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 bool inet_rcv_saddr_any(const struct sock *sk)
 111 {
 112 #if IS_ENABLED(CONFIG_IPV6)
 113         if (sk->sk_family == AF_INET6)
 114                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 115 #endif
 116         return !sk->sk_rcv_saddr;
 117 }
 118
 119 void inet_get_local_port_range(struct net *net, int *low, int *high)
 120 {
 121         unsigned int seq;
 122
 123         do {
 124                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 125
 126                 *low = net->ipv4.ip_local_ports.range[0];
 127                 *high = net->ipv4.ip_local_ports.range[1];
 128         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 129 }
 130 EXPORT_SYMBOL(inet_get_local_port_range);
 131
 132 static int inet_csk_bind_conflict(const struct sock *sk,
 133                                   const struct inet_bind_bucket *tb,
 134                                   bool relax, bool reuseport_ok)
 135 {
 136         struct sock *sk2;
 137         bool reuse = sk->sk_reuse;
 138         bool reuseport = !!sk->sk_reuseport && reuseport_ok;
 139         kuid_t uid = sock_i_uid((struct sock *)sk);
 140
 141         /*
 142          * Unlike other sk lookup places we do not check
 143          * for sk_net here, since _all_ the socks listed
 144          * in tb->owners list belong to the same net - the
 145          * one this bucket belongs to.
 146          */
 147
 148         sk_for_each_bound(sk2, &tb->owners) {
 149                 if (sk != sk2 &&
 150                     (!sk->sk_bound_dev_if ||
 151                      !sk2->sk_bound_dev_if ||
 152                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 153                         if ((!reuse || !sk2->sk_reuse ||
 154                             sk2->sk_state == TCP_LISTEN) &&
 155                             (!reuseport || !sk2->sk_reuseport ||
 156                              rcu_access_pointer(sk->sk_reuseport_cb) ||
 157                              (sk2->sk_state != TCP_TIME_WAIT &&
 158                              !uid_eq(uid, sock_i_uid(sk2))))) {
 159                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 160                                         break;
 161                         }
 162                         if (!relax && reuse && sk2->sk_reuse &&
 163                             sk2->sk_state != TCP_LISTEN) {
 164                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 165                                         break;
 166                         }
 167                 }
 168         }
 169         return sk2 != NULL;
 170 }
 171
 172 /*
 173  * Find an open port number for the socket.  Returns with the
 174  * inet_bind_hashbucket lock held.
 175  */
 176 static struct inet_bind_hashbucket *
 177 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 178 {
 179         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 180         int port = 0;
 181         struct inet_bind_hashbucket *head;
 182         struct net *net = sock_net(sk);
 183         int i, low, high, attempt_half;
 184         struct inet_bind_bucket *tb;
 185         u32 remaining, offset;
 186
 187         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 188 other_half_scan:
 189         inet_get_local_port_range(net, &low, &high);
 190         high++; /* [32768, 60999] -> [32768, 61000[ */
 191         if (high - low < 4)
 192                 attempt_half = 0;
 193         if (attempt_half) {
 194                 int half = low + (((high - low) >> 2) << 1);
 195
 196                 if (attempt_half == 1)
 197                         high = half;
 198                 else
 199                         low = half;
 200         }
 201         remaining = high - low;
 202         if (likely(remaining > 1))
 203                 remaining &= ~1U;
 204
 205         offset = prandom_u32() % remaining;
 206         /* __inet_hash_connect() favors ports having @low parity
 207          * We do the opposite to not pollute connect() users.
