git.oblomov.eu Git - linux-2.6/blob - fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is it's
  25  * responsibility. It is this layer's
  26  * responsibility to resolve these into IP address or
  27  * whatever it needs for inter-node communication.
  28  *
  29  * The comms level is two kernel threads that deal mainly with
  30  * the receiving of messages from other nodes and passing them
  31  * up to the mid-level comms layer (which understands the
  32  * message format) for execution by the locking core, and
  33  * a send thread which does all the setting up of connections
  34  * to remote nodes and the sending of data. Threads are not allowed
  35  * to send their own data because it may cause them to wait in times
  36  * of high load. Also, this way, the sending thread can collect together
  37  * messages bound for one node and send them in one block.
  38  *
  39  * lowcomms will choose to use wither TCP or SCTP as its transport layer
  40  * depending on the configuration variable 'protocol'. This should be set
  41  * to 0 (default) for TCP or 1 for SCTP. It shouldbe configured using a
  42  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  43  * for the DLM to function.
  44  *
  45  */
  46
  47 #include <asm/ioctls.h>
  48 #include <net/sock.h>
  49 #include <net/tcp.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/idr.h>
  52 #include <linux/file.h>
  53 #include <linux/mutex.h>
  54 #include <linux/sctp.h>
  55 #include <net/sctp/user.h>
  56
  57 #include "dlm_internal.h"
  58 #include "lowcomms.h"
  59 #include "midcomms.h"
  60 #include "config.h"
  61
  62 #define NEEDED_RMEM (4*1024*1024)
  63
  64 struct cbuf {
  65         unsigned int base;
  66         unsigned int len;
  67         unsigned int mask;
  68 };
  69
  70 static void cbuf_add(struct cbuf *cb, int n)
  71 {
  72         cb->len += n;
  73 }
  74
  75 static int cbuf_data(struct cbuf *cb)
  76 {
  77         return ((cb->base + cb->len) & cb->mask);
  78 }
  79
  80 static void cbuf_init(struct cbuf *cb, int size)
  81 {
  82         cb->base = cb->len = 0;
  83         cb->mask = size-1;
  84 }
  85
  86 static void cbuf_eat(struct cbuf *cb, int n)
  87 {
  88         cb->len  -= n;
  89         cb->base += n;
  90         cb->base &= cb->mask;
  91 }
  92
  93 static bool cbuf_empty(struct cbuf *cb)
  94 {
  95         return cb->len == 0;
  96 }
  97
  98 struct connection {
  99         struct socket *sock;    /* NULL if not connected */
 100         uint32_t nodeid;        /* So we know who we are in the list */
 101         struct mutex sock_mutex;
 102         unsigned long flags;
 103 #define CF_READ_PENDING 1
 104 #define CF_WRITE_PENDING 2
 105 #define CF_CONNECT_PENDING 3
 106 #define CF_INIT_PENDING 4
 107 #define CF_IS_OTHERCON 5
 108         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 109         spinlock_t writequeue_lock;
 110         int (*rx_action) (struct connection *); /* What to do when active */
 111         void (*connect_action) (struct connection *);   /* What to do to connect */
 112         struct page *rx_page;
 113         struct cbuf cb;
 114         int retries;
 115 #define MAX_CONNECT_RETRIES 3
 116         int sctp_assoc;
 117         struct connection *othercon;
 118         struct work_struct rwork; /* Receive workqueue */
 119         struct work_struct swork; /* Send workqueue */
 120 };
 121 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 122
 123 /* An entry waiting to be sent */
 124 struct writequeue_entry {
 125         struct list_head list;
 126         struct page *page;
 127         int offset;
 128         int len;
 129         int end;
 130         int users;
 131         struct connection *con;
 132 };
 133
 134 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 135 static int dlm_local_count;
 136
 137 /* Work queues */
 138 static struct workqueue_struct *recv_workqueue;
 139 static struct workqueue_struct *send_workqueue;
 140
 141 static DEFINE_IDR(connections_idr);
 142 static DEFINE_MUTEX(connections_lock);
 143 static int max_nodeid;
 144 static struct kmem_cache *con_cache;
 145
 146 static void process_recv_sockets(struct work_struct *work);
 147 static void process_send_sockets(struct work_struct *work);
 148
 149 /*
 150  * If 'allocation' is zero then we don't attempt to create a new
 151  * connection structure for this node.
