fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is 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/idr.h>
  52 #include <linux/file.h>
  53 #include <linux/mutex.h>
  54 #include <linux/sctp.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
  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         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 110         spinlock_t writequeue_lock;
 111         int (*rx_action) (struct connection *); /* What to do when active */
 112         void (*connect_action) (struct connection *);   /* What to do to connect */
 113         struct page *rx_page;
 114         struct cbuf cb;
 115         int retries;
 116 #define MAX_CONNECT_RETRIES 3
 117         int sctp_assoc;
 118         struct connection *othercon;
 119         struct work_struct rwork; /* Receive workqueue */
 120         struct work_struct swork; /* Send workqueue */
 121 };
 122 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 123
 124 /* An entry waiting to be sent */
 125 struct writequeue_entry {
 126         struct list_head list;
 127         struct page *page;
 128         int offset;
 129         int len;
 130         int end;
 131         int users;
 132         struct connection *con;
 133 };
 134
 135 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 136 static int dlm_local_count;
 137
 138 /* Work queues */
 139 static struct workqueue_struct *recv_workqueue;
 140 static struct workqueue_struct *send_workqueue;
 141
 142 static DEFINE_IDR(connections_idr);
 143 static DEFINE_MUTEX(connections_lock);
 144 static int max_nodeid;
 145 static struct kmem_cache *con_cache;
 146
 147 static void process_recv_sockets(struct work_struct *work);
 148 static void process_send_sockets(struct work_struct *work);
 149
 150 /*
 151  * If 'allocation' is zero then we don't attempt to create a new
 152  * connection structure for this node.
 153  */
 154 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 155 {
 156         struct connection *con = NULL;
 157         int r;
 158         int n;
 159
 160         con = idr_find(&connections_idr, nodeid);
 161         if (con || !alloc)
 162                 return con;
 163
 164         r = idr_pre_get(&connections_idr, alloc);
 165         if (!r)
 166                 return NULL;
 167
 168         con = kmem_cache_zalloc(con_cache, alloc);
 169         if (!con)
 170                 return NULL;
 171
 172         r = idr_get_new_above(&connections_idr, con, nodeid, &n);
 173         if (r) {
 174                 kmem_cache_free(con_cache, con);
 175                 return NULL;
 176         }
 177
 178         if (n != nodeid) {
 179                 idr_remove(&connections_idr, n);
 180                 kmem_cache_free(con_cache, con);
 181                 return NULL;
 182         }
 183
 184         con->nodeid = nodeid;
 185         mutex_init(&con->sock_mutex);
 186         INIT_LIST_HEAD(&con->writequeue);
 187         spin_lock_init(&con->writequeue_lock);
 188         INIT_WORK(&con->swork, process_send_sockets);
 189         INIT_WORK(&con->rwork, process_recv_sockets);
 190
 191         /* Setup action pointers for child sockets */
 192         if (con->nodeid) {
 193                 struct connection *zerocon = idr_find(&connections_idr, 0);
 194
 195                 con->connect_action = zerocon->connect_action;
 196                 if (!con->rx_action)
 197                         con->rx_action = zerocon->rx_action;
 198         }
 199
 200         if (nodeid > max_nodeid)
 201                 max_nodeid = nodeid;
 202
 203         return con;
 204 }
 205
 206 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 207 {
 208         struct connection *con;
 209
 210         mutex_lock(&connections_lock);
 211         con = __nodeid2con(nodeid, allocation);
 212         mutex_unlock(&connections_lock);
 213
 214         return con;
 215 }
 216
 217 /* This is a bit drastic, but only called when things go wrong */
 218 static struct connection *assoc2con(int assoc_id)
 219 {
 220         int i;
 221         struct connection *con;
 222
 223         mutex_lock(&connections_lock);
 224         for (i=0; i<=max_nodeid; i++) {
 225                 con = __nodeid2con(i, 0);
 226                 if (con && con->sctp_assoc == assoc_id) {
 227                         mutex_unlock(&connections_lock);
 228                         return con;
 229                 }
 230         }
 231         mutex_unlock(&connections_lock);
 232         return NULL;
 233 }
 234
 235 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 236 {
 237         struct sockaddr_storage addr;
 238         int error;
 239
 240         if (!dlm_local_count)
 241                 return -1;
 242
 243         error = dlm_nodeid_to_addr(nodeid, &addr);
 244         if (error)
 245                 return error;
 246
 247         if (dlm_local_addr[0]->ss_family == AF_INET) {
 248                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 249                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 250                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 251         } else {
 252                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 253                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 254                 ipv6_addr_copy(&ret6->sin6_addr, &in6->sin6_addr);
 255         }
 256
 257         return 0;
 258 }
 259
 260 /* Data available on socket or listen socket received a connect */
 261 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 262 {
 263         struct connection *con = sock2con(sk);
 264         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 265                 queue_work(recv_workqueue, &con->rwork);
 266 }
 267
 268 static void lowcomms_write_space(struct sock *sk)
 269 {
 270         struct connection *con = sock2con(sk);
 271
 272         if (con && !test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 273                 queue_work(send_workqueue, &con->swork);
