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_sk*_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses
  28  *                              if IPv6 only, and any IPv4 addresses
  29  *                              if not IPv6 only
  30  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  31  *                              IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  32  *                              and 0.0.0.0 equals to 0.0.0.0 only
  33  */
  34 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  35                                  const struct in6_addr *sk2_rcv_saddr6,
  36                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  37                                  bool sk1_ipv6only, bool sk2_ipv6only,
  38                                  bool match_sk1_wildcard,
  39                                  bool match_sk2_wildcard)
  40 {
  41         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  42         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  43
  44         /* if both are mapped, treat as IPv4 */
  45         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  46                 if (!sk2_ipv6only) {
  47                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  48                                 return true;
  49                         return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  50                                 (match_sk2_wildcard && !sk2_rcv_saddr);
  51                 }
  52                 return false;
  53         }
  54
  55         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  56                 return true;
  57
  58         if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard &&
  59             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  60                 return true;
  61
  62         if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard &&
  63             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  64                 return true;
  65
  66         if (sk2_rcv_saddr6 &&
  67             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  68                 return true;
  69
  70         return false;
  71 }
  72 #endif
  73
  74 /* match_sk*_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  75  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  76  *                              0.0.0.0 only equals to 0.0.0.0
  77  */
  78 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  79                                  bool sk2_ipv6only, bool match_sk1_wildcard,
  80                                  bool match_sk2_wildcard)
  81 {
  82         if (!sk2_ipv6only) {
  83                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  84                         return true;
  85                 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  86                         (match_sk2_wildcard && !sk2_rcv_saddr);
  87         }
  88         return false;
  89 }
  90
  91 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  92                           bool match_wildcard)
  93 {
  94 #if IS_ENABLED(CONFIG_IPV6)
  95         if (sk->sk_family == AF_INET6)
  96                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
  97                                             inet6_rcv_saddr(sk2),
  98                                             sk->sk_rcv_saddr,
  99                                             sk2->sk_rcv_saddr,
 100                                             ipv6_only_sock(sk),
 101                                             ipv6_only_sock(sk2),
 102                                             match_wildcard,
 103                                             match_wildcard);
 104 #endif
 105         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 106                                     ipv6_only_sock(sk2), match_wildcard,
 107                                     match_wildcard);
 108 }
 109 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 110
 111 bool inet_rcv_saddr_any(const struct sock *sk)
 112 {
 113 #if IS_ENABLED(CONFIG_IPV6)
 114         if (sk->sk_family == AF_INET6)
 115                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 116 #endif
 117         return !sk->sk_rcv_saddr;
 118 }
 119
 120 void inet_get_local_port_range(struct net *net, int *low, int *high)
 121 {
 122         unsigned int seq;
 123
 124         do {
 125                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 126
 127                 *low = net->ipv4.ip_local_ports.range[0];
 128                 *high = net->ipv4.ip_local_ports.range[1];
 129         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 130 }
 131 EXPORT_SYMBOL(inet_get_local_port_range);
 132
 133 static int inet_csk_bind_conflict(const struct sock *sk,
 134                                   const struct inet_bind_bucket *tb,
 135                                   bool relax, bool reuseport_ok)
 136 {
 137         struct sock *sk2;
 138         bool reuse = sk->sk_reuse;
 139         bool reuseport = !!sk->sk_reuseport;
 140         kuid_t uid = sock_i_uid((struct sock *)sk);
 141
 142         /*
 143          * Unlike other sk lookup places we do not check
 144          * for sk_net here, since _all_ the socks listed
 145          * in tb->owners list belong to the same net - the
 146          * one this bucket belongs to.
