net/ipv4/inet_connection_sock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Support for INET connection oriented protocols.
   8  *
   9  * Authors:     See the TCP sources
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/jhash.h>
  14
  15 #include <net/inet_connection_sock.h>
  16 #include <net/inet_hashtables.h>
  17 #include <net/inet_timewait_sock.h>
  18 #include <net/ip.h>
  19 #include <net/route.h>
  20 #include <net/tcp_states.h>
  21 #include <net/xfrm.h>
  22 #include <net/tcp.h>
  23 #include <net/sock_reuseport.h>
  24 #include <net/addrconf.h>
  25
  26 #if IS_ENABLED(CONFIG_IPV6)
  27 /* match_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses if IPv6
  28  *                          only, and any IPv4 addresses if not IPv6 only
  29  * match_wildcard == false: addresses must be exactly the same, i.e.
  30  *                          IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  31  *                          and 0.0.0.0 equals to 0.0.0.0 only
  32  */
  33 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  34                                  const struct in6_addr *sk2_rcv_saddr6,
  35                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  36                                  bool sk1_ipv6only, bool sk2_ipv6only,
  37                                  bool match_wildcard)
  38 {
  39         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  40         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  41
  42         /* if both are mapped, treat as IPv4 */
  43         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  44                 if (!sk2_ipv6only) {
  45                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  46                                 return true;
  47                         if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  48                                 return match_wildcard;
  49                 }
  50                 return false;
  51         }
  52
  53         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  54                 return true;
  55
  56         if (addr_type2 == IPV6_ADDR_ANY && match_wildcard &&
  57             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  58                 return true;
  59
  60         if (addr_type == IPV6_ADDR_ANY && match_wildcard &&
  61             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  62                 return true;
  63
  64         if (sk2_rcv_saddr6 &&
  65             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  66                 return true;
  67
  68         return false;
  69 }
  70 #endif
  71
  72 /* match_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  73  * match_wildcard == false: addresses must be exactly the same, i.e.
  74  *                          0.0.0.0 only equals to 0.0.0.0
  75  */
  76 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  77                                  bool sk2_ipv6only, bool match_wildcard)
  78 {
  79         if (!sk2_ipv6only) {
  80                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  81                         return true;
  82                 if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  83                         return match_wildcard;
  84         }
  85         return false;
  86 }
  87
  88 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  89                           bool match_wildcard)
  90 {
  91 #if IS_ENABLED(CONFIG_IPV6)
  92         if (sk->sk_family == AF_INET6)
  93                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
  94                                             inet6_rcv_saddr(sk2),
  95                                             sk->sk_rcv_saddr,
  96                                             sk2->sk_rcv_saddr,
  97                                             ipv6_only_sock(sk),
  98                                             ipv6_only_sock(sk2),
  99                                             match_wildcard);
 100 #endif
 101         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 102                                     ipv6_only_sock(sk2), match_wildcard);
 103 }
 104 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 105
 106 bool inet_rcv_saddr_any(const struct sock *sk)
 107 {
 108 #if IS_ENABLED(CONFIG_IPV6)
 109         if (sk->sk_family == AF_INET6)
 110                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 111 #endif
 112         return !sk->sk_rcv_saddr;
 113 }
 114
 115 void inet_get_local_port_range(struct net *net, int *low, int *high)
 116 {
 117         unsigned int seq;
 118
 119         do {
 120                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 121
 122                 *low = net->ipv4.ip_local_ports.range[0];
 123                 *high = net->ipv4.ip_local_ports.range[1];
 124         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 125 }
 126 EXPORT_SYMBOL(inet_get_local_port_range);
 127
 128 static int inet_csk_bind_conflict(const struct sock *sk,
 129                                   const struct inet_bind_bucket *tb,
 130                                   bool relax, bool reuseport_ok)
 131 {
 132         struct sock *sk2;
 133         bool reuse = sk->sk_reuse;
 134         bool reuseport = !!sk->sk_reuseport;
 135         kuid_t uid = sock_i_uid((struct sock *)sk);
 136
 137         /*
 138          * Unlike other sk lookup places we do not check
 139          * for sk_net here, since _all_ the socks listed
 140          * in tb->owners list belong to the same net - the
 141          * one this bucket belongs to.