 208          */
 209         offset |= 1U;
 210
 211 other_parity_scan:
 212         port = low + offset;
 213         for (i = 0; i < remaining; i += 2, port += 2) {
 214                 if (unlikely(port >= high))
 215                         port -= remaining;
 216                 if (inet_is_local_reserved_port(net, port))
 217                         continue;
 218                 head = &hinfo->bhash[inet_bhashfn(net, port,
 219                                                   hinfo->bhash_size)];
 220                 spin_lock_bh(&head->lock);
 221                 inet_bind_bucket_for_each(tb, &head->chain)
 222                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 223                                 if (!inet_csk_bind_conflict(sk, tb, false, false))
 224                                         goto success;
 225                                 goto next_port;
 226                         }
 227                 tb = NULL;
 228                 goto success;
 229 next_port:
 230                 spin_unlock_bh(&head->lock);
 231                 cond_resched();
 232         }
 233
 234         offset--;
 235         if (!(offset & 1))
 236                 goto other_parity_scan;
 237
 238         if (attempt_half == 1) {
 239                 /* OK we now try the upper half of the range */
 240                 attempt_half = 2;
 241                 goto other_half_scan;
 242         }
 243         return NULL;
 244 success:
 245         *port_ret = port;
 246         *tb_ret = tb;
 247         return head;
 248 }
 249
 250 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 251                                      struct sock *sk)
 252 {
 253         kuid_t uid = sock_i_uid(sk);
 254
 255         if (tb->fastreuseport <= 0)
 256                 return 0;
 257         if (!sk->sk_reuseport)
 258                 return 0;
 259         if (rcu_access_pointer(sk->sk_reuseport_cb))
 260                 return 0;
 261         if (!uid_eq(tb->fastuid, uid))
 262                 return 0;
 263         /* We only need to check the rcv_saddr if this tb was once marked
 264          * without fastreuseport and then was reset, as we can only know that
 265          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 266          * owners list.
 267          */
 268         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 269                 return 1;
 270 #if IS_ENABLED(CONFIG_IPV6)
 271         if (tb->fast_sk_family == AF_INET6)
 272                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 273                                             inet6_rcv_saddr(sk),
 274                                             tb->fast_rcv_saddr,
 275                                             sk->sk_rcv_saddr,
 276                                             tb->fast_ipv6_only,
 277                                             ipv6_only_sock(sk), true);
 278 #endif
 279         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 280                                     ipv6_only_sock(sk), true);
 281 }
 282
 283 /* Obtain a reference to a local port for the given sock,
 284  * if snum is zero it means select any available local port.
 285  * We try to allocate an odd port (and leave even ports for connect())
 286  */
 287 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 288 {
 289         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 290         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 291         int ret = 1, port = snum;
 292         struct inet_bind_hashbucket *head;
 293         struct net *net = sock_net(sk);
 294         struct inet_bind_bucket *tb = NULL;
 295         kuid_t uid = sock_i_uid(sk);
 296
 297         if (!port) {
 298                 head = inet_csk_find_open_port(sk, &tb, &port);
 299                 if (!head)
 300                         return ret;
 301                 if (!tb)
 302                         goto tb_not_found;
 303                 goto success;
 304         }
 305         head = &hinfo->bhash[inet_bhashfn(net, port,
 306                                           hinfo->bhash_size)];
 307         spin_lock_bh(&head->lock);
 308         inet_bind_bucket_for_each(tb, &head->chain)
 309                 if (net_eq(ib_net(tb), net) && tb->port == port)
 310                         goto tb_found;
 311 tb_not_found:
 312         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 313                                      net, head, port);
 314         if (!tb)
 315                 goto fail_unlock;
 316 tb_found:
 317         if (!hlist_empty(&tb->owners)) {
 318                 if (sk->sk_reuse == SK_FORCE_REUSE)
 319                         goto success;
 320
 321                 if ((tb->fastreuse > 0 && reuse) ||
 322                     sk_reuseport_match(tb, sk))
 323                         goto success;
 324                 if (inet_csk_bind_conflict(sk, tb, true, true))
 325                         goto fail_unlock;
 326         }
 327 success:
 328         if (hlist_empty(&tb->owners)) {
 329                 tb->fastreuse = reuse;
 330                 if (sk->sk_reuseport) {
 331                         tb->fastreuseport = FASTREUSEPORT_ANY;
 332                         tb->fastuid = uid;
 333                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 334                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 335                         tb->fast_sk_family = sk->sk_family;
 336 #if IS_ENABLED(CONFIG_IPV6)
 337                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 338 #endif
 339                 } else {
 340                         tb->fastreuseport = 0;
 341                 }
 342         } else {
 343                 if (!reuse)
 344                         tb->fastreuse = 0;
 345                 if (sk->sk_reuseport) {
 346                         /* We didn't match or we don't have fastreuseport set on
 347                          * the tb, but we have sk_reuseport set on this socket
 348                          * and we know that there are no bind conflicts with
 349                          * this socket in this tb, so reset our tb's reuseport
 350                          * settings so that any subsequent sockets that match
 351                          * our current socket will be put on the fast path.
