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