net/ipv4/tcp_ipv4.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  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Version:     $Id: tcp_ipv4.c,v 1.240 2002/02/01 22:01:04 davem Exp $
   9  *
  10  *              IPv4 specific functions
  11  *
  12  *
  13  *              code split from:
  14  *              linux/ipv4/tcp.c
  15  *              linux/ipv4/tcp_input.c
  16  *              linux/ipv4/tcp_output.c
  17  *
  18  *              See tcp.c for author information
  19  *
  20  *      This program is free software; you can redistribute it and/or
  21  *      modify it under the terms of the GNU General Public License
  22  *      as published by the Free Software Foundation; either version
  23  *      2 of the License, or (at your option) any later version.
  24  */
  25
  26 /*
  27  * Changes:
  28  *              David S. Miller :       New socket lookup architecture.
  29  *                                      This code is dedicated to John Dyson.
  30  *              David S. Miller :       Change semantics of established hash,
  31  *                                      half is devoted to TIME_WAIT sockets
  32  *                                      and the rest go in the other half.
  33  *              Andi Kleen :            Add support for syncookies and fixed
  34  *                                      some bugs: ip options weren't passed to
  35  *                                      the TCP layer, missed a check for an
  36  *                                      ACK bit.
  37  *              Andi Kleen :            Implemented fast path mtu discovery.
  38  *                                      Fixed many serious bugs in the
  39  *                                      open_request handling and moved
  40  *                                      most of it into the af independent code.
  41  *                                      Added tail drop and some other bugfixes.
  42  *                                      Added new listen sematics.
  43  *              Mike McLagan    :       Routing by source
  44  *      Juan Jose Ciarlante:            ip_dynaddr bits
  45  *              Andi Kleen:             various fixes.
  46  *      Vitaly E. Lavrov        :       Transparent proxy revived after year
  47  *                                      coma.
  48  *      Andi Kleen              :       Fix new listen.
  49  *      Andi Kleen              :       Fix accept error reporting.
  50  *      YOSHIFUJI Hideaki @USAGI and:   Support IPV6_V6ONLY socket option, which
  51  *      Alexey Kuznetsov                allow both IPv4 and IPv6 sockets to bind
  52  *                                      a single port at the same time.
  53  */
  54
  55 #include <linux/config.h>
  56
  57 #include <linux/types.h>
  58 #include <linux/fcntl.h>
  59 #include <linux/module.h>
  60 #include <linux/random.h>
  61 #include <linux/cache.h>
  62 #include <linux/jhash.h>
  63 #include <linux/init.h>
  64 #include <linux/times.h>
  65
  66 #include <net/icmp.h>
  67 #include <net/tcp.h>
  68 #include <net/ipv6.h>
  69 #include <net/inet_common.h>
  70 #include <net/xfrm.h>
  71
  72 #include <linux/inet.h>
  73 #include <linux/ipv6.h>
  74 #include <linux/stddef.h>
  75 #include <linux/proc_fs.h>
  76 #include <linux/seq_file.h>
  77 #include <linux/vserver/debug.h>
  78
  79 extern int sysctl_ip_dynaddr;
  80 int sysctl_tcp_tw_reuse;
  81 int sysctl_tcp_low_latency;
  82
  83 /* Check TCP sequence numbers in ICMP packets. */
  84 #define ICMP_MIN_LENGTH 8
  85
  86 /* Socket used for sending RSTs */
  87 static struct socket *tcp_socket;
  88
  89 void tcp_v4_send_check(struct sock *sk, struct tcphdr *th, int len,
  90                        struct sk_buff *skb);
  91
  92 struct tcp_hashinfo __cacheline_aligned tcp_hashinfo = {
  93         .__tcp_lhash_lock       =       RW_LOCK_UNLOCKED,
  94         .__tcp_lhash_users      =       ATOMIC_INIT(0),
  95         .__tcp_lhash_wait
  96           = __WAIT_QUEUE_HEAD_INITIALIZER(tcp_hashinfo.__tcp_lhash_wait),
  97         .__tcp_portalloc_lock   =       SPIN_LOCK_UNLOCKED
  98 };
  99
 100 /*
 101  * This array holds the first and last local port number.
 102  * For high-usage systems, use sysctl to change this to
 103  * 32768-61000
 104  */
 105 int sysctl_local_port_range[2] = { 1024, 4999 };
 106 int tcp_port_rover = 1024 - 1;
 107
 108 static __inline__ int tcp_hashfn(__u32 laddr, __u16 lport,
 109                                  __u32 faddr, __u16 fport)
 110 {
 111         int h = (laddr ^ lport) ^ (faddr ^ fport);
 112         h ^= h >> 16;
 113         h ^= h >> 8;
 114         return h & (tcp_ehash_size - 1);
 115 }
 116
 117 static __inline__ int tcp_sk_hashfn(struct sock *sk)
 118 {
 119         struct inet_opt *inet = inet_sk(sk);
 120         __u32 laddr = inet->rcv_saddr;
 121         __u16 lport = inet->num;
 122         __u32 faddr = inet->daddr;
 123         __u16 fport = inet->dport;
 124
 125         return tcp_hashfn(laddr, lport, faddr, fport);
 126 }
 127
 128 /* Allocate and initialize a new TCP local port bind bucket.
 129  * The bindhash mutex for snum's hash chain must be held here.
 130  */
 131 struct tcp_bind_bucket *tcp_bucket_create(struct tcp_bind_hashbucket *head,
 132                                           unsigned short snum)
 133 {
 134         struct tcp_bind_bucket *tb = kmem_cache_alloc(tcp_bucket_cachep,
 135                                                       SLAB_ATOMIC);
 136         if (tb) {
 137                 tb->port = snum;
 138                 tb->fastreuse = 0;
 139                 INIT_HLIST_HEAD(&tb->owners);
 140                 hlist_add_head(&tb->node, &head->chain);
 141         }
 142         return tb;
 143 }
 144
 145 /* Caller must hold hashbucket lock for this tb with local BH disabled */
 146 void tcp_bucket_destroy(struct tcp_bind_bucket *tb)
 147 {
 148         if (hlist_empty(&tb->owners)) {
 149                 __hlist_del(&tb->node);
 150                 kmem_cache_free(tcp_bucket_cachep, tb);
 151         }
 152 }
 153
 154 /* Caller must disable local BH processing. */
 155 static __inline__ void __tcp_inherit_port(struct sock *sk, struct sock *child)
 156 {
 157         struct tcp_bind_hashbucket *head =
 158                                 &tcp_bhash[tcp_bhashfn(inet_sk(child)->num)];
 159         struct tcp_bind_bucket *tb;
 160
 161         spin_lock(&head->lock);
 162         tb = tcp_sk(sk)->bind_hash;
 163         sk_add_bind_node(child, &tb->owners);
 164         tcp_sk(child)->bind_hash = tb;
 165         spin_unlock(&head->lock);
 166 }
 167
 168 inline void tcp_inherit_port(struct sock *sk, struct sock *child)
 169 {
 170         local_bh_disable();
 171         __tcp_inherit_port(sk, child);
 172         local_bh_enable();
 173 }
 174
 175 void tcp_bind_hash(struct sock *sk, struct tcp_bind_bucket *tb,
 176                    unsigned short snum)
 177 {
 178         inet_sk(sk)->num = snum;
 179         sk_add_bind_node(sk, &tb->owners);
 180         tcp_sk(sk)->bind_hash = tb;
 181 }
 182
 183 static inline int tcp_bind_conflict(struct sock *sk, struct tcp_bind_bucket *tb)
 184 {
 185         const u32 sk_rcv_saddr = tcp_v4_rcv_saddr(sk);
 186         struct sock *sk2;
 187         struct hlist_node *node;
 188         int reuse = sk->sk_reuse;
 189
 190         sk_for_each_bound(sk2, node, &tb->owners) {
 191                 if (sk != sk2 &&
 192                     !tcp_v6_ipv6only(sk2) &&
 193                     (!sk->sk_bound_dev_if ||
 194                      !sk2->sk_bound_dev_if ||
 195                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
 196                         if (!reuse || !sk2->sk_reuse ||
 197                             sk2->sk_state == TCP_LISTEN) {
 198                                 const u32 sk2_rcv_saddr = tcp_v4_rcv_saddr(sk2);
 199                                 if (!sk2_rcv_saddr || !sk_rcv_saddr ||
 200                                     sk2_rcv_saddr == sk_rcv_saddr)
 201                                         break;
 202                         }
 203                 }
 204         }
 205         return node != NULL;
 206 }
 207
 208 /* Obtain a reference to a local port for the given sock,
 209  * if snum is zero it means select any available local port.
 210  */
 211 static int tcp_v4_get_port(struct sock *sk, unsigned short snum)
 212 {
 213         struct tcp_bind_hashbucket *head;
 214         struct hlist_node *node;
 215         struct tcp_bind_bucket *tb;
 216         int ret;
 217
 218         local_bh_disable();
 219         if (!snum) {
 220                 int low = sysctl_local_port_range[0];
 221                 int high = sysctl_local_port_range[1];
 222                 int remaining = (high - low) + 1;
 223                 int rover;
 224
 225                 spin_lock(&tcp_portalloc_lock);
 226                 rover = tcp_port_rover;
 227                 do {
 228                         rover++;
 229                         if (rover < low || rover > high)
 230                                 rover = low;
 231                         head = &tcp_bhash[tcp_bhashfn(rover)];
 232                         spin_lock(&head->lock);
 233                         tb_for_each(tb, node, &head->chain)
 234                                 if (tb->port == rover)
 235                                         goto next;
 236                         break;
 237                 next:
 238                         spin_unlock(&head->lock);
 239                 } while (--remaining > 0);
 240                 tcp_port_rover = rover;
 241                 spin_unlock(&tcp_portalloc_lock);
 242
 243                 /* Exhausted local port range during search? */
 244                 ret = 1;
 245                 if (remaining <= 0)
 246                         goto fail;
 247
 248                 /* OK, here is the one we will use.  HEAD is
 249                  * non-NULL and we hold it's mutex.
 250                  */
 251                 snum = rover;
 252         } else {
 253                 head = &tcp_bhash[tcp_bhashfn(snum)];
 254                 spin_lock(&head->lock);
 255                 tb_for_each(tb, node, &head->chain)
 256                         if (tb->port == snum)
 257                                 goto tb_found;
 258         }
 259         tb = NULL;
 260         goto tb_not_found;
 261 tb_found:
 262         if (!hlist_empty(&tb->owners)) {
 263                 if (sk->sk_reuse > 1)
 264                         goto success;
 265                 if (tb->fastreuse > 0 &&
 266                     sk->sk_reuse && sk->sk_state != TCP_LISTEN) {
 267                         goto success;
 268                 } else {
 269                         ret = 1;
 270                         if (tcp_bind_conflict(sk, tb))
 271                                 goto fail_unlock;
 272                 }
 273         }
 274 tb_not_found:
 275         ret = 1;
 276         if (!tb && (tb = tcp_bucket_create(head, snum)) == NULL)
 277                 goto fail_unlock;
 278         if (hlist_empty(&tb->owners)) {
 279                 if (sk->sk_reuse && sk->sk_state != TCP_LISTEN)
 280                         tb->fastreuse = 1;
 281                 else
 282                         tb->fastreuse = 0;
 283         } else if (tb->fastreuse &&
 284                    (!sk->sk_reuse || sk->sk_state == TCP_LISTEN))
 285                 tb->fastreuse = 0;
 286 success:
 287         if (!tcp_sk(sk)->bind_hash)
 288                 tcp_bind_hash(sk, tb, snum);
 289         BUG_TRAP(tcp_sk(sk)->bind_hash == tb);
 290         ret = 0;
 291
 292 fail_unlock:
 293         spin_unlock(&head->lock);
 294 fail:
 295         local_bh_enable();
 296         return ret;
 297 }
 298
 299 /* Get rid of any references to a local port held by the
 300  * given sock.
 301  */
 302 static void __tcp_put_port(struct sock *sk)
 303 {
 304         struct inet_opt *inet = inet_sk(sk);
 305         struct tcp_bind_hashbucket *head = &tcp_bhash[tcp_bhashfn(inet->num)];
 306         struct tcp_bind_bucket *tb;
 307
 308         spin_lock(&head->lock);
 309         tb = tcp_sk(sk)->bind_hash;
 310         __sk_del_bind_node(sk);
 311         tcp_sk(sk)->bind_hash = NULL;
 312         inet->num = 0;
 313         tcp_bucket_destroy(tb);
 314         spin_unlock(&head->lock);
 315 }
 316
 317 void tcp_put_port(struct sock *sk)
 318 {
 319         local_bh_disable();
 320         __tcp_put_port(sk);
 321         local_bh_enable();
 322 }
 323
 324 /* This lock without WQ_FLAG_EXCLUSIVE is good on UP and it can be very bad on SMP.
 325  * Look, when several writers sleep and reader wakes them up, all but one
 326  * immediately hit write lock and grab all the cpus. Exclusive sleep solves
 327  * this, _but_ remember, it adds useless work on UP machines (wake up each
 328  * exclusive lock release). It should be ifdefed really.