 274 }
 275
 276 static inline void lowcomms_connect_sock(struct connection *con)
 277 {
 278         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 279                 queue_work(send_workqueue, &con->swork);
 280 }
 281
 282 static void lowcomms_state_change(struct sock *sk)
 283 {
 284         if (sk->sk_state == TCP_ESTABLISHED)
 285                 lowcomms_write_space(sk);
 286 }
 287
 288 /* Make a socket active */
 289 static int add_sock(struct socket *sock, struct connection *con)
 290 {
 291         con->sock = sock;
 292
 293         /* Install a data_ready callback */
 294         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 295         con->sock->sk->sk_write_space = lowcomms_write_space;
 296         con->sock->sk->sk_state_change = lowcomms_state_change;
 297         con->sock->sk->sk_user_data = con;
 298         con->sock->sk->sk_allocation = GFP_NOFS;
 299         return 0;
 300 }
 301
 302 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 303    length */
 304 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 305                           int *addr_len)
 306 {
 307         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 308         if (saddr->ss_family == AF_INET) {
 309                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 310                 in4_addr->sin_port = cpu_to_be16(port);
 311                 *addr_len = sizeof(struct sockaddr_in);
 312                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 313         } else {
 314                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 315                 in6_addr->sin6_port = cpu_to_be16(port);
 316                 *addr_len = sizeof(struct sockaddr_in6);
 317         }
 318         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 319 }
 320
 321 /* Close a remote connection and tidy up */
 322 static void close_connection(struct connection *con, bool and_other)
 323 {
 324         mutex_lock(&con->sock_mutex);
 325
 326         if (con->sock) {
 327                 sock_release(con->sock);
 328                 con->sock = NULL;
 329         }
 330         if (con->othercon && and_other) {
 331                 /* Will only re-enter once. */
 332                 close_connection(con->othercon, false);
 333         }
 334         if (con->rx_page) {
 335                 __free_page(con->rx_page);
 336                 con->rx_page = NULL;
 337         }
 338
 339         con->retries = 0;
 340         mutex_unlock(&con->sock_mutex);
 341 }
 342
 343 /* We only send shutdown messages to nodes that are not part of the cluster */
 344 static void sctp_send_shutdown(sctp_assoc_t associd)
 345 {
 346         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 347         struct msghdr outmessage;
 348         struct cmsghdr *cmsg;
 349         struct sctp_sndrcvinfo *sinfo;
 350         int ret;
 351         struct connection *con;
 352
 353         con = nodeid2con(0,0);
 354         BUG_ON(con == NULL);
 355
 356         outmessage.msg_name = NULL;
 357         outmessage.msg_namelen = 0;
 358         outmessage.msg_control = outcmsg;
 359         outmessage.msg_controllen = sizeof(outcmsg);
 360         outmessage.msg_flags = MSG_EOR;
 361
 362         cmsg = CMSG_FIRSTHDR(&outmessage);
 363         cmsg->cmsg_level = IPPROTO_SCTP;
 364         cmsg->cmsg_type = SCTP_SNDRCV;
 365         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 366         outmessage.msg_controllen = cmsg->cmsg_len;
 367         sinfo = CMSG_DATA(cmsg);
 368         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 369
 370         sinfo->sinfo_flags |= MSG_EOF;
 371         sinfo->sinfo_assoc_id = associd;
 372
 373         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 374
 375         if (ret != 0)
 376                 log_print("send EOF to node failed: %d", ret);
 377 }
 378
 379 /* INIT failed but we don't know which node...
 380    restart INIT on all pending nodes */
 381 static void sctp_init_failed(void)
 382 {
 383         int i;
 384         struct connection *con;
 385
 386         mutex_lock(&connections_lock);
 387         for (i=1; i<=max_nodeid; i++) {
 388                 con = __nodeid2con(i, 0);
 389                 if (!con)
 390                         continue;
 391                 con->sctp_assoc = 0;
 392                 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 393                         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
 394                                 queue_work(send_workqueue, &con->swork);
 395                         }
 396                 }
 397         }
 398         mutex_unlock(&connections_lock);
 399 }
 400
 401 /* Something happened to an association */
 402 static void process_sctp_notification(struct connection *con,
 403                                       struct msghdr *msg, char *buf)
 404 {
 405         union sctp_notification *sn = (union sctp_notification *)buf;
 406
 407         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 408                 switch (sn->sn_assoc_change.sac_state) {
 409
 410                 case SCTP_COMM_UP:
 411                 case SCTP_RESTART:
 412                 {
 413                         /* Check that the new node is in the lockspace */
 414                         struct sctp_prim prim;
 415                         int nodeid;
 416                         int prim_len, ret;
 417                         int addr_len;
 418                         struct connection *new_con;
 419                         struct file *file;
 420                         sctp_peeloff_arg_t parg;
 421                         int parglen = sizeof(parg);
 422
 423                         /*
 424                          * We get this before any data for an association.
 425                          * We verify that the node is in the cluster and
 426                          * then peel off a socket for it.
