net/ipv4/inet_connection_sock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Support for INET connection oriented protocols.
   8  *
   9  * Authors:     See the TCP sources
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/jhash.h>
  14
  15 #include <net/inet_connection_sock.h>
  16 #include <net/inet_hashtables.h>
  17 #include <net/inet_timewait_sock.h>
  18 #include <net/ip.h>
  19 #include <net/route.h>
  20 #include <net/tcp_states.h>
  21 #include <net/xfrm.h>
  22 #include <net/tcp.h>
  23 #include <net/sock_reuseport.h>
  24 #include <net/addrconf.h>
  25
  26 #if IS_ENABLED(CONFIG_IPV6)
  27 /* match_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses if IPv6
  28  *                          only, and any IPv4 addresses if not IPv6 only
  29  * match_wildcard == false: addresses must be exactly the same, i.e.
  30  *                          IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  31  *                          and 0.0.0.0 equals to 0.0.0.0 only
  32  */
  33 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  34                                  const struct in6_addr *sk2_rcv_saddr6,
  35                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  36                                  bool sk1_ipv6only, bool sk2_ipv6only,
  37                                  bool match_wildcard)
  38 {
  39         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  40         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  41
  42         /* if both are mapped, treat as IPv4 */
  43         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  44                 if (!sk2_ipv6only) {
  45                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  46                                 return true;
  47                         if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  48                                 return match_wildcard;
  49                 }
  50                 return false;
  51         }
  52
  53         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  54                 return true;
  55
  56         if (addr_type2 == IPV6_ADDR_ANY && match_wildcard &&
  57             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  58                 return true;
  59
  60         if (addr_type == IPV6_ADDR_ANY && match_wildcard &&
  61             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  62                 return true;
  63
  64         if (sk2_rcv_saddr6 &&
  65             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  66                 return true;
  67
  68         return false;
  69 }
  70 #endif
  71
  72 /* match_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  73  * match_wildcard == false: addresses must be exactly the same, i.e.
  74  *                          0.0.0.0 only equals to 0.0.0.0
  75  */
  76 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  77                                  bool sk2_ipv6only, bool match_wildcard)
  78 {
  79         if (!sk2_ipv6only) {
  80                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  81                         return true;
  82                 if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  83                         return match_wildcard;
  84         }
  85         return false;
  86 }
  87
  88 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  89                           bool match_wildcard)
  90 {
  91 #if IS_ENABLED(CONFIG_IPV6)
  92         if (sk->sk_family == AF_INET6)
  93                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
  94                                             inet6_rcv_saddr(sk2),
  95                                             sk->sk_rcv_saddr,
  96                                             sk2->sk_rcv_saddr,
  97                                             ipv6_only_sock(sk),
  98                                             ipv6_only_sock(sk2),
  99                                             match_wildcard);
 100 #endif
 101         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 102                                     ipv6_only_sock(sk2), match_wildcard);
 103 }
 104 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 105
 106 bool inet_rcv_saddr_any(const struct sock *sk)
 107 {
 108 #if IS_ENABLED(CONFIG_IPV6)
 109         if (sk->sk_family == AF_INET6)
 110                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 111 #endif
 112         return !sk->sk_rcv_saddr;
 113 }
 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         int l3mdev;
 183
 184         l3mdev = inet_sk_bound_l3mdev(sk);
 185         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 186 other_half_scan:
 187         inet_get_local_port_range(net, &low, &high);
 188         high++; /* [32768, 60999] -> [32768, 61000[ */
 189         if (high - low < 4)
 190                 attempt_half = 0;
 191         if (attempt_half) {
 192                 int half = low + (((high - low) >> 2) << 1);
 193
 194                 if (attempt_half == 1)
 195                         high = half;
 196                 else
 197                         low = half;
 198         }
 199         remaining = high - low;
 200         if (likely(remaining > 1))
 201                 remaining &= ~1U;
 202
 203         offset = prandom_u32() % remaining;
 204         /* __inet_hash_connect() favors ports having @low parity
 205          * We do the opposite to not pollute connect() users.
