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(const 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 void inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high)
 134 {
 135         const struct inet_sock *inet = inet_sk(sk);
 136         const struct net *net = sock_net(sk);
 137         int lo, hi, sk_lo, sk_hi;
 138
 139         inet_get_local_port_range(net, &lo, &hi);
 140
 141         sk_lo = inet->local_port_range.lo;
 142         sk_hi = inet->local_port_range.hi;
 143
 144         if (unlikely(lo <= sk_lo && sk_lo <= hi))
 145                 lo = sk_lo;
 146         if (unlikely(lo <= sk_hi && sk_hi <= hi))
 147                 hi = sk_hi;
 148
 149         *low = lo;
 150         *high = hi;
 151 }
 152 EXPORT_SYMBOL(inet_sk_get_local_port_range);
 153
 154 static bool inet_use_bhash2_on_bind(const struct sock *sk)
 155 {
 156 #if IS_ENABLED(CONFIG_IPV6)
 157         if (sk->sk_family == AF_INET6) {
 158                 int addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 159
 160                 return addr_type != IPV6_ADDR_ANY &&
 161                         addr_type != IPV6_ADDR_MAPPED;
 162         }
 163 #endif
 164         return sk->sk_rcv_saddr != htonl(INADDR_ANY);
 165 }
 166
 167 static bool inet_bind_conflict(const struct sock *sk, struct sock *sk2,
 168                                kuid_t sk_uid, bool relax,
 169                                bool reuseport_cb_ok, bool reuseport_ok)
 170 {
 171         int bound_dev_if2;
 172
 173         if (sk == sk2)
 174                 return false;
 175
 176         bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
 177
 178         if (!sk->sk_bound_dev_if || !bound_dev_if2 ||
 179             sk->sk_bound_dev_if == bound_dev_if2) {
 180                 if (sk->sk_reuse && sk2->sk_reuse &&
 181                     sk2->sk_state != TCP_LISTEN) {
 182                         if (!relax || (!reuseport_ok && sk->sk_reuseport &&
 183                                        sk2->sk_reuseport && reuseport_cb_ok &&
 184                                        (sk2->sk_state == TCP_TIME_WAIT ||
 185                                         uid_eq(sk_uid, sock_i_uid(sk2)))))
 186                                 return true;
 187                 } else if (!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                 }
 193         }
 194         return false;
 195 }
 196
 197 static bool __inet_bhash2_conflict(const struct sock *sk, struct sock *sk2,
 198                                    kuid_t sk_uid, bool relax,
 199                                    bool reuseport_cb_ok, bool reuseport_ok)
 200 {
 201         if (sk->sk_family == AF_INET && ipv6_only_sock(sk2))
 202                 return false;
 203
 204         return inet_bind_conflict(sk, sk2, sk_uid, relax,
 205                                   reuseport_cb_ok, reuseport_ok);
 206 }
 207
 208 static bool inet_bhash2_conflict(const struct sock *sk,
 209                                  const struct inet_bind2_bucket *tb2,
 210                                  kuid_t sk_uid,
 211                                  bool relax, bool reuseport_cb_ok,
 212                                  bool reuseport_ok)
 213 {
 214         struct inet_timewait_sock *tw2;
 215         struct sock *sk2;
 216
 217         sk_for_each_bound_bhash2(sk2, &tb2->owners) {
 218                 if (__inet_bhash2_conflict(sk, sk2, sk_uid, relax,
 219                                            reuseport_cb_ok, reuseport_ok))
 220                         return true;
 221         }
 222
 223         twsk_for_each_bound_bhash2(tw2, &tb2->deathrow) {
 224                 sk2 = (struct sock *)tw2;
 225
 226                 if (__inet_bhash2_conflict(sk, sk2, sk_uid, relax,
 227                                            reuseport_cb_ok, reuseport_ok))
 228                         return true;
 229         }
 230
 231         return false;
 232 }
 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         bool reuseport_cb_ok;
 241         struct sock_reuseport *reuseport_cb;
 242         kuid_t uid = sock_i_uid((struct sock *)sk);
 243
 244         rcu_read_lock();
 245         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 246         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 247         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 248         rcu_read_unlock();
 249
 250         /*
 251          * Unlike other sk lookup places we do not check
 252          * for sk_net here, since _all_ the socks listed
 253          * in tb->owners and tb2->owners list belong
 254          * to the same net - the one this bucket belongs to.
 255          */
 256
 257         if (!inet_use_bhash2_on_bind(sk)) {
 258                 struct sock *sk2;
 259
 260                 sk_for_each_bound(sk2, &tb->owners)
 261                         if (inet_bind_conflict(sk, sk2, uid, relax,
 262                                                reuseport_cb_ok, reuseport_ok) &&
 263                             inet_rcv_saddr_equal(sk, sk2, true))
 264                                 return true;
 265
 266                 return false;
 267         }
 268
 269         /* Conflicts with an existing IPV6_ADDR_ANY (if ipv6) or INADDR_ANY (if
 270          * ipv4) should have been checked already. We need to do these two
 271          * checks separately because their spinlocks have to be acquired/released
 272          * independently of each other, to prevent possible deadlocks
 273          */
 274         return tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
 275                                            reuseport_ok);
 276 }
 277
 278 /* Determine if there is a bind conflict with an existing IPV6_ADDR_ANY (if ipv6) or
 279  * INADDR_ANY (if ipv4) socket.
