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 void inet_get_local_port_range(struct net *net, int *low, int *high)
  33 {
  34         unsigned int seq;
  35
  36         do {
  37                 seq = read_seqbegin(&net->ipv4.ip_local_ports.lock);
  38
  39                 *low = net->ipv4.ip_local_ports.range[0];
  40                 *high = net->ipv4.ip_local_ports.range[1];
  41         } while (read_seqretry(&net->ipv4.ip_local_ports.lock, seq));
  42 }
  43 EXPORT_SYMBOL(inet_get_local_port_range);
  44
  45 int inet_csk_bind_conflict(const struct sock *sk,
  46                            const struct inet_bind_bucket *tb, bool relax)
  47 {
  48         struct sock *sk2;
  49         int reuse = sk->sk_reuse;
  50         int reuseport = sk->sk_reuseport;
  51         kuid_t uid = sock_i_uid((struct sock *)sk);
  52
  53         /*
  54          * Unlike other sk lookup places we do not check
  55          * for sk_net here, since _all_ the socks listed
  56          * in tb->owners list belong to the same net - the
  57          * one this bucket belongs to.
  58          */
  59
  60         sk_for_each_bound(sk2, &tb->owners) {
  61                 if (sk != sk2 &&
  62                     !inet_v6_ipv6only(sk2) &&
  63                     (!sk->sk_bound_dev_if ||
  64                      !sk2->sk_bound_dev_if ||
  65                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
  66                         if ((!reuse || !sk2->sk_reuse ||
  67                             sk2->sk_state == TCP_LISTEN) &&
  68                             (!reuseport || !sk2->sk_reuseport ||
  69                             (sk2->sk_state != TCP_TIME_WAIT &&
  70                              !uid_eq(uid, sock_i_uid(sk2))))) {
  71
  72                                 if (!sk2->sk_rcv_saddr || !sk->sk_rcv_saddr ||
  73                                     sk2->sk_rcv_saddr == sk->sk_rcv_saddr)
  74                                         break;
  75                         }
  76                         if (!relax && reuse && sk2->sk_reuse &&
  77                             sk2->sk_state != TCP_LISTEN) {
  78
  79                                 if (!sk2->sk_rcv_saddr || !sk->sk_rcv_saddr ||
  80                                     sk2->sk_rcv_saddr == sk->sk_rcv_saddr)
  81                                         break;
  82                         }
  83                 }
  84         }
  85         return sk2 != NULL;
  86 }
  87 EXPORT_SYMBOL_GPL(inet_csk_bind_conflict);
  88
  89 /* Obtain a reference to a local port for the given sock,
  90  * if snum is zero it means select any available local port.
  91  */
  92 int inet_csk_get_port(struct sock *sk, unsigned short snum)
  93 {
  94         struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo;
  95         struct inet_bind_hashbucket *head;
  96         struct inet_bind_bucket *tb;
  97         int ret, attempts = 5;
  98         struct net *net = sock_net(sk);
  99         int smallest_size = -1, smallest_rover;
 100         kuid_t uid = sock_i_uid(sk);
 101
 102         local_bh_disable();
 103         if (!snum) {
 104                 int remaining, rover, low, high;
 105
 106 again:
 107                 inet_get_local_port_range(net, &low, &high);
 108                 remaining = (high - low) + 1;
 109                 smallest_rover = rover = prandom_u32() % remaining + low;
 110
 111                 smallest_size = -1;
 112                 do {
 113                         if (inet_is_local_reserved_port(net, rover))
 114                                 goto next_nolock;
 115                         head = &hashinfo->bhash[inet_bhashfn(net, rover,
 116                                         hashinfo->bhash_size)];
 117                         spin_lock(&head->lock);
 118                         inet_bind_bucket_for_each(tb, &head->chain)
 119                                 if (net_eq(ib_net(tb), net) && tb->port == rover) {
 120                                         if (((tb->fastreuse > 0 &&
 121                                               sk->sk_reuse &&
 122                                               sk->sk_state != TCP_LISTEN) ||
 123                                              (tb->fastreuseport > 0 &&
 124                                               sk->sk_reuseport &&
 125                                               uid_eq(tb->fastuid, uid))) &&
 126                                             (tb->num_owners < smallest_size || smallest_size == -1)) {
 127                                                 smallest_size = tb->num_owners;
 128                                                 smallest_rover = rover;
 129                                                 if (atomic_read(&hashinfo->bsockets) > (high - low) + 1 &&
 130                                                     !inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, false)) {
 131                                                         snum = smallest_rover;
 132                                                         goto tb_found;
 133                                                 }
 134                                         }
 135                                         if (!inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, false)) {
 136                                                 snum = rover;
 137                                                 goto tb_found;
 138                                         }
 139                                         goto next;
 140                                 }
 141                         break;
 142                 next:
 143                         spin_unlock(&head->lock);
 144                 next_nolock:
 145                         if (++rover > high)
 146                                 rover = low;
 147                 } while (--remaining > 0);
 148
 149                 /* Exhausted local port range during search?  It is not
 150                  * possible for us to be holding one of the bind hash
 151                  * locks if this test triggers, because if 'remaining'
 152                  * drops to zero, we broke out of the do/while loop at
 153                  * the top level, not from the 'break;' statement.
