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                 /* This is deprecated, but if someone generated it,
1041                  * we have no reasons to ignore it.
1042                  */
1043                 if (!sock_owned_by_user(sk))
1044                         tcp_enter_cwr(tp);
1045                 goto out;
1046         case ICMP_PARAMETERPROB:
1047                 err = EPROTO;
1048                 break;
1049         case ICMP_DEST_UNREACH:
1050                 if (code > NR_ICMP_UNREACH)
1051                         goto out;
1052
1053                 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
1054                         if (!sock_owned_by_user(sk))
1055                                 do_pmtu_discovery(sk, iph, info);
1056                         goto out;
1057                 }
1058
1059                 err = icmp_err_convert[code].errno;
1060                 break;
1061         case ICMP_TIME_EXCEEDED:
1062                 err = EHOSTUNREACH;
1063                 break;
1064         default:
1065                 goto out;
1066         }
1067
1068         switch (sk->sk_state) {
1069                 struct open_request *req, **prev;
1070         case TCP_LISTEN:
1071                 if (sock_owned_by_user(sk))
1072                         goto out;
1073
1074                 req = tcp_v4_search_req(tp, &prev, th->dest,
1075                                         iph->daddr, iph->saddr);
1076                 if (!req)
1077                         goto out;
1078
1079                 /* ICMPs are not backlogged, hence we cannot get
1080                    an established socket here.
1081                  */
1082                 BUG_TRAP(!req->sk);
1083
1084                 if (seq != req->snt_isn) {
1085                         NET_INC_STATS_BH(LINUX_MIB_OUTOFWINDOWICMPS);
1086                         goto out;
1087                 }
1088
1089                 /*
1090                  * Still in SYN_RECV, just remove it silently.
1091                  * There is no good way to pass the error to the newly
1092                  * created socket, and POSIX does not want network
1093                  * errors returned from accept().
1094                  */
1095                 tcp_synq_drop(sk, req, prev);
1096                 goto out;
1097
1098         case TCP_SYN_SENT:
1099         case TCP_SYN_RECV:  /* Cannot happen.
1100                                It can f.e. if SYNs crossed.
1101                              */
1102                 if (!sock_owned_by_user(sk)) {
1103                         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
1104                         sk->sk_err = err;
1105
1106                         sk->sk_error_report(sk);
1107
1108                         tcp_done(sk);
1109                 } else {
1110                         sk->sk_err_soft = err;
1111                 }
1112                 goto out;
1113         }
1114
1115         /* If we've already connected we will keep trying
1116          * until we time out, or the user gives up.
1117          *
1118          * rfc1122 4.2.3.9 allows to consider as hard errors
1119          * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
1120          * but it is obsoleted by pmtu discovery).
1121          *
1122          * Note, that in modern internet, where routing is unreliable
1123          * and in each dark corner broken firewalls sit, sending random
1124          * errors ordered by their masters even this two messages finally lose
1125          * their original sense (even Linux sends invalid PORT_UNREACHs)
1126          *
1127          * Now we are in compliance with RFCs.
1128          *                                                      --ANK (980905)
1129          */
1130
1131         inet = inet_sk(sk);
1132         if (!sock_owned_by_user(sk) && inet->recverr) {
1133                 sk->sk_err = err;
1134                 sk->sk_error_report(sk);
1135         } else  { /* Only an error on timeout */
1136                 sk->sk_err_soft = err;
1137         }
1138
1139 out:
1140         bh_unlock_sock(sk);
1141         sock_put(sk);
1142 }
1143
1144 /* This routine computes an IPv4 TCP checksum. */
1145 void tcp_v4_send_check(struct sock *sk, struct tcphdr *th, int len,
1146                        struct sk_buff *skb)
1147 {
1148         struct inet_opt *inet = inet_sk(sk);
1149
1150         if (skb->ip_summed == CHECKSUM_HW) {
1151                 th->check = ~tcp_v4_check(th, len, inet->saddr, inet->daddr, 0);
1152                 skb->csum = offsetof(struct tcphdr, check);
1153         } else {
1154                 th->check = tcp_v4_check(th, len, inet->saddr, inet->daddr,
1155                                          csum_partial((char *)th,
1156                                                       th->doff << 2,
1157                                                       skb->csum));
1158         }
1159 }
1160
1161 /*
1162  *      This routine will send an RST to the other tcp.
1163  *
1164  *      Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
1165  *                    for reset.
1166  *      Answer: if a packet caused RST, it is not for a socket
1167  *              existing in our system, if it is matched to a socket,
1168  *              it is just duplicate segment or bug in other side's TCP.
1169  *              So that we build reply only basing on parameters
1170  *              arrived with segment.
1171  *      Exception: precedence violation. We do not implement it in any case.