 147          */
 148
 149         sk_for_each_bound(sk2, &tb->owners) {
 150                 if (sk != sk2 &&
 151                     (!sk->sk_bound_dev_if ||
 152                      !sk2->sk_bound_dev_if ||
 153                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 154                         if (reuse && sk2->sk_reuse &&
 155                             sk2->sk_state != TCP_LISTEN) {
 156                                 if ((!relax ||
 157                                      (!reuseport_ok &&
 158                                       reuseport && sk2->sk_reuseport &&
 159                                       !rcu_access_pointer(sk->sk_reuseport_cb) &&
 160                                       (sk2->sk_state == TCP_TIME_WAIT ||
 161                                        uid_eq(uid, sock_i_uid(sk2))))) &&
 162                                     inet_rcv_saddr_equal(sk, sk2, true))
 163                                         break;
 164                         } else if (!reuseport_ok ||
 165                                    !reuseport || !sk2->sk_reuseport ||
 166                                    rcu_access_pointer(sk->sk_reuseport_cb) ||
 167                                    (sk2->sk_state != TCP_TIME_WAIT &&
 168                                     !uid_eq(uid, sock_i_uid(sk2)))) {
 169                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 170                                         break;
 171                         }
 172                 }
 173         }
 174         return sk2 != NULL;
 175 }
 176
 177 /*
 178  * Find an open port number for the socket.  Returns with the
 179  * inet_bind_hashbucket lock held.
 180  */
 181 static struct inet_bind_hashbucket *
 182 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 183 {
 184         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 185         int port = 0;
 186         struct inet_bind_hashbucket *head;
 187         struct net *net = sock_net(sk);
 188         bool relax = false;
 189         int i, low, high, attempt_half;
 190         struct inet_bind_bucket *tb;
 191         u32 remaining, offset;
 192         int l3mdev;
 193
 194         l3mdev = inet_sk_bound_l3mdev(sk);
 195 ports_exhausted:
 196         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 197 other_half_scan:
 198         inet_get_local_port_range(net, &low, &high);
 199         high++; /* [32768, 60999] -> [32768, 61000[ */
 200         if (high - low < 4)
 201                 attempt_half = 0;
 202         if (attempt_half) {
 203                 int half = low + (((high - low) >> 2) << 1);
 204
 205                 if (attempt_half == 1)
 206                         high = half;
 207                 else
 208                         low = half;
 209         }
 210         remaining = high - low;
 211         if (likely(remaining > 1))
 212                 remaining &= ~1U;
 213
 214         offset = prandom_u32() % remaining;
 215         /* __inet_hash_connect() favors ports having @low parity
 216          * We do the opposite to not pollute connect() users.
 217          */
 218         offset |= 1U;
 219
 220 other_parity_scan:
 221         port = low + offset;
 222         for (i = 0; i < remaining; i += 2, port += 2) {
 223                 if (unlikely(port >= high))
 224                         port -= remaining;
 225                 if (inet_is_local_reserved_port(net, port))
 226                         continue;
 227                 head = &hinfo->bhash[inet_bhashfn(net, port,
 228                                                   hinfo->bhash_size)];
 229                 spin_lock_bh(&head->lock);
 230                 inet_bind_bucket_for_each(tb, &head->chain)
 231                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 232                             tb->port == port) {
 233                                 if (!inet_csk_bind_conflict(sk, tb, relax, false))
 234                                         goto success;
 235                                 goto next_port;
 236                         }
 237                 tb = NULL;
 238                 goto success;
 239 next_port:
 240                 spin_unlock_bh(&head->lock);
 241                 cond_resched();
 242         }
 243
 244         offset--;
 245         if (!(offset & 1))
 246                 goto other_parity_scan;
 247
 248         if (attempt_half == 1) {
 249                 /* OK we now try the upper half of the range */
 250                 attempt_half = 2;
 251                 goto other_half_scan;
 252         }
 253
 254         if (net->ipv4.sysctl_ip_autobind_reuse && !relax) {
 255                 /* We still have a chance to connect to different destinations */
 256                 relax = true;
 257                 goto ports_exhausted;
 258         }
 259         return NULL;
 260 success:
 261         *port_ret = port;
 262         *tb_ret = tb;
 263         return head;
 264 }
 265
 266 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 267                                      struct sock *sk)
 268 {
 269         kuid_t uid = sock_i_uid(sk);
 270
 271         if (tb->fastreuseport <= 0)
 272                 return 0;
 273         if (!sk->sk_reuseport)
 274                 return 0;
 275         if (rcu_access_pointer(sk->sk_reuseport_cb))
 276                 return 0;
 277         if (!uid_eq(tb->fastuid, uid))
 278                 return 0;
 279         /* We only need to check the rcv_saddr if this tb was once marked
 280          * without fastreuseport and then was reset, as we can only know that
 281          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 282          * owners list.
