net/ipv4/inet_connection_sock.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Support for INET connection oriented protocols.
   7  *
   8  * Authors:     See the TCP sources
   9  *
  10  *              This program is free software; you can redistribute it and/or
  11  *              modify it under the terms of the GNU General Public License
  12  *              as published by the Free Software Foundation; either version
  13  *              2 of the License, or(at your option) any later version.
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/jhash.h>
  18
  19 #include <net/inet_connection_sock.h>
  20 #include <net/inet_hashtables.h>
  21 #include <net/inet_timewait_sock.h>
  22 #include <net/ip.h>
  23 #include <net/route.h>
  24 #include <net/tcp_states.h>
  25 #include <net/xfrm.h>
  26 #include <net/tcp.h>
  27 #include <net/sock_reuseport.h>
  28 #include <net/addrconf.h>
  29
  30 #if IS_ENABLED(CONFIG_IPV6)
  31 /* match_sk*_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses
  32  *                              if IPv6 only, and any IPv4 addresses
  33  *                              if not IPv6 only
  34  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  35  *                              IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  36  *                              and 0.0.0.0 equals to 0.0.0.0 only
  37  */
  38 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  39                                  const struct in6_addr *sk2_rcv_saddr6,
  40                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  41                                  bool sk1_ipv6only, bool sk2_ipv6only,
  42                                  bool match_sk1_wildcard,
  43                                  bool match_sk2_wildcard)
  44 {
  45         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  46         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  47
  48         /* if both are mapped, treat as IPv4 */
  49         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  50                 if (!sk2_ipv6only) {
  51                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  52                                 return true;
  53                         return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  54                                 (match_sk2_wildcard && !sk2_rcv_saddr);
  55                 }
  56                 return false;
  57         }
  58
  59         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  60                 return true;
  61
  62         if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard &&
  63             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  64                 return true;
  65
  66         if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard &&
  67             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  68                 return true;
  69
  70         if (sk2_rcv_saddr6 &&
  71             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  72                 return true;
  73
  74         return false;
  75 }
  76 #endif
  77
  78 /* match_sk*_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  79  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  80  *                              0.0.0.0 only equals to 0.0.0.0
  81  */
  82 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  83                                  bool sk2_ipv6only, bool match_sk1_wildcard,
  84                                  bool match_sk2_wildcard)
  85 {
  86         if (!sk2_ipv6only) {
  87                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  88                         return true;
  89                 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  90                         (match_sk2_wildcard && !sk2_rcv_saddr);
  91         }
  92         return false;
  93 }
  94
  95 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  96                           bool match_wildcard)
  97 {
  98 #if IS_ENABLED(CONFIG_IPV6)
  99         if (sk->sk_family == AF_INET6)
 100                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
 101                                             inet6_rcv_saddr(sk2),
 102                                             sk->sk_rcv_saddr,
 103                                             sk2->sk_rcv_saddr,
 104                                             ipv6_only_sock(sk),
 105                                             ipv6_only_sock(sk2),
 106                                             match_wildcard,
 107                                             match_wildcard);
 108 #endif
 109         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 110                                     ipv6_only_sock(sk2), match_wildcard,
 111                                     match_wildcard);
 112 }
 113 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 114
 115 bool inet_rcv_saddr_any(const struct sock *sk)
 116 {
 117 #if IS_ENABLED(CONFIG_IPV6)
 118         if (sk->sk_family == AF_INET6)
 119                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 120 #endif
 121         return !sk->sk_rcv_saddr;
 122 }
 123
 124 void inet_get_local_port_range(struct net *net, int *low, int *high)
 125 {
 126         unsigned int seq;
 127
 128         do {
 129                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 130
 131                 *low = net->ipv4.ip_local_ports.range[0];
 132                 *high = net->ipv4.ip_local_ports.range[1];
 133         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 134 }
 135 EXPORT_SYMBOL(inet_get_local_port_range);
 136
 137 static int inet_csk_bind_conflict(const struct sock *sk,
 138                                   const struct inet_bind_bucket *tb,
 139                                   bool relax, bool reuseport_ok)
 140 {
 141         struct sock *sk2;
 142         bool reuse = sk->sk_reuse;
 143         bool reuseport = !!sk->sk_reuseport && reuseport_ok;
 144         kuid_t uid = sock_i_uid((struct sock *)sk);
 145
 146         /*
 147          * Unlike other sk lookup places we do not check
 148          * for sk_net here, since _all_ the socks listed
 149          * in tb->owners list belong to the same net - the
 150          * one this bucket belongs to.