1172  */
1173
1174 static void tcp_v4_send_reset(struct sk_buff *skb)
1175 {
1176         struct tcphdr *th = skb->h.th;
1177         struct tcphdr rth;
1178         struct ip_reply_arg arg;
1179
1180         /* Never send a reset in response to a reset. */
1181         if (th->rst)
1182                 return;
1183
1184         if (((struct rtable *)skb->dst)->rt_type != RTN_LOCAL)
1185                 return;
1186
1187         /* Swap the send and the receive. */
1188         memset(&rth, 0, sizeof(struct tcphdr));
1189         rth.dest   = th->source;
1190         rth.source = th->dest;
1191         rth.doff   = sizeof(struct tcphdr) / 4;
1192         rth.rst    = 1;
1193
1194         if (th->ack) {
1195                 rth.seq = th->ack_seq;
1196         } else {
1197                 rth.ack = 1;
1198                 rth.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
1199                                     skb->len - (th->doff << 2));
1200         }
1201
1202         memset(&arg, 0, sizeof arg);
1203         arg.iov[0].iov_base = (unsigned char *)&rth;
1204         arg.iov[0].iov_len  = sizeof rth;
1205         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
1206                                       skb->nh.iph->saddr, /*XXX*/
1207                                       sizeof(struct tcphdr), IPPROTO_TCP, 0);
1208         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1209
1210         ip_send_reply(tcp_socket->sk, skb, &arg, sizeof rth);
1211
1212         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
1213         TCP_INC_STATS_BH(TCP_MIB_OUTRSTS);
1214 }
1215
1216 /* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
1217    outside socket context is ugly, certainly. What can I do?
1218  */
1219
1220 static void tcp_v4_send_ack(struct sk_buff *skb, u32 seq, u32 ack,
1221                             u32 win, u32 ts)
1222 {
1223         struct tcphdr *th = skb->h.th;
1224         struct {
1225                 struct tcphdr th;
1226                 u32 tsopt[3];
1227         } rep;
1228         struct ip_reply_arg arg;
1229
1230         memset(&rep.th, 0, sizeof(struct tcphdr));
1231         memset(&arg, 0, sizeof arg);
1232
1233         arg.iov[0].iov_base = (unsigned char *)&rep;
1234         arg.iov[0].iov_len  = sizeof(rep.th);
1235         if (ts) {
1236                 rep.tsopt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1237                                      (TCPOPT_TIMESTAMP << 8) |
1238                                      TCPOLEN_TIMESTAMP);
1239                 rep.tsopt[1] = htonl(tcp_time_stamp);
1240                 rep.tsopt[2] = htonl(ts);
1241                 arg.iov[0].iov_len = sizeof(rep);
1242         }
1243
1244         /* Swap the send and the receive. */
1245         rep.th.dest    = th->source;
1246         rep.th.source  = th->dest;
1247         rep.th.doff    = arg.iov[0].iov_len / 4;
1248         rep.th.seq     = htonl(seq);
1249         rep.th.ack_seq = htonl(ack);
1250         rep.th.ack     = 1;
1251         rep.th.window  = htons(win);
1252
1253         arg.csum = csum_tcpudp_nofold(skb->nh.iph->daddr,
1254                                       skb->nh.iph->saddr, /*XXX*/
1255                                       arg.iov[0].iov_len, IPPROTO_TCP, 0);
1256         arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1257
1258         ip_send_reply(tcp_socket->sk, skb, &arg, arg.iov[0].iov_len);
1259
1260         TCP_INC_STATS_BH(TCP_MIB_OUTSEGS);
1261 }
1262
1263 static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
1264 {
1265         struct tcp_tw_bucket *tw = (struct tcp_tw_bucket *)sk;
1266
1267         tcp_v4_send_ack(skb, tw->tw_snd_nxt, tw->tw_rcv_nxt,
1268                         tw->tw_rcv_wnd >> tw->tw_rcv_wscale, tw->tw_ts_recent);
1269
1270         tcp_tw_put(tw);
1271 }
1272
1273 static void tcp_v4_or_send_ack(struct sk_buff *skb, struct open_request *req)
1274 {
1275         tcp_v4_send_ack(skb, req->snt_isn + 1, req->rcv_isn + 1, req->rcv_wnd,
1276                         req->ts_recent);
1277 }
1278
1279 static struct dst_entry* tcp_v4_route_req(struct sock *sk,
1280                                           struct open_request *req)
1281 {
1282         struct rtable *rt;
1283         struct ip_options *opt = req->af.v4_req.opt;
1284         struct flowi fl = { .oif = sk->sk_bound_dev_if,
1285                             .nl_u = { .ip4_u =
1286                                       { .daddr = ((opt && opt->srr) ?
1287                                                   opt->faddr :
1288                                                   req->af.v4_req.rmt_addr),
1289                                         .saddr = req->af.v4_req.loc_addr,
1290                                         .tos = RT_CONN_FLAGS(sk) } },
1291                             .proto = IPPROTO_TCP,
1292                             .uli_u = { .ports =
1293                                        { .sport = inet_sk(sk)->sport,
1294                                          .dport = req->rmt_port } } };
1295
1296         if (ip_route_output_flow(&rt, &fl, sk, 0)) {
1297                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
1298                 return NULL;
1299         }
1300         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway) {
1301                 ip_rt_put(rt);
1302                 IP_INC_STATS_BH(IPSTATS_MIB_OUTNOROUTES);
1303                 return NULL;
1304         }
1305         return &rt->u.dst;
1306 }
1307
1308 /*
1309  *      Send a SYN-ACK after having received an ACK.
1310  *      This still operates on a open_request only, not on a big
1311  *      socket.