 154                  */
 155                 ret = 1;
 156                 if (remaining <= 0) {
 157                         if (smallest_size != -1) {
 158                                 snum = smallest_rover;
 159                                 goto have_snum;
 160                         }
 161                         goto fail;
 162                 }
 163                 /* OK, here is the one we will use.  HEAD is
 164                  * non-NULL and we hold it's mutex.
 165                  */
 166                 snum = rover;
 167         } else {
 168 have_snum:
 169                 head = &hashinfo->bhash[inet_bhashfn(net, snum,
 170                                 hashinfo->bhash_size)];
 171                 spin_lock(&head->lock);
 172                 inet_bind_bucket_for_each(tb, &head->chain)
 173                         if (net_eq(ib_net(tb), net) && tb->port == snum)
 174                                 goto tb_found;
 175         }
 176         tb = NULL;
 177         goto tb_not_found;
 178 tb_found:
 179         if (!hlist_empty(&tb->owners)) {
 180                 if (sk->sk_reuse == SK_FORCE_REUSE)
 181                         goto success;
 182
 183                 if (((tb->fastreuse > 0 &&
 184                       sk->sk_reuse && sk->sk_state != TCP_LISTEN) ||
 185                      (tb->fastreuseport > 0 &&
 186                       sk->sk_reuseport && uid_eq(tb->fastuid, uid))) &&
 187                     smallest_size == -1) {
 188                         goto success;
 189                 } else {
 190                         ret = 1;
 191                         if (inet_csk(sk)->icsk_af_ops->bind_conflict(sk, tb, true)) {
 192                                 if (((sk->sk_reuse && sk->sk_state != TCP_LISTEN) ||
 193                                      (tb->fastreuseport > 0 &&
 194                                       sk->sk_reuseport && uid_eq(tb->fastuid, uid))) &&
 195                                     smallest_size != -1 && --attempts >= 0) {
 196                                         spin_unlock(&head->lock);
 197                                         goto again;
 198                                 }
 199
 200                                 goto fail_unlock;
 201                         }
 202                 }
 203         }
 204 tb_not_found:
 205         ret = 1;
 206         if (!tb && (tb = inet_bind_bucket_create(hashinfo->bind_bucket_cachep,
 207                                         net, head, snum)) == NULL)
 208                 goto fail_unlock;
 209         if (hlist_empty(&tb->owners)) {
 210                 if (sk->sk_reuse && sk->sk_state != TCP_LISTEN)
 211                         tb->fastreuse = 1;
 212                 else
 213                         tb->fastreuse = 0;
 214                 if (sk->sk_reuseport) {
 215                         tb->fastreuseport = 1;
 216                         tb->fastuid = uid;
 217                 } else
 218                         tb->fastreuseport = 0;
 219         } else {
 220                 if (tb->fastreuse &&
 221                     (!sk->sk_reuse || sk->sk_state == TCP_LISTEN))
 222                         tb->fastreuse = 0;
 223                 if (tb->fastreuseport &&
 224                     (!sk->sk_reuseport || !uid_eq(tb->fastuid, uid)))
 225                         tb->fastreuseport = 0;
 226         }
 227 success:
 228         if (!inet_csk(sk)->icsk_bind_hash)
 229                 inet_bind_hash(sk, tb, snum);
 230         WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
 231         ret = 0;
 232
 233 fail_unlock:
 234         spin_unlock(&head->lock);
 235 fail:
 236         local_bh_enable();
 237         return ret;
 238 }
 239 EXPORT_SYMBOL_GPL(inet_csk_get_port);
 240
 241 /*
 242  * Wait for an incoming connection, avoid race conditions. This must be called
 243  * with the socket locked.
 244  */
 245 static int inet_csk_wait_for_connect(struct sock *sk, long timeo)
 246 {
 247         struct inet_connection_sock *icsk = inet_csk(sk);
 248         DEFINE_WAIT(wait);
 249         int err;
 250
 251         /*
 252          * True wake-one mechanism for incoming connections: only
 253          * one process gets woken up, not the 'whole herd'.