 352                          *
 353                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 354                          * do extra checking for all subsequent sk_reuseport
 355                          * socks.
 356                          */
 357                         if (!sk_reuseport_match(tb, sk)) {
 358                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 359                                 tb->fastuid = uid;
 360                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 361                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 362                                 tb->fast_sk_family = sk->sk_family;
 363 #if IS_ENABLED(CONFIG_IPV6)
 364                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 365 #endif
 366                         }
 367                 } else {
 368                         tb->fastreuseport = 0;
 369                 }
 370         }
 371         if (!inet_csk(sk)->icsk_bind_hash)
 372                 inet_bind_hash(sk, tb, port);
 373         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 374         ret = 0;
 375
 376 fail_unlock:
 377         spin_unlock_bh(&head->lock);
 378         return ret;
 379 }
 380 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 381
 382 /*
 383  * Wait for an incoming connection, avoid race conditions. This must be called
 384  * with the socket locked.
 385  */
 386 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 387 {
 388         struct inet_connection_sock *icsk = inet_csk(sk);
 389         DEFINE_WAIT(wait);
 390         int err;
 391
 392         /*
 393          * True wake-one mechanism for incoming connections: only
 394          * one process gets woken up, not the 'whole herd'.
 395          * Since we do not 'race & poll' for established sockets
 396          * anymore, the common case will execute the loop only once.
 397          *
 398          * Subtle issue: "add_wait_queue_exclusive()" will be added
 399          * after any current non-exclusive waiters, and we know that
 400          * it will always _stay_ after any new non-exclusive waiters
 401          * because all non-exclusive waiters are added at the
 402          * beginning of the wait-queue. As such, it's ok to "drop"
 403          * our exclusiveness temporarily when we get woken up without
 404          * having to remove and re-insert us on the wait queue.
 405          */
 406         for (;;) {
 407                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 408                                           TASK_INTERRUPTIBLE);
 409                 release_sock(sk);
 410                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 411                         timeo = schedule_timeout(timeo);
 412                 sched_annotate_sleep();
 413                 lock_sock(sk);
 414                 err = 0;
 415                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 416                         break;
 417                 err = -EINVAL;
 418                 if (sk->sk_state != TCP_LISTEN)
 419                         break;
 420                 err = sock_intr_errno(timeo);
 421                 if (signal_pending(current))
 422                         break;
 423                 err = -EAGAIN;
 424                 if (!timeo)
 425                         break;
 426         }
 427         finish_wait(sk_sleep(sk), &wait);
 428         return err;
 429 }
 430
 431 /*
 432  * This will accept the next outstanding connection.
 433  */
 434 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 435 {
 436         struct inet_connection_sock *icsk = inet_csk(sk);
 437         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 438         struct request_sock *req;
 439         struct sock *newsk;
 440         int error;
 441
 442         lock_sock(sk);
 443
 444         /* We need to make sure that this socket is listening,
 445          * and that it has something pending.
 446          */
 447         error = -EINVAL;
 448         if (sk->sk_state != TCP_LISTEN)
 449                 goto out_err;
 450
 451         /* Find already established connection */
 452         if (reqsk_queue_empty(queue)) {
 453                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 454
 455                 /* If this is a non blocking socket don't sleep */
 456                 error = -EAGAIN;
 457                 if (!timeo)
 458                         goto out_err;
 459
 460                 error = inet_csk_wait_for_connect(sk, timeo);
 461                 if (error)
 462                         goto out_err;
 463         }
 464         req = reqsk_queue_remove(queue, sk);
 465         newsk = req->sk;
 466
 467         if (sk->sk_protocol == IPPROTO_TCP &&
 468             tcp_rsk(req)->tfo_listener) {
 469                 spin_lock_bh(&queue->fastopenq.lock);
 470                 if (tcp_rsk(req)->tfo_listener) {
 471                         /* We are still waiting for the final ACK from 3WHS
 472                          * so can't free req now. Instead, we set req->sk to
 473                          * NULL to signify that the child socket is taken
 474                          * so reqsk_fastopen_remove() will free the req
 475                          * when 3WHS finishes (or is aborted).
