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
  27 #ifdef INET_CSK_DEBUG
  28 const char inet_csk_timer_bug_msg[] = "inet_csk BUG: unknown timer value\n";
  29 EXPORT_SYMBOL(inet_csk_timer_bug_msg);
  30 #endif
  31
  32 /*
  33  * This struct holds the first and last local port number.
  34  */
  35 struct local_ports sysctl_local_ports __read_mostly = {
  36         .lock = __SEQLOCK_UNLOCKED(sysctl_local_ports.lock),
  37         .range = { 32768, 61000 },
  38 };
  39
  40 unsigned long *sysctl_local_reserved_ports;
  41 EXPORT_SYMBOL(sysctl_local_reserved_ports);
  42
  43 void inet_get_local_port_range(int *low, int *high)
  44 {
  45         unsigned int seq;
  46
  47         do {
  48                 seq = read_seqbegin(&sysctl_local_ports.lock);
  49
  50                 *low = sysctl_local_ports.range[0];
  51                 *high = sysctl_local_ports.range[1];
  52         } while (read_seqretry(&sysctl_local_ports.lock, seq));
  53 }
  54 EXPORT_SYMBOL(inet_get_local_port_range);
  55
  56 int inet_csk_bind_conflict(const struct sock *sk,
  57                            const struct inet_bind_bucket *tb, bool relax)
  58 {
  59         struct sock *sk2;
  60         int reuse = sk->sk_reuse;
  61         int reuseport = sk->sk_reuseport;
  62         kuid_t uid = sock_i_uid((struct sock *)sk);
  63
  64         /*
  65          * Unlike other sk lookup places we do not check
  66          * for sk_net here, since _all_ the socks listed
  67          * in tb->owners list belong to the same net - the
  68          * one this bucket belongs to.
  69          */
  70
  71         sk_for_each_bound(sk2, &tb->owners) {
  72                 if (sk != sk2 &&
  73                     !inet_v6_ipv6only(sk2) &&
  74                     (!sk->sk_bound_dev_if ||
  75                      !sk2->sk_bound_dev_if ||
  76                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
  77                         if ((!reuse || !sk2->sk_reuse ||
  78                             sk2->sk_state == TCP_LISTEN) &&
  79                             (!reuseport || !sk2->sk_reuseport ||
  80                             (sk2->sk_state != TCP_TIME_WAIT &&
  81                              !uid_eq(uid, sock_i_uid(sk2))))) {
  82                                 const __be32 sk2_rcv_saddr = sk_rcv_saddr(sk2);
  83                                 if (!sk2_rcv_saddr || !sk_rcv_saddr(sk) ||
  84                                     sk2_rcv_saddr == sk_rcv_saddr(sk))
  85                                         break;
  86                         }
  87                         if (!relax && reuse && sk2->sk_reuse &&
  88                             sk2->sk_state != TCP_LISTEN) {
  89                                 const __be32 sk2_rcv_saddr = sk_rcv_saddr(sk2);
  90
  91                                 if (!sk2_rcv_saddr || !sk_rcv_saddr(sk) ||
  92                                     sk2_rcv_saddr == sk_rcv_saddr(sk))
  93                                         break;
  94                         }
  95                 }
  96         }
  97         return sk2 != NULL;
  98 }
  99 EXPORT_SYMBOL_GPL(inet_csk_bind_conflict);
 100
 101 /* Obtain a reference to a local port for the given sock,
 102  * if snum is zero it means select any available local port.