 142          */
 143
 144         sk_for_each_bound(sk2, &tb->owners) {
 145                 if (sk != sk2 &&
 146                     (!sk->sk_bound_dev_if ||
 147                      !sk2->sk_bound_dev_if ||
 148                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 149                         if (reuse && sk2->sk_reuse &&
 150                             sk2->sk_state != TCP_LISTEN) {
 151                                 if ((!relax ||
 152                                      (!reuseport_ok &&
 153                                       reuseport && sk2->sk_reuseport &&
 154                                       !rcu_access_pointer(sk->sk_reuseport_cb) &&
 155                                       (sk2->sk_state == TCP_TIME_WAIT ||
 156                                        uid_eq(uid, sock_i_uid(sk2))))) &&
 157                                     inet_rcv_saddr_equal(sk, sk2, true))
 158                                         break;
 159                         } else if (!reuseport_ok ||
 160                                    !reuseport || !sk2->sk_reuseport ||
 161                                    rcu_access_pointer(sk->sk_reuseport_cb) ||
 162                                    (sk2->sk_state != TCP_TIME_WAIT &&
 163                                     !uid_eq(uid, sock_i_uid(sk2)))) {
 164                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 165                                         break;
 166                         }
 167                 }
 168         }
 169         return sk2 != NULL;
 170 }
 171
 172 /*
 173  * Find an open port number for the socket.  Returns with the
 174  * inet_bind_hashbucket lock held.
 175  */
 176 static struct inet_bind_hashbucket *
 177 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 178 {
 179         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 180         int port = 0;
 181         struct inet_bind_hashbucket *head;
 182         struct net *net = sock_net(sk);
 183         bool relax = false;
 184         int i, low, high, attempt_half;
 185         struct inet_bind_bucket *tb;
 186         u32 remaining, offset;
 187         int l3mdev;
 188
 189         l3mdev = inet_sk_bound_l3mdev(sk);
 190 ports_exhausted:
 191         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 192 other_half_scan:
 193         inet_get_local_port_range(net, &low, &high);
 194         high++; /* [32768, 60999] -> [32768, 61000[ */
 195         if (high - low < 4)
 196                 attempt_half = 0;
 197         if (attempt_half) {
 198                 int half = low + (((high - low) >> 2) << 1);
 199
 200                 if (attempt_half == 1)
 201                         high = half;
 202                 else
 203                         low = half;
 204         }
 205         remaining = high - low;
 206         if (likely(remaining > 1))
 207                 remaining &= ~1U;
 208
 209         offset = prandom_u32() % remaining;
 210         /* __inet_hash_connect() favors ports having @low parity
 211          * We do the opposite to not pollute connect() users.
 212          */
 213         offset |= 1U;
 214
 215 other_parity_scan:
 216         port = low + offset;
 217         for (i = 0; i < remaining; i += 2, port += 2) {
 218                 if (unlikely(port >= high))
 219                         port -= remaining;
 220                 if (inet_is_local_reserved_port(net, port))
 221                         continue;
 222                 head = &hinfo->bhash[inet_bhashfn(net, port,
 223                                                   hinfo->bhash_size)];
 224                 spin_lock_bh(&head->lock);
 225                 inet_bind_bucket_for_each(tb, &head->chain)
 226                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 227                             tb->port == port) {
 228                                 if (!inet_csk_bind_conflict(sk, tb, relax, false))
 229                                         goto success;
 230                                 goto next_port;
 231                         }
 232                 tb = NULL;
 233                 goto success;
 234 next_port:
 235                 spin_unlock_bh(&head->lock);
 236                 cond_resched();
 237         }
 238
 239         offset--;
 240         if (!(offset & 1))
 241                 goto other_parity_scan;
 242
 243         if (attempt_half == 1) {
 244                 /* OK we now try the upper half of the range */
 245                 attempt_half = 2;
 246                 goto other_half_scan;
 247         }
 248
 249         if (net->ipv4.sysctl_ip_autobind_reuse && !relax) {
 250                 /* We still have a chance to connect to different destinations */
 251                 relax = true;
 252                 goto ports_exhausted;
 253         }
 254         return NULL;
 255 success:
 256         *port_ret = port;
 257         *tb_ret = tb;
 258         return head;
 259 }
 260
 261 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 262                                      struct sock *sk)
 263 {
 264         kuid_t uid = sock_i_uid(sk);
 265
 266         if (tb->fastreuseport <= 0)
 267                 return 0;
 268         if (!sk->sk_reuseport)
 269                 return 0;
 270         if (rcu_access_pointer(sk->sk_reuseport_cb))
 271                 return 0;
 272         if (!uid_eq(tb->fastuid, uid))
 273                 return 0;
 274         /* We only need to check the rcv_saddr if this tb was once marked
 275          * without fastreuseport and then was reset, as we can only know that
 276          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 277          * owners list.
