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                 if (ipv6_addr_any(&sk->sk_v6_rcv_saddr))
 161                         return false;
 162
 163                 if (!ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr))
 164                         return true;
 165         }
 166 #endif
 167         return sk->sk_rcv_saddr != htonl(INADDR_ANY);
 168 }
 169
 170 static bool inet_bind_conflict(const struct sock *sk, struct sock *sk2,
 171                                kuid_t sk_uid, bool relax,
 172                                bool reuseport_cb_ok, bool reuseport_ok)
 173 {
 174         int bound_dev_if2;
 175
 176         if (sk == sk2)
 177                 return false;
 178
 179         bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if);
 180
 181         if (!sk->sk_bound_dev_if || !bound_dev_if2 ||
 182             sk->sk_bound_dev_if == bound_dev_if2) {
 183                 if (sk->sk_reuse && sk2->sk_reuse &&
 184                     sk2->sk_state != TCP_LISTEN) {
 185                         if (!relax || (!reuseport_ok && sk->sk_reuseport &&
 186                                        sk2->sk_reuseport && reuseport_cb_ok &&
 187                                        (sk2->sk_state == TCP_TIME_WAIT ||
 188                                         uid_eq(sk_uid, sock_i_uid(sk2)))))
 189                                 return true;
 190                 } else if (!reuseport_ok || !sk->sk_reuseport ||
 191                            !sk2->sk_reuseport || !reuseport_cb_ok ||
 192                            (sk2->sk_state != TCP_TIME_WAIT &&
 193                             !uid_eq(sk_uid, sock_i_uid(sk2)))) {
 194                         return true;
 195                 }
 196         }
 197         return false;
 198 }
 199
 200 static bool __inet_bhash2_conflict(const struct sock *sk, struct sock *sk2,
 201                                    kuid_t sk_uid, bool relax,
 202                                    bool reuseport_cb_ok, bool reuseport_ok)
 203 {
 204         if (ipv6_only_sock(sk2)) {
 205                 if (sk->sk_family == AF_INET)
 206                         return false;
 207
 208 #if IS_ENABLED(CONFIG_IPV6)
 209                 if (ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr))
 210                         return false;
 211 #endif
 212         }
 213
 214         return inet_bind_conflict(sk, sk2, sk_uid, relax,
 215                                   reuseport_cb_ok, reuseport_ok);
 216 }
 217
 218 static bool inet_bhash2_conflict(const struct sock *sk,
 219                                  const struct inet_bind2_bucket *tb2,
 220                                  kuid_t sk_uid,
 221                                  bool relax, bool reuseport_cb_ok,
 222                                  bool reuseport_ok)
 223 {
 224         struct sock *sk2;
 225
 226         sk_for_each_bound(sk2, &tb2->owners) {
 227                 if (__inet_bhash2_conflict(sk, sk2, sk_uid, relax,
 228                                            reuseport_cb_ok, reuseport_ok))
 229                         return true;
 230         }
 231
 232         return false;
 233 }
 234
 235 #define sk_for_each_bound_bhash(__sk, __tb2, __tb)                      \
 236         hlist_for_each_entry(__tb2, &(__tb)->bhash2, bhash_node)        \
 237                 sk_for_each_bound((__sk), &(__tb2)->owners)
 238
 239 /* This should be called only when the tb and tb2 hashbuckets' locks are held */
 240 static int inet_csk_bind_conflict(const struct sock *sk,
 241                                   const struct inet_bind_bucket *tb,
 242                                   const struct inet_bind2_bucket *tb2, /* may be null */
 243                                   bool relax, bool reuseport_ok)
 244 {
 245         kuid_t uid = sock_i_uid((struct sock *)sk);
 246         struct sock_reuseport *reuseport_cb;
 247         bool reuseport_cb_ok;
 248         struct sock *sk2;
 249
 250         rcu_read_lock();
 251         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 252         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 253         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 254         rcu_read_unlock();
 255
 256         /* Conflicts with an existing IPV6_ADDR_ANY (if ipv6) or INADDR_ANY (if
 257          * ipv4) should have been checked already. We need to do these two
 258          * checks separately because their spinlocks have to be acquired/released
 259          * independently of each other, to prevent possible deadlocks
 260          */
 261         if (inet_use_bhash2_on_bind(sk))
 262                 return tb2 && inet_bhash2_conflict(sk, tb2, uid, relax,
 263                                                    reuseport_cb_ok, reuseport_ok);
 264
 265         /* Unlike other sk lookup places we do not check
 266          * for sk_net here, since _all_ the socks listed
 267          * in tb->owners and tb2->owners list belong
 268          * to the same net - the one this bucket belongs to.
 269          */
 270         sk_for_each_bound_bhash(sk2, tb2, tb) {
 271                 if (!inet_bind_conflict(sk, sk2, uid, relax, reuseport_cb_ok, reuseport_ok))
 272                         continue;
 273
 274                 if (inet_rcv_saddr_equal(sk, sk2, true))
 275                         return true;
 276         }
 277
 278         return false;
 279 }
 280
 281 /* Determine if there is a bind conflict with an existing IPV6_ADDR_ANY (if ipv6) or
 282  * INADDR_ANY (if ipv4) socket.
 283  *
 284  * Caller must hold bhash hashbucket lock with local bh disabled, to protect
 285  * against concurrent binds on the port for addr any
 286  */
 287 static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l3mdev,
 288                                           bool relax, bool reuseport_ok)
 289 {
 290         kuid_t uid = sock_i_uid((struct sock *)sk);
 291         const struct net *net = sock_net(sk);
 292         struct sock_reuseport *reuseport_cb;
 293         struct inet_bind_hashbucket *head2;
 294         struct inet_bind2_bucket *tb2;
 295         bool conflict = false;
 296         bool reuseport_cb_ok;
 297
 298         rcu_read_lock();
 299         reuseport_cb = rcu_dereference(sk->sk_reuseport_cb);
 300         /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */
 301         reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks);
 302         rcu_read_unlock();
 303
 304         head2 = inet_bhash2_addr_any_hashbucket(sk, net, port);
 305
 306         spin_lock(&head2->lock);
 307
 308         inet_bind_bucket_for_each(tb2, &head2->chain) {
 309                 if (!inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
 310                         continue;
 311
 312                 if (!inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok, reuseport_ok))
 313                         continue;
 314
 315                 conflict = true;
 316                 break;
 317         }
 318
 319         spin_unlock(&head2->lock);
 320
 321         return conflict;
 322 }
 323
 324 /*
 325  * Find an open port number for the socket.  Returns with the
 326  * inet_bind_hashbucket locks held if successful.
