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 /* Obtain a reference to a local port for the given sock,
 300  * if snum is zero it means select any available local port.
 301  * We try to allocate an odd port (and leave even ports for connect())
 302  */
 303 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 304 {
 305         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 306         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 307         int ret = 1, port = snum;
 308         struct inet_bind_hashbucket *head;
 309         struct net *net = sock_net(sk);
 310         struct inet_bind_bucket *tb = NULL;
 311         kuid_t uid = sock_i_uid(sk);
 312         int l3mdev;
 313
 314         l3mdev = inet_sk_bound_l3mdev(sk);
 315
 316         if (!port) {
 317                 head = inet_csk_find_open_port(sk, &tb, &port);
 318                 if (!head)
 319                         return ret;
 320                 if (!tb)
 321                         goto tb_not_found;
 322                 goto success;
 323         }
 324         head = &hinfo->bhash[inet_bhashfn(net, port,
 325                                           hinfo->bhash_size)];
 326         spin_lock_bh(&head->lock);
 327         inet_bind_bucket_for_each(tb, &head->chain)
 328                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 329                     tb->port == port)
 330                         goto tb_found;
 331 tb_not_found:
 332         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 333                                      net, head, port, l3mdev);
 334         if (!tb)
 335                 goto fail_unlock;
 336 tb_found:
 337         if (!hlist_empty(&tb->owners)) {
 338                 if (sk->sk_reuse == SK_FORCE_REUSE)
 339                         goto success;
 340
 341                 if ((tb->fastreuse > 0 && reuse) ||
 342                     sk_reuseport_match(tb, sk))
 343                         goto success;
 344                 if (inet_csk_bind_conflict(sk, tb, true, true))
 345                         goto fail_unlock;
 346         }
 347 success:
 348         if (hlist_empty(&tb->owners)) {
 349                 tb->fastreuse = reuse;
 350                 if (sk->sk_reuseport) {
 351                         tb->fastreuseport = FASTREUSEPORT_ANY;
 352                         tb->fastuid = uid;
 353                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 354                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 355                         tb->fast_sk_family = sk->sk_family;
 356 #if IS_ENABLED(CONFIG_IPV6)
 357                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 358 #endif
 359                 } else {
 360                         tb->fastreuseport = 0;
 361                 }
 362         } else {
 363                 if (!reuse)
 364                         tb->fastreuse = 0;
 365                 if (sk->sk_reuseport) {
 366                         /* We didn't match or we don't have fastreuseport set on
 367                          * the tb, but we have sk_reuseport set on this socket
 368                          * and we know that there are no bind conflicts with
 369                          * this socket in this tb, so reset our tb's reuseport
 370                          * settings so that any subsequent sockets that match
 371                          * our current socket will be put on the fast path.
 372                          *
 373                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 374                          * do extra checking for all subsequent sk_reuseport
 375                          * socks.
 376                          */
 377                         if (!sk_reuseport_match(tb, sk)) {
 378                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 379                                 tb->fastuid = uid;
 380                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 381                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 382                                 tb->fast_sk_family = sk->sk_family;
 383 #if IS_ENABLED(CONFIG_IPV6)
 384                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 385 #endif
 386                         }
 387                 } else {
 388                         tb->fastreuseport = 0;
 389                 }
 390         }
 391         if (!inet_csk(sk)->icsk_bind_hash)
 392                 inet_bind_hash(sk, tb, port);
 393         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 394         ret = 0;
 395
 396 fail_unlock:
 397         spin_unlock_bh(&head->lock);
 398         return ret;
 399 }
 400 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 401
 402 /*
 403  * Wait for an incoming connection, avoid race conditions. This must be called
 404  * with the socket locked.
 405  */
 406 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 407 {
 408         struct inet_connection_sock *icsk = inet_csk(sk);
 409         DEFINE_WAIT(wait);
 410         int err;
 411
 412         /*
 413          * True wake-one mechanism for incoming connections: only
 414          * one process gets woken up, not the 'whole herd'.
 415          * Since we do not 'race & poll' for established sockets
 416          * anymore, the common case will execute the loop only once.
 417          *
 418          * Subtle issue: "add_wait_queue_exclusive()" will be added
 419          * after any current non-exclusive waiters, and we know that
 420          * it will always _stay_ after any new non-exclusive waiters
 421          * because all non-exclusive waiters are added at the
 422          * beginning of the wait-queue. As such, it's ok to "drop"
 423          * our exclusiveness temporarily when we get woken up without
 424          * having to remove and re-insert us on the wait queue.