 427                          */
 428                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 429                                 log_print("COMM_UP for invalid assoc ID %d",
 430                                          (int)sn->sn_assoc_change.sac_assoc_id);
 431                                 sctp_init_failed();
 432                                 return;
 433                         }
 434                         memset(&prim, 0, sizeof(struct sctp_prim));
 435                         prim_len = sizeof(struct sctp_prim);
 436                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 437
 438                         ret = kernel_getsockopt(con->sock,
 439                                                 IPPROTO_SCTP,
 440                                                 SCTP_PRIMARY_ADDR,
 441                                                 (char*)&prim,
 442                                                 &prim_len);
 443                         if (ret < 0) {
 444                                 log_print("getsockopt/sctp_primary_addr on "
 445                                           "new assoc %d failed : %d",
 446                                           (int)sn->sn_assoc_change.sac_assoc_id,
 447                                           ret);
 448
 449                                 /* Retry INIT later */
 450                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 451                                 if (new_con)
 452                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 453                                 return;
 454                         }
 455                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 456                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 457                                 int i;
 458                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 459                                 log_print("reject connect from unknown addr");
 460                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 461                                         printk("%02x ", b[i]);
 462                                 printk("\n");
 463                                 sctp_send_shutdown(prim.ssp_assoc_id);
 464                                 return;
 465                         }
 466
 467                         new_con = nodeid2con(nodeid, GFP_KERNEL);
 468                         if (!new_con)
 469                                 return;
 470
 471                         /* Peel off a new sock */
 472                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 473                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 474                                                 SCTP_SOCKOPT_PEELOFF,
 475                                                 (void *)&parg, &parglen);
 476                         if (ret) {
 477                                 log_print("Can't peel off a socket for "
 478                                           "connection %d to node %d: err=%d\n",
 479                                           parg.associd, nodeid, ret);
 480                         }
 481                         file = fget(parg.sd);
 482                         new_con->sock = SOCKET_I(file->f_dentry->d_inode);
 483                         add_sock(new_con->sock, new_con);
 484                         fput(file);
 485                         put_unused_fd(parg.sd);
 486
 487                         log_print("got new/restarted association %d nodeid %d",
 488                                  (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
 489
 490                         /* Send any pending writes */
 491                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 492                         clear_bit(CF_INIT_PENDING, &con->flags);
 493                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 494                                 queue_work(send_workqueue, &new_con->swork);
 495                         }
 496                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 497                                 queue_work(recv_workqueue, &new_con->rwork);
 498                 }
 499                 break;
 500
 501                 case SCTP_COMM_LOST:
 502                 case SCTP_SHUTDOWN_COMP:
 503                 {
 504                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 505                         if (con) {
 506                                 con->sctp_assoc = 0;
 507                         }
 508                 }
 509                 break;
 510
 511                 /* We don't know which INIT failed, so clear the PENDING flags
 512                  * on them all.  if assoc_id is zero then it will then try
 513                  * again */
 514
 515                 case SCTP_CANT_STR_ASSOC:
 516                 {
 517                         log_print("Can't start SCTP association - retrying");
 518                         sctp_init_failed();
 519                 }
 520                 break;
 521
 522                 default:
 523                         log_print("unexpected SCTP assoc change id=%d state=%d",
 524                                   (int)sn->sn_assoc_change.sac_assoc_id,
 525                                   sn->sn_assoc_change.sac_state);
 526                 }
 527         }
 528 }
 529
 530 /* Data received from remote end */
 531 static int receive_from_sock(struct connection *con)
 532 {
 533         int ret = 0;
 534         struct msghdr msg = {};
 535         struct kvec iov[2];
 536         unsigned len;
 537         int r;
 538         int call_again_soon = 0;
 539         int nvec;
 540         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 541
 542         mutex_lock(&con->sock_mutex);
 543
 544         if (con->sock == NULL) {
 545                 ret = -EAGAIN;
 546                 goto out_close;
 547         }
 548
 549         if (con->rx_page == NULL) {
 550                 /*
 551                  * This doesn't need to be atomic, but I think it should
 552                  * improve performance if it is.
 553                  */
 554                 con->rx_page = alloc_page(GFP_ATOMIC);
 555                 if (con->rx_page == NULL)
 556                         goto out_resched;
 557                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 558         }
 559
 560         /* Only SCTP needs these really */
 561         memset(&incmsg, 0, sizeof(incmsg));
 562         msg.msg_control = incmsg;
 563         msg.msg_controllen = sizeof(incmsg);
 564
 565         /*
 566          * iov[0] is the bit of the circular buffer between the current end
 567          * point (cb.base + cb.len) and the end of the buffer.