 329  */
 330
 331 void tcp_listen_wlock(void)
 332 {
 333         write_lock(&tcp_lhash_lock);
 334
 335         if (atomic_read(&tcp_lhash_users)) {
 336                 DEFINE_WAIT(wait);
 337
 338                 for (;;) {
 339                         prepare_to_wait_exclusive(&tcp_lhash_wait,
 340                                                 &wait, TASK_UNINTERRUPTIBLE);
 341                         if (!atomic_read(&tcp_lhash_users))
 342                                 break;
 343                         write_unlock_bh(&tcp_lhash_lock);
 344                         schedule();
 345                         write_lock_bh(&tcp_lhash_lock);
 346                 }
 347
 348                 finish_wait(&tcp_lhash_wait, &wait);
 349         }
 350 }
 351
 352 static __inline__ void __tcp_v4_hash(struct sock *sk, const int listen_possible)
 353 {
 354         struct hlist_head *list;
 355         rwlock_t *lock;
 356
 357         BUG_TRAP(sk_unhashed(sk));
 358         if (listen_possible && sk->sk_state == TCP_LISTEN) {
 359                 list = &tcp_listening_hash[tcp_sk_listen_hashfn(sk)];
 360                 lock = &tcp_lhash_lock;
 361                 tcp_listen_wlock();
 362         } else {
 363                 list = &tcp_ehash[(sk->sk_hashent = tcp_sk_hashfn(sk))].chain;
 364                 lock = &tcp_ehash[sk->sk_hashent].lock;
 365                 write_lock(lock);
 366         }
 367         __sk_add_node(sk, list);
 368         sock_prot_inc_use(sk->sk_prot);
 369         write_unlock(lock);
 370         if (listen_possible && sk->sk_state == TCP_LISTEN)
 371                 wake_up(&tcp_lhash_wait);
 372 }
 373
 374 static void tcp_v4_hash(struct sock *sk)
 375 {
 376         if (sk->sk_state != TCP_CLOSE) {
 377                 local_bh_disable();
 378                 __tcp_v4_hash(sk, 1);
 379                 local_bh_enable();
 380         }
 381 }
 382
 383 void tcp_unhash(struct sock *sk)
 384 {
 385         rwlock_t *lock;
 386
 387         if (sk_unhashed(sk))
 388                 goto ende;
 389
 390         if (sk->sk_state == TCP_LISTEN) {
 391                 local_bh_disable();
 392                 tcp_listen_wlock();
 393                 lock = &tcp_lhash_lock;
 394         } else {
 395                 struct tcp_ehash_bucket *head = &tcp_ehash[sk->sk_hashent];
 396                 lock = &head->lock;
 397                 write_lock_bh(&head->lock);
 398         }
 399
 400         if (__sk_del_node_init(sk))
 401                 sock_prot_dec_use(sk->sk_prot);
 402         write_unlock_bh(lock);
 403
 404  ende:
 405         if (sk->sk_state == TCP_LISTEN)
 406                 wake_up(&tcp_lhash_wait);
 407 }
 408
 409 /* Don't inline this cruft.  Here are some nice properties to
 410  * exploit here.  The BSD API does not allow a listening TCP
 411  * to specify the remote port nor the remote address for the
 412  * connection.  So always assume those are both wildcarded
 413  * during the search since they can never be otherwise.
 414  */
 415 static struct sock *__tcp_v4_lookup_listener(struct hlist_head *head, u32 daddr,
 416                                              unsigned short hnum, int dif)
 417 {
 418         struct sock *result = NULL, *sk;
 419         struct hlist_node *node;
 420         int score, hiscore;
 421
 422         hiscore=-1;
 423         sk_for_each(sk, node, head) {
 424                 struct inet_opt *inet = inet_sk(sk);
 425
 426                 if (inet->num == hnum && !ipv6_only_sock(sk)) {
 427                         __u32 rcv_saddr = inet->rcv_saddr;
 428
 429                         score = (sk->sk_family == PF_INET ? 1 : 0);
 430                         if (rcv_saddr) {
 431                                 if (rcv_saddr != daddr)
 432                                         continue;
 433                                 score+=2;
 434                         }
 435                         if (sk->sk_bound_dev_if) {
 436                                 if (sk->sk_bound_dev_if != dif)
 437                                         continue;
 438                                 score+=2;
 439                         }
 440                         if (score == 5)
 441                                 return sk;
 442                         if (score > hiscore) {
 443                                 hiscore = score;
 444                                 result = sk;
 445                         }
 446                 }
 447         }
 448         return result;
 449 }
 450
 451 /* Optimize the common listener case. */
 452 inline struct sock *tcp_v4_lookup_listener(u32 daddr, unsigned short hnum,
 453                                            int dif)
 454 {
 455         struct sock *sk = NULL;
 456         struct hlist_head *head;
 457
 458         read_lock(&tcp_lhash_lock);
 459         head = &tcp_listening_hash[tcp_lhashfn(hnum)];
 460         if (!hlist_empty(head)) {
 461                 struct inet_opt *inet = inet_sk((sk = __sk_head(head)));
 462                 if (inet->num == hnum && !sk->sk_node.next &&
 463                     (!inet->rcv_saddr || inet->rcv_saddr == daddr) &&
 464                     (sk->sk_family == PF_INET || !ipv6_only_sock(sk)) &&
 465                     !sk->sk_bound_dev_if)
 466                         goto sherry_cache;
 467                 sk = __tcp_v4_lookup_listener(head, daddr, hnum, dif);
 468         }
 469         if (sk) {
 470 sherry_cache:
 471                 sock_hold(sk);
 472         }
 473         read_unlock(&tcp_lhash_lock);
 474         return sk;
 475 }
 476
 477 /* Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
 478  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
 479  *
 480  * Local BH must be disabled here.
 481  */
 482
 483 static inline struct sock *__tcp_v4_lookup_established(u32 saddr, u16 sport,
 484                                                        u32 daddr, u16 hnum,
 485                                                        int dif)
 486 {
 487         struct tcp_ehash_bucket *head;
 488         TCP_V4_ADDR_COOKIE(acookie, saddr, daddr)
 489         __u32 ports = TCP_COMBINED_PORTS(sport, hnum);
 490         struct sock *sk;
 491         struct hlist_node *node;
 492         /* Optimize here for direct hit, only listening connections can
 493          * have wildcards anyways.
 494          */
 495         int hash = tcp_hashfn(daddr, hnum, saddr, sport);
 496         head = &tcp_ehash[hash];
 497         read_lock(&head->lock);
 498         sk_for_each(sk, node, &head->chain) {
 499                 if (TCP_IPV4_MATCH(sk, acookie, saddr, daddr, ports, dif))
 500                         goto hit; /* You sunk my battleship! */
 501         }
 502
 503         /* Must check for a TIME_WAIT'er before going to listener hash. */
 504         sk_for_each(sk, node, &(head + tcp_ehash_size)->chain) {
 505                 if (TCP_IPV4_TW_MATCH(sk, acookie, saddr, daddr, ports, dif))
 506                         goto hit;
 507         }
 508         sk = NULL;
 509 out:
 510         read_unlock(&head->lock);
 511         return sk;
 512 hit:
 513         sock_hold(sk);
 514         goto out;
 515 }
 516
 517 static inline struct sock *__tcp_v4_lookup(u32 saddr, u16 sport,
 518                                            u32 daddr, u16 hnum, int dif)
 519 {
 520         struct sock *sk = __tcp_v4_lookup_established(saddr, sport,
 521                                                       daddr, hnum, dif);
 522
 523         return sk ? : tcp_v4_lookup_listener(daddr, hnum, dif);
 524 }
 525
 526 inline struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr,
 527                                   u16 dport, int dif)
 528 {
 529         struct sock *sk;
 530
 531         local_bh_disable();
 532         sk = __tcp_v4_lookup(saddr, sport, daddr, ntohs(dport), dif);
 533         local_bh_enable();
 534
 535         return sk;
 536 }
 537
 538 static inline __u32 tcp_v4_init_sequence(struct sock *sk, struct sk_buff *skb)
 539 {
 540         return secure_tcp_sequence_number(skb->nh.iph->daddr,
 541                                           skb->nh.iph->saddr,
 542                                           skb->h.th->dest,
 543                                           skb->h.th->source);
 544 }
 545
 546 /* called with local bh disabled */
 547 static int __tcp_v4_check_established(struct sock *sk, __u16 lport,
 548                                       struct tcp_tw_bucket **twp)
 549 {
 550         struct inet_opt *inet = inet_sk(sk);
 551         u32 daddr = inet->rcv_saddr;
 552         u32 saddr = inet->daddr;
 553         int dif = sk->sk_bound_dev_if;
 554         TCP_V4_ADDR_COOKIE(acookie, saddr, daddr)
 555         __u32 ports = TCP_COMBINED_PORTS(inet->dport, lport);
 556         int hash = tcp_hashfn(daddr, lport, saddr, inet->dport);
 557         struct tcp_ehash_bucket *head = &tcp_ehash[hash];
 558         struct sock *sk2;
 559         struct hlist_node *node;
 560         struct tcp_tw_bucket *tw;
 561
 562         write_lock(&head->lock);
 563
 564         /* Check TIME-WAIT sockets first. */
 565         sk_for_each(sk2, node, &(head + tcp_ehash_size)->chain) {
 566                 tw = (struct tcp_tw_bucket *)sk2;
 567
 568                 if (TCP_IPV4_TW_MATCH(sk2, acookie, saddr, daddr, ports, dif)) {
 569                         struct tcp_opt *tp = tcp_sk(sk);
 570
 571                         /* With PAWS, it is safe from the viewpoint
 572                            of data integrity. Even without PAWS it
 573                            is safe provided sequence spaces do not
 574                            overlap i.e. at data rates <= 80Mbit/sec.
 575
 576                            Actually, the idea is close to VJ's one,
 577                            only timestamp cache is held not per host,
 578                            but per port pair and TW bucket is used
 579                            as state holder.
 580
 581                            If TW bucket has been already destroyed we
 582                            fall back to VJ's scheme and use initial
 583                            timestamp retrieved from peer table.
 584                          */
 585                         if (tw->tw_ts_recent_stamp &&
 586                             (!twp || (sysctl_tcp_tw_reuse &&
 587                                       xtime.tv_sec -
 588                                       tw->tw_ts_recent_stamp > 1))) {
 589                                 if ((tp->write_seq =
 590                                                 tw->tw_snd_nxt + 65535 + 2) == 0)
 591                                         tp->write_seq = 1;
 592                                 tp->ts_recent       = tw->tw_ts_recent;
 593                                 tp->ts_recent_stamp = tw->tw_ts_recent_stamp;
 594                                 sock_hold(sk2);
 595                                 goto unique;
 596                         } else
 597                                 goto not_unique;
 598                 }
 599         }
 600         tw = NULL;
 601
 602         /* And established part... */
 603         sk_for_each(sk2, node, &head->chain) {
 604                 if (TCP_IPV4_MATCH(sk2, acookie, saddr, daddr, ports, dif))
 605                         goto not_unique;
 606         }
 607
 608 unique:
 609         /* Must record num and sport now. Otherwise we will see
 610          * in hash table socket with a funny identity. */
 611         inet->num = lport;
 612         inet->sport = htons(lport);
 613         sk->sk_hashent = hash;
 614         BUG_TRAP(sk_unhashed(sk));
 615         __sk_add_node(sk, &head->chain);
 616         sock_prot_inc_use(sk->sk_prot);
 617         write_unlock(&head->lock);
 618
 619         if (twp) {
 620                 *twp = tw;
 621                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 622         } else if (tw) {
 623                 /* Silly. Should hash-dance instead... */
 624                 tcp_tw_deschedule(tw);
 625                 NET_INC_STATS_BH(LINUX_MIB_TIMEWAITRECYCLED);
 626
 627                 tcp_tw_put(tw);
 628         }
 629
 630         return 0;
 631
 632 not_unique:
 633         write_unlock(&head->lock);
 634         return -EADDRNOTAVAIL;
 635 }
 636
 637 /*
 638  * Bind a port for a connect operation and hash it.
 639  */
 640 static int tcp_v4_hash_connect(struct sock *sk)
 641 {
 642         unsigned short snum = inet_sk(sk)->num;
 643         struct tcp_bind_hashbucket *head;
 644         struct tcp_bind_bucket *tb;
 645         int ret;
 646
 647         if (!snum) {
 648                 int rover;
 649                 int low = sysctl_local_port_range[0];
 650                 int high = sysctl_local_port_range[1];
 651                 int remaining = (high - low) + 1;
 652                 struct hlist_node *node;
 653                 struct tcp_tw_bucket *tw = NULL;
 654
 655                 local_bh_disable();
 656
 657                 /* TODO. Actually it is not so bad idea to remove
 658                  * tcp_portalloc_lock before next submission to Linus.
 659                  * As soon as we touch this place at all it is time to think.
 660                  *
 661                  * Now it protects single _advisory_ variable tcp_port_rover,
 662                  * hence it is mostly useless.
