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                 int bound_dev_if2;
 151
 152                 if (sk == sk2)
 153                         continue;
 154                 bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
 155                 if ((!sk->sk_bound_dev_if ||
 156                      !bound_dev_if2 ||
 157                      sk->sk_bound_dev_if == bound_dev_if2)) {
 158                         if (reuse && sk2->sk_reuse &&
 159                             sk2->sk_state != TCP_LISTEN) {
 160                                 if ((!relax ||
 161                                      (!reuseport_ok &&
 162                                       reuseport && sk2->sk_reuseport &&
 163                                       !rcu_access_pointer(sk->sk_reuseport_cb) &&
 164                                       (sk2->sk_state == TCP_TIME_WAIT ||
 165                                        uid_eq(uid, sock_i_uid(sk2))))) &&
 166                                     inet_rcv_saddr_equal(sk, sk2, true))
 167                                         break;
 168                         } else if (!reuseport_ok ||
 169                                    !reuseport || !sk2->sk_reuseport ||
 170                                    rcu_access_pointer(sk->sk_reuseport_cb) ||
 171                                    (sk2->sk_state != TCP_TIME_WAIT &&
 172                                     !uid_eq(uid, sock_i_uid(sk2)))) {
 173                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 174                                         break;
 175                         }
 176                 }
 177         }
 178         return sk2 != NULL;
 179 }
 180
 181 /*
 182  * Find an open port number for the socket.  Returns with the
 183  * inet_bind_hashbucket lock held.
 184  */
 185 static struct inet_bind_hashbucket *
 186 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 187 {
 188         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 189         int port = 0;
 190         struct inet_bind_hashbucket *head;
 191         struct net *net = sock_net(sk);
 192         bool relax = false;
 193         int i, low, high, attempt_half;
 194         struct inet_bind_bucket *tb;
 195         u32 remaining, offset;
 196         int l3mdev;
 197
 198         l3mdev = inet_sk_bound_l3mdev(sk);
 199 ports_exhausted:
 200         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 201 other_half_scan:
 202         inet_get_local_port_range(net, &low, &high);
 203         high++; /* [32768, 60999] -> [32768, 61000[ */
 204         if (high - low < 4)
 205                 attempt_half = 0;
 206         if (attempt_half) {
 207                 int half = low + (((high - low) >> 2) << 1);
 208
 209                 if (attempt_half == 1)
 210                         high = half;
 211                 else
 212                         low = half;
 213         }
 214         remaining = high - low;
 215         if (likely(remaining > 1))
 216                 remaining &= ~1U;
 217
 218         offset = prandom_u32() % remaining;
 219         /* __inet_hash_connect() favors ports having @low parity
 220          * We do the opposite to not pollute connect() users.
 221          */
 222         offset |= 1U;
 223
 224 other_parity_scan:
 225         port = low + offset;
 226         for (i = 0; i < remaining; i += 2, port += 2) {
 227                 if (unlikely(port >= high))
 228                         port -= remaining;
 229                 if (inet_is_local_reserved_port(net, port))
 230                         continue;
 231                 head = &hinfo->bhash[inet_bhashfn(net, port,
 232                                                   hinfo->bhash_size)];
 233                 spin_lock_bh(&head->lock);
 234                 inet_bind_bucket_for_each(tb, &head->chain)
 235                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 236                             tb->port == port) {
 237                                 if (!inet_csk_bind_conflict(sk, tb, relax, false))
 238                                         goto success;
 239                                 goto next_port;
 240                         }
 241                 tb = NULL;
 242                 goto success;
 243 next_port:
 244                 spin_unlock_bh(&head->lock);
 245                 cond_resched();
 246         }
 247
 248         offset--;
 249         if (!(offset & 1))
 250                 goto other_parity_scan;
 251
 252         if (attempt_half == 1) {
 253                 /* OK we now try the upper half of the range */
 254                 attempt_half = 2;
 255                 goto other_half_scan;
 256         }
 257
 258         if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) {
 259                 /* We still have a chance to connect to different destinations */
 260                 relax = true;
 261                 goto ports_exhausted;
 262         }
 263         return NULL;
 264 success:
 265         *port_ret = port;
 266         *tb_ret = tb;
 267         return head;
 268 }
 269
 270 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 271                                      struct sock *sk)
 272 {
 273         kuid_t uid = sock_i_uid(sk);
 274
 275         if (tb->fastreuseport <= 0)
 276                 return 0;
 277         if (!sk->sk_reuseport)
 278                 return 0;
 279         if (rcu_access_pointer(sk->sk_reuseport_cb))
 280                 return 0;
 281         if (!uid_eq(tb->fastuid, uid))
 282                 return 0;
 283         /* We only need to check the rcv_saddr if this tb was once marked
 284          * without fastreuseport and then was reset, as we can only know that
 285          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 286          * owners list.