 283          */
 284         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 285                 return 1;
 286 #if IS_ENABLED(CONFIG_IPV6)
 287         if (tb->fast_sk_family == AF_INET6)
 288                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 289                                             inet6_rcv_saddr(sk),
 290                                             tb->fast_rcv_saddr,
 291                                             sk->sk_rcv_saddr,
 292                                             tb->fast_ipv6_only,
 293                                             ipv6_only_sock(sk), true, false);
 294 #endif
 295         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 296                                     ipv6_only_sock(sk), true, false);
 297 }
 298
 299 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 300                                struct sock *sk)
 301 {
 302         kuid_t uid = sock_i_uid(sk);
 303         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 304
 305         if (hlist_empty(&tb->owners)) {
 306                 tb->fastreuse = reuse;
 307                 if (sk->sk_reuseport) {
 308                         tb->fastreuseport = FASTREUSEPORT_ANY;
 309                         tb->fastuid = uid;
 310                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 311                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 312                         tb->fast_sk_family = sk->sk_family;
 313 #if IS_ENABLED(CONFIG_IPV6)
 314                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 315 #endif
 316                 } else {
 317                         tb->fastreuseport = 0;
 318                 }
 319         } else {
 320                 if (!reuse)
 321                         tb->fastreuse = 0;
 322                 if (sk->sk_reuseport) {
 323                         /* We didn't match or we don't have fastreuseport set on
 324                          * the tb, but we have sk_reuseport set on this socket
 325                          * and we know that there are no bind conflicts with
 326                          * this socket in this tb, so reset our tb's reuseport
 327                          * settings so that any subsequent sockets that match
 328                          * our current socket will be put on the fast path.
 329                          *
 330                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 331                          * do extra checking for all subsequent sk_reuseport
 332                          * socks.
 333                          */
 334                         if (!sk_reuseport_match(tb, sk)) {
 335                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 336                                 tb->fastuid = uid;
 337                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 338                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 339                                 tb->fast_sk_family = sk->sk_family;
 340 #if IS_ENABLED(CONFIG_IPV6)
 341                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 342 #endif
 343                         }
 344                 } else {
 345                         tb->fastreuseport = 0;
 346                 }
 347         }
 348 }
 349
 350 /* Obtain a reference to a local port for the given sock,
 351  * if snum is zero it means select any available local port.
 352  * We try to allocate an odd port (and leave even ports for connect())
 353  */
 354 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 355 {
 356         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 357         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 358         int ret = 1, port = snum;
 359         struct inet_bind_hashbucket *head;
 360         struct net *net = sock_net(sk);
 361         struct inet_bind_bucket *tb = NULL;
 362         int l3mdev;
 363
 364         l3mdev = inet_sk_bound_l3mdev(sk);
 365
 366         if (!port) {
 367                 head = inet_csk_find_open_port(sk, &tb, &port);
 368                 if (!head)
 369                         return ret;
 370                 if (!tb)
 371                         goto tb_not_found;
 372                 goto success;
 373         }
 374         head = &hinfo->bhash[inet_bhashfn(net, port,
 375                                           hinfo->bhash_size)];
 376         spin_lock_bh(&head->lock);
 377         inet_bind_bucket_for_each(tb, &head->chain)
 378                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 379                     tb->port == port)
 380                         goto tb_found;
 381 tb_not_found:
 382         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 383                                      net, head, port, l3mdev);
 384         if (!tb)
 385                 goto fail_unlock;
 386 tb_found:
 387         if (!hlist_empty(&tb->owners)) {
 388                 if (sk->sk_reuse == SK_FORCE_REUSE)
 389                         goto success;
 390
 391                 if ((tb->fastreuse > 0 && reuse) ||
 392                     sk_reuseport_match(tb, sk))
 393                         goto success;
 394                 if (inet_csk_bind_conflict(sk, tb, true, true))
 395                         goto fail_unlock;
 396         }
 397 success:
 398         inet_csk_update_fastreuse(tb, sk);
 399
 400         if (!inet_csk(sk)->icsk_bind_hash)
 401                 inet_bind_hash(sk, tb, port);
 402         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 403         ret = 0;
 404
 405 fail_unlock:
 406         spin_unlock_bh(&head->lock);
 407         return ret;
 408 }
 409 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 410
 411 /*
 412  * Wait for an incoming connection, avoid race conditions. This must be called
 413  * with the socket locked.
