Pileus Git - ~andy/linux/blob - 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 (con && !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 (con && !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                 kmem_cache_free(con_cache, con->othercon);
 331                 con->othercon = NULL;
 332         }
 333         if (con->rx_page) {
 334                 __free_page(con->rx_page);
 335                 con->rx_page = NULL;
 336         }
 337         con->retries = 0;
 338         mutex_unlock(&con->sock_mutex);
 339 }
 340
 341 /* We only send shutdown messages to nodes that are not part of the cluster */
 342 static void sctp_send_shutdown(sctp_assoc_t associd)
 343 {
 344         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 345         struct msghdr outmessage;
 346         struct cmsghdr *cmsg;
 347         struct sctp_sndrcvinfo *sinfo;
 348         int ret;
 349         struct connection *con;
 350
 351         con = nodeid2con(0,0);
 352         BUG_ON(con == NULL);
 353
 354         outmessage.msg_name = NULL;
 355         outmessage.msg_namelen = 0;
 356         outmessage.msg_control = outcmsg;
 357         outmessage.msg_controllen = sizeof(outcmsg);
 358         outmessage.msg_flags = MSG_EOR;
 359
 360         cmsg = CMSG_FIRSTHDR(&outmessage);
 361         cmsg->cmsg_level = IPPROTO_SCTP;
 362         cmsg->cmsg_type = SCTP_SNDRCV;
 363         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 364         outmessage.msg_controllen = cmsg->cmsg_len;
 365         sinfo = CMSG_DATA(cmsg);
 366         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 367
 368         sinfo->sinfo_flags |= MSG_EOF;
 369         sinfo->sinfo_assoc_id = associd;
 370
 371         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 372
 373         if (ret != 0)
 374                 log_print("send EOF to node failed: %d", ret);
 375 }
 376
 377 /* INIT failed but we don't know which node...
 378    restart INIT on all pending nodes */
 379 static void sctp_init_failed(void)
 380 {
 381         int i;
 382         struct connection *con;
 383
 384         down(&connections_lock);
 385         for (i=1; i<=max_nodeid; i++) {
 386                 con = __nodeid2con(i, 0);
 387                 if (!con)
 388                         continue;
 389                 con->sctp_assoc = 0;
 390                 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 391                         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
 392                                 queue_work(send_workqueue, &con->swork);
 393                         }
 394                 }
 395         }
 396         up(&connections_lock);
 397 }
 398
 399 /* Something happened to an association */
 400 static void process_sctp_notification(struct connection *con,
 401                                       struct msghdr *msg, char *buf)
 402 {
 403         union sctp_notification *sn = (union sctp_notification *)buf;
 404
 405         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 406                 switch (sn->sn_assoc_change.sac_state) {
 407
 408                 case SCTP_COMM_UP:
 409                 case SCTP_RESTART:
 410                 {
 411                         /* Check that the new node is in the lockspace */
 412                         struct sctp_prim prim;
 413                         int nodeid;
 414                         int prim_len, ret;
 415                         int addr_len;
 416                         struct connection *new_con;
 417                         struct file *file;
 418                         sctp_peeloff_arg_t parg;
 419                         int parglen = sizeof(parg);
 420
 421                         /*
 422                          * We get this before any data for an association.
 423                          * We verify that the node is in the cluster and
 424                          * then peel off a socket for it.
