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
  29 #ifdef INET_CSK_DEBUG
  30 const char inet_csk_timer_bug_msg[] = "inet_csk BUG: unknown timer value\n";
  31 EXPORT_SYMBOL(inet_csk_timer_bug_msg);
  32 #endif
  33
  34 #if IS_ENABLED(CONFIG_IPV6)
  35 /* match_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses if IPv6
  36  *                          only, and any IPv4 addresses if not IPv6 only
  37  * match_wildcard == false: addresses must be exactly the same, i.e.
  38  *                          IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  39  *                          and 0.0.0.0 equals to 0.0.0.0 only
  40  */
  41 static int ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  42                                 const struct in6_addr *sk2_rcv_saddr6,
  43                                 __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  44                                 bool sk1_ipv6only, bool sk2_ipv6only,
  45                                 bool match_wildcard)
  46 {
  47         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  48         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  49
  50         /* if both are mapped, treat as IPv4 */
  51         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  52                 if (!sk2_ipv6only) {
  53                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  54                                 return 1;
  55                         if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  56                                 return match_wildcard;
  57                 }
  58                 return 0;
  59         }
  60
  61         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  62                 return 1;
  63
  64         if (addr_type2 == IPV6_ADDR_ANY && match_wildcard &&
  65             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  66                 return 1;
  67
  68         if (addr_type == IPV6_ADDR_ANY && match_wildcard &&
  69             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  70                 return 1;
  71
  72         if (sk2_rcv_saddr6 &&
  73             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  74                 return 1;
  75
  76         return 0;
  77 }
  78 #endif
  79
  80 /* match_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  81  * match_wildcard == false: addresses must be exactly the same, i.e.
  82  *                          0.0.0.0 only equals to 0.0.0.0
  83  */
  84 static int ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  85                                 bool sk2_ipv6only, bool match_wildcard)
  86 {
  87         if (!sk2_ipv6only) {
  88                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  89                         return 1;
  90                 if (!sk1_rcv_saddr || !sk2_rcv_saddr)
  91                         return match_wildcard;
  92         }
  93         return 0;
  94 }
  95
  96 int inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  97                          bool match_wildcard)
  98 {
  99 #if IS_ENABLED(CONFIG_IPV6)
 100         if (sk->sk_family == AF_INET6)
 101                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
 102                                             inet6_rcv_saddr(sk2),
 103                                             sk->sk_rcv_saddr,
 104                                             sk2->sk_rcv_saddr,
 105                                             ipv6_only_sock(sk),
 106                                             ipv6_only_sock(sk2),
 107                                             match_wildcard);
 108 #endif
 109         return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr,
 110                                     ipv6_only_sock(sk2), match_wildcard);
 111 }
 112 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 113
 114 void inet_get_local_port_range(struct net *net, int *low, int *high)
 115 {
 116         unsigned int seq;
 117
 118         do {
 119                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 120
 121                 *low = net->ipv4.ip_local_ports.range[0];
 122                 *high = net->ipv4.ip_local_ports.range[1];
 123         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 124 }
 125 EXPORT_SYMBOL(inet_get_local_port_range);
 126
 127 static int inet_csk_bind_conflict(const struct sock *sk,
 128                                   const struct inet_bind_bucket *tb,
 129                                   bool relax, bool reuseport_ok)
 130 {
 131         struct sock *sk2;
 132         bool reuse = sk->sk_reuse;
 133         bool reuseport = !!sk->sk_reuseport && reuseport_ok;
 134         kuid_t uid = sock_i_uid((struct sock *)sk);
 135
 136         /*
 137          * Unlike other sk lookup places we do not check
 138          * for sk_net here, since _all_ the socks listed
 139          * in tb->owners list belong to the same net - the
 140          * one this bucket belongs to.
 141          */
 142
 143         sk_for_each_bound(sk2, &tb->owners) {
 144                 if (sk != sk2 &&
 145                     (!sk->sk_bound_dev_if ||
 146                      !sk2->sk_bound_dev_if ||
 147                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 148                         if ((!reuse || !sk2->sk_reuse ||
 149                             sk2->sk_state == TCP_LISTEN) &&
 150                             (!reuseport || !sk2->sk_reuseport ||
 151                              rcu_access_pointer(sk->sk_reuseport_cb) ||
 152                              (sk2->sk_state != TCP_TIME_WAIT &&
 153                              !uid_eq(uid, sock_i_uid(sk2))))) {
 154                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 155                                         break;
 156                         }
 157                         if (!relax && reuse && sk2->sk_reuse &&
 158                             sk2->sk_state != TCP_LISTEN) {
 159                                 if (inet_rcv_saddr_equal(sk, sk2, true))
 160                                         break;
 161                         }
 162                 }
 163         }
 164         return sk2 != NULL;
 165 }
 166
 167 /*
 168  * Find an open port number for the socket.  Returns with the
 169  * inet_bind_hashbucket lock held.
