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 /**
 121  *      inet_sk_get_local_port_range - fetch ephemeral ports range
 122  *      @sk: socket
 123  *      @low: pointer to low port
 124  *      @high: pointer to high port
 125  *
 126  *      Fetch netns port range (/proc/sys/net/ipv4/ip_local_port_range)
 127  *      Range can be overridden if socket got IP_LOCAL_PORT_RANGE option.
 128  *      Returns true if IP_LOCAL_PORT_RANGE was set on this socket.
 129  */
 130 bool inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
 131 {
 132         int lo, hi, sk_lo, sk_hi;
 133         bool local_range = false;
 134         u32 sk_range;
 135
 136         inet_get_local_port_range(sock_net(sk), &lo, &hi);
 137
 138         sk_range = READ_ONCE(inet_sk(sk)->local_port_range);
 139         if (unlikely(sk_range)) {
 140                 sk_lo = sk_range & 0xffff;
 141                 sk_hi = sk_range >> 16;
 142
 143                 if (lo <= sk_lo && sk_lo <= hi)
 144                         lo = sk_lo;
 145                 if (lo <= sk_hi && sk_hi <= hi)
 146                         hi = sk_hi;
 147                 local_range = true;
 148         }
 149
 150         *low = lo;
 151         *high = hi;
 152         return local_range;
 153 }
 154 EXPORT_SYMBOL(inet_sk_get_local_port_range);
 155
 156 static bool inet_use_bhash2_on_bind(const struct sock *sk)
 157 {
 158 #if IS_ENABLED(CONFIG_IPV6)
 159         if (sk->sk_family == AF_INET6) {
 160                 int addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 161
 162                 if (addr_type == IPV6_ADDR_ANY)
 163                         return false;
 164
 165                 if (addr_type != IPV6_ADDR_MAPPED)
 166                         return true;
 167         }
 168 #endif
 169         return sk->sk_rcv_saddr != htonl(INADDR_ANY);
 170 }
 171
 172 static bool inet_bind_conflict(const struct sock *sk, struct sock *sk2,
 173                                kuid_t sk_uid, bool relax,
 174                                bool reuseport_cb_ok, bool reuseport_ok)
 175 {
 176         int bound_dev_if2;
 177
 178         if (sk == sk2)
 179                 return false;
 180
 181         bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
 182
 183         if (!sk->sk_bound_dev_if || !bound_dev_if2 ||
 184             sk->sk_bound_dev_if == bound_dev_if2) {
 185                 if (sk->sk_reuse && sk2->sk_reuse &&
 186                     sk2->sk_state != TCP_LISTEN) {
 187                         if (!relax || (!reuseport_ok && sk->sk_reuseport &&
 188                                        sk2->sk_reuseport && reuseport_cb_ok &&
 189                                        (sk2->sk_state == TCP_TIME_WAIT ||
 190                                         uid_eq(sk_uid, sock_i_uid(sk2)))))
 191                                 return true;
 192                 } else if (!reuseport_ok || !sk->sk_reuseport ||
 193                            !sk2->sk_reuseport || !reuseport_cb_ok ||
 194                            (sk2->sk_state != TCP_TIME_WAIT &&
 195                             !uid_eq(sk_uid, sock_i_uid(sk2)))) {
 196                         return true;
 197                 }
 198         }
 199         return false;
 200 }
 201
 202 static bool __inet_bhash2_conflict(const struct sock *sk, struct sock *sk2,
 203                                    kuid_t sk_uid, bool relax,
 204                                    bool reuseport_cb_ok, bool reuseport_ok)
 205 {
 206         if (sk->sk_family == AF_INET && ipv6_only_sock(sk2))
 207                 return false;
 208
 209         return inet_bind_conflict(sk, sk2, sk_uid, relax,
 210                                   reuseport_cb_ok, reuseport_ok);
 211 }
 212
 213 static bool inet_bhash2_conflict(const struct sock *sk,
 214                                  const struct inet_bind2_bucket *tb2,
 215                                  kuid_t sk_uid,
 216                                  bool relax, bool reuseport_cb_ok,
 217                                  bool reuseport_ok)
 218 {
 219         struct sock *sk2;
 220
 221         sk_for_each_bound(sk2, &tb2->owners) {
 222                 if (__inet_bhash2_conflict(sk, sk2, sk_uid, relax,
 223                                            reuseport_cb_ok, reuseport_ok))
 224                         return true;
 225         }
 226
 227         return false;
 228 }
 229
 230 #define sk_for_each_bound_bhash(__sk, __tb2, __tb)                      \
 231         hlist_for_each_entry(__tb2, &(__tb)->bhash2, bhash_node)        \
 232                 sk_for_each_bound(sk2, &(__tb2)->owners)
 233
 234 /* This should be called only when the tb and tb2 hashbuckets' locks are held */
 235 static int inet_csk_bind_conflict(const struct sock *sk,
 236                                   const struct inet_bind_bucket *tb,
 237                                   const struct inet_bind2_bucket *tb2, /* may be null */
 238                                   bool relax, bool reuseport_ok)
 239 {
 240         kuid_t uid = sock_i_uid((struct sock *)sk);
 241         struct sock_reuseport *reuseport_cb;
 242         bool reuseport_cb_ok;
 243         struct sock *sk2;
 244
 245         rcu_read_lock();
 246         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 247         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 248         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 249         rcu_read_unlock();
 250
 251         /* Conflicts with an existing IPV6_ADDR_ANY (if ipv6) or INADDR_ANY (if
 252          * ipv4) should have been checked already. We need to do these two
 253          * checks separately because their spinlocks have to be acquired/released
 254          * independently of each other, to prevent possible deadlocks
 255          */
 256         if (inet_use_bhash2_on_bind(sk))
 257                 return tb2 && inet_bhash2_conflict(sk, tb2, uid, relax,
 258                                                    reuseport_cb_ok, reuseport_ok);
 259
 260         /* Unlike other sk lookup places we do not check
 261          * for sk_net here, since _all_ the socks listed
 262          * in tb->owners and tb2->owners list belong
 263          * to the same net - the one this bucket belongs to.