 206          */
 207         offset |= 1U;
 208
 209 other_parity_scan:
 210         port = low + offset;
 211         for (i = 0; i < remaining; i += 2, port += 2) {
 212                 if (unlikely(port >= high))
 213                         port -= remaining;
 214                 if (inet_is_local_reserved_port(net, port))
 215                         continue;
 216                 head = &hinfo->bhash[inet_bhashfn(net, port,
 217                                                   hinfo->bhash_size)];
 218                 spin_lock_bh(&head->lock);
 219                 inet_bind_bucket_for_each(tb, &head->chain)
 220                         if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 221                             tb->port == port) {
 222                                 if (!inet_csk_bind_conflict(sk, tb, false, false))
 223                                         goto success;
 224                                 goto next_port;
 225                         }
 226                 tb = NULL;
 227                 goto success;
 228 next_port:
 229                 spin_unlock_bh(&head->lock);
 230                 cond_resched();
 231         }
 232
 233         offset--;
 234         if (!(offset & 1))
 235                 goto other_parity_scan;
 236
 237         if (attempt_half == 1) {
 238                 /* OK we now try the upper half of the range */
 239                 attempt_half = 2;
 240                 goto other_half_scan;
 241         }
 242         return NULL;
 243 success:
 244         *port_ret = port;
 245         *tb_ret = tb;
 246         return head;
 247 }
 248
 249 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 250                                      struct sock *sk)
 251 {
 252         kuid_t uid = sock_i_uid(sk);
 253
 254         if (tb->fastreuseport <= 0)
 255                 return 0;
 256         if (!sk->sk_reuseport)
 257                 return 0;
 258         if (rcu_access_pointer(sk->sk_reuseport_cb))
 259                 return 0;
 260         if (!uid_eq(tb->fastuid, uid))
 261                 return 0;
 262         /* We only need to check the rcv_saddr if this tb was once marked
 263          * without fastreuseport and then was reset, as we can only know that
 264          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 265          * owners list.
 266          */
 267         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 268                 return 1;
 269 #if IS_ENABLED(CONFIG_IPV6)
 270         if (tb->fast_sk_family == AF_INET6)
 271                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 272                                             inet6_rcv_saddr(sk),
 273                                             tb->fast_rcv_saddr,
 274                                             sk->sk_rcv_saddr,
 275                                             tb->fast_ipv6_only,
 276                                             ipv6_only_sock(sk), true);
 277 #endif
 278         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 279                                     ipv6_only_sock(sk), true);
 280 }
 281
 282 /* Obtain a reference to a local port for the given sock,
 283  * if snum is zero it means select any available local port.
 284  * We try to allocate an odd port (and leave even ports for connect())
 285  */
 286 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 287 {
 288         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 289         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 290         int ret = 1, port = snum;
 291         struct inet_bind_hashbucket *head;
 292         struct net *net = sock_net(sk);
 293         struct inet_bind_bucket *tb = NULL;
 294         kuid_t uid = sock_i_uid(sk);
 295         int l3mdev;
 296
 297         l3mdev = inet_sk_bound_l3mdev(sk);
 298
 299         if (!port) {
 300                 head = inet_csk_find_open_port(sk, &tb, &port);
 301                 if (!head)
 302                         return ret;
 303                 if (!tb)
 304                         goto tb_not_found;
 305                 goto success;
 306         }
 307         head = &hinfo->bhash[inet_bhashfn(net, port,
 308                                           hinfo->bhash_size)];
 309         spin_lock_bh(&head->lock);
 310         inet_bind_bucket_for_each(tb, &head->chain)
 311                 if (net_eq(ib_net(tb), net) && tb->l3mdev == l3mdev &&
 312                     tb->port == port)
 313                         goto tb_found;
 314 tb_not_found:
 315         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 316                                      net, head, port, l3mdev);
 317         if (!tb)
 318                 goto fail_unlock;
 319 tb_found:
 320         if (!hlist_empty(&tb->owners)) {
 321                 if (sk->sk_reuse == SK_FORCE_REUSE)
 322                         goto success;
 323
 324                 if ((tb->fastreuse > 0 && reuse) ||
 325                     sk_reuseport_match(tb, sk))
 326                         goto success;
 327                 if (inet_csk_bind_conflict(sk, tb, true, true))
 328                         goto fail_unlock;
 329         }
 330 success:
 331         if (hlist_empty(&tb->owners)) {
 332                 tb->fastreuse = reuse;
 333                 if (sk->sk_reuseport) {
 334                         tb->fastreuseport = FASTREUSEPORT_ANY;
 335                         tb->fastuid = uid;
 336                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 337                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 338                         tb->fast_sk_family = sk->sk_family;
 339 #if IS_ENABLED(CONFIG_IPV6)
 340                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 341 #endif
 342                 } else {
 343                         tb->fastreuseport = 0;
 344                 }
 345         } else {
 346                 if (!reuse)
 347                         tb->fastreuse = 0;
 348                 if (sk->sk_reuseport) {
 349                         /* We didn't match or we don't have fastreuseport set on
 350                          * the tb, but we have sk_reuseport set on this socket
 351                          * and we know that there are no bind conflicts with
 352                          * this socket in this tb, so reset our tb's reuseport
 353                          * settings so that any subsequent sockets that match
 354                          * our current socket will be put on the fast path.
