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
  67 struct cbuf {
  68         unsigned int base;
  69         unsigned int len;
  70         unsigned int mask;
  71 };
  72
  73 static void cbuf_add(struct cbuf *cb, int n)
  74 {
  75         cb->len += n;
  76 }
  77
  78 static int cbuf_data(struct cbuf *cb)
  79 {
  80         return ((cb->base + cb->len) & cb->mask);
  81 }
  82
  83 static void cbuf_init(struct cbuf *cb, int size)
  84 {
  85         cb->base = cb->len = 0;
  86         cb->mask = size-1;
  87 }
  88
  89 static void cbuf_eat(struct cbuf *cb, int n)
  90 {
  91         cb->len  -= n;
  92         cb->base += n;
  93         cb->base &= cb->mask;
  94 }
  95
  96 static bool cbuf_empty(struct cbuf *cb)
  97 {
  98         return cb->len == 0;
  99 }
 100
 101 struct connection {
 102         struct socket *sock;    /* NULL if not connected */
 103         uint32_t nodeid;        /* So we know who we are in the list */
 104         struct mutex sock_mutex;
 105         unsigned long flags;
 106 #define CF_READ_PENDING 1
 107 #define CF_WRITE_PENDING 2
 108 #define CF_INIT_PENDING 4
 109 #define CF_IS_OTHERCON 5
 110 #define CF_CLOSE 6
 111 #define CF_APP_LIMITED 7
 112 #define CF_CLOSING 8
 113         struct list_head writequeue;  /* List of outgoing writequeue_entries */
 114         spinlock_t writequeue_lock;
 115         int (*rx_action) (struct connection *); /* What to do when active */
 116         void (*connect_action) (struct connection *);   /* What to do to connect */
 117         struct page *rx_page;
 118         struct cbuf cb;
 119         int retries;
 120 #define MAX_CONNECT_RETRIES 3
 121         struct hlist_node list;
 122         struct connection *othercon;
 123         struct work_struct rwork; /* Receive workqueue */
 124         struct work_struct swork; /* Send workqueue */
 125 };
 126 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 127
 128 /* An entry waiting to be sent */
 129 struct writequeue_entry {
 130         struct list_head list;
 131         struct page *page;
 132         int offset;
 133         int len;
 134         int end;
 135         int users;
 136         struct connection *con;
 137 };
 138
 139 struct dlm_node_addr {
 140         struct list_head list;
 141         int nodeid;
 142         int addr_count;
 143         int curr_addr_index;
 144         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 145 };
 146
 147 static struct listen_sock_callbacks {
 148         void (*sk_error_report)(struct sock *);
 149         void (*sk_data_ready)(struct sock *);
 150         void (*sk_state_change)(struct sock *);
 151         void (*sk_write_space)(struct sock *);
 152 } listen_sock;
 153
 154 static LIST_HEAD(dlm_node_addrs);
 155 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 156
 157 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 158 static int dlm_local_count;
 159 static int dlm_allow_conn;
 160
 161 /* Work queues */
 162 static struct workqueue_struct *recv_workqueue;
 163 static struct workqueue_struct *send_workqueue;
 164
 165 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 166 static DEFINE_MUTEX(connections_lock);
 167 static struct kmem_cache *con_cache;
 168
 169 static void process_recv_sockets(struct work_struct *work);
 170 static void process_send_sockets(struct work_struct *work);
 171
 172
 173 /* This is deliberately very simple because most clusters have simple
 174    sequential nodeids, so we should be able to go straight to a connection
 175    struct in the array */
 176 static inline int nodeid_hash(int nodeid)
 177 {
 178         return nodeid & (CONN_HASH_SIZE-1);
 179 }
 180
 181 static struct connection *__find_con(int nodeid)
 182 {
 183         int r;
 184         struct connection *con;
 185
 186         r = nodeid_hash(nodeid);
 187
 188         hlist_for_each_entry(con, &connection_hash[r], list) {
 189                 if (con->nodeid == nodeid)
 190                         return con;
 191         }
 192         return NULL;
 193 }
 194
 195 /*
 196  * If 'allocation' is zero then we don't attempt to create a new
 197  * connection structure for this node.
 198  */
 199 static struct connection *__nodeid2con(int nodeid, gfp_t alloc)
 200 {
 201         struct connection *con = NULL;
 202         int r;
 203
 204         con = __find_con(nodeid);
 205         if (con || !alloc)
 206                 return con;
 207
 208         con = kmem_cache_zalloc(con_cache, alloc);
 209         if (!con)
 210                 return NULL;
 211
 212         r = nodeid_hash(nodeid);
 213         hlist_add_head(&con->list, &connection_hash[r]);
 214
 215         con->nodeid = nodeid;
 216         mutex_init(&con->sock_mutex);
 217         INIT_LIST_HEAD(&con->writequeue);
 218         spin_lock_init(&con->writequeue_lock);
 219         INIT_WORK(&con->swork, process_send_sockets);
 220         INIT_WORK(&con->rwork, process_recv_sockets);
 221
 222         /* Setup action pointers for child sockets */
 223         if (con->nodeid) {
 224                 struct connection *zerocon = __find_con(0);
 225
 226                 con->connect_action = zerocon->connect_action;
 227                 if (!con->rx_action)
 228                         con->rx_action = zerocon->rx_action;
 229         }
 230
 231         return con;
 232 }
 233
 234 /* Loop round all connections */
 235 static void foreach_conn(void (*conn_func)(struct connection *c))
 236 {
 237         int i;
 238         struct hlist_node *n;
 239         struct connection *con;
 240
 241         for (i = 0; i < CONN_HASH_SIZE; i++) {
 242                 hlist_for_each_entry_safe(con, n, &connection_hash[i], list)
 243                         conn_func(con);
 244         }
 245 }
 246
 247 static struct connection *nodeid2con(int nodeid, gfp_t allocation)
 248 {
 249         struct connection *con;
 250
 251         mutex_lock(&connections_lock);
 252         con = __nodeid2con(nodeid, allocation);
 253         mutex_unlock(&connections_lock);
 254
 255         return con;
 256 }
 257
 258 static struct dlm_node_addr *find_node_addr(int nodeid)
 259 {
 260         struct dlm_node_addr *na;
 261
 262         list_for_each_entry(na, &dlm_node_addrs, list) {
 263                 if (na->nodeid == nodeid)
 264                         return na;
 265         }
 266         return NULL;
 267 }
 268
 269 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y)
 270 {
 271         switch (x->ss_family) {
 272         case AF_INET: {
 273                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 274                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 275                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 276                         return 0;
 277                 if (sinx->sin_port != siny->sin_port)
 278                         return 0;
 279                 break;
 280         }
 281         case AF_INET6: {
 282                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 283                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 284                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 285                         return 0;
