fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is its
  25  * responsibility. It is this layer's
  26  * responsibility to resolve these into IP address or
  27  * whatever it needs for inter-node communication.
  28  *
  29  * The comms level is two kernel threads that deal mainly with
  30  * the receiving of messages from other nodes and passing them
  31  * up to the mid-level comms layer (which understands the
  32  * message format) for execution by the locking core, and
  33  * a send thread which does all the setting up of connections
  34  * to remote nodes and the sending of data. Threads are not allowed
  35  * to send their own data because it may cause them to wait in times
  36  * of high load. Also, this way, the sending thread can collect together
  37  * messages bound for one node and send them in one block.
  38  *
  39  * lowcomms will choose to use either TCP or SCTP as its transport layer
  40  * depending on the configuration variable 'protocol'. This should be set
  41  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  42  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  43  * for the DLM to function.
  44  *
  45  */
  46
  47 #include <asm/ioctls.h>
  48 #include <net/sock.h>
  49 #include <net/tcp.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/file.h>
  52 #include <linux/mutex.h>
  53 #include <linux/sctp.h>
  54 #include <linux/slab.h>
  55 #include <net/sctp/user.h>
  56 #include <net/ipv6.h>
  57
  58 #include "dlm_internal.h"
  59 #include "lowcomms.h"
  60 #include "midcomms.h"
  61 #include "config.h"
  62
  63 #define NEEDED_RMEM (4*1024*1024)
  64 #define CONN_HASH_SIZE 32
  65
  66 struct cbuf {
  67         unsigned int base;
  68         unsigned int len;
  69         unsigned int mask;
  70 };
  71
  72 static void cbuf_add(struct cbuf *cb, int n)
  73 {
  74         cb->len += n;
  75 }
  76
  77 static int cbuf_data(struct cbuf *cb)
  78 {
  79         return ((cb->base + cb->len) & cb->mask);
  80 }
  81
  82 static void cbuf_init(struct cbuf *cb, int size)
  83 {
  84         cb->base = cb->len = 0;
  85         cb->mask = size-1;
  86 }
  87
  88 static void cbuf_eat(struct cbuf *cb, int n)
  89 {
  90         cb->len  -= n;
  91         cb->base += n;
  92         cb->base &= cb->mask;
  93 }
  94
  95 static bool cbuf_empty(struct cbuf *cb)
  96 {
  97         return cb->len == 0;
  98 }
  99
 100 struct connection {
 101         struct socket *sock;    /* NULL if not connected */
 102         uint32_t nodeid;        /* So we know who we are in the list */
 103         struct mutex sock_mutex;
 104         unsigned long flags;
 105 #define CF_READ_PENDING 1
 106 #define CF_WRITE_PENDING 2
 107 #define CF_CONNECT_PENDING 3
 108 #define CF_INIT_PENDING 4
 109 #define CF_IS_OTHERCON 5
 110 #define CF_CLOSE 6
 111 #define CF_APP_LIMITED 7
 112         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 113         spinlock_t writequeue_lock;
 114         int (*rx_action) (struct connection *); /* What to do when active */
 115         void (*connect_action) (struct connection *);   /* What to do to connect */
 116         struct page *rx_page;
 117         struct cbuf cb;
 118         int retries;
 119 #define MAX_CONNECT_RETRIES 3
 120         int sctp_assoc;
 121         struct hlist_node list;
 122         struct connection *othercon;
 123         struct work_struct rwork; /* Receive workqueue */
 124         struct work_struct swork; /* Send workqueue */
 125 };
 126 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 127
 128 /* An entry waiting to be sent */
 129 struct writequeue_entry {
 130         struct list_head list;
 131         struct page *page;
 132         int offset;
 133         int len;
 134         int end;
 135         int users;
 136         struct connection *con;
 137 };
 138
 139 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 140 static int dlm_local_count;
 141
 142 /* Work queues */
 143 static struct workqueue_struct *recv_workqueue;
 144 static struct workqueue_struct *send_workqueue;
 145
 146 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 147 static DEFINE_MUTEX(connections_lock);
 148 static struct kmem_cache *con_cache;
 149
 150 static void process_recv_sockets(struct work_struct *work);
 151 static void process_send_sockets(struct work_struct *work);
 152
 153
 154 /* This is deliberately very simple because most clusters have simple
 155    sequential nodeids, so we should be able to go straight to a connection
 156    struct in the array */
 157 static inline int nodeid_hash(int nodeid)
 158 {
 159         return nodeid & (CONN_HASH_SIZE-1);
 160 }
 161
 162 static struct connection *__find_con(int nodeid)
 163 {
 164         int r;
 165         struct hlist_node *h;
 166         struct connection *con;
 167
 168         r = nodeid_hash(nodeid);
 169
 170         hlist_for_each_entry(con, h, &connection_hash[r], list) {
 171                 if (con->nodeid == nodeid)
 172                         return con;
 173         }
 174         return NULL;
 175 }
 176
 177 /*
 178  * If 'allocation' is zero then we don't attempt to create a new
 179  * connection structure for this node.