 476                          */
 477                         req->sk = NULL;
 478                         req = NULL;
 479                 }
 480                 spin_unlock_bh(&queue->fastopenq.lock);
 481         }
 482 out:
 483         release_sock(sk);
 484         if (req)
 485                 reqsk_put(req);
 486         return newsk;
 487 out_err:
 488         newsk = NULL;
 489         req = NULL;
 490         *err = error;
 491         goto out;
 492 }
 493 EXPORT_SYMBOL(inet_csk_accept);
 494
 495 /*
 496  * Using different timers for retransmit, delayed acks and probes
 497  * We may wish use just one timer maintaining a list of expire jiffies
 498  * to optimize.
 499  */
 500 void inet_csk_init_xmit_timers(struct sock *sk,
 501                                void (*retransmit_handler)(struct timer_list *t),
 502                                void (*delack_handler)(struct timer_list *t),
 503                                void (*keepalive_handler)(struct timer_list *t))
 504 {
 505         struct inet_connection_sock *icsk = inet_csk(sk);
 506
 507         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 508         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 509         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 510         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 511 }
 512 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 513
 514 void inet_csk_clear_xmit_timers(struct sock *sk)
 515 {
 516         struct inet_connection_sock *icsk = inet_csk(sk);
 517
 518         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 519
 520         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 521         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 522         sk_stop_timer(sk, &sk->sk_timer);
 523 }
 524 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 525
 526 void inet_csk_delete_keepalive_timer(struct sock *sk)
 527 {
 528         sk_stop_timer(sk, &sk->sk_timer);
 529 }
 530 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 531
 532 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 533 {
 534         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 535 }
 536 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 537
 538 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 539                                      struct flowi4 *fl4,
 540                                      const struct request_sock *req)
 541 {
 542         const struct inet_request_sock *ireq = inet_rsk(req);
 543         struct net *net = read_pnet(&ireq->ireq_net);
 544         struct ip_options_rcu *opt;
 545         struct rtable *rt;
 546
 547         opt = ireq_opt_deref(ireq);
 548
 549         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 550                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 551                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 552                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 553                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 554                            htons(ireq->ir_num), sk->sk_uid);
 555         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 556         rt = ip_route_output_flow(net, fl4, sk);
 557         if (IS_ERR(rt))
 558                 goto no_route;
 559         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 560                 goto route_err;
 561         return &rt->dst;
 562
 563 route_err:
 564         ip_rt_put(rt);
 565 no_route:
 566         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 567         return NULL;
 568 }
 569 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 570
 571 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 572                                             struct sock *newsk,
 573                                             const struct request_sock *req)
 574 {
 575         const struct inet_request_sock *ireq = inet_rsk(req);
 576         struct net *net = read_pnet(&ireq->ireq_net);
 577         struct inet_sock *newinet = inet_sk(newsk);
 578         struct ip_options_rcu *opt;
 579         struct flowi4 *fl4;
 580         struct rtable *rt;
 581
 582         opt = rcu_dereference(ireq->ireq_opt);
 583         fl4 = &newinet->cork.fl.u.ip4;
 584
 585         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 586                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 587                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 588                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 589                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 590                            htons(ireq->ir_num), sk->sk_uid);
 591         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 592         rt = ip_route_output_flow(net, fl4, sk);
 593         if (IS_ERR(rt))
 594                 goto no_route;
 595         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 596                 goto route_err;
 597         return &rt->dst;
 598
 599 route_err:
 600         ip_rt_put(rt);
 601 no_route:
 602         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 603         return NULL;
 604 }
 605 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 606
 607 #if IS_ENABLED(CONFIG_IPV6)
 608 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 609 #else
 610 #define AF_INET_FAMILY(fam) true
 611 #endif
 612
 613 /* Decide when to expire the request and when to resend SYN-ACK */
 614 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 615                                   const int max_retries,
 616                                   const u8 rskq_defer_accept,
 617                                   int *expire, int *resend)
 618 {
 619         if (!rskq_defer_accept) {
 620                 *expire = req->num_timeout >= thresh;
 621                 *resend = 1;
 622                 return;
 623         }
 624         *expire = req->num_timeout >= thresh &&
 625                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 626         /*
 627          * Do not resend while waiting for data after ACK,
 628          * start to resend on end of deferring period to give
 629          * last chance for data or ACK to create established socket.
