
  79 static void cbuf_init(struct cbuf *cb, int size)
  80 {
  81         cb->base = cb->len = 0;
  82         cb->mask = size-1;
  83 }
  84
  85 static void cbuf_eat(struct cbuf *cb, int n)
  86 {
  87         cb->len  -= n;
  88         cb->base += n;
  89         cb->base &= cb->mask;
  90 }
  91
  92 static bool cbuf_empty(struct cbuf *cb)
  93 {
  94         return cb->len == 0;
  95 }
  96
  97 struct connection {
  98         struct socket *sock;    /* NULL if not connected */
  99         uint32_t nodeid;        /* So we know who we are in the list */
 100         struct mutex sock_mutex;
 101         unsigned long flags;
 102 #define CF_READ_PENDING 1
 103 #define CF_WRITE_PENDING 2
 104 #define CF_CONNECT_PENDING 3
 105 #define CF_INIT_PENDING 4
 106 #define CF_IS_OTHERCON 5
 107         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 108         spinlock_t writequeue_lock;
 109         int (*rx_action) (struct connection *); /* What to do when active */
 110         void (*connect_action) (struct connection *);   /* What to do to connect */
 111         struct page *rx_page;
 112         struct cbuf cb;
 113         int retries;
 114 #define MAX_CONNECT_RETRIES 3
 115         int sctp_assoc;
 116         struct connection *othercon;
 117         struct work_struct rwork; /* Receive workqueue */
 118         struct work_struct swork; /* Send workqueue */
 119 };
 120 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 121
 122 /* An entry waiting to be sent */
 123 struct writequeue_entry {
 124         struct list_head list;
 125         struct page *page;
 126         int offset;
 127         int len;
 128         int end;
 129         int users;
 130         struct connection *con;
 131 };
 132
 133 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 134 static int dlm_local_count;
 135
 136 /* Work queues */
 137 static struct workqueue_struct *recv_workqueue;
 138 static struct workqueue_struct *send_workqueue;
 139
 140 static DEFINE_IDR(connections_idr);
 141 static DECLARE_MUTEX(connections_lock);
 142 static int max_nodeid;
 143 static struct kmem_cache *con_cache;
 144
 145 static void process_recv_sockets(struct work_struct *work);
 146 static void process_send_sockets(struct work_struct *work);
 147
 148 /*
 149  * If 'allocation' is zero then we don't attempt to create a new
 150  * connection structure for this node.
 151  */
 152 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 153 {
 154         struct connection *con = NULL;
 155         int r;
 156         int n;
 157
 158         con = idr_find(&connections_idr, nodeid);
 159         if (con || !alloc)
 160                 return con;
 161
 162         r = idr_pre_get(&connections_idr, alloc);
 163         if (!r)
 164                 return NULL;
 165
 166         con = kmem_cache_zalloc(con_cache, alloc);
 167         if (!con)
 168                 return NULL;
 169
 170         r = idr_get_new_above(&connections_idr, con, nodeid, &n);
 171         if (r) {
 172                 kmem_cache_free(con_cache, con);
 173                 return NULL;
 174         }
 175
 176         if (n != nodeid) {
 177                 idr_remove(&connections_idr, n);
 178                 kmem_cache_free(con_cache, con);
 179                 return NULL;
 180         }
 181
 182         con->nodeid = nodeid;
 183         mutex_init(&con->sock_mutex);
 184         INIT_LIST_HEAD(&con->writequeue);
 185         spin_lock_init(&con->writequeue_lock);
 186         INIT_WORK(&con->swork, process_send_sockets);
 187         INIT_WORK(&con->rwork, process_recv_sockets);
 188
 189         /* Setup action pointers for child sockets */
 190         if (con->nodeid) {
 191                 struct connection *zerocon = idr_find(&connections_idr, 0);
 192
 193                 con->connect_action = zerocon->connect_action;
 194                 if (!con->rx_action)
 195                         con->rx_action = zerocon->rx_action;
 196         }
 197
 198         if (nodeid > max_nodeid)
 199                 max_nodeid = nodeid;
 200
 201         return con;
 202 }
 203
 204 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 205 {
 206         struct connection *con;
 207
 208         down(&connections_lock);
 209         con = __nodeid2con(nodeid, allocation);
 210         up(&connections_lock);
 211
 212         return con;
 213 }
 214
 215 /* This is a bit drastic, but only called when things go wrong */
 216 static struct connection *assoc2con(int assoc_id)
 217 {
 218         int i;
 219         struct connection *con;
 220
 221         down(&connections_lock);
 222         for (i=0; i<=max_nodeid; i++) {
 223                 con = __nodeid2con(i, 0);
 224                 if (con && con->sctp_assoc == assoc_id) {
 225                         up(&connections_lock);
 226                         return con;
 227                 }
 228         }
 229         up(&connections_lock);
 230         return NULL;
 231 }
 232
 233 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 234 {
 235         struct sockaddr_storage addr;
 236         int error;
 237
 238         if (!dlm_local_count)
 239                 return -1;
 240
 241         error = dlm_nodeid_to_addr(nodeid, &addr);
 242         if (error)
 243                 return error;
 244
 245         if (dlm_local_addr[0]->ss_family == AF_INET) {
 246                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 247                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 248                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 249         } else {
 250                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 251                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 252                 memcpy(&ret6->sin6_addr, &in6->sin6_addr,
 253                        sizeof(in6->sin6_addr));
 254         }
 255
 256         return 0;
 257 }
 258
 259 /* Data available on socket or listen socket received a connect */
 260 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 261 {
 262         struct connection *con = sock2con(sk);
