fs/dlm/lowcomms.c

   1 /******************************************************************************
   2 *******************************************************************************
   3 **
   4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   5 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
   6 **
   7 **  This copyrighted material is made available to anyone wishing to use,
   8 **  modify, copy, or redistribute it subject to the terms and conditions
   9 **  of the GNU General Public License v.2.
  10 **
  11 *******************************************************************************
  12 ******************************************************************************/
  13
  14 /*
  15  * lowcomms.c
  16  *
  17  * This is the "low-level" comms layer.
  18  *
  19  * It is responsible for sending/receiving messages
  20  * from other nodes in the cluster.
  21  *
  22  * Cluster nodes are referred to by their nodeids. nodeids are
  23  * simply 32 bit numbers to the locking module - if they need to
  24  * be expanded for the cluster infrastructure then that is its
  25  * responsibility. It is this layer's
  26  * responsibility to resolve these into IP address or
  27  * whatever it needs for inter-node communication.
  28  *
  29  * The comms level is two kernel threads that deal mainly with
  30  * the receiving of messages from other nodes and passing them
  31  * up to the mid-level comms layer (which understands the
  32  * message format) for execution by the locking core, and
  33  * a send thread which does all the setting up of connections
  34  * to remote nodes and the sending of data. Threads are not allowed
  35  * to send their own data because it may cause them to wait in times
  36  * of high load. Also, this way, the sending thread can collect together
  37  * messages bound for one node and send them in one block.
  38  *
  39  * lowcomms will choose to use either TCP or SCTP as its transport layer
  40  * depending on the configuration variable 'protocol'. This should be set
  41  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  42  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  43  * for the DLM to function.
  44  *
  45  */
  46
  47 #include <asm/ioctls.h>
  48 #include <net/sock.h>
  49 #include <net/tcp.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/file.h>
  52 #include <linux/mutex.h>
  53 #include <linux/sctp.h>
  54 #include <linux/slab.h>
  55 #include <net/sctp/user.h>
  56 #include <net/ipv6.h>
  57
  58 #include "dlm_internal.h"
  59 #include "lowcomms.h"
  60 #include "midcomms.h"
  61 #include "config.h"
  62
  63 #define NEEDED_RMEM (4*1024*1024)
  64 #define CONN_HASH_SIZE 32
  65
  66 /* Number of messages to send before rescheduling */
  67 #define MAX_SEND_MSG_COUNT 25
  68
  69 struct cbuf {
  70         unsigned int base;
  71         unsigned int len;
  72         unsigned int mask;
  73 };
  74
  75 static void cbuf_add(struct cbuf *cb, int n)
  76 {
  77         cb->len += n;
  78 }
  79
  80 static int cbuf_data(struct cbuf *cb)
  81 {
  82         return ((cb->base + cb->len) & cb->mask);
  83 }
  84
  85 static void cbuf_init(struct cbuf *cb, int size)
  86 {
  87         cb->base = cb->len = 0;
  88         cb->mask = size-1;
  89 }
  90
  91 static void cbuf_eat(struct cbuf *cb, int n)
  92 {
  93         cb->len  -= n;
  94         cb->base += n;
  95         cb->base &= cb->mask;
  96 }
  97
  98 static bool cbuf_empty(struct cbuf *cb)
  99 {
 100         return cb->len == 0;
 101 }
 102
 103 struct connection {
 104         struct socket *sock;    /* NULL if not connected */
 105         uint32_t nodeid;        /* So we know who we are in the list */
 106         struct mutex sock_mutex;
 107         unsigned long flags;
 108 #define CF_READ_PENDING 1
 109 #define CF_WRITE_PENDING 2
 110 #define CF_CONNECT_PENDING 3
 111 #define CF_INIT_PENDING 4
 112 #define CF_IS_OTHERCON 5
 113 #define CF_CLOSE 6
 114 #define CF_APP_LIMITED 7
 115         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 116         spinlock_t writequeue_lock;
 117         int (*rx_action) (struct connection *); /* What to do when active */
 118         void (*connect_action) (struct connection *);   /* What to do to connect */
 119         struct page *rx_page;
 120         struct cbuf cb;
 121         int retries;
 122 #define MAX_CONNECT_RETRIES 3
 123         int sctp_assoc;
 124         struct hlist_node list;
 125         struct connection *othercon;
 126         struct work_struct rwork; /* Receive workqueue */
 127         struct work_struct swork; /* Send workqueue */
 128 };
 129 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 130
 131 /* An entry waiting to be sent */
 132 struct writequeue_entry {
 133         struct list_head list;
 134         struct page *page;
 135         int offset;
 136         int len;
 137         int end;
 138         int users;
 139         struct connection *con;
 140 };
 141
 142 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 143 static int dlm_local_count;
 144
 145 /* Work queues */
 146 static struct workqueue_struct *recv_workqueue;
 147 static struct workqueue_struct *send_workqueue;
 148
 149 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 150 static DEFINE_MUTEX(connections_lock);
 151 static struct kmem_cache *con_cache;
 152
 153 static void process_recv_sockets(struct work_struct *work);
 154 static void process_send_sockets(struct work_struct *work);
 155
 156
 157 /* This is deliberately very simple because most clusters have simple
 158    sequential nodeids, so we should be able to go straight to a connection
 159    struct in the array */
 160 static inline int nodeid_hash(int nodeid)
 161 {
 162         return nodeid & (CONN_HASH_SIZE-1);
 163 }
 164
 165 static struct connection *__find_con(int nodeid)
 166 {
 167         int r;
 168         struct hlist_node *h;
 169         struct connection *con;
 170
 171         r = nodeid_hash(nodeid);
 172
 173         hlist_for_each_entry(con, h, &connection_hash[r], list) {
 174                 if (con->nodeid == nodeid)
 175                         return con;
 176         }
 177         return NULL;
 178 }
 179
 180 /*
 181  * If 'allocation' is zero then we don't attempt to create a new
 182  * connection structure for this node.