 170  */
 171 static struct inet_bind_hashbucket *
 172 inet_csk_find_open_port(struct sock *sk, struct inet_bind_bucket **tb_ret, int *port_ret)
 173 {
 174         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 175         int port = 0;
 176         struct inet_bind_hashbucket *head;
 177         struct net *net = sock_net(sk);
 178         int i, low, high, attempt_half;
 179         struct inet_bind_bucket *tb;
 180         u32 remaining, offset;
 181
 182         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 183 other_half_scan:
 184         inet_get_local_port_range(net, &low, &high);
 185         high++; /* [32768, 60999] -> [32768, 61000[ */
 186         if (high - low < 4)
 187                 attempt_half = 0;
 188         if (attempt_half) {
 189                 int half = low + (((high - low) >> 2) << 1);
 190
 191                 if (attempt_half == 1)
 192                         high = half;
 193                 else
 194                         low = half;
 195         }
 196         remaining = high - low;
 197         if (likely(remaining > 1))
 198                 remaining &= ~1U;
 199
 200         offset = prandom_u32() % remaining;
 201         /* __inet_hash_connect() favors ports having @low parity
 202          * We do the opposite to not pollute connect() users.
 203          */
 204         offset |= 1U;
 205
 206 other_parity_scan:
 207         port = low + offset;
 208         for (i = 0; i < remaining; i += 2, port += 2) {
 209                 if (unlikely(port >= high))
 210                         port -= remaining;
 211                 if (inet_is_local_reserved_port(net, port))
 212                         continue;
 213                 head = &hinfo->bhash[inet_bhashfn(net, port,
 214                                                   hinfo->bhash_size)];
 215                 spin_lock_bh(&head->lock);
 216                 inet_bind_bucket_for_each(tb, &head->chain)
 217                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 218                                 if (!inet_csk_bind_conflict(sk, tb, false, false))
 219                                         goto success;
 220                                 goto next_port;
 221                         }
 222                 tb = NULL;
 223                 goto success;
 224 next_port:
 225                 spin_unlock_bh(&head->lock);
 226                 cond_resched();
 227         }
 228
 229         offset--;
 230         if (!(offset & 1))
 231                 goto other_parity_scan;
 232
 233         if (attempt_half == 1) {
 234                 /* OK we now try the upper half of the range */
 235                 attempt_half = 2;
 236                 goto other_half_scan;
 237         }
 238         return NULL;
 239 success:
 240         *port_ret = port;
 241         *tb_ret = tb;
 242         return head;
 243 }
 244
 245 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 246                                      struct sock *sk)
 247 {
 248         kuid_t uid = sock_i_uid(sk);
 249
 250         if (tb->fastreuseport <= 0)
 251                 return 0;
 252         if (!sk->sk_reuseport)
 253                 return 0;
 254         if (rcu_access_pointer(sk->sk_reuseport_cb))
 255                 return 0;
 256         if (!uid_eq(tb->fastuid, uid))
 257                 return 0;
 258         /* We only need to check the rcv_saddr if this tb was once marked
 259          * without fastreuseport and then was reset, as we can only know that
 260          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 261          * owners list.
 262          */
 263         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 264                 return 1;
 265 #if IS_ENABLED(CONFIG_IPV6)
 266         if (tb->fast_sk_family == AF_INET6)
 267                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 268                                             &sk->sk_v6_rcv_saddr,
 269                                             tb->fast_rcv_saddr,
 270                                             sk->sk_rcv_saddr,
 271                                             tb->fast_ipv6_only,
 272                                             ipv6_only_sock(sk), true);
 273 #endif
 274         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 275                                     ipv6_only_sock(sk), true);
 276 }
 277
 278 /* Obtain a reference to a local port for the given sock,
 279  * if snum is zero it means select any available local port.