 568          */
 569         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 570         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 571         iov[1].iov_len = 0;
 572         nvec = 1;
 573
 574         /*
 575          * iov[1] is the bit of the circular buffer between the start of the
 576          * buffer and the start of the currently used section (cb.base)
 577          */
 578         if (cbuf_data(&con->cb) >= con->cb.base) {
 579                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 580                 iov[1].iov_len = con->cb.base;
 581                 iov[1].iov_base = page_address(con->rx_page);
 582                 nvec = 2;
 583         }
 584         len = iov[0].iov_len + iov[1].iov_len;
 585
 586         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 587                                MSG_DONTWAIT | MSG_NOSIGNAL);
 588         if (ret <= 0)
 589                 goto out_close;
 590
 591         /* Process SCTP notifications */
 592         if (msg.msg_flags & MSG_NOTIFICATION) {
 593                 msg.msg_control = incmsg;
 594                 msg.msg_controllen = sizeof(incmsg);
 595
 596                 process_sctp_notification(con, &msg,
 597                                 page_address(con->rx_page) + con->cb.base);
 598                 mutex_unlock(&con->sock_mutex);
 599                 return 0;
 600         }
 601         BUG_ON(con->nodeid == 0);
 602
 603         if (ret == len)
 604                 call_again_soon = 1;
 605         cbuf_add(&con->cb, ret);
 606         ret = dlm_process_incoming_buffer(con->nodeid,
 607                                           page_address(con->rx_page),
 608                                           con->cb.base, con->cb.len,
 609                                           PAGE_CACHE_SIZE);
 610         if (ret == -EBADMSG) {
 611                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 612                           "iov_len=%u, iov_base[0]=%p, read=%d",
 613                           page_address(con->rx_page), con->cb.base, con->cb.len,
 614                           len, iov[0].iov_base, r);
 615         }
 616         if (ret < 0)
 617                 goto out_close;
 618         cbuf_eat(&con->cb, ret);
 619
 620         if (cbuf_empty(&con->cb) && !call_again_soon) {
 621                 __free_page(con->rx_page);
 622                 con->rx_page = NULL;
 623         }
 624
 625         if (call_again_soon)
 626                 goto out_resched;
 627         mutex_unlock(&con->sock_mutex);
 628         return 0;
 629
 630 out_resched:
 631         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 632                 queue_work(recv_workqueue, &con->rwork);
 633         mutex_unlock(&con->sock_mutex);
 634         return -EAGAIN;
 635
 636 out_close:
 637         mutex_unlock(&con->sock_mutex);
 638         if (ret != -EAGAIN) {
 639                 close_connection(con, false);
 640                 /* Reconnect when there is something to send */
 641         }
 642         /* Don't return success if we really got EOF */
 643         if (ret == 0)
 644                 ret = -EAGAIN;
 645
 646         return ret;
 647 }
 648
 649 /* Listening socket is busy, accept a connection */
 650 static int tcp_accept_from_sock(struct connection *con)
 651 {
 652         int result;
 653         struct sockaddr_storage peeraddr;
 654         struct socket *newsock;
 655         int len;
 656         int nodeid;
 657         struct connection *newcon;
 658         struct connection *addcon;
 659
 660         memset(&peeraddr, 0, sizeof(peeraddr));
 661         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 662                                   IPPROTO_TCP, &newsock);
 663         if (result < 0)
 664                 return -ENOMEM;
 665
 666         mutex_lock_nested(&con->sock_mutex, 0);
 667
 668         result = -ENOTCONN;
 669         if (con->sock == NULL)
 670                 goto accept_err;
 671
 672         newsock->type = con->sock->type;
 673         newsock->ops = con->sock->ops;
 674
 675         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 676         if (result < 0)
 677                 goto accept_err;
 678
 679         /* Get the connected socket's peer */
 680         memset(&peeraddr, 0, sizeof(peeraddr));
 681         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 682                                   &len, 2)) {
 683                 result = -ECONNABORTED;
 684                 goto accept_err;
 685         }
 686
 687         /* Get the new node's NODEID */
 688         make_sockaddr(&peeraddr, 0, &len);
 689         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 690                 log_print("connect from non cluster node");
 691                 sock_release(newsock);
 692                 mutex_unlock(&con->sock_mutex);
 693                 return -1;
 694         }
 695
 696         log_print("got connection from %d", nodeid);
 697
 698         /*  Check to see if we already have a connection to this node. This
 699          *  could happen if the two nodes initiate a connection at roughly
 700          *  the same time and the connections cross on the wire.
 701          *  In this case we store the incoming one in "othercon"
 702          */
 703         newcon = nodeid2con(nodeid, GFP_KERNEL);
 704         if (!newcon) {
 705                 result = -ENOMEM;
 706                 goto accept_err;
 707         }
 708         mutex_lock_nested(&newcon->sock_mutex, 1);
 709         if (newcon->sock) {
 710                 struct connection *othercon = newcon->othercon;
 711
 712                 if (!othercon) {
 713                         othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
 714                         if (!othercon) {
 715                                 log_print("failed to allocate incoming socket");
 716                                 mutex_unlock(&newcon->sock_mutex);
 717                                 result = -ENOMEM;
 718                                 goto accept_err;
 719                         }
 720                         othercon->nodeid = nodeid;
 721                         othercon->rx_action = receive_from_sock;
 722                         mutex_init(&othercon->sock_mutex);
 723                         INIT_WORK(&othercon->swork, process_send_sockets);
 724                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 725                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 726                 }
 727                 if (!othercon->sock) {
 728                         newcon->othercon = othercon;
 729                         othercon->sock = newsock;
 730                         newsock->sk->sk_user_data = othercon;
 731                         add_sock(newsock, othercon);
 732                         addcon = othercon;
 733                 }
 734                 else {
 735                         printk("Extra connection from node %d attempted\n", nodeid);
 736                         result = -EAGAIN;
 737                         mutex_unlock(&newcon->sock_mutex);
 738                         goto accept_err;
 739                 }
 740         }
 741         else {
 742                 newsock->sk->sk_user_data = newcon;
 743                 newcon->rx_action = receive_from_sock;
 744                 add_sock(newsock, newcon);
 745                 addcon = newcon;
 746         }
 747
 748         mutex_unlock(&newcon->sock_mutex);
 749
 750         /*
 751          * Add it to the active queue in case we got data
 752          * beween processing the accept adding the socket
 753          * to the read_sockets list
 754          */
 755         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 756                 queue_work(recv_workqueue, &addcon->rwork);
 757         mutex_unlock(&con->sock_mutex);
 758
 759         return 0;
 760
 761 accept_err:
 762         mutex_unlock(&con->sock_mutex);
 763         sock_release(newsock);
 764
 765         if (result != -EAGAIN)
 766                 log_print("error accepting connection from node: %d", result);
 767         return result;
 768 }
 769
 770 static void free_entry(struct writequeue_entry *e)
 771 {
 772         __free_page(e->page);
 773         kfree(e);
 774 }
 775
 776 /* Initiate an SCTP association.