 264          */
 265         sk_for_each_bound_bhash(sk2, tb2, tb) {
 266                 if (!inet_bind_conflict(sk, sk2, uid, relax, reuseport_cb_ok, reuseport_ok))
 267                         continue;
 268
 269                 if (inet_rcv_saddr_equal(sk, sk2, true))
 270                         return true;
 271         }
 272
 273         return false;
 274 }
 275
 276 /* Determine if there is a bind conflict with an existing IPV6_ADDR_ANY (if ipv6) or
 277  * INADDR_ANY (if ipv4) socket.
 278  *
 279  * Caller must hold bhash hashbucket lock with local bh disabled, to protect
 280  * against concurrent binds on the port for addr any
 281  */
 282 static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l3mdev,
 283                                           bool relax, bool reuseport_ok)
 284 {
 285         kuid_t uid = sock_i_uid((struct sock *)sk);
 286         const struct net *net = sock_net(sk);
 287         struct sock_reuseport *reuseport_cb;
 288         struct inet_bind_hashbucket *head2;
 289         struct inet_bind2_bucket *tb2;
 290         bool reuseport_cb_ok;
 291
 292         rcu_read_lock();
 293         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 294         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 295         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 296         rcu_read_unlock();
 297
 298         head2 = inet_bhash2_addr_any_hashbucket(sk, net, port);
 299
 300         spin_lock(&head2->lock);
 301
 302         inet_bind_bucket_for_each(tb2, &head2->chain)
 303                 if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
 304                         break;
 305
 306         if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
 307                                         reuseport_ok)) {
 308                 spin_unlock(&head2->lock);
 309                 return true;
 310         }
 311
 312         spin_unlock(&head2->lock);
 313         return false;
 314 }
 315
 316 /*
 317  * Find an open port number for the socket.  Returns with the
 318  * inet_bind_hashbucket locks held if successful.
 319  */
 320 static struct inet_bind_hashbucket *
 321 inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 322                         struct inet_bind2_bucket **tb2_ret,
 323                         struct inet_bind_hashbucket **head2_ret, int *port_ret)
 324 {
 325         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 326         int i, low, high, attempt_half, port, l3mdev;
 327         struct inet_bind_hashbucket *head, *head2;
 328         struct net *net = sock_net(sk);
 329         struct inet_bind2_bucket *tb2;
 330         struct inet_bind_bucket *tb;
 331         u32 remaining, offset;
 332         bool relax = false;
 333
 334         l3mdev = inet_sk_bound_l3mdev(sk);
 335 ports_exhausted:
 336         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 337 other_half_scan:
 338         inet_sk_get_local_port_range(sk, &low, &high);
 339         high++; /* [32768, 60999] -> [32768, 61000[ */
 340         if (high - low < 4)
 341                 attempt_half = 0;
 342         if (attempt_half) {
 343                 int half = low + (((high - low) >> 2) << 1);
 344
 345                 if (attempt_half == 1)
 346                         high = half;
 347                 else
 348                         low = half;
 349         }
 350         remaining = high - low;
 351         if (likely(remaining > 1))
 352                 remaining &= ~1U;
 353
 354         offset = get_random_u32_below(remaining);
 355         /* __inet_hash_connect() favors ports having @low parity
 356          * We do the opposite to not pollute connect() users.
 357          */
 358         offset |= 1U;
 359
 360 other_parity_scan:
 361         port = low + offset;
 362         for (i = 0; i < remaining; i += 2, port += 2) {
 363                 if (unlikely(port >= high))
 364                         port -= remaining;
 365                 if (inet_is_local_reserved_port(net, port))
 366                         continue;
 367                 head = &hinfo->bhash[inet_bhashfn(net, port,
 368                                                   hinfo->bhash_size)];
 369                 spin_lock_bh(&head->lock);
 370                 if (inet_use_bhash2_on_bind(sk)) {
 371                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, relax, false))
 372                                 goto next_port;
 373                 }
 374
 375                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 376                 spin_lock(&head2->lock);
 377                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 378                 inet_bind_bucket_for_each(tb, &head->chain)
 379                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
 380                                 if (!inet_csk_bind_conflict(sk, tb, tb2,
 381                                                             relax, false))
 382                                         goto success;
 383                                 spin_unlock(&head2->lock);
 384                                 goto next_port;
 385                         }
 386                 tb = NULL;
 387                 goto success;
 388 next_port:
 389                 spin_unlock_bh(&head->lock);
 390                 cond_resched();
 391         }
 392
 393         offset--;
 394         if (!(offset & 1))
 395                 goto other_parity_scan;
 396
 397         if (attempt_half == 1) {
 398                 /* OK we now try the upper half of the range */
 399                 attempt_half = 2;
 400                 goto other_half_scan;
 401         }
 402
 403         if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) {
 404                 /* We still have a chance to connect to different destinations */
 405                 relax = true;
 406                 goto ports_exhausted;
 407         }
 408         return NULL;
 409 success:
 410         *port_ret = port;
 411         *tb_ret = tb;
 412         *tb2_ret = tb2;
 413         *head2_ret = head2;
 414         return head;
 415 }
 416
 417 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 418                                      struct sock *sk)
 419 {
 420         kuid_t uid = sock_i_uid(sk);
 421
 422         if (tb->fastreuseport <= 0)
 423                 return 0;
 424         if (!sk->sk_reuseport)
 425                 return 0;
 426         if (rcu_access_pointer(sk->sk_reuseport_cb))
 427                 return 0;
 428         if (!uid_eq(tb->fastuid, uid))
 429                 return 0;
 430         /* We only need to check the rcv_saddr if this tb was once marked
 431          * without fastreuseport and then was reset, as we can only know that
 432          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 433          * owners list.