 151          */
 152
 153         sk_for_each_bound(sk2, &tb->owners) {
 154                 if (sk != sk2 &&
 155                     (!sk->sk_bound_dev_if ||
 156                      !sk2->sk_bound_dev_if ||
 157                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 158                         if ((!reuse || !sk2->sk_reuse ||
 159                             sk2->sk_state == TCP_LISTEN) &&
 160                             (!reuseport || !sk2->sk_reuseport ||
 161                              rcu_access_pointer(sk->sk_reuseport_cb) ||
 162                              (sk2->sk_state != TCP_TIME_WAIT &&
 163                              !uid_eq(uid, sock_i_uid(sk2))))) {
 164                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 165                                         break;
 166                         }
 167                         if (!relax && reuse && sk2->sk_reuse &&
 168                             sk2->sk_state != TCP_LISTEN) {
 169                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 170                                         break;
 171                         }
 172                 }
 173         }
 174         return sk2 != NULL;
 175 }
 176
 177 /*
 178  * Find an open port number for the socket.  Returns with the
 179  * inet_bind_hashbucket lock held.
 180  */
 181 static struct inet_bind_hashbucket *
 182 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 183 {
 184         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 185         int port = 0;
 186         struct inet_bind_hashbucket *head;
 187         struct net *net = sock_net(sk);
 188         int i, low, high, attempt_half;
 189         struct inet_bind_bucket *tb;
 190         u32 remaining, offset;
 191
 192         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 193 other_half_scan:
 194         inet_get_local_port_range(net, &low, &high);
 195         high++; /* [32768, 60999] -> [32768, 61000[ */
 196         if (high - low < 4)
 197                 attempt_half = 0;
 198         if (attempt_half) {
 199                 int half = low + (((high - low) >> 2) << 1);
 200
 201                 if (attempt_half == 1)
 202                         high = half;
 203                 else
 204                         low = half;
 205         }
 206         remaining = high - low;
 207         if (likely(remaining > 1))
 208                 remaining &= ~1U;
 209
 210         offset = prandom_u32() % remaining;
 211         /* __inet_hash_connect() favors ports having @low parity
 212          * We do the opposite to not pollute connect() users.
 213          */
 214         offset |= 1U;
 215
 216 other_parity_scan:
 217         port = low + offset;
 218         for (i = 0; i < remaining; i += 2, port += 2) {
 219                 if (unlikely(port >= high))
 220                         port -= remaining;
 221                 if (inet_is_local_reserved_port(net, port))
 222                         continue;
 223                 head = &hinfo->bhash[inet_bhashfn(net, port,
 224                                                   hinfo->bhash_size)];
 225                 spin_lock_bh(&head->lock);
 226                 inet_bind_bucket_for_each(tb, &head->chain)
 227                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 228                                 if (!inet_csk_bind_conflict(sk, tb, false, false))
 229                                         goto success;
 230                                 goto next_port;
 231                         }
 232                 tb = NULL;
 233                 goto success;
 234 next_port:
 235                 spin_unlock_bh(&head->lock);
 236                 cond_resched();
 237         }
 238
 239         offset--;
 240         if (!(offset & 1))
 241                 goto other_parity_scan;
 242
 243         if (attempt_half == 1) {
 244                 /* OK we now try the upper half of the range */
 245                 attempt_half = 2;
 246                 goto other_half_scan;
 247         }
 248         return NULL;
 249 success:
 250         *port_ret = port;
 251         *tb_ret = tb;
 252         return head;
 253 }
 254
 255 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 256                                      struct sock *sk)
 257 {
 258         kuid_t uid = sock_i_uid(sk);
 259
 260         if (tb->fastreuseport <= 0)
 261                 return 0;
 262         if (!sk->sk_reuseport)
 263                 return 0;
 264         if (rcu_access_pointer(sk->sk_reuseport_cb))
 265                 return 0;
 266         if (!uid_eq(tb->fastuid, uid))
 267                 return 0;
 268         /* We only need to check the rcv_saddr if this tb was once marked
 269          * without fastreuseport and then was reset, as we can only know that
 270          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 271          * owners list.