 287          */
 288         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 289                 return 1;
 290 #if IS_ENABLED(CONFIG_IPV6)
 291         if (tb->fast_sk_family == AF_INET6)
 292                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 293                                             inet6_rcv_saddr(sk),
 294                                             tb->fast_rcv_saddr,
 295                                             sk->sk_rcv_saddr,
 296                                             tb->fast_ipv6_only,
 297                                             ipv6_only_sock(sk), true, false);
 298 #endif
 299         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 300                                     ipv6_only_sock(sk), true, false);
 301 }
 302
 303 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 304                                struct sock *sk)
 305 {
 306         kuid_t uid = sock_i_uid(sk);
 307         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 308
 309         if (hlist_empty(&tb->owners)) {
 310                 tb->fastreuse = reuse;
 311                 if (sk->sk_reuseport) {
 312                         tb->fastreuseport = FASTREUSEPORT_ANY;
 313                         tb->fastuid = uid;
 314                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 315                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 316                         tb->fast_sk_family = sk->sk_family;
 317 #if IS_ENABLED(CONFIG_IPV6)
 318                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 319 #endif
 320                 } else {
 321                         tb->fastreuseport = 0;
 322                 }
 323         } else {
 324                 if (!reuse)
 325                         tb->fastreuse = 0;
 326                 if (sk->sk_reuseport) {
 327                         /* We didn't match or we don't have fastreuseport set on
 328                          * the tb, but we have sk_reuseport set on this socket
 329                          * and we know that there are no bind conflicts with
 330                          * this socket in this tb, so reset our tb's reuseport
 331                          * settings so that any subsequent sockets that match
 332                          * our current socket will be put on the fast path.
 333                          *
 334                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 335                          * do extra checking for all subsequent sk_reuseport
 336                          * socks.
 337                          */
 338                         if (!sk_reuseport_match(tb, sk)) {
 339                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 340                                 tb->fastuid = uid;
 341                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 342                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 343                                 tb->fast_sk_family = sk->sk_family;
 344 #if IS_ENABLED(CONFIG_IPV6)
 345                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 346 #endif
 347                         }
 348                 } else {
 349                         tb->fastreuseport = 0;
 350                 }
 351         }
 352 }
 353
 354 /* Obtain a reference to a local port for the given sock,
 355  * if snum is zero it means select any available local port.
