SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is its
  25  * responsibility. It is this layer's
  26  * responsibility to resolve these into IP address or
  27  * whatever it needs for inter-node communication.
  28  *
  29  * The comms level is two kernel threads that deal mainly with
  30  * the receiving of messages from other nodes and passing them
  31  * up to the mid-level comms layer (which understands the
  32  * message format) for execution by the locking core, and
  33  * a send thread which does all the setting up of connections
  34  * to remote nodes and the sending of data. Threads are not allowed
  35  * to send their own data because it may cause them to wait in times
  36  * of high load. Also, this way, the sending thread can collect together
  37  * messages bound for one node and send them in one block.
  38  *
  39  * lowcomms will choose to use either TCP or SCTP as its transport layer
  40  * depending on the configuration variable 'protocol'. This should be set
  41  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  42  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  43  * for the DLM to function.
  44  *
  45  */
  46
  47 #include <asm/ioctls.h>
  48 #include <net/sock.h>
  49 #include <net/tcp.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/file.h>
  52 #include <linux/mutex.h>
  53 #include <linux/sctp.h>
  54 #include <net/sctp/user.h>
  55 #include <net/ipv6.h>
  56
  57 #include "dlm_internal.h"
  58 #include "lowcomms.h"
  59 #include "midcomms.h"
  60 #include "config.h"
  61
  62 #define NEEDED_RMEM (4*1024*1024)
  63 #define CONN_HASH_SIZE 32
  64
  65 struct cbuf {
  66         unsigned int base;
  67         unsigned int len;
  68         unsigned int mask;
  69 };
  70
  71 static void cbuf_add(struct cbuf *cb, int n)
  72 {
  73         cb->len += n;
  74 }
  75
  76 static int cbuf_data(struct cbuf *cb)
  77 {
  78         return ((cb->base + cb->len) & cb->mask);
  79 }
  80
  81 static void cbuf_init(struct cbuf *cb, int size)
  82 {
  83         cb->base = cb->len = 0;
  84         cb->mask = size-1;
  85 }
  86
  87 static void cbuf_eat(struct cbuf *cb, int n)
  88 {
  89         cb->len  -= n;
  90         cb->base += n;
  91         cb->base &= cb->mask;
  92 }
  93
  94 static bool cbuf_empty(struct cbuf *cb)
  95 {
  96         return cb->len == 0;
  97 }
  98
  99 struct connection {
 100         struct socket *sock;    /* NULL if not connected */
 101         uint32_t nodeid;        /* So we know who we are in the list */
 102         struct mutex sock_mutex;
 103         unsigned long flags;
 104 #define CF_READ_PENDING 1
 105 #define CF_WRITE_PENDING 2
 106 #define CF_CONNECT_PENDING 3
 107 #define CF_INIT_PENDING 4
 108 #define CF_IS_OTHERCON 5
 109 #define CF_CLOSE 6
 110         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 111         spinlock_t writequeue_lock;
 112         int (*rx_action) (struct connection *); /* What to do when active */
 113         void (*connect_action) (struct connection *);   /* What to do to connect */
 114         struct page *rx_page;
 115         struct cbuf cb;
 116         int retries;
 117 #define MAX_CONNECT_RETRIES 3
 118         int sctp_assoc;
 119         struct hlist_node list;
 120         struct connection *othercon;
 121         struct work_struct rwork; /* Receive workqueue */
 122         struct work_struct swork; /* Send workqueue */
 123 };
 124 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 125
 126 /* An entry waiting to be sent */
 127 struct writequeue_entry {
 128         struct list_head list;
 129         struct page *page;
 130         int offset;
 131         int len;
 132         int end;
 133         int users;
 134         struct connection *con;
 135 };
 136
 137 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 138 static int dlm_local_count;
 139
 140 /* Work queues */
 141 static struct workqueue_struct *recv_workqueue;
 142 static struct workqueue_struct *send_workqueue;
 143
 144 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 145 static DEFINE_MUTEX(connections_lock);
 146 static struct kmem_cache *con_cache;
 147
 148 static void process_recv_sockets(struct work_struct *work);
 149 static void process_send_sockets(struct work_struct *work);
 150
 151
 152 /* This is deliberately very simple because most clusters have simple
 153    sequential nodeids, so we should be able to go straight to a connection
 154    struct in the array */
 155 static inline int nodeid_hash(int nodeid)
 156 {
 157         return nodeid & (CONN_HASH_SIZE-1);
 158 }
 159
 160 static struct connection *__find_con(int nodeid)
 161 {
 162         int r;
 163         struct hlist_node *h;
 164         struct connection *con;
 165
 166         r = nodeid_hash(nodeid);
 167
 168         hlist_for_each_entry(con, h, &connection_hash[r], list) {
 169                 if (con->nodeid == nodeid)
 170                         return con;
 171         }
 172         return NULL;
 173 }
 174
 175 /*
 176  * If 'allocation' is zero then we don't attempt to create a new
 177  * connection structure for this node.