 103  */
 104 int inet_csk_get_port(struct sock *sk, unsigned short snum)
 105 {
 106         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
 107         struct inet_bind_hashbucket *head;
 108         struct inet_bind_bucket *tb;
 109         int ret, attempts = 5;
 110         struct net *net = sock_net(sk);
 111         int smallest_size = -1, smallest_rover;
 112         kuid_t uid = sock_i_uid(sk);
 113
 114         local_bh_disable();
 115         if (!snum) {
 116                 int remaining, rover, low, high;
 117
 118 again:
 119                 inet_get_local_port_range(&low, &high);
 120                 remaining = (high - low) + 1;
 121                 smallest_rover = rover = net_random() % remaining + low;
 122
 123                 smallest_size = -1;
 124                 do {
 125                         if (inet_is_reserved_local_port(rover))
 126                                 goto next_nolock;
 127                         head = &hashinfo->bhash[inet_bhashfn(net, rover,
 128                                         hashinfo->bhash_size)];
 129                         spin_lock(&head->lock);
 130                         inet_bind_bucket_for_each(tb, &head->chain)
 131                                 if (net_eq(ib_net(tb), net) && tb->port == rover) {
 132                                         if (((tb->fastreuse > 0 &&
 133                                               sk->sk_reuse &&
 134                                               sk->sk_state != TCP_LISTEN) ||
 135                                              (tb->fastreuseport > 0 &&
 136                                               sk->sk_reuseport &&
 137                                               uid_eq(tb->fastuid, uid))) &&
 138                                             (tb->num_owners < smallest_size || smallest_size == -1)) {
 139                                                 smallest_size = tb->num_owners;
 140                                                 smallest_rover = rover;
 141                                                 if (atomic_read(&hashinfo->bsockets) > (high - low) + 1 &&
 142                                                     !inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, false)) {
 143                                                         snum = smallest_rover;
 144                                                         goto tb_found;
 145                                                 }
 146                                         }
 147                                         if (!inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, false)) {
 148                                                 snum = rover;
 149                                                 goto tb_found;
 150                                         }
 151                                         goto next;
 152                                 }
 153                         break;
 154                 next:
 155                         spin_unlock(&head->lock);
 156                 next_nolock:
 157                         if (++rover > high)
 158                                 rover = low;
 159                 } while (--remaining > 0);
 160
 161                 /* Exhausted local port range during search?  It is not
 162                  * possible for us to be holding one of the bind hash
 163                  * locks if this test triggers, because if 'remaining'
 164                  * drops to zero, we broke out of the do/while loop at
 165                  * the top level, not from the 'break;' statement.
 166                  */
 167                 ret = 1;
 168                 if (remaining <= 0) {
 169                         if (smallest_size != -1) {
 170                                 snum = smallest_rover;
 171                                 goto have_snum;
 172                         }
 173                         goto fail;
 174                 }
 175                 /* OK, here is the one we will use.  HEAD is
 176                  * non-NULL and we hold it's mutex.
 177                  */
 178                 snum = rover;
 179         } else {
 180 have_snum:
 181                 head = &hashinfo->bhash[inet_bhashfn(net, snum,
 182                                 hashinfo->bhash_size)];
 183                 spin_lock(&head->lock);
 184                 inet_bind_bucket_for_each(tb, &head->chain)
 185                         if (net_eq(ib_net(tb), net) && tb->port == snum)
 186                                 goto tb_found;
 187         }
 188         tb = NULL;
 189         goto tb_not_found;
 190 tb_found:
 191         if (!hlist_empty(&tb->owners)) {
 192                 if (sk->sk_reuse == SK_FORCE_REUSE)
 193                         goto success;
 194
 195                 if (((tb->fastreuse > 0 &&
 196                       sk->sk_reuse && sk->sk_state != TCP_LISTEN) ||
 197                      (tb->fastreuseport > 0 &&
 198                       sk->sk_reuseport && uid_eq(tb->fastuid, uid))) &&
 199                     smallest_size == -1) {
 200                         goto success;
 201                 } else {
 202                         ret = 1;
 203                         if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, true)) {
 204                                 if (((sk->sk_reuse && sk->sk_state != TCP_LISTEN) ||
 205                                      (tb->fastreuseport > 0 &&
 206                                       sk->sk_reuseport && uid_eq(tb->fastuid, uid))) &&
 207                                     smallest_size != -1 && --attempts >= 0) {
 208                                         spin_unlock(&head->lock);
 209                                         goto again;
 210                                 }
 211
 212                                 goto fail_unlock;
 213                         }
 214                 }
 215         }
 216 tb_not_found:
 217         ret = 1;
 218         if (!tb && (tb = inet_bind_bucket_create(hashinfo->bind_bucket_cachep,
 219                                         net, head, snum)) == NULL)
 220                 goto fail_unlock;
 221         if (hlist_empty(&tb->owners)) {
 222                 if (sk->sk_reuse && sk->sk_state != TCP_LISTEN)
 223                         tb->fastreuse = 1;
 224                 else
 225                         tb->fastreuse = 0;
 226                 if (sk->sk_reuseport) {
 227                         tb->fastreuseport = 1;
 228                         tb->fastuid = uid;
 229                 } else
 230                         tb->fastreuseport = 0;
 231         } else {
 232                 if (tb->fastreuse &&
 233                     (!sk->sk_reuse || sk->sk_state == TCP_LISTEN))
 234                         tb->fastreuse = 0;
 235                 if (tb->fastreuseport &&
 236                     (!sk->sk_reuseport || !uid_eq(tb->fastuid, uid)))
 237                         tb->fastreuseport = 0;
 238         }
 239 success:
 240         if (!inet_csk(sk)->icsk_bind_hash)
 241                 inet_bind_hash(sk, tb, snum);
 242         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 243         ret = 0;
 244
 245 fail_unlock:
 246         spin_unlock(&head->lock);
 247 fail:
 248         local_bh_enable();
 249         return ret;
 250 }
 251 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 252
 253 /*
 254  * Wait for an incoming connection, avoid race conditions. This must be called
 255  * with the socket locked.