 630          */
 631         *resend = !inet_rsk(req)->acked ||
 632                   req->num_timeout >= rskq_defer_accept - 1;
 633 }
 634
 635 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 636 {
 637         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 638
 639         if (!err)
 640                 req->num_retrans++;
 641         return err;
 642 }
 643 EXPORT_SYMBOL(inet_rtx_syn_ack);
 644
 645 /* return true if req was found in the ehash table */
 646 static bool reqsk_queue_unlink(struct request_sock_queue *queue,
 647                                struct request_sock *req)
 648 {
 649         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 650         bool found = false;
 651
 652         if (sk_hashed(req_to_sk(req))) {
 653                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 654
 655                 spin_lock(lock);
 656                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 657                 spin_unlock(lock);
 658         }
 659         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 660                 reqsk_put(req);
 661         return found;
 662 }
 663
 664 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 665 {
 666         if (reqsk_queue_unlink(&inet_csk(sk)->icsk_accept_queue, req)) {
 667                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 668                 reqsk_put(req);
 669         }
 670 }
 671 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 672
 673 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 674 {
 675         inet_csk_reqsk_queue_drop(sk, req);
 676         reqsk_put(req);
 677 }
 678 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 679
 680 static void reqsk_timer_handler(struct timer_list *t)
 681 {
 682         struct request_sock *req = from_timer(req, t, rsk_timer);
 683         struct sock *sk_listener = req->rsk_listener;
 684         struct net *net = sock_net(sk_listener);
 685         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 686         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 687         int qlen, expire = 0, resend = 0;
 688         int max_retries, thresh;
 689         u8 defer_accept;
 690
 691         if (inet_sk_state_load(sk_listener) != TCP_LISTEN)
 692                 goto drop;
 693
 694         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 695         thresh = max_retries;
 696         /* Normally all the openreqs are young and become mature
 697          * (i.e. converted to established socket) for first timeout.
 698          * If synack was not acknowledged for 1 second, it means
 699          * one of the following things: synack was lost, ack was lost,
 700          * rtt is high or nobody planned to ack (i.e. synflood).
 701          * When server is a bit loaded, queue is populated with old
 702          * open requests, reducing effective size of queue.
 703          * When server is well loaded, queue size reduces to zero
 704          * after several minutes of work. It is not synflood,
 705          * it is normal operation. The solution is pruning
 706          * too old entries overriding normal timeout, when
 707          * situation becomes dangerous.
 708          *
 709          * Essentially, we reserve half of room for young
 710          * embrions; and abort old ones without pity, if old
 711          * ones are about to clog our table.
 712          */
 713         qlen = reqsk_queue_len(queue);
 714         if ((qlen << 1) > max(8U, sk_listener->sk_max_ack_backlog)) {
 715                 int young = reqsk_queue_len_young(queue) << 1;
 716
 717                 while (thresh > 2) {
 718                         if (qlen < young)
 719                                 break;
 720                         thresh--;
 721                         young <<= 1;
 722                 }
 723         }
 724         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 725         if (defer_accept)
 726                 max_retries = defer_accept;
 727         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 728                        &expire, &resend);
 729         req->rsk_ops->syn_ack_timeout(req);
 730         if (!expire &&
 731             (!resend ||
 732              !inet_rtx_syn_ack(sk_listener, req) ||
 733              inet_rsk(req)->acked)) {
 734                 unsigned long timeo;
 735
 736                 if (req->num_timeout++ == 0)
 737                         atomic_dec(&queue->young);
 738                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 739                 mod_timer(&req->rsk_timer, jiffies + timeo);
 740                 return;
 741         }
 742 drop:
 743         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 744 }
 745
 746 static void reqsk_queue_hash_req(struct request_sock *req,
 747                                  unsigned long timeout)
 748 {
 749         req->num_retrans = 0;
 750         req->num_timeout = 0;
 751         req->sk = NULL;
 752
 753         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 754         mod_timer(&req->rsk_timer, jiffies + timeout);
 755
 756         inet_ehash_insert(req_to_sk(req), NULL);
 757         /* before letting lookups find us, make sure all req fields
 758          * are committed to memory and refcnt initialized.