 356  * We try to allocate an odd port (and leave even ports for connect())
 357  */
 358 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 359 {
 360         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 361         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 362         int ret = 1, port = snum;
 363         struct inet_bind_hashbucket *head;
 364         struct net *net = sock_net(sk);
 365         struct inet_bind_bucket *tb = NULL;
 366         int l3mdev;
 367
 368         l3mdev = inet_sk_bound_l3mdev(sk);
 369
 370         if (!port) {
 371                 head = inet_csk_find_open_port(sk, &tb, &port);
 372                 if (!head)
 373                         return ret;
 374                 if (!tb)
 375                         goto tb_not_found;
 376                 goto success;
 377         }
 378         head = &hinfo->bhash[inet_bhashfn(net, port,
 379                                           hinfo->bhash_size)];
 380         spin_lock_bh(&head->lock);
 381         inet_bind_bucket_for_each(tb, &head->chain)
 382                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 383                     tb->port == port)
 384                         goto tb_found;
 385 tb_not_found:
 386         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 387                                      net, head, port, l3mdev);
 388         if (!tb)
 389                 goto fail_unlock;
 390 tb_found:
 391         if (!hlist_empty(&tb->owners)) {
 392                 if (sk->sk_reuse == SK_FORCE_REUSE)
 393                         goto success;
 394
 395                 if ((tb->fastreuse > 0 && reuse) ||
 396                     sk_reuseport_match(tb, sk))
 397                         goto success;
 398                 if (inet_csk_bind_conflict(sk, tb, true, true))
 399                         goto fail_unlock;
 400         }
 401 success:
 402         inet_csk_update_fastreuse(tb, sk);
 403
 404         if (!inet_csk(sk)->icsk_bind_hash)
 405                 inet_bind_hash(sk, tb, port);
 406         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 407         ret = 0;
 408
 409 fail_unlock:
 410         spin_unlock_bh(&head->lock);
 411         return ret;
 412 }
 413 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 414
 415 /*
 416  * Wait for an incoming connection, avoid race conditions. This must be called
 417  * with the socket locked.
 418  */
 419 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 420 {
 421         struct inet_connection_sock *icsk = inet_csk(sk);
 422         DEFINE_WAIT(wait);
 423         int err;
 424
 425         /*
 426          * True wake-one mechanism for incoming connections: only
 427          * one process gets woken up, not the 'whole herd'.
 428          * Since we do not 'race & poll' for established sockets
 429          * anymore, the common case will execute the loop only once.
 430          *
 431          * Subtle issue: "add_wait_queue_exclusive()" will be added
 432          * after any current non-exclusive waiters, and we know that
 433          * it will always _stay_ after any new non-exclusive waiters
 434          * because all non-exclusive waiters are added at the
 435          * beginning of the wait-queue. As such, it's ok to "drop"
 436          * our exclusiveness temporarily when we get woken up without
 437          * having to remove and re-insert us on the wait queue.
 438          */
 439         for (;;) {
 440                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 441                                           TASK_INTERRUPTIBLE);
 442                 release_sock(sk);
 443                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 444                         timeo = schedule_timeout(timeo);
 445                 sched_annotate_sleep();
 446                 lock_sock(sk);
 447                 err = 0;
 448                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 449                         break;
 450                 err = -EINVAL;
 451                 if (sk->sk_state != TCP_LISTEN)
 452                         break;
 453                 err = sock_intr_errno(timeo);
 454                 if (signal_pending(current))
 455                         break;
 456                 err = -EAGAIN;
 457                 if (!timeo)
 458                         break;
 459         }
 460         finish_wait(sk_sleep(sk), &wait);
 461         return err;
 462 }
 463
 464 /*
 465  * This will accept the next outstanding connection.
 466  */
 467 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 468 {
 469         struct inet_connection_sock *icsk = inet_csk(sk);
 470         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 471         struct request_sock *req;
 472         struct sock *newsk;
 473         int error;
 474
 475         lock_sock(sk);
 476
 477         /* We need to make sure that this socket is listening,
 478          * and that it has something pending.
 479          */
 480         error = -EINVAL;
 481         if (sk->sk_state != TCP_LISTEN)
 482                 goto out_err;
 483
 484         /* Find already established connection */
 485         if (reqsk_queue_empty(queue)) {
 486                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 487
 488                 /* If this is a non blocking socket don't sleep */
 489                 error = -EAGAIN;
 490                 if (!timeo)
 491                         goto out_err;
 492
 493                 error = inet_csk_wait_for_connect(sk, timeo);
 494                 if (error)
 495                         goto out_err;
 496         }
 497         req = reqsk_queue_remove(queue, sk);
 498         newsk = req->sk;
 499
 500         if (sk->sk_protocol == IPPROTO_TCP &&
 501             tcp_rsk(req)->tfo_listener) {
 502                 spin_lock_bh(&queue->fastopenq.lock);
 503                 if (tcp_rsk(req)->tfo_listener) {
 504                         /* We are still waiting for the final ACK from 3WHS
 505                          * so can't free req now. Instead, we set req->sk to
 506                          * NULL to signify that the child socket is taken
 507                          * so reqsk_fastopen_remove() will free the req
 508                          * when 3WHS finishes (or is aborted).
