net/ipv4/inet_connection_sock.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Support for INET connection oriented protocols.
   7  *
   8  * Authors:     See the TCP sources
   9  *
  10  *              This program is free software; you can redistribute it and/or
  11  *              modify it under the terms of the GNU General Public License
  12  *              as published by the Free Software Foundation; either version
  13  *              2 of the License, or(at your option) any later version.
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/jhash.h>
  18
  19 #include <net/inet_connection_sock.h>
  20 #include <net/inet_hashtables.h>
  21 #include <net/inet_timewait_sock.h>
  22 #include <net/ip.h>
  23 #include <net/route.h>
  24 #include <net/tcp_states.h>
  25 #include <net/xfrm.h>
  26 #include <net/tcp.h>
  27 #include <net/sock_reuseport.h>
  28
  29 #ifdef INET_CSK_DEBUG
  30 const char inet_csk_timer_bug_msg[] = "inet_csk BUG: unknown timer value\n";
  31 EXPORT_SYMBOL(inet_csk_timer_bug_msg);
  32 #endif
  33
  34 void inet_get_local_port_range(struct net *net, int *low, int *high)
  35 {
  36         unsigned int seq;
  37
  38         do {
  39                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
  40
  41                 *low = net->ipv4.ip_local_ports.range[0];
  42                 *high = net->ipv4.ip_local_ports.range[1];
  43         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
  44 }
  45 EXPORT_SYMBOL(inet_get_local_port_range);
  46
  47 int inet_csk_bind_conflict(const struct sock *sk,
  48                            const struct inet_bind_bucket *tb, bool relax,
  49                            bool reuseport_ok)
  50 {
  51         struct sock *sk2;
  52         bool reuse = sk->sk_reuse;
  53         bool reuseport = !!sk->sk_reuseport && reuseport_ok;
  54         kuid_t uid = sock_i_uid((struct sock *)sk);
  55
  56         /*
  57          * Unlike other sk lookup places we do not check
  58          * for sk_net here, since _all_ the socks listed
  59          * in tb->owners list belong to the same net - the
  60          * one this bucket belongs to.
  61          */
  62
  63         sk_for_each_bound(sk2, &tb->owners) {
  64                 if (sk != sk2 &&
  65                     !inet_v6_ipv6only(sk2) &&
  66                     (!sk->sk_bound_dev_if ||
  67                      !sk2->sk_bound_dev_if ||
  68                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
  69                         if ((!reuse || !sk2->sk_reuse ||
  70                             sk2->sk_state == TCP_LISTEN) &&
  71                             (!reuseport || !sk2->sk_reuseport ||
  72                              rcu_access_pointer(sk->sk_reuseport_cb) ||
  73                              (sk2->sk_state != TCP_TIME_WAIT &&
  74                              !uid_eq(uid, sock_i_uid(sk2))))) {
  75
  76                                 if (!sk2->sk_rcv_saddr || !sk->sk_rcv_saddr ||
  77                                     sk2->sk_rcv_saddr == sk->sk_rcv_saddr)
  78                                         break;
  79                         }
  80                         if (!relax && reuse && sk2->sk_reuse &&
  81                             sk2->sk_state != TCP_LISTEN) {
  82
  83                                 if (!sk2->sk_rcv_saddr || !sk->sk_rcv_saddr ||
  84                                     sk2->sk_rcv_saddr == sk->sk_rcv_saddr)
  85                                         break;
  86                         }
  87                 }
  88         }
  89         return sk2 != NULL;
  90 }
  91 EXPORT_SYMBOL_GPL(inet_csk_bind_conflict);
  92
  93 /* Obtain a reference to a local port for the given sock,
  94  * if snum is zero it means select any available local port.