 280  *
 281  * Caller must hold bhash hashbucket lock with local bh disabled, to protect
 282  * against concurrent binds on the port for addr any
 283  */
 284 static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l3mdev,
 285                                           bool relax, bool reuseport_ok)
 286 {
 287         kuid_t uid = sock_i_uid((struct sock *)sk);
 288         const struct net *net = sock_net(sk);
 289         struct sock_reuseport *reuseport_cb;
 290         struct inet_bind_hashbucket *head2;
 291         struct inet_bind2_bucket *tb2;
 292         bool conflict = false;
 293         bool reuseport_cb_ok;
 294
 295         rcu_read_lock();
 296         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 297         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 298         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 299         rcu_read_unlock();
 300
 301         head2 = inet_bhash2_addr_any_hashbucket(sk, net, port);
 302
 303         spin_lock(&head2->lock);
 304
 305         inet_bind_bucket_for_each(tb2, &head2->chain) {
 306                 if (!inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
 307                         continue;
 308
 309                 if (!inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok, reuseport_ok))
 310                         continue;
 311
 312                 conflict = true;
 313                 break;
 314         }
 315
 316         spin_unlock(&head2->lock);
 317
 318         return conflict;
 319 }
 320
 321 /*
 322  * Find an open port number for the socket.  Returns with the
 323  * inet_bind_hashbucket locks held if successful.
 324  */
 325 static struct inet_bind_hashbucket *
 326 inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 327                         struct inet_bind2_bucket **tb2_ret,
 328                         struct inet_bind_hashbucket **head2_ret, int *port_ret)
 329 {
 330         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 331         int i, low, high, attempt_half, port, l3mdev;
 332         struct inet_bind_hashbucket *head, *head2;
 333         struct net *net = sock_net(sk);
 334         struct inet_bind2_bucket *tb2;
 335         struct inet_bind_bucket *tb;
 336         u32 remaining, offset;
 337         bool relax = false;
 338
 339         l3mdev = inet_sk_bound_l3mdev(sk);
 340 ports_exhausted:
 341         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 342 other_half_scan:
 343         inet_sk_get_local_port_range(sk, &low, &high);
 344         high++; /* [32768, 60999] -> [32768, 61000[ */
 345         if (high - low < 4)
 346                 attempt_half = 0;
 347         if (attempt_half) {
 348                 int half = low + (((high - low) >> 2) << 1);
 349
 350                 if (attempt_half == 1)
 351                         high = half;
 352                 else
 353                         low = half;
 354         }
 355         remaining = high - low;
 356         if (likely(remaining > 1))
 357                 remaining &= ~1U;
 358
 359         offset = prandom_u32_max(remaining);
 360         /* __inet_hash_connect() favors ports having @low parity
 361          * We do the opposite to not pollute connect() users.
 362          */
 363         offset |= 1U;
 364
 365 other_parity_scan:
 366         port = low + offset;
 367         for (i = 0; i < remaining; i += 2, port += 2) {
 368                 if (unlikely(port >= high))
 369                         port -= remaining;
 370                 if (inet_is_local_reserved_port(net, port))
 371                         continue;
 372                 head = &hinfo->bhash[inet_bhashfn(net, port,
 373                                                   hinfo->bhash_size)];
 374                 spin_lock_bh(&head->lock);
 375                 if (inet_use_bhash2_on_bind(sk)) {
 376                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, relax, false))
 377                                 goto next_port;
 378                 }
 379
 380                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 381                 spin_lock(&head2->lock);
 382                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 383                 inet_bind_bucket_for_each(tb, &head->chain)
 384                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
 385                                 if (!inet_csk_bind_conflict(sk, tb, tb2,
 386                                                             relax, false))
 387                                         goto success;
 388                                 spin_unlock(&head2->lock);
 389                                 goto next_port;
 390                         }
 391                 tb = NULL;
 392                 goto success;
 393 next_port:
 394                 spin_unlock_bh(&head->lock);
 395                 cond_resched();
 396         }
 397
 398         offset--;
 399         if (!(offset & 1))
 400                 goto other_parity_scan;
 401
 402         if (attempt_half == 1) {
 403                 /* OK we now try the upper half of the range */
 404                 attempt_half = 2;
 405                 goto other_half_scan;
 406         }
 407
 408         if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) {
 409                 /* We still have a chance to connect to different destinations */
 410                 relax = true;
 411                 goto ports_exhausted;
 412         }
 413         return NULL;
 414 success:
 415         *port_ret = port;
 416         *tb_ret = tb;
 417         *tb2_ret = tb2;
 418         *head2_ret = head2;
 419         return head;
 420 }
 421
 422 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 423                                      struct sock *sk)
 424 {
 425         kuid_t uid = sock_i_uid(sk);
 426
 427         if (tb->fastreuseport <= 0)
 428                 return 0;
 429         if (!sk->sk_reuseport)
 430                 return 0;
 431         if (rcu_access_pointer(sk->sk_reuseport_cb))
 432                 return 0;
 433         if (!uid_eq(tb->fastuid, uid))
 434                 return 0;
 435         /* We only need to check the rcv_saddr if this tb was once marked
 436          * without fastreuseport and then was reset, as we can only know that
 437          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 438          * owners list.
 439          */
 440         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 441                 return 1;
 442 #if IS_ENABLED(CONFIG_IPV6)
 443         if (tb->fast_sk_family == AF_INET6)
 444                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 445                                             inet6_rcv_saddr(sk),
 446                                             tb->fast_rcv_saddr,
 447                                             sk->sk_rcv_saddr,
 448                                             tb->fast_ipv6_only,
 449                                             ipv6_only_sock(sk), true, false);
 450 #endif
 451         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 452                                     ipv6_only_sock(sk), true, false);
 453 }
 454
 455 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 456                                struct sock *sk)
 457 {
 458         kuid_t uid = sock_i_uid(sk);
 459         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 460
 461         if (hlist_empty(&tb->owners)) {
 462                 tb->fastreuse = reuse;
 463                 if (sk->sk_reuseport) {
 464                         tb->fastreuseport = FASTREUSEPORT_ANY;
 465                         tb->fastuid = uid;
 466                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 467                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 468                         tb->fast_sk_family = sk->sk_family;
 469 #if IS_ENABLED(CONFIG_IPV6)
 470                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 471 #endif
 472                 } else {
 473                         tb->fastreuseport = 0;
 474                 }
 475         } else {
 476                 if (!reuse)
 477                         tb->fastreuse = 0;
 478                 if (sk->sk_reuseport) {
 479                         /* We didn't match or we don't have fastreuseport set on
 480                          * the tb, but we have sk_reuseport set on this socket
 481                          * and we know that there are no bind conflicts with
 482                          * this socket in this tb, so reset our tb's reuseport
 483                          * settings so that any subsequent sockets that match
 484                          * our current socket will be put on the fast path.