 509                          */
 510                         req->sk = NULL;
 511                         req = NULL;
 512                 }
 513                 spin_unlock_bh(&queue->fastopenq.lock);
 514         }
 515
 516 out:
 517         release_sock(sk);
 518         if (newsk && mem_cgroup_sockets_enabled) {
 519                 int amt;
 520
 521                 /* atomically get the memory usage, set and charge the
 522                  * newsk->sk_memcg.
 523                  */
 524                 lock_sock(newsk);
 525
 526                 /* The socket has not been accepted yet, no need to look at
 527                  * newsk->sk_wmem_queued.
 528                  */
 529                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 530                                    atomic_read(&newsk->sk_rmem_alloc));
 531                 mem_cgroup_sk_alloc(newsk);
 532                 if (newsk->sk_memcg && amt)
 533                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 534
 535                 release_sock(newsk);
 536         }
 537         if (req)
 538                 reqsk_put(req);
 539
 540         if (newsk)
 541                 inet_init_csk_locks(newsk);
 542
 543         return newsk;
 544 out_err:
 545         newsk = NULL;
 546         req = NULL;
 547         *err = error;
 548         goto out;
 549 }
 550 EXPORT_SYMBOL(inet_csk_accept);
 551
 552 /*
 553  * Using different timers for retransmit, delayed acks and probes
 554  * We may wish use just one timer maintaining a list of expire jiffies
 555  * to optimize.
 556  */
 557 void inet_csk_init_xmit_timers(struct sock *sk,
 558                                void (*retransmit_handler)(struct timer_list *t),
 559                                void (*delack_handler)(struct timer_list *t),
 560                                void (*keepalive_handler)(struct timer_list *t))
 561 {
 562         struct inet_connection_sock *icsk = inet_csk(sk);
 563
 564         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 565         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 566         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 567         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 568 }
 569 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 570
 571 void inet_csk_clear_xmit_timers(struct sock *sk)
 572 {
 573         struct inet_connection_sock *icsk = inet_csk(sk);
 574
 575         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 576
 577         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 578         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 579         sk_stop_timer(sk, &sk->sk_timer);
 580 }
 581 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 582
 583 void inet_csk_clear_xmit_timers_sync(struct sock *sk)
 584 {
 585         struct inet_connection_sock *icsk = inet_csk(sk);
 586
 587         /* ongoing timer handlers need to acquire socket lock. */
 588         sock_not_owned_by_me(sk);
 589
 590         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 591
 592         sk_stop_timer_sync(sk, &icsk->icsk_retransmit_timer);
 593         sk_stop_timer_sync(sk, &icsk->icsk_delack_timer);
 594         sk_stop_timer_sync(sk, &sk->sk_timer);
 595 }
 596
 597 void inet_csk_delete_keepalive_timer(struct sock *sk)
 598 {
 599         sk_stop_timer(sk, &sk->sk_timer);
 600 }
 601 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 602
 603 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 604 {
 605         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 606 }
 607 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 608
 609 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 610                                      struct flowi4 *fl4,
 611                                      const struct request_sock *req)
 612 {
 613         const struct inet_request_sock *ireq = inet_rsk(req);
 614         struct net *net = read_pnet(&ireq->ireq_net);
 615         struct ip_options_rcu *opt;
 616         struct rtable *rt;
 617
 618         rcu_read_lock();
 619         opt = rcu_dereference(ireq->ireq_opt);
 620
 621         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 622                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 623                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 624                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 625                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 626                            htons(ireq->ir_num), sk->sk_uid);
 627         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 628         rt = ip_route_output_flow(net, fl4, sk);
 629         if (IS_ERR(rt))
 630                 goto no_route;
 631         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 632                 goto route_err;
 633         rcu_read_unlock();
 634         return &rt->dst;
 635
 636 route_err:
 637         ip_rt_put(rt);
 638 no_route:
 639         rcu_read_unlock();
 640         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 641         return NULL;
 642 }
 643 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 644
 645 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 646                                             struct sock *newsk,
 647                                             const struct request_sock *req)
 648 {
