net/ipv4/inet_connection_sock.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Support for INET connection oriented protocols.
   8  *
   9  * Authors:     See the TCP sources
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/jhash.h>
  14
  15 #include <net/inet_connection_sock.h>
  16 #include <net/inet_hashtables.h>
  17 #include <net/inet_timewait_sock.h>
  18 #include <net/ip.h>
  19 #include <net/route.h>
  20 #include <net/tcp_states.h>
  21 #include <net/xfrm.h>
  22 #include <net/tcp.h>
  23 #include <net/sock_reuseport.h>
  24 #include <net/addrconf.h>
  25
  26 #if IS_ENABLED(CONFIG_IPV6)
  27 /* match_sk*_wildcard == true:  IPV6_ADDR_ANY equals to any IPv6 addresses
  28  *                              if IPv6 only, and any IPv4 addresses
  29  *                              if not IPv6 only
  30  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  31  *                              IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY,
  32  *                              and 0.0.0.0 equals to 0.0.0.0 only
  33  */
  34 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6,
  35                                  const struct in6_addr *sk2_rcv_saddr6,
  36                                  __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  37                                  bool sk1_ipv6only, bool sk2_ipv6only,
  38                                  bool match_sk1_wildcard,
  39                                  bool match_sk2_wildcard)
  40 {
  41         int addr_type = ipv6_addr_type(sk1_rcv_saddr6);
  42         int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED;
  43
  44         /* if both are mapped, treat as IPv4 */
  45         if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) {
  46                 if (!sk2_ipv6only) {
  47                         if (sk1_rcv_saddr == sk2_rcv_saddr)
  48                                 return true;
  49                         return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  50                                 (match_sk2_wildcard && !sk2_rcv_saddr);
  51                 }
  52                 return false;
  53         }
  54
  55         if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY)
  56                 return true;
  57
  58         if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard &&
  59             !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED))
  60                 return true;
  61
  62         if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard &&
  63             !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED))
  64                 return true;
  65
  66         if (sk2_rcv_saddr6 &&
  67             ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6))
  68                 return true;
  69
  70         return false;
  71 }
  72 #endif
  73
  74 /* match_sk*_wildcard == true:  0.0.0.0 equals to any IPv4 addresses
  75  * match_sk*_wildcard == false: addresses must be exactly the same, i.e.
  76  *                              0.0.0.0 only equals to 0.0.0.0
  77  */
  78 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr,
  79                                  bool sk2_ipv6only, bool match_sk1_wildcard,
  80                                  bool match_sk2_wildcard)
  81 {
  82         if (!sk2_ipv6only) {
  83                 if (sk1_rcv_saddr == sk2_rcv_saddr)
  84                         return true;
  85                 return (match_sk1_wildcard && !sk1_rcv_saddr) ||
  86                         (match_sk2_wildcard && !sk2_rcv_saddr);
  87         }
  88         return false;
  89 }
  90
  91 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2,
  92                           bool match_wildcard)
  93 {
  94 #if IS_ENABLED(CONFIG_IPV6)
  95         if (sk->sk_family == AF_INET6)
  96                 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr,
  97                                             inet6_rcv_saddr(sk2),
  98                                             sk->sk_rcv_saddr,
  99                                             sk2->sk_rcv_saddr,
 100                                             ipv6_only_sock(sk),
 101                                             ipv6_only_sock(sk2),
 102                                             match_wildcard,
 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                                     match_wildcard);
 108 }
 109 EXPORT_SYMBOL(inet_rcv_saddr_equal);
 110
 111 bool inet_rcv_saddr_any(const struct sock *sk)
 112 {
 113 #if IS_ENABLED(CONFIG_IPV6)
 114         if (sk->sk_family == AF_INET6)
 115                 return ipv6_addr_any(&sk->sk_v6_rcv_saddr);
 116 #endif
 117         return !sk->sk_rcv_saddr;
 118 }
 119
 120 void inet_get_local_port_range(struct net *net, int *low, int *high)
 121 {
 122         unsigned int seq;
 123
 124         do {
 125                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
 126
 127                 *low = net->ipv4.ip_local_ports.range[0];
 128                 *high = net->ipv4.ip_local_ports.range[1];
 129         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
 130 }
 131 EXPORT_SYMBOL(inet_get_local_port_range);
 132
 133 static bool inet_use_bhash2_on_bind(const struct sock *sk)
 134 {
 135 #if IS_ENABLED(CONFIG_IPV6)
 136         if (sk->sk_family == AF_INET6) {
 137                 int addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 138
 139                 return addr_type != IPV6_ADDR_ANY &&
 140                         addr_type != IPV6_ADDR_MAPPED;
 141         }
 142 #endif
 143         return sk->sk_rcv_saddr != htonl(INADDR_ANY);
 144 }
 145
 146 static bool inet_bind_conflict(const struct sock *sk, struct sock *sk2,
 147                                kuid_t sk_uid, bool relax,
 148                                bool reuseport_cb_ok, bool reuseport_ok)
 149 {
 150         int bound_dev_if2;
 151
 152         if (sk == sk2)
 153                 return false;
 154
 155         bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
 156
 157         if (!sk->sk_bound_dev_if || !bound_dev_if2 ||
 158             sk->sk_bound_dev_if == bound_dev_if2) {
 159                 if (sk->sk_reuse && sk2->sk_reuse &&
 160                     sk2->sk_state != TCP_LISTEN) {
 161                         if (!relax || (!reuseport_ok && sk->sk_reuseport &&
 162                                        sk2->sk_reuseport && reuseport_cb_ok &&
 163                                        (sk2->sk_state == TCP_TIME_WAIT ||
 164                                         uid_eq(sk_uid, sock_i_uid(sk2)))))
 165                                 return true;
 166                 } else if (!reuseport_ok || !sk->sk_reuseport ||
 167                            !sk2->sk_reuseport || !reuseport_cb_ok ||
 168                            (sk2->sk_state != TCP_TIME_WAIT &&
 169                             !uid_eq(sk_uid, sock_i_uid(sk2)))) {
 170                         return true;
 171                 }
 172         }
 173         return false;
 174 }
 175
 176 static bool inet_bhash2_conflict(const struct sock *sk,
 177                                  const struct inet_bind2_bucket *tb2,
 178                                  kuid_t sk_uid,
 179                                  bool relax, bool reuseport_cb_ok,
 180                                  bool reuseport_ok)
 181 {
 182         struct sock *sk2;
 183
 184         sk_for_each_bound_bhash2(sk2, &tb2->owners) {
 185                 if (sk->sk_family == AF_INET && ipv6_only_sock(sk2))
 186                         continue;
 187
 188                 if (inet_bind_conflict(sk, sk2, sk_uid, relax,
 189                                        reuseport_cb_ok, reuseport_ok))
 190                         return true;
 191         }
 192         return false;
 193 }
 194
 195 /* This should be called only when the tb and tb2 hashbuckets' locks are held */
 196 static int inet_csk_bind_conflict(const struct sock *sk,
 197                                   const struct inet_bind_bucket *tb,
 198                                   const struct inet_bind2_bucket *tb2, /* may be null */
 199                                   bool relax, bool reuseport_ok)
 200 {
 201         bool reuseport_cb_ok;
 202         struct sock_reuseport *reuseport_cb;
 203         kuid_t uid = sock_i_uid((struct sock *)sk);
 204
 205         rcu_read_lock();
 206         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 207         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 208         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 209         rcu_read_unlock();
 210
 211         /*
 212          * Unlike other sk lookup places we do not check
 213          * for sk_net here, since _all_ the socks listed
 214          * in tb->owners and tb2->owners list belong
 215          * to the same net - the one this bucket belongs to.