  95  * We try to allocate an odd port (and leave even ports for connect())
  96  */
  97 int inet_csk_get_port(struct sock *sk, unsigned short snum)
  98 {
  99         bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN;
 100         struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;
 101         int ret = 1, attempts = 5, port = snum;
 102         int smallest_size = -1, smallest_port;
 103         struct inet_bind_hashbucket *head;
 104         struct net *net = sock_net(sk);
 105         int i, low, high, attempt_half;
 106         struct inet_bind_bucket *tb;
 107         kuid_t uid = sock_i_uid(sk);
 108         u32 remaining, offset;
 109         bool reuseport_ok = !!snum;
 110
 111         if (port) {
 112 have_port:
 113                 head = &hinfo->bhash[inet_bhashfn(net, port,
 114                                                   hinfo->bhash_size)];
 115                 spin_lock_bh(&head->lock);
 116                 inet_bind_bucket_for_each(tb, &head->chain)
 117                         if (net_eq(ib_net(tb), net) && tb->port == port)
 118                                 goto tb_found;
 119
 120                 goto tb_not_found;
 121         }
 122 again:
 123         attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 124 other_half_scan:
 125         inet_get_local_port_range(net, &low, &high);
 126         high++; /* [32768, 60999] -> [32768, 61000[ */
 127         if (high - low < 4)
 128                 attempt_half = 0;
 129         if (attempt_half) {
 130                 int half = low + (((high - low) >> 2) << 1);
 131
 132                 if (attempt_half == 1)
 133                         high = half;
 134                 else
 135                         low = half;
 136         }
 137         remaining = high - low;
 138         if (likely(remaining > 1))
 139                 remaining &= ~1U;
 140
 141         offset = prandom_u32() % remaining;
 142         /* __inet_hash_connect() favors ports having @low parity
 143          * We do the opposite to not pollute connect() users.
 144          */
 145         offset |= 1U;
 146         smallest_size = -1;
 147         smallest_port = low; /* avoid compiler warning */
 148
 149 other_parity_scan:
 150         port = low + offset;
 151         for (i = 0; i < remaining; i += 2, port += 2) {
 152                 if (unlikely(port >= high))
 153                         port -= remaining;
 154                 if (inet_is_local_reserved_port(net, port))
 155                         continue;
 156                 head = &hinfo->bhash[inet_bhashfn(net, port,
 157                                                   hinfo->bhash_size)];
 158                 spin_lock_bh(&head->lock);
 159                 inet_bind_bucket_for_each(tb, &head->chain)
 160                         if (net_eq(ib_net(tb), net) && tb->port == port) {
 161                                 if (((tb->fastreuse > 0 && reuse) ||
 162                                      (tb->fastreuseport > 0 &&
 163                                       sk->sk_reuseport &&
 164                                       !rcu_access_pointer(sk->sk_reuseport_cb) &&
 165                                       uid_eq(tb->fastuid, uid))) &&
 166                                     (tb->num_owners < smallest_size || smallest_size == -1)) {
 167                                         smallest_size = tb->num_owners;
 168                                         smallest_port = port;
 169                                 }
 170                                 if (!inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, false,
 171                                                                               reuseport_ok))
 172                                         goto tb_found;
 173                                 goto next_port;
 174                         }
 175                 goto tb_not_found;
 176 next_port:
 177                 spin_unlock_bh(&head->lock);
 178                 cond_resched();
 179         }
 180
 181         if (smallest_size != -1) {
 182                 port = smallest_port;
 183                 goto have_port;
 184         }
 185         offset--;
 186         if (!(offset & 1))
 187                 goto other_parity_scan;
 188
 189         if (attempt_half == 1) {
 190                 /* OK we now try the upper half of the range */
 191                 attempt_half = 2;
 192                 goto other_half_scan;
 193         }
 194         return ret;
 195
 196 tb_not_found:
 197         tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 198                                      net, head, port);
 199         if (!tb)
 200                 goto fail_unlock;
 201 tb_found:
 202         if (!hlist_empty(&tb->owners)) {
 203                 if (sk->sk_reuse == SK_FORCE_REUSE)
 204                         goto success;
 205
 206                 if (((tb->fastreuse > 0 && reuse) ||
 207                      (tb->fastreuseport > 0 &&
 208                       !rcu_access_pointer(sk->sk_reuseport_cb) &&
 209                       sk->sk_reuseport && uid_eq(tb->fastuid, uid))) &&
 210                     smallest_size == -1)
 211                         goto success;
 212                 if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, true,
 213                                                              reuseport_ok)) {
 214                         if ((reuse ||
 215                              (tb->fastreuseport > 0 &&
 216                               sk->sk_reuseport &&
 217                               !rcu_access_pointer(sk->sk_reuseport_cb) &&
 218                               uid_eq(tb->fastuid, uid))) &&
 219                             !snum && smallest_size != -1 && --attempts >= 0) {
 220                                 spin_unlock_bh(&head->lock);
 221                                 goto again;
 222                         }
 223                         goto fail_unlock;
 224                 }
 225                 if (!reuse)
 226                         tb->fastreuse = 0;
 227                 if (!sk->sk_reuseport || !uid_eq(tb->fastuid, uid))
 228                         tb->fastreuseport = 0;
 229         } else {
 230                 tb->fastreuse = reuse;
 231                 if (sk->sk_reuseport) {
 232                         tb->fastreuseport = 1;
 233                         tb->fastuid = uid;
 234                 } else {
 235                         tb->fastreuseport = 0;
 236                 }
 237         }
 238 success:
 239         if (!inet_csk(sk)->icsk_bind_hash)
 240                 inet_bind_hash(sk, tb, port);
 241         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 242         ret = 0;
 243
 244 fail_unlock:
 245         spin_unlock_bh(&head->lock);
 246         return ret;
 247 }
 248 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 249
 250 /*
 251  * Wait for an incoming connection, avoid race conditions. This must be called
 252  * with the socket locked.