 663                  * Code will work nicely if we just delete it, but
 664                  * I am afraid in contented case it will work not better or
 665                  * even worse: another cpu just will hit the same bucket
 666                  * and spin there.
 667                  * So some cpu salt could remove both contention and
 668                  * memory pingpong. Any ideas how to do this in a nice way?
 669                  */
 670                 spin_lock(&tcp_portalloc_lock);
 671                 rover = tcp_port_rover;
 672
 673                 do {
 674                         rover++;
 675                         if ((rover < low) || (rover > high))
 676                                 rover = low;
 677                         head = &tcp_bhash[tcp_bhashfn(rover)];
 678                         spin_lock(&head->lock);
 679
 680                         /* Does not bother with rcv_saddr checks,
 681                          * because the established check is already
 682                          * unique enough.
 683                          */
 684                         tb_for_each(tb, node, &head->chain) {
 685                                 if (tb->port == rover) {
 686                                         BUG_TRAP(!hlist_empty(&tb->owners));
 687                                         if (tb->fastreuse >= 0)
 688                                                 goto next_port;
 689                                         if (!__tcp_v4_check_established(sk,
 690                                                                         rover,
 691                                                                         &tw))
 692                                                 goto ok;
 693                                         goto next_port;
 694                                 }
 695                         }
 696
 697                         tb = tcp_bucket_create(head, rover);
 698                         if (!tb) {
 699                                 spin_unlock(&head->lock);
 700                                 break;
 701                         }
 702                         tb->fastreuse = -1;
 703                         goto ok;
 704
 705                 next_port:
 706                         spin_unlock(&head->lock);
 707                 } while (--remaining > 0);
 708                 tcp_port_rover = rover;
 709                 spin_unlock(&tcp_portalloc_lock);
 710
 711                 local_bh_enable();
 712
 713                 return -EADDRNOTAVAIL;
 714
 715 ok:
 716                 /* All locks still held and bhs disabled */
 717                 tcp_port_rover = rover;
 718                 spin_unlock(&tcp_portalloc_lock);
 719
 720                 tcp_bind_hash(sk, tb, rover);
 721                 if (sk_unhashed(sk)) {
 722                         inet_sk(sk)->sport = htons(rover);
 723                         __tcp_v4_hash(sk, 0);
 724                 }
 725                 spin_unlock(&head->lock);
 726
 727                 if (tw) {
 728                         tcp_tw_deschedule(tw);
 729                         tcp_tw_put(tw);
 730                 }
 731
 732                 ret = 0;
 733                 goto out;
 734         }
 735
 736         head  = &tcp_bhash[tcp_bhashfn(snum)];
 737         tb  = tcp_sk(sk)->bind_hash;
 738         spin_lock_bh(&head->lock);
 739         if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 740                 __tcp_v4_hash(sk, 0);
 741                 spin_unlock_bh(&head->lock);
 742                 return 0;
 743         } else {
 744                 spin_unlock(&head->lock);
 745                 /* No definite answer... Walk to established hash table */
 746                 ret = __tcp_v4_check_established(sk, snum, NULL);
 747 out:
 748                 local_bh_enable();
 749                 return ret;
 750         }
 751 }
 752
 753 /* This will initiate an outgoing connection. */
 754 int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
 755 {
 756         struct inet_opt *inet = inet_sk(sk);
 757         struct tcp_opt *tp = tcp_sk(sk);
 758         struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
 759         struct rtable *rt;
 760         u32 daddr, nexthop;
 761         int tmp;
 762         int err;
 763
 764         if (addr_len < sizeof(struct sockaddr_in))
 765                 return -EINVAL;
 766
 767         if (usin->sin_family != AF_INET)
 768                 return -EAFNOSUPPORT;
 769
 770         nexthop = daddr = usin->sin_addr.s_addr;
 771         if (inet->opt && inet->opt->srr) {
 772                 if (!daddr)
 773                         return -EINVAL;
 774                 nexthop = inet->opt->faddr;
 775         }
 776
 777         tmp = ip_route_connect(&rt, nexthop, inet->saddr,
 778                                RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
 779                                IPPROTO_TCP,
 780                                inet->sport, usin->sin_port, sk);
 781         if (tmp < 0)
 782                 return tmp;
 783
 784         if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
 785                 ip_rt_put(rt);
 786                 return -ENETUNREACH;
 787         }
 788
 789         if (!inet->opt || !inet->opt->srr)
 790                 daddr = rt->rt_dst;
 791
 792         if (!inet->saddr)
 793                 inet->saddr = rt->rt_src;
 794         inet->rcv_saddr = inet->saddr;
 795
 796         if (tp->ts_recent_stamp && inet->daddr != daddr) {
 797                 /* Reset inherited state */
 798                 tp->ts_recent       = 0;
 799                 tp->ts_recent_stamp = 0;
 800                 tp->write_seq       = 0;
 801         }
 802
 803         if (sysctl_tcp_tw_recycle &&
 804             !tp->ts_recent_stamp && rt->rt_dst == daddr) {
 805                 struct inet_peer *peer = rt_get_peer(rt);
 806
 807                 /* VJ's idea. We save last timestamp seen from
 808                  * the destination in peer table, when entering state TIME-WAIT
 809                  * and initialize ts_recent from it, when trying new connection.
 810                  */
 811
 812                 if (peer && peer->tcp_ts_stamp + TCP_PAWS_MSL >= xtime.tv_sec) {
 813                         tp->ts_recent_stamp = peer->tcp_ts_stamp;
 814                         tp->ts_recent = peer->tcp_ts;
 815                 }
 816         }
 817
 818         inet->dport = usin->sin_port;
 819         inet->daddr = daddr;
 820
 821         tp->ext_header_len = 0;
 822         if (inet->opt)
 823                 tp->ext_header_len = inet->opt->optlen;
 824
 825         tp->mss_clamp = 536;
 826
 827         /* Socket identity is still unknown (sport may be zero).
 828          * However we set state to SYN-SENT and not releasing socket
 829          * lock select source port, enter ourselves into the hash tables and
 830          * complete initialization after this.
 831          */
 832         tcp_set_state(sk, TCP_SYN_SENT);
 833         err = tcp_v4_hash_connect(sk);
 834         if (err)
 835                 goto failure;
 836
 837         err = ip_route_newports(&rt, inet->sport, inet->dport, sk);
 838         if (err)
 839                 goto failure;
 840
 841         /* OK, now commit destination to socket.  */
 842         __sk_dst_set(sk, &rt->u.dst);
 843         tcp_v4_setup_caps(sk, &rt->u.dst);
 844         tp->ext2_header_len = rt->u.dst.header_len;
 845
 846         if (!tp->write_seq)
 847                 tp->write_seq = secure_tcp_sequence_number(inet->saddr,
 848                                                            inet->daddr,
 849                                                            inet->sport,
 850                                                            usin->sin_port);
 851
 852         inet->id = tp->write_seq ^ jiffies;
 853
 854         err = tcp_connect(sk);
 855         rt = NULL;
 856         if (err)
 857                 goto failure;
 858
 859         return 0;
 860
 861 failure:
 862         /* This unhashes the socket and releases the local port, if necessary. */
 863         tcp_set_state(sk, TCP_CLOSE);
 864         ip_rt_put(rt);
 865         sk->sk_route_caps = 0;
 866         inet->dport = 0;
 867         return err;
 868 }
 869
 870 static __inline__ int tcp_v4_iif(struct sk_buff *skb)
 871 {
 872         return ((struct rtable *)skb->dst)->rt_iif;
 873 }
 874
 875 static __inline__ u32 tcp_v4_synq_hash(u32 raddr, u16 rport, u32 rnd)
 876 {
 877         return (jhash_2words(raddr, (u32) rport, rnd) & (TCP_SYNQ_HSIZE - 1));
 878 }
 879
 880 static struct open_request *tcp_v4_search_req(struct tcp_opt *tp,
 881                                               struct open_request ***prevp,
 882                                               __u16 rport,
 883                                               __u32 raddr, __u32 laddr)
 884 {
 885         struct tcp_listen_opt *lopt = tp->listen_opt;
 886         struct open_request *req, **prev;
 887
 888         for (prev = &lopt->syn_table[tcp_v4_synq_hash(raddr, rport, lopt->hash_rnd)];
 889              (req = *prev) != NULL;
 890              prev = &req->dl_next) {
 891                 if (req->rmt_port == rport &&
 892                     req->af.v4_req.rmt_addr == raddr &&
 893                     req->af.v4_req.loc_addr == laddr &&
 894                     TCP_INET_FAMILY(req->class->family)) {
 895                         BUG_TRAP(!req->sk);
 896                         *prevp = prev;
 897                         break;
 898                 }
 899         }
 900
 901         return req;
 902 }
 903
 904 static void tcp_v4_synq_add(struct sock *sk, struct open_request *req)
 905 {
 906         struct tcp_opt *tp = tcp_sk(sk);
 907         struct tcp_listen_opt *lopt = tp->listen_opt;
 908         u32 h = tcp_v4_synq_hash(req->af.v4_req.rmt_addr, req->rmt_port, lopt->hash_rnd);
 909
 910         req->expires = jiffies + TCP_TIMEOUT_INIT;
 911         req->retrans = 0;
 912         req->sk = NULL;
 913         req->dl_next = lopt->syn_table[h];
 914
 915         write_lock(&tp->syn_wait_lock);
 916         lopt->syn_table[h] = req;
 917         write_unlock(&tp->syn_wait_lock);
 918
 919 #ifdef CONFIG_ACCEPT_QUEUES
 920         tcp_synq_added(sk, req);
 921 #else
 922         tcp_synq_added(sk);
 923 #endif
 924 }
 925
 926
 927 /*
 928  * This routine does path mtu discovery as defined in RFC1191.
 929  */
 930 static inline void do_pmtu_discovery(struct sock *sk, struct iphdr *iph,
 931                                      u32 mtu)
 932 {
 933         struct dst_entry *dst;
 934         struct inet_opt *inet = inet_sk(sk);
 935         struct tcp_opt *tp = tcp_sk(sk);
 936
 937         /* We are not interested in TCP_LISTEN and open_requests (SYN-ACKs
 938          * send out by Linux are always <576bytes so they should go through
 939          * unfragmented).
 940          */
 941         if (sk->sk_state == TCP_LISTEN)
 942                 return;
 943
 944         /* We don't check in the destentry if pmtu discovery is forbidden
 945          * on this route. We just assume that no packet_to_big packets
 946          * are send back when pmtu discovery is not active.
 947          * There is a small race when the user changes this flag in the
 948          * route, but I think that's acceptable.
 949          */
 950         if ((dst = __sk_dst_check(sk, 0)) == NULL)
 951                 return;
 952
 953         dst->ops->update_pmtu(dst, mtu);
 954
 955         /* Something is about to be wrong... Remember soft error
 956          * for the case, if this connection will not able to recover.
 957          */
 958         if (mtu < dst_pmtu(dst) && ip_dont_fragment(sk, dst))
 959                 sk->sk_err_soft = EMSGSIZE;
 960
 961         mtu = dst_pmtu(dst);
 962
 963         if (inet->pmtudisc != IP_PMTUDISC_DONT &&
 964             tp->pmtu_cookie > mtu) {
 965                 tcp_sync_mss(sk, mtu);
 966
 967                 /* Resend the TCP packet because it's
 968                  * clear that the old packet has been
 969                  * dropped. This is the new "fast" path mtu
 970                  * discovery.
 971                  */
 972                 tcp_simple_retransmit(sk);
 973         } /* else let the usual retransmit timer handle it */
 974 }
 975
 976 /*
 977  * This routine is called by the ICMP module when it gets some
 978  * sort of error condition.  If err < 0 then the socket should
 979  * be closed and the error returned to the user.  If err > 0
 980  * it's just the icmp type << 8 | icmp code.  After adjustment
 981  * header points to the first 8 bytes of the tcp header.  We need
 982  * to find the appropriate port.
 983  *
 984  * The locking strategy used here is very "optimistic". When
 985  * someone else accesses the socket the ICMP is just dropped
 986  * and for some paths there is no check at all.
 987  * A more general error queue to queue errors for later handling
 988  * is probably better.
 989  *
 990  */
 991
 992 void tcp_v4_err(struct sk_buff *skb, u32 info)
 993 {
 994         struct iphdr *iph = (struct iphdr *)skb->data;
 995         struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
 996         struct tcp_opt *tp;
 997         struct inet_opt *inet;
 998         int type = skb->h.icmph->type;
 999         int code = skb->h.icmph->code;
1000         struct sock *sk;
1001         __u32 seq;
1002         int err;
1003
1004         if (skb->len < (iph->ihl << 2) + 8) {
1005                 ICMP_INC_STATS_BH(ICMP_MIB_INERRORS);
1006                 return;
1007         }
1008
1009         sk = tcp_v4_lookup(iph->daddr, th->dest, iph->saddr,
1010                            th->source, tcp_v4_iif(skb));
1011         if (!sk) {
1012                 ICMP_INC_STATS_BH(ICMP_MIB_INERRORS);
1013                 return;
1014         }
1015         if (sk->sk_state == TCP_TIME_WAIT) {
1016                 tcp_tw_put((struct tcp_tw_bucket *)sk);
1017                 return;
1018         }
1019
1020         bh_lock_sock(sk);
1021         /* If too many ICMPs get dropped on busy
1022          * servers this needs to be solved differently.