 263         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 264                 queue_work(recv_workqueue, &con->rwork);
 265 }
 266
 267 static void lowcomms_write_space(struct sock *sk)
 268 {
 269         struct connection *con = sock2con(sk);
 270
 271         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 272                 queue_work(send_workqueue, &con->swork);
 273 }
 274
 275 static inline void lowcomms_connect_sock(struct connection *con)
 276 {
 277         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 278                 queue_work(send_workqueue, &con->swork);
 279 }
 280
 281 static void lowcomms_state_change(struct sock *sk)
 282 {
 283         if (sk->sk_state == TCP_ESTABLISHED)
 284                 lowcomms_write_space(sk);
 285 }
 286
 287 /* Make a socket active */
 288 static int add_sock(struct socket *sock, struct connection *con)
 289 {
 290         con->sock = sock;
 291
 292         /* Install a data_ready callback */
 293         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 294         con->sock->sk->sk_write_space = lowcomms_write_space;
 295         con->sock->sk->sk_state_change = lowcomms_state_change;
 296         con->sock->sk->sk_user_data = con;
 297         return 0;
 298 }
 299
 300 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 301    length */
 302 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 303                           int *addr_len)
 304 {
 305         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 306         if (saddr->ss_family == AF_INET) {
 307                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 308                 in4_addr->sin_port = cpu_to_be16(port);
 309                 *addr_len = sizeof(struct sockaddr_in);
 310                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 311         } else {
 312                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 313                 in6_addr->sin6_port = cpu_to_be16(port);
 314                 *addr_len = sizeof(struct sockaddr_in6);
 315         }
 316 }
 317
 318 /* Close a remote connection and tidy up */
 319 static void close_connection(struct connection *con, bool and_other)
 320 {
 321         mutex_lock(&con->sock_mutex);
 322
 323         if (con->sock) {
 324                 sock_release(con->sock);
 325                 con->sock = NULL;
 326         }
 327         if (con->othercon && and_other) {
 328                 /* Will only re-enter once. */
 329                 close_connection(con->othercon, false);
 330         }
 331         if (con->rx_page) {
 332                 __free_page(con->rx_page);
 333                 con->rx_page = NULL;
 334         }
 335         con->retries = 0;
 336         mutex_unlock(&con->sock_mutex);
 337 }
 338
 339 /* We only send shutdown messages to nodes that are not part of the cluster */
 340 static void sctp_send_shutdown(sctp_assoc_t associd)
 341 {
 342         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 343         struct msghdr outmessage;
 344         struct cmsghdr *cmsg;
 345         struct sctp_sndrcvinfo *sinfo;
 346         int ret;
 347         struct connection *con;
 348
 349         con = nodeid2con(0,0);
 350         BUG_ON(con == NULL);
 351
 352         outmessage.msg_name = NULL;
 353         outmessage.msg_namelen = 0;
 354         outmessage.msg_control = outcmsg;
 355         outmessage.msg_controllen = sizeof(outcmsg);
 356         outmessage.msg_flags = MSG_EOR;
 357
 358         cmsg = CMSG_FIRSTHDR(&outmessage);
 359         cmsg->cmsg_level = IPPROTO_SCTP;
 360         cmsg->cmsg_type = SCTP_SNDRCV;
 361         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 362         outmessage.msg_controllen = cmsg->cmsg_len;
 363         sinfo = CMSG_DATA(cmsg);
 364         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 365
 366         sinfo->sinfo_flags |= MSG_EOF;
 367         sinfo->sinfo_assoc_id = associd;
 368
 369         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 370
 371         if (ret != 0)
 372                 log_print("send EOF to node failed: %d", ret);
 373 }
 374
 375 /* INIT failed but we don't know which node...
 376    restart INIT on all pending nodes */
 377 static void sctp_init_failed(void)
 378 {
 379         int i;
 380         struct connection *con;
 381
 382         down(&connections_lock);
 383         for (i=1; i<=max_nodeid; i++) {
 384                 con = __nodeid2con(i, 0);
 385                 if (!con)
 386                         continue;
 387                 con->sctp_assoc = 0;
 388                 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 389                         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
 390                                 queue_work(send_workqueue, &con->swork);
 391                         }
 392                 }
 393         }
 394         up(&connections_lock);
 395 }
 396
 397 /* Something happened to an association */
 398 static void process_sctp_notification(struct connection *con, struct msghdr *msg, char *buf)
 399 {
 400         union sctp_notification *sn = (union sctp_notification *)buf;
 401
 402         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 403                 switch (sn->sn_assoc_change.sac_state) {
 404
 405                 case SCTP_COMM_UP:
 406                 case SCTP_RESTART:
 407                 {
 408                         /* Check that the new node is in the lockspace */
 409                         struct sctp_prim prim;
 410                         int nodeid;
 411                         int prim_len, ret;
 412                         int addr_len;
 413                         struct connection *new_con;
 414                         struct file *file;
 415                         sctp_peeloff_arg_t parg;
 416                         int parglen = sizeof(parg);
 417
 418                         /*
 419                          * We get this before any data for an association.
 420                          * We verify that the node is in the cluster and
 421                          * then peel off a socket for it.