 278          */
 279         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 280                 return 1;
 281 #if IS_ENABLED(CONFIG_IPV6)
 282         if (tb->fast_sk_family == AF_INET6)
 283                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 284                                             inet6_rcv_saddr(sk),
 285                                             tb->fast_rcv_saddr,
 286                                             sk->sk_rcv_saddr,
 287                                             tb->fast_ipv6_only,
 288                                             ipv6_only_sock(sk), true);
 289 #endif
 290         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 291                                     ipv6_only_sock(sk), true);
 292 }
 293
 294 /* Obtain a reference to a local port for the given sock,
 295  * if snum is zero it means select any available local port.
 296  * We try to allocate an odd port (and leave even ports for connect())
 297  */
 298 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 299 {
 300         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 301         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 302         int ret = 1, port = snum;
 303         struct inet_bind_hashbucket *head;
 304         struct net *net = sock_net(sk);
 305         struct inet_bind_bucket *tb = NULL;
 306         kuid_t uid = sock_i_uid(sk);
 307         int l3mdev;
 308
 309         l3mdev = inet_sk_bound_l3mdev(sk);
 310
 311         if (!port) {
 312                 head = inet_csk_find_open_port(sk, &tb, &port);
 313                 if (!head)
 314                         return ret;
 315                 if (!tb)
 316                         goto tb_not_found;
 317                 goto success;
 318         }
 319         head = &hinfo->bhash[inet_bhashfn(net, port,
 320                                           hinfo->bhash_size)];
 321         spin_lock_bh(&head->lock);
 322         inet_bind_bucket_for_each(tb, &head->chain)
 323                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 324                     tb->port == port)
 325                         goto tb_found;
 326 tb_not_found:
 327         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 328                                      net, head, port, l3mdev);
 329         if (!tb)
 330                 goto fail_unlock;
 331 tb_found:
 332         if (!hlist_empty(&tb->owners)) {
 333                 if (sk->sk_reuse == SK_FORCE_REUSE)
 334                         goto success;
 335
 336                 if ((tb->fastreuse > 0 && reuse) ||
 337                     sk_reuseport_match(tb, sk))
 338                         goto success;
 339                 if (inet_csk_bind_conflict(sk, tb, true, true))
 340                         goto fail_unlock;
 341         }
 342 success:
 343         if (hlist_empty(&tb->owners)) {
 344                 tb->fastreuse = reuse;
 345                 if (sk->sk_reuseport) {
 346                         tb->fastreuseport = FASTREUSEPORT_ANY;
 347                         tb->fastuid = uid;
 348                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 349                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 350                         tb->fast_sk_family = sk->sk_family;
 351 #if IS_ENABLED(CONFIG_IPV6)
 352                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 353 #endif
 354                 } else {
 355                         tb->fastreuseport = 0;
 356                 }
 357         } else {
 358                 if (!reuse)
 359                         tb->fastreuse = 0;
 360                 if (sk->sk_reuseport) {
 361                         /* We didn't match or we don't have fastreuseport set on
 362                          * the tb, but we have sk_reuseport set on this socket
 363                          * and we know that there are no bind conflicts with
 364                          * this socket in this tb, so reset our tb's reuseport
 365                          * settings so that any subsequent sockets that match
 366                          * our current socket will be put on the fast path.
 367                          *
 368                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 369                          * do extra checking for all subsequent sk_reuseport
 370                          * socks.
 371                          */
 372                         if (!sk_reuseport_match(tb, sk)) {
 373                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 374                                 tb->fastuid = uid;
 375                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 376                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 377                                 tb->fast_sk_family = sk->sk_family;
 378 #if IS_ENABLED(CONFIG_IPV6)
 379                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 380 #endif
 381                         }
 382                 } else {
 383                         tb->fastreuseport = 0;
 384                 }
 385         }
 386         if (!inet_csk(sk)->icsk_bind_hash)
 387                 inet_bind_hash(sk, tb, port);
 388         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 389         ret = 0;
 390
 391 fail_unlock:
 392         spin_unlock_bh(&head->lock);
 393         return ret;
 394 }
 395 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 396
 397 /*
 398  * Wait for an incoming connection, avoid race conditions. This must be called
 399  * with the socket locked.
 400  */
 401 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 402 {
 403         struct inet_connection_sock *icsk = inet_csk(sk);
 404         DEFINE_WAIT(wait);
 405         int err;
 406
 407         /*
 408          * True wake-one mechanism for incoming connections: only
 409          * one process gets woken up, not the 'whole herd'.