 272          */
 273         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 274                 return 1;
 275 #if IS_ENABLED(CONFIG_IPV6)
 276         if (tb->fast_sk_family == AF_INET6)
 277                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 278                                             inet6_rcv_saddr(sk),
 279                                             tb->fast_rcv_saddr,
 280                                             sk->sk_rcv_saddr,
 281                                             tb->fast_ipv6_only,
 282                                             ipv6_only_sock(sk), true, false);
 283 #endif
 284         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 285                                     ipv6_only_sock(sk), true, false);
 286 }
 287
 288 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 289                                struct sock *sk)
 290 {
 291         kuid_t uid = sock_i_uid(sk);
 292         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 293
 294         if (hlist_empty(&tb->owners)) {
 295                 tb->fastreuse = reuse;
 296                 if (sk->sk_reuseport) {
 297                         tb->fastreuseport = FASTREUSEPORT_ANY;
 298                         tb->fastuid = uid;
 299                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 300                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 301                         tb->fast_sk_family = sk->sk_family;
 302 #if IS_ENABLED(CONFIG_IPV6)
 303                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 304 #endif
 305                 } else {
 306                         tb->fastreuseport = 0;
 307                 }
 308         } else {
 309                 if (!reuse)
 310                         tb->fastreuse = 0;
 311                 if (sk->sk_reuseport) {
 312                         /* We didn't match or we don't have fastreuseport set on
 313                          * the tb, but we have sk_reuseport set on this socket
 314                          * and we know that there are no bind conflicts with
 315                          * this socket in this tb, so reset our tb's reuseport
 316                          * settings so that any subsequent sockets that match
 317                          * our current socket will be put on the fast path.
 318                          *
 319                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 320                          * do extra checking for all subsequent sk_reuseport
 321                          * socks.
 322                          */
 323                         if (!sk_reuseport_match(tb, sk)) {
 324                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 325                                 tb->fastuid = uid;
 326                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 327                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 328                                 tb->fast_sk_family = sk->sk_family;
 329 #if IS_ENABLED(CONFIG_IPV6)
 330                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 331 #endif
 332                         }
 333                 } else {
 334                         tb->fastreuseport = 0;
 335                 }
 336         }
 337 }
 338
 339 /* Obtain a reference to a local port for the given sock,
 340  * if snum is zero it means select any available local port.
 341  * We try to allocate an odd port (and leave even ports for connect())
 342  */
 343 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 344 {
 345         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 346         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 347         int ret = 1, port = snum;
 348         struct inet_bind_hashbucket *head;
 349         struct net *net = sock_net(sk);
 350         struct inet_bind_bucket *tb = NULL;
 351
 352         if (!port) {
 353                 head = inet_csk_find_open_port(sk, &tb, &port);
 354                 if (!head)
 355                         return ret;
 356                 if (!tb)
 357                         goto tb_not_found;
 358                 goto success;
 359         }
 360         head = &hinfo->bhash[inet_bhashfn(net, port,
 361                                           hinfo->bhash_size)];
 362         spin_lock_bh(&head->lock);
 363         inet_bind_bucket_for_each(tb, &head->chain)
 364                 if (net_eq(ib_net(tb), net) && tb->port == port)
 365                         goto tb_found;
 366 tb_not_found:
 367         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 368                                      net, head, port);
 369         if (!tb)
 370                 goto fail_unlock;
 371 tb_found:
 372         if (!hlist_empty(&tb->owners)) {
 373                 if (sk->sk_reuse == SK_FORCE_REUSE)
 374                         goto success;
 375
 376                 if ((tb->fastreuse > 0 && reuse) ||
 377                     sk_reuseport_match(tb, sk))
 378                         goto success;
 379                 if (inet_csk_bind_conflict(sk, tb, true, true))
 380                         goto fail_unlock;
 381         }
 382 success:
 383         inet_csk_update_fastreuse(tb, sk);
 384
 385         if (!inet_csk(sk)->icsk_bind_hash)
 386                 inet_bind_hash(sk, tb, port);
 387         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 388         ret = 0;
 389
 390 fail_unlock:
 391         spin_unlock_bh(&head->lock);
 392         return ret;
 393 }
 394 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 395
 396 /*
 397  * Wait for an incoming connection, avoid race conditions. This must be called
 398  * with the socket locked.
 399  */
 400 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 401 {
 402         struct inet_connection_sock *icsk = inet_csk(sk);
 403         DEFINE_WAIT(wait);
 404         int err;
 405
 406         /*
 407          * True wake-one mechanism for incoming connections: only
 408          * one process gets woken up, not the 'whole herd'.
 409          * Since we do not 'race & poll' for established sockets
 410          * anymore, the common case will execute the loop only once.
 411          *
 412          * Subtle issue: "add_wait_queue_exclusive()" will be added
 413          * after any current non-exclusive waiters, and we know that
 414          * it will always _stay_ after any new non-exclusive waiters
 415          * because all non-exclusive waiters are added at the
 416          * beginning of the wait-queue. As such, it's ok to "drop"
 417          * our exclusiveness temporarily when we get woken up without
 418          * having to remove and re-insert us on the wait queue.