 256  */
 257 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 258 {
 259         struct inet_connection_sock *icsk = inet_csk(sk);
 260         DEFINE_WAIT(wait);
 261         int err;
 262
 263         /*
 264          * True wake-one mechanism for incoming connections: only
 265          * one process gets woken up, not the 'whole herd'.
 266          * Since we do not 'race & poll' for established sockets
 267          * anymore, the common case will execute the loop only once.
 268          *
 269          * Subtle issue: "add_wait_queue_exclusive()" will be added
 270          * after any current non-exclusive waiters, and we know that
 271          * it will always _stay_ after any new non-exclusive waiters
 272          * because all non-exclusive waiters are added at the
 273          * beginning of the wait-queue. As such, it's ok to "drop"
 274          * our exclusiveness temporarily when we get woken up without
 275          * having to remove and re-insert us on the wait queue.
 276          */
 277         for (;;) {
 278                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 279                                           TASK_INTERRUPTIBLE);
 280                 release_sock(sk);
 281                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 282                         timeo = schedule_timeout(timeo);
 283                 lock_sock(sk);
 284                 err = 0;
 285                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 286                         break;
 287                 err = -EINVAL;
 288                 if (sk->sk_state != TCP_LISTEN)
 289                         break;
 290                 err = sock_intr_errno(timeo);
 291                 if (signal_pending(current))
 292                         break;
 293                 err = -EAGAIN;
 294                 if (!timeo)
 295                         break;
 296         }
 297         finish_wait(sk_sleep(sk), &wait);
 298         return err;
 299 }
 300
 301 /*
 302  * This will accept the next outstanding connection.
 303  */
 304 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err)
 305 {
 306         struct inet_connection_sock *icsk = inet_csk(sk);
 307         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 308         struct sock *newsk;
 309         struct request_sock *req;
 310         int error;
 311
 312         lock_sock(sk);
 313
 314         /* We need to make sure that this socket is listening,
 315          * and that it has something pending.
 316          */
 317         error = -EINVAL;
 318         if (sk->sk_state != TCP_LISTEN)
 319                 goto out_err;
 320
 321         /* Find already established connection */
 322         if (reqsk_queue_empty(queue)) {
 323                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 324
 325                 /* If this is a non blocking socket don't sleep */
 326                 error = -EAGAIN;
 327                 if (!timeo)
 328                         goto out_err;
 329
 330                 error = inet_csk_wait_for_connect(sk, timeo);
 331                 if (error)
 332                         goto out_err;
 333         }
 334         req = reqsk_queue_remove(queue);
 335         newsk = req->sk;
 336
 337         sk_acceptq_removed(sk);
 338         if (sk->sk_protocol == IPPROTO_TCP && queue->fastopenq != NULL) {
 339                 spin_lock_bh(&queue->fastopenq->lock);
 340                 if (tcp_rsk(req)->listener) {
 341                         /* We are still waiting for the final ACK from 3WHS
 342                          * so can't free req now. Instead, we set req->sk to
 343                          * NULL to signify that the child socket is taken
 344                          * so reqsk_fastopen_remove() will free the req
 345                          * when 3WHS finishes (or is aborted).