 410          * Since we do not 'race & poll' for established sockets
 411          * anymore, the common case will execute the loop only once.
 412          *
 413          * Subtle issue: "add_wait_queue_exclusive()" will be added
 414          * after any current non-exclusive waiters, and we know that
 415          * it will always _stay_ after any new non-exclusive waiters
 416          * because all non-exclusive waiters are added at the
 417          * beginning of the wait-queue. As such, it's ok to "drop"
 418          * our exclusiveness temporarily when we get woken up without
 419          * having to remove and re-insert us on the wait queue.
 420          */
 421         for (;;) {
 422                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 423                                           TASK_INTERRUPTIBLE);
 424                 release_sock(sk);
 425                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 426                         timeo = schedule_timeout(timeo);
 427                 sched_annotate_sleep();
 428                 lock_sock(sk);
 429                 err = 0;
 430                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 431                         break;
 432                 err = -EINVAL;
 433                 if (sk->sk_state != TCP_LISTEN)
 434                         break;
 435                 err = sock_intr_errno(timeo);
 436                 if (signal_pending(current))
 437                         break;
 438                 err = -EAGAIN;
 439                 if (!timeo)
 440                         break;
 441         }
 442         finish_wait(sk_sleep(sk), &wait);
 443         return err;
 444 }
 445
 446 /*
 447  * This will accept the next outstanding connection.
 448  */
 449 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 450 {
 451         struct inet_connection_sock *icsk = inet_csk(sk);
 452         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 453         struct request_sock *req;
 454         struct sock *newsk;
 455         int error;
 456
 457         lock_sock(sk);
 458
 459         /* We need to make sure that this socket is listening,
 460          * and that it has something pending.
 461          */
 462         error = -EINVAL;
 463         if (sk->sk_state != TCP_LISTEN)
 464                 goto out_err;
 465
 466         /* Find already established connection */
 467         if (reqsk_queue_empty(queue)) {
 468                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 469
 470                 /* If this is a non blocking socket don't sleep */
 471                 error = -EAGAIN;
 472                 if (!timeo)
 473                         goto out_err;
 474
 475                 error = inet_csk_wait_for_connect(sk, timeo);
 476                 if (error)
 477                         goto out_err;
 478         }
 479         req = reqsk_queue_remove(queue, sk);
 480         newsk = req->sk;
 481
 482         if (sk->sk_protocol == IPPROTO_TCP &&
 483             tcp_rsk(req)->tfo_listener) {
 484                 spin_lock_bh(&queue->fastopenq.lock);
 485                 if (tcp_rsk(req)->tfo_listener) {
 486                         /* We are still waiting for the final ACK from 3WHS
 487                          * so can't free req now. Instead, we set req->sk to
 488                          * NULL to signify that the child socket is taken
 489                          * so reqsk_fastopen_remove() will free the req
 490                          * when 3WHS finishes (or is aborted).
 491                          */
 492                         req->sk = NULL;
 493                         req = NULL;
 494                 }
 495                 spin_unlock_bh(&queue->fastopenq.lock);
 496         }
 497
 498 out:
 499         release_sock(sk);
 500         if (newsk && mem_cgroup_sockets_enabled) {
 501                 int amt;
 502
 503                 /* atomically get the memory usage, set and charge the
 504                  * newsk->sk_memcg.
 505                  */
 506                 lock_sock(newsk);
 507
 508                 /* The socket has not been accepted yet, no need to look at
 509                  * newsk->sk_wmem_queued.
 510                  */
 511                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 512                                    atomic_read(&newsk->sk_rmem_alloc));
 513                 mem_cgroup_sk_alloc(newsk);
 514                 if (newsk->sk_memcg && amt)
 515                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 516
 517                 release_sock(newsk);
 518         }
 519         if (req)
 520                 reqsk_put(req);
 521         return newsk;
 522 out_err:
 523         newsk = NULL;
 524         req = NULL;
 525         *err = error;
 526         goto out;
 527 }
 528 EXPORT_SYMBOL(inet_csk_accept);
 529
 530 /*
 531  * Using different timers for retransmit, delayed acks and probes
 532  * We may wish use just one timer maintaining a list of expire jiffies
 533  * to optimize.