 254          * Since we do not 'race & poll' for established sockets
 255          * anymore, the common case will execute the loop only once.
 256          *
 257          * Subtle issue: "add_wait_queue_exclusive()" will be added
 258          * after any current non-exclusive waiters, and we know that
 259          * it will always _stay_ after any new non-exclusive waiters
 260          * because all non-exclusive waiters are added at the
 261          * beginning of the wait-queue. As such, it's ok to "drop"
 262          * our exclusiveness temporarily when we get woken up without
 263          * having to remove and re-insert us on the wait queue.
 264          */
 265         for (;;) {
 266                 prepare_to_wait_exclusive(sk_sleep(sk), &wait,
 267                                           TASK_INTERRUPTIBLE);
 268                 release_sock(sk);
 269                 if (reqsk_queue_empty(&icsk->icsk_accept_queue))
 270                         timeo = schedule_timeout(timeo);
 271                 sched_annotate_sleep();
 272                 lock_sock(sk);
 273                 err = 0;
 274                 if (!reqsk_queue_empty(&icsk->icsk_accept_queue))
 275                         break;
 276                 err = -EINVAL;
 277                 if (sk->sk_state != TCP_LISTEN)
 278                         break;
 279                 err = sock_intr_errno(timeo);
 280                 if (signal_pending(current))
 281                         break;
 282                 err = -EAGAIN;
 283                 if (!timeo)
 284                         break;
 285         }
 286         finish_wait(sk_sleep(sk), &wait);
 287         return err;
 288 }
 289
 290 /*
 291  * This will accept the next outstanding connection.
 292  */
 293 struct sock *inet_csk_accept(struct sock *sk, int flags, int *err)
 294 {
 295         struct inet_connection_sock *icsk = inet_csk(sk);
 296         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 297         struct sock *newsk;
 298         struct request_sock *req;
 299         int error;
 300
 301         lock_sock(sk);
 302
 303         /* We need to make sure that this socket is listening,
 304          * and that it has something pending.
 305          */
 306         error = -EINVAL;
 307         if (sk->sk_state != TCP_LISTEN)
 308                 goto out_err;
 309
 310         /* Find already established connection */
 311         if (reqsk_queue_empty(queue)) {
 312                 long timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
 313
 314                 /* If this is a non blocking socket don't sleep */
 315                 error = -EAGAIN;
 316                 if (!timeo)
 317                         goto out_err;
 318
 319                 error = inet_csk_wait_for_connect(sk, timeo);
 320                 if (error)
 321                         goto out_err;
 322         }
 323         req = reqsk_queue_remove(queue);
 324         newsk = req->sk;
 325
 326         sk_acceptq_removed(sk);
 327         if (sk->sk_protocol == IPPROTO_TCP && queue->fastopenq != NULL) {
 328                 spin_lock_bh(&queue->fastopenq->lock);
 329                 if (tcp_rsk(req)->listener) {
 330                         /* We are still waiting for the final ACK from 3WHS
 331                          * so can't free req now. Instead, we set req->sk to
 332                          * NULL to signify that the child socket is taken
 333                          * so reqsk_fastopen_remove() will free the req
 334                          * when 3WHS finishes (or is aborted).
 335                          */
 336                         req->sk = NULL;
 337                         req = NULL;
 338                 }
 339                 spin_unlock_bh(&queue->fastopenq->lock);
 340         }
 341 out:
 342         release_sock(sk);
 343         if (req)
 344                 __reqsk_free(req);
 345         return newsk;
 346 out_err:
 347         newsk = NULL;
 348         req = NULL;
 349         *err = error;
 350         goto out;
 351 }
 352 EXPORT_SYMBOL(inet_csk_accept);
 353
 354 /*
 355  * Using different timers for retransmit, delayed acks and probes
 356  * We may wish use just one timer maintaining a list of expire jiffies
 357  * to optimize.