 425                          */
 426                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 427                                 log_print("COMM_UP for invalid assoc ID %d",
 428                                          (int)sn->sn_assoc_change.sac_assoc_id);
 429                                 sctp_init_failed();
 430                                 return;
 431                         }
 432                         memset(&prim, 0, sizeof(struct sctp_prim));
 433                         prim_len = sizeof(struct sctp_prim);
 434                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 435
 436                         ret = kernel_getsockopt(con->sock,
 437                                                 IPPROTO_SCTP,
 438                                                 SCTP_PRIMARY_ADDR,
 439                                                 (char*)&prim,
 440                                                 &prim_len);
 441                         if (ret < 0) {
 442                                 log_print("getsockopt/sctp_primary_addr on "
 443                                           "new assoc %d failed : %d",
 444                                           (int)sn->sn_assoc_change.sac_assoc_id,
 445                                           ret);
 446
 447                                 /* Retry INIT later */
 448                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 449                                 if (new_con)
 450                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 451                                 return;
 452                         }
 453                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 454                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 455                                 int i;
 456                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 457                                 log_print("reject connect from unknown addr");
 458                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 459                                         printk("%02x ", b[i]);
 460                                 printk("\n");
 461                                 sctp_send_shutdown(prim.ssp_assoc_id);
 462                                 return;
 463                         }
 464
 465                         new_con = nodeid2con(nodeid, GFP_KERNEL);
 466                         if (!new_con)
 467                                 return;
 468
 469                         /* Peel off a new sock */
 470                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 471                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 472                                                 SCTP_SOCKOPT_PEELOFF,
 473                                                 (void *)&parg, &parglen);
 474                         if (ret) {
 475                                 log_print("Can't peel off a socket for "
 476                                           "connection %d to node %d: err=%d\n",
 477                                           parg.associd, nodeid, ret);
 478                         }
 479                         file = fget(parg.sd);
 480                         new_con->sock = SOCKET_I(file->f_dentry->d_inode);
 481                         add_sock(new_con->sock, new_con);
 482                         fput(file);
 483                         put_unused_fd(parg.sd);
 484
 485                         log_print("got new/restarted association %d nodeid %d",
 486                                  (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
 487
 488                         /* Send any pending writes */
 489                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 490                         clear_bit(CF_INIT_PENDING, &con->flags);
 491                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 492                                 queue_work(send_workqueue, &new_con->swork);
 493                         }
 494                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 495                                 queue_work(recv_workqueue, &new_con->rwork);
 496                 }
 497                 break;
 498
 499                 case SCTP_COMM_LOST:
 500                 case SCTP_SHUTDOWN_COMP:
 501                 {
 502                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 503                         if (con) {
 504                                 con->sctp_assoc = 0;
 505                         }
 506                 }
 507                 break;
 508
 509                 /* We don't know which INIT failed, so clear the PENDING flags
 510                  * on them all.  if assoc_id is zero then it will then try
 511                  * again */
 512
 513                 case SCTP_CANT_STR_ASSOC:
 514                 {
 515                         log_print("Can't start SCTP association - retrying");
 516                         sctp_init_failed();
 517                 }
 518                 break;
 519
 520                 default:
 521                         log_print("unexpected SCTP assoc change id=%d state=%d",
 522                                   (int)sn->sn_assoc_change.sac_assoc_id,
 523                                   sn->sn_assoc_change.sac_state);
 524                 }
 525         }
 526 }
 527
 528 /* Data received from remote end */
 529 static int receive_from_sock(struct connection *con)
 530 {
 531         int ret = 0;
 532         struct msghdr msg = {};
 533         struct kvec iov[2];
 534         unsigned len;
 535         int r;
 536         int call_again_soon = 0;
 537         int nvec;
 538         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 539
 540         mutex_lock(&con->sock_mutex);
 541
 542         if (con->sock == NULL) {
 543                 ret = -EAGAIN;
 544                 goto out_close;
 545         }
 546
 547         if (con->rx_page == NULL) {
 548                 /*
 549                  * This doesn't need to be atomic, but I think it should
 550                  * improve performance if it is.
 551                  */
 552                 con->rx_page = alloc_page(GFP_ATOMIC);
 553                 if (con->rx_page == NULL)
 554                         goto out_resched;
 555                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 556         }
 557
 558         /* Only SCTP needs these really */
 559         memset(&incmsg, 0, sizeof(incmsg));
 560         msg.msg_control = incmsg;
 561         msg.msg_controllen = sizeof(incmsg);
 562
 563         /*
 564          * iov[0] is the bit of the circular buffer between the current end
 565          * point (cb.base + cb.len) and the end of the buffer.