 346                          */
 347                         req->sk = NULL;
 348                         req = NULL;
 349                 }
 350                 spin_unlock_bh(&queue->fastopenq->lock);
 351         }
 352 out:
 353         release_sock(sk);
 354         if (req)
 355                 __reqsk_free(req);
 356         return newsk;
 357 out_err:
 358         newsk = NULL;
 359         req = NULL;
 360         *err = error;
 361         goto out;
 362 }
 363 EXPORT_SYMBOL(inet_csk_accept);
 364
 365 /*
 366  * Using different timers for retransmit, delayed acks and probes
 367  * We may wish use just one timer maintaining a list of expire jiffies
 368  * to optimize.
 369  */
 370 void inet_csk_init_xmit_timers(struct sock *sk,
 371                                void (*retransmit_handler)(unsigned long),
 372                                void (*delack_handler)(unsigned long),
 373                                void (*keepalive_handler)(unsigned long))
 374 {
 375         struct inet_connection_sock *icsk = inet_csk(sk);
 376
 377         setup_timer(&icsk->icsk_retransmit_timer, retransmit_handler,
 378                         (unsigned long)sk);
 379         setup_timer(&icsk->icsk_delack_timer, delack_handler,
 380                         (unsigned long)sk);
 381         setup_timer(&sk->sk_timer, keepalive_handler, (unsigned long)sk);
 382         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 383 }
 384 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 385
 386 void inet_csk_clear_xmit_timers(struct sock *sk)
 387 {
 388         struct inet_connection_sock *icsk = inet_csk(sk);
 389
 390         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 391
 392         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 393         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 394         sk_stop_timer(sk, &sk->sk_timer);
 395 }
 396 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 397
 398 void inet_csk_delete_keepalive_timer(struct sock *sk)
 399 {
 400         sk_stop_timer(sk, &sk->sk_timer);
 401 }
 402 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 403
 404 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 405 {
 406         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 407 }
 408 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 409
 410 struct dst_entry *inet_csk_route_req(struct sock *sk,
 411                                      struct flowi4 *fl4,
 412                                      const struct request_sock *req)
 413 {
 414         struct rtable *rt;
 415         const struct inet_request_sock *ireq = inet_rsk(req);
 416         struct ip_options_rcu *opt = inet_rsk(req)->opt;
 417         struct net *net = sock_net(sk);
 418         int flags = inet_sk_flowi_flags(sk);
 419
 420         flowi4_init_output(fl4, sk->sk_bound_dev_if, sk->sk_mark,
 421                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 422                            sk->sk_protocol,
 423                            flags,
 424                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->rmt_addr,
 425                            ireq->loc_addr, ireq->rmt_port, inet_sk(sk)->inet_sport);
 426         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 427         rt = ip_route_output_flow(net, fl4, sk);
 428         if (IS_ERR(rt))
 429                 goto no_route;
 430         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 431                 goto route_err;
 432         return &rt->dst;
 433
 434 route_err:
 435         ip_rt_put(rt);
 436 no_route:
 437         IP_INC_STATS_BH(net, IPSTATS_MIB_OUTNOROUTES);
 438         return NULL;
 439 }
 440 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 441
 442 struct dst_entry *inet_csk_route_child_sock(struct sock *sk,
 443                                             struct sock *newsk,
 444                                             const struct request_sock *req)
 445 {
 446         const struct inet_request_sock *ireq = inet_rsk(req);
 447         struct inet_sock *newinet = inet_sk(newsk);
 448         struct ip_options_rcu *opt;
 449         struct net *net = sock_net(sk);
 450         struct flowi4 *fl4;
 451         struct rtable *rt;
 452
 453         fl4 = &newinet->cork.fl.u.ip4;
 454
 455         rcu_read_lock();
 456         opt = rcu_dereference(newinet->inet_opt);
 457         flowi4_init_output(fl4, sk->sk_bound_dev_if, sk->sk_mark,
 458                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 459                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 460                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->rmt_addr,
 461                            ireq->loc_addr, ireq->rmt_port, inet_sk(sk)->inet_sport);
 462         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 463         rt = ip_route_output_flow(net, fl4, sk);
 464         if (IS_ERR(rt))
 465                 goto no_route;
 466         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 467                 goto route_err;