1023          */
1024         if (sock_owned_by_user(sk))
1025                 NET_INC_STATS_BH(LINUX_MIB_LOCKDROPPEDICMPS);
1026
1027         if (sk->sk_state == TCP_CLOSE)
1028                 goto out;
1029
1030         tp = tcp_sk(sk);
1031         seq = ntohl(th->seq);
1032         if (sk->sk_state != TCP_LISTEN &&
1033             !between(seq, tp->snd_una, tp->snd_nxt)) {
1034                 NET_INC_STATS(LINUX_MIB_OUTOFWINDOWICMPS);
1035                 goto out;
1036         }
1037
1038         switch (type) {
1039         case ICMP_SOURCE_QUENCH:
1040                 /* Just silently ignore these. */
1041                 goto out;
1042         case ICMP_PARAMETERPROB:
1043                 err = EPROTO;
1044                 break;
1045         case ICMP_DEST_UNREACH:
1046                 if (code > NR_ICMP_UNREACH)
1047                         goto out;
1048
1049                 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
1050                         if (!sock_owned_by_user(sk))
1051                                 do_pmtu_discovery(sk, iph, info);
1052                         goto out;
1053                 }
1054
1055                 err = icmp_err_convert[code].errno;
1056                 break;
1057         case ICMP_TIME_EXCEEDED:
1058                 err = EHOSTUNREACH;
1059                 break;
1060         default:
1061                 goto out;
1062         }
1063
1064         switch (sk->sk_state) {
1065                 struct open_request *req, **prev;
1066         case TCP_LISTEN:
1067                 if (sock_owned_by_user(sk))
1068                         goto out;
1069
1070                 req = tcp_v4_search_req(tp, &prev, th->dest,
1071                                         iph->daddr, iph->saddr);
1072                 if (!req)
1073                         goto out;
1074
1075                 /* ICMPs are not backlogged, hence we cannot get
1076                    an established socket here.
1077                  */
1078                 BUG_TRAP(!req->sk);
1079
1080                 if (seq != req->snt_isn) {
1081                         NET_INC_STATS_BH(LINUX_MIB_OUTOFWINDOWICMPS);
1082                         goto out;
1083                 }
1084
1085                 /*
1086                  * Still in SYN_RECV, just remove it silently.
1087                  * There is no good way to pass the error to the newly
1088                  * created socket, and POSIX does not want network
1089                  * errors returned from accept().
1090                  */
1091                 tcp_synq_drop(sk, req, prev);
1092                 goto out;
1093
1094         case TCP_SYN_SENT:
1095         case TCP_SYN_RECV:  /* Cannot happen.
1096                                It can f.e. if SYNs crossed.
1097                              */
1098                 if (!sock_owned_by_user(sk)) {
1099                         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
1100                         sk->sk_err = err;
1101
1102                         sk->sk_error_report(sk);
1103
1104                         tcp_done(sk);
1105                 } else {
1106                         sk->sk_err_soft = err;
1107                 }
1108                 goto out;
1109         }
1110
1111         /* If we've already connected we will keep trying
1112          * until we time out, or the user gives up.
1113          *
1114          * rfc1122 4.2.3.9 allows to consider as hard errors
1115          * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
1116          * but it is obsoleted by pmtu discovery).
1117          *
1118          * Note, that in modern internet, where routing is unreliable
1119          * and in each dark corner broken firewalls sit, sending random
1120          * errors ordered by their masters even this two messages finally lose
1121          * their original sense (even Linux sends invalid PORT_UNREACHs)
1122          *
1123          * Now we are in compliance with RFCs.
1124          *                                                      --ANK (980905)
1125          */
1126
1127         inet = inet_sk(sk);
1128         if (!sock_owned_by_user(sk) && inet->recverr) {
1129                 sk->sk_err = err;
1130                 sk->sk_error_report(sk);
1131         } else  { /* Only an error on timeout */
1132                 sk->sk_err_soft = err;
1133         }
1134
1135 out:
1136         bh_unlock_sock(sk);
1137         sock_put(sk);
1138 }
1139
1140 /* This routine computes an IPv4 TCP checksum. */
1141 void tcp_v4_send_check(struct sock *sk, struct tcphdr *th, int len,
1142                        struct sk_buff *skb)
1143 {
1144         struct inet_opt *inet = inet_sk(sk);
1145
1146         if (skb->ip_summed == CHECKSUM_HW) {
1147                 th->check = ~tcp_v4_check(th, len, inet->saddr, inet->daddr, 0);
1148                 skb->csum = offsetof(struct tcphdr, check);
1149         } else {
1150                 th->check = tcp_v4_check(th, len, inet->saddr, inet->daddr,
1151                                          csum_partial((char *)th,
1152                                                       th->doff << 2,
1153                                                       skb->csum));
1154         }
1155 }
1156
1157 /*
1158  *      This routine will send an RST to the other tcp.
1159  *
1160  *      Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
1161  *                    for reset.
1162  *      Answer: if a packet caused RST, it is not for a socket
1163  *              existing in our system, if it is matched to a socket,
1164  *              it is just duplicate segment or bug in other side's TCP.
1165  *              So that we build reply only basing on parameters
1166  *              arrived with segment.
1167  *      Exception: precedence violation. We do not implement it in any case.
1168  */
1169
1170 static void tcp_v4_send_reset(struct sk_buff *skb)
1171 {
1172         struct tcphdr *th = skb->h.th;
1173         struct tcphdr rth;
1174         struct ip_reply_arg arg;
1175
1176         /* Never send a reset in response to a reset. */
1177         if (th->rst)
1178                 return;
1179
1180         if (((struct rtable *)skb->dst)->rt_type != RTN_LOCAL)
1181                 return;
1182
1183         /* Swap the send and the receive. */
1184         memset(&rth, 0, sizeof(struct tcphdr));
1185         rth.dest   = th->source;
1186         rth.source = th->dest;
1187         rth.doff   = sizeof(struct tcphdr) / 4;
1188         rth.rst    = 1;
1189
1190         if (th->ack) {
1191                 rth.seq = th->ack_seq;
1192         } else {
1193                 rth.ack = 1;
1194                 rth.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
1195                                     skb->len - (th->doff << 2));
1196         }
1197
1198         memset(&arg, 0, sizeof arg);
1199         arg.iov[0].iov_base = (unsigned char *)&rth;
1200         arg.iov[0].iov_len  = sizeof rth;
1201         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
1202                                       skb->nh.iph->saddr, /*XXX*/
1203                                       sizeof(struct tcphdr), IPPROTO_TCP, 0);
1204         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1205
1206         ip_send_reply(tcp_socket->sk, skb, &arg, sizeof rth);
1207
1208         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
1209         TCP_INC_STATS_BH(TCP_MIB_OUTRSTS);
1210 }
1211
1212 /* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
1213    outside socket context is ugly, certainly. What can I do?
1214  */
1215
1216 static void tcp_v4_send_ack(struct sk_buff *skb, u32 seq, u32 ack,
1217                             u32 win, u32 ts)
1218 {
1219         struct tcphdr *th = skb->h.th;
1220         struct {
1221                 struct tcphdr th;
1222                 u32 tsopt[3];
1223         } rep;
1224         struct ip_reply_arg arg;
1225
1226         memset(&rep.th, 0, sizeof(struct tcphdr));
1227         memset(&arg, 0, sizeof arg);
1228
1229         arg.iov[0].iov_base = (unsigned char *)&rep;
1230         arg.iov[0].iov_len  = sizeof(rep.th);
1231         if (ts) {
1232                 rep.tsopt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1233                                      (TCPOPT_TIMESTAMP << 8) |
1234                                      TCPOLEN_TIMESTAMP);
1235                 rep.tsopt[1] = htonl(tcp_time_stamp);
1236                 rep.tsopt[2] = htonl(ts);
1237                 arg.iov[0].iov_len = sizeof(rep);
1238         }
1239
1240         /* Swap the send and the receive. */
1241         rep.th.dest    = th->source;
1242         rep.th.source  = th->dest;
1243         rep.th.doff    = arg.iov[0].iov_len / 4;
1244         rep.th.seq     = htonl(seq);
1245         rep.th.ack_seq = htonl(ack);
1246         rep.th.ack     = 1;
1247         rep.th.window  = htons(win);
1248
1249         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
1250                                       skb->nh.iph->saddr, /*XXX*/
1251                                       arg.iov[0].iov_len, IPPROTO_TCP, 0);
1252         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1253
1254         ip_send_reply(tcp_socket->sk, skb, &arg, arg.iov[0].iov_len);
1255
1256         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
1257 }
1258
1259 static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
1260 {
1261         struct tcp_tw_bucket *tw = (struct tcp_tw_bucket *)sk;
1262
1263         tcp_v4_send_ack(skb, tw->tw_snd_nxt, tw->tw_rcv_nxt,
1264                         tw->tw_rcv_wnd >> tw->tw_rcv_wscale, tw->tw_ts_recent);
1265
1266         tcp_tw_put(tw);
1267 }
1268
1269 static void tcp_v4_or_send_ack(struct sk_buff *skb, struct open_request *req)
1270 {
1271         tcp_v4_send_ack(skb, req->snt_isn + 1, req->rcv_isn + 1, req->rcv_wnd,
1272                         req->ts_recent);
1273 }
1274
1275 static struct dst_entry* tcp_v4_route_req(struct sock *sk,
1276                                           struct open_request *req)
1277 {
1278         struct rtable *rt;
1279         struct ip_options *opt = req->af.v4_req.opt;
1280         struct flowi fl = { .oif = sk->sk_bound_dev_if,
1281                             .nl_u = { .ip4_u =
1282                                       { .daddr = ((opt && opt->srr) ?
1283                                                   opt->faddr :
1284                                                   req->af.v4_req.rmt_addr),
1285                                         .saddr = req->af.v4_req.loc_addr,
1286                                         .tos = RT_CONN_FLAGS(sk) } },
1287                             .proto = IPPROTO_TCP,
1288                             .uli_u = { .ports =
1289                                        { .sport = inet_sk(sk)->sport,
1290                                          .dport = req->rmt_port } } };
1291
1292         if (ip_route_output_flow(&rt, &fl, sk, 0)) {
1293                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
1294                 return NULL;
1295         }
1296         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway) {
1297                 ip_rt_put(rt);
1298                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
1299                 return NULL;
1300         }
1301         return &rt->u.dst;
1302 }
1303
1304 /*
1305  *      Send a SYN-ACK after having received an ACK.
1306  *      This still operates on a open_request only, not on a big
1307  *      socket.
1308  */
1309 static int tcp_v4_send_synack(struct sock *sk, struct open_request *req,
1310                               struct dst_entry *dst)
1311 {
1312         int err = -1;
1313         struct sk_buff * skb;
1314
1315         /* First, grab a route. */
1316         if (!dst && (dst = tcp_v4_route_req(sk, req)) == NULL)
1317                 goto out;
1318
1319         skb = tcp_make_synack(sk, dst, req);
1320
1321         if (skb) {
1322                 struct tcphdr *th = skb->h.th;
1323
1324                 th->check = tcp_v4_check(th, skb->len,
1325                                          req->af.v4_req.loc_addr,
1326                                          req->af.v4_req.rmt_addr,
1327                                          csum_partial((char *)th, skb->len,
1328                                                       skb->csum));
1329
1330                 err = ip_build_and_send_pkt(skb, sk, req->af.v4_req.loc_addr,
1331                                             req->af.v4_req.rmt_addr,
1332                                             req->af.v4_req.opt);
1333                 if (err == NET_XMIT_CN)
1334                         err = 0;
1335         }
1336
1337 out:
1338         dst_release(dst);
1339         return err;
1340 }
1341
1342 /*
1343  *      IPv4 open_request destructor.
1344  */
1345 static void tcp_v4_or_free(struct open_request *req)
1346 {
1347         if (req->af.v4_req.opt)
1348                 kfree(req->af.v4_req.opt);
1349 }
1350
1351 static inline void syn_flood_warning(struct sk_buff *skb)
1352 {
1353         static unsigned long warntime;
1354
1355         if (time_after(jiffies, (warntime + HZ * 60))) {
1356                 warntime = jiffies;
1357                 printk(KERN_INFO
1358                        "possible SYN flooding on port %d. Sending cookies.\n",
1359                        ntohs(skb->h.th->dest));
1360         }
1361 }
1362
1363 /*
1364  * Save and compile IPv4 options into the open_request if needed.