 434          */
 435         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 436                 return 1;
 437 #if IS_ENABLED(CONFIG_IPV6)
 438         if (tb->fast_sk_family == AF_INET6)
 439                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 440                                             inet6_rcv_saddr(sk),
 441                                             tb->fast_rcv_saddr,
 442                                             sk->sk_rcv_saddr,
 443                                             tb->fast_ipv6_only,
 444                                             ipv6_only_sock(sk), true, false);
 445 #endif
 446         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 447                                     ipv6_only_sock(sk), true, false);
 448 }
 449
 450 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 451                                struct sock *sk)
 452 {
 453         kuid_t uid = sock_i_uid(sk);
 454         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 455
 456         if (hlist_empty(&tb->bhash2)) {
 457                 tb->fastreuse = reuse;
 458                 if (sk->sk_reuseport) {
 459                         tb->fastreuseport = FASTREUSEPORT_ANY;
 460                         tb->fastuid = uid;
 461                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 462                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 463                         tb->fast_sk_family = sk->sk_family;
 464 #if IS_ENABLED(CONFIG_IPV6)
 465                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 466 #endif
 467                 } else {
 468                         tb->fastreuseport = 0;
 469                 }
 470         } else {
 471                 if (!reuse)
 472                         tb->fastreuse = 0;
 473                 if (sk->sk_reuseport) {
 474                         /* We didn't match or we don't have fastreuseport set on
 475                          * the tb, but we have sk_reuseport set on this socket
 476                          * and we know that there are no bind conflicts with
 477                          * this socket in this tb, so reset our tb's reuseport
 478                          * settings so that any subsequent sockets that match
 479                          * our current socket will be put on the fast path.
 480                          *
 481                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 482                          * do extra checking for all subsequent sk_reuseport
 483                          * socks.
 484                          */
 485                         if (!sk_reuseport_match(tb, sk)) {
 486                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 487                                 tb->fastuid = uid;
 488                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 489                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 490                                 tb->fast_sk_family = sk->sk_family;
 491 #if IS_ENABLED(CONFIG_IPV6)
 492                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 493 #endif
 494                         }
 495                 } else {
 496                         tb->fastreuseport = 0;
 497                 }
 498         }
 499 }
 500
 501 /* Obtain a reference to a local port for the given sock,
 502  * if snum is zero it means select any available local port.
 503  * We try to allocate an odd port (and leave even ports for connect())
 504  */
 505 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 506 {
 507         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 508         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 509         bool found_port = false, check_bind_conflict = true;
 510         bool bhash_created = false, bhash2_created = false;
 511         int ret = -EADDRINUSE, port = snum, l3mdev;
 512         struct inet_bind_hashbucket *head, *head2;
 513         struct inet_bind2_bucket *tb2 = NULL;
 514         struct inet_bind_bucket *tb = NULL;
 515         bool head2_lock_acquired = false;
 516         struct net *net = sock_net(sk);
 517
 518         l3mdev = inet_sk_bound_l3mdev(sk);
 519
 520         if (!port) {
 521                 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port);
 522                 if (!head)
 523                         return ret;
 524
 525                 head2_lock_acquired = true;
 526
 527                 if (tb && tb2)
 528                         goto success;
 529                 found_port = true;
 530         } else {
 531                 head = &hinfo->bhash[inet_bhashfn(net, port,
 532                                                   hinfo->bhash_size)];
 533                 spin_lock_bh(&head->lock);
 534                 inet_bind_bucket_for_each(tb, &head->chain)
 535                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 536                                 break;
 537         }
 538
 539         if (!tb) {
 540                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net,
 541                                              head, port, l3mdev);
 542                 if (!tb)
 543                         goto fail_unlock;
 544                 bhash_created = true;
 545         }
 546
 547         if (!found_port) {
 548                 if (!hlist_empty(&tb->bhash2)) {
 549                         if (sk->sk_reuse == SK_FORCE_REUSE ||
 550                             (tb->fastreuse > 0 && reuse) ||
 551                             sk_reuseport_match(tb, sk))
 552                                 check_bind_conflict = false;
 553                 }
 554
 555                 if (check_bind_conflict && inet_use_bhash2_on_bind(sk)) {
 556                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, true, true))
 557                                 goto fail_unlock;
 558                 }
 559
 560                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 561                 spin_lock(&head2->lock);
 562                 head2_lock_acquired = true;
 563                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 564         }
 565
 566         if (!tb2) {
 567                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep,
 568                                                net, head2, tb, sk);
 569                 if (!tb2)
 570                         goto fail_unlock;
 571                 bhash2_created = true;
 572         }
 573
 574         if (!found_port && check_bind_conflict) {
 575                 if (inet_csk_bind_conflict(sk, tb, tb2, true, true))
 576                         goto fail_unlock;
 577         }
 578
 579 success:
 580         inet_csk_update_fastreuse(tb, sk);
 581
 582         if (!inet_csk(sk)->icsk_bind_hash)
 583                 inet_bind_hash(sk, tb, tb2, port);
 584         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 585         WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
 586         ret = 0;
 587
 588 fail_unlock:
 589         if (ret) {
 590                 if (bhash2_created)
 591                         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, tb2);
 592                 if (bhash_created)
 593                         inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
 594         }
 595         if (head2_lock_acquired)
 596                 spin_unlock(&head2->lock);
 597         spin_unlock_bh(&head->lock);
 598         return ret;
 599 }
 600 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 601
 602 /*
 603  * Wait for an incoming connection, avoid race conditions. This must be called
 604  * with the socket locked.