 468         rcu_read_unlock();
 469         return &rt->dst;
 470
 471 route_err:
 472         ip_rt_put(rt);
 473 no_route:
 474         rcu_read_unlock();
 475         IP_INC_STATS_BH(net, IPSTATS_MIB_OUTNOROUTES);
 476         return NULL;
 477 }
 478 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 479
 480 static inline u32 inet_synq_hash(const __be32 raddr, const __be16 rport,
 481                                  const u32 rnd, const u32 synq_hsize)
 482 {
 483         return jhash_2words((__force u32)raddr, (__force u32)rport, rnd) & (synq_hsize - 1);
 484 }
 485
 486 #if IS_ENABLED(CONFIG_IPV6)
 487 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 488 #else
 489 #define AF_INET_FAMILY(fam) 1
 490 #endif
 491
 492 struct request_sock *inet_csk_search_req(const struct sock *sk,
 493                                          struct request_sock ***prevp,
 494                                          const __be16 rport, const __be32 raddr,
 495                                          const __be32 laddr)
 496 {
 497         const struct inet_connection_sock *icsk = inet_csk(sk);
 498         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
 499         struct request_sock *req, **prev;
 500
 501         for (prev = &lopt->syn_table[inet_synq_hash(raddr, rport, lopt->hash_rnd,
 502                                                     lopt->nr_table_entries)];
 503              (req = *prev) != NULL;
 504              prev = &req->dl_next) {
 505                 const struct inet_request_sock *ireq = inet_rsk(req);
 506
 507                 if (ireq->rmt_port == rport &&
 508                     ireq->rmt_addr == raddr &&
 509                     ireq->loc_addr == laddr &&
 510                     AF_INET_FAMILY(req->rsk_ops->family)) {
 511                         WARN_ON(req->sk);
 512                         *prevp = prev;
 513                         break;
 514                 }
 515         }
 516
 517         return req;
 518 }
 519 EXPORT_SYMBOL_GPL(inet_csk_search_req);
 520
 521 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 522                                    unsigned long timeout)
 523 {
 524         struct inet_connection_sock *icsk = inet_csk(sk);
 525         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
 526         const u32 h = inet_synq_hash(inet_rsk(req)->rmt_addr, inet_rsk(req)->rmt_port,
 527                                      lopt->hash_rnd, lopt->nr_table_entries);
 528
 529         reqsk_queue_hash_req(&icsk->icsk_accept_queue, h, req, timeout);
 530         inet_csk_reqsk_queue_added(sk, timeout);
 531 }
 532 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 533
 534 /* Only thing we need from tcp.h */
 535 extern int sysctl_tcp_synack_retries;
 536
 537
 538 /* Decide when to expire the request and when to resend SYN-ACK */
 539 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 540                                   const int max_retries,
 541                                   const u8 rskq_defer_accept,
 542                                   int *expire, int *resend)
 543 {
 544         if (!rskq_defer_accept) {
 545                 *expire = req->num_timeout >= thresh;
 546                 *resend = 1;
 547                 return;
 548         }
 549         *expire = req->num_timeout >= thresh &&
 550                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 551         /*
 552          * Do not resend while waiting for data after ACK,
 553          * start to resend on end of deferring period to give
 554          * last chance for data or ACK to create established socket.
 555          */
 556         *resend = !inet_rsk(req)->acked ||
 557                   req->num_timeout >= rskq_defer_accept - 1;
 558 }
 559
 560 int inet_rtx_syn_ack(struct sock *parent, struct request_sock *req)
 561 {
 562         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 563
 564         if (!err)
 565                 req->num_retrans++;
 566         return err;
 567 }
 568 EXPORT_SYMBOL(inet_rtx_syn_ack);
 569
 570 void inet_csk_reqsk_queue_prune(struct sock *parent,
 571                                 const unsigned long interval,
 572                                 const unsigned long timeout,
 573                                 const unsigned long max_rto)
 574 {
 575         struct inet_connection_sock *icsk = inet_csk(parent);
 576         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 577         struct listen_sock *lopt = queue->listen_opt;
 578         int max_retries = icsk->icsk_syn_retries ? : sysctl_tcp_synack_retries;
 579         int thresh = max_retries;
 580         unsigned long now = jiffies;
 581         struct request_sock **reqp, *req;
 582         int i, budget;
 583
 584         if (lopt == NULL || lopt->qlen == 0)
 585                 return;
 586
 587         /* Normally all the openreqs are young and become mature
 588          * (i.e. converted to established socket) for first timeout.