 152  */
 153 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 154 {
 155         struct connection *con = NULL;
 156         int r;
 157         int n;
 158
 159         con = idr_find(&connections_idr, nodeid);
 160         if (con || !alloc)
 161                 return con;
 162
 163         r = idr_pre_get(&connections_idr, alloc);
 164         if (!r)
 165                 return NULL;
 166
 167         con = kmem_cache_zalloc(con_cache, alloc);
 168         if (!con)
 169                 return NULL;
 170
 171         r = idr_get_new_above(&connections_idr, con, nodeid, &n);
 172         if (r) {
 173                 kmem_cache_free(con_cache, con);
 174                 return NULL;
 175         }
 176
 177         if (n != nodeid) {
 178                 idr_remove(&connections_idr, n);
 179                 kmem_cache_free(con_cache, con);
 180                 return NULL;
 181         }
 182
 183         con->nodeid = nodeid;
 184         mutex_init(&con->sock_mutex);
 185         INIT_LIST_HEAD(&con->writequeue);
 186         spin_lock_init(&con->writequeue_lock);
 187         INIT_WORK(&con->swork, process_send_sockets);
 188         INIT_WORK(&con->rwork, process_recv_sockets);
 189
 190         /* Setup action pointers for child sockets */
 191         if (con->nodeid) {
 192                 struct connection *zerocon = idr_find(&connections_idr, 0);
 193
 194                 con->connect_action = zerocon->connect_action;
 195                 if (!con->rx_action)
 196                         con->rx_action = zerocon->rx_action;
 197         }
 198
 199         if (nodeid > max_nodeid)
 200                 max_nodeid = nodeid;
 201
 202         return con;
 203 }
 204
 205 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 206 {
 207         struct connection *con;
 208
 209         mutex_lock(&connections_lock);
 210         con = __nodeid2con(nodeid, allocation);
 211         mutex_unlock(&connections_lock);
 212
 213         return con;
 214 }
 215
 216 /* This is a bit drastic, but only called when things go wrong */
 217 static struct connection *assoc2con(int assoc_id)
 218 {
 219         int i;
 220         struct connection *con;
 221
 222         mutex_lock(&connections_lock);
 223         for (i=0; i<=max_nodeid; i++) {
 224                 con = __nodeid2con(i, 0);
 225                 if (con && con->sctp_assoc == assoc_id) {
 226                         mutex_unlock(&connections_lock);
 227                         return con;
 228                 }
 229         }
 230         mutex_unlock(&connections_lock);
 231         return NULL;
 232 }
 233
 234 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 235 {
 236         struct sockaddr_storage addr;
 237         int error;
 238
 239         if (!dlm_local_count)
 240                 return -1;
 241
 242         error = dlm_nodeid_to_addr(nodeid, &addr);
 243         if (error)
 244                 return error;
 245
 246         if (dlm_local_addr[0]->ss_family == AF_INET) {
 247                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 248                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 249                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 250         } else {
 251                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 252                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 253                 memcpy(&ret6->sin6_addr, &in6->sin6_addr,
 254                        sizeof(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         return 0;
 299 }
 300
 301 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 302    length */
 303 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 304                           int *addr_len)
 305 {
 306         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 307         if (saddr->ss_family == AF_INET) {
 308                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 309                 in4_addr->sin_port = cpu_to_be16(port);
 310                 *addr_len = sizeof(struct sockaddr_in);
 311                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 312         } else {
 313                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 314                 in6_addr->sin6_port = cpu_to_be16(port);
 315                 *addr_len = sizeof(struct sockaddr_in6);
 316         }
 317         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 318 }
 319
 320 /* Close a remote connection and tidy up */
 321 static void close_connection(struct connection *con, bool and_other)
 322 {
 323         mutex_lock(&con->sock_mutex);
 324
 325         if (con->sock) {
 326                 sock_release(con->sock);
 327                 con->sock = NULL;
 328         }
 329         if (con->othercon && and_other) {
 330                 /* Will only re-enter once. */
 331                 close_connection(con->othercon, false);
 332         }
 333         if (con->rx_page) {
 334                 __free_page(con->rx_page);
 335                 con->rx_page = NULL;
 336         }
 337
 338         con->retries = 0;
 339         mutex_unlock(&con->sock_mutex);
 340 }
 341
 342 /* We only send shutdown messages to nodes that are not part of the cluster */
 343 static void sctp_send_shutdown(sctp_assoc_t associd)
 344 {
 345         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 346         struct msghdr outmessage;
 347         struct cmsghdr *cmsg;
 348         struct sctp_sndrcvinfo *sinfo;
 349         int ret;
 350         struct connection *con;
 351
 352         con = nodeid2con(0,0);
 353         BUG_ON(con == NULL);
 354
 355         outmessage.msg_name = NULL;
 356         outmessage.msg_namelen = 0;
 357         outmessage.msg_control = outcmsg;
 358         outmessage.msg_controllen = sizeof(outcmsg);
 359         outmessage.msg_flags = MSG_EOR;
 360
 361         cmsg = CMSG_FIRSTHDR(&outmessage);
 362         cmsg->cmsg_level = IPPROTO_SCTP;
 363         cmsg->cmsg_type = SCTP_SNDRCV;
 364         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 365         outmessage.msg_controllen = cmsg->cmsg_len;
 366         sinfo = CMSG_DATA(cmsg);
 367         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 368
 369         sinfo->sinfo_flags |= MSG_EOF;
 370         sinfo->sinfo_assoc_id = associd;
 371
 372         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 373
 374         if (ret != 0)
 375                 log_print("send EOF to node failed: %d", ret);
 376 }
 377
 378 /* INIT failed but we don't know which node...
 379    restart INIT on all pending nodes */
 380 static void sctp_init_failed(void)
 381 {
 382         int i;
 383         struct connection *con;
 384
 385         mutex_lock(&connections_lock);
 386         for (i=1; i<=max_nodeid; i++) {
 387                 con = __nodeid2con(i, 0);
 388                 if (!con)
 389                         continue;
 390                 con->sctp_assoc = 0;
 391                 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 392                         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
 393                                 queue_work(send_workqueue, &con->swork);
 394                         }
 395                 }
 396         }
 397         mutex_unlock(&connections_lock);
 398 }
 399
 400 /* Something happened to an association */
 401 static void process_sctp_notification(struct connection *con,
 402                                       struct msghdr *msg, char *buf)
 403 {
 404         union sctp_notification *sn = (union sctp_notification *)buf;
 405
 406         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 407                 switch (sn->sn_assoc_change.sac_state) {
 408
 409                 case SCTP_COMM_UP:
 410                 case SCTP_RESTART:
 411                 {
 412                         /* Check that the new node is in the lockspace */
 413                         struct sctp_prim prim;
 414                         int nodeid;
 415                         int prim_len, ret;
 416                         int addr_len;
 417                         struct connection *new_con;
 418                         struct file *file;
 419                         sctp_peeloff_arg_t parg;
 420                         int parglen = sizeof(parg);
 421
 422                         /*
 423                          * We get this before any data for an association.