 327  */
 328 static struct inet_bind_hashbucket *
 329 inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 330                         struct inet_bind2_bucket **tb2_ret,
 331                         struct inet_bind_hashbucket **head2_ret, int *port_ret)
 332 {
 333         struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk);
 334         int i, low, high, attempt_half, port, l3mdev;
 335         struct inet_bind_hashbucket *head, *head2;
 336         struct net *net = sock_net(sk);
 337         struct inet_bind2_bucket *tb2;
 338         struct inet_bind_bucket *tb;
 339         u32 remaining, offset;
 340         bool relax = false;
 341
 342         l3mdev = inet_sk_bound_l3mdev(sk);
 343 ports_exhausted:
 344         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 345 other_half_scan:
 346         inet_sk_get_local_port_range(sk, &low, &high);
 347         high++; /* [32768, 60999] -> [32768, 61000[ */
 348         if (high - low < 4)
 349                 attempt_half = 0;
 350         if (attempt_half) {
 351                 int half = low + (((high - low) >> 2) << 1);
 352
 353                 if (attempt_half == 1)
 354                         high = half;
 355                 else
 356                         low = half;
 357         }
 358         remaining = high - low;
 359         if (likely(remaining > 1))
 360                 remaining &= ~1U;
 361
 362         offset = get_random_u32_below(remaining);
 363         /* __inet_hash_connect() favors ports having @low parity
 364          * We do the opposite to not pollute connect() users.
 365          */
 366         offset |= 1U;
 367
 368 other_parity_scan:
 369         port = low + offset;
 370         for (i = 0; i < remaining; i += 2, port += 2) {
 371                 if (unlikely(port >= high))
 372                         port -= remaining;
 373                 if (inet_is_local_reserved_port(net, port))
 374                         continue;
 375                 head = &hinfo->bhash[inet_bhashfn(net, port,
 376                                                   hinfo->bhash_size)];
 377                 spin_lock_bh(&head->lock);
 378                 if (inet_use_bhash2_on_bind(sk)) {
 379                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, relax, false))
 380                                 goto next_port;
 381                 }
 382
 383                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 384                 spin_lock(&head2->lock);
 385                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 386                 inet_bind_bucket_for_each(tb, &head->chain)
 387                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
 388                                 if (!inet_csk_bind_conflict(sk, tb, tb2,
 389                                                             relax, false))
 390                                         goto success;
 391                                 spin_unlock(&head2->lock);
 392                                 goto next_port;
 393                         }
 394                 tb = NULL;
 395                 goto success;
 396 next_port:
 397                 spin_unlock_bh(&head->lock);
 398                 cond_resched();
 399         }
 400
 401         offset--;
 402         if (!(offset & 1))
 403                 goto other_parity_scan;
 404
 405         if (attempt_half == 1) {
 406                 /* OK we now try the upper half of the range */
 407                 attempt_half = 2;
 408                 goto other_half_scan;
 409         }
 410
 411         if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) {
 412                 /* We still have a chance to connect to different destinations */
 413                 relax = true;
 414                 goto ports_exhausted;
 415         }
 416         return NULL;
 417 success:
 418         *port_ret = port;
 419         *tb_ret = tb;
 420         *tb2_ret = tb2;
 421         *head2_ret = head2;
 422         return head;
 423 }
 424
 425 static inline int sk_reuseport_match(struct inet_bind_bucket *tb,
 426                                      struct sock *sk)
 427 {
 428         kuid_t uid = sock_i_uid(sk);
 429
 430         if (tb->fastreuseport <= 0)
 431                 return 0;
 432         if (!sk->sk_reuseport)
 433                 return 0;
 434         if (rcu_access_pointer(sk->sk_reuseport_cb))
 435                 return 0;
 436         if (!uid_eq(tb->fastuid, uid))
 437                 return 0;
 438         /* We only need to check the rcv_saddr if this tb was once marked
 439          * without fastreuseport and then was reset, as we can only know that
 440          * the fast_*rcv_saddr doesn't have any conflicts with the socks on the
 441          * owners list.
 442          */
 443         if (tb->fastreuseport == FASTREUSEPORT_ANY)
 444                 return 1;
 445 #if IS_ENABLED(CONFIG_IPV6)
 446         if (tb->fast_sk_family == AF_INET6)
 447                 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr,
 448                                             inet6_rcv_saddr(sk),
 449                                             tb->fast_rcv_saddr,
 450                                             sk->sk_rcv_saddr,
 451                                             tb->fast_ipv6_only,
 452                                             ipv6_only_sock(sk), true, false);
 453 #endif
 454         return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr,
 455                                     ipv6_only_sock(sk), true, false);
 456 }
 457
 458 void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,
 459                                struct sock *sk)
 460 {
 461         kuid_t uid = sock_i_uid(sk);
 462         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 463
 464         if (hlist_empty(&tb->bhash2)) {
 465                 tb->fastreuse = reuse;
 466                 if (sk->sk_reuseport) {
 467                         tb->fastreuseport = FASTREUSEPORT_ANY;
 468                         tb->fastuid = uid;
 469                         tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 470                         tb->fast_ipv6_only = ipv6_only_sock(sk);
 471                         tb->fast_sk_family = sk->sk_family;
 472 #if IS_ENABLED(CONFIG_IPV6)
 473                         tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 474 #endif
 475                 } else {
 476                         tb->fastreuseport = 0;
 477                 }
 478         } else {
 479                 if (!reuse)
 480                         tb->fastreuse = 0;
 481                 if (sk->sk_reuseport) {
 482                         /* We didn't match or we don't have fastreuseport set on
 483                          * the tb, but we have sk_reuseport set on this socket
 484                          * and we know that there are no bind conflicts with
 485                          * this socket in this tb, so reset our tb's reuseport
 486                          * settings so that any subsequent sockets that match
 487                          * our current socket will be put on the fast path.
 488                          *
 489                          * If we reset we need to set FASTREUSEPORT_STRICT so we
 490                          * do extra checking for all subsequent sk_reuseport
 491                          * socks.