 566          */
 567         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 568         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 569         iov[1].iov_len = 0;
 570         nvec = 1;
 571
 572         /*
 573          * iov[1] is the bit of the circular buffer between the start of the
 574          * buffer and the start of the currently used section (cb.base)
 575          */
 576         if (cbuf_data(&con->cb) >= con->cb.base) {
 577                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 578                 iov[1].iov_len = con->cb.base;
 579                 iov[1].iov_base = page_address(con->rx_page);
 580                 nvec = 2;
 581         }
 582         len = iov[0].iov_len + iov[1].iov_len;
 583
 584         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 585                                MSG_DONTWAIT | MSG_NOSIGNAL);
 586         if (ret <= 0)
 587                 goto out_close;
 588
 589         /* Process SCTP notifications */
 590         if (msg.msg_flags & MSG_NOTIFICATION) {
 591                 msg.msg_control = incmsg;
 592                 msg.msg_controllen = sizeof(incmsg);
 593
 594                 process_sctp_notification(con, &msg,
 595                                 page_address(con->rx_page) + con->cb.base);
 596                 mutex_unlock(&con->sock_mutex);
 597                 return 0;
 598         }
 599         BUG_ON(con->nodeid == 0);
 600
 601         if (ret == len)
 602                 call_again_soon = 1;
 603         cbuf_add(&con->cb, ret);
 604         ret = dlm_process_incoming_buffer(con->nodeid,
 605                                           page_address(con->rx_page),
 606                                           con->cb.base, con->cb.len,
 607                                           PAGE_CACHE_SIZE);
 608         if (ret == -EBADMSG) {
 609                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 610                           "iov_len=%u, iov_base[0]=%p, read=%d",
 611                           page_address(con->rx_page), con->cb.base, con->cb.len,
 612                           len, iov[0].iov_base, r);
 613         }
 614         if (ret < 0)
 615                 goto out_close;
 616         cbuf_eat(&con->cb, ret);
 617
 618         if (cbuf_empty(&con->cb) && !call_again_soon) {
 619                 __free_page(con->rx_page);
 620                 con->rx_page = NULL;
 621         }
 622
 623         if (call_again_soon)
 624                 goto out_resched;
 625         mutex_unlock(&con->sock_mutex);
 626         return 0;
 627
 628 out_resched:
 629         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 630                 queue_work(recv_workqueue, &con->rwork);
 631         mutex_unlock(&con->sock_mutex);
 632         return -EAGAIN;
 633
 634 out_close:
 635         mutex_unlock(&con->sock_mutex);
 636         if (ret != -EAGAIN && !test_bit(CF_IS_OTHERCON, &con->flags)) {
 637                 close_connection(con, false);
 638                 /* Reconnect when there is something to send */
 639         }
 640         /* Don't return success if we really got EOF */
 641         if (ret == 0)
 642                 ret = -EAGAIN;
 643
 644         return ret;
 645 }
 646
 647 /* Listening socket is busy, accept a connection */
 648 static int tcp_accept_from_sock(struct connection *con)
 649 {
 650         int result;
 651         struct sockaddr_storage peeraddr;
 652         struct socket *newsock;
 653         int len;
 654         int nodeid;
 655         struct connection *newcon;
 656         struct connection *addcon;
 657
 658         memset(&peeraddr, 0, sizeof(peeraddr));
 659         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 660                                   IPPROTO_TCP, &newsock);
 661         if (result < 0)
 662                 return -ENOMEM;
 663
 664         mutex_lock_nested(&con->sock_mutex, 0);
 665
 666         result = -ENOTCONN;
 667         if (con->sock == NULL)
 668                 goto accept_err;
 669
 670         newsock->type = con->sock->type;
 671         newsock->ops = con->sock->ops;
 672
 673         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 674         if (result < 0)
 675                 goto accept_err;
 676
 677         /* Get the connected socket's peer */
 678         memset(&peeraddr, 0, sizeof(peeraddr));
 679         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 680                                   &len, 2)) {
 681                 result = -ECONNABORTED;
 682                 goto accept_err;
 683         }
 684
 685         /* Get the new node's NODEID */
 686         make_sockaddr(&peeraddr, 0, &len);
 687         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 688                 log_print("connect from non cluster node");
 689                 sock_release(newsock);
 690                 mutex_unlock(&con->sock_mutex);
 691                 return -1;
 692         }
 693
 694         log_print("got connection from %d", nodeid);
 695
 696         /*  Check to see if we already have a connection to this node. This
 697          *  could happen if the two nodes initiate a connection at roughly
 698          *  the same time and the connections cross on the wire.