 425          */
 426         for (;;) {
 427                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 428                                           TASK_INTERRUPTIBLE);
 429                 release_sock(sk);
 430                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 431                         timeo = schedule_timeout(timeo);
 432                 sched_annotate_sleep();
 433                 lock_sock(sk);
 434                 err = 0;
 435                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 436                         break;
 437                 err = -EINVAL;
 438                 if (sk->sk_state != TCP_LISTEN)
 439                         break;
 440                 err = sock_intr_errno(timeo);
 441                 if (signal_pending(current))
 442                         break;
 443                 err = -EAGAIN;
 444                 if (!timeo)
 445                         break;
 446         }
 447         finish_wait(sk_sleep(sk), &wait);
 448         return err;
 449 }
 450
 451 /*
 452  * This will accept the next outstanding connection.
 453  */
 454 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 455 {
 456         struct inet_connection_sock *icsk = inet_csk(sk);
 457         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 458         struct request_sock *req;
 459         struct sock *newsk;
 460         int error;
 461
 462         lock_sock(sk);
 463
 464         /* We need to make sure that this socket is listening,
 465          * and that it has something pending.
 466          */
 467         error = -EINVAL;
 468         if (sk->sk_state != TCP_LISTEN)
 469                 goto out_err;
 470
 471         /* Find already established connection */
 472         if (reqsk_queue_empty(queue)) {
 473                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 474
 475                 /* If this is a non blocking socket don't sleep */
 476                 error = -EAGAIN;
 477                 if (!timeo)
 478                         goto out_err;
 479
 480                 error = inet_csk_wait_for_connect(sk, timeo);
 481                 if (error)
 482                         goto out_err;
 483         }
 484         req = reqsk_queue_remove(queue, sk);
 485         newsk = req->sk;
 486
 487         if (sk->sk_protocol == IPPROTO_TCP &&
 488             tcp_rsk(req)->tfo_listener) {
 489                 spin_lock_bh(&queue->fastopenq.lock);
 490                 if (tcp_rsk(req)->tfo_listener) {
 491                         /* We are still waiting for the final ACK from 3WHS
 492                          * so can't free req now. Instead, we set req->sk to
 493                          * NULL to signify that the child socket is taken
 494                          * so reqsk_fastopen_remove() will free the req
 495                          * when 3WHS finishes (or is aborted).
 496                          */
 497                         req->sk = NULL;
 498                         req = NULL;
 499                 }
 500                 spin_unlock_bh(&queue->fastopenq.lock);
 501         }
 502
 503 out:
 504         release_sock(sk);
 505         if (newsk && mem_cgroup_sockets_enabled) {
 506                 int amt;
 507
 508                 /* atomically get the memory usage, set and charge the
 509                  * newsk->sk_memcg.
 510                  */
 511                 lock_sock(newsk);
 512
 513                 /* The socket has not been accepted yet, no need to look at
 514                  * newsk->sk_wmem_queued.
 515                  */
 516                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 517                                    atomic_read(&newsk->sk_rmem_alloc));
 518                 mem_cgroup_sk_alloc(newsk);
 519                 if (newsk->sk_memcg && amt)
 520                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 521
 522                 release_sock(newsk);
 523         }
 524         if (req)
 525                 reqsk_put(req);
 526         return newsk;
 527 out_err:
 528         newsk = NULL;
 529         req = NULL;
 530         *err = error;
 531         goto out;
 532 }
 533 EXPORT_SYMBOL(inet_csk_accept);
 534
 535 /*
 536  * Using different timers for retransmit, delayed acks and probes
 537  * We may wish use just one timer maintaining a list of expire jiffies
 538  * to optimize.
