fs/dlm/lowcomms.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /******************************************************************************
   3 *******************************************************************************
   4 **
   5 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   6 **  Copyright (C) 2004-2009 Red Hat, Inc.  All rights reserved.
   7 **
   8 **
   9 *******************************************************************************
  10 ******************************************************************************/
  11
  12 /*
  13  * lowcomms.c
  14  *
  15  * This is the "low-level" comms layer.
  16  *
  17  * It is responsible for sending/receiving messages
  18  * from other nodes in the cluster.
  19  *
  20  * Cluster nodes are referred to by their nodeids. nodeids are
  21  * simply 32 bit numbers to the locking module - if they need to
  22  * be expanded for the cluster infrastructure then that is its
  23  * responsibility. It is this layer's
  24  * responsibility to resolve these into IP address or
  25  * whatever it needs for inter-node communication.
  26  *
  27  * The comms level is two kernel threads that deal mainly with
  28  * the receiving of messages from other nodes and passing them
  29  * up to the mid-level comms layer (which understands the
  30  * message format) for execution by the locking core, and
  31  * a send thread which does all the setting up of connections
  32  * to remote nodes and the sending of data. Threads are not allowed
  33  * to send their own data because it may cause them to wait in times
  34  * of high load. Also, this way, the sending thread can collect together
  35  * messages bound for one node and send them in one block.
  36  *
  37  * lowcomms will choose to use either TCP or SCTP as its transport layer
  38  * depending on the configuration variable 'protocol'. This should be set
  39  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  40  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  41  * for the DLM to function.
  42  *
  43  */
  44
  45 #include <asm/ioctls.h>
  46 #include <net/sock.h>
  47 #include <net/tcp.h>
  48 #include <linux/pagemap.h>
  49 #include <linux/file.h>
  50 #include <linux/mutex.h>
  51 #include <linux/sctp.h>
  52 #include <linux/slab.h>
  53 #include <net/sctp/sctp.h>
  54 #include <net/ipv6.h>
  55
  56 #include "dlm_internal.h"
  57 #include "lowcomms.h"
  58 #include "midcomms.h"
  59 #include "config.h"
  60
  61 #define NEEDED_RMEM (4*1024*1024)
  62 #define CONN_HASH_SIZE 32
  63
  64 /* Number of messages to send before rescheduling */
  65 #define MAX_SEND_MSG_COUNT 25
  66 #define DLM_SHUTDOWN_WAIT_TIMEOUT msecs_to_jiffies(10000)
  67
  68 struct connection {
  69         struct socket *sock;    /* NULL if not connected */
  70         uint32_t nodeid;        /* So we know who we are in the list */
  71         struct mutex sock_mutex;
  72         unsigned long flags;
  73 #define CF_READ_PENDING 1
  74 #define CF_WRITE_PENDING 2
  75 #define CF_INIT_PENDING 4
  76 #define CF_IS_OTHERCON 5
  77 #define CF_CLOSE 6
  78 #define CF_APP_LIMITED 7
  79 #define CF_CLOSING 8
  80 #define CF_SHUTDOWN 9
  81 #define CF_CONNECTED 10
  82         struct list_head writequeue;  /* List of outgoing writequeue_entries */
  83         spinlock_t writequeue_lock;
  84         void (*connect_action) (struct connection *);   /* What to do to connect */
  85         void (*shutdown_action)(struct connection *con); /* What to do to shutdown */
  86         int retries;
  87 #define MAX_CONNECT_RETRIES 3
  88         struct hlist_node list;
  89         struct connection *othercon;
  90         struct work_struct rwork; /* Receive workqueue */
  91         struct work_struct swork; /* Send workqueue */
  92         wait_queue_head_t shutdown_wait; /* wait for graceful shutdown */
  93         unsigned char *rx_buf;
  94         int rx_buflen;
  95         int rx_leftover;
  96         struct rcu_head rcu;
  97 };
  98 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
  99
 100 struct listen_connection {
 101         struct socket *sock;
 102         struct work_struct rwork;
 103 };
 104
 105 /* An entry waiting to be sent */
 106 struct writequeue_entry {
 107         struct list_head list;
 108         struct page *page;
 109         int offset;
 110         int len;
 111         int end;
 112         int users;
 113         struct connection *con;
 114 };
 115
 116 struct dlm_node_addr {
 117         struct list_head list;
 118         int nodeid;
 119         int mark;
 120         int addr_count;
 121         int curr_addr_index;
 122         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 123 };
 124
 125 static struct listen_sock_callbacks {
 126         void (*sk_error_report)(struct sock *);
 127         void (*sk_data_ready)(struct sock *);
 128         void (*sk_state_change)(struct sock *);
 129         void (*sk_write_space)(struct sock *);
 130 } listen_sock;
 131
 132 static LIST_HEAD(dlm_node_addrs);
 133 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 134
 135 static struct listen_connection listen_con;
 136 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 137 static int dlm_local_count;
 138 static int dlm_allow_conn;
 139
 140 /* Work queues */
 141 static struct workqueue_struct *recv_workqueue;
 142 static struct workqueue_struct *send_workqueue;
 143
 144 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 145 static DEFINE_SPINLOCK(connections_lock);
 146 DEFINE_STATIC_SRCU(connections_srcu);
 147
 148 static void process_recv_sockets(struct work_struct *work);
 149 static void process_send_sockets(struct work_struct *work);
 150
 151 static void sctp_connect_to_sock(struct connection *con);
 152 static void tcp_connect_to_sock(struct connection *con);
 153 static void dlm_tcp_shutdown(struct connection *con);
 154
 155 /* This is deliberately very simple because most clusters have simple
 156    sequential nodeids, so we should be able to go straight to a connection
 157    struct in the array */
 158 static inline int nodeid_hash(int nodeid)
 159 {
 160         return nodeid & (CONN_HASH_SIZE-1);
 161 }
 162
 163 static struct connection *__find_con(int nodeid)
 164 {
 165         int r, idx;
 166         struct connection *con;
 167
 168         r = nodeid_hash(nodeid);
 169
 170         idx = srcu_read_lock(&connections_srcu);
 171         hlist_for_each_entry_rcu(con, &connection_hash[r], list) {
 172                 if (con->nodeid == nodeid) {
 173                         srcu_read_unlock(&connections_srcu, idx);
 174                         return con;
 175                 }
 176         }
 177         srcu_read_unlock(&connections_srcu, idx);
 178
 179         return NULL;
 180 }
 181
 182 static int dlm_con_init(struct connection *con, int nodeid)
 183 {
 184         con->rx_buflen = dlm_config.ci_buffer_size;
 185         con->rx_buf = kmalloc(con->rx_buflen, GFP_NOFS);
 186         if (!con->rx_buf)
 187                 return -ENOMEM;
 188
 189         con->nodeid = nodeid;
 190         mutex_init(&con->sock_mutex);
 191         INIT_LIST_HEAD(&con->writequeue);
 192         spin_lock_init(&con->writequeue_lock);
 193         INIT_WORK(&con->swork, process_send_sockets);
 194         INIT_WORK(&con->rwork, process_recv_sockets);
 195         init_waitqueue_head(&con->shutdown_wait);
 196
 197         if (dlm_config.ci_protocol == 0) {
 198                 con->connect_action = tcp_connect_to_sock;
 199                 con->shutdown_action = dlm_tcp_shutdown;
 200         } else {
 201                 con->connect_action = sctp_connect_to_sock;
 202         }
 203
 204         return 0;
 205 }
 206
 207 /*
 208  * If 'allocation' is zero then we don't attempt to create a new
 209  * connection structure for this node.