 358  */
 359 void inet_csk_init_xmit_timers(struct sock *sk,
 360                                void (*retransmit_handler)(unsigned long),
 361                                void (*delack_handler)(unsigned long),
 362                                void (*keepalive_handler)(unsigned long))
 363 {
 364         struct inet_connection_sock *icsk = inet_csk(sk);
 365
 366         setup_timer(&icsk->icsk_retransmit_timer, retransmit_handler,
 367                         (unsigned long)sk);
 368         setup_timer(&icsk->icsk_delack_timer, delack_handler,
 369                         (unsigned long)sk);
 370         setup_timer(&sk->sk_timer, keepalive_handler, (unsigned long)sk);
 371         icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 372 }
 373 EXPORT_SYMBOL(inet_csk_init_xmit_timers);
 374
 375 void inet_csk_clear_xmit_timers(struct sock *sk)
 376 {
 377         struct inet_connection_sock *icsk = inet_csk(sk);
 378
 379         icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
 380
 381         sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 382         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 383         sk_stop_timer(sk, &sk->sk_timer);
 384 }
 385 EXPORT_SYMBOL(inet_csk_clear_xmit_timers);
 386
 387 void inet_csk_delete_keepalive_timer(struct sock *sk)
 388 {
 389         sk_stop_timer(sk, &sk->sk_timer);
 390 }
 391 EXPORT_SYMBOL(inet_csk_delete_keepalive_timer);
 392
 393 void inet_csk_reset_keepalive_timer(struct sock *sk, unsigned long len)
 394 {
 395         sk_reset_timer(sk, &sk->sk_timer, jiffies + len);
 396 }
 397 EXPORT_SYMBOL(inet_csk_reset_keepalive_timer);
 398
 399 struct dst_entry *inet_csk_route_req(struct sock *sk,
 400                                      struct flowi4 *fl4,
 401                                      const struct request_sock *req)
 402 {
 403         struct rtable *rt;
 404         const struct inet_request_sock *ireq = inet_rsk(req);
 405         struct ip_options_rcu *opt = inet_rsk(req)->opt;
 406         struct net *net = sock_net(sk);
 407         int flags = inet_sk_flowi_flags(sk);
 408
 409         flowi4_init_output(fl4, sk->sk_bound_dev_if, ireq->ir_mark,
 410                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 411                            sk->sk_protocol,
 412                            flags,
 413                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 414                            ireq->ir_loc_addr, ireq->ir_rmt_port, inet_sk(sk)->inet_sport);
 415         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 416         rt = ip_route_output_flow(net, fl4, sk);
 417         if (IS_ERR(rt))
 418                 goto no_route;
 419         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 420                 goto route_err;
 421         return &rt->dst;
 422
 423 route_err:
 424         ip_rt_put(rt);
 425 no_route:
 426         IP_INC_STATS_BH(net, IPSTATS_MIB_OUTNOROUTES);
 427         return NULL;
 428 }
 429 EXPORT_SYMBOL_GPL(inet_csk_route_req);
 430
 431 struct dst_entry *inet_csk_route_child_sock(struct sock *sk,
 432                                             struct sock *newsk,
 433                                             const struct request_sock *req)
 434 {
 435         const struct inet_request_sock *ireq = inet_rsk(req);
 436         struct inet_sock *newinet = inet_sk(newsk);
 437         struct ip_options_rcu *opt;
 438         struct net *net = sock_net(sk);
 439         struct flowi4 *fl4;
 440         struct rtable *rt;
 441
 442         fl4 = &newinet->cork.fl.u.ip4;
 443
 444         rcu_read_lock();
 445         opt = rcu_dereference(newinet->inet_opt);
 446         flowi4_init_output(fl4, sk->sk_bound_dev_if, inet_rsk(req)->ir_mark,
 447                            RT_CONN_FLAGS(sk), RT_SCOPE_UNIVERSE,
 448                            sk->sk_protocol, inet_sk_flowi_flags(sk),
 449                            (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr,
 450                            ireq->ir_loc_addr, ireq->ir_rmt_port, inet_sk(sk)->inet_sport);
 451         security_req_classify_flow(req, flowi4_to_flowi(fl4));
 452         rt = ip_route_output_flow(net, fl4, sk);
 453         if (IS_ERR(rt))
 454                 goto no_route;
 455         if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway)
 456                 goto route_err;
 457         rcu_read_unlock();
 458         return &rt->dst;
 459
 460 route_err:
 461         ip_rt_put(rt);
 462 no_route:
 463         rcu_read_unlock();
 464         IP_INC_STATS_BH(net, IPSTATS_MIB_OUTNOROUTES);
 465         return NULL;
 466 }
 467 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock);
 468
 469 static inline u32 inet_synq_hash(const __be32 raddr, const __be16 rport,
 470                                  const u32 rnd, const u32 synq_hsize)
 471 {
 472         return jhash_2words((__force u32)raddr, (__force u32)rport, rnd) & (synq_hsize - 1);
 473 }
 474
 475 #if IS_ENABLED(CONFIG_IPV6)
 476 #define AF_INET_FAMILY(fam) ((fam) == AF_INET)
 477 #else
 478 #define AF_INET_FAMILY(fam) 1
 479 #endif
 480
 481 struct request_sock *inet_csk_search_req(const struct sock *sk,