 253  */
 254 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 255 {
 256         struct inet_connection_sock *icsk = inet_csk(sk);
 257         DEFINE_WAIT(wait);
 258         int err;
 259
 260         /*
 261          * True wake-one mechanism for incoming connections: only
 262          * one process gets woken up, not the 'whole herd'.
 263          * Since we do not 'race & poll' for established sockets
 264          * anymore, the common case will execute the loop only once.
 265          *
 266          * Subtle issue: "add_wait_queue_exclusive()" will be added
 267          * after any current non-exclusive waiters, and we know that
 268          * it will always _stay_ after any new non-exclusive waiters
 269          * because all non-exclusive waiters are added at the
 270          * beginning of the wait-queue. As such, it's ok to "drop"
 271          * our exclusiveness temporarily when we get woken up without
 272          * having to remove and re-insert us on the wait queue.
 273          */
 274         for (;;) {
 275                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 276                                           TASK_INTERRUPTIBLE);
 277                 release_sock(sk);
 278                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 279                         timeo = schedule_timeout(timeo);
 280                 sched_annotate_sleep();
 281                 lock_sock(sk);
 282                 err = 0;
 283                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 284                         break;
 285                 err = -EINVAL;
 286                 if (sk->sk_state != TCP_LISTEN)
 287                         break;
 288                 err = sock_intr_errno(timeo);
 289                 if (signal_pending(current))
 290                         break;
 291                 err = -EAGAIN;
 292                 if (!timeo)
 293                         break;
 294         }
 295         finish_wait(sk_sleep(sk), &wait);
 296         return err;
 297 }
 298
 299 /*
 300  * This will accept the next outstanding connection.
 301  */
 302 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err)
 303 {
 304         struct inet_connection_sock *icsk = inet_csk(sk);
 305         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 306         struct request_sock *req;
 307         struct sock *newsk;
 308         int error;
 309
 310         lock_sock(sk);
 311
 312         /* We need to make sure that this socket is listening,
 313          * and that it has something pending.
 314          */
 315         error = -EINVAL;
 316         if (sk->sk_state != TCP_LISTEN)
 317                 goto out_err;
 318
 319         /* Find already established connection */
 320         if (reqsk_queue_empty(queue)) {
 321                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 322
 323                 /* If this is a non blocking socket don't sleep */
 324                 error = -EAGAIN;
 325                 if (!timeo)
 326                         goto out_err;
 327
 328                 error = inet_csk_wait_for_connect(sk, timeo);
 329                 if (error)
 330                         goto out_err;
 331         }
 332         req = reqsk_queue_remove(queue, sk);
 333         newsk = req->sk;
 334
 335         if (sk->sk_protocol == IPPROTO_TCP &&
 336             tcp_rsk(req)->tfo_listener) {
 337                 spin_lock_bh(&queue->fastopenq.lock);
 338                 if (tcp_rsk(req)->tfo_listener) {
 339                         /* We are still waiting for the final ACK from 3WHS
 340                          * so can't free req now. Instead, we set req->sk to
 341                          * NULL to signify that the child socket is taken
 342                          * so reqsk_fastopen_remove() will free the req
 343                          * when 3WHS finishes (or is aborted).