 210  */
 211 static struct connection *nodeid2con(int nodeid, gfp_t alloc)
 212 {
 213         struct connection *con, *tmp;
 214         int r, ret;
 215
 216         con = __find_con(nodeid);
 217         if (con || !alloc)
 218                 return con;
 219
 220         con = kzalloc(sizeof(*con), alloc);
 221         if (!con)
 222                 return NULL;
 223
 224         ret = dlm_con_init(con, nodeid);
 225         if (ret) {
 226                 kfree(con);
 227                 return NULL;
 228         }
 229
 230         r = nodeid_hash(nodeid);
 231
 232         spin_lock(&connections_lock);
 233         /* Because multiple workqueues/threads calls this function it can
 234          * race on multiple cpu's. Instead of locking hot path __find_con()
 235          * we just check in rare cases of recently added nodes again
 236          * under protection of connections_lock. If this is the case we
 237          * abort our connection creation and return the existing connection.
 238          */
 239         tmp = __find_con(nodeid);
 240         if (tmp) {
 241                 spin_unlock(&connections_lock);
 242                 kfree(con->rx_buf);
 243                 kfree(con);
 244                 return tmp;
 245         }
 246
 247         hlist_add_head_rcu(&con->list, &connection_hash[r]);
 248         spin_unlock(&connections_lock);
 249
 250         return con;
 251 }
 252
 253 /* Loop round all connections */
 254 static void foreach_conn(void (*conn_func)(struct connection *c))
 255 {
 256         int i, idx;
 257         struct connection *con;
 258
 259         idx = srcu_read_lock(&connections_srcu);
 260         for (i = 0; i < CONN_HASH_SIZE; i++) {
 261                 hlist_for_each_entry_rcu(con, &connection_hash[i], list)
 262                         conn_func(con);
 263         }
 264         srcu_read_unlock(&connections_srcu, idx);
 265 }
 266
 267 static struct dlm_node_addr *find_node_addr(int nodeid)
 268 {
 269         struct dlm_node_addr *na;
 270
 271         list_for_each_entry(na, &dlm_node_addrs, list) {
 272                 if (na->nodeid == nodeid)
 273                         return na;
 274         }
 275         return NULL;
 276 }
 277
 278 static int addr_compare(const struct sockaddr_storage *x,
 279                         const struct sockaddr_storage *y)
 280 {
 281         switch (x->ss_family) {
 282         case AF_INET: {
 283                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 284                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 285                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 286                         return 0;
 287                 if (sinx->sin_port != siny->sin_port)
 288                         return 0;
 289                 break;
 290         }
 291         case AF_INET6: {
 292                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 293                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 294                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 295                         return 0;
 296                 if (sinx->sin6_port != siny->sin6_port)
 297                         return 0;
 298                 break;
 299         }
 300         default:
 301                 return 0;
 302         }
 303         return 1;
 304 }
 305
 306 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 307                           struct sockaddr *sa_out, bool try_new_addr,
 308                           unsigned int *mark)
 309 {
 310         struct sockaddr_storage sas;
 311         struct dlm_node_addr *na;
 312
 313         if (!dlm_local_count)
 314                 return -1;
 315
 316         spin_lock(&dlm_node_addrs_spin);
 317         na = find_node_addr(nodeid);
 318         if (na && na->addr_count) {
 319                 memcpy(&sas, na->addr[na->curr_addr_index],
 320                        sizeof(struct sockaddr_storage));
 321
 322                 if (try_new_addr) {
 323                         na->curr_addr_index++;
 324                         if (na->curr_addr_index == na->addr_count)
 325                                 na->curr_addr_index = 0;
 326                 }
 327         }
 328         spin_unlock(&dlm_node_addrs_spin);
 329
 330         if (!na)
 331                 return -EEXIST;
 332
 333         if (!na->addr_count)
 334                 return -ENOENT;
 335
 336         *mark = na->mark;
 337
 338         if (sas_out)
 339                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 340
 341         if (!sa_out)
 342                 return 0;
 343
 344         if (dlm_local_addr[0]->ss_family == AF_INET) {
 345                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 346                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 347                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 348         } else {
 349                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 350                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 351                 ret6->sin6_addr = in6->sin6_addr;
 352         }
 353
 354         return 0;
 355 }
 356
 357 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid,
 358                           unsigned int *mark)
 359 {
 360         struct dlm_node_addr *na;
 361         int rv = -EEXIST;
 362         int addr_i;
 363
 364         spin_lock(&dlm_node_addrs_spin);
 365         list_for_each_entry(na, &dlm_node_addrs, list) {
 366                 if (!na->addr_count)
 367                         continue;
 368
 369                 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
 370                         if (addr_compare(na->addr[addr_i], addr)) {
 371                                 *nodeid = na->nodeid;
 372                                 *mark = na->mark;
 373                                 rv = 0;
 374                                 goto unlock;
 375                         }
 376                 }
 377         }
 378 unlock:
 379         spin_unlock(&dlm_node_addrs_spin);
 380         return rv;
 381 }
 382
 383 /* caller need to held dlm_node_addrs_spin lock */
 384 static bool dlm_lowcomms_na_has_addr(const struct dlm_node_addr *na,
 385                                      const struct sockaddr_storage *addr)
 386 {
 387         int i;
 388
 389         for (i = 0; i < na->addr_count; i++) {
 390                 if (addr_compare(na->addr[i], addr))
 391                         return true;
 392         }
 393
 394         return false;
 395 }
 396
 397 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 398 {
 399         struct sockaddr_storage *new_addr;
 400         struct dlm_node_addr *new_node, *na;
 401         bool ret;
 402
 403         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 404         if (!new_node)
 405                 return -ENOMEM;
 406
 407         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 408         if (!new_addr) {
 409                 kfree(new_node);
 410                 return -ENOMEM;
 411         }
 412
 413         memcpy(new_addr, addr, len);
 414
 415         spin_lock(&dlm_node_addrs_spin);
 416         na = find_node_addr(nodeid);
 417         if (!na) {
 418                 new_node->nodeid = nodeid;
 419                 new_node->addr[0] = new_addr;
 420                 new_node->addr_count = 1;
 421                 new_node->mark = dlm_config.ci_mark;
 422                 list_add(&new_node->list, &dlm_node_addrs);