 534  */
 535 void inet_csk_init_xmit_timers(struct sock *sk,
 536                                void (*retransmit_handler)(struct timer_list *t),
 537                                void (*delack_handler)(struct timer_list *t),
 538                                void (*keepalive_handler)(struct timer_list *t))
 539 {
 540         struct inet_connection_sock *icsk = inet_csk(sk);
 541
 542         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 543         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 544         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 545         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 546 }
 547 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 548
 549 void inet_csk_clear_xmit_timers(struct sock *sk)
 550 {
 551         struct inet_connection_sock *icsk = inet_csk(sk);
 552
 553         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 554
 555         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 556         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 557         sk_stop_timer(sk, &sk->sk_timer);
 558 }
 559 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 560
 561 void inet_csk_delete_keepalive_timer(struct sock *sk)
 562 {
 563         sk_stop_timer(sk, &sk->sk_timer);
 564 }
 565 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 566
 567 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 568 {
 569         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 570 }
 571 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 572
 573 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 574                                      struct flowi4 *fl4,
 575                                      const struct request_sock *req)
 576 {
 577         const struct inet_request_sock *ireq = inet_rsk(req);
 578         struct net *net = read_pnet(&ireq->ireq_net);
 579         struct ip_options_rcu *opt;
 580         struct rtable *rt;
 581
 582         rcu_read_lock();
 583         opt = rcu_dereference(ireq->ireq_opt);
 584
 585         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 586                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 587                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 588                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 589                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 590                            htons(ireq->ir_num), sk->sk_uid);
 591         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 592         rt = ip_route_output_flow(net, fl4, sk);
 593         if (IS_ERR(rt))
 594                 goto no_route;
 595         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 596                 goto route_err;
 597         rcu_read_unlock();
 598         return &rt->dst;
 599
 600 route_err:
 601         ip_rt_put(rt);
 602 no_route:
 603         rcu_read_unlock();
 604         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 605         return NULL;
 606 }
 607 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 608
 609 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 610                                             struct sock *newsk,
 611                                             const struct request_sock *req)
 612 {
 613         const struct inet_request_sock *ireq = inet_rsk(req);
 614         struct net *net = read_pnet(&ireq->ireq_net);
 615         struct inet_sock *newinet = inet_sk(newsk);
 616         struct ip_options_rcu *opt;
 617         struct flowi4 *fl4;
 618         struct rtable *rt;
 619
 620         opt = rcu_dereference(ireq->ireq_opt);
 621         fl4 = &newinet->cork.fl.u.ip4;
 622
 623         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 624                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 625                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 626                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 627                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 628                            htons(ireq->ir_num), sk->sk_uid);
 629         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 630         rt = ip_route_output_flow(net, fl4, sk);
 631         if (IS_ERR(rt))
 632                 goto no_route;
 633         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 634                 goto route_err;
 635         return &rt->dst;
 636
 637 route_err:
 638         ip_rt_put(rt);
 639 no_route:
 640         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 641         return NULL;
 642 }
 643 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 644
 645 /* Decide when to expire the request and when to resend SYN-ACK */
 646 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 647                                   const int max_retries,
 648                                   const u8 rskq_defer_accept,
 649                                   int *expire, int *resend)
 650 {
 651         if (!rskq_defer_accept) {
 652                 *expire = req->num_timeout >= thresh;
 653                 *resend = 1;
 654                 return;
 655         }
 656         *expire = req->num_timeout >= thresh &&
 657                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 658         /*
 659          * Do not resend while waiting for data after ACK,
 660          * start to resend on end of deferring period to give
 661          * last chance for data or ACK to create established socket.
 662          */
 663         *resend = !inet_rsk(req)->acked ||
 664                   req->num_timeout >= rskq_defer_accept - 1;
 665 }
 666
 667 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 668 {
 669         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 670
 671         if (!err)
 672                 req->num_retrans++;
 673         return err;
 674 }
 675 EXPORT_SYMBOL(inet_rtx_syn_ack);
 676
 677 /* return true if req was found in the ehash table */
 678 static bool reqsk_queue_unlink(struct request_sock *req)
 679 {
 680         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 681         bool found = false;
 682
 683         if (sk_hashed(req_to_sk(req))) {
 684                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 685
 686                 spin_lock(lock);
 687                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 688                 spin_unlock(lock);
 689         }
 690         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 691                 reqsk_put(req);
 692         return found;
 693 }
 694
 695 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 696 {
 697         if (reqsk_queue_unlink(req)) {
 698                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 699                 reqsk_put(req);
 700         }
 701 }
 702 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 703
 704 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 705 {
 706         inet_csk_reqsk_queue_drop(sk, req);
 707         reqsk_put(req);
 708 }
 709 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 710
 711 static void reqsk_timer_handler(struct timer_list *t)
 712 {
 713         struct request_sock *req = from_timer(req, t, rsk_timer);
 714         struct sock *sk_listener = req->rsk_listener;
 715         struct net *net = sock_net(sk_listener);
 716         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 717         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 718         int qlen, expire = 0, resend = 0;
 719         int max_retries, thresh;
 720         u8 defer_accept;
 721
 722         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 723                 goto drop;
 724
 725         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 726         thresh = max_retries;
 727         /* Normally all the openreqs are young and become mature
 728          * (i.e. converted to established socket) for first timeout.