 492                          */
 493                         if (!sk_reuseport_match(tb, sk)) {
 494                                 tb->fastreuseport = FASTREUSEPORT_STRICT;
 495                                 tb->fastuid = uid;
 496                                 tb->fast_rcv_saddr = sk->sk_rcv_saddr;
 497                                 tb->fast_ipv6_only = ipv6_only_sock(sk);
 498                                 tb->fast_sk_family = sk->sk_family;
 499 #if IS_ENABLED(CONFIG_IPV6)
 500                                 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 501 #endif
 502                         }
 503                 } else {
 504                         tb->fastreuseport = 0;
 505                 }
 506         }
 507 }
 508
 509 /* Obtain a reference to a local port for the given sock,
 510  * if snum is zero it means select any available local port.
 511  * We try to allocate an odd port (and leave even ports for connect())
 512  */
 513 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 514 {
 515         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 516         bool found_port = false, check_bind_conflict = true;
 517         bool bhash_created = false, bhash2_created = false;
 518         struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk);
 519         int ret = -EADDRINUSE, port = snum, l3mdev;
 520         struct inet_bind_hashbucket *head, *head2;
 521         struct inet_bind2_bucket *tb2 = NULL;
 522         struct inet_bind_bucket *tb = NULL;
 523         bool head2_lock_acquired = false;
 524         struct net *net = sock_net(sk);
 525
 526         l3mdev = inet_sk_bound_l3mdev(sk);
 527
 528         if (!port) {
 529                 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port);
 530                 if (!head)
 531                         return ret;
 532
 533                 head2_lock_acquired = true;
 534
 535                 if (tb && tb2)
 536                         goto success;
 537                 found_port = true;
 538         } else {
 539                 head = &hinfo->bhash[inet_bhashfn(net, port,
 540                                                   hinfo->bhash_size)];
 541                 spin_lock_bh(&head->lock);
 542                 inet_bind_bucket_for_each(tb, &head->chain)
 543                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 544                                 break;
 545         }
 546
 547         if (!tb) {
 548                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net,
 549                                              head, port, l3mdev);
 550                 if (!tb)
 551                         goto fail_unlock;
 552                 bhash_created = true;
 553         }
 554
 555         if (!found_port) {
 556                 if (!hlist_empty(&tb->bhash2)) {
 557                         if (sk->sk_reuse == SK_FORCE_REUSE ||
 558                             (tb->fastreuse > 0 && reuse) ||
 559                             sk_reuseport_match(tb, sk))
 560                                 check_bind_conflict = false;
 561                 }
 562
 563                 if (check_bind_conflict && inet_use_bhash2_on_bind(sk)) {
 564                         if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, true, true))
 565                                 goto fail_unlock;
 566                 }
 567
 568                 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 569                 spin_lock(&head2->lock);
 570                 head2_lock_acquired = true;
 571                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 572         }
 573
 574         if (!tb2) {
 575                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep,
 576                                                net, head2, tb, sk);
 577                 if (!tb2)
 578                         goto fail_unlock;
 579                 bhash2_created = true;
 580         }
 581
 582         if (!found_port && check_bind_conflict) {
 583                 if (inet_csk_bind_conflict(sk, tb, tb2, true, true))
 584                         goto fail_unlock;
 585         }
 586
 587 success:
 588         inet_csk_update_fastreuse(tb, sk);
 589
 590         if (!inet_csk(sk)->icsk_bind_hash)
 591                 inet_bind_hash(sk, tb, tb2, port);
 592         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 593         WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
 594         ret = 0;
 595
 596 fail_unlock:
 597         if (ret) {
 598                 if (bhash2_created)
 599                         inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, tb2);
 600                 if (bhash_created)
 601                         inet_bind_bucket_destroy(tb);
 602         }
 603         if (head2_lock_acquired)
 604                 spin_unlock(&head2->lock);
 605         spin_unlock_bh(&head->lock);
 606         return ret;
 607 }
 608 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 609
 610 /*
 611  * Wait for an incoming connection, avoid race conditions. This must be called
 612  * with the socket locked.
 613  */
 614 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 615 {
 616         struct inet_connection_sock *icsk = inet_csk(sk);
 617         DEFINE_WAIT(wait);
 618         int err;
 619
 620         /*
 621          * True wake-one mechanism for incoming connections: only
 622          * one process gets woken up, not the 'whole herd'.
 623          * Since we do not 'race & poll' for established sockets
 624          * anymore, the common case will execute the loop only once.
 625          *
 626          * Subtle issue: "add_wait_queue_exclusive()" will be added
 627          * after any current non-exclusive waiters, and we know that
 628          * it will always _stay_ after any new non-exclusive waiters
 629          * because all non-exclusive waiters are added at the
 630          * beginning of the wait-queue. As such, it's ok to "drop"
 631          * our exclusiveness temporarily when we get woken up without
 632          * having to remove and re-insert us on the wait queue.
 633          */
 634         for (;;) {
 635                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 636                                           TASK_INTERRUPTIBLE);
 637                 release_sock(sk);
 638                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 639                         timeo = schedule_timeout(timeo);
 640                 sched_annotate_sleep();
 641                 lock_sock(sk);
 642                 err = 0;
 643                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 644                         break;
 645                 err = -EINVAL;
 646                 if (sk->sk_state != TCP_LISTEN)
 647                         break;
 648                 err = sock_intr_errno(timeo);
 649                 if (signal_pending(current))
 650                         break;
 651                 err = -EAGAIN;
 652                 if (!timeo)
 653                         break;
 654         }
 655         finish_wait(sk_sleep(sk), &wait);
 656         return err;
 657 }
 658
 659 /*
 660  * This will accept the next outstanding connection.
 661  */
 662 struct sock *inet_csk_accept(struct sock *sk, struct proto_accept_arg *arg)
 663 {
 664         struct inet_connection_sock *icsk = inet_csk(sk);
 665         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 666         struct request_sock *req;
 667         struct sock *newsk;
 668         int error;
 669
 670         lock_sock(sk);
 671
 672         /* We need to make sure that this socket is listening,
 673          * and that it has something pending.