 355                          *
 356                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 357                          * do extra checking for all subsequent sk_reuseport
 358                          * socks.
 359                          */
 360                         if (!sk_reuseport_match(tb, sk)) {
 361                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 362                                 tb->fastuid = uid;
 363                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 364                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 365                                 tb->fast_sk_family = sk->sk_family;
 366 #if IS_ENABLED(CONFIG_IPV6)
 367                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 368 #endif
 369                         }
 370                 } else {
 371                         tb->fastreuseport = 0;
 372                 }
 373         }
 374         if (!inet_csk(sk)->icsk_bind_hash)
 375                 inet_bind_hash(sk, tb, port);
 376         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 377         ret = 0;
 378
 379 fail_unlock:
 380         spin_unlock_bh(&head->lock);
 381         return ret;
 382 }
 383 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 384
 385 /*
 386  * Wait for an incoming connection, avoid race conditions. This must be called
 387  * with the socket locked.
 388  */
 389 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 390 {
 391         struct inet_connection_sock *icsk = inet_csk(sk);
 392         DEFINE_WAIT(wait);
 393         int err;
 394
 395         /*
 396          * True wake-one mechanism for incoming connections: only
 397          * one process gets woken up, not the 'whole herd'.
 398          * Since we do not 'race & poll' for established sockets
 399          * anymore, the common case will execute the loop only once.
 400          *
 401          * Subtle issue: "add_wait_queue_exclusive()" will be added
 402          * after any current non-exclusive waiters, and we know that
 403          * it will always _stay_ after any new non-exclusive waiters
 404          * because all non-exclusive waiters are added at the
 405          * beginning of the wait-queue. As such, it's ok to "drop"
 406          * our exclusiveness temporarily when we get woken up without
 407          * having to remove and re-insert us on the wait queue.
 408          */
 409         for (;;) {
 410                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 411                                           TASK_INTERRUPTIBLE);
 412                 release_sock(sk);
 413                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 414                         timeo = schedule_timeout(timeo);
 415                 sched_annotate_sleep();
 416                 lock_sock(sk);
 417                 err = 0;
 418                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 419                         break;
 420                 err = -EINVAL;
 421                 if (sk->sk_state != TCP_LISTEN)
 422                         break;
 423                 err = sock_intr_errno(timeo);
 424                 if (signal_pending(current))
 425                         break;
 426                 err = -EAGAIN;
 427                 if (!timeo)
 428                         break;
 429         }
 430         finish_wait(sk_sleep(sk), &wait);
 431         return err;
 432 }
 433
 434 /*
 435  * This will accept the next outstanding connection.
 436  */
 437 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 438 {
 439         struct inet_connection_sock *icsk = inet_csk(sk);
 440         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 441         struct request_sock *req;
 442         struct sock *newsk;
 443         int error;
 444
 445         lock_sock(sk);
 446
 447         /* We need to make sure that this socket is listening,
 448          * and that it has something pending.