 729          * If synack was not acknowledged for 1 second, it means
 730          * one of the following things: synack was lost, ack was lost,
 731          * rtt is high or nobody planned to ack (i.e. synflood).
 732          * When server is a bit loaded, queue is populated with old
 733          * open requests, reducing effective size of queue.
 734          * When server is well loaded, queue size reduces to zero
 735          * after several minutes of work. It is not synflood,
 736          * it is normal operation. The solution is pruning
 737          * too old entries overriding normal timeout, when
 738          * situation becomes dangerous.
 739          *
 740          * Essentially, we reserve half of room for young
 741          * embrions; and abort old ones without pity, if old
 742          * ones are about to clog our table.
 743          */
 744         qlen = reqsk_queue_len(queue);
 745         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
 746                 int young = reqsk_queue_len_young(queue) << 1;
 747
 748                 while (thresh > 2) {
 749                         if (qlen < young)
 750                                 break;
 751                         thresh--;
 752                         young <<= 1;
 753                 }
 754         }
 755         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 756         if (defer_accept)
 757                 max_retries = defer_accept;
 758         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 759                        &expire, &resend);
 760         req->rsk_ops->syn_ack_timeout(req);
 761         if (!expire &&
 762             (!resend ||
 763              !inet_rtx_syn_ack(sk_listener, req) ||
 764              inet_rsk(req)->acked)) {
 765                 unsigned long timeo;
 766
 767                 if (req->num_timeout++ == 0)
 768                         atomic_dec(&queue->young);
 769                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 770                 mod_timer(&req->rsk_timer, jiffies + timeo);
 771                 return;
 772         }
 773 drop:
 774         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 775 }
 776
 777 static void reqsk_queue_hash_req(struct request_sock *req,
 778                                  unsigned long timeout)
 779 {
 780         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 781         mod_timer(&req->rsk_timer, jiffies + timeout);
 782
 783         inet_ehash_insert(req_to_sk(req), NULL);
 784         /* before letting lookups find us, make sure all req fields
 785          * are committed to memory and refcnt initialized.
 786          */
 787         smp_wmb();
 788         refcount_set(&req->rsk_refcnt, 2 + 1);
 789 }
 790
 791 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 792                                    unsigned long timeout)
 793 {
 794         reqsk_queue_hash_req(req, timeout);
 795         inet_csk_reqsk_queue_added(sk);
 796 }
 797 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 798
 799 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
 800                            const gfp_t priority)
 801 {
 802         struct inet_connection_sock *icsk = inet_csk(newsk);
 803
 804         if (!icsk->icsk_ulp_ops)
 805                 return;
 806
 807         if (icsk->icsk_ulp_ops->clone)
 808                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
 809 }
 810
 811 /**
 812  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 813  *      @sk: the socket to clone
 814  *      @req: request_sock
 815  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 816  *
 817  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 818  */
 819 struct sock *inet_csk_clone_lock(const struct sock *sk,
 820                                  const struct request_sock *req,
 821                                  const gfp_t priority)
 822 {
 823         struct sock *newsk = sk_clone_lock(sk, priority);
 824
 825         if (newsk) {
 826                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 827
 828                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 829                 newicsk->icsk_bind_hash = NULL;
 830
 831                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 832                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 833                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 834
 835                 /* listeners have SOCK_RCU_FREE, not the children */
 836                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 837
 838                 inet_sk(newsk)->mc_list = NULL;
 839
 840                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 841                 atomic64_set(&newsk->sk_cookie,
 842                              atomic64_read(&inet_rsk(req)->ir_cookie));
 843
 844                 newicsk->icsk_retransmits = 0;
 845                 newicsk->icsk_backoff     = 0;
 846                 newicsk->icsk_probes_out  = 0;
 847
 848                 /* Deinitialize accept_queue to trap illegal accesses. */
 849                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 850
 851                 inet_clone_ulp(req, newsk, priority);
 852
 853                 security_inet_csk_clone(newsk, req);
 854         }
 855         return newsk;
 856 }
 857 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 858
 859 /*
 860  * At this point, there should be no process reference to this
 861  * socket, and thus no user references at all.  Therefore we
 862  * can assume the socket waitqueue is inactive and nobody will
 863  * try to jump onto it.