 539  */
 540 void inet_csk_init_xmit_timers(struct sock *sk,
 541                                void (*retransmit_handler)(struct timer_list *t),
 542                                void (*delack_handler)(struct timer_list *t),
 543                                void (*keepalive_handler)(struct timer_list *t))
 544 {
 545         struct inet_connection_sock *icsk = inet_csk(sk);
 546
 547         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 548         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 549         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 550         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 551 }
 552 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 553
 554 void inet_csk_clear_xmit_timers(struct sock *sk)
 555 {
 556         struct inet_connection_sock *icsk = inet_csk(sk);
 557
 558         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 559
 560         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 561         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 562         sk_stop_timer(sk, &sk->sk_timer);
 563 }
 564 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 565
 566 void inet_csk_delete_keepalive_timer(struct sock *sk)
 567 {
 568         sk_stop_timer(sk, &sk->sk_timer);
 569 }
 570 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 571
 572 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 573 {
 574         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 575 }
 576 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 577
 578 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 579                                      struct flowi4 *fl4,
 580                                      const struct request_sock *req)
 581 {
 582         const struct inet_request_sock *ireq = inet_rsk(req);
 583         struct net *net = read_pnet(&ireq->ireq_net);
 584         struct ip_options_rcu *opt;
 585         struct rtable *rt;
 586
 587         rcu_read_lock();
 588         opt = rcu_dereference(ireq->ireq_opt);
 589
 590         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 591                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 592                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 593                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 594                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 595                            htons(ireq->ir_num), sk->sk_uid);
 596         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 597         rt = ip_route_output_flow(net, fl4, sk);
 598         if (IS_ERR(rt))
 599                 goto no_route;
 600         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 601                 goto route_err;
 602         rcu_read_unlock();
 603         return &rt->dst;
 604
 605 route_err:
 606         ip_rt_put(rt);
 607 no_route:
 608         rcu_read_unlock();
 609         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 610         return NULL;
 611 }
 612 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 613
 614 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 615                                             struct sock *newsk,
 616                                             const struct request_sock *req)
 617 {
 618         const struct inet_request_sock *ireq = inet_rsk(req);
 619         struct net *net = read_pnet(&ireq->ireq_net);
 620         struct inet_sock *newinet = inet_sk(newsk);
 621         struct ip_options_rcu *opt;
 622         struct flowi4 *fl4;
 623         struct rtable *rt;
 624
 625         opt = rcu_dereference(ireq->ireq_opt);
 626         fl4 = &newinet->cork.fl.u.ip4;
 627
 628         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 629                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 630                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 631                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 632                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 633                            htons(ireq->ir_num), sk->sk_uid);
 634         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 635         rt = ip_route_output_flow(net, fl4, sk);
 636         if (IS_ERR(rt))
 637                 goto no_route;
 638         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 639                 goto route_err;
 640         return &rt->dst;
 641
 642 route_err:
 643         ip_rt_put(rt);
 644 no_route:
 645         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 646         return NULL;
 647 }
 648 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 649
 650 /* Decide when to expire the request and when to resend SYN-ACK */
 651 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 652                                   const int max_retries,
 653                                   const u8 rskq_defer_accept,
 654                                   int *expire, int *resend)
 655 {
 656         if (!rskq_defer_accept) {
 657                 *expire = req->num_timeout >= thresh;
 658                 *resend = 1;
 659                 return;
 660         }
 661         *expire = req->num_timeout >= thresh &&
 662                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 663         /*
 664          * Do not resend while waiting for data after ACK,
 665          * start to resend on end of deferring period to give
 666          * last chance for data or ACK to create established socket.
 667          */
 668         *resend = !inet_rsk(req)->acked ||
 669                   req->num_timeout >= rskq_defer_accept - 1;
 670 }
 671
 672 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 673 {
 674         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 675
 676         if (!err)
 677                 req->num_retrans++;
 678         return err;
 679 }
 680 EXPORT_SYMBOL(inet_rtx_syn_ack);
 681
 682 /* return true if req was found in the ehash table */
 683 static bool reqsk_queue_unlink(struct request_sock *req)
 684 {
 685         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 686         bool found = false;
 687
 688         if (sk_hashed(req_to_sk(req))) {
 689                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 690
 691                 spin_lock(lock);
 692                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 693                 spin_unlock(lock);
 694         }
 695         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 696                 reqsk_put(req);
 697         return found;
 698 }
 699
 700 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 701 {
 702         if (reqsk_queue_unlink(req)) {
 703                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 704                 reqsk_put(req);
 705         }
 706 }
 707 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 708
 709 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 710 {
 711         inet_csk_reqsk_queue_drop(sk, req);
 712         reqsk_put(req);
 713 }
 714 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 715
 716 static void reqsk_timer_handler(struct timer_list *t)
 717 {
 718         struct request_sock *req = from_timer(req, t, rsk_timer);
 719         struct sock *sk_listener = req->rsk_listener;
 720         struct net *net = sock_net(sk_listener);
 721         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 722         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 723         int qlen, expire = 0, resend = 0;
 724         int max_retries, thresh;
 725         u8 defer_accept;
 726
 727         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 728                 goto drop;
 729
 730         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 731         thresh = max_retries;
 732         /* Normally all the openreqs are young and become mature
 733          * (i.e. converted to established socket) for first timeout.
