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