 180  */
 181 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 182 {
 183         struct connection *con = NULL;
 184         int r;
 185
 186         con = __find_con(nodeid);
 187         if (con || !alloc)
 188                 return con;
 189
 190         con = kmem_cache_zalloc(con_cache, alloc);
 191         if (!con)
 192                 return NULL;
 193
 194         r = nodeid_hash(nodeid);
 195         hlist_add_head(&con->list, &connection_hash[r]);
 196
 197         con->nodeid = nodeid;
 198         mutex_init(&con->sock_mutex);
 199         INIT_LIST_HEAD(&con->writequeue);
 200         spin_lock_init(&con->writequeue_lock);
 201         INIT_WORK(&con->swork, process_send_sockets);
 202         INIT_WORK(&con->rwork, process_recv_sockets);
 203
 204         /* Setup action pointers for child sockets */
 205         if (con->nodeid) {
 206                 struct connection *zerocon = __find_con(0);
 207
 208                 con->connect_action = zerocon->connect_action;
 209                 if (!con->rx_action)
 210                         con->rx_action = zerocon->rx_action;
 211         }
 212
 213         return con;
 214 }
 215
 216 /* Loop round all connections */
 217 static void foreach_conn(void (*conn_func)(struct connection *c))
 218 {
 219         int i;
 220         struct hlist_node *h, *n;
 221         struct connection *con;
 222
 223         for (i = 0; i < CONN_HASH_SIZE; i++) {
 224                 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
 225                         conn_func(con);
 226                 }
 227         }
 228 }
 229
 230 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 231 {
 232         struct connection *con;
 233
 234         mutex_lock(&connections_lock);
 235         con = __nodeid2con(nodeid, allocation);
 236         mutex_unlock(&connections_lock);
 237
 238         return con;
 239 }
 240
 241 /* This is a bit drastic, but only called when things go wrong */
 242 static struct connection *assoc2con(int assoc_id)
 243 {
 244         int i;
 245         struct hlist_node *h;
 246         struct connection *con;
 247
 248         mutex_lock(&connections_lock);
 249
 250         for (i = 0 ; i < CONN_HASH_SIZE; i++) {
 251                 hlist_for_each_entry(con, h, &connection_hash[i], list) {
 252                         if (con->sctp_assoc == assoc_id) {
 253                                 mutex_unlock(&connections_lock);
 254                                 return con;
 255                         }
 256                 }
 257         }
 258         mutex_unlock(&connections_lock);
 259         return NULL;
 260 }
 261
 262 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 263 {
 264         struct sockaddr_storage addr;
 265         int error;
 266
 267         if (!dlm_local_count)
 268                 return -1;
 269
 270         error = dlm_nodeid_to_addr(nodeid, &addr);
 271         if (error)
 272                 return error;
 273
 274         if (dlm_local_addr[0]->ss_family == AF_INET) {
 275                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 276                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 277                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 278         } else {
 279                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 280                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 281                 ipv6_addr_copy(&ret6->sin6_addr, &in6->sin6_addr);
 282         }
 283
 284         return 0;
 285 }
 286
 287 /* Data available on socket or listen socket received a connect */
 288 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 289 {
 290         struct connection *con = sock2con(sk);
 291         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 292                 queue_work(recv_workqueue, &con->rwork);
 293 }
 294
 295 static void lowcomms_write_space(struct sock *sk)
 296 {
 297         struct connection *con = sock2con(sk);
 298
 299         if (!con)
 300                 return;
 301
 302         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 303
 304         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 305                 con->sock->sk->sk_write_pending--;
 306                 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
 307         }
 308
 309         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 310                 queue_work(send_workqueue, &con->swork);
 311 }
 312
 313 static inline void lowcomms_connect_sock(struct connection *con)
 314 {
 315         if (test_bit(CF_CLOSE, &con->flags))
 316                 return;
 317         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 318                 queue_work(send_workqueue, &con->swork);
 319 }
 320
 321 static void lowcomms_state_change(struct sock *sk)
 322 {
 323         if (sk->sk_state == TCP_ESTABLISHED)
 324                 lowcomms_write_space(sk);
 325 }
 326
 327 int dlm_lowcomms_connect_node(int nodeid)
 328 {
 329         struct connection *con;
 330
 331         /* with sctp there's no connecting without sending */
 332         if (dlm_config.ci_protocol != 0)
 333                 return 0;
 334
 335         if (nodeid == dlm_our_nodeid())
 336                 return 0;
 337
 338         con = nodeid2con(nodeid, GFP_NOFS);
 339         if (!con)
 340                 return -ENOMEM;
 341         lowcomms_connect_sock(con);
 342         return 0;
 343 }
 344
 345 /* Make a socket active */
 346 static int add_sock(struct socket *sock, struct connection *con)
 347 {
 348         con->sock = sock;
 349
 350         /* Install a data_ready callback */
 351         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 352         con->sock->sk->sk_write_space = lowcomms_write_space;
 353         con->sock->sk->sk_state_change = lowcomms_state_change;
 354         con->sock->sk->sk_user_data = con;
 355         con->sock->sk->sk_allocation = GFP_NOFS;
 356         return 0;
 357 }
 358
 359 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 360    length */
 361 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 362                           int *addr_len)
 363 {
 364         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 365         if (saddr->ss_family == AF_INET) {
 366                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 367                 in4_addr->sin_port = cpu_to_be16(port);
 368                 *addr_len = sizeof(struct sockaddr_in);
 369                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 370         } else {
 371                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 372                 in6_addr->sin6_port = cpu_to_be16(port);
 373                 *addr_len = sizeof(struct sockaddr_in6);
 374         }
 375         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 376 }
 377
 378 /* Close a remote connection and tidy up */
 379 static void close_connection(struct connection *con, bool and_other)
 380 {
 381         mutex_lock(&con->sock_mutex);
 382
 383         if (con->sock) {
 384                 sock_release(con->sock);
 385                 con->sock = NULL;
 386         }
 387         if (con->othercon && and_other) {
 388                 /* Will only re-enter once. */
 389                 close_connection(con->othercon, false);
 390         }
 391         if (con->rx_page) {
 392                 __free_page(con->rx_page);
 393                 con->rx_page = NULL;
 394         }
 395
 396         con->retries = 0;
 397         mutex_unlock(&con->sock_mutex);
 398 }
 399
 400 /* We only send shutdown messages to nodes that are not part of the cluster */
 401 static void sctp_send_shutdown(sctp_assoc_t associd)
 402 {
 403         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 404         struct msghdr outmessage;
 405         struct cmsghdr *cmsg;
 406         struct sctp_sndrcvinfo *sinfo;
 407         int ret;
 408         struct connection *con;
 409
 410         con = nodeid2con(0,0);
 411         BUG_ON(con == NULL);
 412
 413         outmessage.msg_name = NULL;
 414         outmessage.msg_namelen = 0;
 415         outmessage.msg_control = outcmsg;
 416         outmessage.msg_controllen = sizeof(outcmsg);
 417         outmessage.msg_flags = MSG_EOR;
 418
 419         cmsg = CMSG_FIRSTHDR(&outmessage);
 420         cmsg->cmsg_level = IPPROTO_SCTP;
 421         cmsg->cmsg_type = SCTP_SNDRCV;
 422         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 423         outmessage.msg_controllen = cmsg->cmsg_len;
 424         sinfo = CMSG_DATA(cmsg);
 425         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 426
 427         sinfo->sinfo_flags |= MSG_EOF;
 428         sinfo->sinfo_assoc_id = associd;
 429
 430         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 431
 432         if (ret != 0)
 433                 log_print("send EOF to node failed: %d", ret);
 434 }
 435
 436 static void sctp_init_failed_foreach(struct connection *con)
 437 {
 438         con->sctp_assoc = 0;
 439         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 440                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 441                         queue_work(send_workqueue, &con->swork);
 442         }
 443 }
 444
 445 /* INIT failed but we don't know which node...