 419          */
 420         for (;;) {
 421                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 422                                           TASK_INTERRUPTIBLE);
 423                 release_sock(sk);
 424                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 425                         timeo = schedule_timeout(timeo);
 426                 sched_annotate_sleep();
 427                 lock_sock(sk);
 428                 err = 0;
 429                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 430                         break;
 431                 err = -EINVAL;
 432                 if (sk->sk_state != TCP_LISTEN)
 433                         break;
 434                 err = sock_intr_errno(timeo);
 435                 if (signal_pending(current))
 436                         break;
 437                 err = -EAGAIN;
 438                 if (!timeo)
 439                         break;
 440         }
 441         finish_wait(sk_sleep(sk), &wait);
 442         return err;
 443 }
 444
 445 /*
 446  * This will accept the next outstanding connection.
 447  */
 448 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 449 {
 450         struct inet_connection_sock *icsk = inet_csk(sk);
 451         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 452         struct request_sock *req;
 453         struct sock *newsk;
 454         int error;
 455
 456         lock_sock(sk);
 457
 458         /* We need to make sure that this socket is listening,
 459          * and that it has something pending.
 460          */
 461         error = -EINVAL;
 462         if (sk->sk_state != TCP_LISTEN)
 463                 goto out_err;
 464
 465         /* Find already established connection */
 466         if (reqsk_queue_empty(queue)) {
 467                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 468
 469                 /* If this is a non blocking socket don't sleep */
 470                 error = -EAGAIN;
 471                 if (!timeo)
 472                         goto out_err;
 473
 474                 error = inet_csk_wait_for_connect(sk, timeo);
 475                 if (error)
 476                         goto out_err;
 477         }
 478         req = reqsk_queue_remove(queue, sk);
 479         newsk = req->sk;
 480
 481         if (sk->sk_protocol == IPPROTO_TCP &&
 482             tcp_rsk(req)->tfo_listener) {
 483                 spin_lock_bh(&queue->fastopenq.lock);
 484                 if (tcp_rsk(req)->tfo_listener) {
 485                         /* We are still waiting for the final ACK from 3WHS
 486                          * so can't free req now. Instead, we set req->sk to
 487                          * NULL to signify that the child socket is taken
 488                          * so reqsk_fastopen_remove() will free the req
 489                          * when 3WHS finishes (or is aborted).
 490                          */
 491                         req->sk = NULL;
 492                         req = NULL;
 493                 }
 494                 spin_unlock_bh(&queue->fastopenq.lock);
 495         }
 496
 497 out:
 498         release_sock(sk);
 499         if (newsk && mem_cgroup_sockets_enabled) {
 500                 int amt;
 501
 502                 /* atomically get the memory usage, set and charge the
 503                  * newsk->sk_memcg.
 504                  */
 505                 lock_sock(newsk);
 506
 507                 /* The socket has not been accepted yet, no need to look at
 508                  * newsk->sk_wmem_queued.
 509                  */
 510                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 511                                    atomic_read(&newsk->sk_rmem_alloc));
 512                 mem_cgroup_sk_alloc(newsk);
 513                 if (newsk->sk_memcg && amt)
 514                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
 515
 516                 release_sock(newsk);
 517         }
 518         if (req)
 519                 reqsk_put(req);
 520         return newsk;
 521 out_err:
 522         newsk = NULL;
 523         req = NULL;
 524         *err = error;
 525         goto out;
 526 }
 527 EXPORT_SYMBOL(inet_csk_accept);
 528
 529 /*
 530  * Using different timers for retransmit, delayed acks and probes
 531  * We may wish use just one timer maintaining a list of expire jiffies
 532  * to optimize.