 423                 spin_unlock(&dlm_node_addrs_spin);
 424                 return 0;
 425         }
 426
 427         ret = dlm_lowcomms_na_has_addr(na, addr);
 428         if (ret) {
 429                 spin_unlock(&dlm_node_addrs_spin);
 430                 kfree(new_addr);
 431                 kfree(new_node);
 432                 return -EEXIST;
 433         }
 434
 435         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 436                 spin_unlock(&dlm_node_addrs_spin);
 437                 kfree(new_addr);
 438                 kfree(new_node);
 439                 return -ENOSPC;
 440         }
 441
 442         na->addr[na->addr_count++] = new_addr;
 443         spin_unlock(&dlm_node_addrs_spin);
 444         kfree(new_node);
 445         return 0;
 446 }
 447
 448 /* Data available on socket or listen socket received a connect */
 449 static void lowcomms_data_ready(struct sock *sk)
 450 {
 451         struct connection *con;
 452
 453         read_lock_bh(&sk->sk_callback_lock);
 454         con = sock2con(sk);
 455         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 456                 queue_work(recv_workqueue, &con->rwork);
 457         read_unlock_bh(&sk->sk_callback_lock);
 458 }
 459
 460 static void lowcomms_listen_data_ready(struct sock *sk)
 461 {
 462         queue_work(recv_workqueue, &listen_con.rwork);
 463 }
 464
 465 static void lowcomms_write_space(struct sock *sk)
 466 {
 467         struct connection *con;
 468
 469         read_lock_bh(&sk->sk_callback_lock);
 470         con = sock2con(sk);
 471         if (!con)
 472                 goto out;
 473
 474         if (!test_and_set_bit(CF_CONNECTED, &con->flags)) {
 475                 log_print("successful connected to node %d", con->nodeid);
 476                 queue_work(send_workqueue, &con->swork);
 477                 goto out;
 478         }
 479
 480         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 481
 482         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 483                 con->sock->sk->sk_write_pending--;
 484                 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
 485         }
 486
 487         queue_work(send_workqueue, &con->swork);
 488 out:
 489         read_unlock_bh(&sk->sk_callback_lock);
 490 }
 491
 492 static inline void lowcomms_connect_sock(struct connection *con)
 493 {
 494         if (test_bit(CF_CLOSE, &con->flags))
 495                 return;
 496         queue_work(send_workqueue, &con->swork);
 497         cond_resched();
 498 }
 499
 500 static void lowcomms_state_change(struct sock *sk)
 501 {
 502         /* SCTP layer is not calling sk_data_ready when the connection
 503          * is done, so we catch the signal through here. Also, it
 504          * doesn't switch socket state when entering shutdown, so we
 505          * skip the write in that case.
 506          */
 507         if (sk->sk_shutdown) {
 508                 if (sk->sk_shutdown == RCV_SHUTDOWN)
 509                         lowcomms_data_ready(sk);
 510         } else if (sk->sk_state == TCP_ESTABLISHED) {
 511                 lowcomms_write_space(sk);
 512         }
 513 }
 514
 515 int dlm_lowcomms_connect_node(int nodeid)
 516 {
 517         struct connection *con;
 518
 519         if (nodeid == dlm_our_nodeid())
 520                 return 0;
 521
 522         con = nodeid2con(nodeid, GFP_NOFS);
 523         if (!con)
 524                 return -ENOMEM;
 525         lowcomms_connect_sock(con);
 526         return 0;
 527 }
 528
 529 int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark)
 530 {
 531         struct dlm_node_addr *na;
 532
 533         spin_lock(&dlm_node_addrs_spin);
 534         na = find_node_addr(nodeid);
 535         if (!na) {
 536                 spin_unlock(&dlm_node_addrs_spin);
 537                 return -ENOENT;
 538         }
 539
 540         na->mark = mark;
 541         spin_unlock(&dlm_node_addrs_spin);
 542
 543         return 0;
 544 }
 545
 546 static void lowcomms_error_report(struct sock *sk)
 547 {
 548         struct connection *con;
 549         struct sockaddr_storage saddr;
 550         void (*orig_report)(struct sock *) = NULL;
 551
 552         read_lock_bh(&sk->sk_callback_lock);
 553         con = sock2con(sk);
 554         if (con == NULL)
 555                 goto out;
 556
 557         orig_report = listen_sock.sk_error_report;
 558         if (con->sock == NULL ||
 559             kernel_getpeername(con->sock, (struct sockaddr *)&saddr) < 0) {
 560                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 561                                    "sending to node %d, port %d, "
 562                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 563                                    con->nodeid, dlm_config.ci_tcp_port,
 564                                    sk->sk_err, sk->sk_err_soft);
 565         } else if (saddr.ss_family == AF_INET) {
 566                 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
 567
 568                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 569                                    "sending to node %d at %pI4, port %d, "
 570                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 571                                    con->nodeid, &sin4->sin_addr.s_addr,
 572                                    dlm_config.ci_tcp_port, sk->sk_err,
 573                                    sk->sk_err_soft);
 574         } else {
 575                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
 576
 577                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 578                                    "sending to node %d at %u.%u.%u.%u, "
 579                                    "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 580                                    con->nodeid, sin6->sin6_addr.s6_addr32[0],
 581                                    sin6->sin6_addr.s6_addr32[1],
 582                                    sin6->sin6_addr.s6_addr32[2],
 583                                    sin6->sin6_addr.s6_addr32[3],
 584                                    dlm_config.ci_tcp_port, sk->sk_err,
 585                                    sk->sk_err_soft);
 586         }
 587 out:
 588         read_unlock_bh(&sk->sk_callback_lock);
 589         if (orig_report)
 590                 orig_report(sk);
 591 }
 592
 593 /* Note: sk_callback_lock must be locked before calling this function. */
 594 static void save_listen_callbacks(struct socket *sock)
 595 {
 596         struct sock *sk = sock->sk;
 597
 598         listen_sock.sk_data_ready = sk->sk_data_ready;
 599         listen_sock.sk_state_change = sk->sk_state_change;
 600         listen_sock.sk_write_space = sk->sk_write_space;
 601         listen_sock.sk_error_report = sk->sk_error_report;
 602 }
 603
 604 static void restore_callbacks(struct socket *sock)
 605 {
 606         struct sock *sk = sock->sk;
 607
 608         write_lock_bh(&sk->sk_callback_lock);
 609         sk->sk_user_data = NULL;
 610         sk->sk_data_ready = listen_sock.sk_data_ready;
 611         sk->sk_state_change = listen_sock.sk_state_change;
 612         sk->sk_write_space = listen_sock.sk_write_space;
 613         sk->sk_error_report = listen_sock.sk_error_report;
 614         write_unlock_bh(&sk->sk_callback_lock);
 615 }
 616
 617 static void add_listen_sock(struct socket *sock, struct listen_connection *con)
 618 {
 619         struct sock *sk = sock->sk;
 620
 621         write_lock_bh(&sk->sk_callback_lock);