 286                 if (sinx->sin6_port != siny->sin6_port)
 287                         return 0;
 288                 break;
 289         }
 290         default:
 291                 return 0;
 292         }
 293         return 1;
 294 }
 295
 296 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 297                           struct sockaddr *sa_out, bool try_new_addr)
 298 {
 299         struct sockaddr_storage sas;
 300         struct dlm_node_addr *na;
 301
 302         if (!dlm_local_count)
 303                 return -1;
 304
 305         spin_lock(&dlm_node_addrs_spin);
 306         na = find_node_addr(nodeid);
 307         if (na && na->addr_count) {
 308                 memcpy(&sas, na->addr[na->curr_addr_index],
 309                        sizeof(struct sockaddr_storage));
 310
 311                 if (try_new_addr) {
 312                         na->curr_addr_index++;
 313                         if (na->curr_addr_index == na->addr_count)
 314                                 na->curr_addr_index = 0;
 315                 }
 316         }
 317         spin_unlock(&dlm_node_addrs_spin);
 318
 319         if (!na)
 320                 return -EEXIST;
 321
 322         if (!na->addr_count)
 323                 return -ENOENT;
 324
 325         if (sas_out)
 326                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 327
 328         if (!sa_out)
 329                 return 0;
 330
 331         if (dlm_local_addr[0]->ss_family == AF_INET) {
 332                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 333                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 334                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 335         } else {
 336                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 337                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 338                 ret6->sin6_addr = in6->sin6_addr;
 339         }
 340
 341         return 0;
 342 }
 343
 344 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid)
 345 {
 346         struct dlm_node_addr *na;
 347         int rv = -EEXIST;
 348         int addr_i;
 349
 350         spin_lock(&dlm_node_addrs_spin);
 351         list_for_each_entry(na, &dlm_node_addrs, list) {
 352                 if (!na->addr_count)
 353                         continue;
 354
 355                 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
 356                         if (addr_compare(na->addr[addr_i], addr)) {
 357                                 *nodeid = na->nodeid;
 358                                 rv = 0;
 359                                 goto unlock;
 360                         }
 361                 }
 362         }
 363 unlock:
 364         spin_unlock(&dlm_node_addrs_spin);
 365         return rv;
 366 }
 367
 368 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 369 {
 370         struct sockaddr_storage *new_addr;
 371         struct dlm_node_addr *new_node, *na;
 372
 373         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 374         if (!new_node)
 375                 return -ENOMEM;
 376
 377         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 378         if (!new_addr) {
 379                 kfree(new_node);
 380                 return -ENOMEM;
 381         }
 382
 383         memcpy(new_addr, addr, len);
 384
 385         spin_lock(&dlm_node_addrs_spin);
 386         na = find_node_addr(nodeid);
 387         if (!na) {
 388                 new_node->nodeid = nodeid;
 389                 new_node->addr[0] = new_addr;
 390                 new_node->addr_count = 1;
 391                 list_add(&new_node->list, &dlm_node_addrs);
 392                 spin_unlock(&dlm_node_addrs_spin);
 393                 return 0;
 394         }
 395
 396         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 397                 spin_unlock(&dlm_node_addrs_spin);
 398                 kfree(new_addr);
 399                 kfree(new_node);
 400                 return -ENOSPC;
 401         }
 402
 403         na->addr[na->addr_count++] = new_addr;
 404         spin_unlock(&dlm_node_addrs_spin);
 405         kfree(new_node);
 406         return 0;
 407 }
 408
 409 /* Data available on socket or listen socket received a connect */
 410 static void lowcomms_data_ready(struct sock *sk)
 411 {
 412         struct connection *con;
 413
 414         read_lock_bh(&sk->sk_callback_lock);
 415         con = sock2con(sk);
 416         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 417                 queue_work(recv_workqueue, &con->rwork);
 418         read_unlock_bh(&sk->sk_callback_lock);
 419 }
 420
 421 static void lowcomms_write_space(struct sock *sk)
 422 {
 423         struct connection *con;
 424
 425         read_lock_bh(&sk->sk_callback_lock);
 426         con = sock2con(sk);
 427         if (!con)
 428                 goto out;
 429
 430         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 431
 432         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 433                 con->sock->sk->sk_write_pending--;
 434                 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
 435         }
 436
 437         queue_work(send_workqueue, &con->swork);
 438 out:
 439         read_unlock_bh(&sk->sk_callback_lock);
 440 }
 441
 442 static inline void lowcomms_connect_sock(struct connection *con)
 443 {
 444         if (test_bit(CF_CLOSE, &con->flags))
 445                 return;
 446         queue_work(send_workqueue, &con->swork);
 447         cond_resched();
 448 }
 449
 450 static void lowcomms_state_change(struct sock *sk)
 451 {
 452         /* SCTP layer is not calling sk_data_ready when the connection
 453          * is done, so we catch the signal through here. Also, it
 454          * doesn't switch socket state when entering shutdown, so we
 455          * skip the write in that case.
 456          */
 457         if (sk->sk_shutdown) {
 458                 if (sk->sk_shutdown == RCV_SHUTDOWN)
 459                         lowcomms_data_ready(sk);
 460         } else if (sk->sk_state == TCP_ESTABLISHED) {
 461                 lowcomms_write_space(sk);
 462         }
 463 }
 464
 465 int dlm_lowcomms_connect_node(int nodeid)
 466 {
 467         struct connection *con;
 468
 469         if (nodeid == dlm_our_nodeid())
 470                 return 0;
 471
 472         con = nodeid2con(nodeid, GFP_NOFS);
 473         if (!con)
 474                 return -ENOMEM;
 475         lowcomms_connect_sock(con);
 476         return 0;
 477 }
 478
 479 static void lowcomms_error_report(struct sock *sk)
 480 {
 481         struct connection *con;
 482         struct sockaddr_storage saddr;
 483         void (*orig_report)(struct sock *) = NULL;
 484
 485         read_lock_bh(&sk->sk_callback_lock);
 486         con = sock2con(sk);
 487         if (con == NULL)
 488                 goto out;
 489
 490         orig_report = listen_sock.sk_error_report;
 491         if (con->sock == NULL ||