 485                          *
 486                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 487                          * do extra checking for all subsequent sk_reuseport
 488                          * socks.
 489                          */
 490                         if (!sk_reuseport_match(tb, sk)) {
 491                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 492                                 tb->fastuid = uid;
 493                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 494                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 495                                 tb->fast_sk_family = sk->sk_family;
 496 #if IS_ENABLED(CONFIG_IPV6)
 497                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 498 #endif
 499                         }
 500                 } else {
 501                         tb->fastreuseport = 0;
 502                 }
 503         }
 504 }
 505
 506 /* Obtain a reference to a local port for the given sock,
 507  * if snum is zero it means select any available local port.
 508  * We try to allocate an odd port (and leave even ports for connect())
 509  */
 510 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 511 {
 512         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 513         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 514         bool found_port = false, check_bind_conflict = true;
 515         bool bhash_created = false, bhash2_created = false;
 516         int ret = -EADDRINUSE, port = snum, l3mdev;
 517         struct inet_bind_hashbucket *head, *head2;
 518         struct inet_bind2_bucket *tb2 = NULL;
 519         struct inet_bind_bucket *tb = NULL;
 520         bool head2_lock_acquired = false;
 521         struct net *net = sock_net(sk);
 522
 523         l3mdev = inet_sk_bound_l3mdev(sk);
 524
 525         if (!port) {
 526                 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port);
 527                 if (!head)
 528                         return ret;
 529
 530                 head2_lock_acquired = true;
 531
 532                 if (tb && tb2)
 533                         goto success;
 534                 found_port = true;
 535         } else {
 536                 head = &hinfo->bhash[inet_bhashfn(net, port,
 537                                                   hinfo->bhash_size)];
 538                 spin_lock_bh(&head->lock);
 539                 inet_bind_bucket_for_each(tb, &head->chain)
 540                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 541                                 break;
 542         }
 543
 544         if (!tb) {
 545                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net,
 546                                              head, port, l3mdev);
 547                 if (!tb)
 548                         goto fail_unlock;
 549                 bhash_created = true;
 550         }
 551
 552         if (!found_port) {
 553                 if (!hlist_empty(&tb->owners)) {
 554                         if (sk->sk_reuse == SK_FORCE_REUSE ||
 555                             (tb->fastreuse > 0 && reuse) ||
 556                             sk_reuseport_match(tb, sk))
 557                                 check_bind_conflict = false;
 558                 }
 559
 560                 if (check_bind_conflict && inet_use_bhash2_on_bind(sk)) {
 561                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, true, true))
 562                                 goto fail_unlock;
 563                 }
 564
 565                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 566                 spin_lock(&head2->lock);
 567                 head2_lock_acquired = true;
 568                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 569         }
 570
 571         if (!tb2) {
 572                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep,
 573                                                net, head2, port, l3mdev, sk);
 574                 if (!tb2)
 575                         goto fail_unlock;
 576                 bhash2_created = true;
 577         }
 578
 579         if (!found_port && check_bind_conflict) {
 580                 if (inet_csk_bind_conflict(sk, tb, tb2, true, true))
 581                         goto fail_unlock;
 582         }
 583
 584 success:
 585         inet_csk_update_fastreuse(tb, sk);
 586
 587         if (!inet_csk(sk)->icsk_bind_hash)
 588                 inet_bind_hash(sk, tb, tb2, port);
 589         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 590         WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
 591         ret = 0;
 592
 593 fail_unlock:
 594         if (ret) {
 595                 if (bhash_created)
 596                         inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
 597                 if (bhash2_created)
 598                         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep,
 599                                                   tb2);
 600         }
 601         if (head2_lock_acquired)
 602                 spin_unlock(&head2->lock);
 603         spin_unlock_bh(&head->lock);
 604         return ret;
 605 }
 606 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 607
 608 /*
 609  * Wait for an incoming connection, avoid race conditions. This must be called
 610  * with the socket locked.
 611  */
 612 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 613 {
 614         struct inet_connection_sock *icsk = inet_csk(sk);
 615         DEFINE_WAIT(wait);
 616         int err;
 617
 618         /*
 619          * True wake-one mechanism for incoming connections: only
 620          * one process gets woken up, not the 'whole herd'.
 621          * Since we do not 'race & poll' for established sockets
 622          * anymore, the common case will execute the loop only once.
 623          *
 624          * Subtle issue: "add_wait_queue_exclusive()" will be added
 625          * after any current non-exclusive waiters, and we know that
 626          * it will always _stay_ after any new non-exclusive waiters
 627          * because all non-exclusive waiters are added at the
 628          * beginning of the wait-queue. As such, it's ok to "drop"
 629          * our exclusiveness temporarily when we get woken up without
 630          * having to remove and re-insert us on the wait queue.