 183  */
 184 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 185 {
 186         struct connection *con = NULL;
 187         int r;
 188
 189         con = __find_con(nodeid);
 190         if (con || !alloc)
 191                 return con;
 192
 193         con = kmem_cache_zalloc(con_cache, alloc);
 194         if (!con)
 195                 return NULL;
 196
 197         r = nodeid_hash(nodeid);
 198         hlist_add_head(&con->list, &connection_hash[r]);
 199
 200         con->nodeid = nodeid;
 201         mutex_init(&con->sock_mutex);
 202         INIT_LIST_HEAD(&con->writequeue);
 203         spin_lock_init(&con->writequeue_lock);
 204         INIT_WORK(&con->swork, process_send_sockets);
 205         INIT_WORK(&con->rwork, process_recv_sockets);
 206
 207         /* Setup action pointers for child sockets */
 208         if (con->nodeid) {
 209                 struct connection *zerocon = __find_con(0);
 210
 211                 con->connect_action = zerocon->connect_action;
 212                 if (!con->rx_action)
 213                         con->rx_action = zerocon->rx_action;
 214         }
 215
 216         return con;
 217 }
 218
 219 /* Loop round all connections */
 220 static void foreach_conn(void (*conn_func)(struct connection *c))
 221 {
 222         int i;
 223         struct hlist_node *h, *n;
 224         struct connection *con;
 225
 226         for (i = 0; i < CONN_HASH_SIZE; i++) {
 227                 hlist_for_each_entry_safe(con, h, n, &connection_hash[i], list){
 228                         conn_func(con);
 229                 }
 230         }
 231 }
 232
 233 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 234 {
 235         struct connection *con;
 236
 237         mutex_lock(&connections_lock);
 238         con = __nodeid2con(nodeid, allocation);
 239         mutex_unlock(&connections_lock);
 240
 241         return con;
 242 }
 243
 244 /* This is a bit drastic, but only called when things go wrong */
 245 static struct connection *assoc2con(int assoc_id)
 246 {
 247         int i;
 248         struct hlist_node *h;
 249         struct connection *con;
 250
 251         mutex_lock(&connections_lock);
 252
 253         for (i = 0 ; i < CONN_HASH_SIZE; i++) {
 254                 hlist_for_each_entry(con, h, &connection_hash[i], list) {
 255                         if (con->sctp_assoc == assoc_id) {
 256                                 mutex_unlock(&connections_lock);
 257                                 return con;
 258                         }
 259                 }
 260         }
 261         mutex_unlock(&connections_lock);
 262         return NULL;
 263 }
 264
 265 static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr)
 266 {
 267         struct sockaddr_storage addr;
 268         int error;
 269
 270         if (!dlm_local_count)
 271                 return -1;
 272
 273         error = dlm_nodeid_to_addr(nodeid, &addr);
 274         if (error)
 275                 return error;
 276
 277         if (dlm_local_addr[0]->ss_family == AF_INET) {
 278                 struct sockaddr_in *in4  = (struct sockaddr_in *) &addr;
 279                 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr;
 280                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 281         } else {
 282                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &addr;
 283                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr;
 284                 ipv6_addr_copy(&ret6->sin6_addr, &in6->sin6_addr);
 285         }
 286
 287         return 0;
 288 }
 289
 290 /* Data available on socket or listen socket received a connect */
 291 static void lowcomms_data_ready(struct sock *sk, int count_unused)
 292 {
 293         struct connection *con = sock2con(sk);
 294         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 295                 queue_work(recv_workqueue, &con->rwork);
 296 }
 297
 298 static void lowcomms_write_space(struct sock *sk)
 299 {
 300         struct connection *con = sock2con(sk);
 301
 302         if (!con)
 303                 return;
 304
 305         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 306
 307         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 308                 con->sock->sk->sk_write_pending--;
 309                 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags);
 310         }
 311
 312         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 313                 queue_work(send_workqueue, &con->swork);
 314 }
 315
 316 static inline void lowcomms_connect_sock(struct connection *con)
 317 {
 318         if (test_bit(CF_CLOSE, &con->flags))
 319                 return;
 320         if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
 321                 queue_work(send_workqueue, &con->swork);
 322 }
 323
 324 static void lowcomms_state_change(struct sock *sk)
 325 {
 326         if (sk->sk_state == TCP_ESTABLISHED)
 327                 lowcomms_write_space(sk);
 328 }
 329
 330 int dlm_lowcomms_connect_node(int nodeid)
 331 {
 332         struct connection *con;
 333
 334         /* with sctp there's no connecting without sending */
 335         if (dlm_config.ci_protocol != 0)
 336                 return 0;
 337
 338         if (nodeid == dlm_our_nodeid())
 339                 return 0;
 340
 341         con = nodeid2con(nodeid, GFP_NOFS);
 342         if (!con)
 343                 return -ENOMEM;
 344         lowcomms_connect_sock(con);
 345         return 0;
 346 }
 347
 348 /* Make a socket active */
 349 static int add_sock(struct socket *sock, struct connection *con)
 350 {
 351         con->sock = sock;
 352
 353         /* Install a data_ready callback */
 354         con->sock->sk->sk_data_ready = lowcomms_data_ready;
 355         con->sock->sk->sk_write_space = lowcomms_write_space;
 356         con->sock->sk->sk_state_change = lowcomms_state_change;
 357         con->sock->sk->sk_user_data = con;
 358         con->sock->sk->sk_allocation = GFP_NOFS;
 359         return 0;
 360 }
 361
 362 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 363    length */
 364 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 365                           int *addr_len)
 366 {
 367         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 368         if (saddr->ss_family == AF_INET) {
 369                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 370                 in4_addr->sin_port = cpu_to_be16(port);
 371                 *addr_len = sizeof(struct sockaddr_in);
 372                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 373         } else {
 374                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 375                 in6_addr->sin6_port = cpu_to_be16(port);
 376                 *addr_len = sizeof(struct sockaddr_in6);
 377         }
 378         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 379 }
 380
 381 /* Close a remote connection and tidy up */
 382 static void close_connection(struct connection *con, bool and_other)
 383 {
 384         mutex_lock(&con->sock_mutex);
 385
 386         if (con->sock) {
 387                 sock_release(con->sock);
 388                 con->sock = NULL;
 389         }
 390         if (con->othercon && and_other) {
 391                 /* Will only re-enter once. */
 392                 close_connection(con->othercon, false);
 393         }
 394         if (con->rx_page) {
 395                 __free_page(con->rx_page);
 396                 con->rx_page = NULL;
 397         }
 398
 399         con->retries = 0;
 400         mutex_unlock(&con->sock_mutex);
 401 }
 402
 403 /* We only send shutdown messages to nodes that are not part of the cluster */
 404 static void sctp_send_shutdown(sctp_assoc_t associd)
 405 {
 406         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 407         struct msghdr outmessage;
 408         struct cmsghdr *cmsg;
 409         struct sctp_sndrcvinfo *sinfo;
 410         int ret;
 411         struct connection *con;
 412
 413         con = nodeid2con(0,0);
 414         BUG_ON(con == NULL);
 415
 416         outmessage.msg_name = NULL;
 417         outmessage.msg_namelen = 0;
 418         outmessage.msg_control = outcmsg;
 419         outmessage.msg_controllen = sizeof(outcmsg);
 420         outmessage.msg_flags = MSG_EOR;
 421
 422         cmsg = CMSG_FIRSTHDR(&outmessage);
 423         cmsg->cmsg_level = IPPROTO_SCTP;
 424         cmsg->cmsg_type = SCTP_SNDRCV;
 425         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 426         outmessage.msg_controllen = cmsg->cmsg_len;