 699          *  In this case we store the incoming one in "othercon"
 700          */
 701         newcon = nodeid2con(nodeid, GFP_KERNEL);
 702         if (!newcon) {
 703                 result = -ENOMEM;
 704                 goto accept_err;
 705         }
 706         mutex_lock_nested(&newcon->sock_mutex, 1);
 707         if (newcon->sock) {
 708                 struct connection *othercon = newcon->othercon;
 709
 710                 if (!othercon) {
 711                         othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
 712                         if (!othercon) {
 713                                 log_print("failed to allocate incoming socket");
 714                                 mutex_unlock(&newcon->sock_mutex);
 715                                 result = -ENOMEM;
 716                                 goto accept_err;
 717                         }
 718                         othercon->nodeid = nodeid;
 719                         othercon->rx_action = receive_from_sock;
 720                         mutex_init(&othercon->sock_mutex);
 721                         INIT_WORK(&othercon->swork, process_send_sockets);
 722                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 723                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 724                         newcon->othercon = othercon;
 725                         othercon->sock = newsock;
 726                         newsock->sk->sk_user_data = othercon;
 727                         add_sock(newsock, othercon);
 728                         addcon = othercon;
 729                 }
 730                 else {
 731                         printk("Extra connection from node %d attempted\n", nodeid);
 732                         result = -EAGAIN;
 733                         mutex_unlock(&newcon->sock_mutex);
 734                         goto accept_err;
 735                 }
 736         }
 737         else {
 738                 newsock->sk->sk_user_data = newcon;
 739                 newcon->rx_action = receive_from_sock;
 740                 add_sock(newsock, newcon);
 741                 addcon = newcon;
 742         }
 743
 744         mutex_unlock(&newcon->sock_mutex);
 745
 746         /*
 747          * Add it to the active queue in case we got data
 748          * beween processing the accept adding the socket
 749          * to the read_sockets list
 750          */
 751         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 752                 queue_work(recv_workqueue, &addcon->rwork);
 753         mutex_unlock(&con->sock_mutex);
 754
 755         return 0;
 756
 757 accept_err:
 758         mutex_unlock(&con->sock_mutex);
 759         sock_release(newsock);
 760
 761         if (result != -EAGAIN)
 762                 log_print("error accepting connection from node: %d", result);
 763         return result;
 764 }
 765
 766 static void free_entry(struct writequeue_entry *e)
 767 {
 768         __free_page(e->page);
 769         kfree(e);
 770 }
 771
 772 /* Initiate an SCTP association.
 773    This is a special case of send_to_sock() in that we don't yet have a
 774    peeled-off socket for this association, so we use the listening socket
 775    and add the primary IP address of the remote node.
 776  */
 777 static void sctp_init_assoc(struct connection *con)
 778 {
 779         struct sockaddr_storage rem_addr;
 780         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 781         struct msghdr outmessage;
 782         struct cmsghdr *cmsg;
 783         struct sctp_sndrcvinfo *sinfo;
 784         struct connection *base_con;
 785         struct writequeue_entry *e;
 786         int len, offset;
 787         int ret;
 788         int addrlen;
 789         struct kvec iov[1];
 790
 791         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 792                 return;
 793
 794         if (con->retries++ > MAX_CONNECT_RETRIES)
 795                 return;
 796
 797         log_print("Initiating association with node %d", con->nodeid);
 798
 799         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 800                 log_print("no address for nodeid %d", con->nodeid);
 801                 return;
 802         }
 803         base_con = nodeid2con(0, 0);
 804         BUG_ON(base_con == NULL);
 805
 806         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 807
 808         outmessage.msg_name = &rem_addr;
 809         outmessage.msg_namelen = addrlen;
 810         outmessage.msg_control = outcmsg;
 811         outmessage.msg_controllen = sizeof(outcmsg);
 812         outmessage.msg_flags = MSG_EOR;
 813
 814         spin_lock(&con->writequeue_lock);
 815         e = list_entry(con->writequeue.next, struct writequeue_entry,
 816                        list);
 817
 818         BUG_ON((struct list_head *) e == &con->writequeue);
 819
 820         len = e->len;
 821         offset = e->offset;
 822         spin_unlock(&con->writequeue_lock);
 823         kmap(e->page);
 824
 825         /* Send the first block off the write queue */
 826         iov[0].iov_base = page_address(e->page)+offset;
 827         iov[0].iov_len = len;
 828
 829         cmsg = CMSG_FIRSTHDR(&outmessage);
 830         cmsg->cmsg_level = IPPROTO_SCTP;
 831         cmsg->cmsg_type = SCTP_SNDRCV;
 832         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 833         sinfo = CMSG_DATA(cmsg);
 834         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 835         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 836         outmessage.msg_controllen = cmsg->cmsg_len;