 864  */
 865 void inet_csk_destroy_sock(struct sock *sk)
 866 {
 867         WARN_ON(sk->sk_state != TCP_CLOSE);
 868         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 869
 870         /* It cannot be in hash table! */
 871         WARN_ON(!sk_unhashed(sk));
 872
 873         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 874         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 875
 876         sk->sk_prot->destroy(sk);
 877
 878         sk_stream_kill_queues(sk);
 879
 880         xfrm_sk_free_policy(sk);
 881
 882         sk_refcnt_debug_release(sk);
 883
 884         percpu_counter_dec(sk->sk_prot->orphan_count);
 885
 886         sock_put(sk);
 887 }
 888 EXPORT_SYMBOL(inet_csk_destroy_sock);
 889
 890 /* This function allows to force a closure of a socket after the call to
 891  * tcp/dccp_create_openreq_child().
 892  */
 893 void inet_csk_prepare_forced_close(struct sock *sk)
 894         __releases(&sk->sk_lock.slock)
 895 {
 896         /* sk_clone_lock locked the socket and set refcnt to 2 */
 897         bh_unlock_sock(sk);
 898         sock_put(sk);
 899
 900         /* The below has to be done to allow calling inet_csk_destroy_sock */
 901         sock_set_flag(sk, SOCK_DEAD);
 902         percpu_counter_inc(sk->sk_prot->orphan_count);
 903         inet_sk(sk)->inet_num = 0;
 904 }
 905 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 906
 907 int inet_csk_listen_start(struct sock *sk, int backlog)
 908 {
 909         struct inet_connection_sock *icsk = inet_csk(sk);
 910         struct inet_sock *inet = inet_sk(sk);
 911         int err = -EADDRINUSE;
 912
 913         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 914
 915         sk->sk_ack_backlog = 0;
 916         inet_csk_delack_init(sk);
 917
 918         /* There is race window here: we announce ourselves listening,
 919          * but this transition is still not validated by get_port().
 920          * It is OK, because this socket enters to hash table only
 921          * after validation is complete.
 922          */
 923         inet_sk_state_store(sk, TCP_LISTEN);
 924         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 925                 inet->inet_sport = htons(inet->inet_num);
 926
 927                 sk_dst_reset(sk);
 928                 err = sk->sk_prot->hash(sk);
 929
 930                 if (likely(!err))
 931                         return 0;
 932         }
 933
 934         inet_sk_set_state(sk, TCP_CLOSE);
 935         return err;
 936 }
 937 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 938
 939 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 940                               struct sock *child)
 941 {
 942         sk->sk_prot->disconnect(child, O_NONBLOCK);
 943
 944         sock_orphan(child);
 945
 946         percpu_counter_inc(sk->sk_prot->orphan_count);
 947
 948         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 949                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
 950                 BUG_ON(sk != req->rsk_listener);
 951
 952                 /* Paranoid, to prevent race condition if
 953                  * an inbound pkt destined for child is
 954                  * blocked by sock lock in tcp_v4_rcv().
 955                  * Also to satisfy an assertion in
 956                  * tcp_v4_destroy_sock().
 957                  */
 958                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
 959         }
 960         inet_csk_destroy_sock(child);
 961 }
 962
 963 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 964                                       struct request_sock *req,
 965                                       struct sock *child)
 966 {
 967         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 968
 969         spin_lock(&queue->rskq_lock);
 970         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 971                 inet_child_forget(sk, req, child);
 972                 child = NULL;
 973         } else {
 974                 req->sk = child;
 975                 req->dl_next = NULL;
 976                 if (queue->rskq_accept_head == NULL)
 977                         WRITE_ONCE(queue->rskq_accept_head, req);
 978                 else
 979                         queue->rskq_accept_tail->dl_next = req;
 980                 queue->rskq_accept_tail = req;
 981                 sk_acceptq_added(sk);
 982         }
 983         spin_unlock(&queue->rskq_lock);
 984         return child;
 985 }
 986 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 987
 988 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 989                                          struct request_sock *req, bool own_req)
 990 {
 991         if (own_req) {
 992                 inet_csk_reqsk_queue_drop(sk, req);
 993                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 994                 if (inet_csk_reqsk_queue_add(sk, req, child))
 995                         return child;
 996         }
 997         /* Too bad, another child took ownership of the request, undo. */
 998         bh_unlock_sock(child);
 999         sock_put(child);
1000         return NULL;
1001 }
1002 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1003
1004 /*
1005  *      This routine closes sockets which have been at least partially
1006  *      opened, but not yet accepted.
