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 #define CF_RECONNECT 11
  83 #define CF_DELAY_CONNECT 12
  84 #define CF_EOF 13
  85         struct list_head writequeue;  /* List of outgoing writequeue_entries */
  86         spinlock_t writequeue_lock;
  87         atomic_t writequeue_cnt;
  88         void (*connect_action) (struct connection *);   /* What to do to connect */
  89         void (*shutdown_action)(struct connection *con); /* What to do to shutdown */
  90         bool (*eof_condition)(struct connection *con); /* What to do to eof check */
  91         int retries;
  92 #define MAX_CONNECT_RETRIES 3
  93         struct hlist_node list;
  94         struct connection *othercon;
  95         struct connection *sendcon;
  96         struct work_struct rwork; /* Receive workqueue */
  97         struct work_struct swork; /* Send workqueue */
  98         wait_queue_head_t shutdown_wait; /* wait for graceful shutdown */
  99         unsigned char *rx_buf;
 100         int rx_buflen;
 101         int rx_leftover;
 102         struct rcu_head rcu;
 103 };
 104 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 105
 106 struct listen_connection {
 107         struct socket *sock;
 108         struct work_struct rwork;
 109 };
 110
 111 #define DLM_WQ_REMAIN_BYTES(e) (PAGE_SIZE - e->end)
 112 #define DLM_WQ_LENGTH_BYTES(e) (e->end - e->offset)
 113
 114 /* An entry waiting to be sent */
 115 struct writequeue_entry {
 116         struct list_head list;
 117         struct page *page;
 118         int offset;
 119         int len;
 120         int end;
 121         int users;
 122         int idx; /* get()/commit() idx exchange */
 123         struct connection *con;
 124 };
 125
 126 struct dlm_node_addr {
 127         struct list_head list;
 128         int nodeid;
 129         int mark;
 130         int addr_count;
 131         int curr_addr_index;
 132         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 133 };
 134
 135 static struct listen_sock_callbacks {
 136         void (*sk_error_report)(struct sock *);
 137         void (*sk_data_ready)(struct sock *);
 138         void (*sk_state_change)(struct sock *);
 139         void (*sk_write_space)(struct sock *);
 140 } listen_sock;
 141
 142 static LIST_HEAD(dlm_node_addrs);
 143 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 144
 145 static struct listen_connection listen_con;
 146 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 147 static int dlm_local_count;
 148 int dlm_allow_conn;
 149
 150 /* Work queues */
 151 static struct workqueue_struct *recv_workqueue;
 152 static struct workqueue_struct *send_workqueue;
 153
 154 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 155 static DEFINE_SPINLOCK(connections_lock);
 156 DEFINE_STATIC_SRCU(connections_srcu);
 157
 158 static void process_recv_sockets(struct work_struct *work);
 159 static void process_send_sockets(struct work_struct *work);
 160
 161 static void sctp_connect_to_sock(struct connection *con);
 162 static void tcp_connect_to_sock(struct connection *con);
 163 static void dlm_tcp_shutdown(struct connection *con);
 164
 165 /* This is deliberately very simple because most clusters have simple
 166    sequential nodeids, so we should be able to go straight to a connection
 167    struct in the array */
 168 static inline int nodeid_hash(int nodeid)
 169 {
 170         return nodeid & (CONN_HASH_SIZE-1);
 171 }
 172
 173 static struct connection *__find_con(int nodeid, int r)
 174 {
 175         struct connection *con;
 176
 177         hlist_for_each_entry_rcu(con, &connection_hash[r], list) {
 178                 if (con->nodeid == nodeid)
 179                         return con;
 180         }
 181
 182         return NULL;
 183 }
 184
 185 static bool tcp_eof_condition(struct connection *con)
 186 {
 187         return atomic_read(&con->writequeue_cnt);
 188 }
 189
 190 static int dlm_con_init(struct connection *con, int nodeid)
 191 {
 192         con->rx_buflen = dlm_config.ci_buffer_size;
 193         con->rx_buf = kmalloc(con->rx_buflen, GFP_NOFS);
 194         if (!con->rx_buf)
 195                 return -ENOMEM;
 196
 197         con->nodeid = nodeid;
 198         mutex_init(&con->sock_mutex);
 199         INIT_LIST_HEAD(&con->writequeue);
 200         spin_lock_init(&con->writequeue_lock);
 201         atomic_set(&con->writequeue_cnt, 0);
 202         INIT_WORK(&con->swork, process_send_sockets);
 203         INIT_WORK(&con->rwork, process_recv_sockets);
 204         init_waitqueue_head(&con->shutdown_wait);
 205
 206         if (dlm_config.ci_protocol == 0) {
 207                 con->connect_action = tcp_connect_to_sock;
 208                 con->shutdown_action = dlm_tcp_shutdown;
 209                 con->eof_condition = tcp_eof_condition;
 210         } else {
 211                 con->connect_action = sctp_connect_to_sock;
 212         }
 213
 214         return 0;
 215 }
 216
 217 /*
 218  * If 'allocation' is zero then we don't attempt to create a new
 219  * connection structure for this node.
 220  */
 221 static struct connection *nodeid2con(int nodeid, gfp_t alloc)
 222 {
 223         struct connection *con, *tmp;
 224         int r, ret;
 225
 226         r = nodeid_hash(nodeid);
 227         con = __find_con(nodeid, r);
 228         if (con || !alloc)
 229                 return con;
 230
 231         con = kzalloc(sizeof(*con), alloc);
 232         if (!con)
 233                 return NULL;
 234
 235         ret = dlm_con_init(con, nodeid);
 236         if (ret) {
 237                 kfree(con);
 238                 return NULL;
 239         }
 240
 241         spin_lock(&connections_lock);
 242         /* Because multiple workqueues/threads calls this function it can
 243          * race on multiple cpu's. Instead of locking hot path __find_con()
 244          * we just check in rare cases of recently added nodes again
 245          * under protection of connections_lock. If this is the case we
 246          * abort our connection creation and return the existing connection.