 446    restart INIT on all pending nodes */
 447 static void sctp_init_failed(void)
 448 {
 449         mutex_lock(&connections_lock);
 450
 451         foreach_conn(sctp_init_failed_foreach);
 452
 453         mutex_unlock(&connections_lock);
 454 }
 455
 456 /* Something happened to an association */
 457 static void process_sctp_notification(struct connection *con,
 458                                       struct msghdr *msg, char *buf)
 459 {
 460         union sctp_notification *sn = (union sctp_notification *)buf;
 461
 462         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 463                 switch (sn->sn_assoc_change.sac_state) {
 464
 465                 case SCTP_COMM_UP:
 466                 case SCTP_RESTART:
 467                 {
 468                         /* Check that the new node is in the lockspace */
 469                         struct sctp_prim prim;
 470                         int nodeid;
 471                         int prim_len, ret;
 472                         int addr_len;
 473                         struct connection *new_con;
 474                         sctp_peeloff_arg_t parg;
 475                         int parglen = sizeof(parg);
 476                         int err;
 477
 478                         /*
 479                          * We get this before any data for an association.
 480                          * We verify that the node is in the cluster and
 481                          * then peel off a socket for it.
 482                          */
 483                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 484                                 log_print("COMM_UP for invalid assoc ID %d",
 485                                          (int)sn->sn_assoc_change.sac_assoc_id);
 486                                 sctp_init_failed();
 487                                 return;
 488                         }
 489                         memset(&prim, 0, sizeof(struct sctp_prim));
 490                         prim_len = sizeof(struct sctp_prim);
 491                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 492
 493                         ret = kernel_getsockopt(con->sock,
 494                                                 IPPROTO_SCTP,
 495                                                 SCTP_PRIMARY_ADDR,
 496                                                 (char*)&prim,
 497                                                 &prim_len);
 498                         if (ret < 0) {
 499                                 log_print("getsockopt/sctp_primary_addr on "
 500                                           "new assoc %d failed : %d",
 501                                           (int)sn->sn_assoc_change.sac_assoc_id,
 502                                           ret);
 503
 504                                 /* Retry INIT later */
 505                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 506                                 if (new_con)
 507                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 508                                 return;
 509                         }
 510                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 511                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 512                                 int i;
 513                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 514                                 log_print("reject connect from unknown addr");
 515                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 516                                         printk("%02x ", b[i]);
 517                                 printk("\n");
 518                                 sctp_send_shutdown(prim.ssp_assoc_id);
 519                                 return;
 520                         }
 521
 522                         new_con = nodeid2con(nodeid, GFP_NOFS);
 523                         if (!new_con)
 524                                 return;
 525
 526                         /* Peel off a new sock */
 527                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 528                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 529                                                 SCTP_SOCKOPT_PEELOFF,
 530                                                 (void *)&parg, &parglen);
 531                         if (ret < 0) {
 532                                 log_print("Can't peel off a socket for "
 533                                           "connection %d to node %d: err=%d",
 534                                           parg.associd, nodeid, ret);
 535                                 return;
 536                         }
 537                         new_con->sock = sockfd_lookup(parg.sd, &err);
 538                         if (!new_con->sock) {
 539                                 log_print("sockfd_lookup error %d", err);
 540                                 return;
 541                         }
 542                         add_sock(new_con->sock, new_con);
 543                         sockfd_put(new_con->sock);
 544
 545                         log_print("connecting to %d sctp association %d",
 546                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 547
 548                         /* Send any pending writes */
 549                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 550                         clear_bit(CF_INIT_PENDING, &con->flags);
 551                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 552                                 queue_work(send_workqueue, &new_con->swork);
 553                         }
 554                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 555                                 queue_work(recv_workqueue, &new_con->rwork);
 556                 }
 557                 break;
 558
 559                 case SCTP_COMM_LOST:
 560                 case SCTP_SHUTDOWN_COMP:
 561                 {
 562                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 563                         if (con) {
 564                                 con->sctp_assoc = 0;
 565                         }
 566                 }
 567                 break;
 568
 569                 /* We don't know which INIT failed, so clear the PENDING flags
 570                  * on them all.  if assoc_id is zero then it will then try
 571                  * again */
 572
 573                 case SCTP_CANT_STR_ASSOC:
 574                 {
 575                         log_print("Can't start SCTP association - retrying");
 576                         sctp_init_failed();
 577                 }
 578                 break;
 579
 580                 default:
 581                         log_print("unexpected SCTP assoc change id=%d state=%d",
 582                                   (int)sn->sn_assoc_change.sac_assoc_id,
 583                                   sn->sn_assoc_change.sac_state);
 584                 }
 585         }
 586 }
 587
 588 /* Data received from remote end */
 589 static int receive_from_sock(struct connection *con)
 590 {
 591         int ret = 0;
 592         struct msghdr msg = {};
 593         struct kvec iov[2];
 594         unsigned len;
 595         int r;
 596         int call_again_soon = 0;
 597         int nvec;
 598         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 599
 600         mutex_lock(&con->sock_mutex);
 601
 602         if (con->sock == NULL) {
 603                 ret = -EAGAIN;
 604                 goto out_close;
 605         }
 606
 607         if (con->rx_page == NULL) {
 608                 /*
 609                  * This doesn't need to be atomic, but I think it should
 610                  * improve performance if it is.