 216          */
 217
 218         if (!inet_use_bhash2_on_bind(sk)) {
 219                 struct sock *sk2;
 220
 221                 sk_for_each_bound(sk2, &tb->owners)
 222                         if (inet_bind_conflict(sk, sk2, uid, relax,
 223                                                reuseport_cb_ok, reuseport_ok) &&
 224                             inet_rcv_saddr_equal(sk, sk2, true))
 225                                 return true;
 226
 227                 return false;
 228         }
 229
 230         /* Conflicts with an existing IPV6_ADDR_ANY (if ipv6) or INADDR_ANY (if
 231          * ipv4) should have been checked already. We need to do these two
 232          * checks separately because their spinlocks have to be acquired/released
 233          * independently of each other, to prevent possible deadlocks
 234          */
 235         return tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
 236                                            reuseport_ok);
 237 }
 238
 239 /* Determine if there is a bind conflict with an existing IPV6_ADDR_ANY (if ipv6) or
 240  * INADDR_ANY (if ipv4) socket.
 241  *
 242  * Caller must hold bhash hashbucket lock with local bh disabled, to protect
 243  * against concurrent binds on the port for addr any
 244  */
 245 static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l3mdev,
 246                                           bool relax, bool reuseport_ok)
 247 {
 248         kuid_t uid = sock_i_uid((struct sock *)sk);
 249         const struct net *net = sock_net(sk);
 250         struct sock_reuseport *reuseport_cb;
 251         struct inet_bind_hashbucket *head2;
 252         struct inet_bind2_bucket *tb2;
 253         bool reuseport_cb_ok;
 254
 255         rcu_read_lock();
 256         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 257         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 258         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 259         rcu_read_unlock();
 260
 261         head2 = inet_bhash2_addr_any_hashbucket(sk, net, port);
 262
 263         spin_lock(&head2->lock);
 264
 265         inet_bind_bucket_for_each(tb2, &head2->chain)
 266                 if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
 267                         break;
 268
 269         if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
 270                                         reuseport_ok)) {
 271                 spin_unlock(&head2->lock);
 272                 return true;
 273         }
 274
 275         spin_unlock(&head2->lock);
 276         return false;
 277 }
 278
 279 /*
 280  * Find an open port number for the socket.  Returns with the
 281  * inet_bind_hashbucket locks held if successful.
 282  */
 283 static struct inet_bind_hashbucket *
 284 inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 285                         struct inet_bind2_bucket **tb2_ret,
 286                         struct inet_bind_hashbucket **head2_ret, int *port_ret)
 287 {
 288         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 289         int i, low, high, attempt_half, port, l3mdev;
 290         struct inet_bind_hashbucket *head, *head2;
 291         struct net *net = sock_net(sk);
 292         struct inet_bind2_bucket *tb2;
 293         struct inet_bind_bucket *tb;
 294         u32 remaining, offset;
 295         bool relax = false;
 296
 297         l3mdev = inet_sk_bound_l3mdev(sk);
 298 ports_exhausted:
 299         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 300 other_half_scan:
 301         inet_get_local_port_range(net, &low, &high);
 302         high++; /* [32768, 60999] -> [32768, 61000[ */
 303         if (high - low < 4)
 304                 attempt_half = 0;
 305         if (attempt_half) {
 306                 int half = low + (((high - low) >> 2) << 1);
 307
 308                 if (attempt_half == 1)
 309                         high = half;
 310                 else
 311                         low = half;
 312         }
 313         remaining = high - low;
 314         if (likely(remaining > 1))
 315                 remaining &= ~1U;
 316
 317         offset = get_random_u32_below(remaining);
 318         /* __inet_hash_connect() favors ports having @low parity
 319          * We do the opposite to not pollute connect() users.