 837
 838         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 839         if (ret < 0) {
 840                 log_print("Send first packet to node %d failed: %d",
 841                           con->nodeid, ret);
 842
 843                 /* Try again later */
 844                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 845                 clear_bit(CF_INIT_PENDING, &con->flags);
 846         }
 847         else {
 848                 spin_lock(&con->writequeue_lock);
 849                 e->offset += ret;
 850                 e->len -= ret;
 851
 852                 if (e->len == 0 && e->users == 0) {
 853                         list_del(&e->list);
 854                         kunmap(e->page);
 855                         free_entry(e);
 856                 }
 857                 spin_unlock(&con->writequeue_lock);
 858         }
 859 }
 860
 861 /* Connect a new socket to its peer */
 862 static void tcp_connect_to_sock(struct connection *con)
 863 {
 864         int result = -EHOSTUNREACH;
 865         struct sockaddr_storage saddr;
 866         int addr_len;
 867         struct socket *sock;
 868
 869         if (con->nodeid == 0) {
 870                 log_print("attempt to connect sock 0 foiled");
 871                 return;
 872         }
 873
 874         mutex_lock(&con->sock_mutex);
 875         if (con->retries++ > MAX_CONNECT_RETRIES)
 876                 goto out;
 877
 878         /* Some odd races can cause double-connects, ignore them */
 879         if (con->sock) {
 880                 result = 0;
 881                 goto out;
 882         }
 883
 884         /* Create a socket to communicate with */
 885         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 886                                   IPPROTO_TCP, &sock);
 887         if (result < 0)
 888                 goto out_err;
 889
 890         memset(&saddr, 0, sizeof(saddr));
 891         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 892                 goto out_err;
 893
 894         sock->sk->sk_user_data = con;
 895         con->rx_action = receive_from_sock;
 896         con->connect_action = tcp_connect_to_sock;
 897         add_sock(sock, con);
 898
 899         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 900
 901         log_print("connecting to %d", con->nodeid);
 902         result =
 903                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 904                                    O_NONBLOCK);
 905         if (result == -EINPROGRESS)
 906                 result = 0;
 907         if (result == 0)
 908                 goto out;
 909
 910 out_err:
 911         if (con->sock) {
 912                 sock_release(con->sock);
 913                 con->sock = NULL;
 914         }
 915         /*
 916          * Some errors are fatal and this list might need adjusting. For other
 917          * errors we try again until the max number of retries is reached.
 918          */
 919         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 920             result != -ENETDOWN && result != EINVAL
 921             && result != -EPROTONOSUPPORT) {
 922                 lowcomms_connect_sock(con);
 923                 result = 0;
 924         }
 925 out:
 926         mutex_unlock(&con->sock_mutex);
 927         return;
 928 }
 929
 930 static struct socket *tcp_create_listen_sock(struct connection *con,
 931                                              struct sockaddr_storage *saddr)
 932 {
 933         struct socket *sock = NULL;
 934         int result = 0;
 935         int one = 1;
 936         int addr_len;
 937
 938         if (dlm_local_addr[0]->ss_family == AF_INET)
 939                 addr_len = sizeof(struct sockaddr_in);
 940         else
 941                 addr_len = sizeof(struct sockaddr_in6);
 942
 943         /* Create a socket to communicate with */
 944         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 945                                   IPPROTO_TCP, &sock);
 946         if (result < 0) {
 947                 log_print("Can't create listening comms socket");
 948                 goto create_out;
 949         }
 950
 951         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
 952                                    (char *)&one, sizeof(one));
 953
 954         if (result < 0) {
 955                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
 956         }
 957         sock->sk->sk_user_data = con;
 958         con->rx_action = tcp_accept_from_sock;
 959         con->connect_action = tcp_connect_to_sock;
 960         con->sock = sock;
 961
 962         /* Bind to our port */
 963         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
 964         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
 965         if (result < 0) {
 966                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
 967                 sock_release(sock);
 968                 sock = NULL;
 969                 con->sock = NULL;
 970                 goto create_out;
 971         }
 972         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 973                                  (char *)&one, sizeof(one));
 974         if (result < 0) {
 975                 log_print("Set keepalive failed: %d", result);
 976         }
 977
 978         result = sock->ops->listen(sock, 5);
 979         if (result < 0) {
 980                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
 981                 sock_release(sock);
 982                 sock = NULL;