 611                  */
 612                 con->rx_page = alloc_page(GFP_ATOMIC);
 613                 if (con->rx_page == NULL)
 614                         goto out_resched;
 615                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 616         }
 617
 618         /* Only SCTP needs these really */
 619         memset(&incmsg, 0, sizeof(incmsg));
 620         msg.msg_control = incmsg;
 621         msg.msg_controllen = sizeof(incmsg);
 622
 623         /*
 624          * iov[0] is the bit of the circular buffer between the current end
 625          * point (cb.base + cb.len) and the end of the buffer.
 626          */
 627         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 628         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 629         iov[1].iov_len = 0;
 630         nvec = 1;
 631
 632         /*
 633          * iov[1] is the bit of the circular buffer between the start of the
 634          * buffer and the start of the currently used section (cb.base)
 635          */
 636         if (cbuf_data(&con->cb) >= con->cb.base) {
 637                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 638                 iov[1].iov_len = con->cb.base;
 639                 iov[1].iov_base = page_address(con->rx_page);
 640                 nvec = 2;
 641         }
 642         len = iov[0].iov_len + iov[1].iov_len;
 643
 644         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 645                                MSG_DONTWAIT | MSG_NOSIGNAL);
 646         if (ret <= 0)
 647                 goto out_close;
 648
 649         /* Process SCTP notifications */
 650         if (msg.msg_flags & MSG_NOTIFICATION) {
 651                 msg.msg_control = incmsg;
 652                 msg.msg_controllen = sizeof(incmsg);
 653
 654                 process_sctp_notification(con, &msg,
 655                                 page_address(con->rx_page) + con->cb.base);
 656                 mutex_unlock(&con->sock_mutex);
 657                 return 0;
 658         }
 659         BUG_ON(con->nodeid == 0);
 660
 661         if (ret == len)
 662                 call_again_soon = 1;
 663         cbuf_add(&con->cb, ret);
 664         ret = dlm_process_incoming_buffer(con->nodeid,
 665                                           page_address(con->rx_page),
 666                                           con->cb.base, con->cb.len,
 667                                           PAGE_CACHE_SIZE);
 668         if (ret == -EBADMSG) {
 669                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 670                           "iov_len=%u, iov_base[0]=%p, read=%d",
 671                           page_address(con->rx_page), con->cb.base, con->cb.len,
 672                           len, iov[0].iov_base, r);
 673         }
 674         if (ret < 0)
 675                 goto out_close;
 676         cbuf_eat(&con->cb, ret);
 677
 678         if (cbuf_empty(&con->cb) && !call_again_soon) {
 679                 __free_page(con->rx_page);
 680                 con->rx_page = NULL;
 681         }
 682
 683         if (call_again_soon)
 684                 goto out_resched;
 685         mutex_unlock(&con->sock_mutex);
 686         return 0;
 687
 688 out_resched:
 689         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 690                 queue_work(recv_workqueue, &con->rwork);
 691         mutex_unlock(&con->sock_mutex);
 692         return -EAGAIN;
 693
 694 out_close:
 695         mutex_unlock(&con->sock_mutex);
 696         if (ret != -EAGAIN) {
 697                 close_connection(con, false);
 698                 /* Reconnect when there is something to send */
 699         }
 700         /* Don't return success if we really got EOF */
 701         if (ret == 0)
 702                 ret = -EAGAIN;
 703
 704         return ret;
 705 }
 706
 707 /* Listening socket is busy, accept a connection */
 708 static int tcp_accept_from_sock(struct connection *con)
 709 {
 710         int result;
 711         struct sockaddr_storage peeraddr;
 712         struct socket *newsock;
 713         int len;
 714         int nodeid;
 715         struct connection *newcon;
 716         struct connection *addcon;
 717
 718         memset(&peeraddr, 0, sizeof(peeraddr));
 719         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 720                                   IPPROTO_TCP, &newsock);
 721         if (result < 0)
 722                 return -ENOMEM;
 723
 724         mutex_lock_nested(&con->sock_mutex, 0);
 725
 726         result = -ENOTCONN;
 727         if (con->sock == NULL)
 728                 goto accept_err;
 729
 730         newsock->type = con->sock->type;
 731         newsock->ops = con->sock->ops;
 732
 733         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 734         if (result < 0)
 735                 goto accept_err;
 736
 737         /* Get the connected socket's peer */
 738         memset(&peeraddr, 0, sizeof(peeraddr));
 739         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 740                                   &len, 2)) {
 741                 result = -ECONNABORTED;
 742                 goto accept_err;
 743         }
 744
 745         /* Get the new node's NODEID */
 746         make_sockaddr(&peeraddr, 0, &len);
 747         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 748                 log_print("connect from non cluster node");
 749                 sock_release(newsock);
 750                 mutex_unlock(&con->sock_mutex);
 751                 return -1;
 752         }
 753
 754         log_print("got connection from %d", nodeid);
 755
 756         /*  Check to see if we already have a connection to this node. This
 757          *  could happen if the two nodes initiate a connection at roughly
 758          *  the same time and the connections cross on the wire.