 449          */
 450         error = -EINVAL;
 451         if (sk->sk_state != TCP_LISTEN)
 452                 goto out_err;
 453
 454         /* Find already established connection */
 455         if (reqsk_queue_empty(queue)) {
 456                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 457
 458                 /* If this is a non blocking socket don't sleep */
 459                 error = -EAGAIN;
 460                 if (!timeo)
 461                         goto out_err;
 462
 463                 error = inet_csk_wait_for_connect(sk, timeo);
 464                 if (error)
 465                         goto out_err;
 466         }
 467         req = reqsk_queue_remove(queue, sk);
 468         newsk = req->sk;
 469
 470         if (sk->sk_protocol == IPPROTO_TCP &&
 471             tcp_rsk(req)->tfo_listener) {
 472                 spin_lock_bh(&queue->fastopenq.lock);
 473                 if (tcp_rsk(req)->tfo_listener) {
 474                         /* We are still waiting for the final ACK from 3WHS
 475                          * so can't free req now. Instead, we set req->sk to
 476                          * NULL to signify that the child socket is taken
 477                          * so reqsk_fastopen_remove() will free the req
 478                          * when 3WHS finishes (or is aborted).
 479                          */
 480                         req->sk = NULL;
 481                         req = NULL;
 482                 }
 483                 spin_unlock_bh(&queue->fastopenq.lock);
 484         }
 485 out:
 486         release_sock(sk);
 487         if (req)
 488                 reqsk_put(req);
 489         return newsk;
 490 out_err:
 491         newsk = NULL;
 492         req = NULL;
 493         *err = error;
 494         goto out;
 495 }
 496 EXPORT_SYMBOL(inet_csk_accept);
 497
 498 /*
 499  * Using different timers for retransmit, delayed acks and probes
 500  * We may wish use just one timer maintaining a list of expire jiffies
 501  * to optimize.
 502  */
 503 void inet_csk_init_xmit_timers(struct sock *sk,
 504                                void (*retransmit_handler)(struct timer_list *t),
 505                                void (*delack_handler)(struct timer_list *t),
 506                                void (*keepalive_handler)(struct timer_list *t))
 507 {
 508         struct inet_connection_sock *icsk = inet_csk(sk);
 509
 510         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 511         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 512         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 513         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 514 }
 515 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 516
 517 void inet_csk_clear_xmit_timers(struct sock *sk)
 518 {
 519         struct inet_connection_sock *icsk = inet_csk(sk);
 520
 521         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 522
 523         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 524         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 525         sk_stop_timer(sk, &sk->sk_timer);
 526 }
 527 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 528
 529 void inet_csk_delete_keepalive_timer(struct sock *sk)
 530 {
 531         sk_stop_timer(sk, &sk->sk_timer);
 532 }
 533 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 534
 535 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 536 {
 537         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 538 }
 539 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 540
 541 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 542                                      struct flowi4 *fl4,
 543                                      const struct request_sock *req)
 544 {
 545         const struct inet_request_sock *ireq = inet_rsk(req);
 546         struct net *net = read_pnet(&ireq->ireq_net);
 547         struct ip_options_rcu *opt;
 548         struct rtable *rt;
 549
 550         rcu_read_lock();
 551         opt = rcu_dereference(ireq->ireq_opt);
 552
 553         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 554                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 555                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 556                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 557                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 558                            htons(ireq->ir_num), sk->sk_uid);
 559         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 560         rt = ip_route_output_flow(net, fl4, sk);
 561         if (IS_ERR(rt))
 562                 goto no_route;
 563         if (opt && opt->opt.is_strictroute && rt->rt_gw_family)
 564                 goto route_err;
 565         rcu_read_unlock();
 566         return &rt->dst;
 567
 568 route_err:
 569         ip_rt_put(rt);
 570 no_route:
 571         rcu_read_unlock();
 572         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 573         return NULL;
 574 }
 575 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 576
 577 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 578                                             struct sock *newsk,
 579                                             const struct request_sock *req)
 580 {
 581         const struct inet_request_sock *ireq = inet_rsk(req);
 582         struct net *net = read_pnet(&ireq->ireq_net);
 583         struct inet_sock *newinet = inet_sk(newsk);