 422                          */
 423                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 424                                 log_print("COMM_UP for invalid assoc ID %d",
 425                                           (int)sn->sn_assoc_change.sac_assoc_id);
 426                                 sctp_init_failed();
 427                                 return;
 428                         }
 429                         memset(&prim, 0, sizeof(struct sctp_prim));
 430                         prim_len = sizeof(struct sctp_prim);
 431                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 432
 433                         ret = kernel_getsockopt(con->sock,
 434                                                 IPPROTO_SCTP,
 435                                                 SCTP_PRIMARY_ADDR,
 436                                                 (char*)&prim,
 437                                                 &prim_len);
 438                         if (ret < 0) {
 439                                 log_print("getsockopt/sctp_primary_addr on "
 440                                           "new assoc %d failed : %d",
 441                                           (int)sn->sn_assoc_change.sac_assoc_id,
 442                                           ret);
 443
 444                                 /* Retry INIT later */
 445                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 446                                 if (new_con)
 447                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 448                                 return;
 449                         }
 450                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 451                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 452                                 int i;
 453                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 454                                 log_print("reject connect from unknown addr");
 455                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 456                                         printk("%02x ", b[i]);
 457                                 printk("\n");
 458                                 sctp_send_shutdown(prim.ssp_assoc_id);
 459                                 return;
 460                         }
 461
 462                         new_con = nodeid2con(nodeid, GFP_KERNEL);
 463                         if (!new_con)
 464                                 return;
 465
 466                         /* Peel off a new sock */
 467                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 468                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP, SCTP_SOCKOPT_PEELOFF,
 469                                                 (void *)&parg, &parglen);
 470                         if (ret) {
 471                                 log_print("Can't peel off a socket for connection %d to node %d: err=%d\n",
 472                                           parg.associd, nodeid, ret);
 473                         }
 474                         file = fget(parg.sd);
 475                         new_con->sock = SOCKET_I(file->f_dentry->d_inode);
 476                         add_sock(new_con->sock, new_con);
 477                         fput(file);
 478                         put_unused_fd(parg.sd);
 479
 480                         log_print("got new/restarted association %d nodeid %d",
 481                                   (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
 482
 483                         /* Send any pending writes */
 484                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 485                         clear_bit(CF_INIT_PENDING, &con->flags);
 486                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 487                                 queue_work(send_workqueue, &new_con->swork);
 488                         }
 489                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 490                                 queue_work(recv_workqueue, &new_con->rwork);
 491                 }
 492                 break;
 493
 494                 case SCTP_COMM_LOST:
 495                 case SCTP_SHUTDOWN_COMP:
 496                 {
 497                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 498                         if (con) {
 499                                 con->sctp_assoc = 0;
 500                         }
 501                 }
 502                 break;
 503
 504                 /* We don't know which INIT failed, so clear the PENDING flags
 505                  * on them all.  if assoc_id is zero then it will then try
 506                  * again */
 507
 508                 case SCTP_CANT_STR_ASSOC:
 509                 {
 510                         log_print("Can't start SCTP association - retrying");
 511                         sctp_init_failed();
 512                 }
 513                 break;
 514
 515                 default:
 516                         log_print("unexpected SCTP assoc change id=%d state=%d",
 517                                   (int)sn->sn_assoc_change.sac_assoc_id,
 518                                   sn->sn_assoc_change.sac_state);
 519                 }
 520         }
 521 }
 522
 523 /* Data received from remote end */
 524 static int receive_from_sock(struct connection *con)
 525 {
 526         int ret = 0;
 527         struct msghdr msg = {};
 528         struct kvec iov[2];
 529         unsigned len;
 530         int r;
 531         int call_again_soon = 0;
 532         int nvec;
 533         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 534
 535         mutex_lock(&con->sock_mutex);
 536
 537         if (con->sock == NULL) {
 538                 ret = -EAGAIN;
 539                 goto out_close;
 540         }
 541
 542         if (con->rx_page == NULL) {
 543                 /*
 544                  * This doesn't need to be atomic, but I think it should
 545                  * improve performance if it is.
 546                  */
 547                 con->rx_page = alloc_page(GFP_ATOMIC);
 548                 if (con->rx_page == NULL)
 549                         goto out_resched;
 550                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 551         }
 552
 553         /* Only SCTP needs these really */
 554         memset(&incmsg, 0, sizeof(incmsg));
 555         msg.msg_control = incmsg;
 556         msg.msg_controllen = sizeof(incmsg);
 557
 558         /*
 559          * iov[0] is the bit of the circular buffer between the current end
 560          * point (cb.base + cb.len) and the end of the buffer.