1312  */
1313 static int tcp_v4_send_synack(struct sock *sk, struct open_request *req,
1314                               struct dst_entry *dst)
1315 {
1316         int err = -1;
1317         struct sk_buff * skb;
1318
1319         /* First, grab a route. */
1320         if (!dst && (dst = tcp_v4_route_req(sk, req)) == NULL)
1321                 goto out;
1322
1323         skb = tcp_make_synack(sk, dst, req);
1324
1325         if (skb) {
1326                 struct tcphdr *th = skb->h.th;
1327
1328                 th->check = tcp_v4_check(th, skb->len,
1329                                          req->af.v4_req.loc_addr,
1330                                          req->af.v4_req.rmt_addr,
1331                                          csum_partial((char *)th, skb->len,
1332                                                       skb->csum));
1333
1334                 err = ip_build_and_send_pkt(skb, sk, req->af.v4_req.loc_addr,
1335                                             req->af.v4_req.rmt_addr,
1336                                             req->af.v4_req.opt);
1337                 if (err == NET_XMIT_CN)
1338                         err = 0;
1339         }
1340
1341 out:
1342         dst_release(dst);
1343         return err;
1344 }
1345
1346 /*
1347  *      IPv4 open_request destructor.
1348  */
1349 static void tcp_v4_or_free(struct open_request *req)
1350 {
1351         if (req->af.v4_req.opt)
1352                 kfree(req->af.v4_req.opt);
1353 }
1354
1355 static inline void syn_flood_warning(struct sk_buff *skb)
1356 {
1357         static unsigned long warntime;
1358
1359         if (time_after(jiffies, (warntime + HZ * 60))) {
1360                 warntime = jiffies;
1361                 printk(KERN_INFO
1362                        "possible SYN flooding on port %d. Sending cookies.\n",
1363                        ntohs(skb->h.th->dest));
1364         }
1365 }
1366
1367 /*
1368  * Save and compile IPv4 options into the open_request if needed.
1369  */
1370 static inline struct ip_options *tcp_v4_save_options(struct sock *sk,
1371                                                      struct sk_buff *skb)
1372 {
1373         struct ip_options *opt = &(IPCB(skb)->opt);
1374         struct ip_options *dopt = NULL;
1375
1376         if (opt && opt->optlen) {
1377                 int opt_size = optlength(opt);
1378                 dopt = kmalloc(opt_size, GFP_ATOMIC);
1379                 if (dopt) {
1380                         if (ip_options_echo(dopt, skb)) {
1381                                 kfree(dopt);
1382                                 dopt = NULL;
1383                         }
1384                 }
1385         }
1386         return dopt;
1387 }
1388
1389 /*
1390  * Maximum number of SYN_RECV sockets in queue per LISTEN socket.
1391  * One SYN_RECV socket costs about 80bytes on a 32bit machine.
1392  * It would be better to replace it with a global counter for all sockets
1393  * but then some measure against one socket starving all other sockets
1394  * would be needed.
1395  *
1396  * It was 128 by default. Experiments with real servers show, that
1397  * it is absolutely not enough even at 100conn/sec. 256 cures most
1398  * of problems. This value is adjusted to 128 for very small machines
1399  * (<=32Mb of memory) and to 1024 on normal or better ones (>=256Mb).
1400  * Further increasing requires to change hash table size.
1401  */
1402 int sysctl_max_syn_backlog = 256;
1403
1404 struct or_calltable or_ipv4 = {
1405         .family         =       PF_INET,
1406         .rtx_syn_ack    =       tcp_v4_send_synack,
1407         .send_ack       =       tcp_v4_or_send_ack,
1408         .destructor     =       tcp_v4_or_free,
1409         .send_reset     =       tcp_v4_send_reset,
1410 };
1411
1412 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1413 {
1414         struct tcp_opt tp;
1415         struct open_request *req;
1416         __u32 saddr = skb->nh.iph->saddr;
1417         __u32 daddr = skb->nh.iph->daddr;
1418         __u32 isn = TCP_SKB_CB(skb)->when;
1419         struct dst_entry *dst = NULL;
1420 #ifdef CONFIG_ACCEPT_QUEUES
1421         int class = 0;
1422 #endif
1423 #ifdef CONFIG_SYN_COOKIES
1424         int want_cookie = 0;
1425 #else
1426 #define want_cookie 0 /* Argh, why doesn't gcc optimize this :( */
1427 #endif
1428
1429         /* Never answer to SYNs send to broadcast or multicast */
1430         if (((struct rtable *)skb->dst)->rt_flags &
1431             (RTCF_BROADCAST | RTCF_MULTICAST))
1432                 goto drop;
1433
1434         /* TW buckets are converted to open requests without
1435          * limitations, they conserve resources and peer is
1436          * evidently real one.
1437          */
1438         if (tcp_synq_is_full(sk) && !isn) {
1439 #ifdef CONFIG_SYN_COOKIES
1440                 if (sysctl_tcp_syncookies) {
1441                         want_cookie = 1;
1442                 } else
1443 #endif
1444                 goto drop;
1445         }
1446
1447 #ifdef CONFIG_ACCEPT_QUEUES
1448         class = (skb->nfmark <= 0) ? 0 :
1449                 ((skb->nfmark >= NUM_ACCEPT_QUEUES) ? 0: skb->nfmark);
1450         /*
1451          * Accept only if the class has shares set or if the default class
1452          * i.e. class 0 has shares
1453          */
1454         if (!(tcp_sk(sk)->acceptq[class].aq_ratio)) {
1455                 if (tcp_sk(sk)->acceptq[0].aq_ratio)
1456                         class = 0;
1457                 else
1458                         goto drop;
1459         }
1460 #endif
1461
1462         /* Accept backlog is full. If we have already queued enough
1463          * of warm entries in syn queue, drop request. It is better than
1464          * clogging syn queue with openreqs with exponentially increasing
1465          * timeout.