 533  */
 534 void inet_csk_init_xmit_timers(struct sock *sk,
 535                                void (*retransmit_handler)(struct timer_list *t),
 536                                void (*delack_handler)(struct timer_list *t),
 537                                void (*keepalive_handler)(struct timer_list *t))
 538 {
 539         struct inet_connection_sock *icsk = inet_csk(sk);
 540
 541         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 542         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 543         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 544         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 545 }
 546 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 547
 548 void inet_csk_clear_xmit_timers(struct sock *sk)
 549 {
 550         struct inet_connection_sock *icsk = inet_csk(sk);
 551
 552         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 553
 554         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 555         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 556         sk_stop_timer(sk, &sk->sk_timer);
 557 }
 558 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 559
 560 void inet_csk_delete_keepalive_timer(struct sock *sk)
 561 {
 562         sk_stop_timer(sk, &sk->sk_timer);
 563 }
 564 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 565
 566 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 567 {
 568         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 569 }
 570 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 571
 572 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 573                                      struct flowi4 *fl4,
 574                                      const struct request_sock *req)
 575 {
 576         const struct inet_request_sock *ireq = inet_rsk(req);
 577         struct net *net = read_pnet(&ireq->ireq_net);
 578         struct ip_options_rcu *opt;
 579         struct rtable *rt;
 580
 581         rcu_read_lock();
 582         opt = rcu_dereference(ireq->ireq_opt);
 583
 584         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 585                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 586                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 587                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 588                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 589                            htons(ireq->ir_num), sk->sk_uid);
 590         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 591         rt = ip_route_output_flow(net, fl4, sk);
 592         if (IS_ERR(rt))
 593                 goto no_route;
 594         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 595                 goto route_err;
 596         rcu_read_unlock();
 597         return &rt->dst;
 598
 599 route_err:
 600         ip_rt_put(rt);
 601 no_route:
 602         rcu_read_unlock();
 603         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 604         return NULL;
 605 }
 606 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 607
 608 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 609                                             struct sock *newsk,
 610                                             const struct request_sock *req)
 611 {
 612         const struct inet_request_sock *ireq = inet_rsk(req);
 613         struct net *net = read_pnet(&ireq->ireq_net);
 614         struct inet_sock *newinet = inet_sk(newsk);
 615         struct ip_options_rcu *opt;
 616         struct flowi4 *fl4;
 617         struct rtable *rt;
 618
 619         opt = rcu_dereference(ireq->ireq_opt);
 620         fl4 = &newinet->cork.fl.u.ip4;
 621
 622         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 623                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 624                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 625                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 626                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 627                            htons(ireq->ir_num), sk->sk_uid);
 628         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 629         rt = ip_route_output_flow(net, fl4, sk);
 630         if (IS_ERR(rt))
 631                 goto no_route;
 632         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 633                 goto route_err;
 634         return &rt->dst;
 635
 636 route_err:
 637         ip_rt_put(rt);
 638 no_route:
 639         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 640         return NULL;
 641 }
 642 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 643
 644 #if IS_ENABLED(CONFIG_IPV6)
 645 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 646 #else
 647 #define AF_INET_FAMILY(fam) true
 648 #endif
 649
 650 /* Decide when to expire the request and when to resend SYN-ACK */
 651 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 652                                   const int max_retries,
 653                                   const u8 rskq_defer_accept,
 654                                   int *expire, int *resend)
 655 {
 656         if (!rskq_defer_accept) {
 657                 *expire = req->num_timeout >= thresh;
 658                 *resend = 1;
 659                 return;
 660         }
 661         *expire = req->num_timeout >= thresh &&
 662                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 663         /*
 664          * Do not resend while waiting for data after ACK,
 665          * start to resend on end of deferring period to give
 666          * last chance for data or ACK to create established socket.
 667          */
 668         *resend = !inet_rsk(req)->acked ||
 669                   req->num_timeout >= rskq_defer_accept - 1;
 670 }
 671
 672 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 673 {
 674         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 675
 676         if (!err)
 677                 req->num_retrans++;
 678         return err;
 679 }
 680 EXPORT_SYMBOL(inet_rtx_syn_ack);
 681
 682 /* return true if req was found in the ehash table */
 683 static bool reqsk_queue_unlink(struct request_sock_queue *queue,
 684                                struct request_sock *req)
 685 {
 686         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 687         bool found = false;
 688
 689         if (sk_hashed(req_to_sk(req))) {
 690                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 691
 692                 spin_lock(lock);
 693                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 694                 spin_unlock(lock);
 695         }
 696         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 697                 reqsk_put(req);
 698         return found;
 699 }
 700
 701 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 702 {
 703         bool unlinked = reqsk_queue_unlink(&inet_csk(sk)->icsk_accept_queue, req);
 704
 705         if (unlinked) {
 706                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 707                 reqsk_put(req);
 708         }
 709         return unlinked;
 710 }
 711 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 712
 713 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 714 {
 715         inet_csk_reqsk_queue_drop(sk, req);
 716         reqsk_put(req);
 717 }
 718 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 719
 720 static void reqsk_timer_handler(struct timer_list *t)
 721 {
 722         struct request_sock *req = from_timer(req, t, rsk_timer);
 723         struct sock *sk_listener = req->rsk_listener;
 724         struct net *net = sock_net(sk_listener);
 725         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 726         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 727         int qlen, expire = 0, resend = 0;
 728         int max_retries, thresh;
 729         u8 defer_accept;
 730
 731         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 732                 goto drop;
 733
 734         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 735         thresh = max_retries;
 736         /* Normally all the openreqs are young and become mature
 737          * (i.e. converted to established socket) for first timeout.