 561          */
 562         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 563         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 564         iov[1].iov_len = 0;
 565         nvec = 1;
 566
 567         /*
 568          * iov[1] is the bit of the circular buffer between the start of the
 569          * buffer and the start of the currently used section (cb.base)
 570          */
 571         if (cbuf_data(&con->cb) >= con->cb.base) {
 572                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 573                 iov[1].iov_len = con->cb.base;
 574                 iov[1].iov_base = page_address(con->rx_page);
 575                 nvec = 2;
 576         }
 577         len = iov[0].iov_len + iov[1].iov_len;
 578
 579         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 580                                MSG_DONTWAIT | MSG_NOSIGNAL);
 581         if (ret <= 0)
 582                 goto out_close;
 583
 584         /* Process SCTP notifications */
 585         if (msg.msg_flags & MSG_NOTIFICATION) {
 586                 msg.msg_control = incmsg;
 587                 msg.msg_controllen = sizeof(incmsg);
 588
 589                 process_sctp_notification(con, &msg,
 590                                           page_address(con->rx_page) + con->cb.base);
 591                 mutex_unlock(&con->sock_mutex);
 592                 return 0;
 593         }
 594         BUG_ON(con->nodeid == 0);
 595
 596         if (ret == len)
 597                 call_again_soon = 1;
 598         cbuf_add(&con->cb, ret);
 599         ret = dlm_process_incoming_buffer(con->nodeid,
 600                                           page_address(con->rx_page),
 601                                           con->cb.base, con->cb.len,
 602                                           PAGE_CACHE_SIZE);
 603         if (ret == -EBADMSG) {
 604                 printk(KERN_INFO "dlm: lowcomms: addr=%p, base=%u, len=%u, "
 605                        "iov_len=%u, iov_base[0]=%p, read=%d\n",
 606                        page_address(con->rx_page), con->cb.base, con->cb.len,
 607                        len, iov[0].iov_base, r);
 608         }
 609         if (ret < 0)
 610                 goto out_close;
 611         cbuf_eat(&con->cb, ret);
 612
 613         if (cbuf_empty(&con->cb) && !call_again_soon) {
 614                 __free_page(con->rx_page);
 615                 con->rx_page = NULL;
 616         }
 617
 618         if (call_again_soon)
 619                 goto out_resched;
 620         mutex_unlock(&con->sock_mutex);
 621         return 0;
 622
 623 out_resched:
 624         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 625                 queue_work(recv_workqueue, &con->rwork);
 626         mutex_unlock(&con->sock_mutex);
 627         return -EAGAIN;
 628
 629 out_close:
 630         mutex_unlock(&con->sock_mutex);
 631         if (ret != -EAGAIN && !test_bit(CF_IS_OTHERCON, &con->flags)) {
 632                 close_connection(con, false);
 633                 /* Reconnect when there is something to send */
 634         }
 635         /* Don't return success if we really got EOF */
 636         if (ret == 0)
 637                 ret = -EAGAIN;
 638
 639         return ret;
 640 }
 641
 642 /* Listening socket is busy, accept a connection */
 643 static int tcp_accept_from_sock(struct connection *con)
 644 {
 645         int result;
 646         struct sockaddr_storage peeraddr;
 647         struct socket *newsock;
 648         int len;
 649         int nodeid;
 650         struct connection *newcon;
 651         struct connection *addcon;
 652
 653         memset(&peeraddr, 0, sizeof(peeraddr));
 654         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 655                                   IPPROTO_TCP, &newsock);
 656         if (result < 0)
 657                 return -ENOMEM;
 658
 659         mutex_lock_nested(&con->sock_mutex, 0);
 660
 661         result = -ENOTCONN;
 662         if (con->sock == NULL)
 663                 goto accept_err;
 664
 665         newsock->type = con->sock->type;
 666         newsock->ops = con->sock->ops;
 667
 668         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 669         if (result < 0)
 670                 goto accept_err;
 671
 672         /* Get the connected socket's peer */
 673         memset(&peeraddr, 0, sizeof(peeraddr));
 674         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 675                                   &len, 2)) {
 676                 result = -ECONNABORTED;
 677                 goto accept_err;
 678         }
 679
 680         /* Get the new node's NODEID */
 681         make_sockaddr(&peeraddr, 0, &len);
 682         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 683                 printk("dlm: connect from non cluster node\n");
 684                 sock_release(newsock);
 685                 mutex_unlock(&con->sock_mutex);
 686                 return -1;
 687         }
 688
 689         log_print("got connection from %d", nodeid);
 690
 691         /*  Check to see if we already have a connection to this node. This
 692          *  could happen if the two nodes initiate a connection at roughly
 693          *  the same time and the connections cross on the wire.
 694          *  In this case we store the incoming one in "othercon"
 695          */
 696         newcon = nodeid2con(nodeid, GFP_KERNEL);
 697         if (!newcon) {
 698                 result = -ENOMEM;
 699                 goto accept_err;
 700         }
 701         mutex_lock_nested(&newcon->sock_mutex, 1);
 702         if (newcon->sock) {
 703                 struct connection *othercon = newcon->othercon;
 704
 705                 if (!othercon) {
 706                         othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
 707                         if (!othercon) {
 708                                 printk("dlm: failed to allocate incoming socket\n");
 709                                 mutex_unlock(&newcon->sock_mutex);
 710                                 result = -ENOMEM;
 711                                 goto accept_err;
 712                         }
 713                         othercon->nodeid = nodeid;
 714                         othercon->rx_action = receive_from_sock;
 715                         mutex_init(&othercon->sock_mutex);
 716                         INIT_WORK(&othercon->swork, process_send_sockets);
 717                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 718                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 719                         newcon->othercon = othercon;
 720                 }
 721                 othercon->sock = newsock;
 722                 newsock->sk->sk_user_data = othercon;
 723                 add_sock(newsock, othercon);
 724                 addcon = othercon;
 725         }
 726         else {
 727                 newsock->sk->sk_user_data = newcon;
 728                 newcon->rx_action = receive_from_sock;
 729                 add_sock(newsock, newcon);
 730                 addcon = newcon;
 731         }
 732
 733         mutex_unlock(&newcon->sock_mutex);
 734
 735         /*
 736          * Add it to the active queue in case we got data
 737          * beween processing the accept adding the socket
 738          * to the read_sockets list
 739          */
 740         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 741                 queue_work(recv_workqueue, &addcon->rwork);
 742         mutex_unlock(&con->sock_mutex);
 743
 744         return 0;
 745
 746 accept_err:
 747         mutex_unlock(&con->sock_mutex);
 748         sock_release(newsock);
 749
 750         if (result != -EAGAIN)
 751                 printk("dlm: error accepting connection from node: %d\n", result);
 752         return result;
 753 }
 754
 755 static void free_entry(struct writequeue_entry *e)
 756 {
 757         __free_page(e->page);
 758         kfree(e);
 759 }
 760
 761 /* Initiate an SCTP association.