 622         save_listen_callbacks(sock);
 623         con->sock = sock;
 624
 625         sk->sk_user_data = con;
 626         sk->sk_allocation = GFP_NOFS;
 627         /* Install a data_ready callback */
 628         sk->sk_data_ready = lowcomms_listen_data_ready;
 629         write_unlock_bh(&sk->sk_callback_lock);
 630 }
 631
 632 /* Make a socket active */
 633 static void add_sock(struct socket *sock, struct connection *con)
 634 {
 635         struct sock *sk = sock->sk;
 636
 637         write_lock_bh(&sk->sk_callback_lock);
 638         con->sock = sock;
 639
 640         sk->sk_user_data = con;
 641         /* Install a data_ready callback */
 642         sk->sk_data_ready = lowcomms_data_ready;
 643         sk->sk_write_space = lowcomms_write_space;
 644         sk->sk_state_change = lowcomms_state_change;
 645         sk->sk_allocation = GFP_NOFS;
 646         sk->sk_error_report = lowcomms_error_report;
 647         write_unlock_bh(&sk->sk_callback_lock);
 648 }
 649
 650 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 651    length */
 652 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 653                           int *addr_len)
 654 {
 655         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 656         if (saddr->ss_family == AF_INET) {
 657                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 658                 in4_addr->sin_port = cpu_to_be16(port);
 659                 *addr_len = sizeof(struct sockaddr_in);
 660                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 661         } else {
 662                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 663                 in6_addr->sin6_port = cpu_to_be16(port);
 664                 *addr_len = sizeof(struct sockaddr_in6);
 665         }
 666         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 667 }
 668
 669 static void dlm_close_sock(struct socket **sock)
 670 {
 671         if (*sock) {
 672                 restore_callbacks(*sock);
 673                 sock_release(*sock);
 674                 *sock = NULL;
 675         }
 676 }
 677
 678 /* Close a remote connection and tidy up */
 679 static void close_connection(struct connection *con, bool and_other,
 680                              bool tx, bool rx)
 681 {
 682         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 683
 684         if (tx && !closing && cancel_work_sync(&con->swork)) {
 685                 log_print("canceled swork for node %d", con->nodeid);
 686                 clear_bit(CF_WRITE_PENDING, &con->flags);
 687         }
 688         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 689                 log_print("canceled rwork for node %d", con->nodeid);
 690                 clear_bit(CF_READ_PENDING, &con->flags);
 691         }
 692
 693         mutex_lock(&con->sock_mutex);
 694         dlm_close_sock(&con->sock);
 695
 696         if (con->othercon && and_other) {
 697                 /* Will only re-enter once. */
 698                 close_connection(con->othercon, false, true, true);
 699         }
 700
 701         con->rx_leftover = 0;
 702         con->retries = 0;
 703         clear_bit(CF_CONNECTED, &con->flags);
 704         mutex_unlock(&con->sock_mutex);
 705         clear_bit(CF_CLOSING, &con->flags);
 706 }
 707
 708 static void shutdown_connection(struct connection *con)
 709 {
 710         int ret;
 711
 712         if (cancel_work_sync(&con->swork)) {
 713                 log_print("canceled swork for node %d", con->nodeid);
 714                 clear_bit(CF_WRITE_PENDING, &con->flags);
 715         }
 716
 717         mutex_lock(&con->sock_mutex);
 718         /* nothing to shutdown */
 719         if (!con->sock) {
 720                 mutex_unlock(&con->sock_mutex);
 721                 return;
 722         }
 723
 724         set_bit(CF_SHUTDOWN, &con->flags);
 725         ret = kernel_sock_shutdown(con->sock, SHUT_WR);
 726         mutex_unlock(&con->sock_mutex);
 727         if (ret) {
 728                 log_print("Connection %p failed to shutdown: %d will force close",
 729                           con, ret);
 730                 goto force_close;
 731         } else {
 732                 ret = wait_event_timeout(con->shutdown_wait,
 733                                          !test_bit(CF_SHUTDOWN, &con->flags),
 734                                          DLM_SHUTDOWN_WAIT_TIMEOUT);
 735                 if (ret == 0) {
 736                         log_print("Connection %p shutdown timed out, will force close",
 737                                   con);
 738                         goto force_close;
 739                 }
 740         }
 741
 742         return;
 743
 744 force_close:
 745         clear_bit(CF_SHUTDOWN, &con->flags);
 746         close_connection(con, false, true, true);
 747 }
 748
 749 static void dlm_tcp_shutdown(struct connection *con)
 750 {
 751         if (con->othercon)
 752                 shutdown_connection(con->othercon);
 753         shutdown_connection(con);
 754 }
 755
 756 static int con_realloc_receive_buf(struct connection *con, int newlen)
 757 {
 758         unsigned char *newbuf;
 759
 760         newbuf = kmalloc(newlen, GFP_NOFS);
 761         if (!newbuf)
 762                 return -ENOMEM;
 763
 764         /* copy any leftover from last receive */
 765         if (con->rx_leftover)
 766                 memmove(newbuf, con->rx_buf, con->rx_leftover);
 767
 768         /* swap to new buffer space */
 769         kfree(con->rx_buf);
 770         con->rx_buflen = newlen;
 771         con->rx_buf = newbuf;
 772
 773         return 0;
 774 }
 775
 776 /* Data received from remote end */
 777 static int receive_from_sock(struct connection *con)
 778 {
 779         int call_again_soon = 0;
 780         struct msghdr msg;
 781         struct kvec iov;
 782         int ret, buflen;
 783
 784         mutex_lock(&con->sock_mutex);
 785
 786         if (con->sock == NULL) {
 787                 ret = -EAGAIN;
 788                 goto out_close;
 789         }
 790
 791         /* realloc if we get new buffer size to read out */
 792         buflen = dlm_config.ci_buffer_size;
 793         if (con->rx_buflen != buflen && con->rx_leftover <= buflen) {
 794                 ret = con_realloc_receive_buf(con, buflen);
 795                 if (ret < 0)
 796                         goto out_resched;
 797         }
 798
 799         /* calculate new buffer parameter regarding last receive and
 800          * possible leftover bytes
 801          */
 802         iov.iov_base = con->rx_buf + con->rx_leftover;
 803         iov.iov_len = con->rx_buflen - con->rx_leftover;
 804
 805         memset(&msg, 0, sizeof(msg));
 806         msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
 807         ret = kernel_recvmsg(con->sock, &msg, &iov, 1, iov.iov_len,
 808                              msg.msg_flags);
 809         if (ret <= 0)
 810                 goto out_close;
 811         else if (ret == iov.iov_len)
 812                 call_again_soon = 1;
 813
 814         /* new buflen according readed bytes and leftover from last receive */
 815         buflen = ret + con->rx_leftover;
 816         ret = dlm_process_incoming_buffer(con->nodeid, con->rx_buf, buflen);
 817         if (ret < 0)
 818                 goto out_close;
 819
 820         /* calculate leftover bytes from process and put it into begin of
 821          * the receive buffer, so next receive we have the full message
 822          * at the start address of the receive buffer.