1365  */
1366 static inline struct ip_options *tcp_v4_save_options(struct sock *sk,
1367                                                      struct sk_buff *skb)
1368 {
1369         struct ip_options *opt = &(IPCB(skb)->opt);
1370         struct ip_options *dopt = NULL;
1371
1372         if (opt && opt->optlen) {
1373                 int opt_size = optlength(opt);
1374                 dopt = kmalloc(opt_size, GFP_ATOMIC);
1375                 if (dopt) {
1376                         if (ip_options_echo(dopt, skb)) {
1377                                 kfree(dopt);
1378                                 dopt = NULL;
1379                         }
1380                 }
1381         }
1382         return dopt;
1383 }
1384
1385 /*
1386  * Maximum number of SYN_RECV sockets in queue per LISTEN socket.
1387  * One SYN_RECV socket costs about 80bytes on a 32bit machine.
1388  * It would be better to replace it with a global counter for all sockets
1389  * but then some measure against one socket starving all other sockets
1390  * would be needed.
1391  *
1392  * It was 128 by default. Experiments with real servers show, that
1393  * it is absolutely not enough even at 100conn/sec. 256 cures most
1394  * of problems. This value is adjusted to 128 for very small machines
1395  * (<=32Mb of memory) and to 1024 on normal or better ones (>=256Mb).
1396  * Further increasing requires to change hash table size.
1397  */
1398 int sysctl_max_syn_backlog = 256;
1399
1400 struct or_calltable or_ipv4 = {
1401         .family         =       PF_INET,
1402         .rtx_syn_ack    =       tcp_v4_send_synack,
1403         .send_ack       =       tcp_v4_or_send_ack,
1404         .destructor     =       tcp_v4_or_free,
1405         .send_reset     =       tcp_v4_send_reset,
1406 };
1407
1408 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1409 {
1410         struct tcp_opt tp;
1411         struct open_request *req;
1412         __u32 saddr = skb->nh.iph->saddr;
1413         __u32 daddr = skb->nh.iph->daddr;
1414         __u32 isn = TCP_SKB_CB(skb)->when;
1415         struct dst_entry *dst = NULL;
1416 #ifdef CONFIG_ACCEPT_QUEUES
1417         int class = 0;
1418 #endif
1419 #ifdef CONFIG_SYN_COOKIES
1420         int want_cookie = 0;
1421 #else
1422 #define want_cookie 0 /* Argh, why doesn't gcc optimize this :( */
1423 #endif
1424
1425         /* Never answer to SYNs send to broadcast or multicast */
1426         if (((struct rtable *)skb->dst)->rt_flags &
1427             (RTCF_BROADCAST | RTCF_MULTICAST))
1428                 goto drop;
1429
1430         /* TW buckets are converted to open requests without
1431          * limitations, they conserve resources and peer is
1432          * evidently real one.
1433          */
1434         if (tcp_synq_is_full(sk) && !isn) {
1435 #ifdef CONFIG_SYN_COOKIES
1436                 if (sysctl_tcp_syncookies) {
1437                         want_cookie = 1;
1438                 } else
1439 #endif
1440                 goto drop;
1441         }
1442
1443 #ifdef CONFIG_ACCEPT_QUEUES
1444         class = (skb->nfmark <= 0) ? 0 :
1445                 ((skb->nfmark >= NUM_ACCEPT_QUEUES) ? 0: skb->nfmark);
1446         /*
1447          * Accept only if the class has shares set or if the default class
1448          * i.e. class 0 has shares
1449          */
1450         if (!(tcp_sk(sk)->acceptq[class].aq_ratio)) {
1451                 if (tcp_sk(sk)->acceptq[0].aq_ratio)
1452                         class = 0;
1453                 else
1454                         goto drop;
1455         }
1456 #endif
1457
1458         /* Accept backlog is full. If we have already queued enough
1459          * of warm entries in syn queue, drop request. It is better than
1460          * clogging syn queue with openreqs with exponentially increasing
1461          * timeout.
1462          */
1463 #ifdef CONFIG_ACCEPT_QUEUES
1464         if (sk_acceptq_is_full(sk, class) && tcp_synq_young(sk, class) > 1)
1465 #else
1466         if (sk_acceptq_is_full(sk) && tcp_synq_young(sk) > 1)
1467 #endif
1468                 goto drop;
1469
1470         req = tcp_openreq_alloc();
1471         if (!req)
1472                 goto drop;
1473
1474         tcp_clear_options(&tp);
1475         tp.mss_clamp = 536;
1476         tp.user_mss  = tcp_sk(sk)->user_mss;
1477
1478         tcp_parse_options(skb, &tp, 0);
1479
1480         if (want_cookie) {
1481                 tcp_clear_options(&tp);
1482                 tp.saw_tstamp = 0;
1483         }
1484
1485         if (tp.saw_tstamp && !tp.rcv_tsval) {
1486                 /* Some OSes (unknown ones, but I see them on web server, which
1487                  * contains information interesting only for windows'
1488                  * users) do not send their stamp in SYN. It is easy case.
1489                  * We simply do not advertise TS support.
1490                  */
1491                 tp.saw_tstamp = 0;
1492                 tp.tstamp_ok  = 0;
1493         }
1494         tp.tstamp_ok = tp.saw_tstamp;
1495
1496         tcp_openreq_init(req, &tp, skb);
1497 #ifdef CONFIG_ACCEPT_QUEUES
1498         req->acceptq_class = class;
1499         req->acceptq_time_stamp = jiffies;
1500 #endif
1501         req->af.v4_req.loc_addr = daddr;
1502         req->af.v4_req.rmt_addr = saddr;
1503         req->af.v4_req.opt = tcp_v4_save_options(sk, skb);
1504         req->class = &or_ipv4;
1505         if (!want_cookie)
1506                 TCP_ECN_create_request(req, skb->h.th);
1507
1508         if (want_cookie) {
1509 #ifdef CONFIG_SYN_COOKIES
1510                 syn_flood_warning(skb);
1511 #endif
1512                 isn = cookie_v4_init_sequence(sk, skb, &req->mss);
1513         } else if (!isn) {
1514                 struct inet_peer *peer = NULL;
1515
1516                 /* VJ's idea. We save last timestamp seen
1517                  * from the destination in peer table, when entering
1518                  * state TIME-WAIT, and check against it before
1519                  * accepting new connection request.
1520                  *
1521                  * If "isn" is not zero, this request hit alive
1522                  * timewait bucket, so that all the necessary checks
1523                  * are made in the function processing timewait state.
1524                  */
1525                 if (tp.saw_tstamp &&
1526                     sysctl_tcp_tw_recycle &&
1527                     (dst = tcp_v4_route_req(sk, req)) != NULL &&
1528                     (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
1529                     peer->v4daddr == saddr) {
1530                         if (xtime.tv_sec < peer->tcp_ts_stamp + TCP_PAWS_MSL &&
1531                             (s32)(peer->tcp_ts - req->ts_recent) >
1532                                                         TCP_PAWS_WINDOW) {
1533                                 NET_INC_STATS_BH(LINUX_MIB_PAWSPASSIVEREJECTED);
1534                                 dst_release(dst);
1535                                 goto drop_and_free;
1536                         }
1537                 }
1538                 /* Kill the following clause, if you dislike this way. */
1539                 else if (!sysctl_tcp_syncookies &&
1540                          (sysctl_max_syn_backlog - tcp_synq_len(sk) <
1541                           (sysctl_max_syn_backlog >> 2)) &&
1542                          (!peer || !peer->tcp_ts_stamp) &&
1543                          (!dst || !dst_metric(dst, RTAX_RTT))) {
1544                         /* Without syncookies last quarter of
1545                          * backlog is filled with destinations,
1546                          * proven to be alive.
1547                          * It means that we continue to communicate
1548                          * to destinations, already remembered
1549                          * to the moment of synflood.
1550                          */
1551                         NETDEBUG(if (net_ratelimit()) \
1552                                         printk(KERN_DEBUG "TCP: drop open "
1553                                                           "request from %u.%u."
1554                                                           "%u.%u/%u\n", \
1555                                                NIPQUAD(saddr),
1556                                                ntohs(skb->h.th->source)));
1557                         dst_release(dst);
1558                         goto drop_and_free;
1559                 }
1560
1561                 isn = tcp_v4_init_sequence(sk, skb);
1562         }
1563         req->snt_isn = isn;
1564
1565         if (tcp_v4_send_synack(sk, req, dst))
1566                 goto drop_and_free;
1567
1568         if (want_cookie) {
1569                 tcp_openreq_free(req);
1570         } else {
1571                 tcp_v4_synq_add(sk, req);
1572         }
1573         return 0;
1574
1575 drop_and_free:
1576         tcp_openreq_free(req);
1577 drop:
1578         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
1579         return 0;
1580 }
1581
1582
1583 /*
1584  * The three way handshake has completed - we got a valid synack -
1585  * now create the new socket.
1586  */
1587 struct sock *tcp_v4_syn_recv_sock(struct sock *sk, struct sk_buff *skb,
1588                                   struct open_request *req,
1589                                   struct dst_entry *dst)
1590 {
1591         struct inet_opt *newinet;
1592         struct tcp_opt *newtp;
1593         struct sock *newsk;
1594
1595 #ifdef CONFIG_ACCEPT_QUEUES
1596         if (sk_acceptq_is_full(sk, req->acceptq_class))
1597 #else
1598         if (sk_acceptq_is_full(sk))
1599 #endif
1600                 goto exit_overflow;
1601
1602         if (!dst && (dst = tcp_v4_route_req(sk, req)) == NULL)
1603                 goto exit;
1604
1605         newsk = tcp_create_openreq_child(sk, req, skb);
1606         if (!newsk)
1607                 goto exit;
1608
1609         newsk->sk_dst_cache = dst;
1610         tcp_v4_setup_caps(newsk, dst);
1611
1612         newtp                 = tcp_sk(newsk);
1613         newinet               = inet_sk(newsk);
1614         newinet->daddr        = req->af.v4_req.rmt_addr;
1615         newinet->rcv_saddr    = req->af.v4_req.loc_addr;
1616         newinet->saddr        = req->af.v4_req.loc_addr;
1617         newinet->opt          = req->af.v4_req.opt;
1618         req->af.v4_req.opt    = NULL;
1619         newinet->mc_index     = tcp_v4_iif(skb);
1620         newinet->mc_ttl       = skb->nh.iph->ttl;
1621         newtp->ext_header_len = 0;
1622         if (newinet->opt)
1623                 newtp->ext_header_len = newinet->opt->optlen;
1624         newtp->ext2_header_len = dst->header_len;
1625         newinet->id = newtp->write_seq ^ jiffies;
1626
1627         tcp_sync_mss(newsk, dst_pmtu(dst));
1628         newtp->advmss = dst_metric(dst, RTAX_ADVMSS);
1629         tcp_initialize_rcv_mss(newsk);
1630
1631         __tcp_v4_hash(newsk, 0);
1632         __tcp_inherit_port(sk, newsk);
1633
1634         return newsk;
1635
1636 exit_overflow:
1637         NET_INC_STATS_BH(LINUX_MIB_LISTENOVERFLOWS);
1638 exit:
1639         NET_INC_STATS_BH(LINUX_MIB_LISTENDROPS);
1640         dst_release(dst);
1641         return NULL;
1642 }
1643
1644 static struct sock *tcp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
1645 {
1646         struct tcphdr *th = skb->h.th;
1647         struct iphdr *iph = skb->nh.iph;
1648         struct tcp_opt *tp = tcp_sk(sk);
1649         struct sock *nsk;
1650         struct open_request **prev;
1651         /* Find possible connection requests. */
1652         struct open_request *req = tcp_v4_search_req(tp, &prev, th->source,
1653                                                      iph->saddr, iph->daddr);
1654         if (req)
1655                 return tcp_check_req(sk, skb, req, prev);
1656
1657         nsk = __tcp_v4_lookup_established(skb->nh.iph->saddr,
1658                                           th->source,
1659                                           skb->nh.iph->daddr,
1660                                           ntohs(th->dest),
1661                                           tcp_v4_iif(skb));
1662
1663         if (nsk) {
1664                 if (nsk->sk_state != TCP_TIME_WAIT) {
1665                         bh_lock_sock(nsk);
1666                         return nsk;
1667                 }
1668                 tcp_tw_put((struct tcp_tw_bucket *)nsk);
1669                 return NULL;
1670         }
1671
1672 #ifdef CONFIG_SYN_COOKIES
1673         if (!th->rst && !th->syn && th->ack)
1674                 sk = cookie_v4_check(sk, skb, &(IPCB(skb)->opt));
1675 #endif
1676         return sk;
1677 }
1678
1679 static int tcp_v4_checksum_init(struct sk_buff *skb)
1680 {
1681         if (skb->ip_summed == CHECKSUM_HW) {
1682                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1683                 if (!tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
1684                                   skb->nh.iph->daddr, skb->csum))
1685                         return 0;
1686
1687                 NETDEBUG(if (net_ratelimit())
1688                                 printk(KERN_DEBUG "hw tcp v4 csum failed\n"));
1689                 skb->ip_summed = CHECKSUM_NONE;
1690         }
1691         if (skb->len <= 76) {
1692                 if (tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
1693                                  skb->nh.iph->daddr,
1694                                  skb_checksum(skb, 0, skb->len, 0)))
1695                         return -1;
1696                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1697         } else {
1698                 skb->csum = ~tcp_v4_check(skb->h.th, skb->len,
1699                                           skb->nh.iph->saddr,
1700                                           skb->nh.iph->daddr, 0);
1701         }
1702         return 0;
1703 }
1704
1705
1706 /* The socket must have it's spinlock held when we get
1707  * here.