 424                          * We verify that the node is in the cluster and
 425                          * then peel off a socket for it.
 426                          */
 427                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 428                                 log_print("COMM_UP for invalid assoc ID %d",
 429                                          (int)sn->sn_assoc_change.sac_assoc_id);
 430                                 sctp_init_failed();
 431                                 return;
 432                         }
 433                         memset(&prim, 0, sizeof(struct sctp_prim));
 434                         prim_len = sizeof(struct sctp_prim);
 435                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 436
 437                         ret = kernel_getsockopt(con->sock,
 438                                                 IPPROTO_SCTP,
 439                                                 SCTP_PRIMARY_ADDR,
 440                                                 (char*)&prim,
 441                                                 &prim_len);
 442                         if (ret < 0) {
 443                                 log_print("getsockopt/sctp_primary_addr on "
 444                                           "new assoc %d failed : %d",
 445                                           (int)sn->sn_assoc_change.sac_assoc_id,
 446                                           ret);
 447
 448                                 /* Retry INIT later */
 449                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 450                                 if (new_con)
 451                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 452                                 return;
 453                         }
 454                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 455                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 456                                 int i;
 457                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 458                                 log_print("reject connect from unknown addr");
 459                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 460                                         printk("%02x ", b[i]);
 461                                 printk("\n");
 462                                 sctp_send_shutdown(prim.ssp_assoc_id);
 463                                 return;
 464                         }
 465
 466                         new_con = nodeid2con(nodeid, GFP_KERNEL);
 467                         if (!new_con)
 468                                 return;
 469
 470                         /* Peel off a new sock */
 471                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 472                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 473                                                 SCTP_SOCKOPT_PEELOFF,
 474                                                 (void *)&parg, &parglen);
 475                         if (ret) {
 476                                 log_print("Can't peel off a socket for "
 477                                           "connection %d to node %d: err=%d\n",
 478                                           parg.associd, nodeid, ret);
 479                         }
 480                         file = fget(parg.sd);
 481                         new_con->sock = SOCKET_I(file->f_dentry->d_inode);
 482                         add_sock(new_con->sock, new_con);
 483                         fput(file);
 484                         put_unused_fd(parg.sd);
 485
 486                         log_print("got new/restarted association %d nodeid %d",
 487                                  (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
 488
 489                         /* Send any pending writes */
 490                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 491                         clear_bit(CF_INIT_PENDING, &con->flags);
 492                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 493                                 queue_work(send_workqueue, &new_con->swork);
 494                         }
 495                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 496                                 queue_work(recv_workqueue, &new_con->rwork);
 497                 }
 498                 break;
 499
 500                 case SCTP_COMM_LOST:
 501                 case SCTP_SHUTDOWN_COMP:
 502                 {
 503                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 504                         if (con) {
 505                                 con->sctp_assoc = 0;
 506                         }
 507                 }
 508                 break;
 509
 510                 /* We don't know which INIT failed, so clear the PENDING flags
 511                  * on them all.  if assoc_id is zero then it will then try
 512                  * again */
 513
 514                 case SCTP_CANT_STR_ASSOC:
 515                 {
 516                         log_print("Can't start SCTP association - retrying");
 517                         sctp_init_failed();
 518                 }
 519                 break;
 520
 521                 default:
 522                         log_print("unexpected SCTP assoc change id=%d state=%d",
 523                                   (int)sn->sn_assoc_change.sac_assoc_id,
 524                                   sn->sn_assoc_change.sac_state);
 525                 }
 526         }
 527 }
 528
 529 /* Data received from remote end */
 530 static int receive_from_sock(struct connection *con)
 531 {
 532         int ret = 0;
 533         struct msghdr msg = {};
 534         struct kvec iov[2];
 535         unsigned len;
 536         int r;
 537         int call_again_soon = 0;
 538         int nvec;
 539         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 540
 541         mutex_lock(&con->sock_mutex);
 542
 543         if (con->sock == NULL) {
 544                 ret = -EAGAIN;
 545                 goto out_close;
 546         }
 547
 548         if (con->rx_page == NULL) {
 549                 /*
 550                  * This doesn't need to be atomic, but I think it should
 551                  * improve performance if it is.
 552                  */
 553                 con->rx_page = alloc_page(GFP_ATOMIC);
 554                 if (con->rx_page == NULL)
 555                         goto out_resched;
 556                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 557         }
 558
 559         /* Only SCTP needs these really */
 560         memset(&incmsg, 0, sizeof(incmsg));
 561         msg.msg_control = incmsg;
 562         msg.msg_controllen = sizeof(incmsg);
 563
 564         /*
 565          * iov[0] is the bit of the circular buffer between the current end
 566          * point (cb.base + cb.len) and the end of the buffer.