 605  */
 606 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 607 {
 608         struct inet_connection_sock *icsk = inet_csk(sk);
 609         DEFINE_WAIT(wait);
 610         int err;
 611
 612         /*
 613          * True wake-one mechanism for incoming connections: only
 614          * one process gets woken up, not the 'whole herd'.
 615          * Since we do not 'race & poll' for established sockets
 616          * anymore, the common case will execute the loop only once.
 617          *
 618          * Subtle issue: "add_wait_queue_exclusive()" will be added
 619          * after any current non-exclusive waiters, and we know that
 620          * it will always _stay_ after any new non-exclusive waiters
 621          * because all non-exclusive waiters are added at the
 622          * beginning of the wait-queue. As such, it's ok to "drop"
 623          * our exclusiveness temporarily when we get woken up without
 624          * having to remove and re-insert us on the wait queue.
 625          */
 626         for (;;) {
 627                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 628                                           TASK_INTERRUPTIBLE);
 629                 release_sock(sk);
 630                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 631                         timeo = schedule_timeout(timeo);
 632                 sched_annotate_sleep();
 633                 lock_sock(sk);
 634                 err = 0;
 635                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 636                         break;
 637                 err = -EINVAL;
 638                 if (sk->sk_state != TCP_LISTEN)
 639                         break;
 640                 err = sock_intr_errno(timeo);
 641                 if (signal_pending(current))
 642                         break;
 643                 err = -EAGAIN;
 644                 if (!timeo)
 645                         break;
 646         }
 647         finish_wait(sk_sleep(sk), &wait);
 648         return err;
 649 }
 650
 651 /*
 652  * This will accept the next outstanding connection.
 653  */
 654 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 655 {
 656         struct inet_connection_sock *icsk = inet_csk(sk);
 657         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 658         struct request_sock *req;
 659         struct sock *newsk;
 660         int error;
 661
 662         lock_sock(sk);
 663
 664         /* We need to make sure that this socket is listening,
 665          * and that it has something pending.
 666          */
 667         error = -EINVAL;
 668         if (sk->sk_state != TCP_LISTEN)
 669                 goto out_err;
 670
 671         /* Find already established connection */
 672         if (reqsk_queue_empty(queue)) {
 673                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 674
 675                 /* If this is a non blocking socket don't sleep */
 676                 error = -EAGAIN;
 677                 if (!timeo)
 678                         goto out_err;
 679
 680                 error = inet_csk_wait_for_connect(sk, timeo);
 681                 if (error)
 682                         goto out_err;
 683         }
 684         req = reqsk_queue_remove(queue, sk);
 685         newsk = req->sk;
 686
 687         if (sk->sk_protocol == IPPROTO_TCP &&
 688             tcp_rsk(req)->tfo_listener) {
 689                 spin_lock_bh(&queue->fastopenq.lock);
 690                 if (tcp_rsk(req)->tfo_listener) {
 691                         /* We are still waiting for the final ACK from 3WHS
 692                          * so can't free req now. Instead, we set req->sk to
 693                          * NULL to signify that the child socket is taken
 694                          * so reqsk_fastopen_remove() will free the req
 695                          * when 3WHS finishes (or is aborted).
 696                          */
 697                         req->sk = NULL;
 698                         req = NULL;
 699                 }
 700                 spin_unlock_bh(&queue->fastopenq.lock);
 701         }
 702
 703 out:
 704         release_sock(sk);
 705         if (newsk && mem_cgroup_sockets_enabled) {
 706                 int amt = 0;
 707
 708                 /* atomically get the memory usage, set and charge the
 709                  * newsk->sk_memcg.
 710                  */
 711                 lock_sock(newsk);
 712
 713                 mem_cgroup_sk_alloc(newsk);
 714                 if (newsk->sk_memcg) {
 715                         /* The socket has not been accepted yet, no need
 716                          * to look at newsk->sk_wmem_queued.
 717                          */
 718                         amt = sk_mem_pages(newsk->sk_forward_alloc +
 719                                            atomic_read(&newsk->sk_rmem_alloc));
 720                 }
 721
 722                 if (amt)
 723                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt,
 724                                                 GFP_KERNEL | __GFP_NOFAIL);
 725
 726                 release_sock(newsk);
 727         }
 728         if (req)
 729                 reqsk_put(req);
 730         return newsk;
 731 out_err:
 732         newsk = NULL;
 733         req = NULL;
 734         *err = error;
 735         goto out;
 736 }
 737 EXPORT_SYMBOL(inet_csk_accept);
 738
 739 /*
 740  * Using different timers for retransmit, delayed acks and probes
 741  * We may wish use just one timer maintaining a list of expire jiffies
 742  * to optimize.