 762    This is a special case of send_to_sock() in that we don't yet have a
 763    peeled-off socket for this association, so we use the listening socket
 764    and add the primary IP address of the remote node.
 765  */
 766 static void sctp_init_assoc(struct connection *con)
 767 {
 768         struct sockaddr_storage rem_addr;
 769         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 770         struct msghdr outmessage;
 771         struct cmsghdr *cmsg;
 772         struct sctp_sndrcvinfo *sinfo;
 773         struct connection *base_con;
 774         struct writequeue_entry *e;
 775         int len, offset;
 776         int ret;
 777         int addrlen;
 778         struct kvec iov[1];
 779
 780         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 781                 return;
 782
 783         if (con->retries++ > MAX_CONNECT_RETRIES)
 784                 return;
 785
 786         log_print("Initiating association with node %d", con->nodeid);
 787
 788         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 789                 log_print("no address for nodeid %d", con->nodeid);
 790                 return;
 791         }
 792         base_con = nodeid2con(0, 0);
 793         BUG_ON(base_con == NULL);
 794
 795         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 796
 797         outmessage.msg_name = &rem_addr;
 798         outmessage.msg_namelen = addrlen;
 799         outmessage.msg_control = outcmsg;
 800         outmessage.msg_controllen = sizeof(outcmsg);
 801         outmessage.msg_flags = MSG_EOR;
 802
 803         spin_lock(&con->writequeue_lock);
 804         e = list_entry(con->writequeue.next, struct writequeue_entry,
 805                        list);
 806
 807         BUG_ON((struct list_head *) e == &con->writequeue);
 808
 809         len = e->len;
 810         offset = e->offset;
 811         spin_unlock(&con->writequeue_lock);
 812         kmap(e->page);
 813
 814         /* Send the first block off the write queue */
 815         iov[0].iov_base = page_address(e->page)+offset;
 816         iov[0].iov_len = len;
 817
 818         cmsg = CMSG_FIRSTHDR(&outmessage);
 819         cmsg->cmsg_level = IPPROTO_SCTP;
 820         cmsg->cmsg_type = SCTP_SNDRCV;
 821         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 822         sinfo = CMSG_DATA(cmsg);
 823         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 824         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 825         outmessage.msg_controllen = cmsg->cmsg_len;
 826
 827         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 828         if (ret < 0) {
 829                 log_print("Send first packet to node %d failed: %d", con->nodeid, ret);
 830
 831                 /* Try again later */
 832                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 833                 clear_bit(CF_INIT_PENDING, &con->flags);
 834         }
 835         else {
 836                 spin_lock(&con->writequeue_lock);
 837                 e->offset += ret;
 838                 e->len -= ret;
 839
 840                 if (e->len == 0 && e->users == 0) {
 841                         list_del(&e->list);
 842                         kunmap(e->page);
 843                         free_entry(e);
 844                 }
 845                 spin_unlock(&con->writequeue_lock);
 846         }
 847 }
 848
 849 /* Connect a new socket to its peer */
 850 static void tcp_connect_to_sock(struct connection *con)
 851 {
 852         int result = -EHOSTUNREACH;
 853         struct sockaddr_storage saddr;
 854         int addr_len;
 855         struct socket *sock;
 856
 857         if (con->nodeid == 0) {
 858                 log_print("attempt to connect sock 0 foiled");
 859                 return;
 860         }
 861
 862         mutex_lock(&con->sock_mutex);
 863         if (con->retries++ > MAX_CONNECT_RETRIES)
 864                 goto out;
 865
 866         /* Some odd races can cause double-connects, ignore them */
 867         if (con->sock) {
 868                 result = 0;
 869                 goto out;
 870         }
 871
 872         /* Create a socket to communicate with */
 873         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 874                                   IPPROTO_TCP, &sock);
 875         if (result < 0)
 876                 goto out_err;
 877
 878         memset(&saddr, 0, sizeof(saddr));
 879         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 880                 goto out_err;
 881
 882         sock->sk->sk_user_data = con;
 883         con->rx_action = receive_from_sock;
 884         con->connect_action = tcp_connect_to_sock;
 885         add_sock(sock, con);
 886
 887         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 888
 889         log_print("connecting to %d", con->nodeid);
 890         result =
 891                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 892                                    O_NONBLOCK);
 893         if (result == -EINPROGRESS)
 894                 result = 0;
 895         if (result == 0)
 896                 goto out;
 897
 898 out_err:
 899         if (con->sock) {
 900                 sock_release(con->sock);
 901                 con->sock = NULL;
 902         }
 903         /*
 904          * Some errors are fatal and this list might need adjusting. For other
 905          * errors we try again until the max number of retries is reached.