 983                 goto create_out;
 984         }
 985
 986 create_out:
 987         return sock;
 988 }
 989
 990 /* Get local addresses */
 991 static void init_local(void)
 992 {
 993         struct sockaddr_storage sas, *addr;
 994         int i;
 995
 996         dlm_local_count = 0;
 997         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
 998                 if (dlm_our_addr(&sas, i))
 999                         break;
1000
1001                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1002                 if (!addr)
1003                         break;
1004                 memcpy(addr, &sas, sizeof(*addr));
1005                 dlm_local_addr[dlm_local_count++] = addr;
1006         }
1007 }
1008
1009 /* Bind to an IP address. SCTP allows multiple address so it can do
1010    multi-homing */
1011 static int add_sctp_bind_addr(struct connection *sctp_con,
1012                               struct sockaddr_storage *addr,
1013                               int addr_len, int num)
1014 {
1015         int result = 0;
1016
1017         if (num == 1)
1018                 result = kernel_bind(sctp_con->sock,
1019                                      (struct sockaddr *) addr,
1020                                      addr_len);
1021         else
1022                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1023                                            SCTP_SOCKOPT_BINDX_ADD,
1024                                            (char *)addr, addr_len);
1025
1026         if (result < 0)
1027                 log_print("Can't bind to port %d addr number %d",
1028                           dlm_config.ci_tcp_port, num);
1029
1030         return result;
1031 }
1032
1033 /* Initialise SCTP socket and bind to all interfaces */
1034 static int sctp_listen_for_all(void)
1035 {
1036         struct socket *sock = NULL;
1037         struct sockaddr_storage localaddr;
1038         struct sctp_event_subscribe subscribe;
1039         int result = -EINVAL, num = 1, i, addr_len;
1040         struct connection *con = nodeid2con(0, GFP_KERNEL);
1041         int bufsize = NEEDED_RMEM;
1042
1043         if (!con)
1044                 return -ENOMEM;
1045
1046         log_print("Using SCTP for communications");
1047
1048         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1049                                   IPPROTO_SCTP, &sock);
1050         if (result < 0) {
1051                 log_print("Can't create comms socket, check SCTP is loaded");
1052                 goto out;
1053         }
1054
1055         /* Listen for events */
1056         memset(&subscribe, 0, sizeof(subscribe));
1057         subscribe.sctp_data_io_event = 1;
1058         subscribe.sctp_association_event = 1;
1059         subscribe.sctp_send_failure_event = 1;
1060         subscribe.sctp_shutdown_event = 1;
1061         subscribe.sctp_partial_delivery_event = 1;
1062
1063         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
1064                                  (char *)&bufsize, sizeof(bufsize));
1065         if (result)
1066                 log_print("Error increasing buffer space on socket %d", result);
1067
1068         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1069                                    (char *)&subscribe, sizeof(subscribe));
1070         if (result < 0) {
1071                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1072                           result);
1073                 goto create_delsock;
1074         }
1075
1076         /* Init con struct */
1077         sock->sk->sk_user_data = con;
1078         con->sock = sock;
1079         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1080         con->rx_action = receive_from_sock;
1081         con->connect_action = sctp_init_assoc;
1082
1083         /* Bind to all interfaces. */
1084         for (i = 0; i < dlm_local_count; i++) {
1085                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1086                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1087
1088                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1089                 if (result)
1090                         goto create_delsock;
1091                 ++num;
1092         }
1093
1094         result = sock->ops->listen(sock, 5);
1095         if (result < 0) {
1096                 log_print("Can't set socket listening");
1097                 goto create_delsock;
1098         }
1099
1100         return 0;
1101
1102 create_delsock:
1103         sock_release(sock);
1104         con->sock = NULL;
1105 out:
1106         return result;
1107 }
1108
1109 static int tcp_listen_for_all(void)
1110 {
1111         struct socket *sock = NULL;
1112         struct connection *con = nodeid2con(0, GFP_KERNEL);
1113         int result = -EINVAL;
1114
1115         if (!con)
1116                 return -ENOMEM;
1117
1118         /* We don't support multi-homed hosts */
1119         if (dlm_local_addr[1] != NULL) {
1120                 log_print("TCP protocol can't handle multi-homed hosts, "
1121                           "try SCTP");
1122                 return -EINVAL;
1123         }
1124
1125         log_print("Using TCP for communications");
1126
1127         set_bit(CF_IS_OTHERCON, &con->flags);
1128
1129         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1130         if (sock) {
1131                 add_sock(sock, con);
1132                 result = 0;
1133         }
1134         else {
1135                 result = -EADDRINUSE;
1136         }
1137
1138         return result;
1139 }
1140
1141
1142