 777    This is a special case of send_to_sock() in that we don't yet have a
 778    peeled-off socket for this association, so we use the listening socket
 779    and add the primary IP address of the remote node.
 780  */
 781 static void sctp_init_assoc(struct connection *con)
 782 {
 783         struct sockaddr_storage rem_addr;
 784         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 785         struct msghdr outmessage;
 786         struct cmsghdr *cmsg;
 787         struct sctp_sndrcvinfo *sinfo;
 788         struct connection *base_con;
 789         struct writequeue_entry *e;
 790         int len, offset;
 791         int ret;
 792         int addrlen;
 793         struct kvec iov[1];
 794
 795         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 796                 return;
 797
 798         if (con->retries++ > MAX_CONNECT_RETRIES)
 799                 return;
 800
 801         log_print("Initiating association with node %d", con->nodeid);
 802
 803         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 804                 log_print("no address for nodeid %d", con->nodeid);
 805                 return;
 806         }
 807         base_con = nodeid2con(0, 0);
 808         BUG_ON(base_con == NULL);
 809
 810         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 811
 812         outmessage.msg_name = &rem_addr;
 813         outmessage.msg_namelen = addrlen;
 814         outmessage.msg_control = outcmsg;
 815         outmessage.msg_controllen = sizeof(outcmsg);
 816         outmessage.msg_flags = MSG_EOR;
 817
 818         spin_lock(&con->writequeue_lock);
 819         e = list_entry(con->writequeue.next, struct writequeue_entry,
 820                        list);
 821
 822         BUG_ON((struct list_head *) e == &con->writequeue);
 823
 824         len = e->len;
 825         offset = e->offset;
 826         spin_unlock(&con->writequeue_lock);
 827
 828         /* Send the first block off the write queue */
 829         iov[0].iov_base = page_address(e->page)+offset;
 830         iov[0].iov_len = len;
 831
 832         cmsg = CMSG_FIRSTHDR(&outmessage);
 833         cmsg->cmsg_level = IPPROTO_SCTP;
 834         cmsg->cmsg_type = SCTP_SNDRCV;
 835         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 836         sinfo = CMSG_DATA(cmsg);
 837         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 838         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 839         outmessage.msg_controllen = cmsg->cmsg_len;
 840
 841         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 842         if (ret < 0) {
 843                 log_print("Send first packet to node %d failed: %d",
 844                           con->nodeid, ret);
 845
 846                 /* Try again later */
 847                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 848                 clear_bit(CF_INIT_PENDING, &con->flags);
 849         }
 850         else {
 851                 spin_lock(&con->writequeue_lock);
 852                 e->offset += ret;
 853                 e->len -= ret;
 854
 855                 if (e->len == 0 && e->users == 0) {
 856                         list_del(&e->list);
 857                         free_entry(e);
 858                 }
 859                 spin_unlock(&con->writequeue_lock);
 860         }
 861 }
 862
 863 /* Connect a new socket to its peer */
 864 static void tcp_connect_to_sock(struct connection *con)
 865 {
 866         int result = -EHOSTUNREACH;
 867         struct sockaddr_storage saddr, src_addr;
 868         int addr_len;
 869         struct socket *sock;
 870
 871         if (con->nodeid == 0) {
 872                 log_print("attempt to connect sock 0 foiled");
 873                 return;
 874         }
 875
 876         mutex_lock(&con->sock_mutex);
 877         if (con->retries++ > MAX_CONNECT_RETRIES)
 878                 goto out;
 879
 880         /* Some odd races can cause double-connects, ignore them */
 881         if (con->sock) {
 882                 result = 0;
 883                 goto out;
 884         }
 885
 886         /* Create a socket to communicate with */
 887         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 888                                   IPPROTO_TCP, &sock);
 889         if (result < 0)
 890                 goto out_err;
 891
 892         memset(&saddr, 0, sizeof(saddr));
 893         if (dlm_nodeid_to_addr(con->nodeid, &saddr)) {
 894                 sock_release(sock);
 895                 goto out_err;
 896         }
 897
 898         sock->sk->sk_user_data = con;
 899         con->rx_action = receive_from_sock;
 900         con->connect_action = tcp_connect_to_sock;
 901         add_sock(sock, con);
 902
 903         /* Bind to our cluster-known address connecting to avoid
 904            routing problems */
 905         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 906         make_sockaddr(&src_addr, 0, &addr_len);
 907         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 908                                  addr_len);
 909         if (result < 0) {
 910                 log_print("could not bind for connect: %d", result);
 911                 /* This *may* not indicate a critical error */
 912         }
 913
 914         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 915
 916         log_print("connecting to %d", con->nodeid);
 917         result =
 918                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 919                                    O_NONBLOCK);
 920         if (result == -EINPROGRESS)
 921                 result = 0;
 922         if (result == 0)
 923                 goto out;
 924
 925 out_err:
 926         if (con->sock) {
 927                 sock_release(con->sock);
 928                 con->sock = NULL;
 929         }
 930         /*
 931          * Some errors are fatal and this list might need adjusting. For other
 932          * errors we try again until the max number of retries is reached.