 567          */
 568         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 569         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 570         iov[1].iov_len = 0;
 571         nvec = 1;
 572
 573         /*
 574          * iov[1] is the bit of the circular buffer between the start of the
 575          * buffer and the start of the currently used section (cb.base)
 576          */
 577         if (cbuf_data(&con->cb) >= con->cb.base) {
 578                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 579                 iov[1].iov_len = con->cb.base;
 580                 iov[1].iov_base = page_address(con->rx_page);
 581                 nvec = 2;
 582         }
 583         len = iov[0].iov_len + iov[1].iov_len;
 584
 585         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 586                                MSG_DONTWAIT | MSG_NOSIGNAL);
 587         if (ret <= 0)
 588                 goto out_close;
 589
 590         /* Process SCTP notifications */
 591         if (msg.msg_flags & MSG_NOTIFICATION) {
 592                 msg.msg_control = incmsg;
 593                 msg.msg_controllen = sizeof(incmsg);
 594
 595                 process_sctp_notification(con, &msg,
 596                                 page_address(con->rx_page) + con->cb.base);
 597                 mutex_unlock(&con->sock_mutex);
 598                 return 0;
 599         }
 600         BUG_ON(con->nodeid == 0);
 601
 602         if (ret == len)
 603                 call_again_soon = 1;
 604         cbuf_add(&con->cb, ret);
 605         ret = dlm_process_incoming_buffer(con->nodeid,
 606                                           page_address(con->rx_page),
 607                                           con->cb.base, con->cb.len,
 608                                           PAGE_CACHE_SIZE);
 609         if (ret == -EBADMSG) {
 610                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 611                           "iov_len=%u, iov_base[0]=%p, read=%d",
 612                           page_address(con->rx_page), con->cb.base, con->cb.len,
 613                           len, iov[0].iov_base, r);
 614         }
 615         if (ret < 0)
 616                 goto out_close;
 617         cbuf_eat(&con->cb, ret);
 618
 619         if (cbuf_empty(&con->cb) && !call_again_soon) {
 620                 __free_page(con->rx_page);
 621                 con->rx_page = NULL;
 622         }
 623
 624         if (call_again_soon)
 625                 goto out_resched;
 626         mutex_unlock(&con->sock_mutex);
 627         return 0;
 628
 629 out_resched:
 630         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 631                 queue_work(recv_workqueue, &con->rwork);
 632         mutex_unlock(&con->sock_mutex);
 633         return -EAGAIN;
 634
 635 out_close:
 636         mutex_unlock(&con->sock_mutex);
 637         if (ret != -EAGAIN) {
 638                 close_connection(con, false);
 639                 /* Reconnect when there is something to send */
 640         }
 641         /* Don't return success if we really got EOF */
 642         if (ret == 0)
 643                 ret = -EAGAIN;
 644
 645         return ret;
 646 }
 647
 648 /* Listening socket is busy, accept a connection */
 649 static int tcp_accept_from_sock(struct connection *con)
 650 {
 651         int result;
 652         struct sockaddr_storage peeraddr;
 653         struct socket *newsock;
 654         int len;
 655         int nodeid;
 656         struct connection *newcon;
 657         struct connection *addcon;
 658
 659         memset(&peeraddr, 0, sizeof(peeraddr));
 660         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 661                                   IPPROTO_TCP, &newsock);
 662         if (result < 0)
 663                 return -ENOMEM;
 664
 665         mutex_lock_nested(&con->sock_mutex, 0);
 666
 667         result = -ENOTCONN;
 668         if (con->sock == NULL)
 669                 goto accept_err;
 670
 671         newsock->type = con->sock->type;
 672         newsock->ops = con->sock->ops;
 673
 674         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 675         if (result < 0)
 676                 goto accept_err;
 677
 678         /* Get the connected socket's peer */
 679         memset(&peeraddr, 0, sizeof(peeraddr));
 680         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 681                                   &len, 2)) {
 682                 result = -ECONNABORTED;
 683                 goto accept_err;
 684         }
 685
 686         /* Get the new node's NODEID */
 687         make_sockaddr(&peeraddr, 0, &len);
 688         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 689                 log_print("connect from non cluster node");
 690                 sock_release(newsock);
 691                 mutex_unlock(&con->sock_mutex);
 692                 return -1;
 693         }
 694
 695         log_print("got connection from %d", nodeid);
 696
 697         /*  Check to see if we already have a connection to this node. This
 698          *  could happen if the two nodes initiate a connection at roughly
 699          *  the same time and the connections cross on the wire.