 674          */
 675         error = -EINVAL;
 676         if (sk->sk_state != TCP_LISTEN)
 677                 goto out_err;
 678
 679         /* Find already established connection */
 680         if (reqsk_queue_empty(queue)) {
 681                 long timeo = sock_rcvtimeo(sk, arg->flags & O_NONBLOCK);
 682
 683                 /* If this is a non blocking socket don't sleep */
 684                 error = -EAGAIN;
 685                 if (!timeo)
 686                         goto out_err;
 687
 688                 error = inet_csk_wait_for_connect(sk, timeo);
 689                 if (error)
 690                         goto out_err;
 691         }
 692         req = reqsk_queue_remove(queue, sk);
 693         arg->is_empty = reqsk_queue_empty(queue);
 694         newsk = req->sk;
 695
 696         if (sk->sk_protocol == IPPROTO_TCP &&
 697             tcp_rsk(req)->tfo_listener) {
 698                 spin_lock_bh(&queue->fastopenq.lock);
 699                 if (tcp_rsk(req)->tfo_listener) {
 700                         /* We are still waiting for the final ACK from 3WHS
 701                          * so can't free req now. Instead, we set req->sk to
 702                          * NULL to signify that the child socket is taken
 703                          * so reqsk_fastopen_remove() will free the req
 704                          * when 3WHS finishes (or is aborted).
 705                          */
 706                         req->sk = NULL;
 707                         req = NULL;
 708                 }
 709                 spin_unlock_bh(&queue->fastopenq.lock);
 710         }
 711
 712 out:
 713         release_sock(sk);
 714         if (newsk && mem_cgroup_sockets_enabled) {
 715                 gfp_t gfp = GFP_KERNEL | __GFP_NOFAIL;
 716                 int amt = 0;
 717
 718                 /* atomically get the memory usage, set and charge the
 719                  * newsk->sk_memcg.
 720                  */
 721                 lock_sock(newsk);
 722
 723                 mem_cgroup_sk_alloc(newsk);
 724                 if (newsk->sk_memcg) {
 725                         /* The socket has not been accepted yet, no need
 726                          * to look at newsk->sk_wmem_queued.
 727                          */
 728                         amt = sk_mem_pages(newsk->sk_forward_alloc +
 729                                            atomic_read(&newsk->sk_rmem_alloc));
 730                 }
 731
 732                 if (amt)
 733                         mem_cgroup_charge_skmem(newsk->sk_memcg, amt, gfp);
 734                 kmem_cache_charge(newsk, gfp);
 735
 736                 release_sock(newsk);
 737         }
 738         if (req)
 739                 reqsk_put(req);
 740
 741         if (newsk)
 742                 inet_init_csk_locks(newsk);
 743
 744         return newsk;
 745 out_err:
 746         newsk = NULL;
 747         req = NULL;
 748         arg->err = error;
 749         goto out;
 750 }
 751 EXPORT_SYMBOL(inet_csk_accept);
 752
 753 /*
 754  * Using different timers for retransmit, delayed acks and probes
 755  * We may wish use just one timer maintaining a list of expire jiffies
 756  * to optimize.
 757  */
 758 void inet_csk_init_xmit_timers(struct sock *sk,
 759                                void (*retransmit_handler)(struct timer_list *t),
 760                                void (*delack_handler)(struct timer_list *t),
 761                                void (*keepalive_handler)(struct timer_list *t))
 762 {
 763         struct inet_connection_sock *icsk = inet_csk(sk);
 764
 765         timer_setup(&icsk->icsk_retransmit_timer, retransmit_handler, 0);
 766         timer_setup(&icsk->icsk_delack_timer, delack_handler, 0);
 767         timer_setup(&sk->sk_timer, keepalive_handler, 0);
 768         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 769 }
 770
 771 void inet_csk_clear_xmit_timers(struct sock *sk)
 772 {
 773         struct inet_connection_sock *icsk = inet_csk(sk);
 774
 775         smp_store_release(&icsk->icsk_pending, 0);
 776         smp_store_release(&icsk->icsk_ack.pending, 0);
 777
 778         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 779         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 780         sk_stop_timer(sk, &sk->sk_timer);
 781 }
 782
 783 void inet_csk_clear_xmit_timers_sync(struct sock *sk)
 784 {
 785         struct inet_connection_sock *icsk = inet_csk(sk);
 786
 787         /* ongoing timer handlers need to acquire socket lock. */
 788         sock_not_owned_by_me(sk);
 789
 790         smp_store_release(&icsk->icsk_pending, 0);
 791         smp_store_release(&icsk->icsk_ack.pending, 0);
 792
 793         sk_stop_timer_sync(sk, &icsk->icsk_retransmit_timer);
 794         sk_stop_timer_sync(sk, &icsk->icsk_delack_timer);
 795         sk_stop_timer_sync(sk, &sk->sk_timer);
 796 }
 797
 798 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 799                                      struct flowi4 *fl4,
 800                                      const struct request_sock *req)
 801 {
 802         const struct inet_request_sock *ireq = inet_rsk(req);
 803         struct net *net = read_pnet(&ireq->ireq_net);
 804         struct ip_options_rcu *opt;
 805         struct rtable *rt;
 806
 807         rcu_read_lock();
 808         opt = rcu_dereference(ireq->ireq_opt);
 809
 810         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 811                            ip_sock_rt_tos(sk), ip_sock_rt_scope(sk),
 812                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 813                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 814                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 815                            htons(ireq->ir_num), sk->sk_uid);
 816         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 817         rt = ip_route_output_flow(net, fl4, sk);
 818         if (IS_ERR(rt))
 819                 goto no_route;
 820         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 821                 goto route_err;
 822         rcu_read_unlock();
 823         return &rt->dst;
 824
 825 route_err:
 826         ip_rt_put(rt);
 827 no_route:
 828         rcu_read_unlock();
 829         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 830         return NULL;
 831 }
 832
 833 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 834                                             struct sock *newsk,
 835                                             const struct request_sock *req)
 836 {
 837         const struct inet_request_sock *ireq = inet_rsk(req);
 838         struct net *net = read_pnet(&ireq->ireq_net);
 839         struct inet_sock *newinet = inet_sk(newsk);
 840         struct ip_options_rcu *opt;
 841         struct flowi4 *fl4;
 842         struct rtable *rt;
 843
 844         opt = rcu_dereference(ireq->ireq_opt);
 845         fl4 = &newinet->cork.fl.u.ip4;
 846
 847         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 848                            ip_sock_rt_tos(sk), ip_sock_rt_scope(sk),
 849                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 850                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 851                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 852                            htons(ireq->ir_num), sk->sk_uid);
 853         security_req_classify_flow(req, flowi4_to_flowi_common(fl4));
 854         rt = ip_route_output_flow(net, fl4, sk);
 855         if (IS_ERR(rt))
 856                 goto no_route;
 857         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 858                 goto route_err;
 859         return &rt->dst;
 860
 861 route_err:
 862         ip_rt_put(rt);
 863 no_route:
 864         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 865         return NULL;
 866 }
 867 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 868
 869 /* Decide when to expire the request and when to resend SYN-ACK */
 870 static void syn_ack_recalc(struct request_sock *req,
 871                            const int max_syn_ack_retries,
 872                            const u8 rskq_defer_accept,
 873                            int *expire, int *resend)
 874 {
 875         if (!rskq_defer_accept) {
 876                 *expire = req->num_timeout >= max_syn_ack_retries;
 877                 *resend = 1;
 878                 return;
 879         }
 880         *expire = req->num_timeout >= max_syn_ack_retries &&
 881                   (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept);
 882         /* Do not resend while waiting for data after ACK,
 883          * start to resend on end of deferring period to give
 884          * last chance for data or ACK to create established socket.