 759          *  In this case we store the incoming one in "othercon"
 760          */
 761         newcon = nodeid2con(nodeid, GFP_NOFS);
 762         if (!newcon) {
 763                 result = -ENOMEM;
 764                 goto accept_err;
 765         }
 766         mutex_lock_nested(&newcon->sock_mutex, 1);
 767         if (newcon->sock) {
 768                 struct connection *othercon = newcon->othercon;
 769
 770                 if (!othercon) {
 771                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 772                         if (!othercon) {
 773                                 log_print("failed to allocate incoming socket");
 774                                 mutex_unlock(&newcon->sock_mutex);
 775                                 result = -ENOMEM;
 776                                 goto accept_err;
 777                         }
 778                         othercon->nodeid = nodeid;
 779                         othercon->rx_action = receive_from_sock;
 780                         mutex_init(&othercon->sock_mutex);
 781                         INIT_WORK(&othercon->swork, process_send_sockets);
 782                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 783                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 784                 }
 785                 if (!othercon->sock) {
 786                         newcon->othercon = othercon;
 787                         othercon->sock = newsock;
 788                         newsock->sk->sk_user_data = othercon;
 789                         add_sock(newsock, othercon);
 790                         addcon = othercon;
 791                 }
 792                 else {
 793                         printk("Extra connection from node %d attempted\n", nodeid);
 794                         result = -EAGAIN;
 795                         mutex_unlock(&newcon->sock_mutex);
 796                         goto accept_err;
 797                 }
 798         }
 799         else {
 800                 newsock->sk->sk_user_data = newcon;
 801                 newcon->rx_action = receive_from_sock;
 802                 add_sock(newsock, newcon);
 803                 addcon = newcon;
 804         }
 805
 806         mutex_unlock(&newcon->sock_mutex);
 807
 808         /*
 809          * Add it to the active queue in case we got data
 810          * beween processing the accept adding the socket
 811          * to the read_sockets list
 812          */
 813         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 814                 queue_work(recv_workqueue, &addcon->rwork);
 815         mutex_unlock(&con->sock_mutex);
 816
 817         return 0;
 818
 819 accept_err:
 820         mutex_unlock(&con->sock_mutex);
 821         sock_release(newsock);
 822
 823         if (result != -EAGAIN)
 824                 log_print("error accepting connection from node: %d", result);
 825         return result;
 826 }
 827
 828 static void free_entry(struct writequeue_entry *e)
 829 {
 830         __free_page(e->page);
 831         kfree(e);
 832 }
 833
 834 /* Initiate an SCTP association.
 835    This is a special case of send_to_sock() in that we don't yet have a
 836    peeled-off socket for this association, so we use the listening socket
 837    and add the primary IP address of the remote node.
 838  */
 839 static void sctp_init_assoc(struct connection *con)
 840 {
 841         struct sockaddr_storage rem_addr;
 842         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 843         struct msghdr outmessage;
 844         struct cmsghdr *cmsg;
 845         struct sctp_sndrcvinfo *sinfo;
 846         struct connection *base_con;
 847         struct writequeue_entry *e;
 848         int len, offset;
 849         int ret;
 850         int addrlen;
 851         struct kvec iov[1];
 852
 853         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 854                 return;
 855
 856         if (con->retries++ > MAX_CONNECT_RETRIES)
 857                 return;
 858
 859         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 860                 log_print("no address for nodeid %d", con->nodeid);
 861                 return;
 862         }
 863         base_con = nodeid2con(0, 0);
 864         BUG_ON(base_con == NULL);
 865
 866         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 867
 868         outmessage.msg_name = &rem_addr;
 869         outmessage.msg_namelen = addrlen;
 870         outmessage.msg_control = outcmsg;
 871         outmessage.msg_controllen = sizeof(outcmsg);
 872         outmessage.msg_flags = MSG_EOR;
 873
 874         spin_lock(&con->writequeue_lock);
 875
 876         if (list_empty(&con->writequeue)) {
 877                 spin_unlock(&con->writequeue_lock);
 878                 log_print("writequeue empty for nodeid %d", con->nodeid);
 879                 return;
 880         }
 881
 882         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
 883         len = e->len;
 884         offset = e->offset;
 885         spin_unlock(&con->writequeue_lock);
 886
 887         /* Send the first block off the write queue */
 888         iov[0].iov_base = page_address(e->page)+offset;
 889         iov[0].iov_len = len;
 890
 891         cmsg = CMSG_FIRSTHDR(&outmessage);
 892         cmsg->cmsg_level = IPPROTO_SCTP;
 893         cmsg->cmsg_type = SCTP_SNDRCV;
 894         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 895         sinfo = CMSG_DATA(cmsg);
 896         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 897         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 898         outmessage.msg_controllen = cmsg->cmsg_len;
 899
 900         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 901         if (ret < 0) {
 902                 log_print("Send first packet to node %d failed: %d",
 903                           con->nodeid, ret);
 904
 905                 /* Try again later */
 906                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 907                 clear_bit(CF_INIT_PENDING, &con->flags);
 908         }
 909         else {
 910                 spin_lock(&con->writequeue_lock);
 911                 e->offset += ret;
 912                 e->len -= ret;
 913
 914                 if (e->len == 0 && e->users == 0) {
 915                         list_del(&e->list);
 916                         free_entry(e);
 917                 }
 918                 spin_unlock(&con->writequeue_lock);
 919         }
 920 }
 921
 922 /* Connect a new socket to its peer */
 923 static void tcp_connect_to_sock(struct connection *con)
 924 {
 925         int result = -EHOSTUNREACH;
 926         struct sockaddr_storage saddr, src_addr;
 927         int addr_len;
 928         struct socket *sock = NULL;
 929         int one = 1;
 930
 931         if (con->nodeid == 0) {
 932                 log_print("attempt to connect sock 0 foiled");
 933                 return;
 934         }
 935
 936         mutex_lock(&con->sock_mutex);
 937         if (con->retries++ > MAX_CONNECT_RETRIES)
 938                 goto out;
 939
 940         /* Some odd races can cause double-connects, ignore them */
 941         if (con->sock) {
 942                 result = 0;
 943                 goto out;
 944         }
 945
 946         /* Create a socket to communicate with */
 947         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 948                                   IPPROTO_TCP, &sock);
 949         if (result < 0)
 950                 goto out_err;
 951
 952         memset(&saddr, 0, sizeof(saddr));
 953         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 954                 goto out_err;
 955
 956         sock->sk->sk_user_data = con;
 957         con->rx_action = receive_from_sock;
 958         con->connect_action = tcp_connect_to_sock;
 959         add_sock(sock, con);
 960
 961         /* Bind to our cluster-known address connecting to avoid
 962            routing problems */
 963         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 964         make_sockaddr(&src_addr, 0, &addr_len);
 965         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 966                                  addr_len);
 967         if (result < 0) {
 968                 log_print("could not bind for connect: %d", result);
 969                 /* This *may* not indicate a critical error */
 970         }
 971
 972         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 973
 974         log_print("connecting to %d", con->nodeid);
 975
 976         /* Turn off Nagle's algorithm */
 977         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
 978                           sizeof(one));
 979
 980         result =
 981                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 982                                    O_NONBLOCK);
 983         if (result == -EINPROGRESS)
 984                 result = 0;
 985         if (result == 0)
 986                 goto out;
 987
 988 out_err:
 989         if (con->sock) {
 990                 sock_release(con->sock);
 991                 con->sock = NULL;
 992         } else if (sock) {
 993                 sock_release(sock);
 994         }
 995         /*
 996          * Some errors are fatal and this list might need adjusting. For other
 997          * errors we try again until the max number of retries is reached.