1708  *
1709  * We have a potential double-lock case here, so even when
1710  * doing backlog processing we use the BH locking scheme.
1711  * This is because we cannot sleep with the original spinlock
1712  * held.
1713  */
1714 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1715 {
1716         if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1717                 TCP_CHECK_TIMER(sk);
1718                 if (tcp_rcv_established(sk, skb, skb->h.th, skb->len))
1719                         goto reset;
1720                 TCP_CHECK_TIMER(sk);
1721                 return 0;
1722         }
1723
1724         if (skb->len < (skb->h.th->doff << 2) || tcp_checksum_complete(skb))
1725                 goto csum_err;
1726
1727         if (sk->sk_state == TCP_LISTEN) {
1728                 struct sock *nsk = tcp_v4_hnd_req(sk, skb);
1729                 if (!nsk)
1730                         goto discard;
1731
1732                 if (nsk != sk) {
1733                         if (tcp_child_process(sk, nsk, skb))
1734                                 goto reset;
1735                         return 0;
1736                 }
1737         }
1738
1739         TCP_CHECK_TIMER(sk);
1740         if (tcp_rcv_state_process(sk, skb, skb->h.th, skb->len))
1741                 goto reset;
1742         TCP_CHECK_TIMER(sk);
1743         return 0;
1744
1745 reset:
1746         tcp_v4_send_reset(skb);
1747 discard:
1748         kfree_skb(skb);
1749         /* Be careful here. If this function gets more complicated and
1750          * gcc suffers from register pressure on the x86, sk (in %ebx)
1751          * might be destroyed here. This current version compiles correctly,
1752          * but you have been warned.
1753          */
1754         return 0;
1755
1756 csum_err:
1757         TCP_INC_STATS_BH(TCP_MIB_INERRS);
1758         goto discard;
1759 }
1760
1761 extern struct proto_ops inet_stream_ops;
1762
1763 extern int inet_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
1764
1765 /*
1766  *      From tcp_input.c
1767  */
1768
1769 int tcp_v4_rcv(struct sk_buff *skb)
1770 {
1771         struct tcphdr *th;
1772         struct sock *sk;
1773         int ret;
1774
1775         if (skb->pkt_type != PACKET_HOST)
1776                 goto discard_it;
1777
1778         /* Count it even if it's bad */
1779         TCP_INC_STATS_BH(TCP_MIB_INSEGS);
1780
1781         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1782                 goto discard_it;
1783
1784         th = skb->h.th;
1785
1786         if (th->doff < sizeof(struct tcphdr) / 4)
1787                 goto bad_packet;
1788         if (!pskb_may_pull(skb, th->doff * 4))
1789                 goto discard_it;
1790
1791         /* An explanation is required here, I think.
1792          * Packet length and doff are validated by header prediction,
1793          * provided case of th->doff==0 is elimineted.
1794          * So, we defer the checks. */
1795         if ((skb->ip_summed != CHECKSUM_UNNECESSARY &&
1796              tcp_v4_checksum_init(skb) < 0))
1797                 goto bad_packet;
1798
1799         th = skb->h.th;
1800         TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1801         TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1802                                     skb->len - th->doff * 4);
1803         TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1804         TCP_SKB_CB(skb)->when    = 0;
1805         TCP_SKB_CB(skb)->flags   = skb->nh.iph->tos;
1806         TCP_SKB_CB(skb)->sacked  = 0;
1807
1808         sk = __tcp_v4_lookup(skb->nh.iph->saddr, th->source,
1809                              skb->nh.iph->daddr, ntohs(th->dest),
1810                              tcp_v4_iif(skb));
1811
1812         if (!sk)
1813                 goto no_tcp_socket;
1814
1815 process:
1816         /* Silently drop if VNET is active (if INET bind() has been
1817          * overridden) and the context is not entitled to read the
1818          * packet.
1819          */
1820         if (inet_stream_ops.bind != inet_bind &&
1821             (int) sk->sk_xid > 0 && sk->sk_xid != skb->xid)
1822                 goto discard_it;
1823
1824         if (sk->sk_state == TCP_TIME_WAIT)
1825                 goto do_time_wait;
1826
1827         if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1828                 goto discard_and_relse;
1829
1830         if (sk_filter(sk, skb, 0))
1831                 goto discard_and_relse;
1832
1833         skb->dev = NULL;
1834
1835         bh_lock_sock(sk);
1836         ret = 0;
1837         if (!sock_owned_by_user(sk)) {
1838                 if (!tcp_prequeue(sk, skb))
1839                         ret = tcp_v4_do_rcv(sk, skb);
1840         } else
1841                 sk_add_backlog(sk, skb);
1842         bh_unlock_sock(sk);
1843
1844         sock_put(sk);
1845
1846         return ret;
1847
1848 no_tcp_socket:
1849         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1850                 goto discard_it;
1851
1852         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1853 bad_packet:
1854                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1855         } else if (!skb->sk) {
1856                 /* VNET: Suppress RST if the port was bound to a (presumably raw) socket */
1857                 tcp_v4_send_reset(skb);
1858         }
1859
1860 discard_it:
1861         /* Discard frame. */
1862         kfree_skb(skb);
1863         return 0;
1864
1865 discard_and_relse:
1866         sock_put(sk);
1867         goto discard_it;
1868
1869 do_time_wait:
1870         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1871                 tcp_tw_put((struct tcp_tw_bucket *) sk);
1872                 goto discard_it;
1873         }
1874
1875         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1876                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1877                 tcp_tw_put((struct tcp_tw_bucket *) sk);
1878                 goto discard_it;
1879         }
1880         switch (tcp_timewait_state_process((struct tcp_tw_bucket *)sk,
1881                                            skb, th, skb->len)) {
1882         case TCP_TW_SYN: {
1883                 struct sock *sk2 = tcp_v4_lookup_listener(skb->nh.iph->daddr,
1884                                                           ntohs(th->dest),
1885                                                           tcp_v4_iif(skb));
1886                 if (sk2) {
1887                         tcp_tw_deschedule((struct tcp_tw_bucket *)sk);
1888                         tcp_tw_put((struct tcp_tw_bucket *)sk);
1889                         sk = sk2;
1890                         goto process;
1891                 }
1892                 /* Fall through to ACK */
1893         }
1894         case TCP_TW_ACK:
1895                 tcp_v4_timewait_ack(sk, skb);
1896                 break;
1897         case TCP_TW_RST:
1898                 goto no_tcp_socket;
1899         case TCP_TW_SUCCESS:;
1900         }
1901         goto discard_it;
1902 }
1903
1904 /* With per-bucket locks this operation is not-atomic, so that
1905  * this version is not worse.
1906  */
1907 static void __tcp_v4_rehash(struct sock *sk)
1908 {
1909         sk->sk_prot->unhash(sk);
1910         sk->sk_prot->hash(sk);
1911 }
1912
1913 static int tcp_v4_reselect_saddr(struct sock *sk)
1914 {
1915         struct inet_opt *inet = inet_sk(sk);
1916         int err;
1917         struct rtable *rt;
1918         __u32 old_saddr = inet->saddr;
1919         __u32 new_saddr;
1920         __u32 daddr = inet->daddr;
1921
1922         if (inet->opt && inet->opt->srr)
1923                 daddr = inet->opt->faddr;
1924
1925         /* Query new route. */
1926         err = ip_route_connect(&rt, daddr, 0,
1927                                RT_TOS(inet->tos) | sk->sk_localroute,
1928                                sk->sk_bound_dev_if,
1929                                IPPROTO_TCP,
1930                                inet->sport, inet->dport, sk);
1931         if (err)
1932                 return err;
1933
1934         __sk_dst_set(sk, &rt->u.dst);
1935         tcp_v4_setup_caps(sk, &rt->u.dst);
1936         tcp_sk(sk)->ext2_header_len = rt->u.dst.header_len;
1937
1938         new_saddr = rt->rt_src;
1939
1940         if (new_saddr == old_saddr)
1941                 return 0;
1942
1943         if (sysctl_ip_dynaddr > 1) {
1944                 printk(KERN_INFO "tcp_v4_rebuild_header(): shifting inet->"
1945                                  "saddr from %d.%d.%d.%d to %d.%d.%d.%d\n",
1946                        NIPQUAD(old_saddr),
1947                        NIPQUAD(new_saddr));
1948         }
1949
1950         inet->saddr = new_saddr;
1951         inet->rcv_saddr = new_saddr;
1952
1953         /* XXX The only one ugly spot where we need to
1954          * XXX really change the sockets identity after
1955          * XXX it has entered the hashes. -DaveM
1956          *
1957          * Besides that, it does not check for connection
1958          * uniqueness. Wait for troubles.
1959          */
1960         __tcp_v4_rehash(sk);
1961         return 0;
1962 }
1963
1964 int tcp_v4_rebuild_header(struct sock *sk)
1965 {
1966         struct inet_opt *inet = inet_sk(sk);
1967         struct rtable *rt = (struct rtable *)__sk_dst_check(sk, 0);
1968         u32 daddr;
1969         int err;
1970
1971         /* Route is OK, nothing to do. */
1972         if (rt)
1973                 return 0;
1974
1975         /* Reroute. */
1976         daddr = inet->daddr;
1977         if (inet->opt && inet->opt->srr)
1978                 daddr = inet->opt->faddr;
1979
1980         {
1981                 struct flowi fl = { .oif = sk->sk_bound_dev_if,
1982                                     .nl_u = { .ip4_u =
1983                                               { .daddr = daddr,
1984                                                 .saddr = inet->saddr,
1985                                                 .tos = RT_CONN_FLAGS(sk) } },
1986                                     .proto = IPPROTO_TCP,
1987                                     .uli_u = { .ports =
1988                                                { .sport = inet->sport,
1989                                                  .dport = inet->dport } } };
1990
1991                 err = ip_route_output_flow(&rt, &fl, sk, 0);
1992         }
1993         if (!err) {
1994                 __sk_dst_set(sk, &rt->u.dst);
1995                 tcp_v4_setup_caps(sk, &rt->u.dst);
1996                 tcp_sk(sk)->ext2_header_len = rt->u.dst.header_len;
1997                 return 0;
1998         }
1999
2000         /* Routing failed... */
2001         sk->sk_route_caps = 0;
2002
2003         if (!sysctl_ip_dynaddr ||
2004             sk->sk_state != TCP_SYN_SENT ||
2005             (sk->sk_userlocks & SOCK_BINDADDR_LOCK) ||
2006             (err = tcp_v4_reselect_saddr(sk)) != 0)
2007                 sk->sk_err_soft = -err;
2008
2009         return err;
2010 }
2011
2012 static void v4_addr2sockaddr(struct sock *sk, struct sockaddr * uaddr)
2013 {
2014         struct sockaddr_in *sin = (struct sockaddr_in *) uaddr;
2015         struct inet_opt *inet = inet_sk(sk);
2016
2017         sin->sin_family         = AF_INET;
2018         sin->sin_addr.s_addr    = inet->daddr;
2019         sin->sin_port           = inet->dport;
2020 }
2021
2022 /* VJ's idea. Save last timestamp seen from this destination
2023  * and hold it at least for normal timewait interval to use for duplicate
2024  * segment detection in subsequent connections, before they enter synchronized
2025  * state.
2026  */
2027
2028 int tcp_v4_remember_stamp(struct sock *sk)
2029 {
2030         struct inet_opt *inet = inet_sk(sk);
2031         struct tcp_opt *tp = tcp_sk(sk);
2032         struct rtable *rt = (struct rtable *)__sk_dst_get(sk);
2033         struct inet_peer *peer = NULL;
2034         int release_it = 0;
2035
2036         if (!rt || rt->rt_dst != inet->daddr) {
2037                 peer = inet_getpeer(inet->daddr, 1);
2038                 release_it = 1;
2039         } else {
2040                 if (!rt->peer)
2041                         rt_bind_peer(rt, 1);
2042                 peer = rt->peer;
2043         }
2044
2045         if (peer) {
2046                 if ((s32)(peer->tcp_ts - tp->ts_recent) <= 0 ||
2047                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
2048                      peer->tcp_ts_stamp <= tp->ts_recent_stamp)) {
2049                         peer->tcp_ts_stamp = tp->ts_recent_stamp;
2050                         peer->tcp_ts = tp->ts_recent;
2051                 }
2052                 if (release_it)
2053                         inet_putpeer(peer);
2054                 return 1;
2055         }
2056
2057         return 0;
2058 }
2059
2060 int tcp_v4_tw_remember_stamp(struct tcp_tw_bucket *tw)
2061 {
2062         struct inet_peer *peer = NULL;
2063
2064         peer = inet_getpeer(tw->tw_daddr, 1);
2065
2066         if (peer) {
2067                 if ((s32)(peer->tcp_ts - tw->tw_ts_recent) <= 0 ||
2068                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
2069                      peer->tcp_ts_stamp <= tw->tw_ts_recent_stamp)) {
2070                         peer->tcp_ts_stamp = tw->tw_ts_recent_stamp;
2071                         peer->tcp_ts = tw->tw_ts_recent;
2072                 }
2073                 inet_putpeer(peer);
2074                 return 1;
2075         }
2076
2077         return 0;
2078 }
2079
2080 struct tcp_func ipv4_specific = {
2081         .queue_xmit     =       ip_queue_xmit,
2082         .send_check     =       tcp_v4_send_check,
2083         .rebuild_header =       tcp_v4_rebuild_header,
2084         .conn_request   =       tcp_v4_conn_request,
2085         .syn_recv_sock  =       tcp_v4_syn_recv_sock,
2086         .remember_stamp =       tcp_v4_remember_stamp,
2087         .net_header_len =       sizeof(struct iphdr),
2088         .setsockopt     =       ip_setsockopt,
2089         .getsockopt     =       ip_getsockopt,
2090         .addr2sockaddr  =       v4_addr2sockaddr,
2091         .sockaddr_len   =       sizeof(struct sockaddr_in),
2092 };
2093
2094 /* NOTE: A lot of things set to zero explicitly by call to
2095  *       sk_alloc() so need not be done here.