 344                          */
 345                         req->sk = NULL;
 346                         req = NULL;
 347                 }
 348                 spin_unlock_bh(&queue->fastopenq.lock);
 349         }
 350 out:
 351         release_sock(sk);
 352         if (req)
 353                 reqsk_put(req);
 354         return newsk;
 355 out_err:
 356         newsk = NULL;
 357         req = NULL;
 358         *err = error;
 359         goto out;
 360 }
 361 EXPORT_SYMBOL(inet_csk_accept);
 362
 363 /*
 364  * Using different timers for retransmit, delayed acks and probes
 365  * We may wish use just one timer maintaining a list of expire jiffies
 366  * to optimize.
 367  */
 368 void inet_csk_init_xmit_timers(struct sock *sk,
 369                                void (*retransmit_handler)(unsigned long),
 370                                void (*delack_handler)(unsigned long),
 371                                void (*keepalive_handler)(unsigned long))
 372 {
 373         struct inet_connection_sock *icsk = inet_csk(sk);
 374
 375         setup_timer(&icsk->icsk_retransmit_timer, retransmit_handler,
 376                         (unsigned long)sk);
 377         setup_timer(&icsk->icsk_delack_timer, delack_handler,
 378                         (unsigned long)sk);
 379         setup_timer(&sk->sk_timer, keepalive_handler, (unsigned long)sk);
 380         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 381 }
 382 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 383
 384 void inet_csk_clear_xmit_timers(struct sock *sk)
 385 {
 386         struct inet_connection_sock *icsk = inet_csk(sk);
 387
 388         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 389
 390         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 391         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 392         sk_stop_timer(sk, &sk->sk_timer);
 393 }
 394 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 395
 396 void inet_csk_delete_keepalive_timer(struct sock *sk)
 397 {
 398         sk_stop_timer(sk, &sk->sk_timer);
 399 }
 400 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 401
 402 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 403 {
 404         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 405 }
 406 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 407
 408 struct dst_entry *inet_csk_route_req(const struct sock *sk,
 409                                      struct flowi4 *fl4,
 410                                      const struct request_sock *req)
 411 {
 412         const struct inet_request_sock *ireq = inet_rsk(req);
 413         struct net *net = read_pnet(&ireq->ireq_net);
 414         struct ip_options_rcu *opt = ireq->opt;
 415         struct rtable *rt;
 416
 417         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 418                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 419                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 420                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 421                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 422                            htons(ireq->ir_num), sk->sk_uid);
 423         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 424         rt = ip_route_output_flow(net, fl4, sk);
 425         if (IS_ERR(rt))
 426                 goto no_route;
 427         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 428                 goto route_err;
 429         return &rt->dst;
 430
 431 route_err:
 432         ip_rt_put(rt);
 433 no_route:
 434         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 435         return NULL;
 436 }
 437 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 438
 439 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk,
 440                                             struct sock *newsk,
 441                                             const struct request_sock *req)
 442 {
 443         const struct inet_request_sock *ireq = inet_rsk(req);
 444         struct net *net = read_pnet(&ireq->ireq_net);
 445         struct inet_sock *newinet = inet_sk(newsk);
 446         struct ip_options_rcu *opt;
 447         struct flowi4 *fl4;
 448         struct rtable *rt;
 449
 450         fl4 = &newinet->cork.fl.u.ip4;
 451
 452         rcu_read_lock();
 453         opt = rcu_dereference(newinet->inet_opt);
 454         flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark,
 455                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 456                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 457                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 458                            ireq->ir_loc_addr, ireq->ir_rmt_port,
 459                            htons(ireq->ir_num), sk->sk_uid);
 460         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 461         rt = ip_route_output_flow(net, fl4, sk);
 462         if (IS_ERR(rt))
 463                 goto no_route;
 464         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 465                 goto route_err;
 466         rcu_read_unlock();
 467         return &rt->dst;
 468
 469 route_err:
 470         ip_rt_put(rt);
 471 no_route:
 472         rcu_read_unlock();
 473         __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
 474         return NULL;
 475 }
 476 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 477
 478 #if IS_ENABLED(CONFIG_IPV6)
 479 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 480 #else
 481 #define AF_INET_FAMILY(fam) true
 482 #endif
 483
 484 /* Decide when to expire the request and when to resend SYN-ACK */
 485 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 486                                   const int max_retries,
 487                                   const u8 rskq_defer_accept,
 488                                   int *expire, int *resend)
 489 {
 490         if (!rskq_defer_accept) {
 491                 *expire = req->num_timeout >= thresh;
 492                 *resend = 1;
 493                 return;
 494         }
 495         *expire = req->num_timeout >= thresh &&
 496                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 497         /*
 498          * Do not resend while waiting for data after ACK,
 499          * start to resend on end of deferring period to give
 500          * last chance for data or ACK to create established socket.