 482                                          struct request_sock ***prevp,
 483                                          const __be16 rport, const __be32 raddr,
 484                                          const __be32 laddr)
 485 {
 486         const struct inet_connection_sock *icsk = inet_csk(sk);
 487         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
 488         struct request_sock *req, **prev;
 489
 490         for (prev = &lopt->syn_table[inet_synq_hash(raddr, rport, lopt->hash_rnd,
 491                                                     lopt->nr_table_entries)];
 492              (req = *prev) != NULL;
 493              prev = &req->dl_next) {
 494                 const struct inet_request_sock *ireq = inet_rsk(req);
 495
 496                 if (ireq->ir_rmt_port == rport &&
 497                     ireq->ir_rmt_addr == raddr &&
 498                     ireq->ir_loc_addr == laddr &&
 499                     AF_INET_FAMILY(req->rsk_ops->family)) {
 500                         WARN_ON(req->sk);
 501                         *prevp = prev;
 502                         break;
 503                 }
 504         }
 505
 506         return req;
 507 }
 508 EXPORT_SYMBOL_GPL(inet_csk_search_req);
 509
 510 void inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req,
 511                                    unsigned long timeout)
 512 {
 513         struct inet_connection_sock *icsk = inet_csk(sk);
 514         struct listen_sock *lopt = icsk->icsk_accept_queue.listen_opt;
 515         const u32 h = inet_synq_hash(inet_rsk(req)->ir_rmt_addr,
 516                                      inet_rsk(req)->ir_rmt_port,
 517                                      lopt->hash_rnd, lopt->nr_table_entries);
 518
 519         reqsk_queue_hash_req(&icsk->icsk_accept_queue, h, req, timeout);
 520         inet_csk_reqsk_queue_added(sk, timeout);
 521 }
 522 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_hash_add);
 523
 524 /* Only thing we need from tcp.h */
 525 extern int sysctl_tcp_synack_retries;
 526
 527
 528 /* Decide when to expire the request and when to resend SYN-ACK */
 529 static inline void syn_ack_recalc(struct request_sock *req, const int thresh,
 530                                   const int max_retries,
 531                                   const u8 rskq_defer_accept,
 532                                   int *expire, int *resend)
 533 {
 534         if (!rskq_defer_accept) {
 535                 *expire = req->num_timeout >= thresh;
 536                 *resend = 1;
 537                 return;
 538         }
 539         *expire = req->num_timeout >= thresh &&
 540                   (!inet_rsk(req)->acked || req->num_timeout >= max_retries);
 541         /*
 542          * Do not resend while waiting for data after ACK,
 543          * start to resend on end of deferring period to give
 544          * last chance for data or ACK to create established socket.
 545          */
 546         *resend = !inet_rsk(req)->acked ||
 547                   req->num_timeout >= rskq_defer_accept - 1;
 548 }
 549
 550 int inet_rtx_syn_ack(struct sock *parent, struct request_sock *req)
 551 {
 552         int err = req->rsk_ops->rtx_syn_ack(parent, req);
 553
 554         if (!err)
 555                 req->num_retrans++;
 556         return err;
 557 }
 558 EXPORT_SYMBOL(inet_rtx_syn_ack);
 559
 560 void inet_csk_reqsk_queue_prune(struct sock *parent,
 561                                 const unsigned long interval,
 562                                 const unsigned long timeout,
 563                                 const unsigned long max_rto)
 564 {
 565         struct inet_connection_sock *icsk = inet_csk(parent);
 566         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 567         struct listen_sock *lopt = queue->listen_opt;
 568         int max_retries = icsk->icsk_syn_retries ? : sysctl_tcp_synack_retries;
 569         int thresh = max_retries;
 570         unsigned long now = jiffies;
 571         struct request_sock **reqp, *req;
 572         int i, budget;
 573
 574         if (lopt == NULL || lopt->qlen == 0)
 575                 return;
 576
 577         /* Normally all the openreqs are young and become mature
 578          * (i.e. converted to established socket) for first timeout.
 579          * If synack was not acknowledged for 1 second, it means
 580          * one of the following things: synack was lost, ack was lost,
 581          * rtt is high or nobody planned to ack (i.e. synflood).
 582          * When server is a bit loaded, queue is populated with old
 583          * open requests, reducing effective size of queue.
 584          * When server is well loaded, queue size reduces to zero
 585          * after several minutes of work. It is not synflood,
 586          * it is normal operation. The solution is pruning
 587          * too old entries overriding normal timeout, when
 588          * situation becomes dangerous.
 589          *
 590          * Essentially, we reserve half of room for young
 591          * embrions; and abort old ones without pity, if old
 592          * ones are about to clog our table.