 584         struct ip_options_rcu *opt;
 585         struct flowi4 *fl4;
 586         struct rtable *rt;
 587
 588         opt = rcu_dereference(ireq->ireq_opt);
 589         fl4 = &newinet->cork.fl.u.ip4;
 590
 591         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 592                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 593                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 594                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 595                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 596                            htons(ireq->ir_num), sk->sk_uid);
 597         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 598         rt = ip_route_output_flow(net, fl4, sk);
 599         if (IS_ERR(rt))
 600                 goto no_route;
 601         if (opt && opt->opt.is_strictroute && rt->rt_gw_family)
 602                 goto route_err;
 603         return &rt->dst;
 604
 605 route_err:
 606         ip_rt_put(rt);
 607 no_route:
 608         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 609         return NULL;
 610 }
 611 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 612
 613 #if IS_ENABLED(CONFIG_IPV6)
 614 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 615 #else
 616 #define AF_INET_FAMILY(fam) true
 617 #endif
 618
 619 /* Decide when to expire the request and when to resend SYN-ACK */
 620 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 621                                   const int max_retries,
 622                                   const u8 rskq_defer_accept,
 623                                   int *expire, int *resend)
 624 {
 625         if (!rskq_defer_accept) {
 626                 *expire = req->num_timeout >= thresh;
 627                 *resend = 1;
 628                 return;
 629         }
 630         *expire = req->num_timeout >= thresh &&
 631                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 632         /*
 633          * Do not resend while waiting for data after ACK,
 634          * start to resend on end of deferring period to give
 635          * last chance for data or ACK to create established socket.
 636          */
 637         *resend = !inet_rsk(req)->acked ||
 638                   req->num_timeout >= rskq_defer_accept - 1;
 639 }
 640
 641 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 642 {
 643         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 644
 645         if (!err)
 646                 req->num_retrans++;
 647         return err;
 648 }
 649 EXPORT_SYMBOL(inet_rtx_syn_ack);
 650
 651 /* return true if req was found in the ehash table */
 652 static bool reqsk_queue_unlink(struct request_sock *req)
 653 {
 654         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 655         bool found = false;
 656
 657         if (sk_hashed(req_to_sk(req))) {
 658                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 659
 660                 spin_lock(lock);
 661                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 662                 spin_unlock(lock);
 663         }
 664         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 665                 reqsk_put(req);
 666         return found;
 667 }
 668
 669 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 670 {
 671         if (reqsk_queue_unlink(req)) {
 672                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 673                 reqsk_put(req);
 674         }
 675 }
 676 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 677
 678 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 679 {
 680         inet_csk_reqsk_queue_drop(sk, req);
 681         reqsk_put(req);
 682 }
 683 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 684
 685 static void reqsk_timer_handler(struct timer_list *t)
 686 {
 687         struct request_sock *req = from_timer(req, t, rsk_timer);
 688         struct sock *sk_listener = req->rsk_listener;
 689         struct net *net = sock_net(sk_listener);
 690         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 691         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 692         int qlen, expire = 0, resend = 0;
 693         int max_retries, thresh;
 694         u8 defer_accept;
 695
 696         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 697                 goto drop;
 698
 699         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 700         thresh = max_retries;
 701         /* Normally all the openreqs are young and become mature
 702          * (i.e. converted to established socket) for first timeout.
 703          * If synack was not acknowledged for 1 second, it means
 704          * one of the following things: synack was lost, ack was lost,
 705          * rtt is high or nobody planned to ack (i.e. synflood).
 706          * When server is a bit loaded, queue is populated with old
 707          * open requests, reducing effective size of queue.
 708          * When server is well loaded, queue size reduces to zero
 709          * after several minutes of work. It is not synflood,
 710          * it is normal operation. The solution is pruning
 711          * too old entries overriding normal timeout, when
 712          * situation becomes dangerous.
 713          *
 714          * Essentially, we reserve half of room for young
 715          * embrions; and abort old ones without pity, if old
 716          * ones are about to clog our table.