 247          */
 248         tmp = __find_con(nodeid, r);
 249         if (tmp) {
 250                 spin_unlock(&connections_lock);
 251                 kfree(con->rx_buf);
 252                 kfree(con);
 253                 return tmp;
 254         }
 255
 256         hlist_add_head_rcu(&con->list, &connection_hash[r]);
 257         spin_unlock(&connections_lock);
 258
 259         return con;
 260 }
 261
 262 /* Loop round all connections */
 263 static void foreach_conn(void (*conn_func)(struct connection *c))
 264 {
 265         int i;
 266         struct connection *con;
 267
 268         for (i = 0; i < CONN_HASH_SIZE; i++) {
 269                 hlist_for_each_entry_rcu(con, &connection_hash[i], list)
 270                         conn_func(con);
 271         }
 272 }
 273
 274 static struct dlm_node_addr *find_node_addr(int nodeid)
 275 {
 276         struct dlm_node_addr *na;
 277
 278         list_for_each_entry(na, &dlm_node_addrs, list) {
 279                 if (na->nodeid == nodeid)
 280                         return na;
 281         }
 282         return NULL;
 283 }
 284
 285 static int addr_compare(const struct sockaddr_storage *x,
 286                         const struct sockaddr_storage *y)
 287 {
 288         switch (x->ss_family) {
 289         case AF_INET: {
 290                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 291                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 292                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 293                         return 0;
 294                 if (sinx->sin_port != siny->sin_port)
 295                         return 0;
 296                 break;
 297         }
 298         case AF_INET6: {
 299                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 300                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 301                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 302                         return 0;
 303                 if (sinx->sin6_port != siny->sin6_port)
 304                         return 0;
 305                 break;
 306         }
 307         default:
 308                 return 0;
 309         }
 310         return 1;
 311 }
 312
 313 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 314                           struct sockaddr *sa_out, bool try_new_addr,
 315                           unsigned int *mark)
 316 {
 317         struct sockaddr_storage sas;
 318         struct dlm_node_addr *na;
 319
 320         if (!dlm_local_count)
 321                 return -1;
 322
 323         spin_lock(&dlm_node_addrs_spin);
 324         na = find_node_addr(nodeid);
 325         if (na && na->addr_count) {
 326                 memcpy(&sas, na->addr[na->curr_addr_index],
 327                        sizeof(struct sockaddr_storage));
 328
 329                 if (try_new_addr) {
 330                         na->curr_addr_index++;
 331                         if (na->curr_addr_index == na->addr_count)
 332                                 na->curr_addr_index = 0;
 333                 }
 334         }
 335         spin_unlock(&dlm_node_addrs_spin);
 336
 337         if (!na)
 338                 return -EEXIST;
 339
 340         if (!na->addr_count)
 341                 return -ENOENT;
 342
 343         *mark = na->mark;
 344
 345         if (sas_out)
 346                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 347
 348         if (!sa_out)
 349                 return 0;
 350
 351         if (dlm_local_addr[0]->ss_family == AF_INET) {
 352                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 353                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 354                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 355         } else {
 356                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 357                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 358                 ret6->sin6_addr = in6->sin6_addr;
 359         }
 360
 361         return 0;
 362 }
 363
 364 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid,
 365                           unsigned int *mark)
 366 {
 367         struct dlm_node_addr *na;
 368         int rv = -EEXIST;
 369         int addr_i;
 370
 371         spin_lock(&dlm_node_addrs_spin);
 372         list_for_each_entry(na, &dlm_node_addrs, list) {
 373                 if (!na->addr_count)
 374                         continue;
 375
 376                 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
 377                         if (addr_compare(na->addr[addr_i], addr)) {
 378                                 *nodeid = na->nodeid;
 379                                 *mark = na->mark;
 380                                 rv = 0;
 381                                 goto unlock;
 382                         }
 383                 }
 384         }
 385 unlock:
 386         spin_unlock(&dlm_node_addrs_spin);
 387         return rv;
 388 }
 389
 390 /* caller need to held dlm_node_addrs_spin lock */
 391 static bool dlm_lowcomms_na_has_addr(const struct dlm_node_addr *na,
 392                                      const struct sockaddr_storage *addr)
 393 {
 394         int i;
 395
 396         for (i = 0; i < na->addr_count; i++) {
 397                 if (addr_compare(na->addr[i], addr))
 398                         return true;
 399         }
 400
 401         return false;
 402 }
 403
 404 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 405 {
 406         struct sockaddr_storage *new_addr;
 407         struct dlm_node_addr *new_node, *na;
 408         bool ret;
 409
 410         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 411         if (!new_node)
 412                 return -ENOMEM;
 413
 414         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 415         if (!new_addr) {
 416                 kfree(new_node);
 417                 return -ENOMEM;
 418         }
 419
 420         memcpy(new_addr, addr, len);
 421
 422         spin_lock(&dlm_node_addrs_spin);
 423         na = find_node_addr(nodeid);
 424         if (!na) {
 425                 new_node->nodeid = nodeid;
 426                 new_node->addr[0] = new_addr;
 427                 new_node->addr_count = 1;
 428                 new_node->mark = dlm_config.ci_mark;
 429                 list_add(&new_node->list, &dlm_node_addrs);
 430                 spin_unlock(&dlm_node_addrs_spin);
 431                 return 0;
 432         }
 433
 434         ret = dlm_lowcomms_na_has_addr(na, addr);
 435         if (ret) {
 436                 spin_unlock(&dlm_node_addrs_spin);
 437                 kfree(new_addr);
 438                 kfree(new_node);
 439                 return -EEXIST;
 440         }
 441
 442         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 443                 spin_unlock(&dlm_node_addrs_spin);
 444                 kfree(new_addr);
 445                 kfree(new_node);
 446                 return -ENOSPC;
 447         }
 448
 449         na->addr[na->addr_count++] = new_addr;
 450         spin_unlock(&dlm_node_addrs_spin);
 451         kfree(new_node);
 452         return 0;
 453 }
 454
 455 /* Data available on socket or listen socket received a connect */
 456 static void lowcomms_data_ready(struct sock *sk)
 457 {
 458         struct connection *con;
 459
 460         read_lock_bh(&sk->sk_callback_lock);
 461         con = sock2con(sk);
 462         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 463                 queue_work(recv_workqueue, &con->rwork);
 464         read_unlock_bh(&sk->sk_callback_lock);
 465 }
 466
 467 static void lowcomms_listen_data_ready(struct sock *sk)
 468 {
 469         queue_work(recv_workqueue, &listen_con.rwork);
 470 }
 471
 472 static void lowcomms_write_space(struct sock *sk)
 473 {
 474         struct connection *con;
 475
 476         read_lock_bh(&sk->sk_callback_lock);
 477         con = sock2con(sk);
 478         if (!con)
 479                 goto out;
 480
 481         if (!test_and_set_bit(CF_CONNECTED, &con->flags)) {
 482                 log_print("successful connected to node %d", con->nodeid);
 483                 queue_work(send_workqueue, &con->swork);
 484                 goto out;
 485         }
 486
 487         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 488
 489         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 490                 con->sock->sk->sk_write_pending--;
 491                 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
 492         }
 493
 494         queue_work(send_workqueue, &con->swork);
 495 out:
 496         read_unlock_bh(&sk->sk_callback_lock);
 497 }
 498
 499 static inline void lowcomms_connect_sock(struct connection *con)
 500 {
 501         if (test_bit(CF_CLOSE, &con->flags))
 502                 return;
 503         queue_work(send_workqueue, &con->swork);
 504         cond_resched();
 505 }
 506
 507 static void lowcomms_state_change(struct sock *sk)
 508 {
 509         /* SCTP layer is not calling sk_data_ready when the connection
 510          * is done, so we catch the signal through here. Also, it
 511          * doesn't switch socket state when entering shutdown, so we
 512          * skip the write in that case.