 427         sinfo = CMSG_DATA(cmsg);
 428         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 429
 430         sinfo->sinfo_flags |= MSG_EOF;
 431         sinfo->sinfo_assoc_id = associd;
 432
 433         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 434
 435         if (ret != 0)
 436                 log_print("send EOF to node failed: %d", ret);
 437 }
 438
 439 static void sctp_init_failed_foreach(struct connection *con)
 440 {
 441         con->sctp_assoc = 0;
 442         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 443                 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags))
 444                         queue_work(send_workqueue, &con->swork);
 445         }
 446 }
 447
 448 /* INIT failed but we don't know which node...
 449    restart INIT on all pending nodes */
 450 static void sctp_init_failed(void)
 451 {
 452         mutex_lock(&connections_lock);
 453
 454         foreach_conn(sctp_init_failed_foreach);
 455
 456         mutex_unlock(&connections_lock);
 457 }
 458
 459 /* Something happened to an association */
 460 static void process_sctp_notification(struct connection *con,
 461                                       struct msghdr *msg, char *buf)
 462 {
 463         union sctp_notification *sn = (union sctp_notification *)buf;
 464
 465         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 466                 switch (sn->sn_assoc_change.sac_state) {
 467
 468                 case SCTP_COMM_UP:
 469                 case SCTP_RESTART:
 470                 {
 471                         /* Check that the new node is in the lockspace */
 472                         struct sctp_prim prim;
 473                         int nodeid;
 474                         int prim_len, ret;
 475                         int addr_len;
 476                         struct connection *new_con;
 477                         sctp_peeloff_arg_t parg;
 478                         int parglen = sizeof(parg);
 479                         int err;
 480
 481                         /*
 482                          * We get this before any data for an association.
 483                          * We verify that the node is in the cluster and
 484                          * then peel off a socket for it.
 485                          */
 486                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 487                                 log_print("COMM_UP for invalid assoc ID %d",
 488                                          (int)sn->sn_assoc_change.sac_assoc_id);
 489                                 sctp_init_failed();
 490                                 return;
 491                         }
 492                         memset(&prim, 0, sizeof(struct sctp_prim));
 493                         prim_len = sizeof(struct sctp_prim);
 494                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 495
 496                         ret = kernel_getsockopt(con->sock,
 497                                                 IPPROTO_SCTP,
 498                                                 SCTP_PRIMARY_ADDR,
 499                                                 (char*)&prim,
 500                                                 &prim_len);
 501                         if (ret < 0) {
 502                                 log_print("getsockopt/sctp_primary_addr on "
 503                                           "new assoc %d failed : %d",
 504                                           (int)sn->sn_assoc_change.sac_assoc_id,
 505                                           ret);
 506
 507                                 /* Retry INIT later */
 508                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 509                                 if (new_con)
 510                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 511                                 return;
 512                         }
 513                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 514                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 515                                 int i;
 516                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 517                                 log_print("reject connect from unknown addr");
 518                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 519                                         printk("%02x ", b[i]);
 520                                 printk("\n");
 521                                 sctp_send_shutdown(prim.ssp_assoc_id);
 522                                 return;
 523                         }
 524
 525                         new_con = nodeid2con(nodeid, GFP_NOFS);
 526                         if (!new_con)
 527                                 return;
 528
 529                         /* Peel off a new sock */
 530                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 531                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 532                                                 SCTP_SOCKOPT_PEELOFF,
 533                                                 (void *)&parg, &parglen);
 534                         if (ret < 0) {
 535                                 log_print("Can't peel off a socket for "
 536                                           "connection %d to node %d: err=%d",
 537                                           parg.associd, nodeid, ret);
 538                                 return;
 539                         }
 540                         new_con->sock = sockfd_lookup(parg.sd, &err);
 541                         if (!new_con->sock) {
 542                                 log_print("sockfd_lookup error %d", err);
 543                                 return;
 544                         }
 545                         add_sock(new_con->sock, new_con);
 546                         sockfd_put(new_con->sock);
 547
 548                         log_print("connecting to %d sctp association %d",
 549                                  nodeid, (int)sn->sn_assoc_change.sac_assoc_id);
 550
 551                         /* Send any pending writes */
 552                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 553                         clear_bit(CF_INIT_PENDING, &con->flags);
 554                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 555                                 queue_work(send_workqueue, &new_con->swork);
 556                         }
 557                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 558                                 queue_work(recv_workqueue, &new_con->rwork);
 559                 }
 560                 break;
 561
 562                 case SCTP_COMM_LOST:
 563                 case SCTP_SHUTDOWN_COMP:
 564                 {
 565                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 566                         if (con) {
 567                                 con->sctp_assoc = 0;
 568                         }
 569                 }
 570                 break;
 571
 572                 /* We don't know which INIT failed, so clear the PENDING flags
 573                  * on them all.  if assoc_id is zero then it will then try
 574                  * again */
 575
 576                 case SCTP_CANT_STR_ASSOC:
 577                 {
 578                         log_print("Can't start SCTP association - retrying");
 579                         sctp_init_failed();
 580                 }
 581                 break;
 582
 583                 default:
 584                         log_print("unexpected SCTP assoc change id=%d state=%d",
 585                                   (int)sn->sn_assoc_change.sac_assoc_id,
 586                                   sn->sn_assoc_change.sac_state);
 587                 }
 588         }
 589 }
 590
 591 /* Data received from remote end */
 592 static int receive_from_sock(struct connection *con)
 593 {
 594         int ret = 0;
 595         struct msghdr msg = {};
 596         struct kvec iov[2];
 597         unsigned len;
 598         int r;
 599         int call_again_soon = 0;
 600         int nvec;
 601         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 602
 603         mutex_lock(&con->sock_mutex);
 604
 605         if (con->sock == NULL) {
 606                 ret = -EAGAIN;
 607                 goto out_close;
 608         }
 609
 610         if (con->rx_page == NULL) {
 611                 /*
 612                  * This doesn't need to be atomic, but I think it should
 613                  * improve performance if it is.