 885          */
 886         *resend = !inet_rsk(req)->acked ||
 887                   req->num_timeout >= rskq_defer_accept - 1;
 888 }
 889
 890 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 891 {
 892         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 893
 894         if (!err)
 895                 req->num_retrans++;
 896         return err;
 897 }
 898
 899 static struct request_sock *
 900 reqsk_alloc_noprof(const struct request_sock_ops *ops, struct sock *sk_listener,
 901                    bool attach_listener)
 902 {
 903         struct request_sock *req;
 904
 905         req = kmem_cache_alloc_noprof(ops->slab, GFP_ATOMIC | __GFP_NOWARN);
 906         if (!req)
 907                 return NULL;
 908         req->rsk_listener = NULL;
 909         if (attach_listener) {
 910                 if (unlikely(!refcount_inc_not_zero(&sk_listener->sk_refcnt))) {
 911                         kmem_cache_free(ops->slab, req);
 912                         return NULL;
 913                 }
 914                 req->rsk_listener = sk_listener;
 915         }
 916         req->rsk_ops = ops;
 917         req_to_sk(req)->sk_prot = sk_listener->sk_prot;
 918         sk_node_init(&req_to_sk(req)->sk_node);
 919         sk_tx_queue_clear(req_to_sk(req));
 920         req->saved_syn = NULL;
 921         req->syncookie = 0;
 922         req->timeout = 0;
 923         req->num_timeout = 0;
 924         req->num_retrans = 0;
 925         req->sk = NULL;
 926         refcount_set(&req->rsk_refcnt, 0);
 927
 928         return req;
 929 }
 930 #define reqsk_alloc(...)        alloc_hooks(reqsk_alloc_noprof(__VA_ARGS__))
 931
 932 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
 933                                       struct sock *sk_listener,
 934                                       bool attach_listener)
 935 {
 936         struct request_sock *req = reqsk_alloc(ops, sk_listener,
 937                                                attach_listener);
 938
 939         if (req) {
 940                 struct inet_request_sock *ireq = inet_rsk(req);
 941
 942                 ireq->ireq_opt = NULL;
 943 #if IS_ENABLED(CONFIG_IPV6)
 944                 ireq->pktopts = NULL;
 945 #endif
 946                 atomic64_set(&ireq->ir_cookie, 0);
 947                 ireq->ireq_state = TCP_NEW_SYN_RECV;
 948                 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
 949                 ireq->ireq_family = sk_listener->sk_family;
 950                 req->timeout = TCP_TIMEOUT_INIT;
 951         }
 952
 953         return req;
 954 }
 955 EXPORT_SYMBOL(inet_reqsk_alloc);
 956
 957 static struct request_sock *inet_reqsk_clone(struct request_sock *req,
 958                                              struct sock *sk)
 959 {
 960         struct sock *req_sk, *nreq_sk;
 961         struct request_sock *nreq;
 962
 963         nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN);
 964         if (!nreq) {
 965                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
 966
 967                 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */
 968                 sock_put(sk);
 969                 return NULL;
 970         }
 971
 972         req_sk = req_to_sk(req);
 973         nreq_sk = req_to_sk(nreq);
 974
 975         memcpy(nreq_sk, req_sk,
 976                offsetof(struct sock, sk_dontcopy_begin));
 977         unsafe_memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end,
 978                       req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end),
 979                       /* alloc is larger than struct, see above */);
 980
 981         sk_node_init(&nreq_sk->sk_node);
 982         nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping;
 983 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING
 984         nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping;
 985 #endif
 986         nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu;
 987
 988         nreq->rsk_listener = sk;
 989
 990         /* We need not acquire fastopenq->lock
 991          * because the child socket is locked in inet_csk_listen_stop().