 738          * If synack was not acknowledged for 1 second, it means
 739          * one of the following things: synack was lost, ack was lost,
 740          * rtt is high or nobody planned to ack (i.e. synflood).
 741          * When server is a bit loaded, queue is populated with old
 742          * open requests, reducing effective size of queue.
 743          * When server is well loaded, queue size reduces to zero
 744          * after several minutes of work. It is not synflood,
 745          * it is normal operation. The solution is pruning
 746          * too old entries overriding normal timeout, when
 747          * situation becomes dangerous.
 748          *
 749          * Essentially, we reserve half of room for young
 750          * embrions; and abort old ones without pity, if old
 751          * ones are about to clog our table.
 752          */
 753         qlen = reqsk_queue_len(queue);
 754         if ((qlen << 1) > max(8U, sk_listener->sk_max_ack_backlog)) {
 755                 int young = reqsk_queue_len_young(queue) << 1;
 756
 757                 while (thresh > 2) {
 758                         if (qlen < young)
 759                                 break;
 760                         thresh--;
 761                         young <<= 1;
 762                 }
 763         }
 764         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 765         if (defer_accept)
 766                 max_retries = defer_accept;
 767         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 768                        &expire, &resend);
 769         req->rsk_ops->syn_ack_timeout(req);
 770         if (!expire &&
 771             (!resend ||
 772              !inet_rtx_syn_ack(sk_listener, req) ||
 773              inet_rsk(req)->acked)) {
 774                 unsigned long timeo;
 775
 776                 if (req->num_timeout++ == 0)
 777                         atomic_dec(&queue->young);
 778                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 779                 mod_timer(&req->rsk_timer, jiffies + timeo);
 780                 return;
 781         }
 782 drop:
 783         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 784 }
 785
 786 static void reqsk_queue_hash_req(struct request_sock *req,
 787                                  unsigned long timeout)
 788 {
 789         req->num_retrans = 0;
 790         req->num_timeout = 0;
 791         req->sk = NULL;
 792
 793         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 794         mod_timer(&req->rsk_timer, jiffies + timeout);
 795
 796         inet_ehash_insert(req_to_sk(req), NULL, NULL);
 797         /* before letting lookups find us, make sure all req fields
 798          * are committed to memory and refcnt initialized.
 799          */
 800         smp_wmb();
 801         refcount_set(&req->rsk_refcnt, 2 + 1);
 802 }
 803
 804 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 805                                    unsigned long timeout)
 806 {
 807         reqsk_queue_hash_req(req, timeout);
 808         inet_csk_reqsk_queue_added(sk);
 809 }
 810 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 811
 812 /**
 813  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 814  *      @sk: the socket to clone
 815  *      @req: request_sock
 816  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 817  *
 818  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 819  */
 820 struct sock *inet_csk_clone_lock(const struct sock *sk,
 821                                  const struct request_sock *req,
 822                                  const gfp_t priority)
 823 {
 824         struct sock *newsk = sk_clone_lock(sk, priority);
 825
 826         if (newsk) {
 827                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 828
 829                 newsk->sk_wait_pending = 0;
 830                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 831                 newicsk->icsk_bind_hash = NULL;
 832
 833                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 834                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 835                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 836
 837                 /* listeners have SOCK_RCU_FREE, not the children */
 838                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 839
 840                 inet_sk(newsk)->mc_list = NULL;
 841
 842                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 843                 atomic64_set(&newsk->sk_cookie,
 844                              atomic64_read(&inet_rsk(req)->ir_cookie));
 845
 846                 newicsk->icsk_retransmits = 0;
 847                 newicsk->icsk_backoff     = 0;
 848                 newicsk->icsk_probes_out  = 0;
 849
 850                 /* Deinitialize accept_queue to trap illegal accesses. */
 851                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 852
 853                 security_inet_csk_clone(newsk, req);
 854         }
 855         return newsk;
 856 }
 857 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 858
 859 /*
 860  * At this point, there should be no process reference to this
 861  * socket, and thus no user references at all.  Therefore we
 862  * can assume the socket waitqueue is inactive and nobody will
 863  * try to jump onto it.