 320          */
 321         offset |= 1U;
 322
 323 other_parity_scan:
 324         port = low + offset;
 325         for (i = 0; i < remaining; i += 2, port += 2) {
 326                 if (unlikely(port >= high))
 327                         port -= remaining;
 328                 if (inet_is_local_reserved_port(net, port))
 329                         continue;
 330                 head = &hinfo->bhash[inet_bhashfn(net, port,
 331                                                   hinfo->bhash_size)];
 332                 spin_lock_bh(&head->lock);
 333                 if (inet_use_bhash2_on_bind(sk)) {
 334                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, relax, false))
 335                                 goto next_port;
 336                 }
 337
 338                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 339                 spin_lock(&head2->lock);
 340                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 341                 inet_bind_bucket_for_each(tb, &head->chain)
 342                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
 343                                 if (!inet_csk_bind_conflict(sk, tb, tb2,
 344                                                             relax, false))
 345                                         goto success;
 346                                 spin_unlock(&head2->lock);
 347                                 goto next_port;
 348                         }
 349                 tb = NULL;
 350                 goto success;
 351 next_port:
 352                 spin_unlock_bh(&head->lock);
 353                 cond_resched();
 354         }
 355
 356         offset--;
 357         if (!(offset & 1))
 358                 goto other_parity_scan;
 359
 360         if (attempt_half == 1) {
 361                 /* OK we now try the upper half of the range */
 362                 attempt_half = 2;
 363                 goto other_half_scan;
 364         }
 365
 366         if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) {
 367                 /* We still have a chance to connect to different destinations */
 368                 relax = true;
 369                 goto ports_exhausted;
 370         }
 371         return NULL;
 372 success:
 373         *port_ret = port;
 374         *tb_ret = tb;
 375         *tb2_ret = tb2;
 376         *head2_ret = head2;
 377         return head;
 378 }
 379
 380 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 381                                      struct sock *sk)
 382 {
 383         kuid_t uid = sock_i_uid(sk);
 384
 385         if (tb->fastreuseport <= 0)
 386                 return 0;
 387         if (!sk->sk_reuseport)
 388                 return 0;
 389         if (rcu_access_pointer(sk->sk_reuseport_cb))
 390                 return 0;
 391         if (!uid_eq(tb->fastuid, uid))
 392                 return 0;
 393         /* We only need to check the rcv_saddr if this tb was once marked
 394          * without fastreuseport and then was reset, as we can only know that
 395          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 396          * owners list.
 397          */
 398         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 399                 return 1;
 400 #if IS_ENABLED(CONFIG_IPV6)
 401         if (tb->fast_sk_family == AF_INET6)
 402                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 403                                             inet6_rcv_saddr(sk),
 404                                             tb->fast_rcv_saddr,
 405                                             sk->sk_rcv_saddr,
 406                                             tb->fast_ipv6_only,
 407                                             ipv6_only_sock(sk), true, false);
 408 #endif
 409         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 410                                     ipv6_only_sock(sk), true, false);
 411 }
 412
 413 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 414                                struct sock *sk)
 415 {
 416         kuid_t uid = sock_i_uid(sk);
 417         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 418
 419         if (hlist_empty(&tb->owners)) {
 420                 tb->fastreuse = reuse;
 421                 if (sk->sk_reuseport) {
 422                         tb->fastreuseport = FASTREUSEPORT_ANY;
 423                         tb->fastuid = uid;
 424                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 425                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 426                         tb->fast_sk_family = sk->sk_family;
 427 #if IS_ENABLED(CONFIG_IPV6)
 428                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 429 #endif
 430                 } else {
 431                         tb->fastreuseport = 0;
 432                 }
 433         } else {
 434                 if (!reuse)
 435                         tb->fastreuse = 0;
 436                 if (sk->sk_reuseport) {
 437                         /* We didn't match or we don't have fastreuseport set on
 438                          * the tb, but we have sk_reuseport set on this socket
 439                          * and we know that there are no bind conflicts with
 440                          * this socket in this tb, so reset our tb's reuseport
 441                          * settings so that any subsequent sockets that match
 442                          * our current socket will be put on the fast path.
 443                          *
 444                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 445                          * do extra checking for all subsequent sk_reuseport
 446                          * socks.
 447                          */
 448                         if (!sk_reuseport_match(tb, sk)) {
 449                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 450                                 tb->fastuid = uid;
 451                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 452                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 453                                 tb->fast_sk_family = sk->sk_family;
 454 #if IS_ENABLED(CONFIG_IPV6)
 455                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 456 #endif
 457                         }
 458                 } else {
 459                         tb->fastreuseport = 0;
 460                 }
 461         }
 462 }
 463
 464 /* Obtain a reference to a local port for the given sock,
 465  * if snum is zero it means select any available local port.
 466  * We try to allocate an odd port (and leave even ports for connect())
 467  */
 468 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 469 {
 470         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 471         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 472         bool found_port = false, check_bind_conflict = true;
 473         bool bhash_created = false, bhash2_created = false;
 474         int ret = -EADDRINUSE, port = snum, l3mdev;
 475         struct inet_bind_hashbucket *head, *head2;
 476         struct inet_bind2_bucket *tb2 = NULL;
 477         struct inet_bind_bucket *tb = NULL;
 478         bool head2_lock_acquired = false;
 479         struct net *net = sock_net(sk);
 480
 481         l3mdev = inet_sk_bound_l3mdev(sk);
 482
 483         if (!port) {
 484                 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port);
 485                 if (!head)
 486                         return ret;
 487
 488                 head2_lock_acquired = true;
 489
 490                 if (tb && tb2)
 491                         goto success;
 492                 found_port = true;
 493         } else {
 494                 head = &hinfo->bhash[inet_bhashfn(net, port,
 495                                                   hinfo->bhash_size)];
 496                 spin_lock_bh(&head->lock);
 497                 inet_bind_bucket_for_each(tb, &head->chain)
 498                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 499                                 break;
 500         }
 501
 502         if (!tb) {
 503                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net,
 504                                              head, port, l3mdev);
 505                 if (!tb)
 506                         goto fail_unlock;
 507                 bhash_created = true;
 508         }
 509
 510         if (!found_port) {
 511                 if (!hlist_empty(&tb->owners)) {
 512                         if (sk->sk_reuse == SK_FORCE_REUSE ||
 513                             (tb->fastreuse > 0 && reuse) ||
 514                             sk_reuseport_match(tb, sk))
 515                                 check_bind_conflict = false;
 516                 }
 517
 518                 if (check_bind_conflict && inet_use_bhash2_on_bind(sk)) {
 519                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, true, true))
 520                                 goto fail_unlock;
 521                 }
 522
 523                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 524                 spin_lock(&head2->lock);
 525                 head2_lock_acquired = true;
 526                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 527         }
 528
 529         if (!tb2) {
 530                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep,
 531                                                net, head2, port, l3mdev, sk);
 532                 if (!tb2)
 533                         goto fail_unlock;
 534                 bhash2_created = true;
 535         }
 536
 537         if (!found_port && check_bind_conflict) {
 538                 if (inet_csk_bind_conflict(sk, tb, tb2, true, true))
 539                         goto fail_unlock;
 540         }
 541
 542 success:
 543         inet_csk_update_fastreuse(tb, sk);
 544
 545         if (!inet_csk(sk)->icsk_bind_hash)
 546                 inet_bind_hash(sk, tb, tb2, port);
 547         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 548         WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
 549         ret = 0;
 550
 551 fail_unlock:
 552         if (ret) {
 553                 if (bhash_created)
 554                         inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
 555                 if (bhash2_created)
 556                         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep,
 557                                                   tb2);
 558         }
 559         if (head2_lock_acquired)
 560                 spin_unlock(&head2->lock);
 561         spin_unlock_bh(&head->lock);
 562         return ret;
 563 }
 564 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 565
 566 /*
 567  * Wait for an incoming connection, avoid race conditions. This must be called
 568  * with the socket locked.