 513          */
 514         if (sk->sk_shutdown) {
 515                 if (sk->sk_shutdown == RCV_SHUTDOWN)
 516                         lowcomms_data_ready(sk);
 517         } else if (sk->sk_state == TCP_ESTABLISHED) {
 518                 lowcomms_write_space(sk);
 519         }
 520 }
 521
 522 int dlm_lowcomms_connect_node(int nodeid)
 523 {
 524         struct connection *con;
 525         int idx;
 526
 527         if (nodeid == dlm_our_nodeid())
 528                 return 0;
 529
 530         idx = srcu_read_lock(&connections_srcu);
 531         con = nodeid2con(nodeid, GFP_NOFS);
 532         if (!con) {
 533                 srcu_read_unlock(&connections_srcu, idx);
 534                 return -ENOMEM;
 535         }
 536
 537         lowcomms_connect_sock(con);
 538         srcu_read_unlock(&connections_srcu, idx);
 539
 540         return 0;
 541 }
 542
 543 int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark)
 544 {
 545         struct dlm_node_addr *na;
 546
 547         spin_lock(&dlm_node_addrs_spin);
 548         na = find_node_addr(nodeid);
 549         if (!na) {
 550                 spin_unlock(&dlm_node_addrs_spin);
 551                 return -ENOENT;
 552         }
 553
 554         na->mark = mark;
 555         spin_unlock(&dlm_node_addrs_spin);
 556
 557         return 0;
 558 }
 559
 560 static void lowcomms_error_report(struct sock *sk)
 561 {
 562         struct connection *con;
 563         struct sockaddr_storage saddr;
 564         void (*orig_report)(struct sock *) = NULL;
 565
 566         read_lock_bh(&sk->sk_callback_lock);
 567         con = sock2con(sk);
 568         if (con == NULL)
 569                 goto out;
 570
 571         orig_report = listen_sock.sk_error_report;
 572         if (con->sock == NULL ||
 573             kernel_getpeername(con->sock, (struct sockaddr *)&saddr) < 0) {
 574                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 575                                    "sending to node %d, port %d, "
 576                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 577                                    con->nodeid, dlm_config.ci_tcp_port,
 578                                    sk->sk_err, sk->sk_err_soft);
 579         } else if (saddr.ss_family == AF_INET) {
 580                 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
 581
 582                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 583                                    "sending to node %d at %pI4, port %d, "
 584                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 585                                    con->nodeid, &sin4->sin_addr.s_addr,
 586                                    dlm_config.ci_tcp_port, sk->sk_err,
 587                                    sk->sk_err_soft);
 588         } else {
 589                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
 590
 591                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 592                                    "sending to node %d at %u.%u.%u.%u, "
 593                                    "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 594                                    con->nodeid, sin6->sin6_addr.s6_addr32[0],
 595                                    sin6->sin6_addr.s6_addr32[1],
 596                                    sin6->sin6_addr.s6_addr32[2],
 597                                    sin6->sin6_addr.s6_addr32[3],
 598                                    dlm_config.ci_tcp_port, sk->sk_err,
 599                                    sk->sk_err_soft);
 600         }
 601
 602         /* below sendcon only handling */
 603         if (test_bit(CF_IS_OTHERCON, &con->flags))
 604                 con = con->sendcon;
 605
 606         switch (sk->sk_err) {
 607         case ECONNREFUSED:
 608                 set_bit(CF_DELAY_CONNECT, &con->flags);
 609                 break;
 610         default:
 611                 break;
 612         }
 613
 614         if (!test_and_set_bit(CF_RECONNECT, &con->flags))
 615                 queue_work(send_workqueue, &con->swork);
 616
 617 out:
 618         read_unlock_bh(&sk->sk_callback_lock);
 619         if (orig_report)
 620                 orig_report(sk);
 621 }
 622
 623 /* Note: sk_callback_lock must be locked before calling this function. */
 624 static void save_listen_callbacks(struct socket *sock)
 625 {
 626         struct sock *sk = sock->sk;
 627
 628         listen_sock.sk_data_ready = sk->sk_data_ready;
 629         listen_sock.sk_state_change = sk->sk_state_change;
 630         listen_sock.sk_write_space = sk->sk_write_space;
 631         listen_sock.sk_error_report = sk->sk_error_report;
 632 }
 633
 634 static void restore_callbacks(struct socket *sock)
 635 {
 636         struct sock *sk = sock->sk;
 637
 638         write_lock_bh(&sk->sk_callback_lock);
 639         sk->sk_user_data = NULL;
 640         sk->sk_data_ready = listen_sock.sk_data_ready;
 641         sk->sk_state_change = listen_sock.sk_state_change;
 642         sk->sk_write_space = listen_sock.sk_write_space;
 643         sk->sk_error_report = listen_sock.sk_error_report;
 644         write_unlock_bh(&sk->sk_callback_lock);
 645 }
 646
 647 static void add_listen_sock(struct socket *sock, struct listen_connection *con)
 648 {
 649         struct sock *sk = sock->sk;
 650
 651         write_lock_bh(&sk->sk_callback_lock);
 652         save_listen_callbacks(sock);
 653         con->sock = sock;
 654
 655         sk->sk_user_data = con;
 656         sk->sk_allocation = GFP_NOFS;
 657         /* Install a data_ready callback */
 658         sk->sk_data_ready = lowcomms_listen_data_ready;
 659         write_unlock_bh(&sk->sk_callback_lock);
 660 }
 661
 662 /* Make a socket active */
 663 static void add_sock(struct socket *sock, struct connection *con)
 664 {
 665         struct sock *sk = sock->sk;
 666
 667         write_lock_bh(&sk->sk_callback_lock);
 668         con->sock = sock;
 669
 670         sk->sk_user_data = con;
 671         /* Install a data_ready callback */
 672         sk->sk_data_ready = lowcomms_data_ready;
 673         sk->sk_write_space = lowcomms_write_space;
 674         sk->sk_state_change = lowcomms_state_change;
 675         sk->sk_allocation = GFP_NOFS;
 676         sk->sk_error_report = lowcomms_error_report;
 677         write_unlock_bh(&sk->sk_callback_lock);
 678 }
 679
 680 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 681    length */
 682 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 683                           int *addr_len)
 684 {
 685         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 686         if (saddr->ss_family == AF_INET) {
 687                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 688                 in4_addr->sin_port = cpu_to_be16(port);
 689                 *addr_len = sizeof(struct sockaddr_in);
 690                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 691         } else {
 692                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 693                 in6_addr->sin6_port = cpu_to_be16(port);
 694                 *addr_len = sizeof(struct sockaddr_in6);
 695         }
 696         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 697 }
 698
 699 static void dlm_close_sock(struct socket **sock)
 700 {
 701         if (*sock) {
 702                 restore_callbacks(*sock);
 703                 sock_release(*sock);
 704                 *sock = NULL;
 705         }
 706 }
 707
 708 /* Close a remote connection and tidy up */
 709 static void close_connection(struct connection *con, bool and_other,
 710                              bool tx, bool rx)
 711 {
 712         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 713
 714         if (tx && !closing && cancel_work_sync(&con->swork)) {
 715                 log_print("canceled swork for node %d", con->nodeid);
 716                 clear_bit(CF_WRITE_PENDING, &con->flags);
 717         }
 718         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 719                 log_print("canceled rwork for node %d", con->nodeid);
 720                 clear_bit(CF_READ_PENDING, &con->flags);