 743  */
 744 void inet_csk_init_xmit_timers(struct sock *sk,
 745                                void (*retransmit_handler)(struct timer_list *t),
 746                                void (*delack_handler)(struct timer_list *t),
 747                                void (*keepalive_handler)(struct timer_list *t))
 748 {
 749         struct inet_connection_sock *icsk = inet_csk(sk);
 750
 751         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 752         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 753         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 754         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 755 }
 756 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 757
 758 void inet_csk_clear_xmit_timers(struct sock *sk)
 759 {
 760         struct inet_connection_sock *icsk = inet_csk(sk);
 761
 762         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 763
 764         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 765         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 766         sk_stop_timer(sk, &sk->sk_timer);
 767 }
 768 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 769
 770 void inet_csk_delete_keepalive_timer(struct sock *sk)
 771 {
 772         sk_stop_timer(sk, &sk->sk_timer);
 773 }
 774 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 775
 776 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 777 {
 778         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 779 }
 780 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 781
 782 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 783                                      struct flowi4 *fl4,
 784                                      const struct request_sock *req)
 785 {
 786         const struct inet_request_sock *ireq = inet_rsk(req);
 787         struct net *net = read_pnet(&ireq->ireq_net);
 788         struct ip_options_rcu *opt;
 789         struct rtable *rt;
 790
 791         rcu_read_lock();
 792         opt = rcu_dereference(ireq->ireq_opt);
 793
 794         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 795                            ip_sock_rt_tos(sk), ip_sock_rt_scope(sk),
 796                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 797                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 798                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 799                            htons(ireq->ir_num), sk->sk_uid);
 800         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 801         rt = ip_route_output_flow(net, fl4, sk);
 802         if (IS_ERR(rt))
 803                 goto no_route;
 804         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 805                 goto route_err;
 806         rcu_read_unlock();
 807         return &rt->dst;
 808
 809 route_err:
 810         ip_rt_put(rt);
 811 no_route:
 812         rcu_read_unlock();
 813         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 814         return NULL;
 815 }
 816 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 817
 818 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 819                                             struct sock *newsk,
 820                                             const struct request_sock *req)
 821 {
 822         const struct inet_request_sock *ireq = inet_rsk(req);
 823         struct net *net = read_pnet(&ireq->ireq_net);
 824         struct inet_sock *newinet = inet_sk(newsk);
 825         struct ip_options_rcu *opt;
 826         struct flowi4 *fl4;
 827         struct rtable *rt;
 828
 829         opt = rcu_dereference(ireq->ireq_opt);
 830         fl4 = &newinet->cork.fl.u.ip4;
 831
 832         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 833                            ip_sock_rt_tos(sk), ip_sock_rt_scope(sk),
 834                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 835                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 836                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 837                            htons(ireq->ir_num), sk->sk_uid);
 838         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 839         rt = ip_route_output_flow(net, fl4, sk);
 840         if (IS_ERR(rt))
 841                 goto no_route;
 842         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 843                 goto route_err;
 844         return &rt->dst;
 845
 846 route_err:
 847         ip_rt_put(rt);
 848 no_route:
 849         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 850         return NULL;
 851 }
 852 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 853
 854 /* Decide when to expire the request and when to resend SYN-ACK */
 855 static void syn_ack_recalc(struct request_sock *req,
 856                            const int max_syn_ack_retries,
 857                            const u8 rskq_defer_accept,
 858                            int *expire, int *resend)
 859 {
 860         if (!rskq_defer_accept) {
 861                 *expire = req->num_timeout >= max_syn_ack_retries;
 862                 *resend = 1;
 863                 return;
 864         }
 865         *expire = req->num_timeout >= max_syn_ack_retries &&
 866                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 867         /* Do not resend while waiting for data after ACK,
 868          * start to resend on end of deferring period to give
 869          * last chance for data or ACK to create established socket.
 870          */
 871         *resend = !inet_rsk(req)->acked ||
 872                   req->num_timeout >= rskq_defer_accept - 1;
 873 }
 874
 875 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 876 {
 877         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 878
 879         if (!err)
 880                 req->num_retrans++;
 881         return err;
 882 }
 883 EXPORT_SYMBOL(inet_rtx_syn_ack);
 884
 885 static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 886                                              struct sock *sk)
 887 {
 888         struct sock *req_sk, *nreq_sk;
 889         struct request_sock *nreq;
 890
 891         nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN);
 892         if (!nreq) {
 893                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
 894
 895                 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */
 896                 sock_put(sk);
 897                 return NULL;
 898         }
 899
 900         req_sk = req_to_sk(req);
 901         nreq_sk = req_to_sk(nreq);
 902
 903         memcpy(nreq_sk, req_sk,
 904                offsetof(struct sock, sk_dontcopy_begin));
 905         memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end,
 906                req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end));
 907
 908         sk_node_init(&nreq_sk->sk_node);
 909         nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping;
 910 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
 911         nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping;
 912 #endif
 913         nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu;
 914
 915         nreq->rsk_listener = sk;
 916
 917         /* We need not acquire fastopenq->lock
 918          * because the child socket is locked in inet_csk_listen_stop().
 919          */
 920         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener)
 921                 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
 922
 923         return nreq;
 924 }
 925
 926 static void reqsk_queue_migrated(struct request_sock_queue *queue,
 927                                  const struct request_sock *req)
 928 {
 929         if (req->num_timeout == 0)
 930                 atomic_inc(&queue->young);
 931         atomic_inc(&queue->qlen);
 932 }
 933
 934 static void reqsk_migrate_reset(struct request_sock *req)
 935 {
 936         req->saved_syn = NULL;
 937 #if IS_ENABLED(CONFIG_IPV6)
 938         inet_rsk(req)->ipv6_opt = NULL;
 939         inet_rsk(req)->pktopts = NULL;
 940 #else
 941         inet_rsk(req)->ireq_opt = NULL;
 942 #endif
 943 }
 944
 945 /* return true if req was found in the ehash table */
 946 static bool reqsk_queue_unlink(struct request_sock *req)
 947 {
 948         struct sock *sk = req_to_sk(req);
 949         bool found = false;
 950
 951         if (sk_hashed(sk)) {
 952                 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 953                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 954
 955                 spin_lock(lock);
 956                 found = __sk_nulls_del_node_init_rcu(sk);
 957                 spin_unlock(lock);
 958         }
 959         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 960                 reqsk_put(req);
 961         return found;
 962 }
 963
 964 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 965 {
 966         bool unlinked = reqsk_queue_unlink(req);
 967
 968         if (unlinked) {
 969                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 970                 reqsk_put(req);
 971         }
 972         return unlinked;
 973 }
 974 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 975
 976 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 977 {
 978         inet_csk_reqsk_queue_drop(sk, req);
 979         reqsk_put(req);
 980 }
 981 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 982
 983 static void reqsk_timer_handler(struct timer_list *t)
 984 {
 985         struct request_sock *req = from_timer(req, t, rsk_timer);
 986         struct request_sock *nreq = NULL, *oreq = req;
 987         struct sock *sk_listener = req->rsk_listener;
 988         struct inet_connection_sock *icsk;
 989         struct request_sock_queue *queue;
 990         struct net *net;
 991         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
 992
 993         if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
 994                 struct sock *nsk;
 995
 996                 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
 997                 if (!nsk)
 998                         goto drop;
 999
1000                 nreq = inet_reqsk_clone(req, nsk);
1001                 if (!nreq)
1002                         goto drop;
1003
1004                 /* The new timer for the cloned req can decrease the 2
1005                  * by calling inet_csk_reqsk_queue_drop_and_put(), so
1006                  * hold another count to prevent use-after-free and
1007                  * call reqsk_put() just before return.