 759          */
 760         smp_wmb();
 761         refcount_set(&req->rsk_refcnt, 2 + 1);
 762 }
 763
 764 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 765                                    unsigned long timeout)
 766 {
 767         reqsk_queue_hash_req(req, timeout);
 768         inet_csk_reqsk_queue_added(sk);
 769 }
 770 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 771
 772 /**
 773  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 774  *      @sk: the socket to clone
 775  *      @req: request_sock
 776  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 777  *
 778  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 779  */
 780 struct sock *inet_csk_clone_lock(const struct sock *sk,
 781                                  const struct request_sock *req,
 782                                  const gfp_t priority)
 783 {
 784         struct sock *newsk = sk_clone_lock(sk, priority);
 785
 786         if (newsk) {
 787                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 788
 789                 inet_sk_set_state(newsk, TCP_SYN_RECV);
 790                 newicsk->icsk_bind_hash = NULL;
 791
 792                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 793                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 794                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 795
 796                 /* listeners have SOCK_RCU_FREE, not the children */
 797                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 798
 799                 inet_sk(newsk)->mc_list = NULL;
 800
 801                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 802                 atomic64_set(&newsk->sk_cookie,
 803                              atomic64_read(&inet_rsk(req)->ir_cookie));
 804
 805                 newicsk->icsk_retransmits = 0;
 806                 newicsk->icsk_backoff     = 0;
 807                 newicsk->icsk_probes_out  = 0;
 808
 809                 /* Deinitialize accept_queue to trap illegal accesses. */
 810                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 811
 812                 security_inet_csk_clone(newsk, req);
 813         }
 814         return newsk;
 815 }
 816 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 817
 818 /*
 819  * At this point, there should be no process reference to this
 820  * socket, and thus no user references at all.  Therefore we
 821  * can assume the socket waitqueue is inactive and nobody will
 822  * try to jump onto it.
 823  */
 824 void inet_csk_destroy_sock(struct sock *sk)
 825 {
 826         WARN_ON(sk->sk_state != TCP_CLOSE);
 827         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 828
 829         /* It cannot be in hash table! */
 830         WARN_ON(!sk_unhashed(sk));
 831
 832         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 833         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 834
 835         sk->sk_prot->destroy(sk);
 836
 837         sk_stream_kill_queues(sk);
 838
 839         xfrm_sk_free_policy(sk);
 840
 841         sk_refcnt_debug_release(sk);
 842
 843         percpu_counter_dec(sk->sk_prot->orphan_count);
 844
 845         sock_put(sk);
 846 }
 847 EXPORT_SYMBOL(inet_csk_destroy_sock);
 848
 849 /* This function allows to force a closure of a socket after the call to
 850  * tcp/dccp_create_openreq_child().
 851  */
 852 void inet_csk_prepare_forced_close(struct sock *sk)
 853         __releases(&sk->sk_lock.slock)
 854 {
 855         /* sk_clone_lock locked the socket and set refcnt to 2 */
 856         bh_unlock_sock(sk);
 857         sock_put(sk);
 858
 859         /* The below has to be done to allow calling inet_csk_destroy_sock */
 860         sock_set_flag(sk, SOCK_DEAD);
 861         percpu_counter_inc(sk->sk_prot->orphan_count);
 862         inet_sk(sk)->inet_num = 0;
 863 }
 864 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 865
 866 int inet_csk_listen_start(struct sock *sk, int backlog)
 867 {
 868         struct inet_connection_sock *icsk = inet_csk(sk);
 869         struct inet_sock *inet = inet_sk(sk);
 870         int err = -EADDRINUSE;
 871
 872         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 873
 874         sk->sk_max_ack_backlog = backlog;
 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);