 492             kernel_getpeername(con->sock, (struct sockaddr *)&saddr) < 0) {
 493                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 494                                    "sending to node %d, port %d, "
 495                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 496                                    con->nodeid, dlm_config.ci_tcp_port,
 497                                    sk->sk_err, sk->sk_err_soft);
 498         } else if (saddr.ss_family == AF_INET) {
 499                 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr;
 500
 501                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 502                                    "sending to node %d at %pI4, port %d, "
 503                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 504                                    con->nodeid, &sin4->sin_addr.s_addr,
 505                                    dlm_config.ci_tcp_port, sk->sk_err,
 506                                    sk->sk_err_soft);
 507         } else {
 508                 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr;
 509
 510                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 511                                    "sending to node %d at %u.%u.%u.%u, "
 512                                    "port %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 513                                    con->nodeid, sin6->sin6_addr.s6_addr32[0],
 514                                    sin6->sin6_addr.s6_addr32[1],
 515                                    sin6->sin6_addr.s6_addr32[2],
 516                                    sin6->sin6_addr.s6_addr32[3],
 517                                    dlm_config.ci_tcp_port, sk->sk_err,
 518                                    sk->sk_err_soft);
 519         }
 520 out:
 521         read_unlock_bh(&sk->sk_callback_lock);
 522         if (orig_report)
 523                 orig_report(sk);
 524 }
 525
 526 /* Note: sk_callback_lock must be locked before calling this function. */
 527 static void save_listen_callbacks(struct socket *sock)
 528 {
 529         struct sock *sk = sock->sk;
 530
 531         listen_sock.sk_data_ready = sk->sk_data_ready;
 532         listen_sock.sk_state_change = sk->sk_state_change;
 533         listen_sock.sk_write_space = sk->sk_write_space;
 534         listen_sock.sk_error_report = sk->sk_error_report;
 535 }
 536
 537 static void restore_callbacks(struct socket *sock)
 538 {
 539         struct sock *sk = sock->sk;
 540
 541         write_lock_bh(&sk->sk_callback_lock);
 542         sk->sk_user_data = NULL;
 543         sk->sk_data_ready = listen_sock.sk_data_ready;
 544         sk->sk_state_change = listen_sock.sk_state_change;
 545         sk->sk_write_space = listen_sock.sk_write_space;
 546         sk->sk_error_report = listen_sock.sk_error_report;
 547         write_unlock_bh(&sk->sk_callback_lock);
 548 }
 549
 550 /* Make a socket active */
 551 static void add_sock(struct socket *sock, struct connection *con)
 552 {
 553         struct sock *sk = sock->sk;
 554
 555         write_lock_bh(&sk->sk_callback_lock);
 556         con->sock = sock;
 557
 558         sk->sk_user_data = con;
 559         /* Install a data_ready callback */
 560         sk->sk_data_ready = lowcomms_data_ready;
 561         sk->sk_write_space = lowcomms_write_space;
 562         sk->sk_state_change = lowcomms_state_change;
 563         sk->sk_allocation = GFP_NOFS;
 564         sk->sk_error_report = lowcomms_error_report;
 565         write_unlock_bh(&sk->sk_callback_lock);
 566 }
 567
 568 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 569    length */
 570 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 571                           int *addr_len)
 572 {
 573         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 574         if (saddr->ss_family == AF_INET) {
 575                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 576                 in4_addr->sin_port = cpu_to_be16(port);
 577                 *addr_len = sizeof(struct sockaddr_in);
 578                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 579         } else {
 580                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 581                 in6_addr->sin6_port = cpu_to_be16(port);
 582                 *addr_len = sizeof(struct sockaddr_in6);
 583         }
 584         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 585 }
 586
 587 /* Close a remote connection and tidy up */
 588 static void close_connection(struct connection *con, bool and_other,
 589                              bool tx, bool rx)
 590 {
 591         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 592
 593         if (tx && !closing && cancel_work_sync(&con->swork)) {
 594                 log_print("canceled swork for node %d", con->nodeid);
 595                 clear_bit(CF_WRITE_PENDING, &con->flags);
 596         }
 597         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 598                 log_print("canceled rwork for node %d", con->nodeid);
 599                 clear_bit(CF_READ_PENDING, &con->flags);
 600         }
 601
 602         mutex_lock(&con->sock_mutex);
 603         if (con->sock) {
 604                 restore_callbacks(con->sock);
 605                 sock_release(con->sock);
 606                 con->sock = NULL;
 607         }
 608         if (con->othercon && and_other) {
 609                 /* Will only re-enter once. */
 610                 close_connection(con->othercon, false, true, true);
 611         }
 612         if (con->rx_page) {
 613                 __free_page(con->rx_page);
 614                 con->rx_page = NULL;
 615         }
 616
 617         con->retries = 0;
 618         mutex_unlock(&con->sock_mutex);
 619         clear_bit(CF_CLOSING, &con->flags);
 620 }
 621
 622 /* Data received from remote end */
 623 static int receive_from_sock(struct connection *con)
 624 {
 625         int ret = 0;
 626         struct msghdr msg = {};
 627         struct kvec iov[2];
 628         unsigned len;
 629         int r;
 630         int call_again_soon = 0;
 631         int nvec;
 632
 633         mutex_lock(&con->sock_mutex);
 634
 635         if (con->sock == NULL) {
 636                 ret = -EAGAIN;
 637                 goto out_close;
 638         }
 639         if (con->nodeid == 0) {
 640                 ret = -EINVAL;
 641                 goto out_close;
 642         }
 643
 644         if (con->rx_page == NULL) {
 645                 /*
 646                  * This doesn't need to be atomic, but I think it should
 647                  * improve performance if it is.
 648                  */
 649                 con->rx_page = alloc_page(GFP_ATOMIC);
 650                 if (con->rx_page == NULL)
 651                         goto out_resched;
 652                 cbuf_init(&con->cb, PAGE_SIZE);
 653         }
 654
 655         /*
 656          * iov[0] is the bit of the circular buffer between the current end
 657          * point (cb.base + cb.len) and the end of the buffer.