 414  */
 415 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 416 {
 417         struct inet_connection_sock *icsk = inet_csk(sk);
 418         DEFINE_WAIT(wait);
 419         int err;
 420
 421         /*
 422          * True wake-one mechanism for incoming connections: only
 423          * one process gets woken up, not the 'whole herd'.
 424          * Since we do not 'race & poll' for established sockets
 425          * anymore, the common case will execute the loop only once.
 426          *
 427          * Subtle issue: "add_wait_queue_exclusive()" will be added
 428          * after any current non-exclusive waiters, and we know that
 429          * it will always _stay_ after any new non-exclusive waiters
 430          * because all non-exclusive waiters are added at the
 431          * beginning of the wait-queue. As such, it's ok to "drop"
 432          * our exclusiveness temporarily when we get woken up without
 433          * having to remove and re-insert us on the wait queue.
 434          */
 435         for (;;) {
 436                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 437                                           TASK_INTERRUPTIBLE);
 438                 release_sock(sk);
 439                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 440                         timeo = schedule_timeout(timeo);
 441                 sched_annotate_sleep();
 442                 lock_sock(sk);
 443                 err = 0;
 444                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 445                         break;
 446                 err = -EINVAL;
 447                 if (sk->sk_state != TCP_LISTEN)
 448                         break;
 449                 err = sock_intr_errno(timeo);
 450                 if (signal_pending(current))
 451                         break;
 452                 err = -EAGAIN;
 453                 if (!timeo)
 454                         break;
 455         }
 456         finish_wait(sk_sleep(sk), &wait);
 457         return err;
 458 }
 459
 460 /*
 461  * This will accept the next outstanding connection.
 462  */
 463 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 464 {
 465         struct inet_connection_sock *icsk = inet_csk(sk);
 466         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 467         struct request_sock *req;
 468         struct sock *newsk;
 469         int error;
 470
 471         lock_sock(sk);
 472
 473         /* We need to make sure that this socket is listening,
 474          * and that it has something pending.
 475          */
 476         error = -EINVAL;
 477         if (sk->sk_state != TCP_LISTEN)
 478                 goto out_err;
 479
 480         /* Find already established connection */
 481         if (reqsk_queue_empty(queue)) {
 482                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 483
 484                 /* If this is a non blocking socket don't sleep */
 485                 error = -EAGAIN;
 486                 if (!timeo)
 487                         goto out_err;
 488
 489                 error = inet_csk_wait_for_connect(sk, timeo);
 490                 if (error)
 491                         goto out_err;
 492         }
 493         req = reqsk_queue_remove(queue, sk);
 494         newsk = req->sk;
 495
 496         if (sk->sk_protocol == IPPROTO_TCP &&
 497             tcp_rsk(req)->tfo_listener) {
 498                 spin_lock_bh(&queue->fastopenq.lock);
 499                 if (tcp_rsk(req)->tfo_listener) {
 500                         /* We are still waiting for the final ACK from 3WHS
 501                          * so can't free req now. Instead, we set req->sk to
 502                          * NULL to signify that the child socket is taken
 503                          * so reqsk_fastopen_remove() will free the req
 504                          * when 3WHS finishes (or is aborted).
 505                          */
 506                         req->sk = NULL;
 507                         req = NULL;
 508                 }
 509                 spin_unlock_bh(&queue->fastopenq.lock);
 510         }
 511
 512 out:
 513         release_sock(sk);
 514         if (newsk && mem_cgroup_sockets_enabled) {
 515                 int amt;
 516
 517                 /* atomically get the memory usage, set and charge the
 518                  * newsk->sk_memcg.
 519                  */
 520                 lock_sock(newsk);
 521
 522                 /* The socket has not been accepted yet, no need to look at
 523                  * newsk->sk_wmem_queued.
 524                  */
 525                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 526                                    atomic_read(&newsk->sk_rmem_alloc));
 527                 mem_cgroup_sk_alloc(newsk);
 528                 if (newsk->sk_memcg && amt)
 529                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 530
 531                 release_sock(newsk);
 532         }
 533         if (req)
 534                 reqsk_put(req);
 535         return newsk;
 536 out_err:
 537         newsk = NULL;
 538         req = NULL;
 539         *err = error;
 540         goto out;
 541 }
 542 EXPORT_SYMBOL(inet_csk_accept);
 543
 544 /*
 545  * Using different timers for retransmit, delayed acks and probes
 546  * We may wish use just one timer maintaining a list of expire jiffies
 547  * to optimize.