1466          */
1467 #ifdef CONFIG_ACCEPT_QUEUES
1468         if (sk_acceptq_is_full(sk, class) && tcp_synq_young(sk, class) > 1)
1469 #else
1470         if (sk_acceptq_is_full(sk) && tcp_synq_young(sk) > 1)
1471 #endif
1472                 goto drop;
1473
1474         req = tcp_openreq_alloc();
1475         if (!req)
1476                 goto drop;
1477
1478         tcp_clear_options(&tp);
1479         tp.mss_clamp = 536;
1480         tp.user_mss  = tcp_sk(sk)->user_mss;
1481
1482         tcp_parse_options(skb, &tp, 0);
1483
1484         if (want_cookie) {
1485                 tcp_clear_options(&tp);
1486                 tp.saw_tstamp = 0;
1487         }
1488
1489         if (tp.saw_tstamp && !tp.rcv_tsval) {
1490                 /* Some OSes (unknown ones, but I see them on web server, which
1491                  * contains information interesting only for windows'
1492                  * users) do not send their stamp in SYN. It is easy case.
1493                  * We simply do not advertise TS support.
1494                  */
1495                 tp.saw_tstamp = 0;
1496                 tp.tstamp_ok  = 0;
1497         }
1498         tp.tstamp_ok = tp.saw_tstamp;
1499
1500         tcp_openreq_init(req, &tp, skb);
1501 #ifdef CONFIG_ACCEPT_QUEUES
1502         req->acceptq_class = class;
1503         req->acceptq_time_stamp = jiffies;
1504 #endif
1505         req->af.v4_req.loc_addr = daddr;
1506         req->af.v4_req.rmt_addr = saddr;
1507         req->af.v4_req.opt = tcp_v4_save_options(sk, skb);
1508         req->class = &or_ipv4;
1509         if (!want_cookie)
1510                 TCP_ECN_create_request(req, skb->h.th);
1511
1512         if (want_cookie) {
1513 #ifdef CONFIG_SYN_COOKIES
1514                 syn_flood_warning(skb);
1515 #endif
1516                 isn = cookie_v4_init_sequence(sk, skb, &req->mss);
1517         } else if (!isn) {
1518                 struct inet_peer *peer = NULL;
1519
1520                 /* VJ's idea. We save last timestamp seen
1521                  * from the destination in peer table, when entering
1522                  * state TIME-WAIT, and check against it before
1523                  * accepting new connection request.
1524                  *
1525                  * If "isn" is not zero, this request hit alive
1526                  * timewait bucket, so that all the necessary checks
1527                  * are made in the function processing timewait state.
1528                  */
1529                 if (tp.saw_tstamp &&
1530                     sysctl_tcp_tw_recycle &&
1531                     (dst = tcp_v4_route_req(sk, req)) != NULL &&
1532                     (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
1533                     peer->v4daddr == saddr) {
1534                         if (xtime.tv_sec < peer->tcp_ts_stamp + TCP_PAWS_MSL &&
1535                             (s32)(peer->tcp_ts - req->ts_recent) >
1536                                                         TCP_PAWS_WINDOW) {
1537                                 NET_INC_STATS_BH(LINUX_MIB_PAWSPASSIVEREJECTED);
1538                                 dst_release(dst);
1539                                 goto drop_and_free;
1540                         }
1541                 }
1542                 /* Kill the following clause, if you dislike this way. */
1543                 else if (!sysctl_tcp_syncookies &&
1544                          (sysctl_max_syn_backlog - tcp_synq_len(sk) <
1545                           (sysctl_max_syn_backlog >> 2)) &&
1546                          (!peer || !peer->tcp_ts_stamp) &&
1547                          (!dst || !dst_metric(dst, RTAX_RTT))) {
1548                         /* Without syncookies last quarter of
1549                          * backlog is filled with destinations,
1550                          * proven to be alive.
1551                          * It means that we continue to communicate
1552                          * to destinations, already remembered
1553                          * to the moment of synflood.
1554                          */
1555                         NETDEBUG(if (net_ratelimit()) \
1556                                         printk(KERN_DEBUG "TCP: drop open "
1557                                                           "request from %u.%u."
1558                                                           "%u.%u/%u\n", \
1559                                                NIPQUAD(saddr),
1560                                                ntohs(skb->h.th->source)));
1561                         dst_release(dst);
1562                         goto drop_and_free;
1563                 }
1564
1565                 isn = tcp_v4_init_sequence(sk, skb);
1566         }
1567         req->snt_isn = isn;
1568
1569         if (tcp_v4_send_synack(sk, req, dst))
1570                 goto drop_and_free;
1571
1572         if (want_cookie) {
1573                 tcp_openreq_free(req);
1574         } else {
1575                 tcp_v4_synq_add(sk, req);
1576         }
1577         return 0;
1578
1579 drop_and_free:
1580         tcp_openreq_free(req);
1581 drop:
1582         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
1583         return 0;
1584 }
1585
1586
1587 /*
1588  * The three way handshake has completed - we got a valid synack -
1589  * now create the new socket.