 658          */
 659         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 660         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 661         iov[1].iov_len = 0;
 662         nvec = 1;
 663
 664         /*
 665          * iov[1] is the bit of the circular buffer between the start of the
 666          * buffer and the start of the currently used section (cb.base)
 667          */
 668         if (cbuf_data(&con->cb) >= con->cb.base) {
 669                 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb);
 670                 iov[1].iov_len = con->cb.base;
 671                 iov[1].iov_base = page_address(con->rx_page);
 672                 nvec = 2;
 673         }
 674         len = iov[0].iov_len + iov[1].iov_len;
 675         iov_iter_kvec(&msg.msg_iter, READ, iov, nvec, len);
 676
 677         r = ret = sock_recvmsg(con->sock, &msg, MSG_DONTWAIT | MSG_NOSIGNAL);
 678         if (ret <= 0)
 679                 goto out_close;
 680         else if (ret == len)
 681                 call_again_soon = 1;
 682
 683         cbuf_add(&con->cb, ret);
 684         ret = dlm_process_incoming_buffer(con->nodeid,
 685                                           page_address(con->rx_page),
 686                                           con->cb.base, con->cb.len,
 687                                           PAGE_SIZE);
 688         if (ret == -EBADMSG) {
 689                 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d",
 690                           page_address(con->rx_page), con->cb.base,
 691                           con->cb.len, r);
 692         }
 693         if (ret < 0)
 694                 goto out_close;
 695         cbuf_eat(&con->cb, ret);
 696
 697         if (cbuf_empty(&con->cb) && !call_again_soon) {
 698                 __free_page(con->rx_page);
 699                 con->rx_page = NULL;
 700         }
 701
 702         if (call_again_soon)
 703                 goto out_resched;
 704         mutex_unlock(&con->sock_mutex);
 705         return 0;
 706
 707 out_resched:
 708         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 709                 queue_work(recv_workqueue, &con->rwork);
 710         mutex_unlock(&con->sock_mutex);
 711         return -EAGAIN;
 712
 713 out_close:
 714         mutex_unlock(&con->sock_mutex);
 715         if (ret != -EAGAIN) {
 716                 close_connection(con, true, true, false);
 717                 /* Reconnect when there is something to send */
 718         }
 719         /* Don't return success if we really got EOF */
 720         if (ret == 0)
 721                 ret = -EAGAIN;
 722
 723         return ret;
 724 }
 725
 726 /* Listening socket is busy, accept a connection */
 727 static int tcp_accept_from_sock(struct connection *con)
 728 {
 729         int result;
 730         struct sockaddr_storage peeraddr;
 731         struct socket *newsock;
 732         int len;
 733         int nodeid;
 734         struct connection *newcon;
 735         struct connection *addcon;
 736
 737         mutex_lock(&connections_lock);
 738         if (!dlm_allow_conn) {
 739                 mutex_unlock(&connections_lock);
 740                 return -1;
 741         }
 742         mutex_unlock(&connections_lock);
 743
 744         mutex_lock_nested(&con->sock_mutex, 0);
 745
 746         if (!con->sock) {
 747                 mutex_unlock(&con->sock_mutex);
 748                 return -ENOTCONN;
 749         }
 750
 751         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 752         if (result < 0)
 753                 goto accept_err;
 754
 755         /* Get the connected socket's peer */
 756         memset(&peeraddr, 0, sizeof(peeraddr));
 757         len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
 758         if (len < 0) {
 759                 result = -ECONNABORTED;
 760                 goto accept_err;
 761         }
 762
 763         /* Get the new node's NODEID */
 764         make_sockaddr(&peeraddr, 0, &len);
 765         if (addr_to_nodeid(&peeraddr, &nodeid)) {
 766                 unsigned char *b=(unsigned char *)&peeraddr;
 767                 log_print("connect from non cluster node");
 768                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 769                                      b, sizeof(struct sockaddr_storage));
 770                 sock_release(newsock);
 771                 mutex_unlock(&con->sock_mutex);
 772                 return -1;
 773         }
 774
 775         log_print("got connection from %d", nodeid);
 776
 777         /*  Check to see if we already have a connection to this node. This
 778          *  could happen if the two nodes initiate a connection at roughly
 779          *  the same time and the connections cross on the wire.
 780          *  In this case we store the incoming one in "othercon"
 781          */
 782         newcon = nodeid2con(nodeid, GFP_NOFS);
 783         if (!newcon) {
 784                 result = -ENOMEM;
 785                 goto accept_err;
 786         }
 787         mutex_lock_nested(&newcon->sock_mutex, 1);
 788         if (newcon->sock) {
 789                 struct connection *othercon = newcon->othercon;
 790
 791                 if (!othercon) {
 792                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 793                         if (!othercon) {
 794                                 log_print("failed to allocate incoming socket");
 795                                 mutex_unlock(&newcon->sock_mutex);
 796                                 result = -ENOMEM;
 797                                 goto accept_err;
 798                         }
 799                         othercon->nodeid = nodeid;
 800                         othercon->rx_action = receive_from_sock;
 801                         mutex_init(&othercon->sock_mutex);
 802                         INIT_LIST_HEAD(&othercon->writequeue);
 803                         spin_lock_init(&othercon->writequeue_lock);
 804                         INIT_WORK(&othercon->swork, process_send_sockets);
 805                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 806                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 807                 }
 808                 mutex_lock_nested(&othercon->sock_mutex, 2);
 809                 if (!othercon->sock) {
 810                         newcon->othercon = othercon;
 811                         add_sock(newsock, othercon);
 812                         addcon = othercon;
 813                         mutex_unlock(&othercon->sock_mutex);
 814                 }
 815                 else {
 816                         printk("Extra connection from node %d attempted\n", nodeid);
 817                         result = -EAGAIN;
 818                         mutex_unlock(&othercon->sock_mutex);
 819                         mutex_unlock(&newcon->sock_mutex);
 820                         goto accept_err;
 821                 }
 822         }
 823         else {
 824                 newcon->rx_action = receive_from_sock;
 825                 /* accept copies the sk after we've saved the callbacks, so we
 826                    don't want to save them a second time or comm errors will
 827                    result in calling sk_error_report recursively. */
 828                 add_sock(newsock, newcon);
 829                 addcon = newcon;
 830         }
 831
 832         mutex_unlock(&newcon->sock_mutex);
 833
 834         /*
 835          * Add it to the active queue in case we got data
 836          * between processing the accept adding the socket
 837          * to the read_sockets list
 838          */
 839         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 840                 queue_work(recv_workqueue, &addcon->rwork);
 841         mutex_unlock(&con->sock_mutex);
 842
 843         return 0;
 844
 845 accept_err:
 846         mutex_unlock(&con->sock_mutex);
 847         if (newsock)
 848                 sock_release(newsock);
 849
 850         if (result != -EAGAIN)
 851                 log_print("error accepting connection from node: %d", result);
 852         return result;
 853 }
 854
 855 static int sctp_accept_from_sock(struct connection *con)
 856 {
 857         /* Check that the new node is in the lockspace */
 858         struct sctp_prim prim;
 859         int nodeid;
 860         int prim_len, ret;
 861         int addr_len;
 862         struct connection *newcon;
 863         struct connection *addcon;
 864         struct socket *newsock;
 865
 866         mutex_lock(&connections_lock);
 867         if (!dlm_allow_conn) {
 868                 mutex_unlock(&connections_lock);
 869                 return -1;
 870         }
 871         mutex_unlock(&connections_lock);
 872
 873         mutex_lock_nested(&con->sock_mutex, 0);
 874
 875         ret = kernel_accept(con->sock, &newsock, O_NONBLOCK);
 876         if (ret < 0)
 877                 goto accept_err;
 878
 879         memset(&prim, 0, sizeof(struct sctp_prim));
 880         prim_len = sizeof(struct sctp_prim);
 881