 734          * If synack was not acknowledged for 1 second, it means
 735          * one of the following things: synack was lost, ack was lost,
 736          * rtt is high or nobody planned to ack (i.e. synflood).
 737          * When server is a bit loaded, queue is populated with old
 738          * open requests, reducing effective size of queue.
 739          * When server is well loaded, queue size reduces to zero
 740          * after several minutes of work. It is not synflood,
 741          * it is normal operation. The solution is pruning
 742          * too old entries overriding normal timeout, when
 743          * situation becomes dangerous.
 744          *
 745          * Essentially, we reserve half of room for young
 746          * embrions; and abort old ones without pity, if old
 747          * ones are about to clog our table.
 748          */
 749         qlen = reqsk_queue_len(queue);
 750         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
 751                 int young = reqsk_queue_len_young(queue) << 1;
 752
 753                 while (thresh > 2) {
 754                         if (qlen < young)
 755                                 break;
 756                         thresh--;
 757                         young <<= 1;
 758                 }
 759         }
 760         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 761         if (defer_accept)
 762                 max_retries = defer_accept;
 763         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 764                        &expire, &resend);
 765         req->rsk_ops->syn_ack_timeout(req);
 766         if (!expire &&
 767             (!resend ||
 768              !inet_rtx_syn_ack(sk_listener, req) ||
 769              inet_rsk(req)->acked)) {
 770                 unsigned long timeo;
 771
 772                 if (req->num_timeout++ == 0)
 773                         atomic_dec(&queue->young);
 774                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 775                 mod_timer(&req->rsk_timer, jiffies + timeo);
 776                 return;
 777         }
 778 drop:
 779         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 780 }
 781
 782 static void reqsk_queue_hash_req(struct request_sock *req,
 783                                  unsigned long timeout)
 784 {
 785         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 786         mod_timer(&req->rsk_timer, jiffies + timeout);
 787
 788         inet_ehash_insert(req_to_sk(req), NULL);
 789         /* before letting lookups find us, make sure all req fields
 790          * are committed to memory and refcnt initialized.
 791          */
 792         smp_wmb();
 793         refcount_set(&req->rsk_refcnt, 2 + 1);
 794 }
 795
 796 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 797                                    unsigned long timeout)
 798 {
 799         reqsk_queue_hash_req(req, timeout);
 800         inet_csk_reqsk_queue_added(sk);
 801 }
 802 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 803
 804 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
 805                            const gfp_t priority)
 806 {
 807         struct inet_connection_sock *icsk = inet_csk(newsk);
 808
 809         if (!icsk->icsk_ulp_ops)
 810                 return;
 811
 812         if (icsk->icsk_ulp_ops->clone)
 813                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
 814 }
 815
 816 /**
 817  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 818  *      @sk: the socket to clone
 819  *      @req: request_sock
 820  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 821  *
 822  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 823  */
 824 struct sock *inet_csk_clone_lock(const struct sock *sk,
 825                                  const struct request_sock *req,
 826                                  const gfp_t priority)
 827 {
 828         struct sock *newsk = sk_clone_lock(sk, priority);
 829
 830         if (newsk) {
 831                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 832
 833                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 834                 newicsk->icsk_bind_hash = NULL;
 835
 836                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 837                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 838                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 839
 840                 /* listeners have SOCK_RCU_FREE, not the children */
 841                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 842
 843                 inet_sk(newsk)->mc_list = NULL;
 844
 845                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 846                 atomic64_set(&newsk->sk_cookie,
 847                              atomic64_read(&inet_rsk(req)->ir_cookie));
 848
 849                 newicsk->icsk_retransmits = 0;
 850                 newicsk->icsk_backoff     = 0;
 851                 newicsk->icsk_probes_out  = 0;
 852
 853                 /* Deinitialize accept_queue to trap illegal accesses. */
 854                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 855
 856                 inet_clone_ulp(req, newsk, priority);
 857
 858                 security_inet_csk_clone(newsk, req);
 859         }
 860         return newsk;
 861 }
 862 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 863
 864 /*
 865  * At this point, there should be no process reference to this
 866  * socket, and thus no user references at all.  Therefore we
 867  * can assume the socket waitqueue is inactive and nobody will
 868  * try to jump onto it.