1008                  */
1009                 refcount_set(&nreq->rsk_refcnt, 2 + 1);
1010                 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1011                 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
1012
1013                 req = nreq;
1014                 sk_listener = nsk;
1015         }
1016
1017         icsk = inet_csk(sk_listener);
1018         net = sock_net(sk_listener);
1019         max_syn_ack_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
1020                 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries);
1021         /* Normally all the openreqs are young and become mature
1022          * (i.e. converted to established socket) for first timeout.
1023          * If synack was not acknowledged for 1 second, it means
1024          * one of the following things: synack was lost, ack was lost,
1025          * rtt is high or nobody planned to ack (i.e. synflood).
1026          * When server is a bit loaded, queue is populated with old
1027          * open requests, reducing effective size of queue.
1028          * When server is well loaded, queue size reduces to zero
1029          * after several minutes of work. It is not synflood,
1030          * it is normal operation. The solution is pruning
1031          * too old entries overriding normal timeout, when
1032          * situation becomes dangerous.
1033          *
1034          * Essentially, we reserve half of room for young
1035          * embrions; and abort old ones without pity, if old
1036          * ones are about to clog our table.
1037          */
1038         queue = &icsk->icsk_accept_queue;
1039         qlen = reqsk_queue_len(queue);
1040         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
1041                 int young = reqsk_queue_len_young(queue) << 1;
1042
1043                 while (max_syn_ack_retries > 2) {
1044                         if (qlen < young)
1045                                 break;
1046                         max_syn_ack_retries--;
1047                         young <<= 1;
1048                 }
1049         }
1050         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
1051                        &expire, &resend);
1052         req->rsk_ops->syn_ack_timeout(req);
1053         if (!expire &&
1054             (!resend ||
1055              !inet_rtx_syn_ack(sk_listener, req) ||
1056              inet_rsk(req)->acked)) {
1057                 if (req->num_timeout++ == 0)
1058                         atomic_dec(&queue->young);
1059                 mod_timer(&req->rsk_timer, jiffies + reqsk_timeout(req, TCP_RTO_MAX));
1060
1061                 if (!nreq)
1062                         return;
1063
1064                 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
1065                         /* delete timer */
1066                         inet_csk_reqsk_queue_drop(sk_listener, nreq);
1067                         goto no_ownership;
1068                 }
1069
1070                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS);
1071                 reqsk_migrate_reset(oreq);
1072                 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
1073                 reqsk_put(oreq);
1074
1075                 reqsk_put(nreq);
1076                 return;
1077         }
1078
1079         /* Even if we can clone the req, we may need not retransmit any more
1080          * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
1081          * CPU may win the "own_req" race so that inet_ehash_insert() fails.
1082          */
1083         if (nreq) {
1084                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE);
1085 no_ownership:
1086                 reqsk_migrate_reset(nreq);
1087                 reqsk_queue_removed(queue, nreq);
1088                 __reqsk_free(nreq);
1089         }
1090
1091 drop:
1092         inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
1093 }
1094
1095 static void reqsk_queue_hash_req(struct request_sock *req,
1096                                  unsigned long timeout)
1097 {
1098         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1099         mod_timer(&req->rsk_timer, jiffies + timeout);
1100
1101         inet_ehash_insert(req_to_sk(req), NULL, NULL);
1102         /* before letting lookups find us, make sure all req fields
1103          * are committed to memory and refcnt initialized.
1104          */
1105         smp_wmb();
1106         refcount_set(&req->rsk_refcnt, 2 + 1);
1107 }
1108
1109 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
1110                                    unsigned long timeout)
1111 {
1112         reqsk_queue_hash_req(req, timeout);
1113         inet_csk_reqsk_queue_added(sk);
1114 }
1115 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
1116
1117 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
1118                            const gfp_t priority)
1119 {
1120         struct inet_connection_sock *icsk = inet_csk(newsk);
1121
1122         if (!icsk->icsk_ulp_ops)
1123                 return;
1124
1125         icsk->icsk_ulp_ops->clone(req, newsk, priority);
1126 }
1127
1128 /**
1129  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
1130  *      @sk: the socket to clone
1131  *      @req: request_sock
1132  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
1133  *
1134  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
1135  */
1136 struct sock *inet_csk_clone_lock(const struct sock *sk,
1137                                  const struct request_sock *req,
1138                                  const gfp_t priority)
1139 {
1140         struct sock *newsk = sk_clone_lock(sk, priority);
1141
1142         if (newsk) {
1143                 struct inet_connection_sock *newicsk = inet_csk(newsk);
1144
1145                 inet_sk_set_state(newsk, TCP_SYN_RECV);
1146                 newicsk->icsk_bind_hash = NULL;
1147                 newicsk->icsk_bind2_hash = NULL;
1148
1149                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
1150                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
1151                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
1152
1153                 /* listeners have SOCK_RCU_FREE, not the children */
1154                 sock_reset_flag(newsk, SOCK_RCU_FREE);
1155
1156                 inet_sk(newsk)->mc_list = NULL;
1157
1158                 newsk->sk_mark = inet_rsk(req)->ir_mark;
1159                 atomic64_set(&newsk->sk_cookie,
1160                              atomic64_read(&inet_rsk(req)->ir_cookie));
1161
1162                 newicsk->icsk_retransmits = 0;
1163                 newicsk->icsk_backoff     = 0;
1164                 newicsk->icsk_probes_out  = 0;
1165                 newicsk->icsk_probes_tstamp = 0;
1166
1167                 /* Deinitialize accept_queue to trap illegal accesses. */
1168                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
1169
1170                 inet_clone_ulp(req, newsk, priority);
1171
1172                 security_inet_csk_clone(newsk, req);
1173         }
1174         return newsk;
1175 }
1176 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
1177
1178 /*
1179  * At this point, there should be no process reference to this
1180  * socket, and thus no user references at all.  Therefore we
1181  * can assume the socket waitqueue is inactive and nobody will
1182  * try to jump onto it.