 882         ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR,
 883                                 (char *)&prim, &prim_len);
 884         if (ret < 0) {
 885                 log_print("getsockopt/sctp_primary_addr failed: %d", ret);
 886                 goto accept_err;
 887         }
 888
 889         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 890         ret = addr_to_nodeid(&prim.ssp_addr, &nodeid);
 891         if (ret) {
 892                 unsigned char *b = (unsigned char *)&prim.ssp_addr;
 893
 894                 log_print("reject connect from unknown addr");
 895                 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE,
 896                                      b, sizeof(struct sockaddr_storage));
 897                 goto accept_err;
 898         }
 899
 900         newcon = nodeid2con(nodeid, GFP_NOFS);
 901         if (!newcon) {
 902                 ret = -ENOMEM;
 903                 goto accept_err;
 904         }
 905
 906         mutex_lock_nested(&newcon->sock_mutex, 1);
 907
 908         if (newcon->sock) {
 909                 struct connection *othercon = newcon->othercon;
 910
 911                 if (!othercon) {
 912                         othercon = kmem_cache_zalloc(con_cache, GFP_NOFS);
 913                         if (!othercon) {
 914                                 log_print("failed to allocate incoming socket");
 915                                 mutex_unlock(&newcon->sock_mutex);
 916                                 ret = -ENOMEM;
 917                                 goto accept_err;
 918                         }
 919                         othercon->nodeid = nodeid;
 920                         othercon->rx_action = receive_from_sock;
 921                         mutex_init(&othercon->sock_mutex);
 922                         INIT_LIST_HEAD(&othercon->writequeue);
 923                         spin_lock_init(&othercon->writequeue_lock);
 924                         INIT_WORK(&othercon->swork, process_send_sockets);
 925                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 926                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 927                 }
 928                 mutex_lock_nested(&othercon->sock_mutex, 2);
 929                 if (!othercon->sock) {
 930                         newcon->othercon = othercon;
 931                         add_sock(newsock, othercon);
 932                         addcon = othercon;
 933                         mutex_unlock(&othercon->sock_mutex);
 934                 } else {
 935                         printk("Extra connection from node %d attempted\n", nodeid);
 936                         ret = -EAGAIN;
 937                         mutex_unlock(&othercon->sock_mutex);
 938                         mutex_unlock(&newcon->sock_mutex);
 939                         goto accept_err;
 940                 }
 941         } else {
 942                 newcon->rx_action = receive_from_sock;
 943                 add_sock(newsock, newcon);
 944                 addcon = newcon;
 945         }
 946
 947         log_print("connected to %d", nodeid);
 948
 949         mutex_unlock(&newcon->sock_mutex);
 950
 951         /*
 952          * Add it to the active queue in case we got data
 953          * between processing the accept adding the socket
 954          * to the read_sockets list
 955          */
 956         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 957                 queue_work(recv_workqueue, &addcon->rwork);
 958         mutex_unlock(&con->sock_mutex);
 959
 960         return 0;
 961
 962 accept_err:
 963         mutex_unlock(&con->sock_mutex);
 964         if (newsock)
 965                 sock_release(newsock);
 966         if (ret != -EAGAIN)
 967                 log_print("error accepting connection from node: %d", ret);
 968
 969         return ret;
 970 }
 971
 972 static void free_entry(struct writequeue_entry *e)
 973 {
 974         __free_page(e->page);
 975         kfree(e);
 976 }
 977
 978 /*
 979  * writequeue_entry_complete - try to delete and free write queue entry
 980  * @e: write queue entry to try to delete
 981  * @completed: bytes completed
 982  *
 983  * writequeue_lock must be held.
 984  */
 985 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 986 {
 987         e->offset += completed;
 988         e->len -= completed;
 989
 990         if (e->len == 0 && e->users == 0) {
 991                 list_del(&e->list);
 992                 free_entry(e);
 993         }
 994 }
 995
 996 /*
 997  * sctp_bind_addrs - bind a SCTP socket to all our addresses
 998  */
 999 static int sctp_bind_addrs(struct connection *con, uint16_t port)
1000 {
1001         struct sockaddr_storage localaddr;
1002         int i, addr_len, result = 0;
1003
1004         for (i = 0; i < dlm_local_count; i++) {
1005                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1006                 make_sockaddr(&localaddr, port, &addr_len);
1007
1008                 if (!i)
1009                         result = kernel_bind(con->sock,
1010                                              (struct sockaddr *)&localaddr,
1011                                              addr_len);
1012                 else
1013                         result = kernel_setsockopt(con->sock, SOL_SCTP,
1014                                                    SCTP_SOCKOPT_BINDX_ADD,
1015                                                    (char *)&localaddr, addr_len);
1016
1017                 if (result < 0) {
1018                         log_print("Can't bind to %d addr number %d, %d.\n",
1019                                   port, i + 1, result);
1020                         break;
1021                 }
1022         }
1023         return result;
1024 }
1025
1026 /* Initiate an SCTP association.
1027    This is a special case of send_to_sock() in that we don't yet have a
1028    peeled-off socket for this association, so we use the listening socket
1029    and add the primary IP address of the remote node.
1030  */
1031 static void sctp_connect_to_sock(struct connection *con)
1032 {
1033         struct sockaddr_storage daddr;
1034         int one = 1;
1035         int result;
1036         int addr_len;
1037         struct socket *sock;
1038         struct timeval tv = { .tv_sec = 5, .tv_usec = 0 };
1039
1040         if (con->nodeid == 0) {
1041                 log_print("attempt to connect sock 0 foiled");
1042                 return;
1043         }
1044
1045         mutex_lock(&con->sock_mutex);
1046
1047         /* Some odd races can cause double-connects, ignore them */
1048         if (con->retries++ > MAX_CONNECT_RETRIES)
1049                 goto out;
1050
1051         if (con->sock) {
1052                 log_print("node %d already connected.", con->nodeid);
1053                 goto out;
1054         }
1055
1056         memset(&daddr, 0, sizeof(daddr));
1057         result = nodeid_to_addr(con->nodeid, &daddr, NULL, true);
1058         if (result < 0) {
1059                 log_print("no address for nodeid %d", con->nodeid);
1060                 goto out;
1061         }
1062
1063         /* Create a socket to communicate with */
1064         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1065                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1066         if (result < 0)
1067                 goto socket_err;
1068
1069         con->rx_action = receive_from_sock;
1070         con->connect_action = sctp_connect_to_sock;
1071         add_sock(sock, con);
1072
1073         /* Bind to all addresses. */
1074         if (sctp_bind_addrs(con, 0))
1075                 goto bind_err;
1076
1077         make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len);
1078
1079         log_print("connecting to %d", con->nodeid);
1080
1081         /* Turn off Nagle's algorithm */
1082         kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1083                           sizeof(one));
1084
1085         /*
1086          * Make sock->ops->connect() function return in specified time,
1087          * since O_NONBLOCK argument in connect() function does not work here,
1088          * then, we should restore the default value of this attribute.
1089          */
1090         kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
1091                           sizeof(tv));
1092         result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len,
1093                                    0);
1094         memset(&tv, 0, sizeof(tv));
1095         kernel_setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO_OLD, (char *)&tv,
1096                           sizeof(tv));
1097
1098         if (result == -EINPROGRESS)
1099                 result = 0;
1100         if (result == 0)
1101                 goto out;
1102
1103 bind_err:
1104         con->sock = NULL;
1105         sock_release(sock);
1106
1107 socket_err:
1108         /*
1109          * Some errors are fatal and this list might need adjusting. For other
1110          * errors we try again until the max number of retries is reached.