 998          */
 999         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
1000             result != -ENETDOWN && result != -EINVAL
1001             && result != -EPROTONOSUPPORT) {
1002                 lowcomms_connect_sock(con);
1003                 result = 0;
1004         }
1005 out:
1006         mutex_unlock(&con->sock_mutex);
1007         return;
1008 }
1009
1010 static struct socket *tcp_create_listen_sock(struct connection *con,
1011                                              struct sockaddr_storage *saddr)
1012 {
1013         struct socket *sock = NULL;
1014         int result = 0;
1015         int one = 1;
1016         int addr_len;
1017
1018         if (dlm_local_addr[0]->ss_family == AF_INET)
1019                 addr_len = sizeof(struct sockaddr_in);
1020         else
1021                 addr_len = sizeof(struct sockaddr_in6);
1022
1023         /* Create a socket to communicate with */
1024         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1025                                   IPPROTO_TCP, &sock);
1026         if (result < 0) {
1027                 log_print("Can't create listening comms socket");
1028                 goto create_out;
1029         }
1030
1031         /* Turn off Nagle's algorithm */
1032         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1033                           sizeof(one));
1034
1035         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1036                                    (char *)&one, sizeof(one));
1037
1038         if (result < 0) {
1039                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1040         }
1041         sock->sk->sk_user_data = con;
1042         con->rx_action = tcp_accept_from_sock;
1043         con->connect_action = tcp_connect_to_sock;
1044         con->sock = sock;
1045
1046         /* Bind to our port */
1047         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1048         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1049         if (result < 0) {
1050                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1051                 sock_release(sock);
1052                 sock = NULL;
1053                 con->sock = NULL;
1054                 goto create_out;
1055         }
1056         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1057                                  (char *)&one, sizeof(one));
1058         if (result < 0) {
1059                 log_print("Set keepalive failed: %d", result);
1060         }
1061
1062         result = sock->ops->listen(sock, 5);
1063         if (result < 0) {
1064                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1065                 sock_release(sock);
1066                 sock = NULL;
1067                 goto create_out;
1068         }
1069
1070 create_out:
1071         return sock;
1072 }
1073
1074 /* Get local addresses */
1075 static void init_local(void)
1076 {
1077         struct sockaddr_storage sas, *addr;
1078         int i;
1079
1080         dlm_local_count = 0;
1081         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1082                 if (dlm_our_addr(&sas, i))
1083                         break;
1084
1085                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1086                 if (!addr)
1087                         break;
1088                 memcpy(addr, &sas, sizeof(*addr));
1089                 dlm_local_addr[dlm_local_count++] = addr;
1090         }
1091 }
1092
1093 /* Bind to an IP address. SCTP allows multiple address so it can do
1094    multi-homing */
1095 static int add_sctp_bind_addr(struct connection *sctp_con,
1096                               struct sockaddr_storage *addr,
1097                               int addr_len, int num)
1098 {
1099         int result = 0;
1100
1101         if (num == 1)
1102                 result = kernel_bind(sctp_con->sock,
1103                                      (struct sockaddr *) addr,
1104                                      addr_len);
1105         else
1106                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1107                                            SCTP_SOCKOPT_BINDX_ADD,
1108                                            (char *)addr, addr_len);
1109
1110         if (result < 0)
1111                 log_print("Can't bind to port %d addr number %d",
1112                           dlm_config.ci_tcp_port, num);
1113
1114         return result;
1115 }
1116
1117 /* Initialise SCTP socket and bind to all interfaces */
1118 static int sctp_listen_for_all(void)
1119 {
1120         struct socket *sock = NULL;
1121         struct sockaddr_storage localaddr;
1122         struct sctp_event_subscribe subscribe;
1123         int result = -EINVAL, num = 1, i, addr_len;
1124         struct connection *con = nodeid2con(0, GFP_NOFS);
1125         int bufsize = NEEDED_RMEM;
1126
1127         if (!con)
1128                 return -ENOMEM;
1129
1130         log_print("Using SCTP for communications");
1131
1132         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1133                                   IPPROTO_SCTP, &sock);
1134         if (result < 0) {
1135                 log_print("Can't create comms socket, check SCTP is loaded");
1136                 goto out;
1137         }
1138
1139         /* Listen for events */
1140         memset(&subscribe, 0, sizeof(subscribe));
1141         subscribe.sctp_data_io_event = 1;
1142         subscribe.sctp_association_event = 1;
1143         subscribe.sctp_send_failure_event = 1;
1144         subscribe.sctp_shutdown_event = 1;
1145         subscribe.sctp_partial_delivery_event = 1;
1146
1147         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1148                                  (char *)&bufsize, sizeof(bufsize));
1149         if (result)
1150                 log_print("Error increasing buffer space on socket %d", result);
1151
1152         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1153                                    (char *)&subscribe, sizeof(subscribe));
1154         if (result < 0) {
1155                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1156                           result);
1157                 goto create_delsock;
1158         }
1159