 721         }
 722
 723         mutex_lock(&con->sock_mutex);
 724         dlm_close_sock(&con->sock);
 725
 726         if (con->othercon && and_other) {
 727                 /* Will only re-enter once. */
 728                 close_connection(con->othercon, false, tx, rx);
 729         }
 730
 731         con->rx_leftover = 0;
 732         con->retries = 0;
 733         clear_bit(CF_CONNECTED, &con->flags);
 734         clear_bit(CF_DELAY_CONNECT, &con->flags);
 735         clear_bit(CF_RECONNECT, &con->flags);
 736         clear_bit(CF_EOF, &con->flags);
 737         mutex_unlock(&con->sock_mutex);
 738         clear_bit(CF_CLOSING, &con->flags);
 739 }
 740
 741 static void shutdown_connection(struct connection *con)
 742 {
 743         int ret;
 744
 745         flush_work(&con->swork);
 746
 747         mutex_lock(&con->sock_mutex);
 748         /* nothing to shutdown */
 749         if (!con->sock) {
 750                 mutex_unlock(&con->sock_mutex);
 751                 return;
 752         }
 753
 754         set_bit(CF_SHUTDOWN, &con->flags);
 755         ret = kernel_sock_shutdown(con->sock, SHUT_WR);
 756         mutex_unlock(&con->sock_mutex);
 757         if (ret) {
 758                 log_print("Connection %p failed to shutdown: %d will force close",
 759                           con, ret);
 760                 goto force_close;
 761         } else {
 762                 ret = wait_event_timeout(con->shutdown_wait,
 763                                          !test_bit(CF_SHUTDOWN, &con->flags),
 764                                          DLM_SHUTDOWN_WAIT_TIMEOUT);
 765                 if (ret == 0) {
 766                         log_print("Connection %p shutdown timed out, will force close",
 767                                   con);
 768                         goto force_close;
 769                 }
 770         }
 771
 772         return;
 773
 774 force_close:
 775         clear_bit(CF_SHUTDOWN, &con->flags);
 776         close_connection(con, false, true, true);
 777 }
 778
 779 static void dlm_tcp_shutdown(struct connection *con)
 780 {
 781         if (con->othercon)
 782                 shutdown_connection(con->othercon);
 783         shutdown_connection(con);
 784 }
 785
 786 static int con_realloc_receive_buf(struct connection *con, int newlen)
 787 {
 788         unsigned char *newbuf;
 789
 790         newbuf = kmalloc(newlen, GFP_NOFS);
 791         if (!newbuf)
 792                 return -ENOMEM;
 793
 794         /* copy any leftover from last receive */
 795         if (con->rx_leftover)
 796                 memmove(newbuf, con->rx_buf, con->rx_leftover);
 797
 798         /* swap to new buffer space */
 799         kfree(con->rx_buf);
 800         con->rx_buflen = newlen;
 801         con->rx_buf = newbuf;
 802
 803         return 0;
 804 }
 805
 806 /* Data received from remote end */
 807 static int receive_from_sock(struct connection *con)
 808 {
 809         int call_again_soon = 0;
 810         struct msghdr msg;
 811         struct kvec iov;
 812         int ret, buflen;
 813
 814         mutex_lock(&con->sock_mutex);
 815
 816         if (con->sock == NULL) {
 817                 ret = -EAGAIN;
 818                 goto out_close;
 819         }
 820
 821         /* realloc if we get new buffer size to read out */
 822         buflen = dlm_config.ci_buffer_size;
 823         if (con->rx_buflen != buflen && con->rx_leftover <= buflen) {
 824                 ret = con_realloc_receive_buf(con, buflen);
 825                 if (ret < 0)
 826                         goto out_resched;
 827         }
 828
 829         /* calculate new buffer parameter regarding last receive and
 830          * possible leftover bytes
 831          */
 832         iov.iov_base = con->rx_buf + con->rx_leftover;
 833         iov.iov_len = con->rx_buflen - con->rx_leftover;
 834
 835         memset(&msg, 0, sizeof(msg));
 836         msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
 837         ret = kernel_recvmsg(con->sock, &msg, &iov, 1, iov.iov_len,
 838                              msg.msg_flags);
 839         if (ret <= 0)
 840                 goto out_close;
 841         else if (ret == iov.iov_len)
 842                 call_again_soon = 1;
 843
 844         /* new buflen according readed bytes and leftover from last receive */
 845         buflen = ret + con->rx_leftover;
 846         ret = dlm_process_incoming_buffer(con->nodeid, con->rx_buf, buflen);
 847         if (ret < 0)
 848                 goto out_close;
 849
 850         /* calculate leftover bytes from process and put it into begin of
 851          * the receive buffer, so next receive we have the full message
 852          * at the start address of the receive buffer.
 853          */
 854         con->rx_leftover = buflen - ret;
 855         if (con->rx_leftover) {
 856                 memmove(con->rx_buf, con->rx_buf + ret,
 857                         con->rx_leftover);
 858                 call_again_soon = true;
 859         }
 860
 861         if (call_again_soon)
 862                 goto out_resched;
 863
 864         mutex_unlock(&con->sock_mutex);
 865         return 0;
 866
 867 out_resched:
 868         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 869                 queue_work(recv_workqueue, &con->rwork);
 870         mutex_unlock(&con->sock_mutex);
 871         return -EAGAIN;
 872
 873 out_close:
 874         if (ret == 0) {
 875                 log_print("connection %p got EOF from %d",
 876                           con, con->nodeid);
 877
 878                 if (con->eof_condition && con->eof_condition(con)) {
 879                         set_bit(CF_EOF, &con->flags);
 880                         mutex_unlock(&con->sock_mutex);
 881                 } else {
 882                         mutex_unlock(&con->sock_mutex);
 883                         close_connection(con, false, true, false);
 884
 885                         /* handling for tcp shutdown */
 886                         clear_bit(CF_SHUTDOWN, &con->flags);
 887                         wake_up(&con->shutdown_wait);
 888                 }
 889
 890                 /* signal to breaking receive worker */
 891                 ret = -1;
 892         } else {
 893                 mutex_unlock(&con->sock_mutex);
 894         }
 895         return ret;
 896 }
 897
 898 /* Listening socket is busy, accept a connection */
 899 static int accept_from_sock(struct listen_connection *con)
 900 {
 901         int result;
 902         struct sockaddr_storage peeraddr;
 903         struct socket *newsock;
 904         int len, idx;
 905         int nodeid;
 906         struct connection *newcon;
 907         struct connection *addcon;
 908         unsigned int mark;
 909
 910         if (!dlm_allow_conn) {
 911                 return -1;
 912         }
 913
 914         if (!con->sock)
 915                 return -ENOTCONN;
 916
 917         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 918         if (result < 0)
 919                 goto accept_err;
 920
 921         /* Get the connected socket's peer */
 922         memset(&peeraddr, 0, sizeof(peeraddr));
 923         len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
 924         if (len < 0) {
 925                 result = -ECONNABORTED;
 926                 goto accept_err;
 927         }
 928
 929         /* Get the new node's NODEID */
 930         make_sockaddr(&peeraddr, 0, &len);
 931         if (addr_to_nodeid(&peeraddr, &nodeid, &mark)) {
 932                 unsigned char *b=(unsigned char *)&peeraddr;
 933                 log_print("connect from non cluster node");
 934                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 935                                      b, sizeof(struct sockaddr_storage));
 936                 sock_release(newsock);
 937                 return -1;
 938         }
 939
 940         log_print("got connection from %d", nodeid);
 941
 942         /*  Check to see if we already have a connection to this node. This
 943          *  could happen if the two nodes initiate a connection at roughly
 944          *  the same time and the connections cross on the wire.