 906          */
 907         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 908             result != -ENETDOWN && result != EINVAL
 909             && result != -EPROTONOSUPPORT) {
 910                 lowcomms_connect_sock(con);
 911                 result = 0;
 912         }
 913 out:
 914         mutex_unlock(&con->sock_mutex);
 915         return;
 916 }
 917
 918 static struct socket *tcp_create_listen_sock(struct connection *con,
 919                                              struct sockaddr_storage *saddr)
 920 {
 921         struct socket *sock = NULL;
 922         int result = 0;
 923         int one = 1;
 924         int addr_len;
 925
 926         if (dlm_local_addr[0]->ss_family == AF_INET)
 927                 addr_len = sizeof(struct sockaddr_in);
 928         else
 929                 addr_len = sizeof(struct sockaddr_in6);
 930
 931         /* Create a socket to communicate with */
 932         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, IPPROTO_TCP, &sock);
 933         if (result < 0) {
 934                 printk("dlm: Can't create listening comms socket\n");
 935                 goto create_out;
 936         }
 937
 938         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
 939                                    (char *)&one, sizeof(one));
 940
 941         if (result < 0) {
 942                 printk("dlm: Failed to set SO_REUSEADDR on socket: result=%d\n",
 943                        result);
 944         }
 945         sock->sk->sk_user_data = con;
 946         con->rx_action = tcp_accept_from_sock;
 947         con->connect_action = tcp_connect_to_sock;
 948         con->sock = sock;
 949
 950         /* Bind to our port */
 951         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
 952         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
 953         if (result < 0) {
 954                 printk("dlm: Can't bind to port %d\n", dlm_config.ci_tcp_port);
 955                 sock_release(sock);
 956                 sock = NULL;
 957                 con->sock = NULL;
 958                 goto create_out;
 959         }
 960         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 961                                  (char *)&one, sizeof(one));
 962         if (result < 0) {
 963                 printk("dlm: Set keepalive failed: %d\n", result);
 964         }
 965
 966         result = sock->ops->listen(sock, 5);
 967         if (result < 0) {
 968                 printk("dlm: Can't listen on port %d\n", dlm_config.ci_tcp_port);
 969                 sock_release(sock);
 970                 sock = NULL;
 971                 goto create_out;
 972         }
 973
 974 create_out:
 975         return sock;
 976 }
 977
 978 /* Get local addresses */
 979 static void init_local(void)
 980 {
 981         struct sockaddr_storage sas, *addr;
 982         int i;
 983
 984         dlm_local_count = 0;
 985         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
 986                 if (dlm_our_addr(&sas, i))
 987                         break;
 988
 989                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
 990                 if (!addr)
 991                         break;
 992                 memcpy(addr, &sas, sizeof(*addr));
 993                 dlm_local_addr[dlm_local_count++] = addr;
 994         }
 995 }
 996
 997 /* Bind to an IP address. SCTP allows multiple address so it can do multi-homing */
 998 static int add_sctp_bind_addr(struct connection *sctp_con, struct sockaddr_storage *addr, int addr_len, int num)
 999 {
1000         int result = 0;
1001
1002         if (num == 1)
1003                 result = kernel_bind(sctp_con->sock,
1004                                      (struct sockaddr *) addr,
1005                                      addr_len);
1006         else
1007                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1008                                            SCTP_SOCKOPT_BINDX_ADD,
1009                                            (char *)addr, addr_len);
1010
1011         if (result < 0)
1012                 log_print("Can't bind to port %d addr number %d",
1013                           dlm_config.ci_tcp_port, num);
1014
1015         return result;
1016 }
1017
1018 /* Initialise SCTP socket and bind to all interfaces */
1019 static int sctp_listen_for_all(void)
1020 {
1021         struct socket *sock = NULL;
1022         struct sockaddr_storage localaddr;
1023         struct sctp_event_subscribe subscribe;
1024         int result = -EINVAL, num = 1, i, addr_len;
1025         struct connection *con = nodeid2con(0, GFP_KERNEL);
1026         int bufsize = NEEDED_RMEM;
1027
1028         if (!con)
1029                 return -ENOMEM;
1030
1031         log_print("Using SCTP for communications");
1032
1033         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1034                                   IPPROTO_SCTP, &sock);
1035         if (result < 0) {
1036                 log_print("Can't create comms socket, check SCTP is loaded");
1037                 goto out;
1038         }
1039
1040         /* Listen for events */
1041         memset(&subscribe, 0, sizeof(subscribe));
1042         subscribe.sctp_data_io_event = 1;
1043         subscribe.sctp_association_event = 1;
1044         subscribe.sctp_send_failure_event = 1;
1045         subscribe.sctp_shutdown_event = 1;
1046         subscribe.sctp_partial_delivery_event = 1;
1047
1048         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
1049                                  (char *)&bufsize, sizeof(bufsize));
1050         if (result)
1051                 log_print("Error increasing buffer space on socket: %d", result);
1052
1053         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1054                                        (char *)&subscribe, sizeof(subscribe));
1055         if (result < 0) {
1056                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1057                           result);
1058                 goto create_delsock;
1059         }
1060
1061         /* Init con struct */
1062         sock->sk->sk_user_data = con;
1063         con->sock = sock;
1064         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1065         con->rx_action = receive_from_sock;
1066         con->connect_action = sctp_init_assoc;
1067
1068         /* Bind to all interfaces. */