 614                  */
 615                 con->rx_page = alloc_page(GFP_ATOMIC);
 616                 if (con->rx_page == NULL)
 617                         goto out_resched;
 618                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 619         }
 620
 621         /* Only SCTP needs these really */
 622         memset(&incmsg, 0, sizeof(incmsg));
 623         msg.msg_control = incmsg;
 624         msg.msg_controllen = sizeof(incmsg);
 625
 626         /*
 627          * iov[0] is the bit of the circular buffer between the current end
 628          * point (cb.base + cb.len) and the end of the buffer.
 629          */
 630         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 631         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 632         iov[1].iov_len = 0;
 633         nvec = 1;
 634
 635         /*
 636          * iov[1] is the bit of the circular buffer between the start of the
 637          * buffer and the start of the currently used section (cb.base)
 638          */
 639         if (cbuf_data(&con->cb) >= con->cb.base) {
 640                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 641                 iov[1].iov_len = con->cb.base;
 642                 iov[1].iov_base = page_address(con->rx_page);
 643                 nvec = 2;
 644         }
 645         len = iov[0].iov_len + iov[1].iov_len;
 646
 647         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 648                                MSG_DONTWAIT | MSG_NOSIGNAL);
 649         if (ret <= 0)
 650                 goto out_close;
 651
 652         /* Process SCTP notifications */
 653         if (msg.msg_flags & MSG_NOTIFICATION) {
 654                 msg.msg_control = incmsg;
 655                 msg.msg_controllen = sizeof(incmsg);
 656
 657                 process_sctp_notification(con, &msg,
 658                                 page_address(con->rx_page) + con->cb.base);
 659                 mutex_unlock(&con->sock_mutex);
 660                 return 0;
 661         }
 662         BUG_ON(con->nodeid == 0);
 663
 664         if (ret == len)
 665                 call_again_soon = 1;
 666         cbuf_add(&con->cb, ret);
 667         ret = dlm_process_incoming_buffer(con->nodeid,
 668                                           page_address(con->rx_page),
 669                                           con->cb.base, con->cb.len,
 670                                           PAGE_CACHE_SIZE);
 671         if (ret == -EBADMSG) {
 672                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 673                           "iov_len=%u, iov_base[0]=%p, read=%d",
 674                           page_address(con->rx_page), con->cb.base, con->cb.len,
 675                           len, iov[0].iov_base, r);
 676         }
 677         if (ret < 0)
 678                 goto out_close;
 679         cbuf_eat(&con->cb, ret);
 680
 681         if (cbuf_empty(&con->cb) && !call_again_soon) {
 682                 __free_page(con->rx_page);
 683                 con->rx_page = NULL;
 684         }
 685
 686         if (call_again_soon)
 687                 goto out_resched;
 688         mutex_unlock(&con->sock_mutex);
 689         return 0;
 690
 691 out_resched:
 692         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 693                 queue_work(recv_workqueue, &con->rwork);
 694         mutex_unlock(&con->sock_mutex);
 695         return -EAGAIN;
 696
 697 out_close:
 698         mutex_unlock(&con->sock_mutex);
 699         if (ret != -EAGAIN) {
 700                 close_connection(con, false);
 701                 /* Reconnect when there is something to send */
 702         }
 703         /* Don't return success if we really got EOF */
 704         if (ret == 0)
 705                 ret = -EAGAIN;
 706
 707         return ret;
 708 }
 709
 710 /* Listening socket is busy, accept a connection */
 711 static int tcp_accept_from_sock(struct connection *con)
 712 {
 713         int result;
 714         struct sockaddr_storage peeraddr;
 715         struct socket *newsock;
 716         int len;
 717         int nodeid;
 718         struct connection *newcon;
 719         struct connection *addcon;
 720
 721         memset(&peeraddr, 0, sizeof(peeraddr));
 722         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 723                                   IPPROTO_TCP, &newsock);
 724         if (result < 0)
 725                 return -ENOMEM;
 726
 727         mutex_lock_nested(&con->sock_mutex, 0);
 728
 729         result = -ENOTCONN;
 730         if (con->sock == NULL)
 731                 goto accept_err;
 732
 733         newsock->type = con->sock->type;
 734         newsock->ops = con->sock->ops;
 735
 736         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 737         if (result < 0)
 738                 goto accept_err;
 739
 740         /* Get the connected socket's peer */
 741         memset(&peeraddr, 0, sizeof(peeraddr));
 742         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 743                                   &len, 2)) {
 744                 result = -ECONNABORTED;
 745                 goto accept_err;
 746         }
 747
 748         /* Get the new node's NODEID */
 749         make_sockaddr(&peeraddr, 0, &len);
 750         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 751                 log_print("connect from non cluster node");
 752                 sock_release(newsock);
 753                 mutex_unlock(&con->sock_mutex);
 754                 return -1;
 755         }
 756
 757         log_print("got connection from %d", nodeid);
 758
 759         /*  Check to see if we already have a connection to this node. This
 760          *  could happen if the two nodes initiate a connection at roughly
 761          *  the same time and the connections cross on the wire.