 700          *  In this case we store the incoming one in "othercon"
 701          */
 702         newcon = nodeid2con(nodeid, GFP_KERNEL);
 703         if (!newcon) {
 704                 result = -ENOMEM;
 705                 goto accept_err;
 706         }
 707         mutex_lock_nested(&newcon->sock_mutex, 1);
 708         if (newcon->sock) {
 709                 struct connection *othercon = newcon->othercon;
 710
 711                 if (!othercon) {
 712                         othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
 713                         if (!othercon) {
 714                                 log_print("failed to allocate incoming socket");
 715                                 mutex_unlock(&newcon->sock_mutex);
 716                                 result = -ENOMEM;
 717                                 goto accept_err;
 718                         }
 719                         othercon->nodeid = nodeid;
 720                         othercon->rx_action = receive_from_sock;
 721                         mutex_init(&othercon->sock_mutex);
 722                         INIT_WORK(&othercon->swork, process_send_sockets);
 723                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 724                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 725                 }
 726                 if (!othercon->sock) {
 727                         newcon->othercon = othercon;
 728                         othercon->sock = newsock;
 729                         newsock->sk->sk_user_data = othercon;
 730                         add_sock(newsock, othercon);
 731                         addcon = othercon;
 732                 }
 733                 else {
 734                         printk("Extra connection from node %d attempted\n", nodeid);
 735                         result = -EAGAIN;
 736                         mutex_unlock(&newcon->sock_mutex);
 737                         goto accept_err;
 738                 }
 739         }
 740         else {
 741                 newsock->sk->sk_user_data = newcon;
 742                 newcon->rx_action = receive_from_sock;
 743                 add_sock(newsock, newcon);
 744                 addcon = newcon;
 745         }
 746
 747         mutex_unlock(&newcon->sock_mutex);
 748
 749         /*
 750          * Add it to the active queue in case we got data
 751          * beween processing the accept adding the socket
 752          * to the read_sockets list
 753          */
 754         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 755                 queue_work(recv_workqueue, &addcon->rwork);
 756         mutex_unlock(&con->sock_mutex);
 757
 758         return 0;
 759
 760 accept_err:
 761         mutex_unlock(&con->sock_mutex);
 762         sock_release(newsock);
 763
 764         if (result != -EAGAIN)
 765                 log_print("error accepting connection from node: %d", result);
 766         return result;
 767 }
 768
 769 static void free_entry(struct writequeue_entry *e)
 770 {
 771         __free_page(e->page);
 772         kfree(e);
 773 }
 774
 775 /* Initiate an SCTP association.
 776    This is a special case of send_to_sock() in that we don't yet have a
 777    peeled-off socket for this association, so we use the listening socket
 778    and add the primary IP address of the remote node.
 779  */
 780 static void sctp_init_assoc(struct connection *con)
 781 {
 782         struct sockaddr_storage rem_addr;
 783         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 784         struct msghdr outmessage;
 785         struct cmsghdr *cmsg;
 786         struct sctp_sndrcvinfo *sinfo;
 787         struct connection *base_con;
 788         struct writequeue_entry *e;
 789         int len, offset;
 790         int ret;
 791         int addrlen;
 792         struct kvec iov[1];
 793
 794         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 795                 return;
 796
 797         if (con->retries++ > MAX_CONNECT_RETRIES)
 798                 return;
 799
 800         log_print("Initiating association with node %d", con->nodeid);
 801
 802         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 803                 log_print("no address for nodeid %d", con->nodeid);
 804                 return;
 805         }
 806         base_con = nodeid2con(0, 0);
 807         BUG_ON(base_con == NULL);
 808
 809         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 810
 811         outmessage.msg_name = &rem_addr;
 812         outmessage.msg_namelen = addrlen;
 813         outmessage.msg_control = outcmsg;
 814         outmessage.msg_controllen = sizeof(outcmsg);
 815         outmessage.msg_flags = MSG_EOR;
 816
 817         spin_lock(&con->writequeue_lock);
 818         e = list_entry(con->writequeue.next, struct writequeue_entry,
 819                        list);
 820
 821         BUG_ON((struct list_head *) e == &con->writequeue);
 822
 823         len = e->len;
 824         offset = e->offset;
 825         spin_unlock(&con->writequeue_lock);
 826         kmap(e->page);
 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                         kunmap(e->page);
 858                         free_entry(e);
 859                 }
 860                 spin_unlock(&con->writequeue_lock);
 861         }
 862 }
 863
 864 /* Connect a new socket to its peer */
 865 static void tcp_connect_to_sock(struct connection *con)
 866 {
 867         int result = -EHOSTUNREACH;
 868         struct sockaddr_storage saddr, src_addr;
 869         int addr_len;
 870         struct socket *sock;
 871
 872         if (con->nodeid == 0) {
 873                 log_print("attempt to connect sock 0 foiled");
 874                 return;
 875         }
 876
 877         mutex_lock(&con->sock_mutex);
 878         if (con->retries++ > MAX_CONNECT_RETRIES)
 879                 goto out;
 880
 881         /* Some odd races can cause double-connects, ignore them */
 882         if (con->sock) {
 883                 result = 0;
 884                 goto out;
 885         }
 886
 887         /* Create a socket to communicate with */
 888         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 889                                   IPPROTO_TCP, &sock);
 890         if (result < 0)
 891                 goto out_err;
 892
 893         memset(&saddr, 0, sizeof(saddr));
 894         if (dlm_nodeid_to_addr(con->nodeid, &saddr)) {
 895                 sock_release(sock);
 896                 goto out_err;
 897         }
 898
 899         sock->sk->sk_user_data = con;
 900         con->rx_action = receive_from_sock;
 901         con->connect_action = tcp_connect_to_sock;
 902         add_sock(sock, con);
 903
 904         /* Bind to our cluster-known address connecting to avoid
 905            routing problems */
 906         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 907         make_sockaddr(&src_addr, 0, &addr_len);
 908         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 909                                  addr_len);
 910         if (result < 0) {
 911                 log_print("could not bind for connect: %d", result);
 912                 /* This *may* not indicate a critical error */
 913         }
 914
 915         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 916
 917         log_print("connecting to %d", con->nodeid);
 918         result =
 919                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 920                                    O_NONBLOCK);
 921         if (result == -EINPROGRESS)
 922                 result = 0;
 923         if (result == 0)
 924                 goto out;
 925
 926 out_err:
 927         if (con->sock) {
 928                 sock_release(con->sock);
 929                 con->sock = NULL;
 930         }
 931         /*
 932          * Some errors are fatal and this list might need adjusting. For other
 933          * errors we try again until the max number of retries is reached.