 501          */
 502         *resend = !inet_rsk(req)->acked ||
 503                   req->num_timeout >= rskq_defer_accept - 1;
 504 }
 505
 506 int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req)
 507 {
 508         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 509
 510         if (!err)
 511                 req->num_retrans++;
 512         return err;
 513 }
 514 EXPORT_SYMBOL(inet_rtx_syn_ack);
 515
 516 /* return true if req was found in the ehash table */
 517 static bool reqsk_queue_unlink(struct request_sock_queue *queue,
 518                                struct request_sock *req)
 519 {
 520         struct inet_hashinfo *hashinfo = req_to_sk(req)->sk_prot->h.hashinfo;
 521         bool found = false;
 522
 523         if (sk_hashed(req_to_sk(req))) {
 524                 spinlock_t *lock = inet_ehash_lockp(hashinfo, req->rsk_hash);
 525
 526                 spin_lock(lock);
 527                 found = __sk_nulls_del_node_init_rcu(req_to_sk(req));
 528                 spin_unlock(lock);
 529         }
 530         if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
 531                 reqsk_put(req);
 532         return found;
 533 }
 534
 535 void inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
 536 {
 537         if (reqsk_queue_unlink(&inet_csk(sk)->icsk_accept_queue, req)) {
 538                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 539                 reqsk_put(req);
 540         }
 541 }
 542 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 543
 544 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
 545 {
 546         inet_csk_reqsk_queue_drop(sk, req);
 547         reqsk_put(req);
 548 }
 549 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop_and_put);
 550
 551 static void reqsk_timer_handler(unsigned long data)
 552 {
 553         struct request_sock *req = (struct request_sock *)data;
 554         struct sock *sk_listener = req->rsk_listener;
 555         struct net *net = sock_net(sk_listener);
 556         struct inet_connection_sock *icsk = inet_csk(sk_listener);
 557         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 558         int qlen, expire = 0, resend = 0;
 559         int max_retries, thresh;
 560         u8 defer_accept;
 561
 562         if (sk_state_load(sk_listener) != TCP_LISTEN)
 563                 goto drop;
 564
 565         max_retries = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_synack_retries;
 566         thresh = max_retries;
 567         /* Normally all the openreqs are young and become mature
 568          * (i.e. converted to established socket) for first timeout.
 569          * If synack was not acknowledged for 1 second, it means
 570          * one of the following things: synack was lost, ack was lost,
 571          * rtt is high or nobody planned to ack (i.e. synflood).
 572          * When server is a bit loaded, queue is populated with old
 573          * open requests, reducing effective size of queue.
 574          * When server is well loaded, queue size reduces to zero
 575          * after several minutes of work. It is not synflood,
 576          * it is normal operation. The solution is pruning
 577          * too old entries overriding normal timeout, when
 578          * situation becomes dangerous.
 579          *
 580          * Essentially, we reserve half of room for young
 581          * embrions; and abort old ones without pity, if old
 582          * ones are about to clog our table.