 762          *  In this case we store the incoming one in "othercon"
 763          */
 764         newcon = nodeid2con(nodeid, GFP_NOFS);
 765         if (!newcon) {
 766                 result = -ENOMEM;
 767                 goto accept_err;
 768         }
 769         mutex_lock_nested(&newcon->sock_mutex, 1);
 770         if (newcon->sock) {
 771                 struct connection *othercon = newcon->othercon;
 772
 773                 if (!othercon) {
 774                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 775                         if (!othercon) {
 776                                 log_print("failed to allocate incoming socket");
 777                                 mutex_unlock(&newcon->sock_mutex);
 778                                 result = -ENOMEM;
 779                                 goto accept_err;
 780                         }
 781                         othercon->nodeid = nodeid;
 782                         othercon->rx_action = receive_from_sock;
 783                         mutex_init(&othercon->sock_mutex);
 784                         INIT_WORK(&othercon->swork, process_send_sockets);
 785                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 786                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 787                 }
 788                 if (!othercon->sock) {
 789                         newcon->othercon = othercon;
 790                         othercon->sock = newsock;
 791                         newsock->sk->sk_user_data = othercon;
 792                         add_sock(newsock, othercon);
 793                         addcon = othercon;
 794                 }
 795                 else {
 796                         printk("Extra connection from node %d attempted\n", nodeid);
 797                         result = -EAGAIN;
 798                         mutex_unlock(&newcon->sock_mutex);
 799                         goto accept_err;
 800                 }
 801         }
 802         else {
 803                 newsock->sk->sk_user_data = newcon;
 804                 newcon->rx_action = receive_from_sock;
 805                 add_sock(newsock, newcon);
 806                 addcon = newcon;
 807         }
 808
 809         mutex_unlock(&newcon->sock_mutex);
 810
 811         /*
 812          * Add it to the active queue in case we got data
 813          * beween processing the accept adding the socket
 814          * to the read_sockets list
 815          */
 816         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 817                 queue_work(recv_workqueue, &addcon->rwork);
 818         mutex_unlock(&con->sock_mutex);
 819
 820         return 0;
 821
 822 accept_err:
 823         mutex_unlock(&con->sock_mutex);
 824         sock_release(newsock);
 825
 826         if (result != -EAGAIN)
 827                 log_print("error accepting connection from node: %d", result);
 828         return result;
 829 }
 830
 831 static void free_entry(struct writequeue_entry *e)
 832 {
 833         __free_page(e->page);
 834         kfree(e);
 835 }
 836
 837 /* Initiate an SCTP association.
 838    This is a special case of send_to_sock() in that we don't yet have a
 839    peeled-off socket for this association, so we use the listening socket
 840    and add the primary IP address of the remote node.
 841  */
 842 static void sctp_init_assoc(struct connection *con)
 843 {
 844         struct sockaddr_storage rem_addr;
 845         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 846         struct msghdr outmessage;
 847         struct cmsghdr *cmsg;
 848         struct sctp_sndrcvinfo *sinfo;
 849         struct connection *base_con;
 850         struct writequeue_entry *e;
 851         int len, offset;
 852         int ret;
 853         int addrlen;
 854         struct kvec iov[1];
 855
 856         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 857                 return;
 858
 859         if (con->retries++ > MAX_CONNECT_RETRIES)
 860                 return;
 861
 862         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 863                 log_print("no address for nodeid %d", con->nodeid);
 864                 return;
 865         }
 866         base_con = nodeid2con(0, 0);
 867         BUG_ON(base_con == NULL);
 868
 869         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 870
 871         outmessage.msg_name = &rem_addr;
 872         outmessage.msg_namelen = addrlen;
 873         outmessage.msg_control = outcmsg;
 874         outmessage.msg_controllen = sizeof(outcmsg);
 875         outmessage.msg_flags = MSG_EOR;
 876
 877         spin_lock(&con->writequeue_lock);
 878
 879         if (list_empty(&con->writequeue)) {
 880                 spin_unlock(&con->writequeue_lock);
 881                 log_print("writequeue empty for nodeid %d", con->nodeid);
 882                 return;
 883         }
 884
 885         e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
 886         len = e->len;
 887         offset = e->offset;
 888         spin_unlock(&con->writequeue_lock);
 889
 890         /* Send the first block off the write queue */
 891         iov[0].iov_base = page_address(e->page)+offset;
 892         iov[0].iov_len = len;
 893
 894         cmsg = CMSG_FIRSTHDR(&outmessage);
 895         cmsg->cmsg_level = IPPROTO_SCTP;
 896         cmsg->cmsg_type = SCTP_SNDRCV;
 897         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 898         sinfo = CMSG_DATA(cmsg);
 899         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 900         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 901         outmessage.msg_controllen = cmsg->cmsg_len;
 902
 903         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 904         if (ret < 0) {
 905                 log_print("Send first packet to node %d failed: %d",
 906                           con->nodeid, ret);
 907
 908                 /* Try again later */
 909                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 910                 clear_bit(CF_INIT_PENDING, &con->flags);
 911         }
 912         else {
 913                 spin_lock(&con->writequeue_lock);
 914                 e->offset += ret;
 915                 e->len -= ret;
 916
 917                 if (e->len == 0 && e->users == 0) {
 918                         list_del(&e->list);
 919                         free_entry(e);
 920                 }
 921                 spin_unlock(&con->writequeue_lock);
 922         }
 923 }
 924
 925 /* Connect a new socket to its peer */
 926 static void tcp_connect_to_sock(struct connection *con)
 927 {
 928         int result = -EHOSTUNREACH;
 929         struct sockaddr_storage saddr, src_addr;
 930         int addr_len;
 931         struct socket *sock = NULL;
 932         int one = 1;
 933
 934         if (con->nodeid == 0) {
 935                 log_print("attempt to connect sock 0 foiled");
 936                 return;
 937         }
 938
 939         mutex_lock(&con->sock_mutex);
 940         if (con->retries++ > MAX_CONNECT_RETRIES)
 941                 goto out;
 942
 943         /* Some odd races can cause double-connects, ignore them */
 944         if (con->sock) {
 945                 result = 0;
 946                 goto out;
 947         }
 948
 949         /* Create a socket to communicate with */
 950         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 951                                   IPPROTO_TCP, &sock);
 952         if (result < 0)
 953                 goto out_err;
 954
 955         memset(&saddr, 0, sizeof(saddr));
 956         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 957                 goto out_err;
 958
 959         sock->sk->sk_user_data = con;
 960         con->rx_action = receive_from_sock;
 961         con->connect_action = tcp_connect_to_sock;
 962         add_sock(sock, con);
 963
 964         /* Bind to our cluster-known address connecting to avoid
 965            routing problems */
 966         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
 967         make_sockaddr(&src_addr, 0, &addr_len);
 968         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
 969                                  addr_len);
 970         if (result < 0) {
 971                 log_print("could not bind for connect: %d", result);
 972                 /* This *may* not indicate a critical error */
 973         }
 974
 975         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 976
 977         log_print("connecting to %d", con->nodeid);
 978
 979         /* Turn off Nagle's algorithm */
 980         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
 981                           sizeof(one));
 982
 983         result =
 984                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 985                                    O_NONBLOCK);
 986         if (result == -EINPROGRESS)
 987                 result = 0;
 988         if (result == 0)
 989                 goto out;
 990
 991 out_err:
 992         if (con->sock) {
 993                 sock_release(con->sock);
 994                 con->sock = NULL;
 995         } else if (sock) {
 996                 sock_release(sock);
 997         }
 998         /*
 999          * Some errors are fatal and this list might need adjusting. For other
1000          * errors we try again until the max number of retries is reached.