 178  */
 179 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 180 {
 181         struct connection *con = NULL;
 182         int r;
 183
 184         con = __find_con(nodeid);
 185         if (con || !alloc)
 186                 return con;
 187
 188         con = kmem_cache_zalloc(con_cache, alloc);
 189         if (!con)
 190                 return NULL;
 191
 192         r = nodeid_hash(nodeid);
 193         hlist_add_head(&con->list, &connection_hash[r]);
 194
 195         con->nodeid = nodeid;
 196         mutex_init(&con->sock_mutex);
 197         INIT_LIST_HEAD(&con->writequeue);
 198         spin_lock_init(&con->writequeue_lock);
 199         INIT_WORK(&con->swork, process_send_sockets);
 200         INIT_WORK(&con->rwork, process_recv_sockets);
 201
 202         /* Setup action pointers for child sockets */
 203         if (con->nodeid) {
 204                 struct connection *zerocon = __find_con(0);
 205
 206                 con->connect_action = zerocon->connect_action;
 207                 if (!con->rx_action)
 208                         con->rx_action = zerocon->rx_action;
 209         }
 210
 211         return con;
 212 }
 213
 214 /* Loop round all connections */
 215 static void foreach_conn(void (*conn_func)(struct connection *c))
 216 {
 217         int i;
 218         struct hlist_node *h, *n;
 219         struct connection *con;
 220
 221         for (i = 0; i < CONN_HASH_SIZE; i++) {
 222                 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
 223                         conn_func(con);
 224                 }
 225         }
 226 }
 227
 228 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 229 {
 230         struct connection *con;
 231
 232         mutex_lock(&connections_lock);
 233         con = __nodeid2con(nodeid, allocation);
 234         mutex_unlock(&connections_lock);
 235
 236         return con;
 237 }
 238
 239 /* This is a bit drastic, but only called when things go wrong */
 240 static struct connection *assoc2con(int assoc_id)
 241 {
 242         int i;
 243         struct hlist_node *h;
 244         struct connection *con;
 245
 246         mutex_lock(&connections_lock);
 247
 248         for (i = 0 ; i < CONN_HASH_SIZE; i++) {
 249                 hlist_for_each_entry(con, h, &connection_hash[i], list) {
 250                         if (con && con->sctp_assoc == assoc_id) {
 251                                 mutex_unlock(&connections_lock);
 252                                 return con;
 253                         }
 254                 }
 255         }
 256         mutex_unlock(&connections_lock);
 257         return NULL;
 258 }
 259
 260 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 261 {
 262         struct sockaddr_storage addr;
 263         int error;
 264
 265         if (!dlm_local_count)
 266                 return -1;
 267
 268         error = dlm_nodeid_to_addr(nodeid, &addr);
 269         if (error)
 270                 return error;
 271
 272         if (dlm_local_addr[0]->ss_family == AF_INET) {
 273                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 274                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 275                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 276         } else {
 277                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 278                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 279                 ipv6_addr_copy(&ret6->sin6_addr, &in6->sin6_addr);
 280         }
 281
 282         return 0;
 283 }
 284
 285 /* Data available on socket or listen socket received a connect */
 286 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 287 {
 288         struct connection *con = sock2con(sk);
 289         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 290                 queue_work(recv_workqueue, &con->rwork);
 291 }
 292
 293 static void lowcomms_write_space(struct sock *sk)
 294 {
 295         struct connection *con = sock2con(sk);
 296
 297         if (con && !test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 298                 queue_work(send_workqueue, &con->swork);
 299 }
 300
 301 static inline void lowcomms_connect_sock(struct connection *con)
 302 {
 303         if (test_bit(CF_CLOSE, &con->flags))
 304                 return;
 305         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 306                 queue_work(send_workqueue, &con->swork);
 307 }
 308
 309 static void lowcomms_state_change(struct sock *sk)
 310 {
 311         if (sk->sk_state == TCP_ESTABLISHED)
 312                 lowcomms_write_space(sk);
 313 }
 314
 315 int dlm_lowcomms_connect_node(int nodeid)
 316 {
 317         struct connection *con;
 318
 319         /* with sctp there's no connecting without sending */
 320         if (dlm_config.ci_protocol != 0)
 321                 return 0;
 322
 323         if (nodeid == dlm_our_nodeid())
 324                 return 0;
 325
 326         con = nodeid2con(nodeid, GFP_NOFS);
 327         if (!con)
 328                 return -ENOMEM;
 329         lowcomms_connect_sock(con);
 330         return 0;
 331 }
 332
 333 /* Make a socket active */
 334 static int add_sock(struct socket *sock, struct connection *con)
 335 {
 336         con->sock = sock;
 337
 338         /* Install a data_ready callback */
 339         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 340         con->sock->sk->sk_write_space = lowcomms_write_space;
 341         con->sock->sk->sk_state_change = lowcomms_state_change;
 342         con->sock->sk->sk_user_data = con;
 343         con->sock->sk->sk_allocation = GFP_NOFS;
 344         return 0;
 345 }
 346
 347 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 348    length */
 349 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 350                           int *addr_len)
 351 {
 352         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 353         if (saddr->ss_family == AF_INET) {
 354                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 355                 in4_addr->sin_port = cpu_to_be16(port);
 356                 *addr_len = sizeof(struct sockaddr_in);
 357                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 358         } else {
 359                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 360                 in6_addr->sin6_port = cpu_to_be16(port);
 361                 *addr_len = sizeof(struct sockaddr_in6);
 362         }
 363         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 364 }
 365
 366 /* Close a remote connection and tidy up */
 367 static void close_connection(struct connection *con, bool and_other)
 368 {
 369         mutex_lock(&con->sock_mutex);
 370
 371         if (con->sock) {
 372                 sock_release(con->sock);
 373                 con->sock = NULL;
 374         }
 375         if (con->othercon && and_other) {
 376                 /* Will only re-enter once. */
 377                 close_connection(con->othercon, false);
 378         }
 379         if (con->rx_page) {
 380                 __free_page(con->rx_page);
 381                 con->rx_page = NULL;
 382         }
 383
 384         con->retries = 0;
 385         mutex_unlock(&con->sock_mutex);
 386 }
 387
 388 /* We only send shutdown messages to nodes that are not part of the cluster */
 389 static void sctp_send_shutdown(sctp_assoc_t associd)
 390 {
 391         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 392         struct msghdr outmessage;
 393         struct cmsghdr *cmsg;
 394         struct sctp_sndrcvinfo *sinfo;
 395         int ret;
 396         struct connection *con;
 397
 398         con = nodeid2con(0,0);
 399         BUG_ON(con == NULL);
 400
 401         outmessage.msg_name = NULL;
 402         outmessage.msg_namelen = 0;
 403         outmessage.msg_control = outcmsg;
 404         outmessage.msg_controllen = sizeof(outcmsg);
 405         outmessage.msg_flags = MSG_EOR;
 406
 407         cmsg = CMSG_FIRSTHDR(&outmessage);
 408         cmsg->cmsg_level = IPPROTO_SCTP;
 409         cmsg->cmsg_type = SCTP_SNDRCV;
 410         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 411         outmessage.msg_controllen = cmsg->cmsg_len;
 412         sinfo = CMSG_DATA(cmsg);
 413         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 414
 415         sinfo->sinfo_flags |= MSG_EOF;
 416         sinfo->sinfo_assoc_id = associd;
 417
 418         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 419
 420         if (ret != 0)
 421                 log_print("send EOF to node failed: %d", ret);
 422 }
 423
 424 static void sctp_init_failed_foreach(struct connection *con)
 425 {
 426         con->sctp_assoc = 0;
 427         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 428                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 429                         queue_work(send_workqueue, &con->swork);
 430         }
 431 }
 432
 433 /* INIT failed but we don't know which node...