1069         for (i = 0; i < dlm_local_count; i++) {
1070                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1071                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1072
1073                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1074                 if (result)
1075                         goto create_delsock;
1076                 ++num;
1077         }
1078
1079         result = sock->ops->listen(sock, 5);
1080         if (result < 0) {
1081                 log_print("Can't set socket listening");
1082                 goto create_delsock;
1083         }
1084
1085         return 0;
1086
1087 create_delsock:
1088         sock_release(sock);
1089         con->sock = NULL;
1090 out:
1091         return result;
1092 }
1093
1094 static int tcp_listen_for_all(void)
1095 {
1096         struct socket *sock = NULL;
1097         struct connection *con = nodeid2con(0, GFP_KERNEL);
1098         int result = -EINVAL;
1099
1100         if (!con)
1101                 return -ENOMEM;
1102
1103         /* We don't support multi-homed hosts */
1104         if (dlm_local_addr[1] != NULL) {
1105                 log_print("TCP protocol can't handle multi-homed hosts, try SCTP");
1106                 return -EINVAL;
1107         }
1108
1109         log_print("Using TCP for communications");
1110
1111         set_bit(CF_IS_OTHERCON, &con->flags);
1112
1113         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1114         if (sock) {
1115                 add_sock(sock, con);
1116                 result = 0;
1117         }
1118         else {
1119                 result = -EADDRINUSE;
1120         }
1121
1122         return result;
1123 }
1124
1125
1126
1127 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1128                                                      gfp_t allocation)
1129 {
1130         struct writequeue_entry *entry;
1131
1132         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1133         if (!entry)
1134                 return NULL;
1135
1136         entry->page = alloc_page(allocation);
1137         if (!entry->page) {
1138                 kfree(entry);
1139                 return NULL;
1140         }
1141
1142         entry->offset = 0;
1143         entry->len = 0;
1144         entry->end = 0;
1145         entry->users = 0;
1146         entry->con = con;
1147
1148         return entry;
1149 }
1150
1151 void *dlm_lowcomms_get_buffer(int nodeid, int len,
1152                               gfp_t allocation, char **ppc)
1153 {
1154         struct connection *con;
1155         struct writequeue_entry *e;
1156         int offset = 0;
1157         int users = 0;
1158
1159         con = nodeid2con(nodeid, allocation);
1160         if (!con)
1161                 return NULL;
1162
1163         spin_lock(&con->writequeue_lock);
1164         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1165         if ((&e->list == &con->writequeue) ||
1166             (PAGE_CACHE_SIZE - e->end < len)) {
1167                 e = NULL;
1168         } else {
1169                 offset = e->end;
1170                 e->end += len;
1171                 users = e->users++;
1172         }
1173         spin_unlock(&con->writequeue_lock);
1174
1175         if (e) {
1176         got_one:
1177                 if (users == 0)
1178                         kmap(e->page);
1179                 *ppc = page_address(e->page) + offset;
1180                 return e;
1181         }
1182
1183         e = new_writequeue_entry(con, allocation);
1184         if (e) {
1185                 spin_lock(&con->writequeue_lock);
1186                 offset = e->end;
1187                 e->end += len;
1188                 users = e->users++;
1189                 list_add_tail(&e->list, &con->writequeue);
1190                 spin_unlock(&con->writequeue_lock);
1191                 goto got_one;
1192         }
1193         return NULL;
1194 }
1195
1196 void dlm_lowcomms_commit_buffer(void *mh)
1197 {
1198         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1199         struct connection *con = e->con;
1200         int users;
1201
1202         spin_lock(&con->writequeue_lock);
1203         users = --e->users;
1204         if (users)
1205                 goto out;
1206         e->len = e->end - e->offset;
1207         kunmap(e->page);
1208         spin_unlock(&con->writequeue_lock);
1209
1210         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1211                 queue_work(send_workqueue, &con->swork);
1212         }
1213         return;
1214
1215 out:
1216         spin_unlock(&con->writequeue_lock);
1217         return;
1218 }
1219
1220 /* Send a message */
1221 static void send_to_sock(struct connection *con)
1222 {
1223         int ret = 0;
1224         ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1225         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1226         struct writequeue_entry *e;
1227         int len, offset;
1228
1229         mutex_lock(&con->sock_mutex);
1230         if (con->sock == NULL)
1231                 goto out_connect;
1232
1233         sendpage = con->sock->ops->sendpage;
1234
1235         spin_lock(&con->writequeue_lock);
1236         for (;;) {
1237                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1238                                list);
1239                 if ((struct list_head *) e == &con->writequeue)
1240                         break;
1241
1242                 len = e->len;
1243                 offset = e->offset;
1244                 BUG_ON(len == 0 && e->users == 0);
1245                 spin_unlock(&con->writequeue_lock);
1246                 kmap(e->page);
1247
1248                 ret = 0;
1249                 if (len) {
1250                         ret = sendpage(con->sock, e->page, offset, len,
1251                                        msg_flags);
1252                         if (ret == -EAGAIN || ret == 0)
1253                                 goto out;
1254                         if (ret <= 0)
1255                                 goto send_error;
1256                 }
1257                 else {
1258                         /* Don't starve people filling buffers */
1259                         cond_resched();
1260                 }
1261
1262                 spin_lock(&con->writequeue_lock);
1263                 e->offset += ret;
1264                 e->len -= ret;