 649         const struct inet_request_sock *ireq = inet_rsk(req);
 650         struct net *net = read_pnet(&ireq->ireq_net);
 651         struct inet_sock *newinet = inet_sk(newsk);
 652         struct ip_options_rcu *opt;
 653         struct flowi4 *fl4;
 654         struct rtable *rt;
 655
 656         opt = rcu_dereference(ireq->ireq_opt);
 657         fl4 = &newinet->cork.fl.u.ip4;
 658
 659         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 660                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 661                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 662                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 663                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 664                            htons(ireq->ir_num), sk->sk_uid);
 665         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 666         rt = ip_route_output_flow(net, fl4, sk);
 667         if (IS_ERR(rt))
 668                 goto no_route;
 669         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 670                 goto route_err;
 671         return &rt->dst;
 672
 673 route_err:
 674         ip_rt_put(rt);
 675 no_route:
 676         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 677         return NULL;
 678 }
 679 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 680
 681 /* Decide when to expire the request and when to resend SYN-ACK */
 682 static void syn_ack_recalc(struct request_sock *req,
 683                            const int max_syn_ack_retries,
 684                            const u8 rskq_defer_accept,
 685                            int *expire, int *resend)
 686 {
 687         if (!rskq_defer_accept) {
 688                 *expire = req->num_timeout >= max_syn_ack_retries;
 689                 *resend = 1;
 690                 return;
 691         }
 692         *expire = req->num_timeout >= max_syn_ack_retries &&
 693                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 694         /* Do not resend while waiting for data after ACK,
 695          * start to resend on end of deferring period to give
 696          * last chance for data or ACK to create established socket.
 697          */
 698         *resend = !inet_rsk(req)->acked ||
 699                   req->num_timeout >= rskq_defer_accept - 1;
 700 }
 701
 702 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 703 {
 704         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 705
 706         if (!err)
 707                 req->num_retrans++;
 708         return err;
 709 }
 710 EXPORT_SYMBOL(inet_rtx_syn_ack);
 711
 712 /* return true if req was found in the ehash table */
 713 static bool reqsk_queue_unlink(struct request_sock *req)
 714 {
 715         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 716         bool found = false;
 717
 718         if (sk_hashed(req_to_sk(req))) {
 719                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 720
 721                 spin_lock(lock);
 722                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 723                 spin_unlock(lock);
 724         }
 725         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 726                 reqsk_put(req);
 727         return found;
 728 }
 729
 730 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 731 {
 732         bool unlinked = reqsk_queue_unlink(req);
 733
 734         if (unlinked) {
 735                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 736                 reqsk_put(req);
 737         }
 738         return unlinked;
 739 }
 740 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 741
 742 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 743 {
 744         inet_csk_reqsk_queue_drop(sk, req);
 745         reqsk_put(req);
 746 }
 747 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 748
 749 static void reqsk_timer_handler(struct timer_list *t)
 750 {
 751         struct request_sock *req = from_timer(req, t, rsk_timer);
 752         struct sock *sk_listener = req->rsk_listener;
 753         struct net *net = sock_net(sk_listener);
 754         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 755         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 756         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
 757
 758         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 759                 goto drop;
 760
 761         max_syn_ack_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
 762                 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries);
 763         /* Normally all the openreqs are young and become mature
 764          * (i.e. converted to established socket) for first timeout.
 765          * If synack was not acknowledged for 1 second, it means
 766          * one of the following things: synack was lost, ack was lost,
 767          * rtt is high or nobody planned to ack (i.e. synflood).
 768          * When server is a bit loaded, queue is populated with old
 769          * open requests, reducing effective size of queue.
 770          * When server is well loaded, queue size reduces to zero
 771          * after several minutes of work. It is not synflood,
 772          * it is normal operation. The solution is pruning
 773          * too old entries overriding normal timeout, when
 774          * situation becomes dangerous.