 548  */
 549 void inet_csk_init_xmit_timers(struct sock *sk,
 550                                void (*retransmit_handler)(struct timer_list *t),
 551                                void (*delack_handler)(struct timer_list *t),
 552                                void (*keepalive_handler)(struct timer_list *t))
 553 {
 554         struct inet_connection_sock *icsk = inet_csk(sk);
 555
 556         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 557         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 558         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 559         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 560 }
 561 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 562
 563 void inet_csk_clear_xmit_timers(struct sock *sk)
 564 {
 565         struct inet_connection_sock *icsk = inet_csk(sk);
 566
 567         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 568
 569         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 570         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 571         sk_stop_timer(sk, &sk->sk_timer);
 572 }
 573 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 574
 575 void inet_csk_delete_keepalive_timer(struct sock *sk)
 576 {
 577         sk_stop_timer(sk, &sk->sk_timer);
 578 }
 579 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 580
 581 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 582 {
 583         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 584 }
 585 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 586
 587 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 588                                      struct flowi4 *fl4,
 589                                      const struct request_sock *req)
 590 {
 591         const struct inet_request_sock *ireq = inet_rsk(req);
 592         struct net *net = read_pnet(&ireq->ireq_net);
 593         struct ip_options_rcu *opt;
 594         struct rtable *rt;
 595
 596         rcu_read_lock();
 597         opt = rcu_dereference(ireq->ireq_opt);
 598
 599         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 600                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 601                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 602                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 603                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 604                            htons(ireq->ir_num), sk->sk_uid);
 605         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 606         rt = ip_route_output_flow(net, fl4, sk);
 607         if (IS_ERR(rt))
 608                 goto no_route;
 609         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 610                 goto route_err;
 611         rcu_read_unlock();
 612         return &rt->dst;
 613
 614 route_err:
 615         ip_rt_put(rt);
 616 no_route:
 617         rcu_read_unlock();
 618         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 619         return NULL;
 620 }
 621 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 622
 623 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 624                                             struct sock *newsk,
 625                                             const struct request_sock *req)
 626 {
 627         const struct inet_request_sock *ireq = inet_rsk(req);
 628         struct net *net = read_pnet(&ireq->ireq_net);
 629         struct inet_sock *newinet = inet_sk(newsk);
 630         struct ip_options_rcu *opt;
 631         struct flowi4 *fl4;
 632         struct rtable *rt;
 633
 634         opt = rcu_dereference(ireq->ireq_opt);
 635         fl4 = &newinet->cork.fl.u.ip4;
 636
 637         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 638                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 639                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 640                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 641                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 642                            htons(ireq->ir_num), sk->sk_uid);
 643         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 644         rt = ip_route_output_flow(net, fl4, sk);
 645         if (IS_ERR(rt))
 646                 goto no_route;
 647         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 648                 goto route_err;
 649         return &rt->dst;
 650
 651 route_err:
 652         ip_rt_put(rt);
 653 no_route:
 654         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 655         return NULL;
 656 }
 657 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 658
 659 /* Decide when to expire the request and when to resend SYN-ACK */
 660 static void syn_ack_recalc(struct request_sock *req,
 661                            const int max_syn_ack_retries,
 662                            const u8 rskq_defer_accept,
 663                            int *expire, int *resend)
 664 {
 665         if (!rskq_defer_accept) {
 666                 *expire = req->num_timeout >= max_syn_ack_retries;
 667                 *resend = 1;
 668                 return;
 669         }
 670         *expire = req->num_timeout >= max_syn_ack_retries &&
 671                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 672         /* Do not resend while waiting for data after ACK,
 673          * start to resend on end of deferring period to give
 674          * last chance for data or ACK to create established socket.