 593          */
 594         if (lopt->qlen>>(lopt->max_qlen_log-1)) {
 595                 int young = (lopt->qlen_young<<1);
 596
 597                 while (thresh > 2) {
 598                         if (lopt->qlen < young)
 599                                 break;
 600                         thresh--;
 601                         young <<= 1;
 602                 }
 603         }
 604
 605         if (queue->rskq_defer_accept)
 606                 max_retries = queue->rskq_defer_accept;
 607
 608         budget = 2 * (lopt->nr_table_entries / (timeout / interval));
 609         i = lopt->clock_hand;
 610
 611         do {
 612                 reqp=&lopt->syn_table[i];
 613                 while ((req = *reqp) != NULL) {
 614                         if (time_after_eq(now, req->expires)) {
 615                                 int expire = 0, resend = 0;
 616
 617                                 syn_ack_recalc(req, thresh, max_retries,
 618                                                queue->rskq_defer_accept,
 619                                                &expire, &resend);
 620                                 req->rsk_ops->syn_ack_timeout(parent, req);
 621                                 if (!expire &&
 622                                     (!resend ||
 623                                      !inet_rtx_syn_ack(parent, req) ||
 624                                      inet_rsk(req)->acked)) {
 625                                         unsigned long timeo;
 626
 627                                         if (req->num_timeout++ == 0)
 628                                                 lopt->qlen_young--;
 629                                         timeo = min(timeout << req->num_timeout,
 630                                                     max_rto);
 631                                         req->expires = now + timeo;
 632                                         reqp = &req->dl_next;
 633                                         continue;
 634                                 }
 635
 636                                 /* Drop this request */
 637                                 inet_csk_reqsk_queue_unlink(parent, req, reqp);
 638                                 reqsk_queue_removed(queue, req);
 639                                 reqsk_free(req);
 640                                 continue;
 641                         }
 642                         reqp = &req->dl_next;
 643                 }
 644
 645                 i = (i + 1) & (lopt->nr_table_entries - 1);
 646
 647         } while (--budget > 0);
 648
 649         lopt->clock_hand = i;
 650
 651         if (lopt->qlen)
 652                 inet_csk_reset_keepalive_timer(parent, interval);
 653 }
 654 EXPORT_SYMBOL_GPL(inet_csk_reqsk_queue_prune);
 655
 656 /**
 657  *      inet_csk_clone_lock - clone an inet socket, and lock its clone
 658  *      @sk: the socket to clone
 659  *      @req: request_sock
 660  *      @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc)
 661  *
 662  *      Caller must unlock socket even in error path (bh_unlock_sock(newsk))
 663  */
 664 struct sock *inet_csk_clone_lock(const struct sock *sk,
 665                                  const struct request_sock *req,
 666                                  const gfp_t priority)
 667 {
 668         struct sock *newsk = sk_clone_lock(sk, priority);
 669
 670         if (newsk != NULL) {
 671                 struct inet_connection_sock *newicsk = inet_csk(newsk);
 672
 673                 newsk->sk_state = TCP_SYN_RECV;
 674                 newicsk->icsk_bind_hash = NULL;
 675
 676                 inet_sk(newsk)->inet_dport = inet_rsk(req)->ir_rmt_port;
 677                 inet_sk(newsk)->inet_num = inet_rsk(req)->ir_num;
 678                 inet_sk(newsk)->inet_sport = htons(inet_rsk(req)->ir_num);
 679                 newsk->sk_write_space = sk_stream_write_space;
 680
 681                 newsk->sk_mark = inet_rsk(req)->ir_mark;
 682
 683                 newicsk->icsk_retransmits = 0;
 684                 newicsk->icsk_backoff     = 0;
 685                 newicsk->icsk_probes_out  = 0;
 686
 687                 /* Deinitialize accept_queue to trap illegal accesses. */
 688                 memset(&newicsk->icsk_accept_queue, 0, sizeof(newicsk->icsk_accept_queue));
 689
 690                 security_inet_csk_clone(newsk, req);
 691         }
 692         return newsk;
 693 }
 694 EXPORT_SYMBOL_GPL(inet_csk_clone_lock);
 695
 696 /*
 697  * At this point, there should be no process reference to this
 698  * socket, and thus no user references at all.  Therefore we
 699  * can assume the socket waitqueue is inactive and nobody will
 700  * try to jump onto it.