 589          * If synack was not acknowledged for 1 second, it means
 590          * one of the following things: synack was lost, ack was lost,
 591          * rtt is high or nobody planned to ack (i.e. synflood).
 592          * When server is a bit loaded, queue is populated with old
 593          * open requests, reducing effective size of queue.
 594          * When server is well loaded, queue size reduces to zero
 595          * after several minutes of work. It is not synflood,
 596          * it is normal operation. The solution is pruning
 597          * too old entries overriding normal timeout, when
 598          * situation becomes dangerous.
 599          *
 600          * Essentially, we reserve half of room for young
 601          * embrions; and abort old ones without pity, if old
 602          * ones are about to clog our table.
 603          */
 604         if (lopt->qlen>>(lopt->max_qlen_log-1)) {
 605                 int young = (lopt->qlen_young<<1);
 606
 607                 while (thresh > 2) {
 608                         if (lopt->qlen < young)
 609                                 break;
 610                         thresh--;
 611                         young <<= 1;
 612                 }
 613         }
 614
 615         if (queue->rskq_defer_accept)
 616                 max_retries = queue->rskq_defer_accept;
 617
 618         budget = 2 * (lopt->nr_table_entries / (timeout / interval));
 619         i = lopt->clock_hand;
 620
 621         do {
 622                 reqp=&lopt->syn_table[i];
 623                 while ((req = *reqp) != NULL) {
 624                         if (time_after_eq(now, req->expires)) {
 625                                 int expire = 0, resend = 0;
 626
 627                                 syn_ack_recalc(req, thresh, max_retries,
 628                                                queue->rskq_defer_accept,
 629                                                &expire, &resend);
 630                                 req->rsk_ops->syn_ack_timeout(parent, req);
 631                                 if (!expire &&
 632                                     (!resend ||
 633                                      !inet_rtx_syn_ack(parent, req) ||
 634                                      inet_rsk(req)->acked)) {
 635                                         unsigned long timeo;
 636
 637                                         if (req->num_timeout++ == 0)
 638                                                 lopt->qlen_young--;
 639                                         timeo = min(timeout << req->num_timeout,
 640                                                     max_rto);
 641                                         req->expires = now + timeo;
 642                                         reqp = &req->dl_next;
 643                                         continue;
 644                                 }
 645
 646                                 /* Drop this request */
 647                                 inet_csk_reqsk_queue_unlink(parent, req, reqp);
 648                                 reqsk_queue_removed(queue, req);
 649                                 reqsk_free(req);
 650                                 continue;
 651                         }
 652                         reqp = &req->dl_next;
 653                 }
 654
 655                 i = (i + 1) & (lopt->nr_table_entries - 1);
 656
 657         } while (--budget > 0);
 658
 659         lopt->clock_hand = i;
 660
 661         if (lopt->qlen)
 662                 inet_csk_reset_keepalive_timer(parent, interval);
 663 }
 664 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_prune);
 665
 666 /**
 667  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 668  *      @sk: the socket to clone
 669  *      @req: request_sock
 670  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 671  *
 672  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 673  */
 674 struct sock *inet_csk_clone_lock(const struct sock *sk,
 675                                  const struct request_sock *req,
 676                                  const gfp_t priority)
 677 {
 678         struct sock *newsk = sk_clone_lock(sk, priority);
 679
 680         if (newsk != NULL) {
 681                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 682
 683                 newsk->sk_state = TCP_SYN_RECV;
 684                 newicsk->icsk_bind_hash = NULL;
 685
 686                 inet_sk(newsk)->inet_dport = inet_rsk(req)->rmt_port;
 687                 inet_sk(newsk)->inet_num = ntohs(inet_rsk(req)->loc_port);
 688                 inet_sk(newsk)->inet_sport = inet_rsk(req)->loc_port;
 689                 newsk->sk_write_space = sk_stream_write_space;
 690
 691                 newicsk->icsk_retransmits = 0;
 692                 newicsk->icsk_backoff     = 0;
 693                 newicsk->icsk_probes_out  = 0;
 694
 695                 /* Deinitialize accept_queue to trap illegal accesses. */
 696                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 697
 698                 security_inet_csk_clone(newsk, req);
 699         }
 700         return newsk;
 701 }
 702 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 703
 704 /*
 705  * At this point, there should be no process reference to this
 706  * socket, and thus no user references at all.  Therefore we
 707  * can assume the socket waitqueue is inactive and nobody will
 708  * try to jump onto it.