1265
1266                 if (e->len == 0 && e->users == 0) {
1267                         list_del(&e->list);
1268                         kunmap(e->page);
1269                         free_entry(e);
1270                         continue;
1271                 }
1272         }
1273         spin_unlock(&con->writequeue_lock);
1274 out:
1275         mutex_unlock(&con->sock_mutex);
1276         return;
1277
1278 send_error:
1279         mutex_unlock(&con->sock_mutex);
1280         close_connection(con, false);
1281         lowcomms_connect_sock(con);
1282         return;
1283
1284 out_connect:
1285         mutex_unlock(&con->sock_mutex);
1286         if (!test_bit(CF_INIT_PENDING, &con->flags))
1287                 lowcomms_connect_sock(con);
1288         return;
1289 }
1290
1291 static void clean_one_writequeue(struct connection *con)
1292 {
1293         struct list_head *list;
1294         struct list_head *temp;
1295
1296         spin_lock(&con->writequeue_lock);
1297         list_for_each_safe(list, temp, &con->writequeue) {
1298                 struct writequeue_entry *e =
1299                         list_entry(list, struct writequeue_entry, list);
1300                 list_del(&e->list);
1301                 free_entry(e);
1302         }
1303         spin_unlock(&con->writequeue_lock);
1304 }
1305
1306 /* Called from recovery when it knows that a node has
1307    left the cluster */
1308 int dlm_lowcomms_close(int nodeid)
1309 {
1310         struct connection *con;
1311
1312         log_print("closing connection to node %d", nodeid);
1313         con = nodeid2con(nodeid, 0);
1314         if (con) {
1315                 clean_one_writequeue(con);
1316                 close_connection(con, true);
1317         }
1318         return 0;
1319 }
1320
1321 /* Receive workqueue function */
1322 static void process_recv_sockets(struct work_struct *work)
1323 {
1324         struct connection *con = container_of(work, struct connection, rwork);
1325         int err;
1326
1327         clear_bit(CF_READ_PENDING, &con->flags);
1328         do {
1329                 err = con->rx_action(con);
1330         } while (!err);
1331 }
1332
1333 /* Send workqueue function */
1334 static void process_send_sockets(struct work_struct *work)
1335 {
1336         struct connection *con = container_of(work, struct connection, swork);
1337
1338         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1339                 con->connect_action(con);
1340         }
1341         clear_bit(CF_WRITE_PENDING, &con->flags);
1342         send_to_sock(con);
1343 }
1344
1345
1346 /* Discard all entries on the write queues */
1347 static void clean_writequeues(void)
1348 {
1349         int nodeid;
1350
1351         for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1352                 struct connection *con = __nodeid2con(nodeid, 0);
1353
1354                 if (con)
1355                         clean_one_writequeue(con);
1356         }
1357 }
1358
1359 static void work_stop(void)
1360 {
1361         destroy_workqueue(recv_workqueue);
1362         destroy_workqueue(send_workqueue);
1363 }
1364
1365 static int work_start(void)
1366 {
1367         int error;
1368         recv_workqueue = create_workqueue("dlm_recv");
1369         error = IS_ERR(recv_workqueue);
1370         if (error) {
1371                 log_print("can't start dlm_recv %d", error);
1372                 return error;
1373         }
1374
1375         send_workqueue = create_singlethread_workqueue("dlm_send");
1376         error = IS_ERR(send_workqueue);
1377         if (error) {
1378                 log_print("can't start dlm_send %d", error);
1379                 destroy_workqueue(recv_workqueue);
1380                 return error;
1381         }
1382
1383         return 0;
1384 }
1385
1386 void dlm_lowcomms_stop(void)
1387 {
1388         int i;
1389         struct connection *con;
1390
1391         /* Set all the flags to prevent any
1392            socket activity.
1393         */
1394         down(&connections_lock);
1395         for (i = 0; i <= max_nodeid; i++) {
1396                 con = __nodeid2con(i, 0);
1397                 if (con)
1398                         con->flags |= 0xFF;
1399         }
1400         up(&connections_lock);
1401
1402         work_stop();
1403
1404         down(&connections_lock);
1405         clean_writequeues();
1406
1407         for (i = 0; i <= max_nodeid; i++) {
1408                 con = __nodeid2con(i, 0);
1409                 if (con) {
1410                         close_connection(con, true);
1411                         if (con->othercon)
1412                                 kmem_cache_free(con_cache, con->othercon);
1413                         kmem_cache_free(con_cache, con);
1414                 }
1415         }
1416         max_nodeid = 0;
1417         up(&connections_lock);
1418         kmem_cache_destroy(con_cache);
1419         idr_init(&connections_idr);
1420 }
1421
1422 int dlm_lowcomms_start(void)
1423 {
1424         int error = -EINVAL;
1425         struct connection *con;
1426
1427         init_local();
1428         if (!dlm_local_count) {
1429                 log_print("no local IP address has been set");
1430                 goto out;
1431         }
1432
1433         error = -ENOMEM;
1434         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1435                                       __alignof__(struct connection), 0,
1436                                       NULL, NULL);
1437         if (!con_cache)
1438                 goto out;
1439
1440         /* Set some sysctl minima */
1441         if (sysctl_rmem_max < NEEDED_RMEM)
1442                 sysctl_rmem_max = NEEDED_RMEM;
1443
1444         /* Start listening */
1445         if (dlm_config.ci_protocol == 0)
1446                 error = tcp_listen_for_all();
1447         else
1448                 error = sctp_listen_for_all();
1449         if (error)
1450                 goto fail_unlisten;
1451
1452         error = work_start();
1453         if (error)
1454                 goto fail_unlisten;
1455
1456         return 0;
1457
1458 fail_unlisten:
1459         con = nodeid2con(0,0);
1460         if (con) {
1461                 close_connection(con, false);
1462                 kmem_cache_free(con_cache, con);
1463         }
1464         kmem_cache_destroy(con_cache);
1465
1466 out:
1467         return error;
1468 }