1111          */
1112         if (result != -EHOSTUNREACH &&
1113             result != -ENETUNREACH &&
1114             result != -ENETDOWN &&
1115             result != -EINVAL &&
1116             result != -EPROTONOSUPPORT) {
1117                 log_print("connect %d try %d error %d", con->nodeid,
1118                           con->retries, result);
1119                 mutex_unlock(&con->sock_mutex);
1120                 msleep(1000);
1121                 lowcomms_connect_sock(con);
1122                 return;
1123         }
1124
1125 out:
1126         mutex_unlock(&con->sock_mutex);
1127 }
1128
1129 /* Connect a new socket to its peer */
1130 static void tcp_connect_to_sock(struct connection *con)
1131 {
1132         struct sockaddr_storage saddr, src_addr;
1133         int addr_len;
1134         struct socket *sock = NULL;
1135         int one = 1;
1136         int result;
1137
1138         if (con->nodeid == 0) {
1139                 log_print("attempt to connect sock 0 foiled");
1140                 return;
1141         }
1142
1143         mutex_lock(&con->sock_mutex);
1144         if (con->retries++ > MAX_CONNECT_RETRIES)
1145                 goto out;
1146
1147         /* Some odd races can cause double-connects, ignore them */
1148         if (con->sock)
1149                 goto out;
1150
1151         /* Create a socket to communicate with */
1152         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1153                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1154         if (result < 0)
1155                 goto out_err;
1156
1157         memset(&saddr, 0, sizeof(saddr));
1158         result = nodeid_to_addr(con->nodeid, &saddr, NULL, false);
1159         if (result < 0) {
1160                 log_print("no address for nodeid %d", con->nodeid);
1161                 goto out_err;
1162         }
1163
1164         con->rx_action = receive_from_sock;
1165         con->connect_action = tcp_connect_to_sock;
1166         add_sock(sock, con);
1167
1168         /* Bind to our cluster-known address connecting to avoid
1169            routing problems */
1170         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1171         make_sockaddr(&src_addr, 0, &addr_len);
1172         result = sock->ops->bind(sock, (struct sockaddr *) &src_addr,
1173                                  addr_len);
1174         if (result < 0) {
1175                 log_print("could not bind for connect: %d", result);
1176                 /* This *may* not indicate a critical error */
1177         }
1178
1179         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
1180
1181         log_print("connecting to %d", con->nodeid);
1182
1183         /* Turn off Nagle's algorithm */
1184         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1185                           sizeof(one));
1186
1187         result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
1188                                    O_NONBLOCK);
1189         if (result == -EINPROGRESS)
1190                 result = 0;
1191         if (result == 0)
1192                 goto out;
1193
1194 out_err:
1195         if (con->sock) {
1196                 sock_release(con->sock);
1197                 con->sock = NULL;
1198         } else if (sock) {
1199                 sock_release(sock);
1200         }
1201         /*
1202          * Some errors are fatal and this list might need adjusting. For other
1203          * errors we try again until the max number of retries is reached.
1204          */
1205         if (result != -EHOSTUNREACH &&
1206             result != -ENETUNREACH &&
1207             result != -ENETDOWN &&
1208             result != -EINVAL &&
1209             result != -EPROTONOSUPPORT) {
1210                 log_print("connect %d try %d error %d", con->nodeid,
1211                           con->retries, result);
1212                 mutex_unlock(&con->sock_mutex);
1213                 msleep(1000);
1214                 lowcomms_connect_sock(con);
1215                 return;
1216         }
1217 out:
1218         mutex_unlock(&con->sock_mutex);
1219         return;
1220 }
1221
1222 static struct socket *tcp_create_listen_sock(struct connection *con,
1223                                              struct sockaddr_storage *saddr)
1224 {
1225         struct socket *sock = NULL;
1226         int result = 0;
1227         int one = 1;
1228         int addr_len;
1229
1230         if (dlm_local_addr[0]->ss_family == AF_INET)
1231                 addr_len = sizeof(struct sockaddr_in);
1232         else
1233                 addr_len = sizeof(struct sockaddr_in6);
1234
1235         /* Create a socket to communicate with */
1236         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1237                                   SOCK_STREAM, IPPROTO_TCP, &sock);
1238         if (result < 0) {
1239                 log_print("Can't create listening comms socket");
1240                 goto create_out;
1241         }
1242
1243         /* Turn off Nagle's algorithm */
1244         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one,
1245                           sizeof(one));
1246
1247         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
1248                                    (char *)&one, sizeof(one));
1249
1250         if (result < 0) {
1251                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
1252         }
1253         write_lock_bh(&sock->sk->sk_callback_lock);
1254         sock->sk->sk_user_data = con;
1255         save_listen_callbacks(sock);
1256         con->rx_action = tcp_accept_from_sock;
1257         con->connect_action = tcp_connect_to_sock;
1258         write_unlock_bh(&sock->sk->sk_callback_lock);
1259
1260         /* Bind to our port */
1261         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
1262         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
1263         if (result < 0) {
1264                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
1265                 sock_release(sock);
1266                 sock = NULL;
1267                 con->sock = NULL;
1268                 goto create_out;
1269         }
1270         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
1271                                  (char *)&one, sizeof(one));
1272         if (result < 0) {
1273                 log_print("Set keepalive failed: %d", result);
1274         }
1275
1276         result = sock->ops->listen(sock, 5);
1277         if (result < 0) {
1278                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
1279                 sock_release(sock);
1280                 sock = NULL;
1281                 goto create_out;
1282         }
1283
1284 create_out:
1285         return sock;
1286 }
1287
1288 /* Get local addresses */
1289 static void init_local(void)
1290 {
1291         struct sockaddr_storage sas, *addr;
1292         int i;
1293
1294         dlm_local_count = 0;
1295         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1296                 if (dlm_our_addr(&sas, i))
1297                         break;
1298
1299                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1300                 if (!addr)
1301                         break;
1302                 dlm_local_addr[dlm_local_count++] = addr;
1303         }
1304 }
1305