 434    restart INIT on all pending nodes */
 435 static void sctp_init_failed(void)
 436 {
 437         mutex_lock(&connections_lock);
 438
 439         foreach_conn(sctp_init_failed_foreach);
 440
 441         mutex_unlock(&connections_lock);
 442 }
 443
 444 /* Something happened to an association */
 445 static void process_sctp_notification(struct connection *con,
 446                                       struct msghdr *msg, char *buf)
 447 {
 448         union sctp_notification *sn = (union sctp_notification *)buf;
 449
 450         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 451                 switch (sn->sn_assoc_change.sac_state) {
 452
 453                 case SCTP_COMM_UP:
 454                 case SCTP_RESTART:
 455                 {
 456                         /* Check that the new node is in the lockspace */
 457                         struct sctp_prim prim;
 458                         int nodeid;
 459                         int prim_len, ret;
 460                         int addr_len;
 461                         struct connection *new_con;
 462                         sctp_peeloff_arg_t parg;
 463                         int parglen = sizeof(parg);
 464                         int err;
 465
 466                         /*
 467                          * We get this before any data for an association.
 468                          * We verify that the node is in the cluster and
 469                          * then peel off a socket for it.
 470                          */
 471                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 472                                 log_print("COMM_UP for invalid assoc ID %d",
 473                                          (int)sn->sn_assoc_change.sac_assoc_id);
 474                                 sctp_init_failed();
 475                                 return;
 476                         }
 477                         memset(&prim, 0, sizeof(struct sctp_prim));
 478                         prim_len = sizeof(struct sctp_prim);
 479                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 480
 481                         ret = kernel_getsockopt(con->sock,
 482                                                 IPPROTO_SCTP,
 483                                                 SCTP_PRIMARY_ADDR,
 484                                                 (char*)&prim,
 485                                                 &prim_len);
 486                         if (ret < 0) {
 487                                 log_print("getsockopt/sctp_primary_addr on "
 488                                           "new assoc %d failed : %d",
 489                                           (int)sn->sn_assoc_change.sac_assoc_id,
 490                                           ret);
 491
 492                                 /* Retry INIT later */
 493                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 494                                 if (new_con)
 495                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 496                                 return;
 497                         }
 498                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 499                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 500                                 int i;
 501                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 502                                 log_print("reject connect from unknown addr");
 503                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 504                                         printk("%02x ", b[i]);
 505                                 printk("\n");
 506                                 sctp_send_shutdown(prim.ssp_assoc_id);
 507                                 return;
 508                         }
 509
 510                         new_con = nodeid2con(nodeid, GFP_NOFS);
 511                         if (!new_con)
 512                                 return;
 513
 514                         /* Peel off a new sock */
 515                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 516                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 517                                                 SCTP_SOCKOPT_PEELOFF,
 518                                                 (void *)&parg, &parglen);
 519                         if (ret < 0) {
 520                                 log_print("Can't peel off a socket for "
 521                                           "connection %d to node %d: err=%d",
 522                                           parg.associd, nodeid, ret);
 523                                 return;
 524                         }
 525                         new_con->sock = sockfd_lookup(parg.sd, &err);
 526                         if (!new_con->sock) {
 527                                 log_print("sockfd_lookup error %d", err);
 528                                 return;
 529                         }
 530                         add_sock(new_con->sock, new_con);
 531                         sockfd_put(new_con->sock);
 532
 533                         log_print("connecting to %d sctp association %d",
 534                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 535
 536                         /* Send any pending writes */
 537                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 538                         clear_bit(CF_INIT_PENDING, &con->flags);
 539                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 540                                 queue_work(send_workqueue, &new_con->swork);
 541                         }
 542                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 543                                 queue_work(recv_workqueue, &new_con->rwork);
 544                 }
 545                 break;
 546
 547                 case SCTP_COMM_LOST:
 548                 case SCTP_SHUTDOWN_COMP:
 549                 {
 550                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 551                         if (con) {
 552                                 con->sctp_assoc = 0;
 553                         }
 554                 }
 555                 break;
 556
 557                 /* We don't know which INIT failed, so clear the PENDING flags
 558                  * on them all.  if assoc_id is zero then it will then try
 559                  * again */
 560
 561                 case SCTP_CANT_STR_ASSOC:
 562                 {
 563                         log_print("Can't start SCTP association - retrying");
 564                         sctp_init_failed();
 565                 }
 566                 break;
 567
 568                 default:
 569                         log_print("unexpected SCTP assoc change id=%d state=%d",
 570                                   (int)sn->sn_assoc_change.sac_assoc_id,
 571                                   sn->sn_assoc_change.sac_state);
 572                 }
 573         }
 574 }
 575
 576 /* Data received from remote end */
 577 static int receive_from_sock(struct connection *con)
 578 {
 579         int ret = 0;
 580         struct msghdr msg = {};
 581         struct kvec iov[2];
 582         unsigned len;
 583         int r;
 584         int call_again_soon = 0;
 585         int nvec;
 586         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 587
 588         mutex_lock(&con->sock_mutex);
 589
 590         if (con->sock == NULL) {
 591                 ret = -EAGAIN;
 592                 goto out_close;
 593         }
 594
 595         if (con->rx_page == NULL) {
 596                 /*
 597                  * This doesn't need to be atomic, but I think it should
 598                  * improve performance if it is.