 823          */
 824         con->rx_leftover = buflen - ret;
 825         if (con->rx_leftover) {
 826                 memmove(con->rx_buf, con->rx_buf + ret,
 827                         con->rx_leftover);
 828                 call_again_soon = true;
 829         }
 830
 831         if (call_again_soon)
 832                 goto out_resched;
 833
 834         mutex_unlock(&con->sock_mutex);
 835         return 0;
 836
 837 out_resched:
 838         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 839                 queue_work(recv_workqueue, &con->rwork);
 840         mutex_unlock(&con->sock_mutex);
 841         return -EAGAIN;
 842
 843 out_close:
 844         mutex_unlock(&con->sock_mutex);
 845         if (ret != -EAGAIN) {
 846                 /* Reconnect when there is something to send */
 847                 close_connection(con, false, true, false);
 848                 if (ret == 0) {
 849                         log_print("connection %p got EOF from %d",
 850                                   con, con->nodeid);
 851                         /* handling for tcp shutdown */
 852                         clear_bit(CF_SHUTDOWN, &con->flags);
 853                         wake_up(&con->shutdown_wait);
 854                         /* signal to breaking receive worker */
 855                         ret = -1;
 856                 }
 857         }
 858         return ret;
 859 }
 860
 861 /* Listening socket is busy, accept a connection */
 862 static int accept_from_sock(struct listen_connection *con)
 863 {
 864         int result;
 865         struct sockaddr_storage peeraddr;
 866         struct socket *newsock;
 867         int len;
 868         int nodeid;
 869         struct connection *newcon;
 870         struct connection *addcon;
 871         unsigned int mark;
 872
 873         if (!dlm_allow_conn) {
 874                 return -1;
 875         }
 876
 877         if (!con->sock)
 878                 return -ENOTCONN;
 879
 880         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 881         if (result < 0)
 882                 goto accept_err;
 883
 884         /* Get the connected socket's peer */
 885         memset(&peeraddr, 0, sizeof(peeraddr));
 886         len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
 887         if (len < 0) {
 888                 result = -ECONNABORTED;
 889                 goto accept_err;
 890         }
 891
 892         /* Get the new node's NODEID */
 893         make_sockaddr(&peeraddr, 0, &len);
 894         if (addr_to_nodeid(&peeraddr, &nodeid, &mark)) {
 895                 unsigned char *b=(unsigned char *)&peeraddr;
 896                 log_print("connect from non cluster node");
 897                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 898                                      b, sizeof(struct sockaddr_storage));
 899                 sock_release(newsock);
 900                 return -1;
 901         }
 902
 903         log_print("got connection from %d", nodeid);
 904
 905         /*  Check to see if we already have a connection to this node. This
 906          *  could happen if the two nodes initiate a connection at roughly
 907          *  the same time and the connections cross on the wire.
 908          *  In this case we store the incoming one in "othercon"
 909          */
 910         newcon = nodeid2con(nodeid, GFP_NOFS);
 911         if (!newcon) {
 912                 result = -ENOMEM;
 913                 goto accept_err;
 914         }
 915
 916         sock_set_mark(newsock->sk, mark);
 917
 918         mutex_lock(&newcon->sock_mutex);
 919         if (newcon->sock) {
 920                 struct connection *othercon = newcon->othercon;
 921
 922                 if (!othercon) {
 923                         othercon = kzalloc(sizeof(*othercon), GFP_NOFS);
 924                         if (!othercon) {
 925                                 log_print("failed to allocate incoming socket");
 926                                 mutex_unlock(&newcon->sock_mutex);
 927                                 result = -ENOMEM;
 928                                 goto accept_err;
 929                         }
 930
 931                         result = dlm_con_init(othercon, nodeid);
 932                         if (result < 0) {
 933                                 kfree(othercon);
 934                                 goto accept_err;
 935                         }
 936
 937                         lockdep_set_subclass(&othercon->sock_mutex, 1);
 938                         newcon->othercon = othercon;
 939                 } else {
 940                         /* close other sock con if we have something new */
 941                         close_connection(othercon, false, true, false);
 942                 }
 943
 944                 mutex_lock(&othercon->sock_mutex);
 945                 add_sock(newsock, othercon);
 946                 addcon = othercon;
 947                 mutex_unlock(&othercon->sock_mutex);
 948         }
 949         else {
 950                 /* accept copies the sk after we've saved the callbacks, so we
 951                    don't want to save them a second time or comm errors will
 952                    result in calling sk_error_report recursively. */
 953                 add_sock(newsock, newcon);
 954                 addcon = newcon;
 955         }
 956
 957         set_bit(CF_CONNECTED, &addcon->flags);
 958         mutex_unlock(&newcon->sock_mutex);
 959
 960         /*
 961          * Add it to the active queue in case we got data
 962          * between processing the accept adding the socket
 963          * to the read_sockets list
 964          */
 965         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 966                 queue_work(recv_workqueue, &addcon->rwork);
 967
 968         return 0;
 969
 970 accept_err:
 971         if (newsock)
 972                 sock_release(newsock);
 973
 974         if (result != -EAGAIN)
 975                 log_print("error accepting connection from node: %d", result);
 976         return result;
 977 }
 978
 979 static void free_entry(struct writequeue_entry *e)
 980 {
 981         __free_page(e->page);
 982         kfree(e);
 983 }
 984
 985 /*
 986  * writequeue_entry_complete - try to delete and free write queue entry
 987  * @e: write queue entry to try to delete
 988  * @completed: bytes completed
 989  *
 990  * writequeue_lock must be held.
 991  */
 992 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 993 {
 994         e->offset += completed;
 995         e->len -= completed;
 996
 997         if (e->len == 0 && e->users == 0) {
 998                 list_del(&e->list);
 999                 free_entry(e);
1000         }
1001 }
1002
1003 /*
1004  * sctp_bind_addrs - bind a SCTP socket to all our addresses
1005  */
1006 static int sctp_bind_addrs(struct socket *sock, uint16_t port)
1007 {
1008         struct sockaddr_storage localaddr;
1009         struct sockaddr *addr = (struct sockaddr *)&localaddr;
1010         int i, addr_len, result = 0;
1011
1012         for (i = 0; i < dlm_local_count; i++) {
1013                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1014                 make_sockaddr(&localaddr, port, &addr_len);
1015
1016                 if (!i)
1017                         result = kernel_bind(sock, addr, addr_len);
1018                 else
1019                         result = sock_bind_add(sock->sk, addr, addr_len);
1020
1021                 if (result < 0) {
1022                         log_print("Can't bind to %d addr number %d, %d.\n",
1023                                   port, i + 1, result);
1024                         break;
1025                 }
1026         }
1027         return result;
1028 }
1029
1030 /* Initiate an SCTP association.
1031    This is a special case of send_to_sock() in that we don't yet have a
1032    peeled-off socket for this association, so we use the listening socket
1033    and add the primary IP address of the remote node.
1034  */
1035 static void sctp_connect_to_sock(struct connection *con)
1036 {
1037         struct sockaddr_storage daddr;
1038         int result;
1039         int addr_len;
1040         struct socket *sock;
1041         unsigned int mark;
1042
1043         mutex_lock(&con->sock_mutex);
1044
1045         /* Some odd races can cause double-connects, ignore them */
1046         if (con->retries++ > MAX_CONNECT_RETRIES)
1047                 goto out;
1048
1049         if (con->sock) {
1050                 log_print("node %d already connected.", con->nodeid);
1051                 goto out;
1052         }
1053
1054         memset(&daddr, 0, sizeof(daddr));
1055         result = nodeid_to_addr(con->nodeid, &daddr, NULL, true, &mark);
1056         if (result < 0) {
1057                 log_print("no address for nodeid %d", con->nodeid);
1058                 goto out;
1059         }
1060
1061         /* Create a socket to communicate with */
1062         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1063                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1064         if (result < 0)
1065                 goto socket_err;
1066
1067         sock_set_mark(sock->sk, mark);
1068
1069         add_sock(sock, con);
1070
1071         /* Bind to all addresses. */
1072         if (sctp_bind_addrs(con->sock, 0))
1073                 goto bind_err;
1074
1075         make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1076
1077         log_print("connecting to %d", con->nodeid);
1078
1079         /* Turn off Nagle's algorithm */
1080         sctp_sock_set_nodelay(sock->sk);
1081
1082         /*
1083          * Make sock->ops->connect() function return in specified time,
1084          * since O_NONBLOCK argument in connect() function does not work here,
1085          * then, we should restore the default value of this attribute.