 569  */
 570 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 571 {
 572         struct inet_connection_sock *icsk = inet_csk(sk);
 573         DEFINE_WAIT(wait);
 574         int err;
 575
 576         /*
 577          * True wake-one mechanism for incoming connections: only
 578          * one process gets woken up, not the 'whole herd'.
 579          * Since we do not 'race & poll' for established sockets
 580          * anymore, the common case will execute the loop only once.
 581          *
 582          * Subtle issue: "add_wait_queue_exclusive()" will be added
 583          * after any current non-exclusive waiters, and we know that
 584          * it will always _stay_ after any new non-exclusive waiters
 585          * because all non-exclusive waiters are added at the
 586          * beginning of the wait-queue. As such, it's ok to "drop"
 587          * our exclusiveness temporarily when we get woken up without
 588          * having to remove and re-insert us on the wait queue.
 589          */
 590         for (;;) {
 591                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 592                                           TASK_INTERRUPTIBLE);
 593                 release_sock(sk);
 594                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 595                         timeo = schedule_timeout(timeo);
 596                 sched_annotate_sleep();
 597                 lock_sock(sk);
 598                 err = 0;
 599                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 600                         break;
 601                 err = -EINVAL;
 602                 if (sk->sk_state != TCP_LISTEN)
 603                         break;
 604                 err = sock_intr_errno(timeo);
 605                 if (signal_pending(current))
 606                         break;
 607                 err = -EAGAIN;
 608                 if (!timeo)
 609                         break;
 610         }
 611         finish_wait(sk_sleep(sk), &wait);
 612         return err;
 613 }
 614
 615 /*
 616  * This will accept the next outstanding connection.
 617  */
 618 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 619 {
 620         struct inet_connection_sock *icsk = inet_csk(sk);
 621         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 622         struct request_sock *req;
 623         struct sock *newsk;
 624         int error;
 625
 626         lock_sock(sk);
 627
 628         /* We need to make sure that this socket is listening,
 629          * and that it has something pending.
 630          */
 631         error = -EINVAL;
 632         if (sk->sk_state != TCP_LISTEN)
 633                 goto out_err;
 634
 635         /* Find already established connection */
 636         if (reqsk_queue_empty(queue)) {
 637                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 638
 639                 /* If this is a non blocking socket don't sleep */
 640                 error = -EAGAIN;
 641                 if (!timeo)
 642                         goto out_err;
 643
 644                 error = inet_csk_wait_for_connect(sk, timeo);
 645                 if (error)
 646                         goto out_err;
 647         }
 648         req = reqsk_queue_remove(queue, sk);
 649         newsk = req->sk;
 650
 651         if (sk->sk_protocol == IPPROTO_TCP &&
 652             tcp_rsk(req)->tfo_listener) {
 653                 spin_lock_bh(&queue->fastopenq.lock);
 654                 if (tcp_rsk(req)->tfo_listener) {
 655                         /* We are still waiting for the final ACK from 3WHS
 656                          * so can't free req now. Instead, we set req->sk to
 657                          * NULL to signify that the child socket is taken
 658                          * so reqsk_fastopen_remove() will free the req
 659                          * when 3WHS finishes (or is aborted).
 660                          */
 661                         req->sk = NULL;
 662                         req = NULL;
 663                 }
 664                 spin_unlock_bh(&queue->fastopenq.lock);
 665         }
 666
 667 out:
 668         release_sock(sk);
 669         if (newsk && mem_cgroup_sockets_enabled) {
 670                 int amt;
 671
 672                 /* atomically get the memory usage, set and charge the
 673                  * newsk->sk_memcg.
 674                  */
 675                 lock_sock(newsk);
 676
 677                 /* The socket has not been accepted yet, no need to look at
 678                  * newsk->sk_wmem_queued.
 679                  */
 680                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 681                                    atomic_read(&newsk->sk_rmem_alloc));
 682                 mem_cgroup_sk_alloc(newsk);
 683                 if (newsk->sk_memcg && amt)
 684                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt,
 685                                                 GFP_KERNEL | __GFP_NOFAIL);
 686
 687                 release_sock(newsk);
 688         }
 689         if (req)
 690                 reqsk_put(req);
 691         return newsk;
 692 out_err:
 693         newsk = NULL;
 694         req = NULL;
 695         *err = error;
 696         goto out;
 697 }
 698 EXPORT_SYMBOL(inet_csk_accept);
 699
 700 /*
 701  * Using different timers for retransmit, delayed acks and probes
 702  * We may wish use just one timer maintaining a list of expire jiffies
 703  * to optimize.