1183  */
1184 void inet_csk_destroy_sock(struct sock *sk)
1185 {
1186         WARN_ON(sk->sk_state != TCP_CLOSE);
1187         WARN_ON(!sock_flag(sk, SOCK_DEAD));
1188
1189         /* It cannot be in hash table! */
1190         WARN_ON(!sk_unhashed(sk));
1191
1192         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
1193         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
1194
1195         sk->sk_prot->destroy(sk);
1196
1197         sk_stream_kill_queues(sk);
1198
1199         xfrm_sk_free_policy(sk);
1200
1201         this_cpu_dec(*sk->sk_prot->orphan_count);
1202
1203         sock_put(sk);
1204 }
1205 EXPORT_SYMBOL(inet_csk_destroy_sock);
1206
1207 /* This function allows to force a closure of a socket after the call to
1208  * tcp/dccp_create_openreq_child().
1209  */
1210 void inet_csk_prepare_forced_close(struct sock *sk)
1211         __releases(&sk->sk_lock.slock)
1212 {
1213         /* sk_clone_lock locked the socket and set refcnt to 2 */
1214         bh_unlock_sock(sk);
1215         sock_put(sk);
1216         inet_csk_prepare_for_destroy_sock(sk);
1217         inet_sk(sk)->inet_num = 0;
1218 }
1219 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
1220
1221 static int inet_ulp_can_listen(const struct sock *sk)
1222 {
1223         const struct inet_connection_sock *icsk = inet_csk(sk);
1224
1225         if (icsk->icsk_ulp_ops && !icsk->icsk_ulp_ops->clone)
1226                 return -EINVAL;
1227
1228         return 0;
1229 }
1230
1231 int inet_csk_listen_start(struct sock *sk)
1232 {
1233         struct inet_connection_sock *icsk = inet_csk(sk);
1234         struct inet_sock *inet = inet_sk(sk);
1235         int err;
1236
1237         err = inet_ulp_can_listen(sk);
1238         if (unlikely(err))
1239                 return err;
1240
1241         reqsk_queue_alloc(&icsk->icsk_accept_queue);
1242
1243         sk->sk_ack_backlog = 0;
1244         inet_csk_delack_init(sk);
1245
1246         /* There is race window here: we announce ourselves listening,
1247          * but this transition is still not validated by get_port().
1248          * It is OK, because this socket enters to hash table only
1249          * after validation is complete.
1250          */
1251         inet_sk_state_store(sk, TCP_LISTEN);
1252         err = sk->sk_prot->get_port(sk, inet->inet_num);
1253         if (!err) {
1254                 inet->inet_sport = htons(inet->inet_num);
1255
1256                 sk_dst_reset(sk);
1257                 err = sk->sk_prot->hash(sk);
1258
1259                 if (likely(!err))
1260                         return 0;
1261         }
1262
1263         inet_sk_set_state(sk, TCP_CLOSE);
1264         return err;
1265 }
1266 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
1267
1268 static void inet_child_forget(struct sock *sk, struct request_sock *req,
1269                               struct sock *child)
1270 {
1271         sk->sk_prot->disconnect(child, O_NONBLOCK);
1272
1273         sock_orphan(child);
1274
1275         this_cpu_inc(*sk->sk_prot->orphan_count);
1276
1277         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
1278                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
1279                 BUG_ON(sk != req->rsk_listener);
1280
1281                 /* Paranoid, to prevent race condition if
1282                  * an inbound pkt destined for child is
1283                  * blocked by sock lock in tcp_v4_rcv().
1284                  * Also to satisfy an assertion in
1285                  * tcp_v4_destroy_sock().
1286                  */
1287                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1288         }
1289         inet_csk_destroy_sock(child);
1290 }
1291
1292 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1293                                       struct request_sock *req,
1294                                       struct sock *child)
1295 {
1296         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1297
1298         spin_lock(&queue->rskq_lock);
1299         if (unlikely(sk->sk_state != TCP_LISTEN)) {
1300                 inet_child_forget(sk, req, child);
1301                 child = NULL;
1302         } else {
1303                 req->sk = child;
1304                 req->dl_next = NULL;
1305                 if (queue->rskq_accept_head == NULL)
1306                         WRITE_ONCE(queue->rskq_accept_head, req);
1307                 else
1308                         queue->rskq_accept_tail->dl_next = req;
1309                 queue->rskq_accept_tail = req;
1310                 sk_acceptq_added(sk);
1311         }
1312         spin_unlock(&queue->rskq_lock);
1313         return child;
1314 }
1315 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1316
1317 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1318                                          struct request_sock *req, bool own_req)
1319 {
1320         if (own_req) {
1321                 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
1322                 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
1323
1324                 if (sk != req->rsk_listener) {
1325                         /* another listening sk has been selected,
1326                          * migrate the req to it.