 933          */
 934         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 935             result != -ENETDOWN && result != -EINVAL
 936             && result != -EPROTONOSUPPORT) {
 937                 lowcomms_connect_sock(con);
 938                 result = 0;
 939         }
 940 out:
 941         mutex_unlock(&con->sock_mutex);
 942         return;
 943 }
 944
 945 static struct socket *tcp_create_listen_sock(struct connection *con,
 946                                              struct sockaddr_storage *saddr)
 947 {
 948         struct socket *sock = NULL;
 949         int result = 0;
 950         int one = 1;
 951         int addr_len;
 952
 953         if (dlm_local_addr[0]->ss_family == AF_INET)
 954                 addr_len = sizeof(struct sockaddr_in);
 955         else
 956                 addr_len = sizeof(struct sockaddr_in6);
 957
 958         /* Create a socket to communicate with */
 959         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 960                                   IPPROTO_TCP, &sock);
 961         if (result < 0) {
 962                 log_print("Can't create listening comms socket");
 963                 goto create_out;
 964         }
 965
 966         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
 967                                    (char *)&one, sizeof(one));
 968
 969         if (result < 0) {
 970                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
 971         }
 972         sock->sk->sk_user_data = con;
 973         con->rx_action = tcp_accept_from_sock;
 974         con->connect_action = tcp_connect_to_sock;
 975         con->sock = sock;
 976
 977         /* Bind to our port */
 978         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
 979         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
 980         if (result < 0) {
 981                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
 982                 sock_release(sock);
 983                 sock = NULL;
 984                 con->sock = NULL;
 985                 goto create_out;
 986         }
 987         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 988                                  (char *)&one, sizeof(one));
 989         if (result < 0) {
 990                 log_print("Set keepalive failed: %d", result);
 991         }
 992
 993         result = sock->ops->listen(sock, 5);
 994         if (result < 0) {
 995                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
 996                 sock_release(sock);
 997                 sock = NULL;
 998                 goto create_out;
 999         }
1000
1001 create_out:
1002         return sock;
1003 }
1004
1005 /* Get local addresses */
1006 static void init_local(void)
1007 {
1008         struct sockaddr_storage sas, *addr;
1009         int i;
1010
1011         dlm_local_count = 0;
1012         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1013                 if (dlm_our_addr(&sas, i))
1014                         break;
1015
1016                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1017                 if (!addr)
1018                         break;
1019                 memcpy(addr, &sas, sizeof(*addr));
1020                 dlm_local_addr[dlm_local_count++] = addr;
1021         }
1022 }
1023
1024 /* Bind to an IP address. SCTP allows multiple address so it can do
1025    multi-homing */
1026 static int add_sctp_bind_addr(struct connection *sctp_con,
1027                               struct sockaddr_storage *addr,
1028                               int addr_len, int num)
1029 {
1030         int result = 0;
1031
1032         if (num == 1)
1033                 result = kernel_bind(sctp_con->sock,
1034                                      (struct sockaddr *) addr,
1035                                      addr_len);
1036         else
1037                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1038                                            SCTP_SOCKOPT_BINDX_ADD,
1039                                            (char *)addr, addr_len);
1040
1041         if (result < 0)
1042                 log_print("Can't bind to port %d addr number %d",
1043                           dlm_config.ci_tcp_port, num);
1044
1045         return result;
1046 }
1047
1048 /* Initialise SCTP socket and bind to all interfaces */
1049 static int sctp_listen_for_all(void)
1050 {
1051         struct socket *sock = NULL;
1052         struct sockaddr_storage localaddr;
1053         struct sctp_event_subscribe subscribe;
1054         int result = -EINVAL, num = 1, i, addr_len;
1055         struct connection *con = nodeid2con(0, GFP_KERNEL);
1056         int bufsize = NEEDED_RMEM;
1057
1058         if (!con)
1059                 return -ENOMEM;
1060
1061         log_print("Using SCTP for communications");
1062
1063         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1064                                   IPPROTO_SCTP, &sock);
1065         if (result < 0) {
1066                 log_print("Can't create comms socket, check SCTP is loaded");
1067                 goto out;
1068         }
1069
1070         /* Listen for events */
1071         memset(&subscribe, 0, sizeof(subscribe));
1072         subscribe.sctp_data_io_event = 1;
1073         subscribe.sctp_association_event = 1;
1074         subscribe.sctp_send_failure_event = 1;
1075         subscribe.sctp_shutdown_event = 1;
1076         subscribe.sctp_partial_delivery_event = 1;
1077
1078         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1079                                  (char *)&bufsize, sizeof(bufsize));
1080         if (result)
1081                 log_print("Error increasing buffer space on socket %d", result);
1082
1083         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1084                                    (char *)&subscribe, sizeof(subscribe));
1085         if (result < 0) {
1086                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1087                           result);
1088                 goto create_delsock;
1089         }
1090
1091         /* Init con struct */
1092         sock->sk->sk_user_data = con;
1093         con->sock = sock;
1094         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1095         con->rx_action = receive_from_sock;