 631          */
 632         for (;;) {
 633                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 634                                           TASK_INTERRUPTIBLE);
 635                 release_sock(sk);
 636                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 637                         timeo = schedule_timeout(timeo);
 638                 sched_annotate_sleep();
 639                 lock_sock(sk);
 640                 err = 0;
 641                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 642                         break;
 643                 err = -EINVAL;
 644                 if (sk->sk_state != TCP_LISTEN)
 645                         break;
 646                 err = sock_intr_errno(timeo);
 647                 if (signal_pending(current))
 648                         break;
 649                 err = -EAGAIN;
 650                 if (!timeo)
 651                         break;
 652         }
 653         finish_wait(sk_sleep(sk), &wait);
 654         return err;
 655 }
 656
 657 /*
 658  * This will accept the next outstanding connection.
 659  */
 660 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
 661 {
 662         struct inet_connection_sock *icsk = inet_csk(sk);
 663         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 664         struct request_sock *req;
 665         struct sock *newsk;
 666         int error;
 667
 668         lock_sock(sk);
 669
 670         /* We need to make sure that this socket is listening,
 671          * and that it has something pending.
 672          */
 673         error = -EINVAL;
 674         if (sk->sk_state != TCP_LISTEN)
 675                 goto out_err;
 676
 677         /* Find already established connection */
 678         if (reqsk_queue_empty(queue)) {
 679                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 680
 681                 /* If this is a non blocking socket don't sleep */
 682                 error = -EAGAIN;
 683                 if (!timeo)
 684                         goto out_err;
 685
 686                 error = inet_csk_wait_for_connect(sk, timeo);
 687                 if (error)
 688                         goto out_err;
 689         }
 690         req = reqsk_queue_remove(queue, sk);
 691         newsk = req->sk;
 692
 693         if (sk->sk_protocol == IPPROTO_TCP &&
 694             tcp_rsk(req)->tfo_listener) {
 695                 spin_lock_bh(&queue->fastopenq.lock);
 696                 if (tcp_rsk(req)->tfo_listener) {
 697                         /* We are still waiting for the final ACK from 3WHS
 698                          * so can't free req now. Instead, we set req->sk to
 699                          * NULL to signify that the child socket is taken
 700                          * so reqsk_fastopen_remove() will free the req
 701                          * when 3WHS finishes (or is aborted).
 702                          */
 703                         req->sk = NULL;
 704                         req = NULL;
 705                 }
 706                 spin_unlock_bh(&queue->fastopenq.lock);
 707         }
 708
 709 out:
 710         release_sock(sk);
 711         if (newsk && mem_cgroup_sockets_enabled) {
 712                 int amt;
 713
 714                 /* atomically get the memory usage, set and charge the
 715                  * newsk->sk_memcg.
 716                  */
 717                 lock_sock(newsk);
 718
 719                 /* The socket has not been accepted yet, no need to look at
 720                  * newsk->sk_wmem_queued.
 721                  */
 722                 amt = sk_mem_pages(newsk->sk_forward_alloc +
 723                                    atomic_read(&newsk->sk_rmem_alloc));
 724                 mem_cgroup_sk_alloc(newsk);
 725                 if (newsk->sk_memcg && amt)
 726                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt,
 727                                                 GFP_KERNEL | __GFP_NOFAIL);
 728
 729                 release_sock(newsk);
 730         }
 731         if (req)
 732                 reqsk_put(req);
 733
 734         if (newsk)
 735                 inet_init_csk_locks(newsk);
 736
 737         return newsk;
 738 out_err:
 739         newsk = NULL;
 740         req = NULL;
 741         *err = error;
 742         goto out;
 743 }
 744 EXPORT_SYMBOL(inet_csk_accept);
 745
 746 /*
 747  * Using different timers for retransmit, delayed acks and probes
 748  * We may wish use just one timer maintaining a list of expire jiffies
 749  * to optimize.
 750  */
 751 void inet_csk_init_xmit_timers(struct sock *sk,
 752                                void (*retransmit_handler)(struct timer_list *t),
 753                                void (*delack_handler)(struct timer_list *t),
 754                                void (*keepalive_handler)(struct timer_list *t))
 755 {
 756         struct inet_connection_sock *icsk = inet_csk(sk);
 757
 758         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 759         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 760         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 761         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 762 }
 763 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 764
 765 void inet_csk_clear_xmit_timers(struct sock *sk)
 766 {
 767         struct inet_connection_sock *icsk = inet_csk(sk);
 768
 769         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 770
 771         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 772         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 773         sk_stop_timer(sk, &sk->sk_timer);
 774 }
 775 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 776
 777 void inet_csk_clear_xmit_timers_sync(struct sock *sk)
 778 {
 779         struct inet_connection_sock *icsk = inet_csk(sk);
 780
 781         /* ongoing timer handlers need to acquire socket lock. */
 782         sock_not_owned_by_me(sk);
 783
 784         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 785
 786         sk_stop_timer_sync(sk, &icsk->icsk_retransmit_timer);
 787         sk_stop_timer_sync(sk, &icsk->icsk_delack_timer);
 788         sk_stop_timer_sync(sk, &sk->sk_timer);
 789 }
 790
 791 void inet_csk_delete_keepalive_timer(struct sock *sk)
 792 {
 793         sk_stop_timer(sk, &sk->sk_timer);
 794 }
 795 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 796
 797 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 798 {
 799         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 800 }
 801 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 802
 803 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 804                                      struct flowi4 *fl4,
 805                                      const struct request_sock *req)
 806 {
 807         const struct inet_request_sock *ireq = inet_rsk(req);
 808         struct net *net = read_pnet(&ireq->ireq_net);
 809         struct ip_options_rcu *opt;
 810         struct rtable *rt;
 811
 812         rcu_read_lock();
 813         opt = rcu_dereference(ireq->ireq_opt);
 814
 815         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 816                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 817                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 818                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 819                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 820                            htons(ireq->ir_num), sk->sk_uid);
 821         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 822         rt = ip_route_output_flow(net, fl4, sk);
 823         if (IS_ERR(rt))
 824                 goto no_route;
 825         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 826                 goto route_err;