 583          */
 584         qlen = reqsk_queue_len(queue);
 585         if ((qlen << 1) > max(8U, sk_listener->sk_max_ack_backlog)) {
 586                 int young = reqsk_queue_len_young(queue) << 1;
 587
 588                 while (thresh > 2) {
 589                         if (qlen < young)
 590                                 break;
 591                         thresh--;
 592                         young <<= 1;
 593                 }
 594         }
 595         defer_accept = READ_ONCE(queue->rskq_defer_accept);
 596         if (defer_accept)
 597                 max_retries = defer_accept;
 598         syn_ack_recalc(req, thresh, max_retries, defer_accept,
 599                        &expire, &resend);
 600         req->rsk_ops->syn_ack_timeout(req);
 601         if (!expire &&
 602             (!resend ||
 603              !inet_rtx_syn_ack(sk_listener, req) ||
 604              inet_rsk(req)->acked)) {
 605                 unsigned long timeo;
 606
 607                 if (req->num_timeout++ == 0)
 608                         atomic_dec(&queue->young);
 609                 timeo = min(TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 610                 mod_timer(&req->rsk_timer, jiffies + timeo);
 611                 return;
 612         }
 613 drop:
 614         inet_csk_reqsk_queue_drop_and_put(sk_listener, req);
 615 }
 616
 617 static void reqsk_queue_hash_req(struct request_sock *req,
 618                                  unsigned long timeout)
 619 {
 620         req->num_retrans = 0;
 621         req->num_timeout = 0;
 622         req->sk = NULL;
 623
 624         setup_pinned_timer(&req->rsk_timer, reqsk_timer_handler,
 625                             (unsigned long)req);
 626         mod_timer(&req->rsk_timer, jiffies + timeout);
 627
 628         inet_ehash_insert(req_to_sk(req), NULL);
 629         /* before letting lookups find us, make sure all req fields
 630          * are committed to memory and refcnt initialized.
 631          */
 632         smp_wmb();
 633         atomic_set(&req->rsk_refcnt, 2 + 1);
 634 }
 635
 636 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 637                                    unsigned long timeout)
 638 {
 639         reqsk_queue_hash_req(req, timeout);
 640         inet_csk_reqsk_queue_added(sk);
 641 }
 642 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 643
 644 /**
 645  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 646  *      @sk: the socket to clone
 647  *      @req: request_sock
 648  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 649  *
 650  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 651  */
 652 struct sock *inet_csk_clone_lock(const struct sock *sk,
 653                                  const struct request_sock *req,
 654                                  const gfp_t priority)
 655 {
 656         struct sock *newsk = sk_clone_lock(sk, priority);
 657
 658         if (newsk) {
 659                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 660
 661                 newsk->sk_state = TCP_SYN_RECV;
 662                 newicsk->icsk_bind_hash = NULL;
 663
 664                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 665                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 666                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 667                 newsk->sk_write_space = sk_stream_write_space;
 668
 669                 /* listeners have SOCK_RCU_FREE, not the children */
 670                 sock_reset_flag(newsk, SOCK_RCU_FREE);
 671
 672                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 673                 atomic64_set(&newsk->sk_cookie,
 674                              atomic64_read(&inet_rsk(req)->ir_cookie));
 675
 676                 newicsk->icsk_retransmits = 0;
 677                 newicsk->icsk_backoff     = 0;
 678                 newicsk->icsk_probes_out  = 0;
 679
 680                 /* Deinitialize accept_queue to trap illegal accesses. */
 681                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 682
 683                 security_inet_csk_clone(newsk, req);
 684         }
 685         return newsk;
 686 }
 687 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 688
 689 /*
 690  * At this point, there should be no process reference to this
 691  * socket, and thus no user references at all.  Therefore we
 692  * can assume the socket waitqueue is inactive and nobody will
 693  * try to jump onto it.
 694  */
 695 void inet_csk_destroy_sock(struct sock *sk)
 696 {
 697         WARN_ON(sk->sk_state != TCP_CLOSE);
 698         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 699
 700         /* It cannot be in hash table! */
 701         WARN_ON(!sk_unhashed(sk));
 702
 703         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 704         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 705
 706         sk->sk_prot->destroy(sk);
 707
 708         sk_stream_kill_queues(sk);
 709
 710         xfrm_sk_free_policy(sk);
 711
 712         sk_refcnt_debug_release(sk);
 713
 714         local_bh_disable();
 715         percpu_counter_dec(sk->sk_prot->orphan_count);
 716         local_bh_enable();
 717         sock_put(sk);
 718 }
 719 EXPORT_SYMBOL(inet_csk_destroy_sock);
 720
 721 /* This function allows to force a closure of a socket after the call to
 722  * tcp/dccp_create_openreq_child().