1086          */
1087         sock_set_sndtimeo(sock->sk, 5);
1088         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1089                                    0);
1090         sock_set_sndtimeo(sock->sk, 0);
1091
1092         if (result == -EINPROGRESS)
1093                 result = 0;
1094         if (result == 0) {
1095                 if (!test_and_set_bit(CF_CONNECTED, &con->flags))
1096                         log_print("successful connected to node %d", con->nodeid);
1097                 goto out;
1098         }
1099
1100 bind_err:
1101         con->sock = NULL;
1102         sock_release(sock);
1103
1104 socket_err:
1105         /*
1106          * Some errors are fatal and this list might need adjusting. For other
1107          * errors we try again until the max number of retries is reached.
1108          */
1109         if (result != -EHOSTUNREACH &&
1110             result != -ENETUNREACH &&
1111             result != -ENETDOWN &&
1112             result != -EINVAL &&
1113             result != -EPROTONOSUPPORT) {
1114                 log_print("connect %d try %d error %d", con->nodeid,
1115                           con->retries, result);
1116                 mutex_unlock(&con->sock_mutex);
1117                 msleep(1000);
1118                 lowcomms_connect_sock(con);
1119                 return;
1120         }
1121
1122 out:
1123         mutex_unlock(&con->sock_mutex);
1124 }
1125
1126 /* Connect a new socket to its peer */
1127 static void tcp_connect_to_sock(struct connection *con)
1128 {
1129         struct sockaddr_storage saddr, src_addr;
1130         unsigned int mark;
1131         int addr_len;
1132         struct socket *sock = NULL;
1133         int result;
1134
1135         mutex_lock(&con->sock_mutex);
1136         if (con->retries++ > MAX_CONNECT_RETRIES)
1137                 goto out;
1138
1139         /* Some odd races can cause double-connects, ignore them */
1140         if (con->sock)
1141                 goto out;
1142
1143         /* Create a socket to communicate with */
1144         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1145                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1146         if (result < 0)
1147                 goto out_err;
1148
1149         memset(&saddr, 0, sizeof(saddr));
1150         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false, &mark);
1151         if (result < 0) {
1152                 log_print("no address for nodeid %d", con->nodeid);
1153                 goto out_err;
1154         }
1155
1156         sock_set_mark(sock->sk, mark);
1157
1158         add_sock(sock, con);
1159
1160         /* Bind to our cluster-known address connecting to avoid
1161            routing problems */
1162         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1163         make_sockaddr(&src_addr, 0, &addr_len);
1164         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1165                                  addr_len);
1166         if (result < 0) {
1167                 log_print("could not bind for connect: %d", result);
1168                 /* This *may* not indicate a critical error */
1169         }
1170
1171         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1172
1173         log_print("connecting to %d", con->nodeid);
1174
1175         /* Turn off Nagle's algorithm */
1176         tcp_sock_set_nodelay(sock->sk);
1177
1178         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1179                                    O_NONBLOCK);
1180         if (result == -EINPROGRESS)
1181                 result = 0;
1182         if (result == 0)
1183                 goto out;
1184
1185 out_err:
1186         if (con->sock) {
1187                 sock_release(con->sock);
1188                 con->sock = NULL;
1189         } else if (sock) {
1190                 sock_release(sock);
1191         }
1192         /*
1193          * Some errors are fatal and this list might need adjusting. For other
1194          * errors we try again until the max number of retries is reached.
1195          */
1196         if (result != -EHOSTUNREACH &&
1197             result != -ENETUNREACH &&
1198             result != -ENETDOWN &&
1199             result != -EINVAL &&
1200             result != -EPROTONOSUPPORT) {
1201                 log_print("connect %d try %d error %d", con->nodeid,
1202                           con->retries, result);
1203                 mutex_unlock(&con->sock_mutex);
1204                 msleep(1000);
1205                 lowcomms_connect_sock(con);
1206                 return;
1207         }
1208 out:
1209         mutex_unlock(&con->sock_mutex);
1210         return;
1211 }
1212
1213 /* On error caller must run dlm_close_sock() for the
1214  * listen connection socket.
1215  */
1216 static int tcp_create_listen_sock(struct listen_connection *con,
1217                                   struct sockaddr_storage *saddr)
1218 {
1219         struct socket *sock = NULL;
1220         int result = 0;
1221         int addr_len;
1222
1223         if (dlm_local_addr[0]->ss_family == AF_INET)
1224                 addr_len = sizeof(struct sockaddr_in);
1225         else
1226                 addr_len = sizeof(struct sockaddr_in6);
1227
1228         /* Create a socket to communicate with */
1229         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1230                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1231         if (result < 0) {
1232                 log_print("Can't create listening comms socket");
1233                 goto create_out;
1234         }
1235
1236         sock_set_mark(sock->sk, dlm_config.ci_mark);
1237
1238         /* Turn off Nagle's algorithm */
1239         tcp_sock_set_nodelay(sock->sk);
1240
1241         sock_set_reuseaddr(sock->sk);
1242
1243         add_listen_sock(sock, con);
1244
1245         /* Bind to our port */
1246         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1247         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1248         if (result < 0) {
1249                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1250                 goto create_out;
1251         }
1252         sock_set_keepalive(sock->sk);
1253
1254         result = sock->ops->listen(sock, 5);
1255         if (result < 0) {
1256                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1257                 goto create_out;
1258         }
1259
1260         return 0;
1261
1262 create_out:
1263         return result;
1264 }
1265
1266 /* Get local addresses */
1267 static void init_local(void)
1268 {
1269         struct sockaddr_storage sas, *addr;
1270         int i;
1271
1272         dlm_local_count = 0;
1273         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1274                 if (dlm_our_addr(&sas, i))
1275                         break;
1276
1277                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1278                 if (!addr)
1279                         break;
1280                 dlm_local_addr[dlm_local_count++] = addr;
1281         }
1282 }
1283
1284 static void deinit_local(void)
1285 {
1286         int i;
1287
1288         for (i = 0; i < dlm_local_count; i++)
1289                 kfree(dlm_local_addr[i]);
1290 }
1291
1292 /* Initialise SCTP socket and bind to all interfaces
1293  * On error caller must run dlm_close_sock() for the
1294  * listen connection socket.