 701  */
 702 void inet_csk_destroy_sock(struct sock *sk)
 703 {
 704         WARN_ON(sk->sk_state != TCP_CLOSE);
 705         WARN_ON(!sock_flag(sk, SOCK_DEAD));
 706
 707         /* It cannot be in hash table! */
 708         WARN_ON(!sk_unhashed(sk));
 709
 710         /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */
 711         WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
 712
 713         sk->sk_prot->destroy(sk);
 714
 715         sk_stream_kill_queues(sk);
 716
 717         xfrm_sk_free_policy(sk);
 718
 719         sk_refcnt_debug_release(sk);
 720
 721         percpu_counter_dec(sk->sk_prot->orphan_count);
 722         sock_put(sk);
 723 }
 724 EXPORT_SYMBOL(inet_csk_destroy_sock);
 725
 726 /* This function allows to force a closure of a socket after the call to
 727  * tcp/dccp_create_openreq_child().
 728  */
 729 void inet_csk_prepare_forced_close(struct sock *sk)
 730         __releases(&sk->sk_lock.slock)
 731 {
 732         /* sk_clone_lock locked the socket and set refcnt to 2 */
 733         bh_unlock_sock(sk);
 734         sock_put(sk);
 735
 736         /* The below has to be done to allow calling inet_csk_destroy_sock */
 737         sock_set_flag(sk, SOCK_DEAD);
 738         percpu_counter_inc(sk->sk_prot->orphan_count);
 739         inet_sk(sk)->inet_num = 0;
 740 }
 741 EXPORT_SYMBOL(inet_csk_prepare_forced_close);
 742
 743 int inet_csk_listen_start(struct sock *sk, const int nr_table_entries)
 744 {
 745         struct inet_sock *inet = inet_sk(sk);
 746         struct inet_connection_sock *icsk = inet_csk(sk);
 747         int rc = reqsk_queue_alloc(&icsk->icsk_accept_queue, nr_table_entries);
 748
 749         if (rc != 0)
 750                 return rc;
 751
 752         sk->sk_max_ack_backlog = 0;
 753         sk->sk_ack_backlog = 0;
 754         inet_csk_delack_init(sk);
 755
 756         /* There is race window here: we announce ourselves listening,
 757          * but this transition is still not validated by get_port().
 758          * It is OK, because this socket enters to hash table only
 759          * after validation is complete.
 760          */
 761         sk->sk_state = TCP_LISTEN;
 762         if (!sk->sk_prot->get_port(sk, inet->inet_num)) {
 763                 inet->inet_sport = htons(inet->inet_num);
 764
 765                 sk_dst_reset(sk);
 766                 sk->sk_prot->hash(sk);
 767
 768                 return 0;
 769         }
 770
 771         sk->sk_state = TCP_CLOSE;
 772         __reqsk_queue_destroy(&icsk->icsk_accept_queue);
 773         return -EADDRINUSE;
 774 }
 775 EXPORT_SYMBOL_GPL(inet_csk_listen_start);
 776
 777 /*
 778  *      This routine closes sockets which have been at least partially
 779  *      opened, but not yet accepted.
 780  */
 781 void inet_csk_listen_stop(struct sock *sk)
 782 {
 783         struct inet_connection_sock *icsk = inet_csk(sk);
 784         struct request_sock_queue *queue = &icsk->icsk_accept_queue;
 785         struct request_sock *acc_req;
 786         struct request_sock *req;
 787
 788         inet_csk_delete_keepalive_timer(sk);
 789
 790         /* make all the listen_opt local to us */
 791         acc_req = reqsk_queue_yank_acceptq(queue);
 792
 793         /* Following specs, it would be better either to send FIN
 794          * (and enter FIN-WAIT-1, it is normal close)
 795          * or to send active reset (abort).
 796          * Certainly, it is pretty dangerous while synflood, but it is
 797          * bad justification for our negligence 8)
 798          * To be honest, we are not able to make either
 799          * of the variants now.                 --ANK
 800          */
 801         reqsk_queue_destroy(queue);
 802
 803         while ((req = acc_req) != NULL) {
 804                 struct sock *child = req->sk;
 805
 806                 acc_req = req->dl_next;
 807
 808                 local_bh_disable();
 809                 bh_lock_sock(child);
 810                 WARN_ON(sock_owned_by_user(child));
 811                 sock_hold(child);
 812
 813                 sk->sk_prot->disconnect(child, O_NONBLOCK);
 814
 815                 sock_orphan(child);
 816
 817                 percpu_counter_inc(sk->sk_prot->orphan_count);
 818
 819                 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->listener) {
 820                         BUG_ON(tcp_sk(child)->fastopen_rsk != req);
 821                         BUG_ON(sk != tcp_rsk(req)->listener);
 822
 823                         /* Paranoid, to prevent race condition if
 824                          * an inbound pkt destined for child is
 825                          * blocked by sock lock in tcp_v4_rcv().
 826                          * Also to satisfy an assertion in
 827                          * tcp_v4_destroy_sock().