1306 /* Initialise SCTP socket and bind to all interfaces */
1307 static int sctp_listen_for_all(void)
1308 {
1309         struct socket *sock = NULL;
1310         int result = -EINVAL;
1311         struct connection *con = nodeid2con(0, GFP_NOFS);
1312         int bufsize = NEEDED_RMEM;
1313         int one = 1;
1314
1315         if (!con)
1316                 return -ENOMEM;
1317
1318         log_print("Using SCTP for communications");
1319
1320         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1321                                   SOCK_STREAM, IPPROTO_SCTP, &sock);
1322         if (result < 0) {
1323                 log_print("Can't create comms socket, check SCTP is loaded");
1324                 goto out;
1325         }
1326
1327         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE,
1328                                  (char *)&bufsize, sizeof(bufsize));
1329         if (result)
1330                 log_print("Error increasing buffer space on socket %d", result);
1331
1332         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one,
1333                                    sizeof(one));
1334         if (result < 0)
1335                 log_print("Could not set SCTP NODELAY error %d\n", result);
1336
1337         write_lock_bh(&sock->sk->sk_callback_lock);
1338         /* Init con struct */
1339         sock->sk->sk_user_data = con;
1340         save_listen_callbacks(sock);
1341         con->sock = sock;
1342         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1343         con->rx_action = sctp_accept_from_sock;
1344         con->connect_action = sctp_connect_to_sock;
1345
1346         write_unlock_bh(&sock->sk->sk_callback_lock);
1347
1348         /* Bind to all addresses. */
1349         if (sctp_bind_addrs(con, dlm_config.ci_tcp_port))
1350                 goto create_delsock;
1351
1352         result = sock->ops->listen(sock, 5);
1353         if (result < 0) {
1354                 log_print("Can't set socket listening");
1355                 goto create_delsock;
1356         }
1357
1358         return 0;
1359
1360 create_delsock:
1361         sock_release(sock);
1362         con->sock = NULL;
1363 out:
1364         return result;
1365 }
1366
1367 static int tcp_listen_for_all(void)
1368 {
1369         struct socket *sock = NULL;
1370         struct connection *con = nodeid2con(0, GFP_NOFS);
1371         int result = -EINVAL;
1372
1373         if (!con)
1374                 return -ENOMEM;
1375
1376         /* We don't support multi-homed hosts */
1377         if (dlm_local_addr[1] != NULL) {
1378                 log_print("TCP protocol can't handle multi-homed hosts, "
1379                           "try SCTP");
1380                 return -EINVAL;
1381         }
1382
1383         log_print("Using TCP for communications");
1384
1385         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1386         if (sock) {
1387                 add_sock(sock, con);
1388                 result = 0;
1389         }
1390         else {
1391                 result = -EADDRINUSE;
1392         }
1393
1394         return result;
1395 }
1396
1397
1398
1399 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1400                                                      gfp_t allocation)
1401 {
1402         struct writequeue_entry *entry;
1403
1404         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1405         if (!entry)
1406                 return NULL;
1407
1408         entry->page = alloc_page(allocation);
1409         if (!entry->page) {
1410                 kfree(entry);
1411                 return NULL;
1412         }
1413
1414         entry->offset = 0;
1415         entry->len = 0;
1416         entry->end = 0;
1417         entry->users = 0;
1418         entry->con = con;
1419
1420         return entry;
1421 }
1422
1423 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1424 {
1425         struct connection *con;
1426         struct writequeue_entry *e;
1427         int offset = 0;
1428
1429         con = nodeid2con(nodeid, allocation);
1430         if (!con)
1431                 return NULL;
1432
1433         spin_lock(&con->writequeue_lock);
1434         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1435         if ((&e->list == &con->writequeue) ||
1436             (PAGE_SIZE - e->end < len)) {
1437                 e = NULL;
1438         } else {
1439                 offset = e->end;
1440                 e->end += len;
1441                 e->users++;
1442         }
1443         spin_unlock(&con->writequeue_lock);
1444
1445         if (e) {
1446         got_one:
1447                 *ppc = page_address(e->page) + offset;
1448                 return e;
1449         }
1450
1451         e = new_writequeue_entry(con, allocation);
1452         if (e) {
1453                 spin_lock(&con->writequeue_lock);
1454                 offset = e->end;
1455                 e->end += len;
1456                 e->users++;
1457                 list_add_tail(&e->list, &con->writequeue);
1458                 spin_unlock(&con->writequeue_lock);
1459                 goto got_one;
1460         }
1461         return NULL;
1462 }
1463
1464 void dlm_lowcomms_commit_buffer(void *mh)
1465 {
1466         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1467         struct connection *con = e->con;
1468         int users;
1469
1470         spin_lock(&con->writequeue_lock);
1471         users = --e->users;
1472         if (users)
1473                 goto out;
1474         e->len = e->end - e->offset;
1475         spin_unlock(&con->writequeue_lock);
1476
1477         queue_work(send_workqueue, &con->swork);
1478         return;
1479
1480 out:
1481         spin_unlock(&con->writequeue_lock);
1482         return;
1483 }
1484
1485 /* Send a message */
1486 static void send_to_sock(struct connection *con)
1487 {
1488         int ret = 0;
1489         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1490         struct writequeue_entry *e;
1491         int len, offset;
1492         int count = 0;
1493
1494         mutex_lock(&con->sock_mutex);
1495         if (con->sock == NULL)
1496                 goto out_connect;
1497
1498         spin_lock(&con->writequeue_lock);
1499         for (;;) {
1500                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1501                                list);
1502                 if ((struct list_head *) e == &con->writequeue)
1503                         break;
1504
1505                 len = e->len;
1506                 offset = e->offset;
1507                 BUG_ON(len == 0 && e->users == 0);
1508                 spin_unlock(&con->writequeue_lock);
1509
1510                 ret = 0;
1511                 if (len) {
1512                         ret = kernel_sendpage(con->sock, e->page, offset, len,
1513                                               msg_flags);
1514                         if (ret == -EAGAIN || ret == 0) {
1515                                 if (ret == -EAGAIN &&
1516                                     test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1517                                     !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1518                                         /* Notify TCP that we're limited by the
1519                                          * application window size.