 723  */
 724 void inet_csk_prepare_forced_close(struct sock *sk)
 725         __releases(&sk->sk_lock.slock)
 726 {
 727         /* sk_clone_lock locked the socket and set refcnt to 2 */
 728         bh_unlock_sock(sk);
 729         sock_put(sk);
 730
 731         /* The below has to be done to allow calling inet_csk_destroy_sock */
 732         sock_set_flag(sk, SOCK_DEAD);
 733         percpu_counter_inc(sk->sk_prot->orphan_count);
 734         inet_sk(sk)->inet_num = 0;
 735 }
 736 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 737
 738 int inet_csk_listen_start(struct sock *sk, int backlog)
 739 {
 740         struct inet_connection_sock *icsk = inet_csk(sk);
 741         struct inet_sock *inet = inet_sk(sk);
 742         int err = -EADDRINUSE;
 743
 744         reqsk_queue_alloc(&icsk->icsk_accept_queue);
 745
 746         sk->sk_max_ack_backlog = backlog;
 747         sk->sk_ack_backlog = 0;
 748         inet_csk_delack_init(sk);
 749
 750         /* There is race window here: we announce ourselves listening,
 751          * but this transition is still not validated by get_port().
 752          * It is OK, because this socket enters to hash table only
 753          * after validation is complete.
 754          */
 755         sk_state_store(sk, TCP_LISTEN);
 756         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 757                 inet->inet_sport = htons(inet->inet_num);
 758
 759                 sk_dst_reset(sk);
 760                 err = sk->sk_prot->hash(sk);
 761
 762                 if (likely(!err))
 763                         return 0;
 764         }
 765
 766         sk->sk_state = TCP_CLOSE;
 767         return err;
 768 }
 769 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 770
 771 static void inet_child_forget(struct sock *sk, struct request_sock *req,
 772                               struct sock *child)
 773 {
 774         sk->sk_prot->disconnect(child, O_NONBLOCK);
 775
 776         sock_orphan(child);
 777
 778         percpu_counter_inc(sk->sk_prot->orphan_count);
 779
 780         if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 781                 BUG_ON(tcp_sk(child)->fastopen_rsk != req);
 782                 BUG_ON(sk != req->rsk_listener);
 783
 784                 /* Paranoid, to prevent race condition if
 785                  * an inbound pkt destined for child is
 786                  * blocked by sock lock in tcp_v4_rcv().
 787                  * Also to satisfy an assertion in
 788                  * tcp_v4_destroy_sock().
 789                  */
 790                 tcp_sk(child)->fastopen_rsk = NULL;
 791         }
 792         inet_csk_destroy_sock(child);
 793         reqsk_put(req);
 794 }
 795
 796 struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
 797                                       struct request_sock *req,
 798                                       struct sock *child)
 799 {
 800         struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
 801
 802         spin_lock(&queue->rskq_lock);
 803         if (unlikely(sk->sk_state != TCP_LISTEN)) {
 804                 inet_child_forget(sk, req, child);
 805                 child = NULL;
 806         } else {
 807                 req->sk = child;
 808                 req->dl_next = NULL;
 809                 if (queue->rskq_accept_head == NULL)
 810                         queue->rskq_accept_head = req;
 811                 else
 812                         queue->rskq_accept_tail->dl_next = req;
 813                 queue->rskq_accept_tail = req;
 814                 sk_acceptq_added(sk);
 815         }
 816         spin_unlock(&queue->rskq_lock);
 817         return child;
 818 }
 819 EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
 820
 821 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child,
 822                                          struct request_sock *req, bool own_req)
 823 {
 824         if (own_req) {
 825                 inet_csk_reqsk_queue_drop(sk, req);
 826                 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 827                 if (inet_csk_reqsk_queue_add(sk, req, child))
 828                         return child;
 829         }
 830         /* Too bad, another child took ownership of the request, undo. */
 831         bh_unlock_sock(child);
 832         sock_put(child);
 833         return NULL;
 834 }
 835 EXPORT_SYMBOL(inet_csk_complete_hashdance);
 836
 837 /*
 838  *      This routine closes sockets which have been at least partially
 839  *      opened, but not yet accepted.