1143 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1144                                                      gfp_t allocation)
1145 {
1146         struct writequeue_entry *entry;
1147
1148         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1149         if (!entry)
1150                 return NULL;
1151
1152         entry->page = alloc_page(allocation);
1153         if (!entry->page) {
1154                 kfree(entry);
1155                 return NULL;
1156         }
1157
1158         entry->offset = 0;
1159         entry->len = 0;
1160         entry->end = 0;
1161         entry->users = 0;
1162         entry->con = con;
1163
1164         return entry;
1165 }
1166
1167 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1168 {
1169         struct connection *con;
1170         struct writequeue_entry *e;
1171         int offset = 0;
1172         int users = 0;
1173
1174         con = nodeid2con(nodeid, allocation);
1175         if (!con)
1176                 return NULL;
1177
1178         spin_lock(&con->writequeue_lock);
1179         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1180         if ((&e->list == &con->writequeue) ||
1181             (PAGE_CACHE_SIZE - e->end < len)) {
1182                 e = NULL;
1183         } else {
1184                 offset = e->end;
1185                 e->end += len;
1186                 users = e->users++;
1187         }
1188         spin_unlock(&con->writequeue_lock);
1189
1190         if (e) {
1191         got_one:
1192                 if (users == 0)
1193                         kmap(e->page);
1194                 *ppc = page_address(e->page) + offset;
1195                 return e;
1196         }
1197
1198         e = new_writequeue_entry(con, allocation);
1199         if (e) {
1200                 spin_lock(&con->writequeue_lock);
1201                 offset = e->end;
1202                 e->end += len;
1203                 users = e->users++;
1204                 list_add_tail(&e->list, &con->writequeue);
1205                 spin_unlock(&con->writequeue_lock);
1206                 goto got_one;
1207         }
1208         return NULL;
1209 }
1210
1211 void dlm_lowcomms_commit_buffer(void *mh)
1212 {
1213         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1214         struct connection *con = e->con;
1215         int users;
1216
1217         spin_lock(&con->writequeue_lock);
1218         users = --e->users;
1219         if (users)
1220                 goto out;
1221         e->len = e->end - e->offset;
1222         kunmap(e->page);
1223         spin_unlock(&con->writequeue_lock);
1224
1225         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1226                 queue_work(send_workqueue, &con->swork);
1227         }
1228         return;
1229
1230 out:
1231         spin_unlock(&con->writequeue_lock);
1232         return;
1233 }
1234
1235 /* Send a message */
1236 static void send_to_sock(struct connection *con)
1237 {
1238         int ret = 0;
1239         ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1240         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1241         struct writequeue_entry *e;
1242         int len, offset;
1243
1244         mutex_lock(&con->sock_mutex);
1245         if (con->sock == NULL)
1246                 goto out_connect;
1247
1248         sendpage = con->sock->ops->sendpage;
1249
1250         spin_lock(&con->writequeue_lock);
1251         for (;;) {
1252                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1253                                list);
1254                 if ((struct list_head *) e == &con->writequeue)
1255                         break;
1256
1257                 len = e->len;
1258                 offset = e->offset;
1259                 BUG_ON(len == 0 && e->users == 0);
1260                 spin_unlock(&con->writequeue_lock);
1261                 kmap(e->page);
1262
1263                 ret = 0;
1264                 if (len) {
1265                         ret = sendpage(con->sock, e->page, offset, len,
1266                                        msg_flags);
1267                         if (ret == -EAGAIN || ret == 0)
1268                                 goto out;
1269                         if (ret <= 0)
1270                                 goto send_error;
1271                 } else {
1272                         /* Don't starve people filling buffers */
1273                         cond_resched();
1274                 }
1275
1276                 spin_lock(&con->writequeue_lock);
1277                 e->offset += ret;
1278                 e->len -= ret;
1279
1280                 if (e->len == 0 && e->users == 0) {
1281                         list_del(&e->list);
1282                         kunmap(e->page);
1283                         free_entry(e);
1284                         continue;
1285                 }
1286         }
1287         spin_unlock(&con->writequeue_lock);
1288 out:
1289         mutex_unlock(&con->sock_mutex);
1290         return;
1291
1292 send_error:
1293         mutex_unlock(&con->sock_mutex);
1294         close_connection(con, false);
1295         lowcomms_connect_sock(con);
1296         return;
1297
1298 out_connect:
1299         mutex_unlock(&con->sock_mutex);
1300         if (!test_bit(CF_INIT_PENDING, &con->flags))
1301                 lowcomms_connect_sock(con);
1302         return;
1303 }
1304
1305 static void clean_one_writequeue(struct connection *con)
1306 {
1307         struct list_head *list;
1308         struct list_head *temp;
1309
1310         spin_lock(&con->writequeue_lock);