 827         rcu_read_unlock();
 828         return &rt->dst;
 829
 830 route_err:
 831         ip_rt_put(rt);
 832 no_route:
 833         rcu_read_unlock();
 834         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 835         return NULL;
 836 }
 837 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 838
 839 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 840                                             struct sock *newsk,
 841                                             const struct request_sock *req)
 842 {
 843         const struct inet_request_sock *ireq = inet_rsk(req);
 844         struct net *net = read_pnet(&ireq->ireq_net);
 845         struct inet_sock *newinet = inet_sk(newsk);
 846         struct ip_options_rcu *opt;
 847         struct flowi4 *fl4;
 848         struct rtable *rt;
 849
 850         opt = rcu_dereference(ireq->ireq_opt);
 851         fl4 = &newinet->cork.fl.u.ip4;
 852
 853         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 854                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 855                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 856                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 857                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 858                            htons(ireq->ir_num), sk->sk_uid);
 859         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 860         rt = ip_route_output_flow(net, fl4, sk);
 861         if (IS_ERR(rt))
 862                 goto no_route;
 863         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 864                 goto route_err;
 865         return &rt->dst;
 866
 867 route_err:
 868         ip_rt_put(rt);
 869 no_route:
 870         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 871         return NULL;
 872 }
 873 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 874
 875 /* Decide when to expire the request and when to resend SYN-ACK */
 876 static void syn_ack_recalc(struct request_sock *req,
 877                            const int max_syn_ack_retries,
 878                            const u8 rskq_defer_accept,
 879                            int *expire, int *resend)
 880 {
 881         if (!rskq_defer_accept) {
 882                 *expire = req->num_timeout >= max_syn_ack_retries;
 883                 *resend = 1;
 884                 return;
 885         }
 886         *expire = req->num_timeout >= max_syn_ack_retries &&
 887                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 888         /* Do not resend while waiting for data after ACK,
 889          * start to resend on end of deferring period to give
 890          * last chance for data or ACK to create established socket.
 891          */
 892         *resend = !inet_rsk(req)->acked ||
 893                   req->num_timeout >= rskq_defer_accept - 1;
 894 }
 895
 896 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 897 {
 898         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 899
 900         if (!err)
 901                 req->num_retrans++;
 902         return err;
 903 }
 904 EXPORT_SYMBOL(inet_rtx_syn_ack);
 905
 906 static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 907                                              struct sock *sk)
 908 {
 909         struct sock *req_sk, *nreq_sk;
 910         struct request_sock *nreq;
 911
 912         nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN);
 913         if (!nreq) {
 914                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
 915
 916                 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */
 917                 sock_put(sk);
 918                 return NULL;
 919         }
 920
 921         req_sk = req_to_sk(req);
 922         nreq_sk = req_to_sk(nreq);
 923
 924         memcpy(nreq_sk, req_sk,
 925                offsetof(struct sock, sk_dontcopy_begin));
 926         memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end,
 927                req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end));
 928
 929         sk_node_init(&nreq_sk->sk_node);
 930         nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping;
 931 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
 932         nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping;
 933 #endif
 934         nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu;
 935
 936         nreq->rsk_listener = sk;
 937
 938         /* We need not acquire fastopenq->lock
 939          * because the child socket is locked in inet_csk_listen_stop().
 940          */
 941         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener)
 942                 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
 943
 944         return nreq;
 945 }
 946
 947 static void reqsk_queue_migrated(struct request_sock_queue *queue,
 948                                  const struct request_sock *req)
 949 {
 950         if (req->num_timeout == 0)
 951                 atomic_inc(&queue->young);
 952         atomic_inc(&queue->qlen);
 953 }
 954
 955 static void reqsk_migrate_reset(struct request_sock *req)
 956 {
 957         req->saved_syn = NULL;
 958 #if IS_ENABLED(CONFIG_IPV6)
 959         inet_rsk(req)->ipv6_opt = NULL;
 960         inet_rsk(req)->pktopts = NULL;
 961 #else
 962         inet_rsk(req)->ireq_opt = NULL;
 963 #endif
 964 }
 965
 966 /* return true if req was found in the ehash table */
 967 static bool reqsk_queue_unlink(struct request_sock *req)
 968 {
 969         struct sock *sk = req_to_sk(req);
 970         bool found = false;
 971
 972         if (sk_hashed(sk)) {
 973                 struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 974                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 975
 976                 spin_lock(lock);
 977                 found = __sk_nulls_del_node_init_rcu(sk);
 978                 spin_unlock(lock);
 979         }
 980         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 981                 reqsk_put(req);
 982         return found;
 983 }
 984
 985 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 986 {
 987         bool unlinked = reqsk_queue_unlink(req);
 988
 989         if (unlinked) {
 990                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 991                 reqsk_put(req);
 992         }
 993         return unlinked;
 994 }
 995 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 996
 997 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 998 {
 999         inet_csk_reqsk_queue_drop(sk, req);
1000         reqsk_put(req);
1001 }
1002 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
1003
1004 static void reqsk_timer_handler(struct timer_list *t)
1005 {
1006         struct request_sock *req = from_timer(req, t, rsk_timer);
1007         struct request_sock *nreq = NULL, *oreq = req;
1008         struct sock *sk_listener = req->rsk_listener;
1009         struct inet_connection_sock *icsk;
1010         struct request_sock_queue *queue;
1011         struct net *net;
1012         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
1013
1014         if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
1015                 struct sock *nsk;
1016
1017                 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
1018                 if (!nsk)
1019                         goto drop;
1020
1021                 nreq = inet_reqsk_clone(req, nsk);
1022                 if (!nreq)
1023                         goto drop;
1024
1025                 /* The new timer for the cloned req can decrease the 2
1026                  * by calling inet_csk_reqsk_queue_drop_and_put(), so
1027                  * hold another count to prevent use-after-free and
1028                  * call reqsk_put() just before return.