2096  */
2097 static int tcp_v4_init_sock(struct sock *sk)
2098 {
2099         struct tcp_opt *tp = tcp_sk(sk);
2100
2101         skb_queue_head_init(&tp->out_of_order_queue);
2102         tcp_init_xmit_timers(sk);
2103         tcp_prequeue_init(tp);
2104
2105         tp->rto  = TCP_TIMEOUT_INIT;
2106         tp->mdev = TCP_TIMEOUT_INIT;
2107
2108         /* So many TCP implementations out there (incorrectly) count the
2109          * initial SYN frame in their delayed-ACK and congestion control
2110          * algorithms that we must have the following bandaid to talk
2111          * efficiently to them.  -DaveM
2112          */
2113         tp->snd_cwnd = 2;
2114
2115         /* See draft-stevens-tcpca-spec-01 for discussion of the
2116          * initialization of these values.
2117          */
2118         tp->snd_ssthresh = 0x7fffffff;  /* Infinity */
2119         tp->snd_cwnd_clamp = ~0;
2120         tp->mss_cache_std = tp->mss_cache = 536;
2121
2122         tp->reordering = sysctl_tcp_reordering;
2123
2124         sk->sk_state = TCP_CLOSE;
2125
2126         sk->sk_write_space = sk_stream_write_space;
2127         sk->sk_use_write_queue = 1;
2128
2129         tp->af_specific = &ipv4_specific;
2130
2131         sk->sk_sndbuf = sysctl_tcp_wmem[1];
2132         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
2133
2134         atomic_inc(&tcp_sockets_allocated);
2135
2136         return 0;
2137 }
2138
2139 int tcp_v4_destroy_sock(struct sock *sk)
2140 {
2141         struct tcp_opt *tp = tcp_sk(sk);
2142
2143         tcp_clear_xmit_timers(sk);
2144
2145         /* Cleanup up the write buffer. */
2146         sk_stream_writequeue_purge(sk);
2147
2148         /* Cleans up our, hopefully empty, out_of_order_queue. */
2149         __skb_queue_purge(&tp->out_of_order_queue);
2150
2151         /* Clean prequeue, it must be empty really */
2152         __skb_queue_purge(&tp->ucopy.prequeue);
2153
2154         /* Clean up a referenced TCP bind bucket. */
2155         if (tp->bind_hash)
2156                 tcp_put_port(sk);
2157
2158         /*
2159          * If sendmsg cached page exists, toss it.
2160          */
2161         if (sk->sk_sndmsg_page) {
2162                 __free_page(sk->sk_sndmsg_page);
2163                 sk->sk_sndmsg_page = NULL;
2164         }
2165
2166         atomic_dec(&tcp_sockets_allocated);
2167
2168         return 0;
2169 }
2170
2171 EXPORT_SYMBOL(tcp_v4_destroy_sock);
2172
2173 #ifdef CONFIG_PROC_FS
2174 /* Proc filesystem TCP sock list dumping. */
2175
2176 static inline struct tcp_tw_bucket *tw_head(struct hlist_head *head)
2177 {
2178         return hlist_empty(head) ? NULL :
2179                 list_entry(head->first, struct tcp_tw_bucket, tw_node);
2180 }
2181
2182 static inline struct tcp_tw_bucket *tw_next(struct tcp_tw_bucket *tw)
2183 {
2184         return tw->tw_node.next ?
2185                 hlist_entry(tw->tw_node.next, typeof(*tw), tw_node) : NULL;
2186 }
2187
2188 static void *listening_get_next(struct seq_file *seq, void *cur)
2189 {
2190         struct tcp_opt *tp;
2191         struct hlist_node *node;
2192         struct sock *sk = cur;
2193         struct tcp_iter_state* st = seq->private;
2194
2195         if (!sk) {
2196                 st->bucket = 0;
2197                 sk = sk_head(&tcp_listening_hash[0]);
2198                 goto get_sk;
2199         }
2200
2201         ++st->num;
2202
2203         if (st->state == TCP_SEQ_STATE_OPENREQ) {
2204                 struct open_request *req = cur;
2205
2206                 tp = tcp_sk(st->syn_wait_sk);
2207                 req = req->dl_next;
2208                 while (1) {
2209                         while (req) {
2210                                 vxdprintk(VXD_CBIT(net, 6),
2211                                         "sk,req: %p [#%d] (from %d)", req->sk,
2212                                         (req->sk)?req->sk->sk_xid:0, current->xid);
2213                                 if (req->sk &&
2214                                         !vx_check(req->sk->sk_xid, VX_IDENT|VX_WATCH))
2215                                         continue;
2216                                 if (req->class->family == st->family) {
2217                                         cur = req;
2218                                         goto out;
2219                                 }
2220                                 req = req->dl_next;
2221                         }
2222                         if (++st->sbucket >= TCP_SYNQ_HSIZE)
2223                                 break;
2224 get_req:
2225                         req = tp->listen_opt->syn_table[st->sbucket];
2226                 }
2227                 sk        = sk_next(st->syn_wait_sk);
2228                 st->state = TCP_SEQ_STATE_LISTENING;
2229                 read_unlock_bh(&tp->syn_wait_lock);
2230         } else {
2231                 tp = tcp_sk(sk);
2232                 read_lock_bh(&tp->syn_wait_lock);
2233                 if (tp->listen_opt && tp->listen_opt->qlen)
2234                         goto start_req;
2235                 read_unlock_bh(&tp->syn_wait_lock);
2236                 sk = sk_next(sk);
2237         }
2238 get_sk:
2239         sk_for_each_from(sk, node) {
2240                 vxdprintk(VXD_CBIT(net, 6), "sk: %p [#%d] (from %d)",
2241                         sk, sk->sk_xid, current->xid);
2242                 if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2243                         continue;
2244                 if (sk->sk_family == st->family) {
2245                         cur = sk;
2246                         goto out;
2247                 }
2248                 tp = tcp_sk(sk);
2249                 read_lock_bh(&tp->syn_wait_lock);
2250                 if (tp->listen_opt && tp->listen_opt->qlen) {
2251 start_req:
2252                         st->uid         = sock_i_uid(sk);
2253                         st->syn_wait_sk = sk;
2254                         st->state       = TCP_SEQ_STATE_OPENREQ;
2255                         st->sbucket     = 0;
2256                         goto get_req;
2257                 }
2258                 read_unlock_bh(&tp->syn_wait_lock);
2259         }
2260         if (++st->bucket < TCP_LHTABLE_SIZE) {
2261                 sk = sk_head(&tcp_listening_hash[st->bucket]);
2262                 goto get_sk;
2263         }
2264         cur = NULL;
2265 out:
2266         return cur;
2267 }
2268
2269 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2270 {
2271         void *rc = listening_get_next(seq, NULL);
2272
2273         while (rc && *pos) {
2274                 rc = listening_get_next(seq, rc);
2275                 --*pos;
2276         }
2277         return rc;
2278 }
2279
2280 static void *established_get_first(struct seq_file *seq)
2281 {
2282         struct tcp_iter_state* st = seq->private;
2283         void *rc = NULL;
2284
2285         for (st->bucket = 0; st->bucket < tcp_ehash_size; ++st->bucket) {
2286                 struct sock *sk;
2287                 struct hlist_node *node;
2288                 struct tcp_tw_bucket *tw;
2289
2290                 read_lock(&tcp_ehash[st->bucket].lock);
2291                 sk_for_each(sk, node, &tcp_ehash[st->bucket].chain) {
2292                         vxdprintk(VXD_CBIT(net, 6),
2293                                 "sk,egf: %p [#%d] (from %d)",
2294                                 sk, sk->sk_xid, current->xid);
2295                         if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2296                                 continue;
2297                         if (sk->sk_family != st->family)
2298                                 continue;
2299                         rc = sk;
2300                         goto out;
2301                 }
2302                 st->state = TCP_SEQ_STATE_TIME_WAIT;
2303                 tw_for_each(tw, node,
2304                             &tcp_ehash[st->bucket + tcp_ehash_size].chain) {
2305                         vxdprintk(VXD_CBIT(net, 6),
2306                                 "tw: %p [#%d] (from %d)",
2307                                 tw, tw->tw_xid, current->xid);
2308                         if (!vx_check(tw->tw_xid, VX_IDENT|VX_WATCH))
2309                                 continue;
2310                         if (tw->tw_family != st->family)
2311                                 continue;
2312                         rc = tw;
2313                         goto out;
2314                 }
2315                 read_unlock(&tcp_ehash[st->bucket].lock);
2316                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2317         }
2318 out:
2319         return rc;
2320 }
2321
2322 static void *established_get_next(struct seq_file *seq, void *cur)
2323 {
2324         struct sock *sk = cur;
2325         struct tcp_tw_bucket *tw;
2326         struct hlist_node *node;
2327         struct tcp_iter_state* st = seq->private;
2328
2329         ++st->num;
2330
2331         if (st->state == TCP_SEQ_STATE_TIME_WAIT) {
2332                 tw = cur;
2333                 tw = tw_next(tw);
2334 get_tw:
2335                 while (tw && (tw->tw_family != st->family ||
2336                         !vx_check(tw->tw_xid, VX_IDENT|VX_WATCH))) {
2337                         tw = tw_next(tw);
2338                 }
2339                 if (tw) {
2340                         cur = tw;
2341                         goto out;
2342                 }
2343                 read_unlock(&tcp_ehash[st->bucket].lock);
2344                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2345                 if (++st->bucket < tcp_ehash_size) {
2346                         read_lock(&tcp_ehash[st->bucket].lock);
2347                         sk = sk_head(&tcp_ehash[st->bucket].chain);
2348                 } else {
2349                         cur = NULL;
2350                         goto out;
2351                 }
2352         } else
2353                 sk = sk_next(sk);
2354
2355         sk_for_each_from(sk, node) {
2356                 vxdprintk(VXD_CBIT(net, 6),
2357                         "sk,egn: %p [#%d] (from %d)",
2358                         sk, sk->sk_xid, current->xid);
2359                 if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2360                         continue;
2361                 if (sk->sk_family == st->family)
2362                         goto found;
2363         }
2364
2365         st->state = TCP_SEQ_STATE_TIME_WAIT;
2366         tw = tw_head(&tcp_ehash[st->bucket + tcp_ehash_size].chain);
2367         goto get_tw;
2368 found:
2369         cur = sk;
2370 out:
2371         return cur;
2372 }
2373
2374 static void *established_get_idx(struct seq_file *seq, loff_t pos)
2375 {
2376         void *rc = established_get_first(seq);
2377
2378         while (rc && pos) {
2379                 rc = established_get_next(seq, rc);
2380                 --pos;
2381         }
2382         return rc;
2383 }
2384
2385 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2386 {
2387         void *rc;
2388         struct tcp_iter_state* st = seq->private;
2389
2390         tcp_listen_lock();
2391         st->state = TCP_SEQ_STATE_LISTENING;
2392         rc        = listening_get_idx(seq, &pos);
2393
2394         if (!rc) {
2395                 tcp_listen_unlock();
2396                 local_bh_disable();
2397                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2398                 rc        = established_get_idx(seq, pos);
2399         }
2400
2401         return rc;
2402 }
2403
2404 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2405 {
2406         struct tcp_iter_state* st = seq->private;
2407         st->state = TCP_SEQ_STATE_LISTENING;
2408         st->num = 0;
2409         return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2410 }
2411
2412 static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2413 {
2414         void *rc = NULL;
2415         struct tcp_iter_state* st;
2416
2417         if (v == SEQ_START_TOKEN) {
2418                 rc = tcp_get_idx(seq, 0);
2419                 goto out;
2420         }
2421         st = seq->private;