 599                  */
 600                 con->rx_page = alloc_page(GFP_ATOMIC);
 601                 if (con->rx_page == NULL)
 602                         goto out_resched;
 603                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 604         }
 605
 606         /* Only SCTP needs these really */
 607         memset(&incmsg, 0, sizeof(incmsg));
 608         msg.msg_control = incmsg;
 609         msg.msg_controllen = sizeof(incmsg);
 610
 611         /*
 612          * iov[0] is the bit of the circular buffer between the current end
 613          * point (cb.base + cb.len) and the end of the buffer.
 614          */
 615         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 616         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 617         iov[1].iov_len = 0;
 618         nvec = 1;
 619
 620         /*
 621          * iov[1] is the bit of the circular buffer between the start of the
 622          * buffer and the start of the currently used section (cb.base)
 623          */
 624         if (cbuf_data(&con->cb) >= con->cb.base) {
 625                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 626                 iov[1].iov_len = con->cb.base;
 627                 iov[1].iov_base = page_address(con->rx_page);
 628                 nvec = 2;
 629         }
 630         len = iov[0].iov_len + iov[1].iov_len;
 631
 632         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 633                                MSG_DONTWAIT | MSG_NOSIGNAL);
 634         if (ret <= 0)
 635                 goto out_close;
 636
 637         /* Process SCTP notifications */
 638         if (msg.msg_flags & MSG_NOTIFICATION) {
 639                 msg.msg_control = incmsg;
 640                 msg.msg_controllen = sizeof(incmsg);
 641
 642                 process_sctp_notification(con, &msg,
 643                                 page_address(con->rx_page) + con->cb.base);
 644                 mutex_unlock(&con->sock_mutex);
 645                 return 0;
 646         }
 647         BUG_ON(con->nodeid == 0);
 648
 649         if (ret == len)
 650                 call_again_soon = 1;
 651         cbuf_add(&con->cb, ret);
 652         ret = dlm_process_incoming_buffer(con->nodeid,
 653                                           page_address(con->rx_page),
 654                                           con->cb.base, con->cb.len,
 655                                           PAGE_CACHE_SIZE);
 656         if (ret == -EBADMSG) {
 657                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 658                           "iov_len=%u, iov_base[0]=%p, read=%d",
 659                           page_address(con->rx_page), con->cb.base, con->cb.len,
 660                           len, iov[0].iov_base, r);
 661         }
 662         if (ret < 0)
 663                 goto out_close;
 664         cbuf_eat(&con->cb, ret);
 665
 666         if (cbuf_empty(&con->cb) && !call_again_soon) {
 667                 __free_page(con->rx_page);
 668                 con->rx_page = NULL;
 669         }
 670
 671         if (call_again_soon)
 672                 goto out_resched;
 673         mutex_unlock(&con->sock_mutex);
 674         return 0;
 675
 676 out_resched:
 677         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 678                 queue_work(recv_workqueue, &con->rwork);
 679         mutex_unlock(&con->sock_mutex);
 680         return -EAGAIN;
 681
 682 out_close:
 683         mutex_unlock(&con->sock_mutex);
 684         if (ret != -EAGAIN) {
 685                 close_connection(con, false);
 686                 /* Reconnect when there is something to send */
 687         }
 688         /* Don't return success if we really got EOF */
 689         if (ret == 0)
 690                 ret = -EAGAIN;
 691
 692         return ret;
 693 }
 694
 695 /* Listening socket is busy, accept a connection */
 696 static int tcp_accept_from_sock(struct connection *con)
 697 {
 698         int result;
 699         struct sockaddr_storage peeraddr;
 700         struct socket *newsock;
 701         int len;
 702         int nodeid;
 703         struct connection *newcon;
 704         struct connection *addcon;
 705
 706         memset(&peeraddr, 0, sizeof(peeraddr));
 707         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 708                                   IPPROTO_TCP, &newsock);
 709         if (result < 0)
 710                 return -ENOMEM;
 711
 712         mutex_lock_nested(&con->sock_mutex, 0);
 713
 714         result = -ENOTCONN;
 715         if (con->sock == NULL)
 716                 goto accept_err;
 717
 718         newsock->type = con->sock->type;
 719         newsock->ops = con->sock->ops;
 720
 721         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 722         if (result < 0)
 723                 goto accept_err;
 724
 725         /* Get the connected socket's peer */
 726         memset(&peeraddr, 0, sizeof(peeraddr));
 727         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 728                                   &len, 2)) {
 729                 result = -ECONNABORTED;
 730                 goto accept_err;
 731         }
 732
 733         /* Get the new node's NODEID */
 734         make_sockaddr(&peeraddr, 0, &len);
 735         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 736                 log_print("connect from non cluster node");
 737                 sock_release(newsock);
 738                 mutex_unlock(&con->sock_mutex);
 739                 return -1;
 740         }
 741
 742         log_print("got connection from %d", nodeid);
 743
 744         /*  Check to see if we already have a connection to this node. This
 745          *  could happen if the two nodes initiate a connection at roughly
 746          *  the same time and the connections cross on the wire.