 717          */
 718         qlen = reqsk_queue_len(queue);
 719         if ((qlen << 1) > max(8U, sk_listener->sk_max_ack_backlog)) {
 720                 int young = reqsk_queue_len_young(queue) << 1;
 721
 722                 while (thresh > 2) {
 723                         if (qlen < young)
 724                                 break;
 725                         thresh--;
 726                         young <<= 1;
 727                 }
 728         }
 729         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 730         if (defer_accept)
 731                 max_retries = defer_accept;
 732         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 733                        &expire, &resend);
 734         req->rsk_ops->syn_ack_timeout(req);
 735         if (!expire &&
 736             (!resend ||
 737              !inet_rtx_syn_ack(sk_listener, req) ||
 738              inet_rsk(req)->acked)) {
 739                 unsigned long timeo;
 740
 741                 if (req->num_timeout++ == 0)
 742                         atomic_dec(&queue->young);
 743                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 744                 mod_timer(&req->rsk_timer, jiffies + timeo);
 745                 return;
 746         }
 747 drop:
 748         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 749 }
 750
 751 static void reqsk_queue_hash_req(struct request_sock *req,
 752                                  unsigned long timeout)
 753 {
 754         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 755         mod_timer(&req->rsk_timer, jiffies + timeout);
 756
 757         inet_ehash_insert(req_to_sk(req), NULL);
 758         /* before letting lookups find us, make sure all req fields
 759          * are committed to memory and refcnt initialized.
 760          */
 761         smp_wmb();
 762         refcount_set(&req->rsk_refcnt, 2 + 1);
 763 }
 764
 765 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 766                                    unsigned long timeout)
 767 {
 768         reqsk_queue_hash_req(req, timeout);
 769         inet_csk_reqsk_queue_added(sk);
 770 }
 771 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 772
 773 /**
 774  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 775  *      @sk: the socket to clone
 776  *      @req: request_sock
 777  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 778  *
 779  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 780  */
 781 struct sock *inet_csk_clone_lock(const struct sock *sk,
 782                                  const struct request_sock *req,
 783                                  const gfp_t priority)
 784 {
 785         struct sock *newsk = sk_clone_lock(sk, priority);
 786
 787         if (newsk) {
 788                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 789
 790                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 791                 newicsk->icsk_bind_hash = NULL;
 792
 793                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 794                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 795                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 796
 797                 /* listeners have SOCK_RCU_FREE, not the children */
 798                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 799
 800                 inet_sk(newsk)->mc_list = NULL;
 801
 802                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 803                 atomic64_set(&newsk->sk_cookie,
 804                              atomic64_read(&inet_rsk(req)->ir_cookie));
 805
 806                 newicsk->icsk_retransmits = 0;
 807                 newicsk->icsk_backoff     = 0;
 808                 newicsk->icsk_probes_out  = 0;
 809
 810                 /* Deinitialize accept_queue to trap illegal accesses. */
 811                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 812
 813                 security_inet_csk_clone(newsk, req);
 814         }
 815         return newsk;
 816 }
 817 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 818
 819 /*
 820  * At this point, there should be no process reference to this
 821  * socket, and thus no user references at all.  Therefore we
 822  * can assume the socket waitqueue is inactive and nobody will
 823  * try to jump onto it.
 824  */
 825 void inet_csk_destroy_sock(struct sock *sk)
 826 {
 827         WARN_ON(sk->sk_state != TCP_CLOSE);
 828         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 829
 830         /* It cannot be in hash table! */
 831         WARN_ON(!sk_unhashed(sk));
 832
 833         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 834         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 835
 836         sk->sk_prot->destroy(sk);
 837
 838         sk_stream_kill_queues(sk);
 839
 840         xfrm_sk_free_policy(sk);
 841
 842         sk_refcnt_debug_release(sk);
 843
 844         percpu_counter_dec(sk->sk_prot->orphan_count);
 845
 846         sock_put(sk);
 847 }
 848 EXPORT_SYMBOL(inet_csk_destroy_sock);
 849
 850 /* This function allows to force a closure of a socket after the call to
 851  * tcp/dccp_create_openreq_child().