 934          */
 935         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 936             result != -ENETDOWN && result != -EINVAL
 937             && result != -EPROTONOSUPPORT) {
 938                 lowcomms_connect_sock(con);
 939                 result = 0;
 940         }
 941 out:
 942         mutex_unlock(&con->sock_mutex);
 943         return;
 944 }
 945
 946 static struct socket *tcp_create_listen_sock(struct connection *con,
 947                                              struct sockaddr_storage *saddr)
 948 {
 949         struct socket *sock = NULL;
 950         int result = 0;
 951         int one = 1;
 952         int addr_len;
 953
 954         if (dlm_local_addr[0]->ss_family == AF_INET)
 955                 addr_len = sizeof(struct sockaddr_in);
 956         else
 957                 addr_len = sizeof(struct sockaddr_in6);
 958
 959         /* Create a socket to communicate with */
 960         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 961                                   IPPROTO_TCP, &sock);
 962         if (result < 0) {
 963                 log_print("Can't create listening comms socket");
 964                 goto create_out;
 965         }
 966
 967         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
 968                                    (char *)&one, sizeof(one));
 969
 970         if (result < 0) {
 971                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
 972         }
 973         sock->sk->sk_user_data = con;
 974         con->rx_action = tcp_accept_from_sock;
 975         con->connect_action = tcp_connect_to_sock;
 976         con->sock = sock;
 977
 978         /* Bind to our port */
 979         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
 980         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
 981         if (result < 0) {
 982                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
 983                 sock_release(sock);
 984                 sock = NULL;
 985                 con->sock = NULL;
 986                 goto create_out;
 987         }
 988         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 989                                  (char *)&one, sizeof(one));
 990         if (result < 0) {
 991                 log_print("Set keepalive failed: %d", result);
 992         }
 993
 994         result = sock->ops->listen(sock, 5);
 995         if (result < 0) {
 996                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
 997                 sock_release(sock);
 998                 sock = NULL;
 999                 goto create_out;
1000         }
1001
1002 create_out:
1003         return sock;
1004 }
1005
1006 /* Get local addresses */
1007 static void init_local(void)
1008 {
1009         struct sockaddr_storage sas, *addr;
1010         int i;
1011
1012         dlm_local_count = 0;
1013         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1014                 if (dlm_our_addr(&sas, i))
1015                         break;
1016
1017                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1018                 if (!addr)
1019                         break;
1020                 memcpy(addr, &sas, sizeof(*addr));
1021                 dlm_local_addr[dlm_local_count++] = addr;
1022         }
1023 }
1024
1025 /* Bind to an IP address. SCTP allows multiple address so it can do
1026    multi-homing */
1027 static int add_sctp_bind_addr(struct connection *sctp_con,
1028                               struct sockaddr_storage *addr,
1029                               int addr_len, int num)
1030 {
1031         int result = 0;
1032
1033         if (num == 1)
1034                 result = kernel_bind(sctp_con->sock,
1035                                      (struct sockaddr *) addr,
1036                                      addr_len);
1037         else
1038                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1039                                            SCTP_SOCKOPT_BINDX_ADD,
1040                                            (char *)addr, addr_len);
1041
1042         if (result < 0)
1043                 log_print("Can't bind to port %d addr number %d",
1044                           dlm_config.ci_tcp_port, num);
1045
1046         return result;
1047 }
1048
1049 /* Initialise SCTP socket and bind to all interfaces */
1050 static int sctp_listen_for_all(void)
1051 {
1052         struct socket *sock = NULL;
1053         struct sockaddr_storage localaddr;
1054         struct sctp_event_subscribe subscribe;
1055         int result = -EINVAL, num = 1, i, addr_len;
1056         struct connection *con = nodeid2con(0, GFP_KERNEL);
1057         int bufsize = NEEDED_RMEM;
1058
1059         if (!con)
1060                 return -ENOMEM;
1061
1062         log_print("Using SCTP for communications");
1063
1064         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1065                                   IPPROTO_SCTP, &sock);
1066         if (result < 0) {
1067                 log_print("Can't create comms socket, check SCTP is loaded");
1068                 goto out;
1069         }
1070
1071         /* Listen for events */
1072         memset(&subscribe, 0, sizeof(subscribe));
1073         subscribe.sctp_data_io_event = 1;
1074         subscribe.sctp_association_event = 1;
1075         subscribe.sctp_send_failure_event = 1;
1076         subscribe.sctp_shutdown_event = 1;
1077         subscribe.sctp_partial_delivery_event = 1;
1078
1079         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1080                                  (char *)&bufsize, sizeof(bufsize));
1081         if (result)
1082                 log_print("Error increasing buffer space on socket %d", result);
1083
1084         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1085                                    (char *)&subscribe, sizeof(subscribe));
1086         if (result < 0) {
1087                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1088                           result);
1089                 goto create_delsock;
1090         }
1091
1092         /* Init con struct */
1093         sock->sk->sk_user_data = con;
1094         con->sock = sock;
1095         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1096         con->rx_action = receive_from_sock;
1097         con->connect_action = sctp_init_assoc;
1098
1099         /* Bind to all interfaces. */