1520                                          */
1521                                         set_bit(SOCK_NOSPACE, &con->sock->flags);
1522                                         con->sock->sk->sk_write_pending++;
1523                                 }
1524                                 cond_resched();
1525                                 goto out;
1526                         } else if (ret < 0)
1527                                 goto send_error;
1528                 }
1529
1530                 /* Don't starve people filling buffers */
1531                 if (++count >= MAX_SEND_MSG_COUNT) {
1532                         cond_resched();
1533                         count = 0;
1534                 }
1535
1536                 spin_lock(&con->writequeue_lock);
1537                 writequeue_entry_complete(e, ret);
1538         }
1539         spin_unlock(&con->writequeue_lock);
1540 out:
1541         mutex_unlock(&con->sock_mutex);
1542         return;
1543
1544 send_error:
1545         mutex_unlock(&con->sock_mutex);
1546         close_connection(con, true, false, true);
1547         /* Requeue the send work. When the work daemon runs again, it will try
1548            a new connection, then call this function again. */
1549         queue_work(send_workqueue, &con->swork);
1550         return;
1551
1552 out_connect:
1553         mutex_unlock(&con->sock_mutex);
1554         queue_work(send_workqueue, &con->swork);
1555         cond_resched();
1556 }
1557
1558 static void clean_one_writequeue(struct connection *con)
1559 {
1560         struct writequeue_entry *e, *safe;
1561
1562         spin_lock(&con->writequeue_lock);
1563         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1564                 list_del(&e->list);
1565                 free_entry(e);
1566         }
1567         spin_unlock(&con->writequeue_lock);
1568 }
1569
1570 /* Called from recovery when it knows that a node has
1571    left the cluster */
1572 int dlm_lowcomms_close(int nodeid)
1573 {
1574         struct connection *con;
1575         struct dlm_node_addr *na;
1576
1577         log_print("closing connection to node %d", nodeid);
1578         con = nodeid2con(nodeid, 0);
1579         if (con) {
1580                 set_bit(CF_CLOSE, &con->flags);
1581                 close_connection(con, true, true, true);
1582                 clean_one_writequeue(con);
1583         }
1584
1585         spin_lock(&dlm_node_addrs_spin);
1586         na = find_node_addr(nodeid);
1587         if (na) {
1588                 list_del(&na->list);
1589                 while (na->addr_count--)
1590                         kfree(na->addr[na->addr_count]);
1591                 kfree(na);
1592         }
1593         spin_unlock(&dlm_node_addrs_spin);
1594
1595         return 0;
1596 }
1597
1598 /* Receive workqueue function */
1599 static void process_recv_sockets(struct work_struct *work)
1600 {
1601         struct connection *con = container_of(work, struct connection, rwork);
1602         int err;
1603
1604         clear_bit(CF_READ_PENDING, &con->flags);
1605         do {
1606                 err = con->rx_action(con);
1607         } while (!err);
1608 }
1609
1610 /* Send workqueue function */
1611 static void process_send_sockets(struct work_struct *work)
1612 {
1613         struct connection *con = container_of(work, struct connection, swork);
1614
1615         clear_bit(CF_WRITE_PENDING, &con->flags);
1616         if (con->sock == NULL) /* not mutex protected so check it inside too */
1617                 con->connect_action(con);
1618         if (!list_empty(&con->writequeue))
1619                 send_to_sock(con);
1620 }
1621
1622
1623 /* Discard all entries on the write queues */
1624 static void clean_writequeues(void)
1625 {
1626         foreach_conn(clean_one_writequeue);
1627 }
1628
1629 static void work_stop(void)
1630 {
1631         if (recv_workqueue)
1632                 destroy_workqueue(recv_workqueue);
1633         if (send_workqueue)
1634                 destroy_workqueue(send_workqueue);
1635 }
1636
1637 static int work_start(void)
1638 {
1639         recv_workqueue = alloc_workqueue("dlm_recv",
1640                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1641         if (!recv_workqueue) {
1642                 log_print("can't start dlm_recv");
1643                 return -ENOMEM;
1644         }
1645
1646         send_workqueue = alloc_workqueue("dlm_send",
1647                                          WQ_UNBOUND | WQ_MEM_RECLAIM, 1);
1648         if (!send_workqueue) {
1649                 log_print("can't start dlm_send");
1650                 destroy_workqueue(recv_workqueue);
1651                 return -ENOMEM;
1652         }
1653
1654         return 0;
1655 }
1656
1657 static void _stop_conn(struct connection *con, bool and_other)
1658 {
1659         mutex_lock(&con->sock_mutex);
1660         set_bit(CF_CLOSE, &con->flags);
1661         set_bit(CF_READ_PENDING, &con->flags);
1662         set_bit(CF_WRITE_PENDING, &con->flags);
1663         if (con->sock && con->sock->sk) {
1664                 write_lock_bh(&con->sock->sk->sk_callback_lock);
1665                 con->sock->sk->sk_user_data = NULL;
1666                 write_unlock_bh(&con->sock->sk->sk_callback_lock);
1667         }
1668         if (con->othercon && and_other)
1669                 _stop_conn(con->othercon, false);
1670         mutex_unlock(&con->sock_mutex);
1671 }
1672
1673 static void stop_conn(struct connection *con)
1674 {
1675         _stop_conn(con, true);
1676 }
1677
1678 static void free_conn(struct connection *con)
1679 {
1680         close_connection(con, true, true, true);
1681         if (con->othercon)
1682                 kmem_cache_free(con_cache, con->othercon);
1683         hlist_del(&con->list);
1684         kmem_cache_free(con_cache, con);
1685 }
1686
1687 static void work_flush(void)
1688 {
1689         int ok;
1690         int i;
1691         struct hlist_node *n;
1692         struct connection *con;
1693
1694         if (recv_workqueue)
1695                 flush_workqueue(recv_workqueue);
1696         if (send_workqueue)
1697                 flush_workqueue(send_workqueue);
1698         do {
1699                 ok = 1;
1700                 foreach_conn(stop_conn);
1701                 if (recv_workqueue)
1702                         flush_workqueue(recv_workqueue);
1703                 if (send_workqueue)
1704                         flush_workqueue(send_workqueue);
1705                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1706                         hlist_for_each_entry_safe(con, n,
1707                                                   &connection_hash[i], list) {
1708                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1709                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1710                                 if (con->othercon) {
1711                                         ok &= test_bit(CF_READ_PENDING,
1712                                                        &con->othercon->flags);
1713                                         ok &= test_bit(CF_WRITE_PENDING,
1714                                                        &con->othercon->flags);
1715                                 }
1716                         }
1717                 }
1718         } while (!ok);
1719 }
1720
1721 void dlm_lowcomms_stop(void)
1722 {
1723         /* Set all the flags to prevent any
1724            socket activity.
1725         */
1726         mutex_lock(&connections_lock);
1727         dlm_allow_conn = 0;
1728         mutex_unlock(&connections_lock);
1729         work_flush();
1730         clean_writequeues();
1731         foreach_conn(free_conn);
1732         work_stop();
1733
1734         kmem_cache_destroy(con_cache);
1735 }
1736
1737 int dlm_lowcomms_start(void)
1738 {
1739         int error = -EINVAL;
1740         struct connection *con;
1741         int i;
1742
1743         for (i = 0; i < CONN_HASH_SIZE; i++)
1744                 INIT_HLIST_HEAD(&connection_hash[i]);
1745
1746         init_local();
1747         if (!dlm_local_count) {
1748                 error = -ENOTCONN;
1749                 log_print("no local IP address has been set");
1750                 goto fail;
1751         }
1752
1753         error = -ENOMEM;
1754         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1755                                       __alignof__(struct connection), 0,
1756                                       NULL);
1757         if (!con_cache)
1758                 goto fail;
1759
1760         error = work_start();
1761         if (error)
1762                 goto fail_destroy;
1763
1764         dlm_allow_conn = 1;
1765
1766         /* Start listening */
1767         if (dlm_config.ci_protocol == 0)
1768                 error = tcp_listen_for_all();
1769         else
1770                 error = sctp_listen_for_all();
1771         if (error)
1772                 goto fail_unlisten;
1773
1774         return 0;
1775
1776 fail_unlisten:
1777         dlm_allow_conn = 0;
1778         con = nodeid2con(0,0);
1779         if (con) {
1780                 close_connection(con, false, true, true);
1781                 kmem_cache_free(con_cache, con);
1782         }
1783 fail_destroy:
1784         kmem_cache_destroy(con_cache);
1785 fail:
1786         return error;
1787 }
1788
1789 void dlm_lowcomms_exit(void)
1790 {
1791         struct dlm_node_addr *na, *safe;
1792
1793         spin_lock(&dlm_node_addrs_spin);
1794         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
1795                 list_del(&na->list);
1796                 while (na->addr_count--)
1797                         kfree(na->addr[na->addr_count]);
1798                 kfree(na);
1799         }
1800         spin_unlock(&dlm_node_addrs_spin);
1801 }