 992          */
 993         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener)
 994                 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq);
 995
 996         return nreq;
 997 }
 998
 999 static void reqsk_queue_migrated(struct request_sock_queue *queue,
1000                                  const struct request_sock *req)
1001 {
1002         if (req->num_timeout == 0)
1003                 atomic_inc(&queue->young);
1004         atomic_inc(&queue->qlen);
1005 }
1006
1007 static void reqsk_migrate_reset(struct request_sock *req)
1008 {
1009         req->saved_syn = NULL;
1010 #if IS_ENABLED(CONFIG_IPV6)
1011         inet_rsk(req)->ipv6_opt = NULL;
1012         inet_rsk(req)->pktopts = NULL;
1013 #else
1014         inet_rsk(req)->ireq_opt = NULL;
1015 #endif
1016 }
1017
1018 /* return true if req was found in the ehash table */
1019 static bool reqsk_queue_unlink(struct request_sock *req)
1020 {
1021         struct sock *sk = req_to_sk(req);
1022         bool found = false;
1023
1024         if (sk_hashed(sk)) {
1025                 struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk);
1026                 spinlock_t *lock;
1027
1028                 lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
1029                 spin_lock(lock);
1030                 found = __sk_nulls_del_node_init_rcu(sk);
1031                 spin_unlock(lock);
1032         }
1033
1034         return found;
1035 }
1036
1037 static bool __inet_csk_reqsk_queue_drop(struct sock *sk,
1038                                         struct request_sock *req,
1039                                         bool from_timer)
1040 {
1041         bool unlinked = reqsk_queue_unlink(req);
1042
1043         if (!from_timer && timer_delete_sync(&req->rsk_timer))
1044                 reqsk_put(req);
1045
1046         if (unlinked) {
1047                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
1048                 reqsk_put(req);
1049         }
1050
1051         return unlinked;
1052 }
1053
1054 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
1055 {
1056         return __inet_csk_reqsk_queue_drop(sk, req, false);
1057 }
1058
1059 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
1060 {
1061         inet_csk_reqsk_queue_drop(sk, req);
1062         reqsk_put(req);
1063 }
1064 EXPORT_IPV6_MOD(inet_csk_reqsk_queue_drop_and_put);
1065
1066 static void reqsk_timer_handler(struct timer_list *t)
1067 {
1068         struct request_sock *req = timer_container_of(req, t, rsk_timer);
1069         struct request_sock *nreq = NULL, *oreq = req;
1070         struct sock *sk_listener = req->rsk_listener;
1071         struct inet_connection_sock *icsk;
1072         struct request_sock_queue *queue;
1073         struct net *net;
1074         int max_syn_ack_retries, qlen, expire = 0, resend = 0;
1075
1076         if (inet_sk_state_load(sk_listener) != TCP_LISTEN) {
1077                 struct sock *nsk;
1078
1079                 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL);
1080                 if (!nsk)
1081                         goto drop;
1082
1083                 nreq = inet_reqsk_clone(req, nsk);
1084                 if (!nreq)
1085                         goto drop;
1086
1087                 /* The new timer for the cloned req can decrease the 2
1088                  * by calling inet_csk_reqsk_queue_drop_and_put(), so
1089                  * hold another count to prevent use-after-free and
1090                  * call reqsk_put() just before return.
1091                  */
1092                 refcount_set(&nreq->rsk_refcnt, 2 + 1);
1093                 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1094                 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req);
1095
1096                 req = nreq;
1097                 sk_listener = nsk;
1098         }
1099
1100         icsk = inet_csk(sk_listener);
1101         net = sock_net(sk_listener);
1102         max_syn_ack_retries = READ_ONCE(icsk->icsk_syn_retries) ? :
1103                 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries);
1104         /* Normally all the openreqs are young and become mature
1105          * (i.e. converted to established socket) for first timeout.
1106          * If synack was not acknowledged for 1 second, it means
1107          * one of the following things: synack was lost, ack was lost,
1108          * rtt is high or nobody planned to ack (i.e. synflood).
1109          * When server is a bit loaded, queue is populated with old
1110          * open requests, reducing effective size of queue.
1111          * When server is well loaded, queue size reduces to zero
1112          * after several minutes of work. It is not synflood,
1113          * it is normal operation. The solution is pruning
1114          * too old entries overriding normal timeout, when
1115          * situation becomes dangerous.
1116          *
1117          * Essentially, we reserve half of room for young
1118          * embrions; and abort old ones without pity, if old
1119          * ones are about to clog our table.
1120          */
1121         queue = &icsk->icsk_accept_queue;
1122         qlen = reqsk_queue_len(queue);
1123         if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) {
1124                 int young = reqsk_queue_len_young(queue) << 1;
1125
1126                 while (max_syn_ack_retries > 2) {
1127                         if (qlen < young)
1128                                 break;
1129                         max_syn_ack_retries--;
1130                         young <<= 1;
1131                 }
1132         }
1133         syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept),
1134                        &expire, &resend);
1135         req->rsk_ops->syn_ack_timeout(req);
1136         if (!expire &&
1137             (!resend ||
1138              !inet_rtx_syn_ack(sk_listener, req) ||
1139              inet_rsk(req)->acked)) {
1140                 if (req->num_timeout++ == 0)
1141                         atomic_dec(&queue->young);
1142                 mod_timer(&req->rsk_timer, jiffies + reqsk_timeout(req, TCP_RTO_MAX));
1143
1144                 if (!nreq)
1145                         return;
1146
1147                 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
1148                         /* delete timer */
1149                         __inet_csk_reqsk_queue_drop(sk_listener, nreq, true);
1150                         goto no_ownership;
1151                 }
1152
1153                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS);
1154                 reqsk_migrate_reset(oreq);
1155                 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq);
1156                 reqsk_put(oreq);
1157
1158                 reqsk_put(nreq);
1159                 return;
1160         }
1161
1162         /* Even if we can clone the req, we may need not retransmit any more
1163          * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another
1164          * CPU may win the "own_req" race so that inet_ehash_insert() fails.
1165          */
1166         if (nreq) {
1167                 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE);
1168 no_ownership:
1169                 reqsk_migrate_reset(nreq);
1170                 reqsk_queue_removed(queue, nreq);
1171                 __reqsk_free(nreq);
1172         }
1173
1174 drop:
1175         __inet_csk_reqsk_queue_drop(sk_listener, oreq, true);
1176         reqsk_put(oreq);
1177 }
1178
1179 static bool reqsk_queue_hash_req(struct request_sock *req,
1180                                  unsigned long timeout)
1181 {
1182         bool found_dup_sk = false;
1183
1184         if (!inet_ehash_insert(req_to_sk(req), NULL, &found_dup_sk))
1185                 return false;
1186
1187         /* The timer needs to be setup after a successful insertion. */
1188         timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED);
1189         mod_timer(&req->rsk_timer, jiffies + timeout);
1190
1191         /* before letting lookups find us, make sure all req fields
1192          * are committed to memory and refcnt initialized.