1295  */
1296 static int sctp_listen_for_all(struct listen_connection *con)
1297 {
1298         struct socket *sock = NULL;
1299         int result = -EINVAL;
1300
1301         log_print("Using SCTP for communications");
1302
1303         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1304                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1305         if (result < 0) {
1306                 log_print("Can't create comms socket, check SCTP is loaded");
1307                 goto out;
1308         }
1309
1310         sock_set_rcvbuf(sock->sk, NEEDED_RMEM);
1311         sock_set_mark(sock->sk, dlm_config.ci_mark);
1312         sctp_sock_set_nodelay(sock->sk);
1313
1314         add_listen_sock(sock, con);
1315
1316         /* Bind to all addresses. */
1317         result = sctp_bind_addrs(con->sock, dlm_config.ci_tcp_port);
1318         if (result < 0)
1319                 goto out;
1320
1321         result = sock->ops->listen(sock, 5);
1322         if (result < 0) {
1323                 log_print("Can't set socket listening");
1324                 goto out;
1325         }
1326
1327         return 0;
1328
1329 out:
1330         return result;
1331 }
1332
1333 static int tcp_listen_for_all(void)
1334 {
1335         /* We don't support multi-homed hosts */
1336         if (dlm_local_count > 1) {
1337                 log_print("TCP protocol can't handle multi-homed hosts, "
1338                           "try SCTP");
1339                 return -EINVAL;
1340         }
1341
1342         log_print("Using TCP for communications");
1343
1344         return tcp_create_listen_sock(&listen_con, dlm_local_addr[0]);
1345 }
1346
1347
1348
1349 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1350                                                      gfp_t allocation)
1351 {
1352         struct writequeue_entry *entry;
1353
1354         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1355         if (!entry)
1356                 return NULL;
1357
1358         entry->page = alloc_page(allocation);
1359         if (!entry->page) {
1360                 kfree(entry);
1361                 return NULL;
1362         }
1363
1364         entry->offset = 0;
1365         entry->len = 0;
1366         entry->end = 0;
1367         entry->users = 0;
1368         entry->con = con;
1369
1370         return entry;
1371 }
1372
1373 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1374 {
1375         struct connection *con;
1376         struct writequeue_entry *e;
1377         int offset = 0;
1378
1379         if (len > LOWCOMMS_MAX_TX_BUFFER_LEN) {
1380                 BUILD_BUG_ON(PAGE_SIZE < LOWCOMMS_MAX_TX_BUFFER_LEN);
1381                 log_print("failed to allocate a buffer of size %d", len);
1382                 return NULL;
1383         }
1384
1385         con = nodeid2con(nodeid, allocation);
1386         if (!con)
1387                 return NULL;
1388
1389         spin_lock(&con->writequeue_lock);
1390         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1391         if ((&e->list == &con->writequeue) ||
1392             (PAGE_SIZE - e->end < len)) {
1393                 e = NULL;
1394         } else {
1395                 offset = e->end;
1396                 e->end += len;
1397                 e->users++;
1398         }
1399         spin_unlock(&con->writequeue_lock);
1400
1401         if (e) {
1402         got_one:
1403                 *ppc = page_address(e->page) + offset;
1404                 return e;
1405         }
1406
1407         e = new_writequeue_entry(con, allocation);
1408         if (e) {
1409                 spin_lock(&con->writequeue_lock);
1410                 offset = e->end;
1411                 e->end += len;
1412                 e->users++;
1413                 list_add_tail(&e->list, &con->writequeue);
1414                 spin_unlock(&con->writequeue_lock);
1415                 goto got_one;
1416         }
1417         return NULL;
1418 }
1419
1420 void dlm_lowcomms_commit_buffer(void *mh)
1421 {
1422         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1423         struct connection *con = e->con;
1424         int users;
1425
1426         spin_lock(&con->writequeue_lock);
1427         users = --e->users;
1428         if (users)
1429                 goto out;
1430         e->len = e->end - e->offset;
1431         spin_unlock(&con->writequeue_lock);
1432
1433         queue_work(send_workqueue, &con->swork);
1434         return;
1435
1436 out:
1437         spin_unlock(&con->writequeue_lock);
1438         return;
1439 }
1440
1441 /* Send a message */
1442 static void send_to_sock(struct connection *con)
1443 {
1444         int ret = 0;
1445         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1446         struct writequeue_entry *e;
1447         int len, offset;
1448         int count = 0;
1449
1450         mutex_lock(&con->sock_mutex);
1451         if (con->sock == NULL)
1452                 goto out_connect;
1453
1454         spin_lock(&con->writequeue_lock);
1455         for (;;) {
1456                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1457                                list);
1458                 if ((struct list_head *) e == &con->writequeue)
1459                         break;
1460
1461                 len = e->len;
1462                 offset = e->offset;
1463                 BUG_ON(len == 0 && e->users == 0);
1464                 spin_unlock(&con->writequeue_lock);
1465
1466                 ret = 0;
1467                 if (len) {
1468                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1469                                               msg_flags);
1470                         if (ret == -EAGAIN || ret == 0) {
1471                                 if (ret == -EAGAIN &&
1472                                     test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1473                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1474                                         /* Notify TCP that we're limited by the
1475                                          * application window size.
1476                                          */
1477                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1478                                         con->sock->sk->sk_write_pending++;
1479                                 }
1480                                 cond_resched();
1481                                 goto out;
1482                         } else if (ret < 0)
1483                                 goto send_error;
1484                 }
1485
1486                 /* Don't starve people filling buffers */
1487                 if (++count >= MAX_SEND_MSG_COUNT) {
1488                         cond_resched();
1489                         count = 0;
1490                 }
1491
1492                 spin_lock(&con->writequeue_lock);
1493                 writequeue_entry_complete(e, ret);
1494         }
1495         spin_unlock(&con->writequeue_lock);
1496 out:
1497         mutex_unlock(&con->sock_mutex);
1498         return;
1499
1500 send_error:
1501         mutex_unlock(&con->sock_mutex);
1502         close_connection(con, false, false, true);
1503         /* Requeue the send work. When the work daemon runs again, it will try
1504            a new connection, then call this function again. */
1505         queue_work(send_workqueue, &con->swork);
1506         return;
1507
1508 out_connect:
1509         mutex_unlock(&con->sock_mutex);
1510         queue_work(send_workqueue, &con->swork);
1511         cond_resched();
1512 }
1513
1514 static void clean_one_writequeue(struct connection *con)
1515 {
1516         struct writequeue_entry *e, *safe;
1517
1518         spin_lock(&con->writequeue_lock);
1519         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1520                 list_del(&e->list);
1521                 free_entry(e);
1522         }
1523         spin_unlock(&con->writequeue_lock);
1524 }
1525
1526 /* Called from recovery when it knows that a node has
1527    left the cluster */
1528 int dlm_lowcomms_close(int nodeid)
1529 {
1530         struct connection *con;