 280  * We try to allocate an odd port (and leave even ports for connect())
 281  */
 282 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 283 {
 284         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 285         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 286         int ret = 1, port = snum;
 287         struct inet_bind_hashbucket *head;
 288         struct net *net = sock_net(sk);
 289         struct inet_bind_bucket *tb = NULL;
 290         kuid_t uid = sock_i_uid(sk);
 291
 292         if (!port) {
 293                 head = inet_csk_find_open_port(sk, &tb, &port);
 294                 if (!head)
 295                         return ret;
 296                 if (!tb)
 297                         goto tb_not_found;
 298                 goto success;
 299         }
 300         head = &hinfo->bhash[inet_bhashfn(net, port,
 301                                           hinfo->bhash_size)];
 302         spin_lock_bh(&head->lock);
 303         inet_bind_bucket_for_each(tb, &head->chain)
 304                 if (net_eq(ib_net(tb), net) && tb->port == port)
 305                         goto tb_found;
 306 tb_not_found:
 307         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 308                                      net, head, port);
 309         if (!tb)
 310                 goto fail_unlock;
 311 tb_found:
 312         if (!hlist_empty(&tb->owners)) {
 313                 if (sk->sk_reuse == SK_FORCE_REUSE)
 314                         goto success;
 315
 316                 if ((tb->fastreuse > 0 && reuse) ||
 317                     sk_reuseport_match(tb, sk))
 318                         goto success;
 319                 if (inet_csk_bind_conflict(sk, tb, true, true))
 320                         goto fail_unlock;
 321         }
 322 success:
 323         if (!hlist_empty(&tb->owners)) {
 324                 tb->fastreuse = reuse;
 325                 if (sk->sk_reuseport) {
 326                         tb->fastreuseport = FASTREUSEPORT_ANY;
 327                         tb->fastuid = uid;
 328                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 329                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 330 #if IS_ENABLED(CONFIG_IPV6)
 331                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 332 #endif
 333                 } else {
 334                         tb->fastreuseport = 0;
 335                 }
 336         } else {
 337                 if (!reuse)
 338                         tb->fastreuse = 0;
 339                 if (sk->sk_reuseport) {
 340                         /* We didn't match or we don't have fastreuseport set on
 341                          * the tb, but we have sk_reuseport set on this socket
 342                          * and we know that there are no bind conflicts with
 343                          * this socket in this tb, so reset our tb's reuseport
 344                          * settings so that any subsequent sockets that match
 345                          * our current socket will be put on the fast path.
 346                          *
 347                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 348                          * do extra checking for all subsequent sk_reuseport
 349                          * socks.
 350                          */
 351                         if (!sk_reuseport_match(tb, sk)) {
 352                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 353                                 tb->fastuid = uid;
 354                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 355                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 356 #if IS_ENABLED(CONFIG_IPV6)
 357                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 358 #endif
 359                         }
 360                 } else {
 361                         tb->fastreuseport = 0;
 362                 }
 363         }
 364         if (!inet_csk(sk)->icsk_bind_hash)
 365                 inet_bind_hash(sk, tb, port);
 366         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 367         ret = 0;
 368
 369 fail_unlock:
 370         spin_unlock_bh(&head->lock);
 371         return ret;
 372 }
 373 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 374
 375 /*
 376  * Wait for an incoming connection, avoid race conditions. This must be called
 377  * with the socket locked.
 378  */
 379 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 380 {
 381         struct inet_connection_sock *icsk = inet_csk(sk);
 382         DEFINE_WAIT(wait);
 383         int err;
 384
 385         /*
 386          * True wake-one mechanism for incoming connections: only
 387          * one process gets woken up, not the 'whole herd'.
 388          * Since we do not 'race & poll' for established sockets
 389          * anymore, the common case will execute the loop only once.
 390          *
 391          * Subtle issue: "add_wait_queue_exclusive()" will be added
 392          * after any current non-exclusive waiters, and we know that
 393          * it will always _stay_ after any new non-exclusive waiters
 394          * because all non-exclusive waiters are added at the
 395          * beginning of the wait-queue. As such, it's ok to "drop"
 396          * our exclusiveness temporarily when we get woken up without
 397          * having to remove and re-insert us on the wait queue.
 398          */
 399         for (;;) {
 400                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 401                                           TASK_INTERRUPTIBLE);
 402                 release_sock(sk);
 403                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 404                         timeo = schedule_timeout(timeo);
 405                 sched_annotate_sleep();
 406                 lock_sock(sk);
 407                 err = 0;
 408                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 409                         break;
 410                 err = -EINVAL;
 411                 if (sk->sk_state != TCP_LISTEN)
 412                         break;
 413                 err = sock_intr_errno(timeo);
 414                 if (signal_pending(current))
 415                         break;
 416                 err = -EAGAIN;
 417                 if (!timeo)
 418                         break;
 419         }
 420         finish_wait(sk_sleep(sk), &wait);
 421         return err;
 422 }
 423
 424 /*
 425  * This will accept the next outstanding connection.