 945          *  In this case we store the incoming one in "othercon"
 946          */
 947         idx = srcu_read_lock(&connections_srcu);
 948         newcon = nodeid2con(nodeid, GFP_NOFS);
 949         if (!newcon) {
 950                 srcu_read_unlock(&connections_srcu, idx);
 951                 result = -ENOMEM;
 952                 goto accept_err;
 953         }
 954
 955         sock_set_mark(newsock->sk, mark);
 956
 957         mutex_lock(&newcon->sock_mutex);
 958         if (newcon->sock) {
 959                 struct connection *othercon = newcon->othercon;
 960
 961                 if (!othercon) {
 962                         othercon = kzalloc(sizeof(*othercon), GFP_NOFS);
 963                         if (!othercon) {
 964                                 log_print("failed to allocate incoming socket");
 965                                 mutex_unlock(&newcon->sock_mutex);
 966                                 srcu_read_unlock(&connections_srcu, idx);
 967                                 result = -ENOMEM;
 968                                 goto accept_err;
 969                         }
 970
 971                         result = dlm_con_init(othercon, nodeid);
 972                         if (result < 0) {
 973                                 kfree(othercon);
 974                                 mutex_unlock(&newcon->sock_mutex);
 975                                 srcu_read_unlock(&connections_srcu, idx);
 976                                 goto accept_err;
 977                         }
 978
 979                         lockdep_set_subclass(&othercon->sock_mutex, 1);
 980                         newcon->othercon = othercon;
 981                         othercon->sendcon = newcon;
 982                 } else {
 983                         /* close other sock con if we have something new */
 984                         close_connection(othercon, false, true, false);
 985                 }
 986
 987                 mutex_lock(&othercon->sock_mutex);
 988                 add_sock(newsock, othercon);
 989                 addcon = othercon;
 990                 mutex_unlock(&othercon->sock_mutex);
 991         }
 992         else {
 993                 /* accept copies the sk after we've saved the callbacks, so we
 994                    don't want to save them a second time or comm errors will
 995                    result in calling sk_error_report recursively. */
 996                 add_sock(newsock, newcon);
 997                 addcon = newcon;
 998         }
 999
1000         set_bit(CF_CONNECTED, &addcon->flags);
1001         mutex_unlock(&newcon->sock_mutex);
1002
1003         /*
1004          * Add it to the active queue in case we got data
1005          * between processing the accept adding the socket
1006          * to the read_sockets list
1007          */
1008         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
1009                 queue_work(recv_workqueue, &addcon->rwork);
1010
1011         srcu_read_unlock(&connections_srcu, idx);
1012
1013         return 0;
1014
1015 accept_err:
1016         if (newsock)
1017                 sock_release(newsock);
1018
1019         if (result != -EAGAIN)
1020                 log_print("error accepting connection from node: %d", result);
1021         return result;
1022 }
1023
1024 static void free_entry(struct writequeue_entry *e)
1025 {
1026         __free_page(e->page);
1027         kfree(e);
1028 }
1029
1030 /*
1031  * writequeue_entry_complete - try to delete and free write queue entry
1032  * @e: write queue entry to try to delete
1033  * @completed: bytes completed
1034  *
1035  * writequeue_lock must be held.
1036  */
1037 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1038 {
1039         e->offset += completed;
1040         e->len -= completed;
1041
1042         if (e->len == 0 && e->users == 0) {
1043                 list_del(&e->list);
1044                 atomic_dec(&e->con->writequeue_cnt);
1045                 free_entry(e);
1046         }
1047 }
1048
1049 /*
1050  * sctp_bind_addrs - bind a SCTP socket to all our addresses
1051  */
1052 static int sctp_bind_addrs(struct socket *sock, uint16_t port)
1053 {
1054         struct sockaddr_storage localaddr;
1055         struct sockaddr *addr = (struct sockaddr *)&localaddr;
1056         int i, addr_len, result = 0;
1057
1058         for (i = 0; i < dlm_local_count; i++) {
1059                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1060                 make_sockaddr(&localaddr, port, &addr_len);
1061
1062                 if (!i)
1063                         result = kernel_bind(sock, addr, addr_len);
1064                 else
1065                         result = sock_bind_add(sock->sk, addr, addr_len);
1066
1067                 if (result < 0) {
1068                         log_print("Can't bind to %d addr number %d, %d.\n",
1069                                   port, i + 1, result);
1070                         break;
1071                 }
1072         }
1073         return result;
1074 }
1075
1076 /* Initiate an SCTP association.
1077    This is a special case of send_to_sock() in that we don't yet have a
1078    peeled-off socket for this association, so we use the listening socket
1079    and add the primary IP address of the remote node.
1080  */
1081 static void sctp_connect_to_sock(struct connection *con)
1082 {
1083         struct sockaddr_storage daddr;
1084         int result;
1085         int addr_len;
1086         struct socket *sock;
1087         unsigned int mark;
1088
1089         mutex_lock(&con->sock_mutex);
1090
1091         /* Some odd races can cause double-connects, ignore them */
1092         if (con->retries++ > MAX_CONNECT_RETRIES)
1093                 goto out;
1094
1095         if (con->sock) {
1096                 log_print("node %d already connected.", con->nodeid);
1097                 goto out;
1098         }
1099
1100         memset(&daddr, 0, sizeof(daddr));
1101         result = nodeid_to_addr(con->nodeid, &daddr, NULL, true, &mark);
1102         if (result < 0) {
1103                 log_print("no address for nodeid %d", con->nodeid);
1104                 goto out;
1105         }
1106
1107         /* Create a socket to communicate with */
1108         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1109                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1110         if (result < 0)
1111                 goto socket_err;
1112
1113         sock_set_mark(sock->sk, mark);
1114
1115         add_sock(sock, con);
1116
1117         /* Bind to all addresses. */
1118         if (sctp_bind_addrs(con->sock, 0))
1119                 goto bind_err;
1120
1121         make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1122
1123         log_print("connecting to %d", con->nodeid);
1124
1125         /* Turn off Nagle's algorithm */
1126         sctp_sock_set_nodelay(sock->sk);
1127
1128         /*
1129          * Make sock->ops->connect() function return in specified time,
1130          * since O_NONBLOCK argument in connect() function does not work here,
1131          * then, we should restore the default value of this attribute.
1132          */
1133         sock_set_sndtimeo(sock->sk, 5);
1134         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1135                                    0);
1136         sock_set_sndtimeo(sock->sk, 0);
1137
1138         if (result == -EINPROGRESS)
1139                 result = 0;
1140         if (result == 0) {
1141                 if (!test_and_set_bit(CF_CONNECTED, &con->flags))
1142                         log_print("successful connected to node %d", con->nodeid);
1143                 goto out;
1144         }
1145
1146 bind_err:
1147         con->sock = NULL;
1148         sock_release(sock);
1149
1150 socket_err:
1151         /*
1152          * Some errors are fatal and this list might need adjusting. For other
1153          * errors we try again until the max number of retries is reached.