1193          */
1194         smp_wmb();
1195         refcount_set(&req->rsk_refcnt, 2 + 1);
1196         return true;
1197 }
1198
1199 bool inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
1200                                    unsigned long timeout)
1201 {
1202         if (!reqsk_queue_hash_req(req, timeout))
1203                 return false;
1204
1205         inet_csk_reqsk_queue_added(sk);
1206         return true;
1207 }
1208
1209 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk,
1210                            const gfp_t priority)
1211 {
1212         struct inet_connection_sock *icsk = inet_csk(newsk);
1213
1214         if (!icsk->icsk_ulp_ops)
1215                 return;
1216
1217         icsk->icsk_ulp_ops->clone(req, newsk, priority);
1218 }
1219
1220 /**
1221  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
1222  *      @sk: the socket to clone
1223  *      @req: request_sock
1224  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
1225  *
1226  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
1227  */
1228 struct sock *inet_csk_clone_lock(const struct sock *sk,
1229                                  const struct request_sock *req,
1230                                  const gfp_t priority)
1231 {
1232         struct sock *newsk = sk_clone_lock(sk, priority);
1233         struct inet_connection_sock *newicsk;
1234         struct inet_request_sock *ireq;
1235         struct inet_sock *newinet;
1236
1237         if (!newsk)
1238                 return NULL;
1239
1240         newicsk = inet_csk(newsk);
1241         newinet = inet_sk(newsk);
1242         ireq = inet_rsk(req);
1243
1244         newicsk->icsk_bind_hash = NULL;
1245         newicsk->icsk_bind2_hash = NULL;
1246
1247         newinet->inet_dport = ireq->ir_rmt_port;
1248         newinet->inet_num = ireq->ir_num;
1249         newinet->inet_sport = htons(ireq->ir_num);
1250
1251         newsk->sk_bound_dev_if = ireq->ir_iif;
1252
1253         newsk->sk_daddr = ireq->ir_rmt_addr;
1254         newsk->sk_rcv_saddr = ireq->ir_loc_addr;
1255         newinet->inet_saddr = ireq->ir_loc_addr;
1256
1257 #if IS_ENABLED(CONFIG_IPV6)
1258         newsk->sk_v6_daddr = ireq->ir_v6_rmt_addr;
1259         newsk->sk_v6_rcv_saddr = ireq->ir_v6_loc_addr;
1260 #endif
1261
1262         /* listeners have SOCK_RCU_FREE, not the children */
1263         sock_reset_flag(newsk, SOCK_RCU_FREE);
1264
1265         inet_sk(newsk)->mc_list = NULL;
1266
1267         newsk->sk_mark = inet_rsk(req)->ir_mark;
1268         atomic64_set(&newsk->sk_cookie,
1269                      atomic64_read(&inet_rsk(req)->ir_cookie));
1270
1271         newicsk->icsk_retransmits = 0;
1272         newicsk->icsk_backoff     = 0;
1273         newicsk->icsk_probes_out  = 0;
1274         newicsk->icsk_probes_tstamp = 0;
1275
1276         /* Deinitialize accept_queue to trap illegal accesses. */
1277         memset(&newicsk->icsk_accept_queue, 0,
1278                sizeof(newicsk->icsk_accept_queue));
1279
1280         inet_sk_set_state(newsk, TCP_SYN_RECV);
1281
1282         inet_clone_ulp(req, newsk, priority);
1283
1284         security_inet_csk_clone(newsk, req);
1285
1286         return newsk;
1287 }
1288
1289 /*
1290  * At this point, there should be no process reference to this
1291  * socket, and thus no user references at all.  Therefore we
1292  * can assume the socket waitqueue is inactive and nobody will
1293  * try to jump onto it.
1294  */
1295 void inet_csk_destroy_sock(struct sock *sk)
1296 {
1297         WARN_ON(sk->sk_state != TCP_CLOSE);
1298         WARN_ON(!sock_flag(sk, SOCK_DEAD));
1299
1300         /* It cannot be in hash table! */
1301         WARN_ON(!sk_unhashed(sk));
1302
1303         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
1304         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
1305
1306         sk->sk_prot->destroy(sk);
1307
1308         sk_stream_kill_queues(sk);
1309
1310         xfrm_sk_free_policy(sk);
1311
1312         this_cpu_dec(*sk->sk_prot->orphan_count);
1313
1314         sock_put(sk);
1315 }
1316 EXPORT_SYMBOL(inet_csk_destroy_sock);
1317
1318 /* This function allows to force a closure of a socket after the call to
1319  * tcp_create_openreq_child().
1320  */
1321 void inet_csk_prepare_forced_close(struct sock *sk)
1322         __releases(&sk->sk_lock.slock)
1323 {
1324         /* sk_clone_lock locked the socket and set refcnt to 2 */
1325         bh_unlock_sock(sk);
1326         sock_put(sk);
1327         inet_csk_prepare_for_destroy_sock(sk);
1328         inet_sk(sk)->inet_num = 0;
1329 }
1330 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
1331
1332 static int inet_ulp_can_listen(const struct sock *sk)
1333 {
1334         const struct inet_connection_sock *icsk = inet_csk(sk);
1335
1336         if (icsk->icsk_ulp_ops && !icsk->icsk_ulp_ops->clone)
1337                 return -EINVAL;
1338
1339         return 0;
1340 }
1341
1342 int inet_csk_listen_start(struct sock *sk)
1343 {
1344         struct inet_connection_sock *icsk = inet_csk(sk);
1345         struct inet_sock *inet = inet_sk(sk);
1346         int err;
1347
1348         err = inet_ulp_can_listen(sk);
1349         if (unlikely(err))
1350                 return err;
1351
1352         reqsk_queue_alloc(&icsk->icsk_accept_queue);
1353
1354         sk->sk_ack_backlog = 0;
1355         inet_csk_delack_init(sk);
1356
1357         /* There is race window here: we announce ourselves listening,
1358          * but this transition is still not validated by get_port().
1359          * It is OK, because this socket enters to hash table only
1360          * after validation is complete.
1361          */
1362         inet_sk_state_store(sk, TCP_LISTEN);
1363         err = sk->sk_prot->get_port(sk, inet->inet_num);
1364         if (!err) {
1365                 inet->inet_sport = htons(inet->inet_num);
1366
1367                 sk_dst_reset(sk);
1368                 err = sk->sk_prot->hash(sk);
1369
1370                 if (likely(!err))
1371                         return 0;
1372         }
1373
1374         inet_sk_set_state(sk, TCP_CLOSE);
1375         return err;
1376 }
1377
1378 static void inet_child_forget(struct sock *sk, struct request_sock *req,
1379                               struct sock *child)
1380 {
1381         sk->sk_prot->disconnect(child, O_NONBLOCK);
1382
1383         sock_orphan(child);
1384
1385         this_cpu_inc(*sk->sk_prot->orphan_count);
1386
1387         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
1388                 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
1389                 BUG_ON(sk != req->rsk_listener);
1390
1391                 /* Paranoid, to prevent race condition if
1392                  * an inbound pkt destined for child is
1393                  * blocked by sock lock in tcp_v4_rcv().
1394                  * Also to satisfy an assertion in
1395                  * tcp_v4_destroy_sock().