 747          *  In this case we store the incoming one in "othercon"
 748          */
 749         newcon = nodeid2con(nodeid, GFP_NOFS);
 750         if (!newcon) {
 751                 result = -ENOMEM;
 752                 goto accept_err;
 753         }
 754         mutex_lock_nested(&newcon->sock_mutex, 1);
 755         if (newcon->sock) {
 756                 struct connection *othercon = newcon->othercon;
 757
 758                 if (!othercon) {
 759                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 760                         if (!othercon) {
 761                                 log_print("failed to allocate incoming socket");
 762                                 mutex_unlock(&newcon->sock_mutex);
 763                                 result = -ENOMEM;
 764                                 goto accept_err;
 765                         }
 766                         othercon->nodeid = nodeid;
 767                         othercon->rx_action = receive_from_sock;
 768                         mutex_init(&othercon->sock_mutex);
 769                         INIT_WORK(&othercon->swork, process_send_sockets);
 770                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 771                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 772                 }
 773                 if (!othercon->sock) {
 774                         newcon->othercon = othercon;
 775                         othercon->sock = newsock;
 776                         newsock->sk->sk_user_data = othercon;
 777                         add_sock(newsock, othercon);
 778                         addcon = othercon;
 779                 }
 780                 else {
 781                         printk("Extra connection from node %d attempted\n", nodeid);
 782                         result = -EAGAIN;
 783                         mutex_unlock(&newcon->sock_mutex);
 784                         goto accept_err;
 785                 }
 786         }
 787         else {
 788                 newsock->sk->sk_user_data = newcon;
 789                 newcon->rx_action = receive_from_sock;
 790                 add_sock(newsock, newcon);
 791                 addcon = newcon;
 792         }
 793
 794         mutex_unlock(&newcon->sock_mutex);
 795
 796         /*
 797          * Add it to the active queue in case we got data
 798          * beween processing the accept adding the socket
 799          * to the read_sockets list
 800          */
 801         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 802                 queue_work(recv_workqueue, &addcon->rwork);
 803         mutex_unlock(&con->sock_mutex);
 804
 805         return 0;
 806
 807 accept_err:
 808         mutex_unlock(&con->sock_mutex);
 809         sock_release(newsock);
 810
 811         if (result != -EAGAIN)
 812                 log_print("error accepting connection from node: %d", result);
 813         return result;
 814 }
 815
 816 static void free_entry(struct writequeue_entry *e)
 817 {
 818         __free_page(e->page);
 819         kfree(e);
 820 }
 821
 822 /* Initiate an SCTP association.
 823    This is a special case of send_to_sock() in that we don't yet have a
 824    peeled-off socket for this association, so we use the listening socket
 825    and add the primary IP address of the remote node.
 826  */
 827 static void sctp_init_assoc(struct connection *con)
 828 {
 829         struct sockaddr_storage rem_addr;
 830         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 831         struct msghdr outmessage;
 832         struct cmsghdr *cmsg;
 833         struct sctp_sndrcvinfo *sinfo;
 834         struct connection *base_con;
 835         struct writequeue_entry *e;
 836         int len, offset;
 837         int ret;
 838         int addrlen;
 839         struct kvec iov[1];
 840
 841         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 842                 return;
 843
 844         if (con->retries++ > MAX_CONNECT_RETRIES)
 845                 return;
 846
 847         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 848                 log_print("no address for nodeid %d", con->nodeid);
 849                 return;
 850         }
 851         base_con = nodeid2con(0, 0);
 852         BUG_ON(base_con == NULL);
 853
 854         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 855
 856         outmessage.msg_name = &rem_addr;
 857         outmessage.msg_namelen = addrlen;
 858         outmessage.msg_control = outcmsg;
 859         outmessage.msg_controllen = sizeof(outcmsg);
 860         outmessage.msg_flags = MSG_EOR;
 861
 862         spin_lock(&con->writequeue_lock);
 863
 864         if (list_empty(&con->writequeue)) {
 865                 spin_unlock(&con->writequeue_lock);
 866                 log_print("writequeue empty for nodeid %d", con->nodeid);
 867                 return;
 868         }
 869
 870         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
 871         len = e->len;
 872         offset = e->offset;
 873         spin_unlock(&con->writequeue_lock);
 874
 875         /* Send the first block off the write queue */
 876         iov[0].iov_base = page_address(e->page)+offset;
 877         iov[0].iov_len = len;
 878
 879         cmsg = CMSG_FIRSTHDR(&outmessage);
 880         cmsg->cmsg_level = IPPROTO_SCTP;
 881         cmsg->cmsg_type = SCTP_SNDRCV;
 882         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 883         sinfo = CMSG_DATA(cmsg);
 884         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 885         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 886         outmessage.msg_controllen = cmsg->cmsg_len;
 887
 888         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 889         if (ret < 0) {
 890                 log_print("Send first packet to node %d failed: %d",
 891                           con->nodeid, ret);
 892
 893                 /* Try again later */
 894                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 895                 clear_bit(CF_INIT_PENDING, &con->flags);
 896         }
 897         else {
 898                 spin_lock(&con->writequeue_lock);
 899                 e->offset += ret;
 900                 e->len -= ret;
 901
 902                 if (e->len == 0 && e->users == 0) {
 903                         list_del(&e->list);
 904                         free_entry(e);
 905                 }
 906                 spin_unlock(&con->writequeue_lock);
 907         }
 908 }
 909
 910 /* Connect a new socket to its peer */
 911 static void tcp_connect_to_sock(struct connection *con)
 912 {
 913         int result = -EHOSTUNREACH;
 914         struct sockaddr_storage saddr, src_addr;
 915         int addr_len;
 916         struct socket *sock = NULL;
 917
 918         if (con->nodeid == 0) {
 919                 log_print("attempt to connect sock 0 foiled");
 920                 return;
 921         }
 922
 923         mutex_lock(&con->sock_mutex);
 924         if (con->retries++ > MAX_CONNECT_RETRIES)
 925                 goto out;
 926
 927         /* Some odd races can cause double-connects, ignore them */
 928         if (con->sock) {
 929                 result = 0;
 930                 goto out;
 931         }
 932
 933         /* Create a socket to communicate with */
 934         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 935                                   IPPROTO_TCP, &sock);
 936         if (result < 0)
 937                 goto out_err;
 938
 939         memset(&saddr, 0, sizeof(saddr));
 940         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 941                 goto out_err;
 942
 943         sock->sk->sk_user_data = con;
 944         con->rx_action = receive_from_sock;
 945         con->connect_action = tcp_connect_to_sock;
 946         add_sock(sock, con);
 947
 948         /* Bind to our cluster-known address connecting to avoid
 949            routing problems */
 950         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 951         make_sockaddr(&src_addr, 0, &addr_len);
 952         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 953                                  addr_len);
 954         if (result < 0) {
 955                 log_print("could not bind for connect: %d", result);
 956                 /* This *may* not indicate a critical error */
 957         }
 958
 959         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 960
 961         log_print("connecting to %d", con->nodeid);
 962         result =
 963                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 964                                    O_NONBLOCK);
 965         if (result == -EINPROGRESS)
 966                 result = 0;
 967         if (result == 0)
 968                 goto out;
 969
 970 out_err:
 971         if (con->sock) {
 972                 sock_release(con->sock);
 973                 con->sock = NULL;
 974         } else if (sock) {
 975                 sock_release(sock);
 976         }
 977         /*
 978          * Some errors are fatal and this list might need adjusting. For other
 979          * errors we try again until the max number of retries is reached.