1154          */
1155         if (result != -EHOSTUNREACH &&
1156             result != -ENETUNREACH &&
1157             result != -ENETDOWN &&
1158             result != -EINVAL &&
1159             result != -EPROTONOSUPPORT) {
1160                 log_print("connect %d try %d error %d", con->nodeid,
1161                           con->retries, result);
1162                 mutex_unlock(&con->sock_mutex);
1163                 msleep(1000);
1164                 lowcomms_connect_sock(con);
1165                 return;
1166         }
1167
1168 out:
1169         mutex_unlock(&con->sock_mutex);
1170 }
1171
1172 /* Connect a new socket to its peer */
1173 static void tcp_connect_to_sock(struct connection *con)
1174 {
1175         struct sockaddr_storage saddr, src_addr;
1176         unsigned int mark;
1177         int addr_len;
1178         struct socket *sock = NULL;
1179         int result;
1180
1181         mutex_lock(&con->sock_mutex);
1182         if (con->retries++ > MAX_CONNECT_RETRIES)
1183                 goto out;
1184
1185         /* Some odd races can cause double-connects, ignore them */
1186         if (con->sock)
1187                 goto out;
1188
1189         /* Create a socket to communicate with */
1190         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1191                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1192         if (result < 0)
1193                 goto out_err;
1194
1195         memset(&saddr, 0, sizeof(saddr));
1196         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false, &mark);
1197         if (result < 0) {
1198                 log_print("no address for nodeid %d", con->nodeid);
1199                 goto out_err;
1200         }
1201
1202         sock_set_mark(sock->sk, mark);
1203
1204         add_sock(sock, con);
1205
1206         /* Bind to our cluster-known address connecting to avoid
1207            routing problems */
1208         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1209         make_sockaddr(&src_addr, 0, &addr_len);
1210         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1211                                  addr_len);
1212         if (result < 0) {
1213                 log_print("could not bind for connect: %d", result);
1214                 /* This *may* not indicate a critical error */
1215         }
1216
1217         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1218
1219         log_print("connecting to %d", con->nodeid);
1220
1221         /* Turn off Nagle's algorithm */
1222         tcp_sock_set_nodelay(sock->sk);
1223
1224         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1225                                    O_NONBLOCK);
1226         if (result == -EINPROGRESS)
1227                 result = 0;
1228         if (result == 0)
1229                 goto out;
1230
1231 out_err:
1232         if (con->sock) {
1233                 sock_release(con->sock);
1234                 con->sock = NULL;
1235         } else if (sock) {
1236                 sock_release(sock);
1237         }
1238         /*
1239          * Some errors are fatal and this list might need adjusting. For other
1240          * errors we try again until the max number of retries is reached.
1241          */
1242         if (result != -EHOSTUNREACH &&
1243             result != -ENETUNREACH &&
1244             result != -ENETDOWN &&
1245             result != -EINVAL &&
1246             result != -EPROTONOSUPPORT) {
1247                 log_print("connect %d try %d error %d", con->nodeid,
1248                           con->retries, result);
1249                 mutex_unlock(&con->sock_mutex);
1250                 msleep(1000);
1251                 lowcomms_connect_sock(con);
1252                 return;
1253         }
1254 out:
1255         mutex_unlock(&con->sock_mutex);
1256         return;
1257 }
1258
1259 /* On error caller must run dlm_close_sock() for the
1260  * listen connection socket.
1261  */
1262 static int tcp_create_listen_sock(struct listen_connection *con,
1263                                   struct sockaddr_storage *saddr)
1264 {
1265         struct socket *sock = NULL;
1266         int result = 0;
1267         int addr_len;
1268
1269         if (dlm_local_addr[0]->ss_family == AF_INET)
1270                 addr_len = sizeof(struct sockaddr_in);
1271         else
1272                 addr_len = sizeof(struct sockaddr_in6);
1273
1274         /* Create a socket to communicate with */
1275         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1276                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1277         if (result < 0) {
1278                 log_print("Can't create listening comms socket");
1279                 goto create_out;
1280         }
1281
1282         sock_set_mark(sock->sk, dlm_config.ci_mark);
1283
1284         /* Turn off Nagle's algorithm */
1285         tcp_sock_set_nodelay(sock->sk);
1286
1287         sock_set_reuseaddr(sock->sk);
1288
1289         add_listen_sock(sock, con);
1290
1291         /* Bind to our port */
1292         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1293         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1294         if (result < 0) {
1295                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1296                 goto create_out;
1297         }
1298         sock_set_keepalive(sock->sk);
1299
1300         result = sock->ops->listen(sock, 5);
1301         if (result < 0) {
1302                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1303                 goto create_out;
1304         }
1305
1306         return 0;
1307
1308 create_out:
1309         return result;
1310 }
1311
1312 /* Get local addresses */
1313 static void init_local(void)
1314 {
1315         struct sockaddr_storage sas, *addr;
1316         int i;
1317
1318         dlm_local_count = 0;
1319         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1320                 if (dlm_our_addr(&sas, i))
1321                         break;
1322
1323                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1324                 if (!addr)
1325                         break;
1326                 dlm_local_addr[dlm_local_count++] = addr;
1327         }
1328 }
1329
1330 static void deinit_local(void)
1331 {
1332         int i;
1333
1334         for (i = 0; i < dlm_local_count; i++)
1335                 kfree(dlm_local_addr[i]);
1336 }
1337
1338 /* Initialise SCTP socket and bind to all interfaces
1339  * On error caller must run dlm_close_sock() for the
1340  * listen connection socket.