1590  */
1591 struct sock *tcp_v4_syn_recv_sock(struct sock *sk, struct sk_buff *skb,
1592                                   struct open_request *req,
1593                                   struct dst_entry *dst)
1594 {
1595         struct inet_opt *newinet;
1596         struct tcp_opt *newtp;
1597         struct sock *newsk;
1598
1599 #ifdef CONFIG_ACCEPT_QUEUES
1600         if (sk_acceptq_is_full(sk, req->acceptq_class))
1601 #else
1602         if (sk_acceptq_is_full(sk))
1603 #endif
1604                 goto exit_overflow;
1605
1606         if (!dst && (dst = tcp_v4_route_req(sk, req)) == NULL)
1607                 goto exit;
1608
1609         newsk = tcp_create_openreq_child(sk, req, skb);
1610         if (!newsk)
1611                 goto exit;
1612
1613         newsk->sk_dst_cache = dst;
1614         tcp_v4_setup_caps(newsk, dst);
1615
1616         newtp                 = tcp_sk(newsk);
1617         newinet               = inet_sk(newsk);
1618         newinet->daddr        = req->af.v4_req.rmt_addr;
1619         newinet->rcv_saddr    = req->af.v4_req.loc_addr;
1620         newinet->saddr        = req->af.v4_req.loc_addr;
1621         newinet->opt          = req->af.v4_req.opt;
1622         req->af.v4_req.opt    = NULL;
1623         newinet->mc_index     = tcp_v4_iif(skb);
1624         newinet->mc_ttl       = skb->nh.iph->ttl;
1625         newtp->ext_header_len = 0;
1626         if (newinet->opt)
1627                 newtp->ext_header_len = newinet->opt->optlen;
1628         newtp->ext2_header_len = dst->header_len;
1629         newinet->id = newtp->write_seq ^ jiffies;
1630
1631         tcp_sync_mss(newsk, dst_pmtu(dst));
1632         newtp->advmss = dst_metric(dst, RTAX_ADVMSS);
1633         tcp_initialize_rcv_mss(newsk);
1634
1635         __tcp_v4_hash(newsk, 0);
1636         __tcp_inherit_port(sk, newsk);
1637
1638         return newsk;
1639
1640 exit_overflow:
1641         NET_INC_STATS_BH(LINUX_MIB_LISTENOVERFLOWS);
1642 exit:
1643         NET_INC_STATS_BH(LINUX_MIB_LISTENDROPS);
1644         dst_release(dst);
1645         return NULL;
1646 }
1647
1648 static struct sock *tcp_v4_hnd_req(struct sock *sk, struct sk_buff *skb)
1649 {
1650         struct tcphdr *th = skb->h.th;
1651         struct iphdr *iph = skb->nh.iph;
1652         struct tcp_opt *tp = tcp_sk(sk);
1653         struct sock *nsk;
1654         struct open_request **prev;
1655         /* Find possible connection requests. */
1656         struct open_request *req = tcp_v4_search_req(tp, &prev, th->source,
1657                                                      iph->saddr, iph->daddr);
1658         if (req)
1659                 return tcp_check_req(sk, skb, req, prev);
1660
1661         nsk = __tcp_v4_lookup_established(skb->nh.iph->saddr,
1662                                           th->source,
1663                                           skb->nh.iph->daddr,
1664                                           ntohs(th->dest),
1665                                           tcp_v4_iif(skb));
1666
1667         if (nsk) {
1668                 if (nsk->sk_state != TCP_TIME_WAIT) {
1669                         bh_lock_sock(nsk);
1670                         return nsk;
1671                 }
1672                 tcp_tw_put((struct tcp_tw_bucket *)nsk);
1673                 return NULL;
1674         }
1675
1676 #ifdef CONFIG_SYN_COOKIES
1677         if (!th->rst && !th->syn && th->ack)
1678                 sk = cookie_v4_check(sk, skb, &(IPCB(skb)->opt));
1679 #endif
1680         return sk;
1681 }
1682
1683 static int tcp_v4_checksum_init(struct sk_buff *skb)
1684 {
1685         if (skb->ip_summed == CHECKSUM_HW) {
1686                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1687                 if (!tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
1688                                   skb->nh.iph->daddr, skb->csum))
1689                         return 0;
1690
1691                 NETDEBUG(if (net_ratelimit())
1692                                 printk(KERN_DEBUG "hw tcp v4 csum failed\n"));
1693                 skb->ip_summed = CHECKSUM_NONE;
1694         }
1695         if (skb->len <= 76) {
1696                 if (tcp_v4_check(skb->h.th, skb->len, skb->nh.iph->saddr,
1697                                  skb->nh.iph->daddr,
1698                                  skb_checksum(skb, 0, skb->len, 0)))
1699                         return -1;
1700                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1701         } else {
1702                 skb->csum = ~tcp_v4_check(skb->h.th, skb->len,
1703                                           skb->nh.iph->saddr,
1704                                           skb->nh.iph->daddr, 0);
1705         }
1706         return 0;
1707 }
1708
1709
1710 /* The socket must have it's spinlock held when we get
1711  * here.
1712  *
1713  * We have a potential double-lock case here, so even when
1714  * doing backlog processing we use the BH locking scheme.
1715  * This is because we cannot sleep with the original spinlock
1716  * held.