 704  */
 705 void inet_csk_init_xmit_timers(struct sock *sk,
 706                                void (*retransmit_handler)(struct timer_list *t),
 707                                void (*delack_handler)(struct timer_list *t),
 708                                void (*keepalive_handler)(struct timer_list *t))
 709 {
 710         struct inet_connection_sock *icsk = inet_csk(sk);
 711
 712         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 713         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 714         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 715         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 716 }
 717 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 718
 719 void inet_csk_clear_xmit_timers(struct sock *sk)
 720 {
 721         struct inet_connection_sock *icsk = inet_csk(sk);
 722
 723         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 724
 725         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 726         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 727         sk_stop_timer(sk, &sk->sk_timer);
 728 }
 729 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 730
 731 void inet_csk_delete_keepalive_timer(struct sock *sk)
 732 {
 733         sk_stop_timer(sk, &sk->sk_timer);
 734 }
 735 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 736
 737 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 738 {
 739         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 740 }
 741 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 742
 743 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 744                                      struct flowi4 *fl4,
 745                                      const struct request_sock *req)
 746 {
 747         const struct inet_request_sock *ireq = inet_rsk(req);
 748         struct net *net = read_pnet(&ireq->ireq_net);
 749         struct ip_options_rcu *opt;
 750         struct rtable *rt;
 751
 752         rcu_read_lock();
 753         opt = rcu_dereference(ireq->ireq_opt);
 754
 755         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 756                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 757                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 758                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 759                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 760                            htons(ireq->ir_num), sk->sk_uid);
 761         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 762         rt = ip_route_output_flow(net, fl4, sk);
 763         if (IS_ERR(rt))
 764                 goto no_route;
 765         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 766                 goto route_err;
 767         rcu_read_unlock();
 768         return &rt->dst;
 769
 770 route_err:
 771         ip_rt_put(rt);
 772 no_route:
 773         rcu_read_unlock();
 774         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 775         return NULL;
 776 }
 777 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 778
 779 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 780                                             struct sock *newsk,
 781                                             const struct request_sock *req)
 782 {
 783         const struct inet_request_sock *ireq = inet_rsk(req);
 784         struct net *net = read_pnet(&ireq->ireq_net);
 785         struct inet_sock *newinet = inet_sk(newsk);
 786         struct ip_options_rcu *opt;
 787         struct flowi4 *fl4;
 788         struct rtable *rt;
 789
 790         opt = rcu_dereference(ireq->ireq_opt);
 791         fl4 = &newinet->cork.fl.u.ip4;
 792
 793         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 794                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 795                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 796                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 797                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 798                            htons(ireq->ir_num), sk->sk_uid);
 799         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 800         rt = ip_route_output_flow(net, fl4, sk);
 801         if (IS_ERR(rt))
 802                 goto no_route;
 803         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 804                 goto route_err;
 805         return &rt->dst;
 806
 807 route_err:
 808         ip_rt_put(rt);
 809 no_route:
 810         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 811         return NULL;
 812 }
 813 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 814
 815 /* Decide when to expire the request and when to resend SYN-ACK */
 816 static void syn_ack_recalc(struct request_sock *req,
 817                            const int max_syn_ack_retries,
 818                            const u8 rskq_defer_accept,
 819                            int *expire, int *resend)
 820 {
 821         if (!rskq_defer_accept) {
 822                 *expire = req->num_timeout >= max_syn_ack_retries;
 823                 *resend = 1;
 824                 return;
 825         }
 826         *expire = req->num_timeout >= max_syn_ack_retries &&
 827                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 828         /* Do not resend while waiting for data after ACK,
 829          * start to resend on end of deferring period to give
 830          * last chance for data or ACK to create established socket.
 831          */
 832         *resend = !inet_rsk(req)->acked ||
 833                   req->num_timeout >= rskq_defer_accept - 1;
 834 }
 835
 836 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 837 {
 838         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 839
 840         if (!err)
 841                 req->num_retrans++;
 842         return err;
 843 }
 844 EXPORT_SYMBOL(inet_rtx_syn_ack);
 845
 846 static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 847                                              struct sock *sk)
 848 {
 849         struct sock *req_sk, *nreq_sk;
 850         struct request_sock *nreq;
 851
 852         nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN);
 853         if (!nreq) {
 854                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
 855
 856                 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */
 857                 sock_put(sk);
 858                 return NULL;
 859         }
 860
 861         req_sk = req_to_sk(req);
 862         nreq_sk = req_to_sk(nreq);
 863
 864         memcpy(nreq_sk, req_sk,
 865                offsetof(struct sock, sk_dontcopy_begin));
 866         memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end,
 867                req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end));
 868
 869         sk_node_init(&nreq_sk->sk_node);
 870         nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping;
 871 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
 872         nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping;
 873 #endif
 874         nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu;
 875
 876         nreq->rsk_listener = sk;
 877
 878         /* We need not acquire fastopenq->lock
 879          * because the child socket is locked in inet_csk_listen_stop().