1341  */
1342 static int sctp_listen_for_all(struct listen_connection *con)
1343 {
1344         struct socket *sock = NULL;
1345         int result = -EINVAL;
1346
1347         log_print("Using SCTP for communications");
1348
1349         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1350                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1351         if (result < 0) {
1352                 log_print("Can't create comms socket, check SCTP is loaded");
1353                 goto out;
1354         }
1355
1356         sock_set_rcvbuf(sock->sk, NEEDED_RMEM);
1357         sock_set_mark(sock->sk, dlm_config.ci_mark);
1358         sctp_sock_set_nodelay(sock->sk);
1359
1360         add_listen_sock(sock, con);
1361
1362         /* Bind to all addresses. */
1363         result = sctp_bind_addrs(con->sock, dlm_config.ci_tcp_port);
1364         if (result < 0)
1365                 goto out;
1366
1367         result = sock->ops->listen(sock, 5);
1368         if (result < 0) {
1369                 log_print("Can't set socket listening");
1370                 goto out;
1371         }
1372
1373         return 0;
1374
1375 out:
1376         return result;
1377 }
1378
1379 static int tcp_listen_for_all(void)
1380 {
1381         /* We don't support multi-homed hosts */
1382         if (dlm_local_count > 1) {
1383                 log_print("TCP protocol can't handle multi-homed hosts, "
1384                           "try SCTP");
1385                 return -EINVAL;
1386         }
1387
1388         log_print("Using TCP for communications");
1389
1390         return tcp_create_listen_sock(&listen_con, dlm_local_addr[0]);
1391 }
1392
1393
1394
1395 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1396                                                      gfp_t allocation)
1397 {
1398         struct writequeue_entry *entry;
1399
1400         entry = kzalloc(sizeof(*entry), allocation);
1401         if (!entry)
1402                 return NULL;
1403
1404         entry->page = alloc_page(allocation | __GFP_ZERO);
1405         if (!entry->page) {
1406                 kfree(entry);
1407                 return NULL;
1408         }
1409
1410         entry->con = con;
1411         entry->users = 1;
1412
1413         return entry;
1414 }
1415
1416 static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
1417                                              gfp_t allocation, char **ppc)
1418 {
1419         struct writequeue_entry *e;
1420
1421         spin_lock(&con->writequeue_lock);
1422         if (!list_empty(&con->writequeue)) {
1423                 e = list_last_entry(&con->writequeue, struct writequeue_entry, list);
1424                 if (DLM_WQ_REMAIN_BYTES(e) >= len) {
1425                         *ppc = page_address(e->page) + e->end;
1426                         e->end += len;
1427                         e->users++;
1428                         spin_unlock(&con->writequeue_lock);
1429
1430                         return e;
1431                 }
1432         }
1433         spin_unlock(&con->writequeue_lock);
1434
1435         e = new_writequeue_entry(con, allocation);
1436         if (!e)
1437                 return NULL;
1438
1439         *ppc = page_address(e->page);
1440         e->end += len;
1441         atomic_inc(&con->writequeue_cnt);
1442
1443         spin_lock(&con->writequeue_lock);
1444         list_add_tail(&e->list, &con->writequeue);
1445         spin_unlock(&con->writequeue_lock);
1446
1447         return e;
1448 };
1449
1450 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1451 {
1452         struct writequeue_entry *e;
1453         struct connection *con;
1454         int idx;
1455
1456         if (len > DEFAULT_BUFFER_SIZE ||
1457             len < sizeof(struct dlm_header)) {
1458                 BUILD_BUG_ON(PAGE_SIZE < DEFAULT_BUFFER_SIZE);
1459                 log_print("failed to allocate a buffer of size %d", len);
1460                 WARN_ON(1);
1461                 return NULL;
1462         }
1463
1464         idx = srcu_read_lock(&connections_srcu);
1465         con = nodeid2con(nodeid, allocation);
1466         if (!con) {
1467                 srcu_read_unlock(&connections_srcu, idx);
1468                 return NULL;
1469         }
1470
1471         e = new_wq_entry(con, len, allocation, ppc);
1472         if (!e) {
1473                 srcu_read_unlock(&connections_srcu, idx);
1474                 return NULL;
1475         }
1476
1477         /* we assume if successful commit must called */
1478         e->idx = idx;
1479
1480         return e;
1481 }
1482
1483 void dlm_lowcomms_commit_buffer(void *mh)
1484 {
1485         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1486         struct connection *con = e->con;
1487         int users;
1488
1489         spin_lock(&con->writequeue_lock);
1490         users = --e->users;
1491         if (users)
1492                 goto out;
1493
1494         e->len = DLM_WQ_LENGTH_BYTES(e);
1495         spin_unlock(&con->writequeue_lock);
1496
1497         queue_work(send_workqueue, &con->swork);
1498         srcu_read_unlock(&connections_srcu, e->idx);
1499         return;
1500
1501 out:
1502         spin_unlock(&con->writequeue_lock);
1503         srcu_read_unlock(&connections_srcu, e->idx);
1504         return;
1505 }
1506
1507 /* Send a message */
1508 static void send_to_sock(struct connection *con)
1509 {
1510         int ret = 0;
1511         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1512         struct writequeue_entry *e;
1513         int len, offset;
1514         int count = 0;
1515
1516         mutex_lock(&con->sock_mutex);
1517         if (con->sock == NULL)
1518                 goto out_connect;
1519
1520         spin_lock(&con->writequeue_lock);
1521         for (;;) {
1522                 if (list_empty(&con->writequeue))
1523                         break;
1524
1525                 e = list_first_entry(&con->writequeue, struct writequeue_entry, list);
1526                 len = e->len;
1527                 offset = e->offset;
1528                 BUG_ON(len == 0 && e->users == 0);
1529                 spin_unlock(&con->writequeue_lock);
1530
1531                 ret = 0;
1532                 if (len) {
1533                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1534                                               msg_flags);
1535                         if (ret == -EAGAIN || ret == 0) {
1536                                 if (ret == -EAGAIN &&
1537                                     test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1538                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1539                                         /* Notify TCP that we're limited by the
1540                                          * application window size.
1541                                          */
1542                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1543                                         con->sock->sk->sk_write_pending++;
1544                                 }
1545                                 cond_resched();
1546                                 goto out;
1547                         } else if (ret < 0)
1548                                 goto out;
1549                 }
1550
1551                 /* Don't starve people filling buffers */
1552                 if (++count >= MAX_SEND_MSG_COUNT) {
1553                         cond_resched();
1554                         count = 0;
1555                 }
1556
1557                 spin_lock(&con->writequeue_lock);
1558                 writequeue_entry_complete(e, ret);
1559         }
1560         spin_unlock(&con->writequeue_lock);
1561
1562         /* close if we got EOF */
1563         if (test_and_clear_bit(CF_EOF, &con->flags)) {
1564                 mutex_unlock(&con->sock_mutex);
1565                 close_connection(con, false, false, true);
1566
1567                 /* handling for tcp shutdown */
1568                 clear_bit(CF_SHUTDOWN, &con->flags);
1569                 wake_up(&con->shutdown_wait);
1570         } else {
1571                 mutex_unlock(&con->sock_mutex);
1572         }
1573
1574         return;
1575
1576 out:
1577         mutex_unlock(&con->sock_mutex);
1578         return;
1579
1580 out_connect:
1581         mutex_unlock(&con->sock_mutex);
1582         queue_work(send_workqueue, &con->swork);
1583         cond_resched();
1584 }
1585
1586 static void clean_one_writequeue(struct connection *con)
1587 {
1588         struct writequeue_entry *e, *safe;
1589
1590         spin_lock(&con->writequeue_lock);
1591         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1592                 list_del(&e->list);
1593                 free_entry(e);
1594         }
1595         spin_unlock(&con->writequeue_lock);
1596 }
1597
1598 /* Called from recovery when it knows that a node has
1599    left the cluster */
1600 int dlm_lowcomms_close(int nodeid)
1601 {
1602         struct connection *con;