1029                  */
1030                 refcount_set(&nreq->rsk_refcnt, 2 + 1);
1031                 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1032                 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
1033
1034                 req = nreq;
1035                 sk_listener = nsk;
1036         }
1037
1038         icsk = inet_csk(sk_listener);
1039         net = sock_net(sk_listener);
1040         max_syn_ack_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
1041                 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries);
1042         /* Normally all the openreqs are young and become mature
1043          * (i.e. converted to established socket) for first timeout.
1044          * If synack was not acknowledged for 1 second, it means
1045          * one of the following things: synack was lost, ack was lost,
1046          * rtt is high or nobody planned to ack (i.e. synflood).
1047          * When server is a bit loaded, queue is populated with old
1048          * open requests, reducing effective size of queue.
1049          * When server is well loaded, queue size reduces to zero
1050          * after several minutes of work. It is not synflood,
1051          * it is normal operation. The solution is pruning
1052          * too old entries overriding normal timeout, when
1053          * situation becomes dangerous.
1054          *
1055          * Essentially, we reserve half of room for young
1056          * embrions; and abort old ones without pity, if old
1057          * ones are about to clog our table.
1058          */
1059         queue = &icsk->icsk_accept_queue;
1060         qlen = reqsk_queue_len(queue);
1061         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
1062                 int young = reqsk_queue_len_young(queue) << 1;
1063
1064                 while (max_syn_ack_retries > 2) {
1065                         if (qlen < young)
1066                                 break;
1067                         max_syn_ack_retries--;
1068                         young <<= 1;
1069                 }
1070         }
1071         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
1072                        &expire, &resend);
1073         req->rsk_ops->syn_ack_timeout(req);
1074         if (!expire &&
1075             (!resend ||
1076              !inet_rtx_syn_ack(sk_listener, req) ||
1077              inet_rsk(req)->acked)) {
1078                 if (req->num_timeout++ == 0)
1079                         atomic_dec(&queue->young);
1080                 mod_timer(&req->rsk_timer, jiffies + reqsk_timeout(req, TCP_RTO_MAX));
1081
1082                 if (!nreq)
1083                         return;
1084
1085                 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
1086                         /* delete timer */
1087                         inet_csk_reqsk_queue_drop(sk_listener, nreq);
1088                         goto no_ownership;
1089                 }
1090
1091                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS);
1092                 reqsk_migrate_reset(oreq);
1093                 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
1094                 reqsk_put(oreq);
1095
1096                 reqsk_put(nreq);
1097                 return;
1098         }
1099
1100         /* Even if we can clone the req, we may need not retransmit any more
1101          * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
1102          * CPU may win the "own_req" race so that inet_ehash_insert() fails.
1103          */
1104         if (nreq) {
1105                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE);
1106 no_ownership:
1107                 reqsk_migrate_reset(nreq);
1108                 reqsk_queue_removed(queue, nreq);
1109                 __reqsk_free(nreq);
1110         }
1111
1112 drop:
1113         inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
1114 }
1115
1116 static void reqsk_queue_hash_req(struct request_sock *req,
1117                                  unsigned long timeout)
1118 {
1119         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1120         mod_timer(&req->rsk_timer, jiffies + timeout);
1121
1122         inet_ehash_insert(req_to_sk(req), NULL, NULL);
1123         /* before letting lookups find us, make sure all req fields
1124          * are committed to memory and refcnt initialized.
1125          */
1126         smp_wmb();
1127         refcount_set(&req->rsk_refcnt, 2 + 1);
1128 }
1129
1130 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
1131                                    unsigned long timeout)
1132 {
1133         reqsk_queue_hash_req(req, timeout);
1134         inet_csk_reqsk_queue_added(sk);
1135 }
1136 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
1137
1138 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
1139                            const gfp_t priority)
1140 {
1141         struct inet_connection_sock *icsk = inet_csk(newsk);
1142
1143         if (!icsk->icsk_ulp_ops)
1144                 return;
1145
1146         if (icsk->icsk_ulp_ops->clone)
1147                 icsk->icsk_ulp_ops->clone(req, newsk, priority);
1148 }
1149
1150 /**
1151  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
1152  *      @sk: the socket to clone
1153  *      @req: request_sock
1154  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
1155  *
1156  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
1157  */
1158 struct sock *inet_csk_clone_lock(const struct sock *sk,
1159                                  const struct request_sock *req,
1160                                  const gfp_t priority)
1161 {
1162         struct sock *newsk = sk_clone_lock(sk, priority);
1163
1164         if (newsk) {
1165                 struct inet_connection_sock *newicsk = inet_csk(newsk);
1166
1167                 inet_sk_set_state(newsk, TCP_SYN_RECV);
1168                 newicsk->icsk_bind_hash = NULL;
1169                 newicsk->icsk_bind2_hash = NULL;
1170
1171                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
1172                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
1173                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
1174
1175                 /* listeners have SOCK_RCU_FREE, not the children */
1176                 sock_reset_flag(newsk, SOCK_RCU_FREE);
1177
1178                 inet_sk(newsk)->mc_list = NULL;
1179
1180                 newsk->sk_mark = inet_rsk(req)->ir_mark;
1181                 atomic64_set(&newsk->sk_cookie,
1182                              atomic64_read(&inet_rsk(req)->ir_cookie));
1183
1184                 newicsk->icsk_retransmits = 0;
1185                 newicsk->icsk_backoff     = 0;
1186                 newicsk->icsk_probes_out  = 0;
1187                 newicsk->icsk_probes_tstamp = 0;
1188
1189                 /* Deinitialize accept_queue to trap illegal accesses. */
1190                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
1191
1192                 inet_clone_ulp(req, newsk, priority);
1193
1194                 security_inet_csk_clone(newsk, req);
1195         }
1196         return newsk;
1197 }
1198 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
1199
1200 /*
1201  * At this point, there should be no process reference to this
1202  * socket, and thus no user references at all.  Therefore we
1203  * can assume the socket waitqueue is inactive and nobody will
1204  * try to jump onto it.