 880          */
 881         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener)
 882                 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
 883
 884         return nreq;
 885 }
 886
 887 static void reqsk_queue_migrated(struct request_sock_queue *queue,
 888                                  const struct request_sock *req)
 889 {
 890         if (req->num_timeout == 0)
 891                 atomic_inc(&queue->young);
 892         atomic_inc(&queue->qlen);
 893 }
 894
 895 static void reqsk_migrate_reset(struct request_sock *req)
 896 {
 897         req->saved_syn = NULL;
 898 #if IS_ENABLED(CONFIG_IPV6)
 899         inet_rsk(req)->ipv6_opt = NULL;
 900         inet_rsk(req)->pktopts = NULL;
 901 #else
 902         inet_rsk(req)->ireq_opt = NULL;
 903 #endif
 904 }
 905
 906 /* return true if req was found in the ehash table */
 907 static bool reqsk_queue_unlink(struct request_sock *req)
 908 {
 909         struct sock *sk = req_to_sk(req);
 910         bool found = false;
 911
 912         if (sk_hashed(sk)) {
 913                 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 914                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 915
 916                 spin_lock(lock);
 917                 found = __sk_nulls_del_node_init_rcu(sk);
 918                 spin_unlock(lock);
 919         }
 920         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 921                 reqsk_put(req);
 922         return found;
 923 }
 924
 925 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 926 {
 927         bool unlinked = reqsk_queue_unlink(req);
 928
 929         if (unlinked) {
 930                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 931                 reqsk_put(req);
 932         }
 933         return unlinked;
 934 }
 935 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 936
 937 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 938 {
 939         inet_csk_reqsk_queue_drop(sk, req);
 940         reqsk_put(req);
 941 }
 942 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 943
 944 static void reqsk_timer_handler(struct timer_list *t)
 945 {
 946         struct request_sock *req = from_timer(req, t, rsk_timer);
 947         struct request_sock *nreq = NULL, *oreq = req;
 948         struct sock *sk_listener = req->rsk_listener;
 949         struct inet_connection_sock *icsk;
 950         struct request_sock_queue *queue;
 951         struct net *net;
 952         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
 953
 954         if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
 955                 struct sock *nsk;
 956
 957                 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
 958                 if (!nsk)
 959                         goto drop;
 960
 961                 nreq = inet_reqsk_clone(req, nsk);
 962                 if (!nreq)
 963                         goto drop;
 964
 965                 /* The new timer for the cloned req can decrease the 2
 966                  * by calling inet_csk_reqsk_queue_drop_and_put(), so
 967                  * hold another count to prevent use-after-free and
 968                  * call reqsk_put() just before return.
 969                  */
 970                 refcount_set(&nreq->rsk_refcnt, 2 + 1);
 971                 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
 972                 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
 973
 974                 req = nreq;
 975                 sk_listener = nsk;
 976         }
 977
 978         icsk = inet_csk(sk_listener);
 979         net = sock_net(sk_listener);
 980         max_syn_ack_retries = icsk->icsk_syn_retries ? :
 981                 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries);
 982         /* Normally all the openreqs are young and become mature
 983          * (i.e. converted to established socket) for first timeout.
 984          * If synack was not acknowledged for 1 second, it means
 985          * one of the following things: synack was lost, ack was lost,
 986          * rtt is high or nobody planned to ack (i.e. synflood).
 987          * When server is a bit loaded, queue is populated with old
 988          * open requests, reducing effective size of queue.
 989          * When server is well loaded, queue size reduces to zero
 990          * after several minutes of work. It is not synflood,
 991          * it is normal operation. The solution is pruning
 992          * too old entries overriding normal timeout, when
 993          * situation becomes dangerous.
 994          *
 995          * Essentially, we reserve half of room for young
 996          * embrions; and abort old ones without pity, if old
 997          * ones are about to clog our table.
 998          */
 999         queue = &icsk->icsk_accept_queue;
1000         qlen = reqsk_queue_len(queue);
1001         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
1002                 int young = reqsk_queue_len_young(queue) << 1;
1003
1004                 while (max_syn_ack_retries > 2) {
1005                         if (qlen < young)
1006                                 break;
1007                         max_syn_ack_retries--;
1008                         young <<= 1;
1009                 }
1010         }
1011         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
1012                        &expire, &resend);
1013         req->rsk_ops->syn_ack_timeout(req);
1014         if (!expire &&
1015             (!resend ||
1016              !inet_rtx_syn_ack(sk_listener, req) ||
1017              inet_rsk(req)->acked)) {
1018                 if (req->num_timeout++ == 0)
1019                         atomic_dec(&queue->young);
1020                 mod_timer(&req->rsk_timer, jiffies + reqsk_timeout(req, TCP_RTO_MAX));
1021
1022                 if (!nreq)
1023                         return;
1024
1025                 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
1026                         /* delete timer */
1027                         inet_csk_reqsk_queue_drop(sk_listener, nreq);
1028                         goto no_ownership;
1029                 }
1030
1031                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS);
1032                 reqsk_migrate_reset(oreq);
1033                 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
1034                 reqsk_put(oreq);
1035
1036                 reqsk_put(nreq);
1037                 return;
1038         }
1039
1040         /* Even if we can clone the req, we may need not retransmit any more
1041          * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
1042          * CPU may win the "own_req" race so that inet_ehash_insert() fails.
1043          */
1044         if (nreq) {
1045                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE);
1046 no_ownership:
1047                 reqsk_migrate_reset(nreq);
1048                 reqsk_queue_removed(queue, nreq);
1049                 __reqsk_free(nreq);
1050         }
1051
1052 drop:
1053         inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
1054 }
1055
1056 static void reqsk_queue_hash_req(struct request_sock *req,
1057                                  unsigned long timeout)
1058 {
1059         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1060         mod_timer(&req->rsk_timer, jiffies + timeout);
1061
1062         inet_ehash_insert(req_to_sk(req), NULL, NULL);
1063         /* before letting lookups find us, make sure all req fields
1064          * are committed to memory and refcnt initialized.