 869  */
 870 void inet_csk_destroy_sock(struct sock *sk)
 871 {
 872         WARN_ON(sk->sk_state != TCP_CLOSE);
 873         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 874
 875         /* It cannot be in hash table! */
 876         WARN_ON(!sk_unhashed(sk));
 877
 878         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 879         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 880
 881         sk->sk_prot->destroy(sk);
 882
 883         sk_stream_kill_queues(sk);
 884
 885         xfrm_sk_free_policy(sk);
 886
 887         sk_refcnt_debug_release(sk);
 888
 889         percpu_counter_dec(sk->sk_prot->orphan_count);
 890
 891         sock_put(sk);
 892 }
 893 EXPORT_SYMBOL(inet_csk_destroy_sock);
 894
 895 /* This function allows to force a closure of a socket after the call to
 896  * tcp/dccp_create_openreq_child().
 897  */
 898 void inet_csk_prepare_forced_close(struct sock *sk)
 899         __releases(&sk->sk_lock.slock)
 900 {
 901         /* sk_clone_lock locked the socket and set refcnt to 2 */
 902         bh_unlock_sock(sk);
 903         sock_put(sk);
 904         inet_csk_prepare_for_destroy_sock(sk);
 905         inet_sk(sk)->inet_num = 0;
 906 }
 907 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 908
 909 int inet_csk_listen_start(struct sock *sk, int backlog)
 910 {
 911         struct inet_connection_sock *icsk = inet_csk(sk);
 912         struct inet_sock *inet = inet_sk(sk);
 913         int err = -EADDRINUSE;
 914
 915         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 916
 917         sk->sk_ack_backlog = 0;
 918         inet_csk_delack_init(sk);
 919
 920         /* There is race window here: we announce ourselves listening,
 921          * but this transition is still not validated by get_port().
 922          * It is OK, because this socket enters to hash table only
 923          * after validation is complete.
 924          */
 925         inet_sk_state_store(sk, TCP_LISTEN);
 926         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 927                 inet->inet_sport = htons(inet->inet_num);
 928
 929                 sk_dst_reset(sk);
 930                 err = sk->sk_prot->hash(sk);
 931
 932                 if (likely(!err))
 933                         return 0;
 934         }
 935
 936         inet_sk_set_state(sk, TCP_CLOSE);
 937         return err;
 938 }
 939 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 940
 941 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 942                               struct sock *child)
 943 {
 944         sk->sk_prot->disconnect(child, O_NONBLOCK);
 945
 946         sock_orphan(child);
 947
 948         percpu_counter_inc(sk->sk_prot->orphan_count);
 949
 950         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 951                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
 952                 BUG_ON(sk != req->rsk_listener);
 953
 954                 /* Paranoid, to prevent race condition if
 955                  * an inbound pkt destined for child is
 956                  * blocked by sock lock in tcp_v4_rcv().
 957                  * Also to satisfy an assertion in
 958                  * tcp_v4_destroy_sock().
 959                  */
 960                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
 961         }
 962         inet_csk_destroy_sock(child);
 963 }
 964
 965 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 966                                       struct request_sock *req,
 967                                       struct sock *child)
 968 {
 969         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 970
 971         spin_lock(&queue->rskq_lock);
 972         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 973                 inet_child_forget(sk, req, child);
 974                 child = NULL;
 975         } else {
 976                 req->sk = child;
 977                 req->dl_next = NULL;
 978                 if (queue->rskq_accept_head == NULL)
 979                         WRITE_ONCE(queue->rskq_accept_head, req);
 980                 else
 981                         queue->rskq_accept_tail->dl_next = req;
 982                 queue->rskq_accept_tail = req;
 983                 sk_acceptq_added(sk);
 984         }
 985         spin_unlock(&queue->rskq_lock);
 986         return child;
 987 }
 988 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 989
 990 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 991                                          struct request_sock *req, bool own_req)
 992 {
 993         if (own_req) {
 994                 inet_csk_reqsk_queue_drop(sk, req);
 995                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 996                 if (inet_csk_reqsk_queue_add(sk, req, child))
 997                         return child;
 998         }
 999         /* Too bad, another child took ownership of the request, undo. */
1000         bh_unlock_sock(child);
1001         sock_put(child);
1002         return NULL;
1003 }
1004 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1005
1006 /*
1007  *      This routine closes sockets which have been at least partially
1008  *      opened, but not yet accepted.