 426  */
 427 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 428 {
 429         struct inet_connection_sock *icsk = inet_csk(sk);
 430         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 431         struct request_sock *req;
 432         struct sock *newsk;
 433         int error;
 434
 435         lock_sock(sk);
 436
 437         /* We need to make sure that this socket is listening,
 438          * and that it has something pending.
 439          */
 440         error = -EINVAL;
 441         if (sk->sk_state != TCP_LISTEN)
 442                 goto out_err;
 443
 444         /* Find already established connection */
 445         if (reqsk_queue_empty(queue)) {
 446                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 447
 448                 /* If this is a non blocking socket don't sleep */
 449                 error = -EAGAIN;
 450                 if (!timeo)
 451                         goto out_err;
 452
 453                 error = inet_csk_wait_for_connect(sk, timeo);
 454                 if (error)
 455                         goto out_err;
 456         }
 457         req = reqsk_queue_remove(queue, sk);
 458         newsk = req->sk;
 459
 460         if (sk->sk_protocol == IPPROTO_TCP &&
 461             tcp_rsk(req)->tfo_listener) {
 462                 spin_lock_bh(&queue->fastopenq.lock);
 463                 if (tcp_rsk(req)->tfo_listener) {
 464                         /* We are still waiting for the final ACK from 3WHS
 465                          * so can't free req now. Instead, we set req->sk to
 466                          * NULL to signify that the child socket is taken
 467                          * so reqsk_fastopen_remove() will free the req
 468                          * when 3WHS finishes (or is aborted).
 469                          */
 470                         req->sk = NULL;
 471                         req = NULL;
 472                 }
 473                 spin_unlock_bh(&queue->fastopenq.lock);
 474         }
 475 out:
 476         release_sock(sk);
 477         if (req)
 478                 reqsk_put(req);
 479         return newsk;
 480 out_err:
 481         newsk = NULL;
 482         req = NULL;
 483         *err = error;
 484         goto out;
 485 }
 486 EXPORT_SYMBOL(inet_csk_accept);
 487
 488 /*
 489  * Using different timers for retransmit, delayed acks and probes
 490  * We may wish use just one timer maintaining a list of expire jiffies
 491  * to optimize.
 492  */
 493 void inet_csk_init_xmit_timers(struct sock *sk,
 494                                void (*retransmit_handler)(unsigned long),
 495                                void (*delack_handler)(unsigned long),
 496                                void (*keepalive_handler)(unsigned long))
 497 {
 498         struct inet_connection_sock *icsk = inet_csk(sk);
 499
 500         setup_timer(&icsk->icsk_retransmit_timer, retransmit_handler,
 501                         (unsigned long)sk);
 502         setup_timer(&icsk->icsk_delack_timer, delack_handler,
 503                         (unsigned long)sk);
 504         setup_timer(&sk->sk_timer, keepalive_handler, (unsigned long)sk);
 505         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 506 }
 507 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 508
 509 void inet_csk_clear_xmit_timers(struct sock *sk)
 510 {
 511         struct inet_connection_sock *icsk = inet_csk(sk);
 512
 513         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 514
 515         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 516         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 517         sk_stop_timer(sk, &sk->sk_timer);
 518 }
 519 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 520
 521 void inet_csk_delete_keepalive_timer(struct sock *sk)
 522 {
 523         sk_stop_timer(sk, &sk->sk_timer);
 524 }
 525 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 526
 527 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 528 {
 529         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 530 }
 531 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 532
 533 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 534                                      struct flowi4 *fl4,
 535                                      const struct request_sock *req)
 536 {
 537         const struct inet_request_sock *ireq = inet_rsk(req);
 538         struct net *net = read_pnet(&ireq->ireq_net);
 539         struct ip_options_rcu *opt = ireq->opt;
 540         struct rtable *rt;
 541
 542         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 543                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 544                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 545                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 546                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 547                            htons(ireq->ir_num), sk->sk_uid);
 548         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 549         rt = ip_route_output_flow(net, fl4, sk);
 550         if (IS_ERR(rt))
 551                 goto no_route;
 552         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 553                 goto route_err;
 554         return &rt->dst;
 555
 556 route_err:
 557         ip_rt_put(rt);
 558 no_route:
 559         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 560         return NULL;
 561 }
 562 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 563
 564 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 565                                             struct sock *newsk,
 566                                             const struct request_sock *req)
 567 {
 568         const struct inet_request_sock *ireq = inet_rsk(req);
 569         struct net *net = read_pnet(&ireq->ireq_net);
 570         struct inet_sock *newinet = inet_sk(newsk);
 571         struct ip_options_rcu *opt;
 572         struct flowi4 *fl4;
 573         struct rtable *rt;
 574
 575         fl4 = &newinet->cork.fl.u.ip4;
 576
 577         rcu_read_lock();
 578         opt = rcu_dereference(newinet->inet_opt);
 579         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 580                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 581                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 582                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 583                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 584                            htons(ireq->ir_num), sk->sk_uid);
 585         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 586         rt = ip_route_output_flow(net, fl4, sk);
 587         if (IS_ERR(rt))
 588                 goto no_route;
 589         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 590                 goto route_err;
 591         rcu_read_unlock();
 592         return &rt->dst;
 593
 594 route_err:
 595         ip_rt_put(rt);
 596 no_route:
 597         rcu_read_unlock();
 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(unsigned long data)
 677 {
 678         struct request_sock *req = (struct request_sock *)data;
 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 (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         setup_pinned_timer(&req->rsk_timer, reqsk_timer_handler,
 750                             (unsigned long)req);
 751         mod_timer(&req->rsk_timer, jiffies + timeout);
 752
 753         inet_ehash_insert(req_to_sk(req), NULL);
 754         /* before letting lookups find us, make sure all req fields
 755          * are committed to memory and refcnt initialized.