 980          */
 981         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 982             result != -ENETDOWN && result != -EINVAL
 983             && result != -EPROTONOSUPPORT) {
 984                 lowcomms_connect_sock(con);
 985                 result = 0;
 986         }
 987 out:
 988         mutex_unlock(&con->sock_mutex);
 989         return;
 990 }
 991
 992 static struct socket *tcp_create_listen_sock(struct connection *con,
 993                                              struct sockaddr_storage *saddr)
 994 {
 995         struct socket *sock = NULL;
 996         int result = 0;
 997         int one = 1;
 998         int addr_len;
 999
1000         if (dlm_local_addr[0]->ss_family == AF_INET)
1001                 addr_len = sizeof(struct sockaddr_in);
1002         else
1003                 addr_len = sizeof(struct sockaddr_in6);
1004
1005         /* Create a socket to communicate with */
1006         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1007                                   IPPROTO_TCP, &sock);
1008         if (result < 0) {
1009                 log_print("Can't create listening comms socket");
1010                 goto create_out;
1011         }
1012
1013         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1014                                    (char *)&one, sizeof(one));
1015
1016         if (result < 0) {
1017                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1018         }
1019         sock->sk->sk_user_data = con;
1020         con->rx_action = tcp_accept_from_sock;
1021         con->connect_action = tcp_connect_to_sock;
1022         con->sock = sock;
1023
1024         /* Bind to our port */
1025         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1026         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1027         if (result < 0) {
1028                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1029                 sock_release(sock);
1030                 sock = NULL;
1031                 con->sock = NULL;
1032                 goto create_out;
1033         }
1034         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1035                                  (char *)&one, sizeof(one));
1036         if (result < 0) {
1037                 log_print("Set keepalive failed: %d", result);
1038         }
1039
1040         result = sock->ops->listen(sock, 5);
1041         if (result < 0) {
1042                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1043                 sock_release(sock);
1044                 sock = NULL;
1045                 goto create_out;
1046         }
1047
1048 create_out:
1049         return sock;
1050 }
1051
1052 /* Get local addresses */
1053 static void init_local(void)
1054 {
1055         struct sockaddr_storage sas, *addr;
1056         int i;
1057
1058         dlm_local_count = 0;
1059         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1060                 if (dlm_our_addr(&sas, i))
1061                         break;
1062
1063                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1064                 if (!addr)
1065                         break;
1066                 memcpy(addr, &sas, sizeof(*addr));
1067                 dlm_local_addr[dlm_local_count++] = addr;
1068         }
1069 }
1070
1071 /* Bind to an IP address. SCTP allows multiple address so it can do
1072    multi-homing */
1073 static int add_sctp_bind_addr(struct connection *sctp_con,
1074                               struct sockaddr_storage *addr,
1075                               int addr_len, int num)
1076 {
1077         int result = 0;
1078
1079         if (num == 1)
1080                 result = kernel_bind(sctp_con->sock,
1081                                      (struct sockaddr *) addr,
1082                                      addr_len);
1083         else
1084                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1085                                            SCTP_SOCKOPT_BINDX_ADD,
1086                                            (char *)addr, addr_len);
1087
1088         if (result < 0)
1089                 log_print("Can't bind to port %d addr number %d",
1090                           dlm_config.ci_tcp_port, num);
1091
1092         return result;
1093 }
1094
1095 /* Initialise SCTP socket and bind to all interfaces */
1096 static int sctp_listen_for_all(void)
1097 {
1098         struct socket *sock = NULL;
1099         struct sockaddr_storage localaddr;
1100         struct sctp_event_subscribe subscribe;
1101         int result = -EINVAL, num = 1, i, addr_len;
1102         struct connection *con = nodeid2con(0, GFP_KERNEL);
1103         int bufsize = NEEDED_RMEM;
1104
1105         if (!con)
1106                 return -ENOMEM;
1107
1108         log_print("Using SCTP for communications");
1109
1110         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1111                                   IPPROTO_SCTP, &sock);
1112         if (result < 0) {
1113                 log_print("Can't create comms socket, check SCTP is loaded");
1114                 goto out;
1115         }
1116
1117         /* Listen for events */
1118         memset(&subscribe, 0, sizeof(subscribe));
1119         subscribe.sctp_data_io_event = 1;
1120         subscribe.sctp_association_event = 1;
1121         subscribe.sctp_send_failure_event = 1;
1122         subscribe.sctp_shutdown_event = 1;
1123         subscribe.sctp_partial_delivery_event = 1;
1124
1125         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1126                                  (char *)&bufsize, sizeof(bufsize));
1127         if (result)
1128                 log_print("Error increasing buffer space on socket %d", result);
1129
1130         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1131                                    (char *)&subscribe, sizeof(subscribe));
1132         if (result < 0) {
1133                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1134                           result);
1135                 goto create_delsock;
1136         }
1137
1138         /* Init con struct */
1139         sock->sk->sk_user_data = con;
1140         con->sock = sock;
1141         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1142         con->rx_action = receive_from_sock;