1603         struct dlm_node_addr *na;
1604         int idx;
1605
1606         log_print("closing connection to node %d", nodeid);
1607         idx = srcu_read_lock(&connections_srcu);
1608         con = nodeid2con(nodeid, 0);
1609         if (con) {
1610                 set_bit(CF_CLOSE, &con->flags);
1611                 close_connection(con, true, true, true);
1612                 clean_one_writequeue(con);
1613                 if (con->othercon)
1614                         clean_one_writequeue(con->othercon);
1615         }
1616         srcu_read_unlock(&connections_srcu, idx);
1617
1618         spin_lock(&dlm_node_addrs_spin);
1619         na = find_node_addr(nodeid);
1620         if (na) {
1621                 list_del(&na->list);
1622                 while (na->addr_count--)
1623                         kfree(na->addr[na->addr_count]);
1624                 kfree(na);
1625         }
1626         spin_unlock(&dlm_node_addrs_spin);
1627
1628         return 0;
1629 }
1630
1631 /* Receive workqueue function */
1632 static void process_recv_sockets(struct work_struct *work)
1633 {
1634         struct connection *con = container_of(work, struct connection, rwork);
1635         int err;
1636
1637         clear_bit(CF_READ_PENDING, &con->flags);
1638         do {
1639                 err = receive_from_sock(con);
1640         } while (!err);
1641 }
1642
1643 static void process_listen_recv_socket(struct work_struct *work)
1644 {
1645         accept_from_sock(&listen_con);
1646 }
1647
1648 /* Send workqueue function */
1649 static void process_send_sockets(struct work_struct *work)
1650 {
1651         struct connection *con = container_of(work, struct connection, swork);
1652
1653         clear_bit(CF_WRITE_PENDING, &con->flags);
1654
1655         if (test_and_clear_bit(CF_RECONNECT, &con->flags))
1656                 close_connection(con, false, false, true);
1657
1658         if (con->sock == NULL) { /* not mutex protected so check it inside too */
1659                 if (test_and_clear_bit(CF_DELAY_CONNECT, &con->flags))
1660                         msleep(1000);
1661                 con->connect_action(con);
1662         }
1663         if (!list_empty(&con->writequeue))
1664                 send_to_sock(con);
1665 }
1666
1667 static void work_stop(void)
1668 {
1669         if (recv_workqueue) {
1670                 destroy_workqueue(recv_workqueue);
1671                 recv_workqueue = NULL;
1672         }
1673
1674         if (send_workqueue) {
1675                 destroy_workqueue(send_workqueue);
1676                 send_workqueue = NULL;
1677         }
1678 }
1679
1680 static int work_start(void)
1681 {
1682         recv_workqueue = alloc_workqueue("dlm_recv",
1683                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1684         if (!recv_workqueue) {
1685                 log_print("can't start dlm_recv");
1686                 return -ENOMEM;
1687         }
1688
1689         send_workqueue = alloc_workqueue("dlm_send",
1690                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1691         if (!send_workqueue) {
1692                 log_print("can't start dlm_send");
1693                 destroy_workqueue(recv_workqueue);
1694                 recv_workqueue = NULL;
1695                 return -ENOMEM;
1696         }
1697
1698         return 0;
1699 }
1700
1701 static void shutdown_conn(struct connection *con)
1702 {
1703         if (con->shutdown_action)
1704                 con->shutdown_action(con);
1705 }
1706
1707 void dlm_lowcomms_shutdown(void)
1708 {
1709         int idx;
1710
1711         /* Set all the flags to prevent any
1712          * socket activity.
1713          */
1714         dlm_allow_conn = 0;
1715
1716         if (recv_workqueue)
1717                 flush_workqueue(recv_workqueue);
1718         if (send_workqueue)
1719                 flush_workqueue(send_workqueue);
1720
1721         dlm_close_sock(&listen_con.sock);
1722
1723         idx = srcu_read_lock(&connections_srcu);
1724         foreach_conn(shutdown_conn);
1725         srcu_read_unlock(&connections_srcu, idx);
1726 }
1727
1728 static void _stop_conn(struct connection *con, bool and_other)
1729 {
1730         mutex_lock(&con->sock_mutex);
1731         set_bit(CF_CLOSE, &con->flags);
1732         set_bit(CF_READ_PENDING, &con->flags);
1733         set_bit(CF_WRITE_PENDING, &con->flags);
1734         if (con->sock && con->sock->sk) {
1735                 write_lock_bh(&con->sock->sk->sk_callback_lock);
1736                 con->sock->sk->sk_user_data = NULL;
1737                 write_unlock_bh(&con->sock->sk->sk_callback_lock);
1738         }
1739         if (con->othercon && and_other)
1740                 _stop_conn(con->othercon, false);
1741         mutex_unlock(&con->sock_mutex);
1742 }
1743
1744 static void stop_conn(struct connection *con)
1745 {
1746         _stop_conn(con, true);
1747 }
1748
1749 static void connection_release(struct rcu_head *rcu)
1750 {
1751         struct connection *con = container_of(rcu, struct connection, rcu);
1752
1753         kfree(con->rx_buf);
1754         kfree(con);
1755 }
1756
1757 static void free_conn(struct connection *con)
1758 {
1759         close_connection(con, true, true, true);
1760         spin_lock(&connections_lock);
1761         hlist_del_rcu(&con->list);
1762         spin_unlock(&connections_lock);
1763         if (con->othercon) {
1764                 clean_one_writequeue(con->othercon);
1765                 call_srcu(&connections_srcu, &con->othercon->rcu,
1766                           connection_release);
1767         }
1768         clean_one_writequeue(con);
1769         call_srcu(&connections_srcu, &con->rcu, connection_release);
1770 }
1771
1772 static void work_flush(void)
1773 {
1774         int ok;
1775         int i;
1776         struct connection *con;
1777
1778         do {
1779                 ok = 1;
1780                 foreach_conn(stop_conn);
1781                 if (recv_workqueue)
1782                         flush_workqueue(recv_workqueue);
1783                 if (send_workqueue)
1784                         flush_workqueue(send_workqueue);
1785                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1786                         hlist_for_each_entry_rcu(con, &connection_hash[i],
1787                                                  list) {
1788                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1789                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1790                                 if (con->othercon) {
1791                                         ok &= test_bit(CF_READ_PENDING,
1792                                                        &con->othercon->flags);
1793                                         ok &= test_bit(CF_WRITE_PENDING,
1794                                                        &con->othercon->flags);
1795                                 }
1796                         }
1797                 }
1798         } while (!ok);
1799 }
1800
1801 void dlm_lowcomms_stop(void)
1802 {
1803         int idx;
1804
1805         idx = srcu_read_lock(&connections_srcu);
1806         work_flush();
1807         foreach_conn(free_conn);
1808         srcu_read_unlock(&connections_srcu, idx);
1809         work_stop();
1810         deinit_local();
1811 }
1812
1813 int dlm_lowcomms_start(void)
1814 {
1815         int error = -EINVAL;
1816         int i;
1817
1818         for (i = 0; i < CONN_HASH_SIZE; i++)
1819                 INIT_HLIST_HEAD(&connection_hash[i]);
1820
1821         init_local();
1822         if (!dlm_local_count) {
1823                 error = -ENOTCONN;
1824                 log_print("no local IP address has been set");
1825                 goto fail;
1826         }
1827
1828         INIT_WORK(&listen_con.rwork, process_listen_recv_socket);
1829
1830         error = work_start();
1831         if (error)
1832                 goto fail_local;
1833
1834         dlm_allow_conn = 1;
1835
1836         /* Start listening */
1837         if (dlm_config.ci_protocol == 0)
1838                 error = tcp_listen_for_all();
1839         else
1840                 error = sctp_listen_for_all(&listen_con);
1841         if (error)
1842                 goto fail_unlisten;
1843
1844         return 0;
1845
1846 fail_unlisten:
1847         dlm_allow_conn = 0;
1848         dlm_close_sock(&listen_con.sock);
1849         work_stop();
1850 fail_local:
1851         deinit_local();
1852 fail:
1853         return error;
1854 }
1855
1856 void dlm_lowcomms_exit(void)
1857 {
1858         struct dlm_node_addr *na, *safe;
1859
1860         spin_lock(&dlm_node_addrs_spin);
1861         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1862                 list_del(&na->list);
1863                 while (na->addr_count--)
1864                         kfree(na->addr[na->addr_count]);
1865                 kfree(na);
1866         }
1867         spin_unlock(&dlm_node_addrs_spin);
1868 }