net/ipv4/inet_connection_sock.c

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