 775          *
 776          * Essentially, we reserve half of room for young
 777          * embrions; and abort old ones without pity, if old
 778          * ones are about to clog our table.
 779          */
 780         qlen = reqsk_queue_len(queue);
 781         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
 782                 int young = reqsk_queue_len_young(queue) << 1;
 783
 784                 while (max_syn_ack_retries > 2) {
 785                         if (qlen < young)
 786                                 break;
 787                         max_syn_ack_retries--;
 788                         young <<= 1;
 789                 }
 790         }
 791         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
 792                        &expire, &resend);
 793         req->rsk_ops->syn_ack_timeout(req);
 794         if (!expire &&
 795             (!resend ||
 796              !inet_rtx_syn_ack(sk_listener, req) ||
 797              inet_rsk(req)->acked)) {
 798                 unsigned long timeo;
 799
 800                 if (req->num_timeout++ == 0)
 801                         atomic_dec(&queue->young);
 802                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 803                 mod_timer(&req->rsk_timer, jiffies + timeo);
 804                 return;
 805         }
 806 drop:
 807         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 808 }
 809
 810 static void reqsk_queue_hash_req(struct request_sock *req,
 811                                  unsigned long timeout)
 812 {
 813         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 814         mod_timer(&req->rsk_timer, jiffies + timeout);
 815
 816         inet_ehash_insert(req_to_sk(req), NULL, NULL);
 817         /* before letting lookups find us, make sure all req fields
 818          * are committed to memory and refcnt initialized.
 819          */
 820         smp_wmb();
 821         refcount_set(&req->rsk_refcnt, 2 + 1);
 822 }
 823
 824 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 825                                    unsigned long timeout)
 826 {
 827         reqsk_queue_hash_req(req, timeout);
 828         inet_csk_reqsk_queue_added(sk);
 829 }
 830 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 831
 832 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
 833                            const gfp_t priority)
 834 {
 835         struct inet_connection_sock *icsk = inet_csk(newsk);
 836
 837         if (!icsk->icsk_ulp_ops)
 838                 return;
 839
 840         if (icsk->icsk_ulp_ops->clone)
 841                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
 842 }
 843
 844 /**
 845  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 846  *      @sk: the socket to clone
 847  *      @req: request_sock
 848  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 849  *
 850  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 851  */
 852 struct sock *inet_csk_clone_lock(const struct sock *sk,
 853                                  const struct request_sock *req,
 854                                  const gfp_t priority)
 855 {
 856         struct sock *newsk = sk_clone_lock(sk, priority);
 857
 858         if (newsk) {
 859                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 860
 861                 newsk->sk_wait_pending = 0;
 862                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 863                 newicsk->icsk_bind_hash = NULL;
 864
 865                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 866                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 867                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 868
 869                 /* listeners have SOCK_RCU_FREE, not the children */
 870                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 871
 872                 inet_sk(newsk)->mc_list = NULL;
 873
 874                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 875                 atomic64_set(&newsk->sk_cookie,
 876                              atomic64_read(&inet_rsk(req)->ir_cookie));
 877
 878                 newicsk->icsk_retransmits = 0;
 879                 newicsk->icsk_backoff     = 0;
 880                 newicsk->icsk_probes_out  = 0;
 881                 newicsk->icsk_probes_tstamp = 0;
 882
 883                 /* Deinitialize accept_queue to trap illegal accesses. */
 884                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 885
 886                 inet_clone_ulp(req, newsk, priority);
 887
 888                 security_inet_csk_clone(newsk, req);
 889         }
 890         return newsk;
 891 }
 892 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 893
 894 /*
 895  * At this point, there should be no process reference to this
 896  * socket, and thus no user references at all.  Therefore we
 897  * can assume the socket waitqueue is inactive and nobody will
 898  * try to jump onto it.
