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