 675          */
 676         *resend = !inet_rsk(req)->acked ||
 677                   req->num_timeout >= rskq_defer_accept - 1;
 678 }
 679
 680 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 681 {
 682         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 683
 684         if (!err)
 685                 req->num_retrans++;
 686         return err;
 687 }
 688 EXPORT_SYMBOL(inet_rtx_syn_ack);
 689
 690 /* return true if req was found in the ehash table */
 691 static bool reqsk_queue_unlink(struct request_sock *req)
 692 {
 693         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 694         bool found = false;
 695
 696         if (sk_hashed(req_to_sk(req))) {
 697                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 698
 699                 spin_lock(lock);
 700                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 701                 spin_unlock(lock);
 702         }
 703         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 704                 reqsk_put(req);
 705         return found;
 706 }
 707
 708 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 709 {
 710         if (reqsk_queue_unlink(req)) {
 711                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 712                 reqsk_put(req);
 713         }
 714 }
 715 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 716
 717 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 718 {
 719         inet_csk_reqsk_queue_drop(sk, req);
 720         reqsk_put(req);
 721 }
 722 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 723
 724 static void reqsk_timer_handler(struct timer_list *t)
 725 {
 726         struct request_sock *req = from_timer(req, t, rsk_timer);
 727         struct sock *sk_listener = req->rsk_listener;
 728         struct net *net = sock_net(sk_listener);
 729         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 730         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 731         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
 732
 733         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 734                 goto drop;
 735
 736         max_syn_ack_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 737         /* Normally all the openreqs are young and become mature
 738          * (i.e. converted to established socket) for first timeout.
 739          * If synack was not acknowledged for 1 second, it means
 740          * one of the following things: synack was lost, ack was lost,
 741          * rtt is high or nobody planned to ack (i.e. synflood).
 742          * When server is a bit loaded, queue is populated with old
 743          * open requests, reducing effective size of queue.
 744          * When server is well loaded, queue size reduces to zero
 745          * after several minutes of work. It is not synflood,
 746          * it is normal operation. The solution is pruning
 747          * too old entries overriding normal timeout, when
 748          * situation becomes dangerous.
 749          *
 750          * Essentially, we reserve half of room for young
 751          * embrions; and abort old ones without pity, if old
 752          * ones are about to clog our table.
 753          */
 754         qlen = reqsk_queue_len(queue);
 755         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
 756                 int young = reqsk_queue_len_young(queue) << 1;
 757
 758                 while (max_syn_ack_retries > 2) {
 759                         if (qlen < young)
 760                                 break;
 761                         max_syn_ack_retries--;
 762                         young <<= 1;
 763                 }
 764         }
 765         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
 766                        &expire, &resend);
 767         req->rsk_ops->syn_ack_timeout(req);
 768         if (!expire &&
 769             (!resend ||
 770              !inet_rtx_syn_ack(sk_listener, req) ||
 771              inet_rsk(req)->acked)) {
 772                 unsigned long timeo;
 773
 774                 if (req->num_timeout++ == 0)
 775                         atomic_dec(&queue->young);
 776                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 777                 mod_timer(&req->rsk_timer, jiffies + timeo);
 778                 return;
 779         }
 780 drop:
 781         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 782 }
 783
 784 static void reqsk_queue_hash_req(struct request_sock *req,
 785                                  unsigned long timeout)
 786 {
 787         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 788         mod_timer(&req->rsk_timer, jiffies + timeout);
 789
 790         inet_ehash_insert(req_to_sk(req), NULL, NULL);
 791         /* before letting lookups find us, make sure all req fields
 792          * are committed to memory and refcnt initialized.
 793          */
 794         smp_wmb();
 795         refcount_set(&req->rsk_refcnt, 2 + 1);
 796 }
 797
 798 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 799                                    unsigned long timeout)
 800 {
 801         reqsk_queue_hash_req(req, timeout);
 802         inet_csk_reqsk_queue_added(sk);
 803 }
 804 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 805
 806 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
 807                            const gfp_t priority)
 808 {
 809         struct inet_connection_sock *icsk = inet_csk(newsk);
 810
 811         if (!icsk->icsk_ulp_ops)
 812                 return;
 813
 814         if (icsk->icsk_ulp_ops->clone)
 815                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
 816 }
 817
 818 /**
 819  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 820  *      @sk: the socket to clone
 821  *      @req: request_sock
 822  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 823  *
 824  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 825  */
 826 struct sock *inet_csk_clone_lock(const struct sock *sk,
 827                                  const struct request_sock *req,
 828                                  const gfp_t priority)
 829 {
 830         struct sock *newsk = sk_clone_lock(sk, priority);
 831
 832         if (newsk) {
 833                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 834
 835                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 836                 newicsk->icsk_bind_hash = NULL;
 837
 838                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 839                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 840                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 841
 842                 /* listeners have SOCK_RCU_FREE, not the children */
 843                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 844
 845                 inet_sk(newsk)->mc_list = NULL;
 846
 847                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 848                 atomic64_set(&newsk->sk_cookie,
 849                              atomic64_read(&inet_rsk(req)->ir_cookie));
 850
 851                 newicsk->icsk_retransmits = 0;
 852                 newicsk->icsk_backoff     = 0;
 853                 newicsk->icsk_probes_out  = 0;
 854
 855                 /* Deinitialize accept_queue to trap illegal accesses. */
 856                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 857
 858                 inet_clone_ulp(req, newsk, priority);
 859
 860                 security_inet_csk_clone(newsk, req);
 861         }
 862         return newsk;
 863 }
 864 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 865
 866 /*
 867  * At this point, there should be no process reference to this
 868  * socket, and thus no user references at all.  Therefore we
 869  * can assume the socket waitqueue is inactive and nobody will
 870  * try to jump onto it.