1009  */
1010 void inet_csk_listen_stop(struct sock *sk)
1011 {
1012         struct inet_connection_sock *icsk = inet_csk(sk);
1013         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1014         struct request_sock *next, *req;
1015
1016         /* Following specs, it would be better either to send FIN
1017          * (and enter FIN-WAIT-1, it is normal close)
1018          * or to send active reset (abort).
1019          * Certainly, it is pretty dangerous while synflood, but it is
1020          * bad justification for our negligence 8)
1021          * To be honest, we are not able to make either
1022          * of the variants now.                 --ANK
1023          */
1024         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1025                 struct sock *child = req->sk;
1026
1027                 local_bh_disable();
1028                 bh_lock_sock(child);
1029                 WARN_ON(sock_owned_by_user(child));
1030                 sock_hold(child);
1031
1032                 inet_child_forget(sk, req, child);
1033                 reqsk_put(req);
1034                 bh_unlock_sock(child);
1035                 local_bh_enable();
1036                 sock_put(child);
1037
1038                 cond_resched();
1039         }
1040         if (queue->fastopenq.rskq_rst_head) {
1041                 /* Free all the reqs queued in rskq_rst_head. */
1042                 spin_lock_bh(&queue->fastopenq.lock);
1043                 req = queue->fastopenq.rskq_rst_head;
1044                 queue->fastopenq.rskq_rst_head = NULL;
1045                 spin_unlock_bh(&queue->fastopenq.lock);
1046                 while (req != NULL) {
1047                         next = req->dl_next;
1048                         reqsk_put(req);
1049                         req = next;
1050                 }
1051         }
1052         WARN_ON_ONCE(sk->sk_ack_backlog);
1053 }
1054 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1055
1056 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1057 {
1058         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1059         const struct inet_sock *inet = inet_sk(sk);
1060
1061         sin->sin_family         = AF_INET;
1062         sin->sin_addr.s_addr    = inet->inet_daddr;
1063         sin->sin_port           = inet->inet_dport;
1064 }
1065 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1066
1067 #ifdef CONFIG_COMPAT
1068 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
1069                                char __user *optval, int __user *optlen)
1070 {
1071         const struct inet_connection_sock *icsk = inet_csk(sk);
1072
1073         if (icsk->icsk_af_ops->compat_getsockopt)
1074                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
1075                                                             optval, optlen);
1076         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
1077                                              optval, optlen);
1078 }
1079 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
1080
1081 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
1082                                char __user *optval, unsigned int optlen)
1083 {
1084         const struct inet_connection_sock *icsk = inet_csk(sk);
1085
1086         if (icsk->icsk_af_ops->compat_setsockopt)
1087                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
1088                                                             optval, optlen);
1089         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1090                                              optval, optlen);
1091 }
1092 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
1093 #endif
1094
1095 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1096 {
1097         const struct inet_sock *inet = inet_sk(sk);
1098         const struct ip_options_rcu *inet_opt;
1099         __be32 daddr = inet->inet_daddr;
1100         struct flowi4 *fl4;
1101         struct rtable *rt;
1102
1103         rcu_read_lock();
1104         inet_opt = rcu_dereference(inet->inet_opt);
1105         if (inet_opt && inet_opt->opt.srr)
1106                 daddr = inet_opt->opt.faddr;
1107         fl4 = &fl->u.ip4;
1108         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1109                                    inet->inet_saddr, inet->inet_dport,
1110                                    inet->inet_sport, sk->sk_protocol,
1111                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1112         if (IS_ERR(rt))
1113                 rt = NULL;
1114         if (rt)
1115                 sk_setup_caps(sk, &rt->dst);
1116         rcu_read_unlock();
1117
1118         return &rt->dst;
1119 }
1120
1121 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1122 {
1123         struct dst_entry *dst = __sk_dst_check(sk, 0);
1124         struct inet_sock *inet = inet_sk(sk);
1125
1126         if (!dst) {
1127                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1128                 if (!dst)
1129                         goto out;
1130         }
1131         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1132
1133         dst = __sk_dst_check(sk, 0);
1134         if (!dst)
1135                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1136 out:
1137         return dst;
1138 }
1139 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);