net/ipv4/inet_connection_sock.c

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