 852  */
 853 void inet_csk_prepare_forced_close(struct sock *sk)
 854         __releases(&sk->sk_lock.slock)
 855 {
 856         /* sk_clone_lock locked the socket and set refcnt to 2 */
 857         bh_unlock_sock(sk);
 858         sock_put(sk);
 859
 860         /* The below has to be done to allow calling inet_csk_destroy_sock */
 861         sock_set_flag(sk, SOCK_DEAD);
 862         percpu_counter_inc(sk->sk_prot->orphan_count);
 863         inet_sk(sk)->inet_num = 0;
 864 }
 865 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 866
 867 int inet_csk_listen_start(struct sock *sk, int backlog)
 868 {
 869         struct inet_connection_sock *icsk = inet_csk(sk);
 870         struct inet_sock *inet = inet_sk(sk);
 871         int err = -EADDRINUSE;
 872
 873         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 874
 875         sk->sk_ack_backlog = 0;
 876         inet_csk_delack_init(sk);
 877
 878         /* There is race window here: we announce ourselves listening,
 879          * but this transition is still not validated by get_port().
 880          * It is OK, because this socket enters to hash table only
 881          * after validation is complete.
 882          */
 883         inet_sk_state_store(sk, TCP_LISTEN);
 884         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 885                 inet->inet_sport = htons(inet->inet_num);
 886
 887                 sk_dst_reset(sk);
 888                 err = sk->sk_prot->hash(sk);
 889
 890                 if (likely(!err))
 891                         return 0;
 892         }
 893
 894         inet_sk_set_state(sk, TCP_CLOSE);
 895         return err;
 896 }
 897 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 898
 899 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 900                               struct sock *child)
 901 {
 902         sk->sk_prot->disconnect(child, O_NONBLOCK);
 903
 904         sock_orphan(child);
 905
 906         percpu_counter_inc(sk->sk_prot->orphan_count);
 907
 908         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 909                 BUG_ON(tcp_sk(child)->fastopen_rsk != req);
 910                 BUG_ON(sk != req->rsk_listener);
 911
 912                 /* Paranoid, to prevent race condition if
 913                  * an inbound pkt destined for child is
 914                  * blocked by sock lock in tcp_v4_rcv().
 915                  * Also to satisfy an assertion in
 916                  * tcp_v4_destroy_sock().
 917                  */
 918                 tcp_sk(child)->fastopen_rsk = NULL;
 919         }
 920         inet_csk_destroy_sock(child);
 921 }
 922
 923 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 924                                       struct request_sock *req,
 925                                       struct sock *child)
 926 {
 927         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 928
 929         spin_lock(&queue->rskq_lock);
 930         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 931                 inet_child_forget(sk, req, child);
 932                 child = NULL;
 933         } else {
 934                 req->sk = child;
 935                 req->dl_next = NULL;
 936                 if (queue->rskq_accept_head == NULL)
 937                         queue->rskq_accept_head = req;
 938                 else
 939                         queue->rskq_accept_tail->dl_next = req;
 940                 queue->rskq_accept_tail = req;
 941                 sk_acceptq_added(sk);
 942         }
 943         spin_unlock(&queue->rskq_lock);
 944         return child;
 945 }
 946 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 947
 948 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 949                                          struct request_sock *req, bool own_req)
 950 {
 951         if (own_req) {
 952                 inet_csk_reqsk_queue_drop(sk, req);
 953                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 954                 if (inet_csk_reqsk_queue_add(sk, req, child))
 955                         return child;
 956         }
 957         /* Too bad, another child took ownership of the request, undo. */
 958         bh_unlock_sock(child);
 959         sock_put(child);
 960         return NULL;
 961 }
 962 EXPORT_SYMBOL(inet_csk_complete_hashdance);
 963
 964 /*
 965  *      This routine closes sockets which have been at least partially
 966  *      opened, but not yet accepted.