1396                  */
1397                 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL);
1398         }
1399         inet_csk_destroy_sock(child);
1400 }
1401
1402 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
1403                                       struct request_sock *req,
1404                                       struct sock *child)
1405 {
1406         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
1407
1408         spin_lock(&queue->rskq_lock);
1409         if (unlikely(sk->sk_state != TCP_LISTEN)) {
1410                 inet_child_forget(sk, req, child);
1411                 child = NULL;
1412         } else {
1413                 req->sk = child;
1414                 req->dl_next = NULL;
1415                 if (queue->rskq_accept_head == NULL)
1416                         WRITE_ONCE(queue->rskq_accept_head, req);
1417                 else
1418                         queue->rskq_accept_tail->dl_next = req;
1419                 queue->rskq_accept_tail = req;
1420                 sk_acceptq_added(sk);
1421         }
1422         spin_unlock(&queue->rskq_lock);
1423         return child;
1424 }
1425 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
1426
1427 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
1428                                          struct request_sock *req, bool own_req)
1429 {
1430         if (own_req) {
1431                 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
1432                 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req);
1433
1434                 if (sk != req->rsk_listener) {
1435                         /* another listening sk has been selected,
1436                          * migrate the req to it.
1437                          */
1438                         struct request_sock *nreq;
1439
1440                         /* hold a refcnt for the nreq->rsk_listener
1441                          * which is assigned in inet_reqsk_clone()
1442                          */
1443                         sock_hold(sk);
1444                         nreq = inet_reqsk_clone(req, sk);
1445                         if (!nreq) {
1446                                 inet_child_forget(sk, req, child);
1447                                 goto child_put;
1448                         }
1449
1450                         refcount_set(&nreq->rsk_refcnt, 1);
1451                         if (inet_csk_reqsk_queue_add(sk, nreq, child)) {
1452                                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS);
1453                                 reqsk_migrate_reset(req);
1454                                 reqsk_put(req);
1455                                 return child;
1456                         }
1457
1458                         __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE);
1459                         reqsk_migrate_reset(nreq);
1460                         __reqsk_free(nreq);
1461                 } else if (inet_csk_reqsk_queue_add(sk, req, child)) {
1462                         return child;
1463                 }
1464         }
1465         /* Too bad, another child took ownership of the request, undo. */
1466 child_put:
1467         bh_unlock_sock(child);
1468         sock_put(child);
1469         return NULL;
1470 }
1471
1472 /*
1473  *      This routine closes sockets which have been at least partially
1474  *      opened, but not yet accepted.
1475  */
1476 void inet_csk_listen_stop(struct sock *sk)
1477 {
1478         struct inet_connection_sock *icsk = inet_csk(sk);
1479         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
1480         struct request_sock *next, *req;
1481
1482         /* Following specs, it would be better either to send FIN
1483          * (and enter FIN-WAIT-1, it is normal close)
1484          * or to send active reset (abort).
1485          * Certainly, it is pretty dangerous while synflood, but it is
1486          * bad justification for our negligence 8)
1487          * To be honest, we are not able to make either
1488          * of the variants now.                 --ANK
1489          */
1490         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
1491                 struct sock *child = req->sk, *nsk;
1492                 struct request_sock *nreq;
1493
1494                 local_bh_disable();
1495                 bh_lock_sock(child);
1496                 WARN_ON(sock_owned_by_user(child));
1497                 sock_hold(child);
1498
1499                 nsk = reuseport_migrate_sock(sk, child, NULL);
1500                 if (nsk) {
1501                         nreq = inet_reqsk_clone(req, nsk);
1502                         if (nreq) {
1503                                 refcount_set(&nreq->rsk_refcnt, 1);
1504
1505                                 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) {
1506                                         __NET_INC_STATS(sock_net(nsk),
1507                                                         LINUX_MIB_TCPMIGRATEREQSUCCESS);
1508                                         reqsk_migrate_reset(req);
1509                                 } else {
1510                                         __NET_INC_STATS(sock_net(nsk),
1511                                                         LINUX_MIB_TCPMIGRATEREQFAILURE);
1512                                         reqsk_migrate_reset(nreq);
1513                                         __reqsk_free(nreq);
1514                                 }
1515
1516                                 /* inet_csk_reqsk_queue_add() has already
1517                                  * called inet_child_forget() on failure case.
1518                                  */
1519                                 goto skip_child_forget;
1520                         }
1521                 }
1522
1523                 inet_child_forget(sk, req, child);
1524 skip_child_forget:
1525                 reqsk_put(req);
1526                 bh_unlock_sock(child);
1527                 local_bh_enable();
1528                 sock_put(child);
1529
1530                 cond_resched();
1531         }
1532         if (queue->fastopenq.rskq_rst_head) {
1533                 /* Free all the reqs queued in rskq_rst_head. */
1534                 spin_lock_bh(&queue->fastopenq.lock);
1535                 req = queue->fastopenq.rskq_rst_head;
1536                 queue->fastopenq.rskq_rst_head = NULL;
1537                 spin_unlock_bh(&queue->fastopenq.lock);
1538                 while (req != NULL) {
1539                         next = req->dl_next;
1540                         reqsk_put(req);
1541                         req = next;
1542                 }
1543         }
1544         WARN_ON_ONCE(sk->sk_ack_backlog);
1545 }
1546 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
1547
1548 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
1549 {
1550         const struct inet_sock *inet = inet_sk(sk);
1551         struct flowi4 *fl4;
1552         struct rtable *rt;
1553
1554         rcu_read_lock();
1555         fl4 = &fl->u.ip4;
1556         inet_sk_init_flowi4(inet, fl4);
1557         rt = ip_route_output_flow(sock_net(sk), fl4, sk);
1558         if (IS_ERR(rt))
1559                 rt = NULL;
1560         if (rt)
1561                 sk_setup_caps(sk, &rt->dst);
1562         rcu_read_unlock();
1563
1564         return &rt->dst;
1565 }
1566
1567 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
1568 {
1569         struct dst_entry *dst = __sk_dst_check(sk, 0);
1570         struct inet_sock *inet = inet_sk(sk);
1571
1572         if (!dst) {
1573                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1574                 if (!dst)
1575                         goto out;
1576         }
1577         dst->ops->update_pmtu(dst, sk, NULL, mtu, true);
1578
1579         dst = __sk_dst_check(sk, 0);
1580         if (!dst)
1581                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
1582 out:
1583         return dst;
1584 }