1160         /* Init con struct */
1161         sock->sk->sk_user_data = con;
1162         con->sock = sock;
1163         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1164         con->rx_action = receive_from_sock;
1165         con->connect_action = sctp_init_assoc;
1166
1167         /* Bind to all interfaces. */
1168         for (i = 0; i < dlm_local_count; i++) {
1169                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1170                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1171
1172                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1173                 if (result)
1174                         goto create_delsock;
1175                 ++num;
1176         }
1177
1178         result = sock->ops->listen(sock, 5);
1179         if (result < 0) {
1180                 log_print("Can't set socket listening");
1181                 goto create_delsock;
1182         }
1183
1184         return 0;
1185
1186 create_delsock:
1187         sock_release(sock);
1188         con->sock = NULL;
1189 out:
1190         return result;
1191 }
1192
1193 static int tcp_listen_for_all(void)
1194 {
1195         struct socket *sock = NULL;
1196         struct connection *con = nodeid2con(0, GFP_NOFS);
1197         int result = -EINVAL;
1198
1199         if (!con)
1200                 return -ENOMEM;
1201
1202         /* We don't support multi-homed hosts */
1203         if (dlm_local_addr[1] != NULL) {
1204                 log_print("TCP protocol can't handle multi-homed hosts, "
1205                           "try SCTP");
1206                 return -EINVAL;
1207         }
1208
1209         log_print("Using TCP for communications");
1210
1211         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1212         if (sock) {
1213                 add_sock(sock, con);
1214                 result = 0;
1215         }
1216         else {
1217                 result = -EADDRINUSE;
1218         }
1219
1220         return result;
1221 }
1222
1223
1224
1225 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1226                                                      gfp_t allocation)
1227 {
1228         struct writequeue_entry *entry;
1229
1230         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1231         if (!entry)
1232                 return NULL;
1233
1234         entry->page = alloc_page(allocation);
1235         if (!entry->page) {
1236                 kfree(entry);
1237                 return NULL;
1238         }
1239
1240         entry->offset = 0;
1241         entry->len = 0;
1242         entry->end = 0;
1243         entry->users = 0;
1244         entry->con = con;
1245
1246         return entry;
1247 }
1248
1249 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1250 {
1251         struct connection *con;
1252         struct writequeue_entry *e;
1253         int offset = 0;
1254         int users = 0;
1255
1256         con = nodeid2con(nodeid, allocation);
1257         if (!con)
1258                 return NULL;
1259
1260         spin_lock(&con->writequeue_lock);
1261         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1262         if ((&e->list == &con->writequeue) ||
1263             (PAGE_CACHE_SIZE - e->end < len)) {
1264                 e = NULL;
1265         } else {
1266                 offset = e->end;
1267                 e->end += len;
1268                 users = e->users++;
1269         }
1270         spin_unlock(&con->writequeue_lock);
1271
1272         if (e) {
1273         got_one:
1274                 *ppc = page_address(e->page) + offset;
1275                 return e;
1276         }
1277
1278         e = new_writequeue_entry(con, allocation);
1279         if (e) {
1280                 spin_lock(&con->writequeue_lock);
1281                 offset = e->end;
1282                 e->end += len;
1283                 users = e->users++;
1284                 list_add_tail(&e->list, &con->writequeue);
1285                 spin_unlock(&con->writequeue_lock);
1286                 goto got_one;
1287         }
1288         return NULL;
1289 }
1290
1291 void dlm_lowcomms_commit_buffer(void *mh)
1292 {
1293         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1294         struct connection *con = e->con;
1295         int users;
1296
1297         spin_lock(&con->writequeue_lock);
1298         users = --e->users;
1299         if (users)
1300                 goto out;
1301         e->len = e->end - e->offset;
1302         spin_unlock(&con->writequeue_lock);
1303
1304         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1305                 queue_work(send_workqueue, &con->swork);
1306         }
1307         return;
1308
1309 out:
1310         spin_unlock(&con->writequeue_lock);
1311         return;
1312 }
1313
1314 /* Send a message */
1315 static void send_to_sock(struct connection *con)
1316 {
1317         int ret = 0;
1318         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1319         struct writequeue_entry *e;
1320         int len, offset;
1321
1322         mutex_lock(&con->sock_mutex);
1323         if (con->sock == NULL)
1324                 goto out_connect;
1325
1326         spin_lock(&con->writequeue_lock);
1327         for (;;) {
1328                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1329                                list);
1330                 if ((struct list_head *) e == &con->writequeue)
1331                         break;
1332
1333                 len = e->len;
1334                 offset = e->offset;
1335                 BUG_ON(len == 0 && e->users == 0);
1336                 spin_unlock(&con->writequeue_lock);
1337
1338                 ret = 0;
1339                 if (len) {
1340                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1341                                               msg_flags);
1342                         if (ret == -EAGAIN || ret == 0) {
1343                                 if (ret == -EAGAIN &&
1344                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1345                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1346                                         /* Notify TCP that we're limited by the
1347                                          * application window size.