1001          */
1002         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
1003             result != -ENETDOWN && result != -EINVAL
1004             && result != -EPROTONOSUPPORT) {
1005                 lowcomms_connect_sock(con);
1006                 result = 0;
1007         }
1008 out:
1009         mutex_unlock(&con->sock_mutex);
1010         return;
1011 }
1012
1013 static struct socket *tcp_create_listen_sock(struct connection *con,
1014                                              struct sockaddr_storage *saddr)
1015 {
1016         struct socket *sock = NULL;
1017         int result = 0;
1018         int one = 1;
1019         int addr_len;
1020
1021         if (dlm_local_addr[0]->ss_family == AF_INET)
1022                 addr_len = sizeof(struct sockaddr_in);
1023         else
1024                 addr_len = sizeof(struct sockaddr_in6);
1025
1026         /* Create a socket to communicate with */
1027         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
1028                                   IPPROTO_TCP, &sock);
1029         if (result < 0) {
1030                 log_print("Can't create listening comms socket");
1031                 goto create_out;
1032         }
1033
1034         /* Turn off Nagle's algorithm */
1035         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1036                           sizeof(one));
1037
1038         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1039                                    (char *)&one, sizeof(one));
1040
1041         if (result < 0) {
1042                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1043         }
1044         sock->sk->sk_user_data = con;
1045         con->rx_action = tcp_accept_from_sock;
1046         con->connect_action = tcp_connect_to_sock;
1047         con->sock = sock;
1048
1049         /* Bind to our port */
1050         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1051         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1052         if (result < 0) {
1053                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1054                 sock_release(sock);
1055                 sock = NULL;
1056                 con->sock = NULL;
1057                 goto create_out;
1058         }
1059         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1060                                  (char *)&one, sizeof(one));
1061         if (result < 0) {
1062                 log_print("Set keepalive failed: %d", result);
1063         }
1064
1065         result = sock->ops->listen(sock, 5);
1066         if (result < 0) {
1067                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1068                 sock_release(sock);
1069                 sock = NULL;
1070                 goto create_out;
1071         }
1072
1073 create_out:
1074         return sock;
1075 }
1076
1077 /* Get local addresses */
1078 static void init_local(void)
1079 {
1080         struct sockaddr_storage sas, *addr;
1081         int i;
1082
1083         dlm_local_count = 0;
1084         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1085                 if (dlm_our_addr(&sas, i))
1086                         break;
1087
1088                 addr = kmalloc(sizeof(*addr), GFP_NOFS);
1089                 if (!addr)
1090                         break;
1091                 memcpy(addr, &sas, sizeof(*addr));
1092                 dlm_local_addr[dlm_local_count++] = addr;
1093         }
1094 }
1095
1096 /* Bind to an IP address. SCTP allows multiple address so it can do
1097    multi-homing */
1098 static int add_sctp_bind_addr(struct connection *sctp_con,
1099                               struct sockaddr_storage *addr,
1100                               int addr_len, int num)
1101 {
1102         int result = 0;
1103
1104         if (num == 1)
1105                 result = kernel_bind(sctp_con->sock,
1106                                      (struct sockaddr *) addr,
1107                                      addr_len);
1108         else
1109                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1110                                            SCTP_SOCKOPT_BINDX_ADD,
1111                                            (char *)addr, addr_len);
1112
1113         if (result < 0)
1114                 log_print("Can't bind to port %d addr number %d",
1115                           dlm_config.ci_tcp_port, num);
1116
1117         return result;
1118 }
1119
1120 /* Initialise SCTP socket and bind to all interfaces */
1121 static int sctp_listen_for_all(void)
1122 {
1123         struct socket *sock = NULL;
1124         struct sockaddr_storage localaddr;
1125         struct sctp_event_subscribe subscribe;
1126         int result = -EINVAL, num = 1, i, addr_len;
1127         struct connection *con = nodeid2con(0, GFP_NOFS);
1128         int bufsize = NEEDED_RMEM;
1129
1130         if (!con)
1131                 return -ENOMEM;
1132
1133         log_print("Using SCTP for communications");
1134
1135         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1136                                   IPPROTO_SCTP, &sock);
1137         if (result < 0) {
1138                 log_print("Can't create comms socket, check SCTP is loaded");
1139                 goto out;
1140         }
1141
1142         /* Listen for events */
1143         memset(&subscribe, 0, sizeof(subscribe));
1144         subscribe.sctp_data_io_event = 1;
1145         subscribe.sctp_association_event = 1;
1146         subscribe.sctp_send_failure_event = 1;
1147         subscribe.sctp_shutdown_event = 1;
1148         subscribe.sctp_partial_delivery_event = 1;
1149
1150         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1151                                  (char *)&bufsize, sizeof(bufsize));
1152         if (result)
1153                 log_print("Error increasing buffer space on socket %d", result);
1154
1155         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1156                                    (char *)&subscribe, sizeof(subscribe));
1157         if (result < 0) {
1158                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1159                           result);
1160                 goto create_delsock;
1161         }
1162
1163         /* Init con struct */