 709  */
 710 void inet_csk_destroy_sock(struct sock *sk)
 711 {
 712         WARN_ON(sk->sk_state != TCP_CLOSE);
 713         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 714
 715         /* It cannot be in hash table! */
 716         WARN_ON(!sk_unhashed(sk));
 717
 718         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 719         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 720
 721         sk->sk_prot->destroy(sk);
 722
 723         sk_stream_kill_queues(sk);
 724
 725         xfrm_sk_free_policy(sk);
 726
 727         sk_refcnt_debug_release(sk);
 728
 729         percpu_counter_dec(sk->sk_prot->orphan_count);
 730         sock_put(sk);
 731 }
 732 EXPORT_SYMBOL(inet_csk_destroy_sock);
 733
 734 /* This function allows to force a closure of a socket after the call to
 735  * tcp/dccp_create_openreq_child().
 736  */
 737 void inet_csk_prepare_forced_close(struct sock *sk)
 738         __releases(&sk->sk_lock.slock)
 739 {
 740         /* sk_clone_lock locked the socket and set refcnt to 2 */
 741         bh_unlock_sock(sk);
 742         sock_put(sk);
 743
 744         /* The below has to be done to allow calling inet_csk_destroy_sock */
 745         sock_set_flag(sk, SOCK_DEAD);
 746         percpu_counter_inc(sk->sk_prot->orphan_count);
 747         inet_sk(sk)->inet_num = 0;
 748 }
 749 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 750
 751 int inet_csk_listen_start(struct sock *sk, const int nr_table_entries)
 752 {
 753         struct inet_sock *inet = inet_sk(sk);
 754         struct inet_connection_sock *icsk = inet_csk(sk);
 755         int rc = reqsk_queue_alloc(&icsk->icsk_accept_queue, nr_table_entries);
 756
 757         if (rc != 0)
 758                 return rc;
 759
 760         sk->sk_max_ack_backlog = 0;
 761         sk->sk_ack_backlog = 0;
 762         inet_csk_delack_init(sk);
 763
 764         /* There is race window here: we announce ourselves listening,
 765          * but this transition is still not validated by get_port().
 766          * It is OK, because this socket enters to hash table only
 767          * after validation is complete.
 768          */
 769         sk->sk_state = TCP_LISTEN;
 770         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 771                 inet->inet_sport = htons(inet->inet_num);
 772
 773                 sk_dst_reset(sk);
 774                 sk->sk_prot->hash(sk);
 775
 776                 return 0;
 777         }
 778
 779         sk->sk_state = TCP_CLOSE;
 780         __reqsk_queue_destroy(&icsk->icsk_accept_queue);
 781         return -EADDRINUSE;
 782 }
 783 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 784
 785 /*
 786  *      This routine closes sockets which have been at least partially
 787  *      opened, but not yet accepted.
 788  */
 789 void inet_csk_listen_stop(struct sock *sk)
 790 {
 791         struct inet_connection_sock *icsk = inet_csk(sk);
 792         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 793         struct request_sock *acc_req;
 794         struct request_sock *req;
 795
 796         inet_csk_delete_keepalive_timer(sk);
 797
 798         /* make all the listen_opt local to us */
 799         acc_req = reqsk_queue_yank_acceptq(queue);
 800
 801         /* Following specs, it would be better either to send FIN
 802          * (and enter FIN-WAIT-1, it is normal close)
 803          * or to send active reset (abort).
 804          * Certainly, it is pretty dangerous while synflood, but it is
 805          * bad justification for our negligence 8)
 806          * To be honest, we are not able to make either
 807          * of the variants now.                 --ANK
 808          */
 809         reqsk_queue_destroy(queue);
 810
 811         while ((req = acc_req) != NULL) {
 812                 struct sock *child = req->sk;
 813
 814                 acc_req = req->dl_next;
 815
 816                 local_bh_disable();
 817                 bh_lock_sock(child);
 818                 WARN_ON(sock_owned_by_user(child));
 819                 sock_hold(child);
 820
 821                 sk->sk_prot->disconnect(child, O_NONBLOCK);
 822
 823                 sock_orphan(child);
 824
 825                 percpu_counter_inc(sk->sk_prot->orphan_count);
 826
 827                 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->listener) {
 828                         BUG_ON(tcp_sk(child)->fastopen_rsk != req);
 829                         BUG_ON(sk != tcp_rsk(req)->listener);
 830
 831                         /* Paranoid, to prevent race condition if
 832                          * an inbound pkt destined for child is
 833                          * blocked by sock lock in tcp_v4_rcv().