1205  */
1206 void inet_csk_destroy_sock(struct sock *sk)
1207 {
1208         WARN_ON(sk->sk_state != TCP_CLOSE);
1209         WARN_ON(!sock_flag(sk, SOCK_DEAD));
1210
1211         /* It cannot be in hash table! */
1212         WARN_ON(!sk_unhashed(sk));
1213
1214         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
1215         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
1216
1217         sk->sk_prot->destroy(sk);
1218
1219         sk_stream_kill_queues(sk);
1220
1221         xfrm_sk_free_policy(sk);
1222
1223         sk_refcnt_debug_release(sk);
1224
1225         this_cpu_dec(*sk->sk_prot->orphan_count);
1226
1227         sock_put(sk);
1228 }
1229 EXPORT_SYMBOL(inet_csk_destroy_sock);
1230
1231 /* This function allows to force a closure of a socket after the call to
1232  * tcp/dccp_create_openreq_child().
1233  */
1234 void inet_csk_prepare_forced_close(struct sock *sk)
1235         __releases(&sk->sk_lock.slock)
1236 {
1237         /* sk_clone_lock locked the socket and set refcnt to 2 */
1238         bh_unlock_sock(sk);
1239         sock_put(sk);
1240         inet_csk_prepare_for_destroy_sock(sk);
1241         inet_sk(sk)->inet_num = 0;
1242 }
1243 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
1244
1245 static int inet_ulp_can_listen(const struct sock *sk)
1246 {
1247         const struct inet_connection_sock *icsk = inet_csk(sk);
1248
1249         if (icsk->icsk_ulp_ops && !icsk->icsk_ulp_ops->clone)
1250                 return -EINVAL;
1251
1252         return 0;
1253 }
1254
1255 int inet_csk_listen_start(struct sock *sk)
1256 {
1257         struct inet_connection_sock *icsk = inet_csk(sk);
1258         struct inet_sock *inet = inet_sk(sk);
1259         int err;
1260
1261         err = inet_ulp_can_listen(sk);
1262         if (unlikely(err))
1263                 return err;
1264
1265         reqsk_queue_alloc(&icsk->icsk_accept_queue);
1266
1267         sk->sk_ack_backlog = 0;
1268         inet_csk_delack_init(sk);
1269
1270         /* There is race window here: we announce ourselves listening,
1271          * but this transition is still not validated by get_port().
1272          * It is OK, because this socket enters to hash table only
1273          * after validation is complete.
1274          */
1275         inet_sk_state_store(sk, TCP_LISTEN);
1276         err = sk->sk_prot->get_port(sk, inet->inet_num);
1277         if (!err) {
1278                 inet->inet_sport = htons(inet->inet_num);
1279
1280                 sk_dst_reset(sk);
1281                 err = sk->sk_prot->hash(sk);
1282
1283                 if (likely(!err))
1284                         return 0;
1285         }
1286
1287         inet_sk_set_state(sk, TCP_CLOSE);
1288         return err;
1289 }
1290 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
1291
1292 static void inet_child_forget(struct sock *sk, struct request_sock *req,
1293                               struct sock *child)
1294 {
1295         sk->sk_prot->disconnect(child, O_NONBLOCK);
1296
1297         sock_orphan(child);
1298
1299         this_cpu_inc(*sk->sk_prot->orphan_count);
1300
1301         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
1302                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
1303                 BUG_ON(sk != req->rsk_listener);
1304
1305                 /* Paranoid, to prevent race condition if
1306                  * an inbound pkt destined for child is
1307                  * blocked by sock lock in tcp_v4_rcv().
1308                  * Also to satisfy an assertion in
1309                  * tcp_v4_destroy_sock().
1310                  */
1311                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1312         }
1313         inet_csk_destroy_sock(child);
1314 }
1315
1316 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1317                                       struct request_sock *req,
1318                                       struct sock *child)
1319 {
1320         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1321
1322         spin_lock(&queue->rskq_lock);
1323         if (unlikely(sk->sk_state != TCP_LISTEN)) {
1324                 inet_child_forget(sk, req, child);
1325                 child = NULL;
1326         } else {
1327                 req->sk = child;
1328                 req->dl_next = NULL;
1329                 if (queue->rskq_accept_head == NULL)
1330                         WRITE_ONCE(queue->rskq_accept_head, req);
1331                 else
1332                         queue->rskq_accept_tail->dl_next = req;
1333                 queue->rskq_accept_tail = req;
1334                 sk_acceptq_added(sk);
1335         }
1336         spin_unlock(&queue->rskq_lock);
1337         return child;
1338 }
1339 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1340
1341 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1342                                          struct request_sock *req, bool own_req)
1343 {
1344         if (own_req) {
1345                 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
1346                 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
1347
1348                 if (sk != req->rsk_listener) {
1349                         /* another listening sk has been selected,
1350                          * migrate the req to it.