 899  */
 900 void inet_csk_destroy_sock(struct sock *sk)
 901 {
 902         WARN_ON(sk->sk_state != TCP_CLOSE);
 903         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 904
 905         /* It cannot be in hash table! */
 906         WARN_ON(!sk_unhashed(sk));
 907
 908         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 909         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 910
 911         sk->sk_prot->destroy(sk);
 912
 913         sk_stream_kill_queues(sk);
 914
 915         xfrm_sk_free_policy(sk);
 916
 917         sk_refcnt_debug_release(sk);
 918
 919         this_cpu_dec(*sk->sk_prot->orphan_count);
 920
 921         sock_put(sk);
 922 }
 923 EXPORT_SYMBOL(inet_csk_destroy_sock);
 924
 925 /* This function allows to force a closure of a socket after the call to
 926  * tcp/dccp_create_openreq_child().
 927  */
 928 void inet_csk_prepare_forced_close(struct sock *sk)
 929         __releases(&sk->sk_lock.slock)
 930 {
 931         /* sk_clone_lock locked the socket and set refcnt to 2 */
 932         bh_unlock_sock(sk);
 933         sock_put(sk);
 934         inet_csk_prepare_for_destroy_sock(sk);
 935         inet_sk(sk)->inet_num = 0;
 936 }
 937 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 938
 939 static int inet_ulp_can_listen(const struct sock *sk)
 940 {
 941         const struct inet_connection_sock *icsk = inet_csk(sk);
 942
 943         if (icsk->icsk_ulp_ops && !icsk->icsk_ulp_ops->clone)
 944                 return -EINVAL;
 945
 946         return 0;
 947 }
 948
 949 int inet_csk_listen_start(struct sock *sk, int backlog)
 950 {
 951         struct inet_connection_sock *icsk = inet_csk(sk);
 952         struct inet_sock *inet = inet_sk(sk);
 953         int err;
 954
 955         err = inet_ulp_can_listen(sk);
 956         if (unlikely(err))
 957                 return err;
 958
 959         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 960
 961         sk->sk_ack_backlog = 0;
 962         inet_csk_delack_init(sk);
 963
 964         /* There is race window here: we announce ourselves listening,
 965          * but this transition is still not validated by get_port().
 966          * It is OK, because this socket enters to hash table only
 967          * after validation is complete.
 968          */
 969         err = -EADDRINUSE;
 970         inet_sk_state_store(sk, TCP_LISTEN);
 971         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 972                 inet->inet_sport = htons(inet->inet_num);
 973
 974                 sk_dst_reset(sk);
 975                 err = sk->sk_prot->hash(sk);
 976
 977                 if (likely(!err))
 978                         return 0;
 979         }
 980
 981         inet_sk_set_state(sk, TCP_CLOSE);
 982         return err;
 983 }
 984 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 985
 986 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 987                               struct sock *child)
 988 {
 989         sk->sk_prot->disconnect(child, O_NONBLOCK);
 990
 991         sock_orphan(child);
 992
 993         this_cpu_inc(*sk->sk_prot->orphan_count);
 994
 995         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 996                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
 997                 BUG_ON(sk != req->rsk_listener);
 998
 999                 /* Paranoid, to prevent race condition if
1000                  * an inbound pkt destined for child is
1001                  * blocked by sock lock in tcp_v4_rcv().
1002                  * Also to satisfy an assertion in
1003                  * tcp_v4_destroy_sock().
1004                  */
1005                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1006         }
1007         inet_csk_destroy_sock(child);
1008 }
1009
1010 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1011                                       struct request_sock *req,
1012                                       struct sock *child)
1013 {
1014         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1015
1016         spin_lock(&queue->rskq_lock);
1017         if (unlikely(sk->sk_state != TCP_LISTEN)) {
1018                 inet_child_forget(sk, req, child);
1019                 child = NULL;
1020         } else {
1021                 req->sk = child;
1022                 req->dl_next = NULL;
1023                 if (queue->rskq_accept_head == NULL)
1024                         WRITE_ONCE(queue->rskq_accept_head, req);
1025                 else
1026                         queue->rskq_accept_tail->dl_next = req;
1027                 queue->rskq_accept_tail = req;
1028                 sk_acceptq_added(sk);
1029         }
1030         spin_unlock(&queue->rskq_lock);
1031         return child;
1032 }
1033 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1034
1035 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1036                                          struct request_sock *req, bool own_req)
1037 {
1038         if (own_req) {
1039                 inet_csk_reqsk_queue_drop(sk, req);
1040                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
1041                 if (inet_csk_reqsk_queue_add(sk, req, child))
1042                         return child;
1043         }
1044         /* Too bad, another child took ownership of the request, undo. */
1045         bh_unlock_sock(child);
1046         sock_put(child);
1047         return NULL;
1048 }
1049 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1050
1051 /*
1052  *      This routine closes sockets which have been at least partially
1053  *      opened, but not yet accepted.