1348                                          */
1349                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1350                                         con->sock->sk->sk_write_pending++;
1351                                 }
1352                                 cond_resched();
1353                                 goto out;
1354                         }
1355                         if (ret <= 0)
1356                                 goto send_error;
1357                 }
1358                         /* Don't starve people filling buffers */
1359                         cond_resched();
1360
1361                 spin_lock(&con->writequeue_lock);
1362                 e->offset += ret;
1363                 e->len -= ret;
1364
1365                 if (e->len == 0 && e->users == 0) {
1366                         list_del(&e->list);
1367                         free_entry(e);
1368                         continue;
1369                 }
1370         }
1371         spin_unlock(&con->writequeue_lock);
1372 out:
1373         mutex_unlock(&con->sock_mutex);
1374         return;
1375
1376 send_error:
1377         mutex_unlock(&con->sock_mutex);
1378         close_connection(con, false);
1379         lowcomms_connect_sock(con);
1380         return;
1381
1382 out_connect:
1383         mutex_unlock(&con->sock_mutex);
1384         if (!test_bit(CF_INIT_PENDING, &con->flags))
1385                 lowcomms_connect_sock(con);
1386         return;
1387 }
1388
1389 static void clean_one_writequeue(struct connection *con)
1390 {
1391         struct writequeue_entry *e, *safe;
1392
1393         spin_lock(&con->writequeue_lock);
1394         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1395                 list_del(&e->list);
1396                 free_entry(e);
1397         }
1398         spin_unlock(&con->writequeue_lock);
1399 }
1400
1401 /* Called from recovery when it knows that a node has
1402    left the cluster */
1403 int dlm_lowcomms_close(int nodeid)
1404 {
1405         struct connection *con;
1406
1407         log_print("closing connection to node %d", nodeid);
1408         con = nodeid2con(nodeid, 0);
1409         if (con) {
1410                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1411                 clear_bit(CF_WRITE_PENDING, &con->flags);
1412                 set_bit(CF_CLOSE, &con->flags);
1413                 if (cancel_work_sync(&con->swork))
1414                         log_print("canceled swork for node %d", nodeid);
1415                 if (cancel_work_sync(&con->rwork))
1416                         log_print("canceled rwork for node %d", nodeid);
1417                 clean_one_writequeue(con);
1418                 close_connection(con, true);
1419         }
1420         return 0;
1421 }
1422
1423 /* Receive workqueue function */
1424 static void process_recv_sockets(struct work_struct *work)
1425 {
1426         struct connection *con = container_of(work, struct connection, rwork);
1427         int err;
1428
1429         clear_bit(CF_READ_PENDING, &con->flags);
1430         do {
1431                 err = con->rx_action(con);
1432         } while (!err);
1433 }
1434
1435 /* Send workqueue function */
1436 static void process_send_sockets(struct work_struct *work)
1437 {
1438         struct connection *con = container_of(work, struct connection, swork);
1439
1440         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1441                 con->connect_action(con);
1442                 set_bit(CF_WRITE_PENDING, &con->flags);
1443         }
1444         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1445                 send_to_sock(con);
1446 }
1447
1448
1449 /* Discard all entries on the write queues */
1450 static void clean_writequeues(void)
1451 {
1452         foreach_conn(clean_one_writequeue);
1453 }
1454
1455 static void work_stop(void)
1456 {
1457         destroy_workqueue(recv_workqueue);
1458         destroy_workqueue(send_workqueue);
1459 }
1460
1461 static int work_start(void)
1462 {
1463         int error;
1464         recv_workqueue = alloc_workqueue("dlm_recv", WQ_MEM_RECLAIM |
1465                                          WQ_HIGHPRI | WQ_FREEZEABLE, 0);
1466         error = IS_ERR(recv_workqueue);
1467         if (error) {
1468                 log_print("can't start dlm_recv %d", error);
1469                 return error;
1470         }
1471
1472         send_workqueue = alloc_workqueue("dlm_send", WQ_MEM_RECLAIM |
1473                                          WQ_HIGHPRI | WQ_FREEZEABLE, 0);
1474         error = IS_ERR(send_workqueue);
1475         if (error) {
1476                 log_print("can't start dlm_send %d", error);
1477                 destroy_workqueue(recv_workqueue);
1478                 return error;
1479         }
1480
1481         return 0;
1482 }
1483
1484 static void stop_conn(struct connection *con)
1485 {
1486         con->flags |= 0x0F;
1487         if (con->sock && con->sock->sk)
1488                 con->sock->sk->sk_user_data = NULL;
1489 }
1490
1491 static void free_conn(struct connection *con)
1492 {
1493         close_connection(con, true);
1494         if (con->othercon)
1495                 kmem_cache_free(con_cache, con->othercon);
1496         hlist_del(&con->list);
1497         kmem_cache_free(con_cache, con);
1498 }
1499
1500 void dlm_lowcomms_stop(void)
1501 {
1502         /* Set all the flags to prevent any
1503            socket activity.
1504         */
1505         mutex_lock(&connections_lock);
1506         foreach_conn(stop_conn);
1507         mutex_unlock(&connections_lock);
1508
1509         work_stop();
1510
1511         mutex_lock(&connections_lock);
1512         clean_writequeues();
1513
1514         foreach_conn(free_conn);
1515
1516         mutex_unlock(&connections_lock);
1517         kmem_cache_destroy(con_cache);
1518 }
1519
1520 int dlm_lowcomms_start(void)
1521 {
1522         int error = -EINVAL;
1523         struct connection *con;
1524         int i;
1525
1526         for (i = 0; i < CONN_HASH_SIZE; i++)
1527                 INIT_HLIST_HEAD(&connection_hash[i]);
1528
1529         init_local();
1530         if (!dlm_local_count) {
1531                 error = -ENOTCONN;
1532                 log_print("no local IP address has been set");
1533                 goto out;
1534         }
1535
1536         error = -ENOMEM;
1537         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1538                                       __alignof__(struct connection), 0,
1539                                       NULL);
1540         if (!con_cache)
1541                 goto out;
1542
1543         /* Start listening */
1544         if (dlm_config.ci_protocol == 0)
1545                 error = tcp_listen_for_all();
1546         else
1547                 error = sctp_listen_for_all();
1548         if (error)
1549                 goto fail_unlisten;
1550
1551         error = work_start();
1552         if (error)
1553                 goto fail_unlisten;
1554
1555         return 0;
1556
1557 fail_unlisten:
1558         con = nodeid2con(0,0);
1559         if (con) {
1560                 close_connection(con, false);
1561                 kmem_cache_free(con_cache, con);
1562         }
1563         kmem_cache_destroy(con_cache);
1564
1565 out:
1566         return error;
1567 }