 871  */
 872 void inet_csk_destroy_sock(struct sock *sk)
 873 {
 874         WARN_ON(sk->sk_state != TCP_CLOSE);
 875         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 876
 877         /* It cannot be in hash table! */
 878         WARN_ON(!sk_unhashed(sk));
 879
 880         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 881         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 882
 883         sk->sk_prot->destroy(sk);
 884
 885         sk_stream_kill_queues(sk);
 886
 887         xfrm_sk_free_policy(sk);
 888
 889         sk_refcnt_debug_release(sk);
 890
 891         percpu_counter_dec(sk->sk_prot->orphan_count);
 892
 893         sock_put(sk);
 894 }
 895 EXPORT_SYMBOL(inet_csk_destroy_sock);
 896
 897 /* This function allows to force a closure of a socket after the call to
 898  * tcp/dccp_create_openreq_child().
 899  */
 900 void inet_csk_prepare_forced_close(struct sock *sk)
 901         __releases(&sk->sk_lock.slock)
 902 {
 903         /* sk_clone_lock locked the socket and set refcnt to 2 */
 904         bh_unlock_sock(sk);
 905         sock_put(sk);
 906         inet_csk_prepare_for_destroy_sock(sk);
 907         inet_sk(sk)->inet_num = 0;
 908 }
 909 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 910
 911 int inet_csk_listen_start(struct sock *sk, int backlog)
 912 {
 913         struct inet_connection_sock *icsk = inet_csk(sk);
 914         struct inet_sock *inet = inet_sk(sk);
 915         int err = -EADDRINUSE;
 916
 917         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 918
 919         sk->sk_ack_backlog = 0;
 920         inet_csk_delack_init(sk);
 921
 922         /* There is race window here: we announce ourselves listening,
 923          * but this transition is still not validated by get_port().
 924          * It is OK, because this socket enters to hash table only
 925          * after validation is complete.
 926          */
 927         inet_sk_state_store(sk, TCP_LISTEN);
 928         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 929                 inet->inet_sport = htons(inet->inet_num);
 930
 931                 sk_dst_reset(sk);
 932                 err = sk->sk_prot->hash(sk);
 933
 934                 if (likely(!err))
 935                         return 0;
 936         }
 937
 938         inet_sk_set_state(sk, TCP_CLOSE);
 939         return err;
 940 }
 941 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 942
 943 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 944                               struct sock *child)
 945 {
 946         sk->sk_prot->disconnect(child, O_NONBLOCK);
 947
 948         sock_orphan(child);
 949
 950         percpu_counter_inc(sk->sk_prot->orphan_count);
 951
 952         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 953                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
 954                 BUG_ON(sk != req->rsk_listener);
 955
 956                 /* Paranoid, to prevent race condition if
 957                  * an inbound pkt destined for child is
 958                  * blocked by sock lock in tcp_v4_rcv().
 959                  * Also to satisfy an assertion in
 960                  * tcp_v4_destroy_sock().