1065          */
1066         smp_wmb();
1067         refcount_set(&req->rsk_refcnt, 2 + 1);
1068 }
1069
1070 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
1071                                    unsigned long timeout)
1072 {
1073         reqsk_queue_hash_req(req, timeout);
1074         inet_csk_reqsk_queue_added(sk);
1075 }
1076 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
1077
1078 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
1079                            const gfp_t priority)
1080 {
1081         struct inet_connection_sock *icsk = inet_csk(newsk);
1082
1083         if (!icsk->icsk_ulp_ops)
1084                 return;
1085
1086         if (icsk->icsk_ulp_ops->clone)
1087                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
1088 }
1089
1090 /**
1091  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
1092  *      @sk: the socket to clone
1093  *      @req: request_sock
1094  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
1095  *
1096  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
1097  */
1098 struct sock *inet_csk_clone_lock(const struct sock *sk,
1099                                  const struct request_sock *req,
1100                                  const gfp_t priority)
1101 {
1102         struct sock *newsk = sk_clone_lock(sk, priority);
1103
1104         if (newsk) {
1105                 struct inet_connection_sock *newicsk = inet_csk(newsk);
1106
1107                 inet_sk_set_state(newsk, TCP_SYN_RECV);
1108                 newicsk->icsk_bind_hash = NULL;
1109                 newicsk->icsk_bind2_hash = NULL;
1110
1111                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
1112                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
1113                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
1114
1115                 /* listeners have SOCK_RCU_FREE, not the children */
1116                 sock_reset_flag(newsk, SOCK_RCU_FREE);
1117
1118                 inet_sk(newsk)->mc_list = NULL;
1119
1120                 newsk->sk_mark = inet_rsk(req)->ir_mark;
1121                 atomic64_set(&newsk->sk_cookie,
1122                              atomic64_read(&inet_rsk(req)->ir_cookie));
1123
1124                 newicsk->icsk_retransmits = 0;
1125                 newicsk->icsk_backoff     = 0;
1126                 newicsk->icsk_probes_out  = 0;
1127                 newicsk->icsk_probes_tstamp = 0;
1128
1129                 /* Deinitialize accept_queue to trap illegal accesses. */
1130                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
1131
1132                 inet_clone_ulp(req, newsk, priority);
1133
1134                 security_inet_csk_clone(newsk, req);
1135         }
1136         return newsk;
1137 }
1138 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
1139
1140 /*
1141  * At this point, there should be no process reference to this
1142  * socket, and thus no user references at all.  Therefore we
1143  * can assume the socket waitqueue is inactive and nobody will
1144  * try to jump onto it.
1145  */
1146 void inet_csk_destroy_sock(struct sock *sk)
1147 {
1148         WARN_ON(sk->sk_state != TCP_CLOSE);
1149         WARN_ON(!sock_flag(sk, SOCK_DEAD));
1150
1151         /* It cannot be in hash table! */
1152         WARN_ON(!sk_unhashed(sk));
1153
1154         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
1155         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
1156
1157         sk->sk_prot->destroy(sk);
1158
1159         sk_stream_kill_queues(sk);
1160
1161         xfrm_sk_free_policy(sk);
1162
1163         sk_refcnt_debug_release(sk);
1164
1165         this_cpu_dec(*sk->sk_prot->orphan_count);
1166
1167         sock_put(sk);
1168 }
1169 EXPORT_SYMBOL(inet_csk_destroy_sock);
1170
1171 /* This function allows to force a closure of a socket after the call to
1172  * tcp/dccp_create_openreq_child().
1173  */
1174 void inet_csk_prepare_forced_close(struct sock *sk)
1175         __releases(&sk->sk_lock.slock)
1176 {
1177         /* sk_clone_lock locked the socket and set refcnt to 2 */
1178         bh_unlock_sock(sk);
1179         sock_put(sk);
1180         inet_csk_prepare_for_destroy_sock(sk);
1181         inet_sk(sk)->inet_num = 0;
1182 }
1183 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
1184
1185 int inet_csk_listen_start(struct sock *sk)
1186 {
1187         struct inet_connection_sock *icsk = inet_csk(sk);
1188         struct inet_sock *inet = inet_sk(sk);
1189         int err;
1190
1191         reqsk_queue_alloc(&icsk->icsk_accept_queue);
1192
1193         sk->sk_ack_backlog = 0;
1194         inet_csk_delack_init(sk);
1195
1196         if (sk->sk_txrehash == SOCK_TXREHASH_DEFAULT)
1197                 sk->sk_txrehash = READ_ONCE(sock_net(sk)->core.sysctl_txrehash);
1198
1199         /* There is race window here: we announce ourselves listening,
1200          * but this transition is still not validated by get_port().
1201          * It is OK, because this socket enters to hash table only
1202          * after validation is complete.
1203          */
1204         inet_sk_state_store(sk, TCP_LISTEN);
1205         err = sk->sk_prot->get_port(sk, inet->inet_num);
1206         if (!err) {
1207                 inet->inet_sport = htons(inet->inet_num);
1208
1209                 sk_dst_reset(sk);
1210                 err = sk->sk_prot->hash(sk);
1211
1212                 if (likely(!err))
1213                         return 0;
1214         }
1215
1216         inet_sk_set_state(sk, TCP_CLOSE);
1217         return err;
1218 }
1219 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
1220
1221 static void inet_child_forget(struct sock *sk, struct request_sock *req,
1222                               struct sock *child)
1223 {
1224         sk->sk_prot->disconnect(child, O_NONBLOCK);
1225
1226         sock_orphan(child);
1227
1228         this_cpu_inc(*sk->sk_prot->orphan_count);
1229
1230         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
1231                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
1232                 BUG_ON(sk != req->rsk_listener);
1233
1234                 /* Paranoid, to prevent race condition if
1235                  * an inbound pkt destined for child is
1236                  * blocked by sock lock in tcp_v4_rcv().
1237                  * Also to satisfy an assertion in
1238                  * tcp_v4_destroy_sock().
1239                  */
1240                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1241         }
1242         inet_csk_destroy_sock(child);
1243 }
1244
1245 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1246                                       struct request_sock *req,
1247                                       struct sock *child)
1248 {
1249         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1250
1251         spin_lock(&queue->rskq_lock);
1252         if (unlikely(sk->sk_state != TCP_LISTEN)) {
1253                 inet_child_forget(sk, req, child);
1254                 child = NULL;
1255         } else {
1256                 req->sk = child;
1257                 req->dl_next = NULL;
1258                 if (queue->rskq_accept_head == NULL)
1259                         WRITE_ONCE(queue->rskq_accept_head, req);
1260                 else
1261                         queue->rskq_accept_tail->dl_next = req;
1262                 queue->rskq_accept_tail = req;
1263                 sk_acceptq_added(sk);
1264         }
1265         spin_unlock(&queue->rskq_lock);
1266         return child;
1267 }
1268 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1269
1270 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1271                                          struct request_sock *req, bool own_req)
1272 {
1273         if (own_req) {
1274                 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
1275                 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
1276
1277                 if (sk != req->rsk_listener) {
1278                         /* another listening sk has been selected,
1279                          * migrate the req to it.