1717  */
1718 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1719 {
1720         if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1721                 TCP_CHECK_TIMER(sk);
1722                 if (tcp_rcv_established(sk, skb, skb->h.th, skb->len))
1723                         goto reset;
1724                 TCP_CHECK_TIMER(sk);
1725                 return 0;
1726         }
1727
1728         if (skb->len < (skb->h.th->doff << 2) || tcp_checksum_complete(skb))
1729                 goto csum_err;
1730
1731         if (sk->sk_state == TCP_LISTEN) {
1732                 struct sock *nsk = tcp_v4_hnd_req(sk, skb);
1733                 if (!nsk)
1734                         goto discard;
1735
1736                 if (nsk != sk) {
1737                         if (tcp_child_process(sk, nsk, skb))
1738                                 goto reset;
1739                         return 0;
1740                 }
1741         }
1742
1743         TCP_CHECK_TIMER(sk);
1744         if (tcp_rcv_state_process(sk, skb, skb->h.th, skb->len))
1745                 goto reset;
1746         TCP_CHECK_TIMER(sk);
1747         return 0;
1748
1749 reset:
1750         tcp_v4_send_reset(skb);
1751 discard:
1752         kfree_skb(skb);
1753         /* Be careful here. If this function gets more complicated and
1754          * gcc suffers from register pressure on the x86, sk (in %ebx)
1755          * might be destroyed here. This current version compiles correctly,
1756          * but you have been warned.
1757          */
1758         return 0;
1759
1760 csum_err:
1761         TCP_INC_STATS_BH(TCP_MIB_INERRS);
1762         goto discard;
1763 }
1764
1765 extern struct proto_ops inet_stream_ops;
1766
1767 extern int inet_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len);
1768
1769 /*
1770  *      From tcp_input.c
1771  */
1772
1773 int tcp_v4_rcv(struct sk_buff *skb)
1774 {
1775         struct tcphdr *th;
1776         struct sock *sk;
1777         int ret;
1778
1779         if (skb->pkt_type != PACKET_HOST)
1780                 goto discard_it;
1781
1782         /* Count it even if it's bad */
1783         TCP_INC_STATS_BH(TCP_MIB_INSEGS);
1784
1785         if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1786                 goto discard_it;
1787
1788         th = skb->h.th;
1789
1790         if (th->doff < sizeof(struct tcphdr) / 4)
1791                 goto bad_packet;
1792         if (!pskb_may_pull(skb, th->doff * 4))
1793                 goto discard_it;
1794
1795         /* An explanation is required here, I think.
1796          * Packet length and doff are validated by header prediction,
1797          * provided case of th->doff==0 is elimineted.
1798          * So, we defer the checks. */
1799         if ((skb->ip_summed != CHECKSUM_UNNECESSARY &&
1800              tcp_v4_checksum_init(skb) < 0))
1801                 goto bad_packet;
1802
1803         th = skb->h.th;
1804         TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1805         TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1806                                     skb->len - th->doff * 4);
1807         TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1808         TCP_SKB_CB(skb)->when    = 0;
1809         TCP_SKB_CB(skb)->flags   = skb->nh.iph->tos;
1810         TCP_SKB_CB(skb)->sacked  = 0;
1811
1812         sk = __tcp_v4_lookup(skb->nh.iph->saddr, th->source,
1813                              skb->nh.iph->daddr, ntohs(th->dest),
1814                              tcp_v4_iif(skb));
1815
1816         if (!sk)
1817                 goto no_tcp_socket;
1818
1819 process:
1820         /* Silently drop if VNET is active (if INET bind() has been
1821          * overridden) and the context is not entitled to read the
1822          * packet.
1823          */
1824         if (inet_stream_ops.bind != inet_bind &&
1825             (int) sk->sk_xid > 0 && sk->sk_xid != skb->xid)
1826                 goto discard_it;
1827
1828         if (sk->sk_state == TCP_TIME_WAIT)
1829                 goto do_time_wait;
1830
1831         if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1832                 goto discard_and_relse;
1833
1834         if (sk_filter(sk, skb, 0))
1835                 goto discard_and_relse;
1836
1837         skb->dev = NULL;
1838
1839         bh_lock_sock(sk);
1840         ret = 0;
1841         if (!sock_owned_by_user(sk)) {
1842                 if (!tcp_prequeue(sk, skb))
1843                         ret = tcp_v4_do_rcv(sk, skb);
1844         } else
1845                 sk_add_backlog(sk, skb);
1846         bh_unlock_sock(sk);
1847
1848         sock_put(sk);
1849
1850         return ret;
1851
1852 no_tcp_socket:
1853         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1854                 goto discard_it;
1855
1856         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1857 bad_packet:
1858                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1859         } else if (!skb->sk) {
1860                 /* VNET: Suppress RST if the port was bound to a (presumably raw) socket */
1861                 tcp_v4_send_reset(skb);
1862         }
1863
1864 discard_it:
1865         /* Discard frame. */
1866         kfree_skb(skb);
1867         return 0;
1868
1869 discard_and_relse:
1870         sock_put(sk);
1871         goto discard_it;
1872
1873 do_time_wait:
1874         if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1875                 tcp_tw_put((struct tcp_tw_bucket *) sk);
1876                 goto discard_it;
1877         }
1878
1879         if (skb->len < (th->doff << 2) || tcp_checksum_complete(skb)) {
1880                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
1881                 tcp_tw_put((struct tcp_tw_bucket *) sk);
1882                 goto discard_it;
1883         }
1884         switch (tcp_timewait_state_process((struct tcp_tw_bucket *)sk,
1885                                            skb, th, skb->len)) {
1886         case TCP_TW_SYN: {
1887                 struct sock *sk2 = tcp_v4_lookup_listener(skb->nh.iph->daddr,
1888                                                           ntohs(th->dest),
1889                                                           tcp_v4_iif(skb));
1890                 if (sk2) {
1891                         tcp_tw_deschedule((struct tcp_tw_bucket *)sk);
1892                         tcp_tw_put((struct tcp_tw_bucket *)sk);
1893                         sk = sk2;
1894                         goto process;
1895                 }
1896                 /* Fall through to ACK */
1897         }
1898         case TCP_TW_ACK:
1899                 tcp_v4_timewait_ack(sk, skb);
1900                 break;
1901         case TCP_TW_RST:
1902                 goto no_tcp_socket;
1903         case TCP_TW_SUCCESS:;
1904         }
1905         goto discard_it;
1906 }
1907
1908 /* With per-bucket locks this operation is not-atomic, so that
1909  * this version is not worse.