1054  */
1055 void inet_csk_listen_stop(struct sock *sk)
1056 {
1057         struct inet_connection_sock *icsk = inet_csk(sk);
1058         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1059         struct request_sock *next, *req;
1060
1061         /* Following specs, it would be better either to send FIN
1062          * (and enter FIN-WAIT-1, it is normal close)
1063          * or to send active reset (abort).
1064          * Certainly, it is pretty dangerous while synflood, but it is
1065          * bad justification for our negligence 8)
1066          * To be honest, we are not able to make either
1067          * of the variants now.                 --ANK
1068          */
1069         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1070                 struct sock *child = req->sk;
1071
1072                 local_bh_disable();
1073                 bh_lock_sock(child);
1074                 WARN_ON(sock_owned_by_user(child));
1075                 sock_hold(child);
1076
1077                 inet_child_forget(sk, req, child);
1078                 reqsk_put(req);
1079                 bh_unlock_sock(child);
1080                 local_bh_enable();
1081                 sock_put(child);
1082
1083                 cond_resched();
1084         }
1085         if (queue->fastopenq.rskq_rst_head) {
1086                 /* Free all the reqs queued in rskq_rst_head. */
1087                 spin_lock_bh(&queue->fastopenq.lock);
1088                 req = queue->fastopenq.rskq_rst_head;
1089                 queue->fastopenq.rskq_rst_head = NULL;
1090                 spin_unlock_bh(&queue->fastopenq.lock);
1091                 while (req != NULL) {
1092                         next = req->dl_next;
1093                         reqsk_put(req);
1094                         req = next;
1095                 }
1096         }
1097         WARN_ON_ONCE(sk->sk_ack_backlog);
1098 }
1099 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1100
1101 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1102 {
1103         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1104         const struct inet_sock *inet = inet_sk(sk);
1105
1106         sin->sin_family         = AF_INET;
1107         sin->sin_addr.s_addr    = inet->inet_daddr;
1108         sin->sin_port           = inet->inet_dport;
1109 }
1110 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1111
1112 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1113 {
1114         const struct inet_sock *inet = inet_sk(sk);
1115         const struct ip_options_rcu *inet_opt;
1116         __be32 daddr = inet->inet_daddr;
1117         struct flowi4 *fl4;
1118         struct rtable *rt;
1119
1120         rcu_read_lock();
1121         inet_opt = rcu_dereference(inet->inet_opt);
1122         if (inet_opt && inet_opt->opt.srr)
1123                 daddr = inet_opt->opt.faddr;
1124         fl4 = &fl->u.ip4;
1125         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1126                                    inet->inet_saddr, inet->inet_dport,
1127                                    inet->inet_sport, sk->sk_protocol,
1128                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1129         if (IS_ERR(rt))
1130                 rt = NULL;
1131         if (rt)
1132                 sk_setup_caps(sk, &rt->dst);
1133         rcu_read_unlock();
1134
1135         return &rt->dst;
1136 }
1137
1138 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1139 {
1140         struct dst_entry *dst = __sk_dst_check(sk, 0);
1141         struct inet_sock *inet = inet_sk(sk);
1142
1143         if (!dst) {
1144                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1145                 if (!dst)
1146                         goto out;
1147         }
1148         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1149
1150         dst = __sk_dst_check(sk, 0);
1151         if (!dst)
1152                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1153 out:
1154         return dst;
1155 }
1156 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);