1327                          */
1328                         struct request_sock *nreq;
1329
1330                         /* hold a refcnt for the nreq->rsk_listener
1331                          * which is assigned in inet_reqsk_clone()
1332                          */
1333                         sock_hold(sk);
1334                         nreq = inet_reqsk_clone(req, sk);
1335                         if (!nreq) {
1336                                 inet_child_forget(sk, req, child);
1337                                 goto child_put;
1338                         }
1339
1340                         refcount_set(&nreq->rsk_refcnt, 1);
1341                         if (inet_csk_reqsk_queue_add(sk, nreq, child)) {
1342                                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS);
1343                                 reqsk_migrate_reset(req);
1344                                 reqsk_put(req);
1345                                 return child;
1346                         }
1347
1348                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
1349                         reqsk_migrate_reset(nreq);
1350                         __reqsk_free(nreq);
1351                 } else if (inet_csk_reqsk_queue_add(sk, req, child)) {
1352                         return child;
1353                 }
1354         }
1355         /* Too bad, another child took ownership of the request, undo. */
1356 child_put:
1357         bh_unlock_sock(child);
1358         sock_put(child);
1359         return NULL;
1360 }
1361 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1362
1363 /*
1364  *      This routine closes sockets which have been at least partially
1365  *      opened, but not yet accepted.
1366  */
1367 void inet_csk_listen_stop(struct sock *sk)
1368 {
1369         struct inet_connection_sock *icsk = inet_csk(sk);
1370         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1371         struct request_sock *next, *req;
1372
1373         /* Following specs, it would be better either to send FIN
1374          * (and enter FIN-WAIT-1, it is normal close)
1375          * or to send active reset (abort).
1376          * Certainly, it is pretty dangerous while synflood, but it is
1377          * bad justification for our negligence 8)
1378          * To be honest, we are not able to make either
1379          * of the variants now.                 --ANK
1380          */
1381         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1382                 struct sock *child = req->sk, *nsk;
1383                 struct request_sock *nreq;
1384
1385                 local_bh_disable();
1386                 bh_lock_sock(child);
1387                 WARN_ON(sock_owned_by_user(child));
1388                 sock_hold(child);
1389
1390                 nsk = reuseport_migrate_sock(sk, child, NULL);
1391                 if (nsk) {
1392                         nreq = inet_reqsk_clone(req, nsk);
1393                         if (nreq) {
1394                                 refcount_set(&nreq->rsk_refcnt, 1);
1395
1396                                 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
1397                                         __NET_INC_STATS(sock_net(nsk),
1398                                                         LINUX_MIB_TCPMIGRATEREQSUCCESS);
1399                                         reqsk_migrate_reset(req);
1400                                 } else {
1401                                         __NET_INC_STATS(sock_net(nsk),
1402                                                         LINUX_MIB_TCPMIGRATEREQFAILURE);
1403                                         reqsk_migrate_reset(nreq);
1404                                         __reqsk_free(nreq);
1405                                 }
1406
1407                                 /* inet_csk_reqsk_queue_add() has already
1408                                  * called inet_child_forget() on failure case.
1409                                  */
1410                                 goto skip_child_forget;
1411                         }
1412                 }
1413
1414                 inet_child_forget(sk, req, child);
1415 skip_child_forget:
1416                 reqsk_put(req);
1417                 bh_unlock_sock(child);
1418                 local_bh_enable();
1419                 sock_put(child);
1420
1421                 cond_resched();
1422         }
1423         if (queue->fastopenq.rskq_rst_head) {
1424                 /* Free all the reqs queued in rskq_rst_head. */
1425                 spin_lock_bh(&queue->fastopenq.lock);
1426                 req = queue->fastopenq.rskq_rst_head;
1427                 queue->fastopenq.rskq_rst_head = NULL;
1428                 spin_unlock_bh(&queue->fastopenq.lock);
1429                 while (req != NULL) {
1430                         next = req->dl_next;
1431                         reqsk_put(req);
1432                         req = next;
1433                 }
1434         }
1435         WARN_ON_ONCE(sk->sk_ack_backlog);
1436 }
1437 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1438
1439 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1440 {
1441         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1442         const struct inet_sock *inet = inet_sk(sk);
1443
1444         sin->sin_family         = AF_INET;
1445         sin->sin_addr.s_addr    = inet->inet_daddr;
1446         sin->sin_port           = inet->inet_dport;
1447 }
1448 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1449
1450 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1451 {
1452         const struct inet_sock *inet = inet_sk(sk);
1453         const struct ip_options_rcu *inet_opt;
1454         __be32 daddr = inet->inet_daddr;
1455         struct flowi4 *fl4;
1456         struct rtable *rt;
1457
1458         rcu_read_lock();
1459         inet_opt = rcu_dereference(inet->inet_opt);
1460         if (inet_opt && inet_opt->opt.srr)
1461                 daddr = inet_opt->opt.faddr;
1462         fl4 = &fl->u.ip4;
1463         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1464                                    inet->inet_saddr, inet->inet_dport,
1465                                    inet->inet_sport, sk->sk_protocol,
1466                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1467         if (IS_ERR(rt))
1468                 rt = NULL;
1469         if (rt)
1470                 sk_setup_caps(sk, &rt->dst);
1471         rcu_read_unlock();
1472
1473         return &rt->dst;
1474 }
1475
1476 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1477 {
1478         struct dst_entry *dst = __sk_dst_check(sk, 0);
1479         struct inet_sock *inet = inet_sk(sk);
1480
1481         if (!dst) {
1482                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1483                 if (!dst)
1484                         goto out;
1485         }
1486         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1487
1488         dst = __sk_dst_check(sk, 0);
1489         if (!dst)
1490                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1491 out:
1492         return dst;
1493 }
1494 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);