1910  */
1911 static void __tcp_v4_rehash(struct sock *sk)
1912 {
1913         sk->sk_prot->unhash(sk);
1914         sk->sk_prot->hash(sk);
1915 }
1916
1917 static int tcp_v4_reselect_saddr(struct sock *sk)
1918 {
1919         struct inet_opt *inet = inet_sk(sk);
1920         int err;
1921         struct rtable *rt;
1922         __u32 old_saddr = inet->saddr;
1923         __u32 new_saddr;
1924         __u32 daddr = inet->daddr;
1925
1926         if (inet->opt && inet->opt->srr)
1927                 daddr = inet->opt->faddr;
1928
1929         /* Query new route. */
1930         err = ip_route_connect(&rt, daddr, 0,
1931                                RT_TOS(inet->tos) | sk->sk_localroute,
1932                                sk->sk_bound_dev_if,
1933                                IPPROTO_TCP,
1934                                inet->sport, inet->dport, sk);
1935         if (err)
1936                 return err;
1937
1938         __sk_dst_set(sk, &rt->u.dst);
1939         tcp_v4_setup_caps(sk, &rt->u.dst);
1940         tcp_sk(sk)->ext2_header_len = rt->u.dst.header_len;
1941
1942         new_saddr = rt->rt_src;
1943
1944         if (new_saddr == old_saddr)
1945                 return 0;
1946
1947         if (sysctl_ip_dynaddr > 1) {
1948                 printk(KERN_INFO "tcp_v4_rebuild_header(): shifting inet->"
1949                                  "saddr from %d.%d.%d.%d to %d.%d.%d.%d\n",
1950                        NIPQUAD(old_saddr),
1951                        NIPQUAD(new_saddr));
1952         }
1953
1954         inet->saddr = new_saddr;
1955         inet->rcv_saddr = new_saddr;
1956
1957         /* XXX The only one ugly spot where we need to
1958          * XXX really change the sockets identity after
1959          * XXX it has entered the hashes. -DaveM
1960          *
1961          * Besides that, it does not check for connection
1962          * uniqueness. Wait for troubles.
1963          */
1964         __tcp_v4_rehash(sk);
1965         return 0;
1966 }
1967
1968 int tcp_v4_rebuild_header(struct sock *sk)
1969 {
1970         struct inet_opt *inet = inet_sk(sk);
1971         struct rtable *rt = (struct rtable *)__sk_dst_check(sk, 0);
1972         u32 daddr;
1973         int err;
1974
1975         /* Route is OK, nothing to do. */
1976         if (rt)
1977                 return 0;
1978
1979         /* Reroute. */
1980         daddr = inet->daddr;
1981         if (inet->opt && inet->opt->srr)
1982                 daddr = inet->opt->faddr;
1983
1984         {
1985                 struct flowi fl = { .oif = sk->sk_bound_dev_if,
1986                                     .nl_u = { .ip4_u =
1987                                               { .daddr = daddr,
1988                                                 .saddr = inet->saddr,
1989                                                 .tos = RT_CONN_FLAGS(sk) } },
1990                                     .proto = IPPROTO_TCP,
1991                                     .uli_u = { .ports =
1992                                                { .sport = inet->sport,
1993                                                  .dport = inet->dport } } };
1994
1995                 err = ip_route_output_flow(&rt, &fl, sk, 0);
1996         }
1997         if (!err) {
1998                 __sk_dst_set(sk, &rt->u.dst);
1999                 tcp_v4_setup_caps(sk, &rt->u.dst);
2000                 tcp_sk(sk)->ext2_header_len = rt->u.dst.header_len;
2001                 return 0;
2002         }
2003
2004         /* Routing failed... */
2005         sk->sk_route_caps = 0;
2006
2007         if (!sysctl_ip_dynaddr ||
2008             sk->sk_state != TCP_SYN_SENT ||
2009             (sk->sk_userlocks & SOCK_BINDADDR_LOCK) ||
2010             (err = tcp_v4_reselect_saddr(sk)) != 0)
2011                 sk->sk_err_soft = -err;
2012
2013         return err;
2014 }
2015
2016 static void v4_addr2sockaddr(struct sock *sk, struct sockaddr * uaddr)
2017 {
2018         struct sockaddr_in *sin = (struct sockaddr_in *) uaddr;
2019         struct inet_opt *inet = inet_sk(sk);
2020
2021         sin->sin_family         = AF_INET;
2022         sin->sin_addr.s_addr    = inet->daddr;
2023         sin->sin_port           = inet->dport;
2024 }
2025
2026 /* VJ's idea. Save last timestamp seen from this destination
2027  * and hold it at least for normal timewait interval to use for duplicate
2028  * segment detection in subsequent connections, before they enter synchronized
2029  * state.