1143         con->connect_action = sctp_init_assoc;
1144
1145         /* Bind to all interfaces. */
1146         for (i = 0; i < dlm_local_count; i++) {
1147                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1148                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1149
1150                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1151                 if (result)
1152                         goto create_delsock;
1153                 ++num;
1154         }
1155
1156         result = sock->ops->listen(sock, 5);
1157         if (result < 0) {
1158                 log_print("Can't set socket listening");
1159                 goto create_delsock;
1160         }
1161
1162         return 0;
1163
1164 create_delsock:
1165         sock_release(sock);
1166         con->sock = NULL;
1167 out:
1168         return result;
1169 }
1170
1171 static int tcp_listen_for_all(void)
1172 {
1173         struct socket *sock = NULL;
1174         struct connection *con = nodeid2con(0, GFP_KERNEL);
1175         int result = -EINVAL;
1176
1177         if (!con)
1178                 return -ENOMEM;
1179
1180         /* We don't support multi-homed hosts */
1181         if (dlm_local_addr[1] != NULL) {
1182                 log_print("TCP protocol can't handle multi-homed hosts, "
1183                           "try SCTP");
1184                 return -EINVAL;
1185         }
1186
1187         log_print("Using TCP for communications");
1188
1189         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1190         if (sock) {
1191                 add_sock(sock, con);
1192                 result = 0;
1193         }
1194         else {
1195                 result = -EADDRINUSE;
1196         }
1197
1198         return result;
1199 }
1200
1201
1202
1203 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1204                                                      gfp_t allocation)
1205 {
1206         struct writequeue_entry *entry;
1207
1208         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1209         if (!entry)
1210                 return NULL;
1211
1212         entry->page = alloc_page(allocation);
1213         if (!entry->page) {
1214                 kfree(entry);
1215                 return NULL;
1216         }
1217
1218         entry->offset = 0;
1219         entry->len = 0;
1220         entry->end = 0;
1221         entry->users = 0;
1222         entry->con = con;
1223
1224         return entry;
1225 }
1226
1227 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1228 {
1229         struct connection *con;
1230         struct writequeue_entry *e;
1231         int offset = 0;
1232         int users = 0;
1233
1234         con = nodeid2con(nodeid, allocation);
1235         if (!con)
1236                 return NULL;
1237
1238         spin_lock(&con->writequeue_lock);
1239         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1240         if ((&e->list == &con->writequeue) ||
1241             (PAGE_CACHE_SIZE - e->end < len)) {
1242                 e = NULL;
1243         } else {
1244                 offset = e->end;
1245                 e->end += len;
1246                 users = e->users++;
1247         }
1248         spin_unlock(&con->writequeue_lock);
1249
1250         if (e) {
1251         got_one:
1252                 *ppc = page_address(e->page) + offset;
1253                 return e;
1254         }
1255
1256         e = new_writequeue_entry(con, allocation);
1257         if (e) {
1258                 spin_lock(&con->writequeue_lock);
1259                 offset = e->end;
1260                 e->end += len;
1261                 users = e->users++;
1262                 list_add_tail(&e->list, &con->writequeue);
1263                 spin_unlock(&con->writequeue_lock);
1264                 goto got_one;
1265         }
1266         return NULL;
1267 }
1268
1269 void dlm_lowcomms_commit_buffer(void *mh)
1270 {
1271         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1272         struct connection *con = e->con;
1273         int users;
1274
1275         spin_lock(&con->writequeue_lock);
1276         users = --e->users;
1277         if (users)
1278                 goto out;
1279         e->len = e->end - e->offset;
1280         spin_unlock(&con->writequeue_lock);
1281
1282         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1283                 queue_work(send_workqueue, &con->swork);
1284         }
1285         return;
1286
1287 out:
1288         spin_unlock(&con->writequeue_lock);
1289         return;
1290 }
1291
1292 /* Send a message */
1293 static void send_to_sock(struct connection *con)
1294 {
1295         int ret = 0;
1296         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1297         struct writequeue_entry *e;
1298         int len, offset;
1299
1300         mutex_lock(&con->sock_mutex);
1301         if (con->sock == NULL)
1302                 goto out_connect;
1303
1304         spin_lock(&con->writequeue_lock);
1305         for (;;) {
1306                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1307                                list);
1308                 if ((struct list_head *) e == &con->writequeue)
1309                         break;
1310
1311                 len = e->len;
1312                 offset = e->offset;
1313                 BUG_ON(len == 0 && e->users == 0);
1314                 spin_unlock(&con->writequeue_lock);
1315
1316                 ret = 0;
1317                 if (len) {
1318                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1319                                               msg_flags);
1320                         if (ret == -EAGAIN || ret == 0) {
1321                                 cond_resched();
1322                                 goto out;
1323                         }
1324                         if (ret <= 0)
1325                                 goto send_error;
1326                 }
1327                         /* Don't starve people filling buffers */
1328                         cond_resched();
1329
1330                 spin_lock(&con->writequeue_lock);
1331                 e->offset += ret;
1332                 e->len -= ret;
1333
1334                 if (e->len == 0 && e->users == 0) {
1335                         list_del(&e->list);