1311         list_for_each_safe(list, temp, &con->writequeue) {
1312                 struct writequeue_entry *e =
1313                         list_entry(list, struct writequeue_entry, list);
1314                 list_del(&e->list);
1315                 free_entry(e);
1316         }
1317         spin_unlock(&con->writequeue_lock);
1318 }
1319
1320 /* Called from recovery when it knows that a node has
1321    left the cluster */
1322 int dlm_lowcomms_close(int nodeid)
1323 {
1324         struct connection *con;
1325
1326         log_print("closing connection to node %d", nodeid);
1327         con = nodeid2con(nodeid, 0);
1328         if (con) {
1329                 clean_one_writequeue(con);
1330                 close_connection(con, true);
1331         }
1332         return 0;
1333 }
1334
1335 /* Receive workqueue function */
1336 static void process_recv_sockets(struct work_struct *work)
1337 {
1338         struct connection *con = container_of(work, struct connection, rwork);
1339         int err;
1340
1341         clear_bit(CF_READ_PENDING, &con->flags);
1342         do {
1343                 err = con->rx_action(con);
1344         } while (!err);
1345 }
1346
1347 /* Send workqueue function */
1348 static void process_send_sockets(struct work_struct *work)
1349 {
1350         struct connection *con = container_of(work, struct connection, swork);
1351
1352         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1353                 con->connect_action(con);
1354         }
1355         clear_bit(CF_WRITE_PENDING, &con->flags);
1356         send_to_sock(con);
1357 }
1358
1359
1360 /* Discard all entries on the write queues */
1361 static void clean_writequeues(void)
1362 {
1363         int nodeid;
1364
1365         for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1366                 struct connection *con = __nodeid2con(nodeid, 0);
1367
1368                 if (con)
1369                         clean_one_writequeue(con);
1370         }
1371 }
1372
1373 static void work_stop(void)
1374 {
1375         destroy_workqueue(recv_workqueue);
1376         destroy_workqueue(send_workqueue);
1377 }
1378
1379 static int work_start(void)
1380 {
1381         int error;
1382         recv_workqueue = create_workqueue("dlm_recv");
1383         error = IS_ERR(recv_workqueue);
1384         if (error) {
1385                 log_print("can't start dlm_recv %d", error);
1386                 return error;
1387         }
1388
1389         send_workqueue = create_singlethread_workqueue("dlm_send");
1390         error = IS_ERR(send_workqueue);
1391         if (error) {
1392                 log_print("can't start dlm_send %d", error);
1393                 destroy_workqueue(recv_workqueue);
1394                 return error;
1395         }
1396
1397         return 0;
1398 }
1399
1400 void dlm_lowcomms_stop(void)
1401 {
1402         int i;
1403         struct connection *con;
1404
1405         /* Set all the flags to prevent any
1406            socket activity.
1407         */
1408         down(&connections_lock);
1409         for (i = 0; i <= max_nodeid; i++) {
1410                 con = __nodeid2con(i, 0);
1411                 if (con) {
1412                         con->flags |= 0xFF;
1413                         if (con->sock)
1414                                 con->sock->sk->sk_user_data = NULL;
1415                 }
1416         }
1417         up(&connections_lock);
1418
1419         work_stop();
1420
1421         down(&connections_lock);
1422         clean_writequeues();
1423
1424         for (i = 0; i <= max_nodeid; i++) {
1425                 con = __nodeid2con(i, 0);
1426                 if (con) {
1427                         close_connection(con, true);
1428                         if (con->othercon)
1429                                 kmem_cache_free(con_cache, con->othercon);
1430                         kmem_cache_free(con_cache, con);
1431                 }
1432         }
1433         max_nodeid = 0;
1434         up(&connections_lock);
1435         kmem_cache_destroy(con_cache);
1436         idr_init(&connections_idr);
1437 }
1438
1439 int dlm_lowcomms_start(void)
1440 {
1441         int error = -EINVAL;
1442         struct connection *con;
1443
1444         init_local();
1445         if (!dlm_local_count) {
1446                 error = -ENOTCONN;
1447                 log_print("no local IP address has been set");
1448                 goto out;
1449         }
1450
1451         error = -ENOMEM;
1452         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1453                                       __alignof__(struct connection), 0,
1454                                       NULL);
1455         if (!con_cache)
1456                 goto out;
1457
1458         /* Set some sysctl minima */
1459         if (sysctl_rmem_max < NEEDED_RMEM)
1460                 sysctl_rmem_max = NEEDED_RMEM;
1461
1462         /* Start listening */
1463         if (dlm_config.ci_protocol == 0)
1464                 error = tcp_listen_for_all();
1465         else
1466                 error = sctp_listen_for_all();
1467         if (error)
1468                 goto fail_unlisten;
1469
1470         error = work_start();
1471         if (error)
1472                 goto fail_unlisten;
1473
1474         return 0;
1475
1476 fail_unlisten:
1477         con = nodeid2con(0,0);
1478         if (con) {
1479                 close_connection(con, false);
1480                 kmem_cache_free(con_cache, con);
1481         }
1482         kmem_cache_destroy(con_cache);
1483
1484 out:
1485         return error;
1486 }