 961                  */
 962                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
 963         }
 964         inet_csk_destroy_sock(child);
 965 }
 966
 967 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 968                                       struct request_sock *req,
 969                                       struct sock *child)
 970 {
 971         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 972
 973         spin_lock(&queue->rskq_lock);
 974         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 975                 inet_child_forget(sk, req, child);
 976                 child = NULL;
 977         } else {
 978                 req->sk = child;
 979                 req->dl_next = NULL;
 980                 if (queue->rskq_accept_head == NULL)
 981                         WRITE_ONCE(queue->rskq_accept_head, req);
 982                 else
 983                         queue->rskq_accept_tail->dl_next = req;
 984                 queue->rskq_accept_tail = req;
 985                 sk_acceptq_added(sk);
 986         }
 987         spin_unlock(&queue->rskq_lock);
 988         return child;
 989 }
 990 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 991
 992 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 993                                          struct request_sock *req, bool own_req)
 994 {
 995         if (own_req) {
 996                 inet_csk_reqsk_queue_drop(sk, req);
 997                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 998                 if (inet_csk_reqsk_queue_add(sk, req, child))
 999                         return child;
1000         }
1001         /* Too bad, another child took ownership of the request, undo. */
1002         bh_unlock_sock(child);
1003         sock_put(child);
1004         return NULL;
1005 }
1006 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1007
1008 /*
1009  *      This routine closes sockets which have been at least partially
1010  *      opened, but not yet accepted.
1011  */
1012 void inet_csk_listen_stop(struct sock *sk)
1013 {
1014         struct inet_connection_sock *icsk = inet_csk(sk);
1015         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1016         struct request_sock *next, *req;
1017
1018         /* Following specs, it would be better either to send FIN
1019          * (and enter FIN-WAIT-1, it is normal close)
1020          * or to send active reset (abort).
1021          * Certainly, it is pretty dangerous while synflood, but it is
1022          * bad justification for our negligence 8)
1023          * To be honest, we are not able to make either
1024          * of the variants now.                 --ANK
1025          */
1026         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1027                 struct sock *child = req->sk;
1028
1029                 local_bh_disable();
1030                 bh_lock_sock(child);
1031                 WARN_ON(sock_owned_by_user(child));
1032                 sock_hold(child);
1033
1034                 inet_child_forget(sk, req, child);
1035                 reqsk_put(req);
1036                 bh_unlock_sock(child);
1037                 local_bh_enable();
1038                 sock_put(child);
1039
1040                 cond_resched();
1041         }
1042         if (queue->fastopenq.rskq_rst_head) {
1043                 /* Free all the reqs queued in rskq_rst_head. */
1044                 spin_lock_bh(&queue->fastopenq.lock);
1045                 req = queue->fastopenq.rskq_rst_head;
1046                 queue->fastopenq.rskq_rst_head = NULL;
1047                 spin_unlock_bh(&queue->fastopenq.lock);
1048                 while (req != NULL) {
1049                         next = req->dl_next;
1050                         reqsk_put(req);
1051                         req = next;
1052                 }
1053         }
1054         WARN_ON_ONCE(sk->sk_ack_backlog);
1055 }
1056 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1057
1058 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1059 {
1060         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1061         const struct inet_sock *inet = inet_sk(sk);
1062
1063         sin->sin_family         = AF_INET;
1064         sin->sin_addr.s_addr    = inet->inet_daddr;
1065         sin->sin_port           = inet->inet_dport;
1066 }
1067 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1068
1069 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1070 {
1071         const struct inet_sock *inet = inet_sk(sk);
1072         const struct ip_options_rcu *inet_opt;
1073         __be32 daddr = inet->inet_daddr;
1074         struct flowi4 *fl4;
1075         struct rtable *rt;
1076
1077         rcu_read_lock();
1078         inet_opt = rcu_dereference(inet->inet_opt);
1079         if (inet_opt && inet_opt->opt.srr)
1080                 daddr = inet_opt->opt.faddr;
1081         fl4 = &fl->u.ip4;
1082         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1083                                    inet->inet_saddr, inet->inet_dport,
1084                                    inet->inet_sport, sk->sk_protocol,
1085                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1086         if (IS_ERR(rt))
1087                 rt = NULL;
1088         if (rt)
1089                 sk_setup_caps(sk, &rt->dst);
1090         rcu_read_unlock();
1091
1092         return &rt->dst;
1093 }
1094
1095 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1096 {
1097         struct dst_entry *dst = __sk_dst_check(sk, 0);
1098         struct inet_sock *inet = inet_sk(sk);
1099
1100         if (!dst) {
1101                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1102                 if (!dst)
1103                         goto out;
1104         }
1105         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1106
1107         dst = __sk_dst_check(sk, 0);
1108         if (!dst)
1109                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1110 out:
1111         return dst;
1112 }
1113 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);