1531         struct dlm_node_addr *na;
1532
1533         log_print("closing connection to node %d", nodeid);
1534         con = nodeid2con(nodeid, 0);
1535         if (con) {
1536                 set_bit(CF_CLOSE, &con->flags);
1537                 close_connection(con, true, true, true);
1538                 clean_one_writequeue(con);
1539                 if (con->othercon)
1540                         clean_one_writequeue(con->othercon);
1541         }
1542
1543         spin_lock(&dlm_node_addrs_spin);
1544         na = find_node_addr(nodeid);
1545         if (na) {
1546                 list_del(&na->list);
1547                 while (na->addr_count--)
1548                         kfree(na->addr[na->addr_count]);
1549                 kfree(na);
1550         }
1551         spin_unlock(&dlm_node_addrs_spin);
1552
1553         return 0;
1554 }
1555
1556 /* Receive workqueue function */
1557 static void process_recv_sockets(struct work_struct *work)
1558 {
1559         struct connection *con = container_of(work, struct connection, rwork);
1560         int err;
1561
1562         clear_bit(CF_READ_PENDING, &con->flags);
1563         do {
1564                 err = receive_from_sock(con);
1565         } while (!err);
1566 }
1567
1568 static void process_listen_recv_socket(struct work_struct *work)
1569 {
1570         accept_from_sock(&listen_con);
1571 }
1572
1573 /* Send workqueue function */
1574 static void process_send_sockets(struct work_struct *work)
1575 {
1576         struct connection *con = container_of(work, struct connection, swork);
1577
1578         clear_bit(CF_WRITE_PENDING, &con->flags);
1579         if (con->sock == NULL) /* not mutex protected so check it inside too */
1580                 con->connect_action(con);
1581         if (!list_empty(&con->writequeue))
1582                 send_to_sock(con);
1583 }
1584
1585 static void work_stop(void)
1586 {
1587         if (recv_workqueue)
1588                 destroy_workqueue(recv_workqueue);
1589         if (send_workqueue)
1590                 destroy_workqueue(send_workqueue);
1591 }
1592
1593 static int work_start(void)
1594 {
1595         recv_workqueue = alloc_workqueue("dlm_recv",
1596                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1597         if (!recv_workqueue) {
1598                 log_print("can't start dlm_recv");
1599                 return -ENOMEM;
1600         }
1601
1602         send_workqueue = alloc_workqueue("dlm_send",
1603                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1604         if (!send_workqueue) {
1605                 log_print("can't start dlm_send");
1606                 destroy_workqueue(recv_workqueue);
1607                 return -ENOMEM;
1608         }
1609
1610         return 0;
1611 }
1612
1613 static void _stop_conn(struct connection *con, bool and_other)
1614 {
1615         mutex_lock(&con->sock_mutex);
1616         set_bit(CF_CLOSE, &con->flags);
1617         set_bit(CF_READ_PENDING, &con->flags);
1618         set_bit(CF_WRITE_PENDING, &con->flags);
1619         if (con->sock && con->sock->sk) {
1620                 write_lock_bh(&con->sock->sk->sk_callback_lock);
1621                 con->sock->sk->sk_user_data = NULL;
1622                 write_unlock_bh(&con->sock->sk->sk_callback_lock);
1623         }
1624         if (con->othercon && and_other)
1625                 _stop_conn(con->othercon, false);
1626         mutex_unlock(&con->sock_mutex);
1627 }
1628
1629 static void stop_conn(struct connection *con)
1630 {
1631         _stop_conn(con, true);
1632 }
1633
1634 static void shutdown_conn(struct connection *con)
1635 {
1636         if (con->shutdown_action)
1637                 con->shutdown_action(con);
1638 }
1639
1640 static void connection_release(struct rcu_head *rcu)
1641 {
1642         struct connection *con = container_of(rcu, struct connection, rcu);
1643
1644         kfree(con->rx_buf);
1645         kfree(con);
1646 }
1647
1648 static void free_conn(struct connection *con)
1649 {
1650         close_connection(con, true, true, true);
1651         spin_lock(&connections_lock);
1652         hlist_del_rcu(&con->list);
1653         spin_unlock(&connections_lock);
1654         if (con->othercon) {
1655                 clean_one_writequeue(con->othercon);
1656                 call_srcu(&connections_srcu, &con->othercon->rcu,
1657                           connection_release);
1658         }
1659         clean_one_writequeue(con);
1660         call_srcu(&connections_srcu, &con->rcu, connection_release);
1661 }
1662
1663 static void work_flush(void)
1664 {
1665         int ok, idx;
1666         int i;
1667         struct connection *con;
1668
1669         do {
1670                 ok = 1;
1671                 foreach_conn(stop_conn);
1672                 if (recv_workqueue)
1673                         flush_workqueue(recv_workqueue);
1674                 if (send_workqueue)
1675                         flush_workqueue(send_workqueue);
1676                 idx = srcu_read_lock(&connections_srcu);
1677                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1678                         hlist_for_each_entry_rcu(con, &connection_hash[i],
1679                                                  list) {
1680                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1681                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1682                                 if (con->othercon) {
1683                                         ok &= test_bit(CF_READ_PENDING,
1684                                                        &con->othercon->flags);
1685                                         ok &= test_bit(CF_WRITE_PENDING,
1686                                                        &con->othercon->flags);
1687                                 }
1688                         }
1689                 }
1690                 srcu_read_unlock(&connections_srcu, idx);
1691         } while (!ok);
1692 }
1693
1694 void dlm_lowcomms_stop(void)
1695 {
1696         /* Set all the flags to prevent any
1697            socket activity.
1698         */
1699         dlm_allow_conn = 0;
1700
1701         if (recv_workqueue)
1702                 flush_workqueue(recv_workqueue);
1703         if (send_workqueue)
1704                 flush_workqueue(send_workqueue);
1705
1706         dlm_close_sock(&listen_con.sock);
1707
1708         foreach_conn(shutdown_conn);
1709         work_flush();
1710         foreach_conn(free_conn);
1711         work_stop();
1712         deinit_local();
1713 }
1714
1715 int dlm_lowcomms_start(void)
1716 {
1717         int error = -EINVAL;
1718         int i;
1719
1720         for (i = 0; i < CONN_HASH_SIZE; i++)
1721                 INIT_HLIST_HEAD(&connection_hash[i]);
1722
1723         init_local();
1724         if (!dlm_local_count) {
1725                 error = -ENOTCONN;
1726                 log_print("no local IP address has been set");
1727                 goto fail;
1728         }
1729
1730         INIT_WORK(&listen_con.rwork, process_listen_recv_socket);
1731
1732         error = work_start();
1733         if (error)
1734                 goto fail;
1735
1736         dlm_allow_conn = 1;
1737
1738         /* Start listening */
1739         if (dlm_config.ci_protocol == 0)
1740                 error = tcp_listen_for_all();
1741         else
1742                 error = sctp_listen_for_all(&listen_con);
1743         if (error)
1744                 goto fail_unlisten;
1745
1746         return 0;
1747
1748 fail_unlisten:
1749         dlm_allow_conn = 0;
1750         dlm_close_sock(&listen_con.sock);
1751 fail:
1752         return error;
1753 }
1754
1755 void dlm_lowcomms_exit(void)
1756 {
1757         struct dlm_node_addr *na, *safe;
1758
1759         spin_lock(&dlm_node_addrs_spin);
1760         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1761                 list_del(&na->list);
1762                 while (na->addr_count--)
1763                         kfree(na->addr[na->addr_count]);
1764                 kfree(na);
1765         }
1766         spin_unlock(&dlm_node_addrs_spin);
1767 }