1336                         free_entry(e);
1337                         continue;
1338                 }
1339         }
1340         spin_unlock(&con->writequeue_lock);
1341 out:
1342         mutex_unlock(&con->sock_mutex);
1343         return;
1344
1345 send_error:
1346         mutex_unlock(&con->sock_mutex);
1347         close_connection(con, false);
1348         lowcomms_connect_sock(con);
1349         return;
1350
1351 out_connect:
1352         mutex_unlock(&con->sock_mutex);
1353         if (!test_bit(CF_INIT_PENDING, &con->flags))
1354                 lowcomms_connect_sock(con);
1355         return;
1356 }
1357
1358 static void clean_one_writequeue(struct connection *con)
1359 {
1360         struct writequeue_entry *e, *safe;
1361
1362         spin_lock(&con->writequeue_lock);
1363         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1364                 list_del(&e->list);
1365                 free_entry(e);
1366         }
1367         spin_unlock(&con->writequeue_lock);
1368 }
1369
1370 /* Called from recovery when it knows that a node has
1371    left the cluster */
1372 int dlm_lowcomms_close(int nodeid)
1373 {
1374         struct connection *con;
1375
1376         log_print("closing connection to node %d", nodeid);
1377         con = nodeid2con(nodeid, 0);
1378         if (con) {
1379                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1380                 clear_bit(CF_WRITE_PENDING, &con->flags);
1381                 set_bit(CF_CLOSE, &con->flags);
1382                 if (cancel_work_sync(&con->swork))
1383                         log_print("canceled swork for node %d", nodeid);
1384                 if (cancel_work_sync(&con->rwork))
1385                         log_print("canceled rwork for node %d", nodeid);
1386                 clean_one_writequeue(con);
1387                 close_connection(con, true);
1388         }
1389         return 0;
1390 }
1391
1392 /* Receive workqueue function */
1393 static void process_recv_sockets(struct work_struct *work)
1394 {
1395         struct connection *con = container_of(work, struct connection, rwork);
1396         int err;
1397
1398         clear_bit(CF_READ_PENDING, &con->flags);
1399         do {
1400                 err = con->rx_action(con);
1401         } while (!err);
1402 }
1403
1404 /* Send workqueue function */
1405 static void process_send_sockets(struct work_struct *work)
1406 {
1407         struct connection *con = container_of(work, struct connection, swork);
1408
1409         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1410                 con->connect_action(con);
1411                 set_bit(CF_WRITE_PENDING, &con->flags);
1412         }
1413         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1414                 send_to_sock(con);
1415 }
1416
1417
1418 /* Discard all entries on the write queues */
1419 static void clean_writequeues(void)
1420 {
1421         foreach_conn(clean_one_writequeue);
1422 }
1423
1424 static void work_stop(void)
1425 {
1426         destroy_workqueue(recv_workqueue);
1427         destroy_workqueue(send_workqueue);
1428 }
1429
1430 static int work_start(void)
1431 {
1432         int error;
1433         recv_workqueue = create_workqueue("dlm_recv");
1434         error = IS_ERR(recv_workqueue);
1435         if (error) {
1436                 log_print("can't start dlm_recv %d", error);
1437                 return error;
1438         }
1439
1440         send_workqueue = create_singlethread_workqueue("dlm_send");
1441         error = IS_ERR(send_workqueue);
1442         if (error) {
1443                 log_print("can't start dlm_send %d", error);
1444                 destroy_workqueue(recv_workqueue);
1445                 return error;
1446         }
1447
1448         return 0;
1449 }
1450
1451 static void stop_conn(struct connection *con)
1452 {
1453         con->flags |= 0x0F;
1454         if (con->sock && con->sock->sk)
1455                 con->sock->sk->sk_user_data = NULL;
1456 }
1457
1458 static void free_conn(struct connection *con)
1459 {
1460         close_connection(con, true);
1461         if (con->othercon)
1462                 kmem_cache_free(con_cache, con->othercon);
1463         hlist_del(&con->list);
1464         kmem_cache_free(con_cache, con);
1465 }
1466
1467 void dlm_lowcomms_stop(void)
1468 {
1469         /* Set all the flags to prevent any
1470            socket activity.
1471         */
1472         mutex_lock(&connections_lock);
1473         foreach_conn(stop_conn);
1474         mutex_unlock(&connections_lock);
1475
1476         work_stop();
1477
1478         mutex_lock(&connections_lock);
1479         clean_writequeues();
1480
1481         foreach_conn(free_conn);
1482
1483         mutex_unlock(&connections_lock);
1484         kmem_cache_destroy(con_cache);
1485 }
1486
1487 int dlm_lowcomms_start(void)
1488 {
1489         int error = -EINVAL;
1490         struct connection *con;
1491         int i;
1492
1493         for (i = 0; i < CONN_HASH_SIZE; i++)
1494                 INIT_HLIST_HEAD(&connection_hash[i]);
1495
1496         init_local();
1497         if (!dlm_local_count) {
1498                 error = -ENOTCONN;
1499                 log_print("no local IP address has been set");
1500                 goto out;
1501         }
1502
1503         error = -ENOMEM;
1504         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1505                                       __alignof__(struct connection), 0,
1506                                       NULL);
1507         if (!con_cache)
1508                 goto out;
1509
1510         /* Start listening */
1511         if (dlm_config.ci_protocol == 0)
1512                 error = tcp_listen_for_all();
1513         else
1514                 error = sctp_listen_for_all();
1515         if (error)
1516                 goto fail_unlisten;
1517
1518         error = work_start();
1519         if (error)
1520                 goto fail_unlisten;
1521
1522         return 0;
1523
1524 fail_unlisten:
1525         con = nodeid2con(0,0);
1526         if (con) {
1527                 close_connection(con, false);
1528                 kmem_cache_free(con_cache, con);
1529         }
1530         kmem_cache_destroy(con_cache);
1531
1532 out:
1533         return error;
1534 }