1007  */
1008 void inet_csk_listen_stop(struct sock *sk)
1009 {
1010         struct inet_connection_sock *icsk = inet_csk(sk);
1011         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1012         struct request_sock *next, *req;
1013
1014         /* Following specs, it would be better either to send FIN
1015          * (and enter FIN-WAIT-1, it is normal close)
1016          * or to send active reset (abort).
1017          * Certainly, it is pretty dangerous while synflood, but it is
1018          * bad justification for our negligence 8)
1019          * To be honest, we are not able to make either
1020          * of the variants now.                 --ANK
1021          */
1022         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1023                 struct sock *child = req->sk;
1024
1025                 local_bh_disable();
1026                 bh_lock_sock(child);
1027                 WARN_ON(sock_owned_by_user(child));
1028                 sock_hold(child);
1029
1030                 inet_child_forget(sk, req, child);
1031                 reqsk_put(req);
1032                 bh_unlock_sock(child);
1033                 local_bh_enable();
1034                 sock_put(child);
1035
1036                 cond_resched();
1037         }
1038         if (queue->fastopenq.rskq_rst_head) {
1039                 /* Free all the reqs queued in rskq_rst_head. */
1040                 spin_lock_bh(&queue->fastopenq.lock);
1041                 req = queue->fastopenq.rskq_rst_head;
1042                 queue->fastopenq.rskq_rst_head = NULL;
1043                 spin_unlock_bh(&queue->fastopenq.lock);
1044                 while (req != NULL) {
1045                         next = req->dl_next;
1046                         reqsk_put(req);
1047                         req = next;
1048                 }
1049         }
1050         WARN_ON_ONCE(sk->sk_ack_backlog);
1051 }
1052 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1053
1054 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1055 {
1056         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1057         const struct inet_sock *inet = inet_sk(sk);
1058
1059         sin->sin_family         = AF_INET;
1060         sin->sin_addr.s_addr    = inet->inet_daddr;
1061         sin->sin_port           = inet->inet_dport;
1062 }
1063 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1064
1065 #ifdef CONFIG_COMPAT
1066 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
1067                                char __user *optval, int __user *optlen)
1068 {
1069         const struct inet_connection_sock *icsk = inet_csk(sk);
1070
1071         if (icsk->icsk_af_ops->compat_getsockopt)
1072                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
1073                                                             optval, optlen);
1074         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
1075                                              optval, optlen);
1076 }
1077 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
1078
1079 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
1080                                char __user *optval, unsigned int optlen)
1081 {
1082         const struct inet_connection_sock *icsk = inet_csk(sk);
1083
1084         if (icsk->icsk_af_ops->compat_setsockopt)
1085                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
1086                                                             optval, optlen);
1087         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1088                                              optval, optlen);
1089 }
1090 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
1091 #endif
1092
1093 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1094 {
1095         const struct inet_sock *inet = inet_sk(sk);
1096         const struct ip_options_rcu *inet_opt;
1097         __be32 daddr = inet->inet_daddr;
1098         struct flowi4 *fl4;
1099         struct rtable *rt;
1100
1101         rcu_read_lock();
1102         inet_opt = rcu_dereference(inet->inet_opt);
1103         if (inet_opt && inet_opt->opt.srr)
1104                 daddr = inet_opt->opt.faddr;
1105         fl4 = &fl->u.ip4;
1106         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1107                                    inet->inet_saddr, inet->inet_dport,
1108                                    inet->inet_sport, sk->sk_protocol,
1109                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1110         if (IS_ERR(rt))
1111                 rt = NULL;
1112         if (rt)
1113                 sk_setup_caps(sk, &rt->dst);
1114         rcu_read_unlock();
1115
1116         return &rt->dst;
1117 }
1118
1119 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1120 {
1121         struct dst_entry *dst = __sk_dst_check(sk, 0);
1122         struct inet_sock *inet = inet_sk(sk);
1123
1124         if (!dst) {
1125                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1126                 if (!dst)
1127                         goto out;
1128         }
1129         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1130
1131         dst = __sk_dst_check(sk, 0);
1132         if (!dst)
1133                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1134 out:
1135         return dst;
1136 }
1137 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);