1164         sock->sk->sk_user_data = con;
1165         con->sock = sock;
1166         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1167         con->rx_action = receive_from_sock;
1168         con->connect_action = sctp_init_assoc;
1169
1170         /* Bind to all interfaces. */
1171         for (i = 0; i < dlm_local_count; i++) {
1172                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1173                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1174
1175                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1176                 if (result)
1177                         goto create_delsock;
1178                 ++num;
1179         }
1180
1181         result = sock->ops->listen(sock, 5);
1182         if (result < 0) {
1183                 log_print("Can't set socket listening");
1184                 goto create_delsock;
1185         }
1186
1187         return 0;
1188
1189 create_delsock:
1190         sock_release(sock);
1191         con->sock = NULL;
1192 out:
1193         return result;
1194 }
1195
1196 static int tcp_listen_for_all(void)
1197 {
1198         struct socket *sock = NULL;
1199         struct connection *con = nodeid2con(0, GFP_NOFS);
1200         int result = -EINVAL;
1201
1202         if (!con)
1203                 return -ENOMEM;
1204
1205         /* We don't support multi-homed hosts */
1206         if (dlm_local_addr[1] != NULL) {
1207                 log_print("TCP protocol can't handle multi-homed hosts, "
1208                           "try SCTP");
1209                 return -EINVAL;
1210         }
1211
1212         log_print("Using TCP for communications");
1213
1214         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1215         if (sock) {
1216                 add_sock(sock, con);
1217                 result = 0;
1218         }
1219         else {
1220                 result = -EADDRINUSE;
1221         }
1222
1223         return result;
1224 }
1225
1226
1227
1228 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1229                                                      gfp_t allocation)
1230 {
1231         struct writequeue_entry *entry;
1232
1233         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1234         if (!entry)
1235                 return NULL;
1236
1237         entry->page = alloc_page(allocation);
1238         if (!entry->page) {
1239                 kfree(entry);
1240                 return NULL;
1241         }
1242
1243         entry->offset = 0;
1244         entry->len = 0;
1245         entry->end = 0;
1246         entry->users = 0;
1247         entry->con = con;
1248
1249         return entry;
1250 }
1251
1252 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1253 {
1254         struct connection *con;
1255         struct writequeue_entry *e;
1256         int offset = 0;
1257         int users = 0;
1258
1259         con = nodeid2con(nodeid, allocation);
1260         if (!con)
1261                 return NULL;
1262
1263         spin_lock(&con->writequeue_lock);
1264         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1265         if ((&e->list == &con->writequeue) ||
1266             (PAGE_CACHE_SIZE - e->end < len)) {
1267                 e = NULL;
1268         } else {
1269                 offset = e->end;
1270                 e->end += len;
1271                 users = e->users++;
1272         }
1273         spin_unlock(&con->writequeue_lock);
1274
1275         if (e) {
1276         got_one:
1277                 *ppc = page_address(e->page) + offset;
1278                 return e;
1279         }
1280
1281         e = new_writequeue_entry(con, allocation);
1282         if (e) {
1283                 spin_lock(&con->writequeue_lock);
1284                 offset = e->end;
1285                 e->end += len;
1286                 users = e->users++;
1287                 list_add_tail(&e->list, &con->writequeue);
1288                 spin_unlock(&con->writequeue_lock);
1289                 goto got_one;
1290         }
1291         return NULL;
1292 }
1293
1294 void dlm_lowcomms_commit_buffer(void *mh)
1295 {
1296         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1297         struct connection *con = e->con;
1298         int users;
1299
1300         spin_lock(&con->writequeue_lock);
1301         users = --e->users;
1302         if (users)
1303                 goto out;
1304         e->len = e->end - e->offset;
1305         spin_unlock(&con->writequeue_lock);
1306
1307         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1308                 queue_work(send_workqueue, &con->swork);
1309         }
1310         return;
1311
1312 out:
1313         spin_unlock(&con->writequeue_lock);
1314         return;
1315 }
1316
1317 /* Send a message */
1318 static void send_to_sock(struct connection *con)
1319 {
1320         int ret = 0;
1321         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1322         struct writequeue_entry *e;
1323         int len, offset;
1324         int count = 0;
1325
1326         mutex_lock(&con->sock_mutex);
1327         if (con->sock == NULL)
1328                 goto out_connect;
1329
1330         spin_lock(&con->writequeue_lock);
1331         for (;;) {
1332                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1333                                list);
1334                 if ((struct list_head *) e == &con->writequeue)
1335                         break;
1336
1337                 len = e->len;
1338                 offset = e->offset;
1339                 BUG_ON(len == 0 && e->users == 0);
1340                 spin_unlock(&con->writequeue_lock);
1341
1342                 ret = 0;
1343                 if (len) {
1344                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1345                                               msg_flags);
1346                         if (ret == -EAGAIN || ret == 0) {
1347                                 if (ret == -EAGAIN &&
1348                                     test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) &&
1349                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1350                                         /* Notify TCP that we're limited by the
1351                                          * application window size.