 967  */
 968 void inet_csk_listen_stop(struct sock *sk)
 969 {
 970         struct inet_connection_sock *icsk = inet_csk(sk);
 971         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 972         struct request_sock *next, *req;
 973
 974         /* Following specs, it would be better either to send FIN
 975          * (and enter FIN-WAIT-1, it is normal close)
 976          * or to send active reset (abort).
 977          * Certainly, it is pretty dangerous while synflood, but it is
 978          * bad justification for our negligence 8)
 979          * To be honest, we are not able to make either
 980          * of the variants now.                 --ANK
 981          */
 982         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
 983                 struct sock *child = req->sk;
 984
 985                 local_bh_disable();
 986                 bh_lock_sock(child);
 987                 WARN_ON(sock_owned_by_user(child));
 988                 sock_hold(child);
 989
 990                 inet_child_forget(sk, req, child);
 991                 reqsk_put(req);
 992                 bh_unlock_sock(child);
 993                 local_bh_enable();
 994                 sock_put(child);
 995
 996                 cond_resched();
 997         }
 998         if (queue->fastopenq.rskq_rst_head) {
 999                 /* Free all the reqs queued in rskq_rst_head. */
1000                 spin_lock_bh(&queue->fastopenq.lock);
1001                 req = queue->fastopenq.rskq_rst_head;
1002                 queue->fastopenq.rskq_rst_head = NULL;
1003                 spin_unlock_bh(&queue->fastopenq.lock);
1004                 while (req != NULL) {
1005                         next = req->dl_next;
1006                         reqsk_put(req);
1007                         req = next;
1008                 }
1009         }
1010         WARN_ON_ONCE(sk->sk_ack_backlog);
1011 }
1012 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1013
1014 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1015 {
1016         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1017         const struct inet_sock *inet = inet_sk(sk);
1018
1019         sin->sin_family         = AF_INET;
1020         sin->sin_addr.s_addr    = inet->inet_daddr;
1021         sin->sin_port           = inet->inet_dport;
1022 }
1023 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1024
1025 #ifdef CONFIG_COMPAT
1026 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
1027                                char __user *optval, int __user *optlen)
1028 {
1029         const struct inet_connection_sock *icsk = inet_csk(sk);
1030
1031         if (icsk->icsk_af_ops->compat_getsockopt)
1032                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
1033                                                             optval, optlen);
1034         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
1035                                              optval, optlen);
1036 }
1037 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
1038
1039 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
1040                                char __user *optval, unsigned int optlen)
1041 {
1042         const struct inet_connection_sock *icsk = inet_csk(sk);
1043
1044         if (icsk->icsk_af_ops->compat_setsockopt)
1045                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
1046                                                             optval, optlen);
1047         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
1048                                              optval, optlen);
1049 }
1050 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
1051 #endif
1052
1053 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1054 {
1055         const struct inet_sock *inet = inet_sk(sk);
1056         const struct ip_options_rcu *inet_opt;
1057         __be32 daddr = inet->inet_daddr;
1058         struct flowi4 *fl4;
1059         struct rtable *rt;
1060
1061         rcu_read_lock();
1062         inet_opt = rcu_dereference(inet->inet_opt);
1063         if (inet_opt && inet_opt->opt.srr)
1064                 daddr = inet_opt->opt.faddr;
1065         fl4 = &fl->u.ip4;
1066         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1067                                    inet->inet_saddr, inet->inet_dport,
1068                                    inet->inet_sport, sk->sk_protocol,
1069                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1070         if (IS_ERR(rt))
1071                 rt = NULL;
1072         if (rt)
1073                 sk_setup_caps(sk, &rt->dst);
1074         rcu_read_unlock();
1075
1076         return &rt->dst;
1077 }
1078
1079 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1080 {
1081         struct dst_entry *dst = __sk_dst_check(sk, 0);
1082         struct inet_sock *inet = inet_sk(sk);
1083
1084         if (!dst) {
1085                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1086                 if (!dst)
1087                         goto out;
1088         }
1089         dst->ops->update_pmtu(dst, sk, NULL, mtu);
1090
1091         dst = __sk_dst_check(sk, 0);
1092         if (!dst)
1093                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1094 out:
1095         return dst;
1096 }
1097 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);