1280                          */
1281                         struct request_sock *nreq;
1282
1283                         /* hold a refcnt for the nreq->rsk_listener
1284                          * which is assigned in inet_reqsk_clone()
1285                          */
1286                         sock_hold(sk);
1287                         nreq = inet_reqsk_clone(req, sk);
1288                         if (!nreq) {
1289                                 inet_child_forget(sk, req, child);
1290                                 goto child_put;
1291                         }
1292
1293                         refcount_set(&nreq->rsk_refcnt, 1);
1294                         if (inet_csk_reqsk_queue_add(sk, nreq, child)) {
1295                                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS);
1296                                 reqsk_migrate_reset(req);
1297                                 reqsk_put(req);
1298                                 return child;
1299                         }
1300
1301                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
1302                         reqsk_migrate_reset(nreq);
1303                         __reqsk_free(nreq);
1304                 } else if (inet_csk_reqsk_queue_add(sk, req, child)) {
1305                         return child;
1306                 }
1307         }
1308         /* Too bad, another child took ownership of the request, undo. */
1309 child_put:
1310         bh_unlock_sock(child);
1311         sock_put(child);
1312         return NULL;
1313 }
1314 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1315
1316 /*
1317  *      This routine closes sockets which have been at least partially
1318  *      opened, but not yet accepted.
1319  */
1320 void inet_csk_listen_stop(struct sock *sk)
1321 {
1322         struct inet_connection_sock *icsk = inet_csk(sk);
1323         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1324         struct request_sock *next, *req;
1325
1326         /* Following specs, it would be better either to send FIN
1327          * (and enter FIN-WAIT-1, it is normal close)
1328          * or to send active reset (abort).
1329          * Certainly, it is pretty dangerous while synflood, but it is
1330          * bad justification for our negligence 8)
1331          * To be honest, we are not able to make either
1332          * of the variants now.                 --ANK
1333          */
1334         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1335                 struct sock *child = req->sk, *nsk;
1336                 struct request_sock *nreq;
1337
1338                 local_bh_disable();
1339                 bh_lock_sock(child);
1340                 WARN_ON(sock_owned_by_user(child));
1341                 sock_hold(child);
1342
1343                 nsk = reuseport_migrate_sock(sk, child, NULL);
1344                 if (nsk) {
1345                         nreq = inet_reqsk_clone(req, nsk);
1346                         if (nreq) {
1347                                 refcount_set(&nreq->rsk_refcnt, 1);
1348
1349                                 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
1350                                         __NET_INC_STATS(sock_net(nsk),
1351                                                         LINUX_MIB_TCPMIGRATEREQSUCCESS);
1352                                         reqsk_migrate_reset(req);
1353                                 } else {
1354                                         __NET_INC_STATS(sock_net(nsk),
1355                                                         LINUX_MIB_TCPMIGRATEREQFAILURE);
1356                                         reqsk_migrate_reset(nreq);
1357                                         __reqsk_free(nreq);
1358                                 }
1359
1360                                 /* inet_csk_reqsk_queue_add() has already
1361                                  * called inet_child_forget() on failure case.
1362                                  */
1363                                 goto skip_child_forget;
1364                         }
1365                 }
1366
1367                 inet_child_forget(sk, req, child);
1368 skip_child_forget:
1369                 reqsk_put(req);
1370                 bh_unlock_sock(child);
1371                 local_bh_enable();
1372                 sock_put(child);
1373
1374                 cond_resched();
1375         }
1376         if (queue->fastopenq.rskq_rst_head) {
1377                 /* Free all the reqs queued in rskq_rst_head. */
1378                 spin_lock_bh(&queue->fastopenq.lock);
1379                 req = queue->fastopenq.rskq_rst_head;
1380                 queue->fastopenq.rskq_rst_head = NULL;
1381                 spin_unlock_bh(&queue->fastopenq.lock);
1382                 while (req != NULL) {
1383                         next = req->dl_next;
1384                         reqsk_put(req);
1385                         req = next;
1386                 }
1387         }
1388         WARN_ON_ONCE(sk->sk_ack_backlog);
1389 }
1390 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1391
1392 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1393 {
1394         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1395         const struct inet_sock *inet = inet_sk(sk);
1396
1397         sin->sin_family         = AF_INET;
1398         sin->sin_addr.s_addr    = inet->inet_daddr;
1399         sin->sin_port           = inet->inet_dport;
1400 }
1401 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1402
1403 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1404 {
1405         const struct inet_sock *inet = inet_sk(sk);
1406         const struct ip_options_rcu *inet_opt;
1407         __be32 daddr = inet->inet_daddr;
1408         struct flowi4 *fl4;
1409         struct rtable *rt;
1410
1411         rcu_read_lock();
1412         inet_opt = rcu_dereference(inet->inet_opt);
1413         if (inet_opt && inet_opt->opt.srr)
1414                 daddr = inet_opt->opt.faddr;
1415         fl4 = &fl->u.ip4;
1416         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1417                                    inet->inet_saddr, inet->inet_dport,
1418                                    inet->inet_sport, sk->sk_protocol,
1419                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1420         if (IS_ERR(rt))
1421                 rt = NULL;
1422         if (rt)
1423                 sk_setup_caps(sk, &rt->dst);
1424         rcu_read_unlock();
1425
1426         return &rt->dst;
1427 }
1428
1429 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1430 {
1431         struct dst_entry *dst = __sk_dst_check(sk, 0);
1432         struct inet_sock *inet = inet_sk(sk);
1433
1434         if (!dst) {
1435                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1436                 if (!dst)
1437                         goto out;
1438         }
1439         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1440
1441         dst = __sk_dst_check(sk, 0);
1442         if (!dst)
1443                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1444 out:
1445         return dst;
1446 }
1447 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);