 756          */
 757         smp_wmb();
 758         atomic_set(&req->rsk_refcnt, 2 + 1);
 759 }
 760
 761 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 762                                    unsigned long timeout)
 763 {
 764         reqsk_queue_hash_req(req, timeout);
 765         inet_csk_reqsk_queue_added(sk);
 766 }
 767 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 768
 769 /**
 770  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 771  *      @sk: the socket to clone
 772  *      @req: request_sock
 773  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 774  *
 775  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 776  */
 777 struct sock *inet_csk_clone_lock(const struct sock *sk,
 778                                  const struct request_sock *req,
 779                                  const gfp_t priority)
 780 {
 781         struct sock *newsk = sk_clone_lock(sk, priority);
 782
 783         if (newsk) {
 784                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 785
 786                 newsk->sk_state = TCP_SYN_RECV;
 787                 newicsk->icsk_bind_hash = NULL;
 788
 789                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 790                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 791                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 792                 newsk->sk_write_space = sk_stream_write_space;
 793
 794                 /* listeners have SOCK_RCU_FREE, not the children */
 795                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 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         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         sk->sk_state = 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         reqsk_put(req);
 918 }
 919
 920 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 921                                       struct request_sock *req,
 922                                       struct sock *child)
 923 {
 924         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 925
 926         spin_lock(&queue->rskq_lock);
 927         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 928                 inet_child_forget(sk, req, child);
 929                 child = NULL;
 930         } else {
 931                 req->sk = child;
 932                 req->dl_next = NULL;
 933                 if (queue->rskq_accept_head == NULL)
 934                         queue->rskq_accept_head = req;
 935                 else
 936                         queue->rskq_accept_tail->dl_next = req;
 937                 queue->rskq_accept_tail = req;
 938                 sk_acceptq_added(sk);
 939         }
 940         spin_unlock(&queue->rskq_lock);
 941         return child;
 942 }
 943 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 944
 945 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 946                                          struct request_sock *req, bool own_req)
 947 {
 948         if (own_req) {
 949                 inet_csk_reqsk_queue_drop(sk, req);
 950                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 951                 if (inet_csk_reqsk_queue_add(sk, req, child))
 952                         return child;
 953         }
 954         /* Too bad, another child took ownership of the request, undo. */
 955         bh_unlock_sock(child);
 956         sock_put(child);
 957         return NULL;
 958 }
 959 EXPORT_SYMBOL(inet_csk_complete_hashdance);
 960
 961 /*
 962  *      This routine closes sockets which have been at least partially
 963  *      opened, but not yet accepted.
 964  */
 965 void inet_csk_listen_stop(struct sock *sk)
 966 {
 967         struct inet_connection_sock *icsk = inet_csk(sk);
 968         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 969         struct request_sock *next, *req;
 970
 971         /* Following specs, it would be better either to send FIN
 972          * (and enter FIN-WAIT-1, it is normal close)
 973          * or to send active reset (abort).
 974          * Certainly, it is pretty dangerous while synflood, but it is
 975          * bad justification for our negligence 8)
 976          * To be honest, we are not able to make either
 977          * of the variants now.                 --ANK
 978          */
 979         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
 980                 struct sock *child = req->sk;
 981
 982                 local_bh_disable();
 983                 bh_lock_sock(child);
 984                 WARN_ON(sock_owned_by_user(child));
 985                 sock_hold(child);
 986
 987                 inet_child_forget(sk, req, child);
 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);