1100         for (i = 0; i < dlm_local_count; i++) {
1101                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1102                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1103
1104                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1105                 if (result)
1106                         goto create_delsock;
1107                 ++num;
1108         }
1109
1110         result = sock->ops->listen(sock, 5);
1111         if (result < 0) {
1112                 log_print("Can't set socket listening");
1113                 goto create_delsock;
1114         }
1115
1116         return 0;
1117
1118 create_delsock:
1119         sock_release(sock);
1120         con->sock = NULL;
1121 out:
1122         return result;
1123 }
1124
1125 static int tcp_listen_for_all(void)
1126 {
1127         struct socket *sock = NULL;
1128         struct connection *con = nodeid2con(0, GFP_KERNEL);
1129         int result = -EINVAL;
1130
1131         if (!con)
1132                 return -ENOMEM;
1133
1134         /* We don't support multi-homed hosts */
1135         if (dlm_local_addr[1] != NULL) {
1136                 log_print("TCP protocol can't handle multi-homed hosts, "
1137                           "try SCTP");
1138                 return -EINVAL;
1139         }
1140
1141         log_print("Using TCP for communications");
1142
1143         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1144         if (sock) {
1145                 add_sock(sock, con);
1146                 result = 0;
1147         }
1148         else {
1149                 result = -EADDRINUSE;
1150         }
1151
1152         return result;
1153 }
1154
1155
1156
1157 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1158                                                      gfp_t allocation)
1159 {
1160         struct writequeue_entry *entry;
1161
1162         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1163         if (!entry)
1164                 return NULL;
1165
1166         entry->page = alloc_page(allocation);
1167         if (!entry->page) {
1168                 kfree(entry);
1169                 return NULL;
1170         }
1171
1172         entry->offset = 0;
1173         entry->len = 0;
1174         entry->end = 0;
1175         entry->users = 0;
1176         entry->con = con;
1177
1178         return entry;
1179 }
1180
1181 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1182 {
1183         struct connection *con;
1184         struct writequeue_entry *e;
1185         int offset = 0;
1186         int users = 0;
1187
1188         con = nodeid2con(nodeid, allocation);
1189         if (!con)
1190                 return NULL;
1191
1192         spin_lock(&con->writequeue_lock);
1193         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1194         if ((&e->list == &con->writequeue) ||
1195             (PAGE_CACHE_SIZE - e->end < len)) {
1196                 e = NULL;
1197         } else {
1198                 offset = e->end;
1199                 e->end += len;
1200                 users = e->users++;
1201         }
1202         spin_unlock(&con->writequeue_lock);
1203
1204         if (e) {
1205         got_one:
1206                 if (users == 0)
1207                         kmap(e->page);
1208                 *ppc = page_address(e->page) + offset;
1209                 return e;
1210         }
1211
1212         e = new_writequeue_entry(con, allocation);
1213         if (e) {
1214                 spin_lock(&con->writequeue_lock);
1215                 offset = e->end;
1216                 e->end += len;
1217                 users = e->users++;
1218                 list_add_tail(&e->list, &con->writequeue);
1219                 spin_unlock(&con->writequeue_lock);
1220                 goto got_one;
1221         }
1222         return NULL;
1223 }
1224
1225 void dlm_lowcomms_commit_buffer(void *mh)
1226 {
1227         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1228         struct connection *con = e->con;
1229         int users;
1230
1231         spin_lock(&con->writequeue_lock);
1232         users = --e->users;
1233         if (users)
1234                 goto out;
1235         e->len = e->end - e->offset;
1236         kunmap(e->page);
1237         spin_unlock(&con->writequeue_lock);
1238
1239         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1240                 queue_work(send_workqueue, &con->swork);
1241         }
1242         return;
1243
1244 out:
1245         spin_unlock(&con->writequeue_lock);
1246         return;
1247 }
1248
1249 /* Send a message */
1250 static void send_to_sock(struct connection *con)
1251 {
1252         int ret = 0;
1253         ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1254         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1255         struct writequeue_entry *e;
1256         int len, offset;
1257
1258         mutex_lock(&con->sock_mutex);
1259         if (con->sock == NULL)
1260                 goto out_connect;
1261
1262         sendpage = con->sock->ops->sendpage;
1263
1264         spin_lock(&con->writequeue_lock);
1265         for (;;) {
1266                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1267                                list);
1268                 if ((struct list_head *) e == &con->writequeue)
1269                         break;
1270
1271                 len = e->len;
1272                 offset = e->offset;
1273                 BUG_ON(len == 0 && e->users == 0);
1274                 spin_unlock(&con->writequeue_lock);
1275                 kmap(e->page);
1276
1277                 ret = 0;
1278                 if (len) {
1279                         ret = sendpage(con->sock, e->page, offset, len,
1280                                        msg_flags);
1281                         if (ret == -EAGAIN || ret == 0) {
1282                                 cond_resched();
1283                                 goto out;
1284                         }
1285                         if (ret <= 0)
1286                                 goto send_error;
1287                 }
1288                         /* Don't starve people filling buffers */
1289                         cond_resched();
1290
1291                 spin_lock(&con->writequeue_lock);
1292                 e->offset += ret;
1293                 e->len -= ret;
1294
1295                 if (e->len == 0 && e->users == 0) {