2422
2423         switch (st->state) {
2424         case TCP_SEQ_STATE_OPENREQ:
2425         case TCP_SEQ_STATE_LISTENING:
2426                 rc = listening_get_next(seq, v);
2427                 if (!rc) {
2428                         tcp_listen_unlock();
2429                         local_bh_disable();
2430                         st->state = TCP_SEQ_STATE_ESTABLISHED;
2431                         rc        = established_get_first(seq);
2432                 }
2433                 break;
2434         case TCP_SEQ_STATE_ESTABLISHED:
2435         case TCP_SEQ_STATE_TIME_WAIT:
2436                 rc = established_get_next(seq, v);
2437                 break;
2438         }
2439 out:
2440         ++*pos;
2441         return rc;
2442 }
2443
2444 static void tcp_seq_stop(struct seq_file *seq, void *v)
2445 {
2446         struct tcp_iter_state* st = seq->private;
2447
2448         switch (st->state) {
2449         case TCP_SEQ_STATE_OPENREQ:
2450                 if (v) {
2451                         struct tcp_opt *tp = tcp_sk(st->syn_wait_sk);
2452                         read_unlock_bh(&tp->syn_wait_lock);
2453                 }
2454         case TCP_SEQ_STATE_LISTENING:
2455                 if (v != SEQ_START_TOKEN)
2456                         tcp_listen_unlock();
2457                 break;
2458         case TCP_SEQ_STATE_TIME_WAIT:
2459         case TCP_SEQ_STATE_ESTABLISHED:
2460                 if (v)
2461                         read_unlock(&tcp_ehash[st->bucket].lock);
2462                 local_bh_enable();
2463                 break;
2464         }
2465 }
2466
2467 static int tcp_seq_open(struct inode *inode, struct file *file)
2468 {
2469         struct tcp_seq_afinfo *afinfo = PDE(inode)->data;
2470         struct seq_file *seq;
2471         struct tcp_iter_state *s;
2472         int rc;
2473
2474         if (unlikely(afinfo == NULL))
2475                 return -EINVAL;
2476
2477         s = kmalloc(sizeof(*s), GFP_KERNEL);
2478         if (!s)
2479                 return -ENOMEM;
2480         memset(s, 0, sizeof(*s));
2481         s->family               = afinfo->family;
2482         s->seq_ops.start        = tcp_seq_start;
2483         s->seq_ops.next         = tcp_seq_next;
2484         s->seq_ops.show         = afinfo->seq_show;
2485         s->seq_ops.stop         = tcp_seq_stop;
2486
2487         rc = seq_open(file, &s->seq_ops);
2488         if (rc)
2489                 goto out_kfree;
2490         seq          = file->private_data;
2491         seq->private = s;
2492 out:
2493         return rc;
2494 out_kfree:
2495         kfree(s);
2496         goto out;
2497 }
2498
2499 int tcp_proc_register(struct tcp_seq_afinfo *afinfo)
2500 {
2501         int rc = 0;
2502         struct proc_dir_entry *p;
2503
2504         if (!afinfo)
2505                 return -EINVAL;
2506         afinfo->seq_fops->owner         = afinfo->owner;
2507         afinfo->seq_fops->open          = tcp_seq_open;
2508         afinfo->seq_fops->read          = seq_read;
2509         afinfo->seq_fops->llseek        = seq_lseek;
2510         afinfo->seq_fops->release       = seq_release_private;
2511
2512         p = proc_net_fops_create(afinfo->name, S_IRUGO, afinfo->seq_fops);
2513         if (p)
2514                 p->data = afinfo;
2515         else
2516                 rc = -ENOMEM;
2517         return rc;
2518 }
2519
2520 void tcp_proc_unregister(struct tcp_seq_afinfo *afinfo)
2521 {
2522         if (!afinfo)
2523                 return;
2524         proc_net_remove(afinfo->name);
2525         memset(afinfo->seq_fops, 0, sizeof(*afinfo->seq_fops));
2526 }
2527
2528 static void get_openreq4(struct sock *sk, struct open_request *req,
2529                          char *tmpbuf, int i, int uid)
2530 {
2531         int ttd = req->expires - jiffies;
2532
2533         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
2534                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %u %d %p",
2535                 i,
2536                 req->af.v4_req.loc_addr,
2537                 ntohs(inet_sk(sk)->sport),
2538                 req->af.v4_req.rmt_addr,
2539                 ntohs(req->rmt_port),
2540                 TCP_SYN_RECV,
2541                 0, 0, /* could print option size, but that is af dependent. */
2542                 1,    /* timers active (only the expire timer) */
2543                 jiffies_to_clock_t(ttd),
2544                 req->retrans,
2545                 uid,
2546                 0,  /* non standard timer */
2547                 0, /* open_requests have no inode */
2548                 atomic_read(&sk->sk_refcnt),
2549                 req);
2550 }
2551
2552 static void get_tcp4_sock(struct sock *sp, char *tmpbuf, int i)
2553 {
2554         int timer_active;
2555         unsigned long timer_expires;
2556         struct tcp_opt *tp = tcp_sk(sp);
2557         struct inet_opt *inet = inet_sk(sp);
2558         unsigned int dest = inet->daddr;
2559         unsigned int src = inet->rcv_saddr;
2560         __u16 destp = ntohs(inet->dport);
2561         __u16 srcp = ntohs(inet->sport);
2562
2563         if (tp->pending == TCP_TIME_RETRANS) {
2564                 timer_active    = 1;
2565                 timer_expires   = tp->timeout;
2566         } else if (tp->pending == TCP_TIME_PROBE0) {
2567                 timer_active    = 4;
2568                 timer_expires   = tp->timeout;
2569         } else if (timer_pending(&sp->sk_timer)) {
2570                 timer_active    = 2;
2571                 timer_expires   = sp->sk_timer.expires;
2572         } else {
2573                 timer_active    = 0;
2574                 timer_expires = jiffies;
2575         }
2576
2577         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2578                         "%08X %5d %8d %lu %d %p %u %u %u %u %d",
2579                 i, src, srcp, dest, destp, sp->sk_state,
2580                 tp->write_seq - tp->snd_una, tp->rcv_nxt - tp->copied_seq,
2581                 timer_active,
2582                 jiffies_to_clock_t(timer_expires - jiffies),
2583                 tp->retransmits,
2584                 sock_i_uid(sp),
2585                 tp->probes_out,
2586                 sock_i_ino(sp),
2587                 atomic_read(&sp->sk_refcnt), sp,
2588                 tp->rto, tp->ack.ato, (tp->ack.quick << 1) | tp->ack.pingpong,
2589                 tp->snd_cwnd,
2590                 tp->snd_ssthresh >= 0xFFFF ? -1 : tp->snd_ssthresh);
2591 }
2592
2593 static void get_timewait4_sock(struct tcp_tw_bucket *tw, char *tmpbuf, int i)
2594 {
2595         unsigned int dest, src;
2596         __u16 destp, srcp;
2597         int ttd = tw->tw_ttd - jiffies;
2598
2599         if (ttd < 0)
2600                 ttd = 0;
2601
2602         dest  = tw->tw_daddr;
2603         src   = tw->tw_rcv_saddr;
2604         destp = ntohs(tw->tw_dport);
2605         srcp  = ntohs(tw->tw_sport);
2606
2607         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
2608                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %p",
2609                 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2610                 3, jiffies_to_clock_t(ttd), 0, 0, 0, 0,
2611                 atomic_read(&tw->tw_refcnt), tw);
2612 }
2613
2614 #define TMPSZ 150
2615
2616 static int tcp4_seq_show(struct seq_file *seq, void *v)
2617 {
2618         struct tcp_iter_state* st;
2619         char tmpbuf[TMPSZ + 1];
2620
2621         if (v == SEQ_START_TOKEN) {
2622                 seq_printf(seq, "%-*s\n", TMPSZ - 1,
2623                            "  sl  local_address rem_address   st tx_queue "
2624                            "rx_queue tr tm->when retrnsmt   uid  timeout "
2625                            "inode");
2626                 goto out;
2627         }
2628         st = seq->private;
2629
2630         switch (st->state) {
2631         case TCP_SEQ_STATE_LISTENING:
2632         case TCP_SEQ_STATE_ESTABLISHED:
2633                 get_tcp4_sock(v, tmpbuf, st->num);
2634                 break;
2635         case TCP_SEQ_STATE_OPENREQ:
2636                 get_openreq4(st->syn_wait_sk, v, tmpbuf, st->num, st->uid);
2637                 break;
2638         case TCP_SEQ_STATE_TIME_WAIT:
2639                 get_timewait4_sock(v, tmpbuf, st->num);
2640                 break;
2641         }
2642         seq_printf(seq, "%-*s\n", TMPSZ - 1, tmpbuf);
2643 out:
2644         return 0;
2645 }
2646
2647 static struct file_operations tcp4_seq_fops;
2648 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
2649         .owner          = THIS_MODULE,
2650         .name           = "tcp",
2651         .family         = AF_INET,
2652         .seq_show       = tcp4_seq_show,
2653         .seq_fops       = &tcp4_seq_fops,
2654 };
2655
2656 int __init tcp4_proc_init(void)
2657 {
2658         return tcp_proc_register(&tcp4_seq_afinfo);
2659 }
2660
2661 void tcp4_proc_exit(void)
2662 {
2663         tcp_proc_unregister(&tcp4_seq_afinfo);
2664 }
2665 #endif /* CONFIG_PROC_FS */
2666
2667 struct proto tcp_prot = {
2668         .name                   = "TCP",
2669         .close                  = tcp_close,
2670         .connect                = tcp_v4_connect,
2671         .disconnect             = tcp_disconnect,
2672         .accept                 = tcp_accept,
2673         .ioctl                  = tcp_ioctl,
2674         .init                   = tcp_v4_init_sock,
2675         .destroy                = tcp_v4_destroy_sock,
2676         .shutdown               = tcp_shutdown,
2677         .setsockopt             = tcp_setsockopt,
2678         .getsockopt             = tcp_getsockopt,
2679         .sendmsg                = tcp_sendmsg,
2680         .recvmsg                = tcp_recvmsg,
2681         .backlog_rcv            = tcp_v4_do_rcv,
2682         .hash                   = tcp_v4_hash,
2683         .unhash                 = tcp_unhash,
2684         .get_port               = tcp_v4_get_port,
2685         .enter_memory_pressure  = tcp_enter_memory_pressure,
2686         .sockets_allocated      = &tcp_sockets_allocated,
2687         .memory_allocated       = &tcp_memory_allocated,
2688         .memory_pressure        = &tcp_memory_pressure,
2689         .sysctl_mem             = sysctl_tcp_mem,
2690         .sysctl_wmem            = sysctl_tcp_wmem,
2691         .sysctl_rmem            = sysctl_tcp_rmem,
2692         .max_header             = MAX_TCP_HEADER,
2693         .slab_obj_size          = sizeof(struct tcp_sock),
2694 };
2695
2696
2697
2698 void __init tcp_v4_init(struct net_proto_family *ops)
2699 {
2700         int err = sock_create_kern(PF_INET, SOCK_RAW, IPPROTO_TCP, &tcp_socket);
2701         if (err < 0)
2702                 panic("Failed to create the TCP control socket.\n");
2703         tcp_socket->sk->sk_allocation   = GFP_ATOMIC;
2704         inet_sk(tcp_socket->sk)->uc_ttl = -1;
2705
2706         /* Unhash it so that IP input processing does not even
2707          * see it, we do not wish this socket to see incoming
2708          * packets.
2709          */
2710         tcp_socket->sk->sk_prot->unhash(tcp_socket->sk);
2711 }
2712
2713 EXPORT_SYMBOL(ipv4_specific);
2714 EXPORT_SYMBOL(tcp_bind_hash);
2715 EXPORT_SYMBOL(tcp_bucket_create);
2716 EXPORT_SYMBOL(tcp_hashinfo);
2717 EXPORT_SYMBOL(tcp_inherit_port);
2718 EXPORT_SYMBOL(tcp_listen_wlock);
2719 EXPORT_SYMBOL(tcp_port_rover);
2720 EXPORT_SYMBOL(tcp_prot);
2721 EXPORT_SYMBOL(tcp_put_port);
2722 EXPORT_SYMBOL(tcp_unhash);
2723 EXPORT_SYMBOL(tcp_v4_conn_request);
2724 EXPORT_SYMBOL(tcp_v4_connect);
2725 EXPORT_SYMBOL(tcp_v4_do_rcv);
2726 EXPORT_SYMBOL_GPL(tcp_v4_lookup_listener);
2727 EXPORT_SYMBOL(tcp_v4_rebuild_header);
2728 EXPORT_SYMBOL(tcp_v4_remember_stamp);
2729 EXPORT_SYMBOL(tcp_v4_send_check);
2730 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
2731
2732 #ifdef CONFIG_PROC_FS
2733 EXPORT_SYMBOL(tcp_proc_register);
2734 EXPORT_SYMBOL(tcp_proc_unregister);
2735 #endif
2736 #ifdef CONFIG_SYSCTL
2737 EXPORT_SYMBOL(sysctl_local_port_range);
2738 EXPORT_SYMBOL(sysctl_max_syn_backlog);
2739 EXPORT_SYMBOL(sysctl_tcp_low_latency);
2740 #endif