1096         con->connect_action = sctp_init_assoc;
1097
1098         /* Bind to all interfaces. */
1099         for (i = 0; i < dlm_local_count; i++) {
1100                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1101                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1102
1103                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1104                 if (result)
1105                         goto create_delsock;
1106                 ++num;
1107         }
1108
1109         result = sock->ops->listen(sock, 5);
1110         if (result < 0) {
1111                 log_print("Can't set socket listening");
1112                 goto create_delsock;
1113         }
1114
1115         return 0;
1116
1117 create_delsock:
1118         sock_release(sock);
1119         con->sock = NULL;
1120 out:
1121         return result;
1122 }
1123
1124 static int tcp_listen_for_all(void)
1125 {
1126         struct socket *sock = NULL;
1127         struct connection *con = nodeid2con(0, GFP_KERNEL);
1128         int result = -EINVAL;
1129
1130         if (!con)
1131                 return -ENOMEM;
1132
1133         /* We don't support multi-homed hosts */
1134         if (dlm_local_addr[1] != NULL) {
1135                 log_print("TCP protocol can't handle multi-homed hosts, "
1136                           "try SCTP");
1137                 return -EINVAL;
1138         }
1139
1140         log_print("Using TCP for communications");
1141
1142         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1143         if (sock) {
1144                 add_sock(sock, con);
1145                 result = 0;
1146         }
1147         else {
1148                 result = -EADDRINUSE;
1149         }
1150
1151         return result;
1152 }
1153
1154
1155
1156 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1157                                                      gfp_t allocation)
1158 {
1159         struct writequeue_entry *entry;
1160
1161         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1162         if (!entry)
1163                 return NULL;
1164
1165         entry->page = alloc_page(allocation);
1166         if (!entry->page) {
1167                 kfree(entry);
1168                 return NULL;
1169         }
1170
1171         entry->offset = 0;
1172         entry->len = 0;
1173         entry->end = 0;
1174         entry->users = 0;
1175         entry->con = con;
1176
1177         return entry;
1178 }
1179
1180 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1181 {
1182         struct connection *con;
1183         struct writequeue_entry *e;
1184         int offset = 0;
1185         int users = 0;
1186
1187         con = nodeid2con(nodeid, allocation);
1188         if (!con)
1189                 return NULL;
1190
1191         spin_lock(&con->writequeue_lock);
1192         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1193         if ((&e->list == &con->writequeue) ||
1194             (PAGE_CACHE_SIZE - e->end < len)) {
1195                 e = NULL;
1196         } else {
1197                 offset = e->end;
1198                 e->end += len;
1199                 users = e->users++;
1200         }
1201         spin_unlock(&con->writequeue_lock);
1202
1203         if (e) {
1204         got_one:
1205                 *ppc = page_address(e->page) + offset;
1206                 return e;
1207         }
1208
1209         e = new_writequeue_entry(con, allocation);
1210         if (e) {
1211                 spin_lock(&con->writequeue_lock);
1212                 offset = e->end;
1213                 e->end += len;
1214                 users = e->users++;
1215                 list_add_tail(&e->list, &con->writequeue);
1216                 spin_unlock(&con->writequeue_lock);
1217                 goto got_one;
1218         }
1219         return NULL;
1220 }
1221
1222 void dlm_lowcomms_commit_buffer(void *mh)
1223 {
1224         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1225         struct connection *con = e->con;
1226         int users;
1227
1228         spin_lock(&con->writequeue_lock);
1229         users = --e->users;
1230         if (users)
1231                 goto out;
1232         e->len = e->end - e->offset;
1233         spin_unlock(&con->writequeue_lock);
1234
1235         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1236                 queue_work(send_workqueue, &con->swork);
1237         }
1238         return;
1239
1240 out:
1241         spin_unlock(&con->writequeue_lock);
1242         return;
1243 }
1244
1245 /* Send a message */
1246 static void send_to_sock(struct connection *con)
1247 {
1248         int ret = 0;
1249         ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1250         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1251         struct writequeue_entry *e;
1252         int len, offset;
1253
1254         mutex_lock(&con->sock_mutex);
1255         if (con->sock == NULL)
1256                 goto out_connect;
1257
1258         sendpage = con->sock->ops->sendpage;
1259
1260         spin_lock(&con->writequeue_lock);
1261         for (;;) {
1262                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1263                                list);
1264                 if ((struct list_head *) e == &con->writequeue)
1265                         break;
1266
1267                 len = e->len;
1268                 offset = e->offset;
1269                 BUG_ON(len == 0 && e->users == 0);
1270                 spin_unlock(&con->writequeue_lock);
1271
1272                 ret = 0;
1273                 if (len) {
1274                         ret = sendpage(con->sock, e->page, offset, len,
1275                                        msg_flags);
1276                         if (ret == -EAGAIN || ret == 0) {
1277                                 cond_resched();
1278                                 goto out;
1279                         }
1280                         if (ret <= 0)
1281                                 goto send_error;
1282                 }
1283                         /* Don't starve people filling buffers */
1284                         cond_resched();
1285
1286                 spin_lock(&con->writequeue_lock);
1287                 e->offset += ret;
1288                 e->len -= ret;
1289
1290                 if (e->len == 0 && e->users == 0) {