1351                          */
1352                         struct request_sock *nreq;
1353
1354                         /* hold a refcnt for the nreq->rsk_listener
1355                          * which is assigned in inet_reqsk_clone()
1356                          */
1357                         sock_hold(sk);
1358                         nreq = inet_reqsk_clone(req, sk);
1359                         if (!nreq) {
1360                                 inet_child_forget(sk, req, child);
1361                                 goto child_put;
1362                         }
1363
1364                         refcount_set(&nreq->rsk_refcnt, 1);
1365                         if (inet_csk_reqsk_queue_add(sk, nreq, child)) {
1366                                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS);
1367                                 reqsk_migrate_reset(req);
1368                                 reqsk_put(req);
1369                                 return child;
1370                         }
1371
1372                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
1373                         reqsk_migrate_reset(nreq);
1374                         __reqsk_free(nreq);
1375                 } else if (inet_csk_reqsk_queue_add(sk, req, child)) {
1376                         return child;
1377                 }
1378         }
1379         /* Too bad, another child took ownership of the request, undo. */
1380 child_put:
1381         bh_unlock_sock(child);
1382         sock_put(child);
1383         return NULL;
1384 }
1385 EXPORT_SYMBOL(inet_csk_complete_hashdance);
1386
1387 /*
1388  *      This routine closes sockets which have been at least partially
1389  *      opened, but not yet accepted.
1390  */
1391 void inet_csk_listen_stop(struct sock *sk)
1392 {
1393         struct inet_connection_sock *icsk = inet_csk(sk);
1394         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1395         struct request_sock *next, *req;
1396
1397         /* Following specs, it would be better either to send FIN
1398          * (and enter FIN-WAIT-1, it is normal close)
1399          * or to send active reset (abort).
1400          * Certainly, it is pretty dangerous while synflood, but it is
1401          * bad justification for our negligence 8)
1402          * To be honest, we are not able to make either
1403          * of the variants now.                 --ANK
1404          */
1405         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1406                 struct sock *child = req->sk, *nsk;
1407                 struct request_sock *nreq;
1408
1409                 local_bh_disable();
1410                 bh_lock_sock(child);
1411                 WARN_ON(sock_owned_by_user(child));
1412                 sock_hold(child);
1413
1414                 nsk = reuseport_migrate_sock(sk, child, NULL);
1415                 if (nsk) {
1416                         nreq = inet_reqsk_clone(req, nsk);
1417                         if (nreq) {
1418                                 refcount_set(&nreq->rsk_refcnt, 1);
1419
1420                                 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
1421                                         __NET_INC_STATS(sock_net(nsk),
1422                                                         LINUX_MIB_TCPMIGRATEREQSUCCESS);
1423                                         reqsk_migrate_reset(req);
1424                                 } else {
1425                                         __NET_INC_STATS(sock_net(nsk),
1426                                                         LINUX_MIB_TCPMIGRATEREQFAILURE);
1427                                         reqsk_migrate_reset(nreq);
1428                                         __reqsk_free(nreq);
1429                                 }
1430
1431                                 /* inet_csk_reqsk_queue_add() has already
1432                                  * called inet_child_forget() on failure case.
1433                                  */
1434                                 goto skip_child_forget;
1435                         }
1436                 }
1437
1438                 inet_child_forget(sk, req, child);
1439 skip_child_forget:
1440                 reqsk_put(req);
1441                 bh_unlock_sock(child);
1442                 local_bh_enable();
1443                 sock_put(child);
1444
1445                 cond_resched();
1446         }
1447         if (queue->fastopenq.rskq_rst_head) {
1448                 /* Free all the reqs queued in rskq_rst_head. */
1449                 spin_lock_bh(&queue->fastopenq.lock);
1450                 req = queue->fastopenq.rskq_rst_head;
1451                 queue->fastopenq.rskq_rst_head = NULL;
1452                 spin_unlock_bh(&queue->fastopenq.lock);
1453                 while (req != NULL) {
1454                         next = req->dl_next;
1455                         reqsk_put(req);
1456                         req = next;
1457                 }
1458         }
1459         WARN_ON_ONCE(sk->sk_ack_backlog);
1460 }
1461 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1462
1463 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
1464 {
1465         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
1466         const struct inet_sock *inet = inet_sk(sk);
1467
1468         sin->sin_family         = AF_INET;
1469         sin->sin_addr.s_addr    = inet->inet_daddr;
1470         sin->sin_port           = inet->inet_dport;
1471 }
1472 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
1473
1474 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1475 {
1476         const struct inet_sock *inet = inet_sk(sk);
1477         const struct ip_options_rcu *inet_opt;
1478         __be32 daddr = inet->inet_daddr;
1479         struct flowi4 *fl4;
1480         struct rtable *rt;
1481
1482         rcu_read_lock();
1483         inet_opt = rcu_dereference(inet->inet_opt);
1484         if (inet_opt && inet_opt->opt.srr)
1485                 daddr = inet_opt->opt.faddr;
1486         fl4 = &fl->u.ip4;
1487         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
1488                                    inet->inet_saddr, inet->inet_dport,
1489                                    inet->inet_sport, sk->sk_protocol,
1490                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
1491         if (IS_ERR(rt))
1492                 rt = NULL;
1493         if (rt)
1494                 sk_setup_caps(sk, &rt->dst);
1495         rcu_read_unlock();
1496
1497         return &rt->dst;
1498 }
1499
1500 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1501 {
1502         struct dst_entry *dst = __sk_dst_check(sk, 0);
1503         struct inet_sock *inet = inet_sk(sk);
1504
1505         if (!dst) {
1506                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1507                 if (!dst)
1508                         goto out;
1509         }
1510         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1511
1512         dst = __sk_dst_check(sk, 0);
1513         if (!dst)
1514                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1515 out:
1516         return dst;
1517 }
1518 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);