1296                         list_del(&e->list);
1297                         kunmap(e->page);
1298                         free_entry(e);
1299                         continue;
1300                 }
1301         }
1302         spin_unlock(&con->writequeue_lock);
1303 out:
1304         mutex_unlock(&con->sock_mutex);
1305         return;
1306
1307 send_error:
1308         mutex_unlock(&con->sock_mutex);
1309         close_connection(con, false);
1310         lowcomms_connect_sock(con);
1311         return;
1312
1313 out_connect:
1314         mutex_unlock(&con->sock_mutex);
1315         if (!test_bit(CF_INIT_PENDING, &con->flags))
1316                 lowcomms_connect_sock(con);
1317         return;
1318 }
1319
1320 static void clean_one_writequeue(struct connection *con)
1321 {
1322         struct list_head *list;
1323         struct list_head *temp;
1324
1325         spin_lock(&con->writequeue_lock);
1326         list_for_each_safe(list, temp, &con->writequeue) {
1327                 struct writequeue_entry *e =
1328                         list_entry(list, struct writequeue_entry, list);
1329                 list_del(&e->list);
1330                 free_entry(e);
1331         }
1332         spin_unlock(&con->writequeue_lock);
1333 }
1334
1335 /* Called from recovery when it knows that a node has
1336    left the cluster */
1337 int dlm_lowcomms_close(int nodeid)
1338 {
1339         struct connection *con;
1340
1341         log_print("closing connection to node %d", nodeid);
1342         con = nodeid2con(nodeid, 0);
1343         if (con) {
1344                 clean_one_writequeue(con);
1345                 close_connection(con, true);
1346         }
1347         return 0;
1348 }
1349
1350 /* Receive workqueue function */
1351 static void process_recv_sockets(struct work_struct *work)
1352 {
1353         struct connection *con = container_of(work, struct connection, rwork);
1354         int err;
1355
1356         clear_bit(CF_READ_PENDING, &con->flags);
1357         do {
1358                 err = con->rx_action(con);
1359         } while (!err);
1360 }
1361
1362 /* Send workqueue function */
1363 static void process_send_sockets(struct work_struct *work)
1364 {
1365         struct connection *con = container_of(work, struct connection, swork);
1366
1367         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1368                 con->connect_action(con);
1369         }
1370         clear_bit(CF_WRITE_PENDING, &con->flags);
1371         send_to_sock(con);
1372 }
1373
1374
1375 /* Discard all entries on the write queues */
1376 static void clean_writequeues(void)
1377 {
1378         int nodeid;
1379
1380         for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1381                 struct connection *con = __nodeid2con(nodeid, 0);
1382
1383                 if (con)
1384                         clean_one_writequeue(con);
1385         }
1386 }
1387
1388 static void work_stop(void)
1389 {
1390         destroy_workqueue(recv_workqueue);
1391         destroy_workqueue(send_workqueue);
1392 }
1393
1394 static int work_start(void)
1395 {
1396         int error;
1397         recv_workqueue = create_workqueue("dlm_recv");
1398         error = IS_ERR(recv_workqueue);
1399         if (error) {
1400                 log_print("can't start dlm_recv %d", error);
1401                 return error;
1402         }
1403
1404         send_workqueue = create_singlethread_workqueue("dlm_send");
1405         error = IS_ERR(send_workqueue);
1406         if (error) {
1407                 log_print("can't start dlm_send %d", error);
1408                 destroy_workqueue(recv_workqueue);
1409                 return error;
1410         }
1411
1412         return 0;
1413 }
1414
1415 void dlm_lowcomms_stop(void)
1416 {
1417         int i;
1418         struct connection *con;
1419
1420         /* Set all the flags to prevent any
1421            socket activity.
1422         */
1423         mutex_lock(&connections_lock);
1424         for (i = 0; i <= max_nodeid; i++) {
1425                 con = __nodeid2con(i, 0);
1426                 if (con) {
1427                         con->flags |= 0x0F;
1428                         if (con->sock)
1429                                 con->sock->sk->sk_user_data = NULL;
1430                 }
1431         }
1432         mutex_unlock(&connections_lock);
1433
1434         work_stop();
1435
1436         mutex_lock(&connections_lock);
1437         clean_writequeues();
1438
1439         for (i = 0; i <= max_nodeid; i++) {
1440                 con = __nodeid2con(i, 0);
1441                 if (con) {
1442                         close_connection(con, true);
1443                         if (con->othercon)
1444                                 kmem_cache_free(con_cache, con->othercon);
1445                         kmem_cache_free(con_cache, con);
1446                 }
1447         }
1448         max_nodeid = 0;
1449         mutex_unlock(&connections_lock);
1450         kmem_cache_destroy(con_cache);
1451         idr_init(&connections_idr);
1452 }
1453
1454 int dlm_lowcomms_start(void)
1455 {
1456         int error = -EINVAL;
1457         struct connection *con;
1458
1459         init_local();
1460         if (!dlm_local_count) {
1461                 error = -ENOTCONN;
1462                 log_print("no local IP address has been set");
1463                 goto out;
1464         }
1465
1466         error = -ENOMEM;
1467         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1468                                       __alignof__(struct connection), 0,
1469                                       NULL);
1470         if (!con_cache)
1471                 goto out;
1472
1473         /* Start listening */
1474         if (dlm_config.ci_protocol == 0)
1475                 error = tcp_listen_for_all();
1476         else
1477                 error = sctp_listen_for_all();
1478         if (error)
1479                 goto fail_unlisten;
1480
1481         error = work_start();
1482         if (error)
1483                 goto fail_unlisten;
1484
1485         return 0;
1486
1487 fail_unlisten:
1488         con = nodeid2con(0,0);
1489         if (con) {
1490                 close_connection(con, false);
1491                 kmem_cache_free(con_cache, con);
1492         }
1493         kmem_cache_destroy(con_cache);
1494
1495 out:
1496         return error;
1497 }