1291                         list_del(&e->list);
1292                         free_entry(e);
1293                         continue;
1294                 }
1295         }
1296         spin_unlock(&con->writequeue_lock);
1297 out:
1298         mutex_unlock(&con->sock_mutex);
1299         return;
1300
1301 send_error:
1302         mutex_unlock(&con->sock_mutex);
1303         close_connection(con, false);
1304         lowcomms_connect_sock(con);
1305         return;
1306
1307 out_connect:
1308         mutex_unlock(&con->sock_mutex);
1309         if (!test_bit(CF_INIT_PENDING, &con->flags))
1310                 lowcomms_connect_sock(con);
1311         return;
1312 }
1313
1314 static void clean_one_writequeue(struct connection *con)
1315 {
1316         struct list_head *list;
1317         struct list_head *temp;
1318
1319         spin_lock(&con->writequeue_lock);
1320         list_for_each_safe(list, temp, &con->writequeue) {
1321                 struct writequeue_entry *e =
1322                         list_entry(list, struct writequeue_entry, list);
1323                 list_del(&e->list);
1324                 free_entry(e);
1325         }
1326         spin_unlock(&con->writequeue_lock);
1327 }
1328
1329 /* Called from recovery when it knows that a node has
1330    left the cluster */
1331 int dlm_lowcomms_close(int nodeid)
1332 {
1333         struct connection *con;
1334
1335         log_print("closing connection to node %d", nodeid);
1336         con = nodeid2con(nodeid, 0);
1337         if (con) {
1338                 clean_one_writequeue(con);
1339                 close_connection(con, true);
1340         }
1341         return 0;
1342 }
1343
1344 /* Receive workqueue function */
1345 static void process_recv_sockets(struct work_struct *work)
1346 {
1347         struct connection *con = container_of(work, struct connection, rwork);
1348         int err;
1349
1350         clear_bit(CF_READ_PENDING, &con->flags);
1351         do {
1352                 err = con->rx_action(con);
1353         } while (!err);
1354 }
1355
1356 /* Send workqueue function */
1357 static void process_send_sockets(struct work_struct *work)
1358 {
1359         struct connection *con = container_of(work, struct connection, swork);
1360
1361         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1362                 con->connect_action(con);
1363         }
1364         clear_bit(CF_WRITE_PENDING, &con->flags);
1365         send_to_sock(con);
1366 }
1367
1368
1369 /* Discard all entries on the write queues */
1370 static void clean_writequeues(void)
1371 {
1372         int nodeid;
1373
1374         for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1375                 struct connection *con = __nodeid2con(nodeid, 0);
1376
1377                 if (con)
1378                         clean_one_writequeue(con);
1379         }
1380 }
1381
1382 static void work_stop(void)
1383 {
1384         destroy_workqueue(recv_workqueue);
1385         destroy_workqueue(send_workqueue);
1386 }
1387
1388 static int work_start(void)
1389 {
1390         int error;
1391         recv_workqueue = create_workqueue("dlm_recv");
1392         error = IS_ERR(recv_workqueue);
1393         if (error) {
1394                 log_print("can't start dlm_recv %d", error);
1395                 return error;
1396         }
1397
1398         send_workqueue = create_singlethread_workqueue("dlm_send");
1399         error = IS_ERR(send_workqueue);
1400         if (error) {
1401                 log_print("can't start dlm_send %d", error);
1402                 destroy_workqueue(recv_workqueue);
1403                 return error;
1404         }
1405
1406         return 0;
1407 }
1408
1409 void dlm_lowcomms_stop(void)
1410 {
1411         int i;
1412         struct connection *con;
1413
1414         /* Set all the flags to prevent any
1415            socket activity.
1416         */
1417         mutex_lock(&connections_lock);
1418         for (i = 0; i <= max_nodeid; i++) {
1419                 con = __nodeid2con(i, 0);
1420                 if (con) {
1421                         con->flags |= 0x0F;
1422                         if (con->sock)
1423                                 con->sock->sk->sk_user_data = NULL;
1424                 }
1425         }
1426         mutex_unlock(&connections_lock);
1427
1428         work_stop();
1429
1430         mutex_lock(&connections_lock);
1431         clean_writequeues();
1432
1433         for (i = 0; i <= max_nodeid; i++) {
1434                 con = __nodeid2con(i, 0);
1435                 if (con) {
1436                         close_connection(con, true);
1437                         if (con->othercon)
1438                                 kmem_cache_free(con_cache, con->othercon);
1439                         kmem_cache_free(con_cache, con);
1440                 }
1441         }
1442         max_nodeid = 0;
1443         mutex_unlock(&connections_lock);
1444         kmem_cache_destroy(con_cache);
1445         idr_init(&connections_idr);
1446 }
1447
1448 int dlm_lowcomms_start(void)
1449 {
1450         int error = -EINVAL;
1451         struct connection *con;
1452
1453         init_local();
1454         if (!dlm_local_count) {
1455                 error = -ENOTCONN;
1456                 log_print("no local IP address has been set");
1457                 goto out;
1458         }
1459
1460         error = -ENOMEM;
1461         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1462                                       __alignof__(struct connection), 0,
1463                                       NULL);
1464         if (!con_cache)
1465                 goto out;
1466
1467         /* Start listening */
1468         if (dlm_config.ci_protocol == 0)
1469                 error = tcp_listen_for_all();
1470         else
1471                 error = sctp_listen_for_all();
1472         if (error)
1473                 goto fail_unlisten;
1474
1475         error = work_start();
1476         if (error)
1477                 goto fail_unlisten;
1478
1479         return 0;
1480
1481 fail_unlisten:
1482         con = nodeid2con(0,0);
1483         if (con) {
1484                 close_connection(con, false);
1485                 kmem_cache_free(con_cache, con);
1486         }
1487         kmem_cache_destroy(con_cache);
1488
1489 out:
1490         return error;
1491 }