1352                                          */
1353                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1354                                         con->sock->sk->sk_write_pending++;
1355                                 }
1356                                 cond_resched();
1357                                 goto out;
1358                         }
1359                         if (ret <= 0)
1360                                 goto send_error;
1361                 }
1362
1363                 /* Don't starve people filling buffers */
1364                 if (++count >= MAX_SEND_MSG_COUNT) {
1365                         cond_resched();
1366                         count = 0;
1367                 }
1368
1369                 spin_lock(&con->writequeue_lock);
1370                 e->offset += ret;
1371                 e->len -= ret;
1372
1373                 if (e->len == 0 && e->users == 0) {
1374                         list_del(&e->list);
1375                         free_entry(e);
1376                         continue;
1377                 }
1378         }
1379         spin_unlock(&con->writequeue_lock);
1380 out:
1381         mutex_unlock(&con->sock_mutex);
1382         return;
1383
1384 send_error:
1385         mutex_unlock(&con->sock_mutex);
1386         close_connection(con, false);
1387         lowcomms_connect_sock(con);
1388         return;
1389
1390 out_connect:
1391         mutex_unlock(&con->sock_mutex);
1392         if (!test_bit(CF_INIT_PENDING, &con->flags))
1393                 lowcomms_connect_sock(con);
1394         return;
1395 }
1396
1397 static void clean_one_writequeue(struct connection *con)
1398 {
1399         struct writequeue_entry *e, *safe;
1400
1401         spin_lock(&con->writequeue_lock);
1402         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1403                 list_del(&e->list);
1404                 free_entry(e);
1405         }
1406         spin_unlock(&con->writequeue_lock);
1407 }
1408
1409 /* Called from recovery when it knows that a node has
1410    left the cluster */
1411 int dlm_lowcomms_close(int nodeid)
1412 {
1413         struct connection *con;
1414
1415         log_print("closing connection to node %d", nodeid);
1416         con = nodeid2con(nodeid, 0);
1417         if (con) {
1418                 clear_bit(CF_CONNECT_PENDING, &con->flags);
1419                 clear_bit(CF_WRITE_PENDING, &con->flags);
1420                 set_bit(CF_CLOSE, &con->flags);
1421                 if (cancel_work_sync(&con->swork))
1422                         log_print("canceled swork for node %d", nodeid);
1423                 if (cancel_work_sync(&con->rwork))
1424                         log_print("canceled rwork for node %d", nodeid);
1425                 clean_one_writequeue(con);
1426                 close_connection(con, true);
1427         }
1428         return 0;
1429 }
1430
1431 /* Receive workqueue function */
1432 static void process_recv_sockets(struct work_struct *work)
1433 {
1434         struct connection *con = container_of(work, struct connection, rwork);
1435         int err;
1436
1437         clear_bit(CF_READ_PENDING, &con->flags);
1438         do {
1439                 err = con->rx_action(con);
1440         } while (!err);
1441 }
1442
1443 /* Send workqueue function */
1444 static void process_send_sockets(struct work_struct *work)
1445 {
1446         struct connection *con = container_of(work, struct connection, swork);
1447
1448         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1449                 con->connect_action(con);
1450                 set_bit(CF_WRITE_PENDING, &con->flags);
1451         }
1452         if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
1453                 send_to_sock(con);
1454 }
1455
1456
1457 /* Discard all entries on the write queues */
1458 static void clean_writequeues(void)
1459 {
1460         foreach_conn(clean_one_writequeue);
1461 }
1462
1463 static void work_stop(void)
1464 {
1465         destroy_workqueue(recv_workqueue);
1466         destroy_workqueue(send_workqueue);
1467 }
1468
1469 static int work_start(void)
1470 {
1471         recv_workqueue = alloc_workqueue("dlm_recv", WQ_MEM_RECLAIM |
1472                                          WQ_HIGHPRI | WQ_FREEZEABLE, 0);
1473         if (!recv_workqueue) {
1474                 log_print("can't start dlm_recv");
1475                 return -ENOMEM;
1476         }
1477
1478         send_workqueue = alloc_workqueue("dlm_send", WQ_MEM_RECLAIM |
1479                                          WQ_HIGHPRI | WQ_FREEZEABLE, 0);
1480         if (!send_workqueue) {
1481                 log_print("can't start dlm_send");
1482                 destroy_workqueue(recv_workqueue);
1483                 return -ENOMEM;
1484         }
1485
1486         return 0;
1487 }
1488
1489 static void stop_conn(struct connection *con)
1490 {
1491         con->flags |= 0x0F;
1492         if (con->sock && con->sock->sk)
1493                 con->sock->sk->sk_user_data = NULL;
1494 }
1495
1496 static void free_conn(struct connection *con)
1497 {
1498         close_connection(con, true);
1499         if (con->othercon)
1500                 kmem_cache_free(con_cache, con->othercon);
1501         hlist_del(&con->list);
1502         kmem_cache_free(con_cache, con);
1503 }
1504
1505 void dlm_lowcomms_stop(void)
1506 {
1507         /* Set all the flags to prevent any
1508            socket activity.
1509         */
1510         mutex_lock(&connections_lock);
1511         foreach_conn(stop_conn);
1512         mutex_unlock(&connections_lock);
1513
1514         work_stop();
1515
1516         mutex_lock(&connections_lock);
1517         clean_writequeues();
1518
1519         foreach_conn(free_conn);
1520
1521         mutex_unlock(&connections_lock);
1522         kmem_cache_destroy(con_cache);
1523 }
1524
1525 int dlm_lowcomms_start(void)
1526 {
1527         int error = -EINVAL;
1528         struct connection *con;
1529         int i;
1530
1531         for (i = 0; i < CONN_HASH_SIZE; i++)
1532                 INIT_HLIST_HEAD(&connection_hash[i]);
1533
1534         init_local();
1535         if (!dlm_local_count) {
1536                 error = -ENOTCONN;
1537                 log_print("no local IP address has been set");
1538                 goto out;
1539         }
1540
1541         error = -ENOMEM;
1542         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1543                                       __alignof__(struct connection), 0,
1544                                       NULL);
1545         if (!con_cache)
1546                 goto out;
1547
1548         /* Start listening */
1549         if (dlm_config.ci_protocol == 0)
1550                 error = tcp_listen_for_all();
1551         else
1552                 error = sctp_listen_for_all();
1553         if (error)
1554                 goto fail_unlisten;
1555
1556         error = work_start();
1557         if (error)
1558                 goto fail_unlisten;
1559
1560         return 0;
1561
1562 fail_unlisten:
1563         con = nodeid2con(0,0);
1564         if (con) {
1565                 close_connection(con, false);
1566                 kmem_cache_free(con_cache, con);
1567         }
1568         kmem_cache_destroy(con_cache);
1569
1570 out:
1571         return error;
1572 }