 834                          * Also to satisfy an assertion in
 835                          * tcp_v4_destroy_sock().
 836                          */
 837                         tcp_sk(child)->fastopen_rsk = NULL;
 838                         sock_put(sk);
 839                 }
 840                 inet_csk_destroy_sock(child);
 841
 842                 bh_unlock_sock(child);
 843                 local_bh_enable();
 844                 sock_put(child);
 845
 846                 sk_acceptq_removed(sk);
 847                 __reqsk_free(req);
 848         }
 849         if (queue->fastopenq != NULL) {
 850                 /* Free all the reqs queued in rskq_rst_head. */
 851                 spin_lock_bh(&queue->fastopenq->lock);
 852                 acc_req = queue->fastopenq->rskq_rst_head;
 853                 queue->fastopenq->rskq_rst_head = NULL;
 854                 spin_unlock_bh(&queue->fastopenq->lock);
 855                 while ((req = acc_req) != NULL) {
 856                         acc_req = req->dl_next;
 857                         __reqsk_free(req);
 858                 }
 859         }
 860         WARN_ON(sk->sk_ack_backlog);
 861 }
 862 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
 863
 864 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
 865 {
 866         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
 867         const struct inet_sock *inet = inet_sk(sk);
 868
 869         sin->sin_family         = AF_INET;
 870         sin->sin_addr.s_addr    = inet->inet_daddr;
 871         sin->sin_port           = inet->inet_dport;
 872 }
 873 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
 874
 875 #ifdef CONFIG_COMPAT
 876 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
 877                                char __user *optval, int __user *optlen)
 878 {
 879         const struct inet_connection_sock *icsk = inet_csk(sk);
 880
 881         if (icsk->icsk_af_ops->compat_getsockopt != NULL)
 882                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
 883                                                             optval, optlen);
 884         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
 885                                              optval, optlen);
 886 }
 887 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
 888
 889 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
 890                                char __user *optval, unsigned int optlen)
 891 {
 892         const struct inet_connection_sock *icsk = inet_csk(sk);
 893
 894         if (icsk->icsk_af_ops->compat_setsockopt != NULL)
 895                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
 896                                                             optval, optlen);
 897         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
 898                                              optval, optlen);
 899 }
 900 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
 901 #endif
 902
 903 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
 904 {
 905         const struct inet_sock *inet = inet_sk(sk);
 906         const struct ip_options_rcu *inet_opt;
 907         __be32 daddr = inet->inet_daddr;
 908         struct flowi4 *fl4;
 909         struct rtable *rt;
 910
 911         rcu_read_lock();
 912         inet_opt = rcu_dereference(inet->inet_opt);
 913         if (inet_opt && inet_opt->opt.srr)
 914                 daddr = inet_opt->opt.faddr;
 915         fl4 = &fl->u.ip4;
 916         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
 917                                    inet->inet_saddr, inet->inet_dport,
 918                                    inet->inet_sport, sk->sk_protocol,
 919                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
 920         if (IS_ERR(rt))
 921                 rt = NULL;
 922         if (rt)
 923                 sk_setup_caps(sk, &rt->dst);
 924         rcu_read_unlock();
 925
 926         return &rt->dst;
 927 }
 928
 929 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
 930 {
 931         struct dst_entry *dst = __sk_dst_check(sk, 0);
 932         struct inet_sock *inet = inet_sk(sk);
 933
 934         if (!dst) {
 935                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 936                 if (!dst)
 937                         goto out;
 938         }
 939         dst->ops->update_pmtu(dst, sk, NULL, mtu);
 940
 941         dst = __sk_dst_check(sk, 0);
 942         if (!dst)
 943                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 944 out:
 945         return dst;
 946 }
 947 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);