 864  */
 865 void inet_csk_destroy_sock(struct sock *sk)
 866 {
 867         WARN_ON(sk->sk_state != TCP_CLOSE);
 868         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 869
 870         /* It cannot be in hash table! */
 871         WARN_ON(!sk_unhashed(sk));
 872
 873         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 874         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 875
 876         sk->sk_prot->destroy(sk);
 877
 878         sk_stream_kill_queues(sk);
 879
 880         xfrm_sk_free_policy(sk);
 881
 882         sk_refcnt_debug_release(sk);
 883
 884         percpu_counter_dec(sk->sk_prot->orphan_count);
 885
 886         sock_put(sk);
 887 }
 888 EXPORT_SYMBOL(inet_csk_destroy_sock);
 889
 890 /* This function allows to force a closure of a socket after the call to
 891  * tcp/dccp_create_openreq_child().
 892  */
 893 void inet_csk_prepare_forced_close(struct sock *sk)
 894         __releases(&sk->sk_lock.slock)
 895 {
 896         /* sk_clone_lock locked the socket and set refcnt to 2 */
 897         bh_unlock_sock(sk);
 898         sock_put(sk);
 899
 900         /* The below has to be done to allow calling inet_csk_destroy_sock */
 901         sock_set_flag(sk, SOCK_DEAD);
 902         percpu_counter_inc(sk->sk_prot->orphan_count);
 903         inet_sk(sk)->inet_num = 0;
 904 }
 905 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 906
 907 static int inet_ulp_can_listen(const struct sock *sk)
 908 {
 909         const struct inet_connection_sock *icsk = inet_csk(sk);
 910
 911         if (icsk->icsk_ulp_ops)
 912                 return -EINVAL;
 913
 914         return 0;
 915 }
 916
 917 int inet_csk_listen_start(struct sock *sk, int backlog)
 918 {
 919         struct inet_connection_sock *icsk = inet_csk(sk);
 920         struct inet_sock *inet = inet_sk(sk);
 921         int err;
 922
 923         err = inet_ulp_can_listen(sk);
 924         if (unlikely(err))
 925                 return err;
 926
 927         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 928
 929         sk->sk_max_ack_backlog = backlog;
 930         sk->sk_ack_backlog = 0;
 931         inet_csk_delack_init(sk);
 932
 933         /* There is race window here: we announce ourselves listening,
 934          * but this transition is still not validated by get_port().
 935          * It is OK, because this socket enters to hash table only
 936          * after validation is complete.
 937          */
 938         err = -EADDRINUSE;
 939         inet_sk_state_store(sk, TCP_LISTEN);
 940         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 941                 inet->inet_sport = htons(inet->inet_num);
 942
 943                 sk_dst_reset(sk);
 944                 err = sk->sk_prot->hash(sk);
 945
 946                 if (likely(!err))
 947                         return 0;
 948         }
 949
 950         inet_sk_set_state(sk, TCP_CLOSE);
 951         return err;
 952 }
 953 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 954
 955 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 956                               struct sock *child)
 957 {
 958         sk->sk_prot->disconnect(child, O_NONBLOCK);
 959
 960         sock_orphan(child);
 961
 962         percpu_counter_inc(sk->sk_prot->orphan_count);
 963
 964         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 965                 BUG_ON(tcp_sk(child)->fastopen_rsk != req);
 966                 BUG_ON(sk != req->rsk_listener);
 967
 968                 /* Paranoid, to prevent race condition if
 969                  * an inbound pkt destined for child is
 970                  * blocked by sock lock in tcp_v4_rcv().
 971                  * Also to satisfy an assertion in
 972                  * tcp_v4_destroy_sock().
 973                  */
 974                 tcp_sk(child)->fastopen_rsk = NULL;
 975         }
 976         inet_csk_destroy_sock(child);
 977 }
 978
 979 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 980                                       struct request_sock *req,
 981                                       struct sock *child)
 982 {
 983         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 984
 985         spin_lock(&queue->rskq_lock);
 986         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 987                 inet_child_forget(sk, req, child);
 988                 child = NULL;
 989         } else {
 990                 req->sk = child;
 991                 req->dl_next = NULL;
 992                 if (queue->rskq_accept_head == NULL)
 993                         WRITE_ONCE(queue->rskq_accept_head, req);
 994                 else
 995                         queue->rskq_accept_tail->dl_next = req;
 996                 queue->rskq_accept_tail = req;
 997                 sk_acceptq_added(sk);
 998         }
 999         spin_unlock(&queue->rskq_lock);
1000         return child;
1001 }
1002 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1003
1004 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1005                                          struct request_sock *req, bool own_req)
1006 {
1007         if (own_req) {
1008                 inet_csk_reqsk_queue_drop(sk, req);
1009                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
1010                 if (inet_csk_reqsk_queue_add(sk, req, child))
1011                         return child;
1012         }
1013         /* Too bad, another child took ownership of the request, undo. */
1014         bh_unlock_sock(child);
1015         sock_put(child);
1016         return NULL;
1017 }
1018 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1019
1020 /*
1021  *      This routine closes sockets which have been at least partially
1022  *      opened, but not yet accepted.