 840  */
 841 void inet_csk_listen_stop(struct sock *sk)
 842 {
 843         struct inet_connection_sock *icsk = inet_csk(sk);
 844         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 845         struct request_sock *next, *req;
 846
 847         /* Following specs, it would be better either to send FIN
 848          * (and enter FIN-WAIT-1, it is normal close)
 849          * or to send active reset (abort).
 850          * Certainly, it is pretty dangerous while synflood, but it is
 851          * bad justification for our negligence 8)
 852          * To be honest, we are not able to make either
 853          * of the variants now.                 --ANK
 854          */
 855         while ((req = reqsk_queue_remove(queue, sk)) != NULL) {
 856                 struct sock *child = req->sk;
 857
 858                 local_bh_disable();
 859                 bh_lock_sock(child);
 860                 WARN_ON(sock_owned_by_user(child));
 861                 sock_hold(child);
 862
 863                 inet_child_forget(sk, req, child);
 864                 bh_unlock_sock(child);
 865                 local_bh_enable();
 866                 sock_put(child);
 867
 868                 cond_resched();
 869         }
 870         if (queue->fastopenq.rskq_rst_head) {
 871                 /* Free all the reqs queued in rskq_rst_head. */
 872                 spin_lock_bh(&queue->fastopenq.lock);
 873                 req = queue->fastopenq.rskq_rst_head;
 874                 queue->fastopenq.rskq_rst_head = NULL;
 875                 spin_unlock_bh(&queue->fastopenq.lock);
 876                 while (req != NULL) {
 877                         next = req->dl_next;
 878                         reqsk_put(req);
 879                         req = next;
 880                 }
 881         }
 882         WARN_ON_ONCE(sk->sk_ack_backlog);
 883 }
 884 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
 885
 886 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
 887 {
 888         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
 889         const struct inet_sock *inet = inet_sk(sk);
 890
 891         sin->sin_family         = AF_INET;
 892         sin->sin_addr.s_addr    = inet->inet_daddr;
 893         sin->sin_port           = inet->inet_dport;
 894 }
 895 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
 896
 897 #ifdef CONFIG_COMPAT
 898 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
 899                                char __user *optval, int __user *optlen)
 900 {
 901         const struct inet_connection_sock *icsk = inet_csk(sk);
 902
 903         if (icsk->icsk_af_ops->compat_getsockopt)
 904                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
 905                                                             optval, optlen);
 906         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
 907                                              optval, optlen);
 908 }
 909 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
 910
 911 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
 912                                char __user *optval, unsigned int optlen)
 913 {
 914         const struct inet_connection_sock *icsk = inet_csk(sk);
 915
 916         if (icsk->icsk_af_ops->compat_setsockopt)
 917                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
 918                                                             optval, optlen);
 919         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
 920                                              optval, optlen);
 921 }
 922 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
 923 #endif
 924
 925 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
 926 {
 927         const struct inet_sock *inet = inet_sk(sk);
 928         const struct ip_options_rcu *inet_opt;
 929         __be32 daddr = inet->inet_daddr;
 930         struct flowi4 *fl4;
 931         struct rtable *rt;
 932
 933         rcu_read_lock();
 934         inet_opt = rcu_dereference(inet->inet_opt);
 935         if (inet_opt && inet_opt->opt.srr)
 936                 daddr = inet_opt->opt.faddr;
 937         fl4 = &fl->u.ip4;
 938         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
 939                                    inet->inet_saddr, inet->inet_dport,
 940                                    inet->inet_sport, sk->sk_protocol,
 941                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
 942         if (IS_ERR(rt))
 943                 rt = NULL;
 944         if (rt)
 945                 sk_setup_caps(sk, &rt->dst);
 946         rcu_read_unlock();
 947
 948         return &rt->dst;
 949 }
 950
 951 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
 952 {
 953         struct dst_entry *dst = __sk_dst_check(sk, 0);
 954         struct inet_sock *inet = inet_sk(sk);
 955
 956         if (!dst) {
 957                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 958                 if (!dst)
 959                         goto out;
 960         }
 961         dst->ops->update_pmtu(dst, sk, NULL, mtu);
 962
 963         dst = __sk_dst_check(sk, 0);
 964         if (!dst)
 965                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 966 out:
 967         return dst;
 968 }
 969 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);