 828                          */
 829                         tcp_sk(child)->fastopen_rsk = NULL;
 830                         sock_put(sk);
 831                 }
 832                 inet_csk_destroy_sock(child);
 833
 834                 bh_unlock_sock(child);
 835                 local_bh_enable();
 836                 sock_put(child);
 837
 838                 sk_acceptq_removed(sk);
 839                 __reqsk_free(req);
 840         }
 841         if (queue->fastopenq != NULL) {
 842                 /* Free all the reqs queued in rskq_rst_head. */
 843                 spin_lock_bh(&queue->fastopenq->lock);
 844                 acc_req = queue->fastopenq->rskq_rst_head;
 845                 queue->fastopenq->rskq_rst_head = NULL;
 846                 spin_unlock_bh(&queue->fastopenq->lock);
 847                 while ((req = acc_req) != NULL) {
 848                         acc_req = req->dl_next;
 849                         __reqsk_free(req);
 850                 }
 851         }
 852         WARN_ON(sk->sk_ack_backlog);
 853 }
 854 EXPORT_SYMBOL_GPL(inet_csk_listen_stop);
 855
 856 void inet_csk_addr2sockaddr(struct sock *sk, struct sockaddr *uaddr)
 857 {
 858         struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
 859         const struct inet_sock *inet = inet_sk(sk);
 860
 861         sin->sin_family         = AF_INET;
 862         sin->sin_addr.s_addr    = inet->inet_daddr;
 863         sin->sin_port           = inet->inet_dport;
 864 }
 865 EXPORT_SYMBOL_GPL(inet_csk_addr2sockaddr);
 866
 867 #ifdef CONFIG_COMPAT
 868 int inet_csk_compat_getsockopt(struct sock *sk, int level, int optname,
 869                                char __user *optval, int __user *optlen)
 870 {
 871         const struct inet_connection_sock *icsk = inet_csk(sk);
 872
 873         if (icsk->icsk_af_ops->compat_getsockopt != NULL)
 874                 return icsk->icsk_af_ops->compat_getsockopt(sk, level, optname,
 875                                                             optval, optlen);
 876         return icsk->icsk_af_ops->getsockopt(sk, level, optname,
 877                                              optval, optlen);
 878 }
 879 EXPORT_SYMBOL_GPL(inet_csk_compat_getsockopt);
 880
 881 int inet_csk_compat_setsockopt(struct sock *sk, int level, int optname,
 882                                char __user *optval, unsigned int optlen)
 883 {
 884         const struct inet_connection_sock *icsk = inet_csk(sk);
 885
 886         if (icsk->icsk_af_ops->compat_setsockopt != NULL)
 887                 return icsk->icsk_af_ops->compat_setsockopt(sk, level, optname,
 888                                                             optval, optlen);
 889         return icsk->icsk_af_ops->setsockopt(sk, level, optname,
 890                                              optval, optlen);
 891 }
 892 EXPORT_SYMBOL_GPL(inet_csk_compat_setsockopt);
 893 #endif
 894
 895 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl)
 896 {
 897         const struct inet_sock *inet = inet_sk(sk);
 898         const struct ip_options_rcu *inet_opt;
 899         __be32 daddr = inet->inet_daddr;
 900         struct flowi4 *fl4;
 901         struct rtable *rt;
 902
 903         rcu_read_lock();
 904         inet_opt = rcu_dereference(inet->inet_opt);
 905         if (inet_opt && inet_opt->opt.srr)
 906                 daddr = inet_opt->opt.faddr;
 907         fl4 = &fl->u.ip4;
 908         rt = ip_route_output_ports(sock_net(sk), fl4, sk, daddr,
 909                                    inet->inet_saddr, inet->inet_dport,
 910                                    inet->inet_sport, sk->sk_protocol,
 911                                    RT_CONN_FLAGS(sk), sk->sk_bound_dev_if);
 912         if (IS_ERR(rt))
 913                 rt = NULL;
 914         if (rt)
 915                 sk_setup_caps(sk, &rt->dst);
 916         rcu_read_unlock();
 917
 918         return &rt->dst;
 919 }
 920
 921 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu)
 922 {
 923         struct dst_entry *dst = __sk_dst_check(sk, 0);
 924         struct inet_sock *inet = inet_sk(sk);
 925
 926         if (!dst) {
 927                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 928                 if (!dst)
 929                         goto out;
 930         }
 931         dst->ops->update_pmtu(dst, sk, NULL, mtu);
 932
 933         dst = __sk_dst_check(sk, 0);
 934         if (!dst)
 935                 dst = inet_csk_rebuild_route(sk, &inet->cork.fl);
 936 out:
 937         return dst;
 938 }
 939 EXPORT_SYMBOL_GPL(inet_csk_update_pmtu);