1023  */
1024 void inet_csk_listen_stop(struct sock *sk)
1025 {
1026         struct inet_connection_sock *icsk = inet_csk(sk);
1027         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1028         struct request_sock *next, *req;
1029
1030         /* Following specs, it would be better either to send FIN
1031          * (and enter FIN-WAIT-1, it is normal close)
1032          * or to send active reset (abort).
1033          * Certainly, it is pretty dangerous while synflood, but it is
1034          * bad justification for our negligence 8)
1035          * To be honest, we are not able to make either
1036          * of the variants now.                 --ANK
1037          */
1038         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1039                 struct sock *child = req->sk;
1040
1041                 local_bh_disable();
1042                 bh_lock_sock(child);
1043                 WARN_ON(sock_owned_by_user(child));
1044                 sock_hold(child);
1045
1046                 inet_child_forget(sk, req, child);
1047                 reqsk_put(req);
1048                 bh_unlock_sock(child);
1049                 local_bh_enable();
1050                 sock_put(child);
1051
1052                 cond_resched();
1053         }
1054         if (queue->fastopenq.rskq_rst_head) {
1055                 /* Free all the reqs queued in rskq_rst_head. */
1056                 spin_lock_bh(&queue->fastopenq.lock);
1057                 req = queue->fastopenq.rskq_rst_head;
1058                 queue->fastopenq.rskq_rst_head = NULL;
1059                 spin_unlock_bh(&queue->fastopenq.lock);
1060                 while (req != NULL) {
1061                         next = req->dl_next;
1062                         reqsk_put(req);
1063                         req = next;
1064                 }
1065         }
1066         WARN_ON_ONCE(sk->sk_ack_backlog);
1067 }
1068 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1069
1070 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1071 {
1072         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1073         const struct inet_sock *inet = inet_sk(sk);
1074
1075         sin->sin_family         = AF_INET;
1076         sin->sin_addr.s_addr    = inet->inet_daddr;
1077         sin->sin_port           = inet->inet_dport;
1078 }
1079 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1080
1081 #ifdef CONFIG_COMPAT
1082 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
1083                                char __user *optval, int __user *optlen)
1084 {
1085         const struct inet_connection_sock *icsk = inet_csk(sk);
1086
1087         if (icsk->icsk_af_ops->compat_getsockopt)
1088                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
1089                                                             optval, optlen);
1090         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
1091                                              optval, optlen);
1092 }
1093 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
1094
1095 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
1096                                char __user *optval, unsigned int optlen)
1097 {
1098         const struct inet_connection_sock *icsk = inet_csk(sk);
1099
1100         if (icsk->icsk_af_ops->compat_setsockopt)
1101                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
1102                                                             optval, optlen);
1103         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1104                                              optval, optlen);
1105 }
1106 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
1107 #endif
1108
1109 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1110 {
1111         const struct inet_sock *inet = inet_sk(sk);
1112         const struct ip_options_rcu *inet_opt;
1113         __be32 daddr = inet->inet_daddr;
1114         struct flowi4 *fl4;
1115         struct rtable *rt;
1116
1117         rcu_read_lock();
1118         inet_opt = rcu_dereference(inet->inet_opt);
1119         if (inet_opt && inet_opt->opt.srr)
1120                 daddr = inet_opt->opt.faddr;
1121         fl4 = &fl->u.ip4;
1122         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1123                                    inet->inet_saddr, inet->inet_dport,
1124                                    inet->inet_sport, sk->sk_protocol,
1125                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1126         if (IS_ERR(rt))
1127                 rt = NULL;
1128         if (rt)
1129                 sk_setup_caps(sk, &rt->dst);
1130         rcu_read_unlock();
1131
1132         return &rt->dst;
1133 }
1134
1135 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1136 {
1137         struct dst_entry *dst = __sk_dst_check(sk, 0);
1138         struct inet_sock *inet = inet_sk(sk);
1139
1140         if (!dst) {
1141                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1142                 if (!dst)
1143                         goto out;
1144         }
1145         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1146
1147         dst = __sk_dst_check(sk, 0);
1148         if (!dst)
1149                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1150 out:
1151         return dst;
1152 }
1153 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);