2030  */
2031
2032 int tcp_v4_remember_stamp(struct sock *sk)
2033 {
2034         struct inet_opt *inet = inet_sk(sk);
2035         struct tcp_opt *tp = tcp_sk(sk);
2036         struct rtable *rt = (struct rtable *)__sk_dst_get(sk);
2037         struct inet_peer *peer = NULL;
2038         int release_it = 0;
2039
2040         if (!rt || rt->rt_dst != inet->daddr) {
2041                 peer = inet_getpeer(inet->daddr, 1);
2042                 release_it = 1;
2043         } else {
2044                 if (!rt->peer)
2045                         rt_bind_peer(rt, 1);
2046                 peer = rt->peer;
2047         }
2048
2049         if (peer) {
2050                 if ((s32)(peer->tcp_ts - tp->ts_recent) <= 0 ||
2051                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
2052                      peer->tcp_ts_stamp <= tp->ts_recent_stamp)) {
2053                         peer->tcp_ts_stamp = tp->ts_recent_stamp;
2054                         peer->tcp_ts = tp->ts_recent;
2055                 }
2056                 if (release_it)
2057                         inet_putpeer(peer);
2058                 return 1;
2059         }
2060
2061         return 0;
2062 }
2063
2064 int tcp_v4_tw_remember_stamp(struct tcp_tw_bucket *tw)
2065 {
2066         struct inet_peer *peer = NULL;
2067
2068         peer = inet_getpeer(tw->tw_daddr, 1);
2069
2070         if (peer) {
2071                 if ((s32)(peer->tcp_ts - tw->tw_ts_recent) <= 0 ||
2072                     (peer->tcp_ts_stamp + TCP_PAWS_MSL < xtime.tv_sec &&
2073                      peer->tcp_ts_stamp <= tw->tw_ts_recent_stamp)) {
2074                         peer->tcp_ts_stamp = tw->tw_ts_recent_stamp;
2075                         peer->tcp_ts = tw->tw_ts_recent;
2076                 }
2077                 inet_putpeer(peer);
2078                 return 1;
2079         }
2080
2081         return 0;
2082 }
2083
2084 struct tcp_func ipv4_specific = {
2085         .queue_xmit     =       ip_queue_xmit,
2086         .send_check     =       tcp_v4_send_check,
2087         .rebuild_header =       tcp_v4_rebuild_header,
2088         .conn_request   =       tcp_v4_conn_request,
2089         .syn_recv_sock  =       tcp_v4_syn_recv_sock,
2090         .remember_stamp =       tcp_v4_remember_stamp,
2091         .net_header_len =       sizeof(struct iphdr),
2092         .setsockopt     =       ip_setsockopt,
2093         .getsockopt     =       ip_getsockopt,
2094         .addr2sockaddr  =       v4_addr2sockaddr,
2095         .sockaddr_len   =       sizeof(struct sockaddr_in),
2096 };
2097
2098 /* NOTE: A lot of things set to zero explicitly by call to
2099  *       sk_alloc() so need not be done here.
2100  */
2101 static int tcp_v4_init_sock(struct sock *sk)
2102 {
2103         struct tcp_opt *tp = tcp_sk(sk);
2104
2105         skb_queue_head_init(&tp->out_of_order_queue);
2106         tcp_init_xmit_timers(sk);
2107         tcp_prequeue_init(tp);
2108
2109         tp->rto  = TCP_TIMEOUT_INIT;
2110         tp->mdev = TCP_TIMEOUT_INIT;
2111
2112         /* So many TCP implementations out there (incorrectly) count the
2113          * initial SYN frame in their delayed-ACK and congestion control
2114          * algorithms that we must have the following bandaid to talk
2115          * efficiently to them.  -DaveM
2116          */
2117         tp->snd_cwnd = 2;
2118
2119         /* See draft-stevens-tcpca-spec-01 for discussion of the
2120          * initialization of these values.
2121          */
2122         tp->snd_ssthresh = 0x7fffffff;  /* Infinity */
2123         tp->snd_cwnd_clamp = ~0;
2124         tp->mss_cache = 536;
2125
2126         tp->reordering = sysctl_tcp_reordering;
2127
2128         sk->sk_state = TCP_CLOSE;
2129
2130         sk->sk_write_space = sk_stream_write_space;
2131         sk->sk_use_write_queue = 1;
2132
2133         tp->af_specific = &ipv4_specific;
2134
2135         sk->sk_sndbuf = sysctl_tcp_wmem[1];
2136         sk->sk_rcvbuf = sysctl_tcp_rmem[1];
2137
2138         atomic_inc(&tcp_sockets_allocated);
2139
2140         return 0;
2141 }
2142
2143 int tcp_v4_destroy_sock(struct sock *sk)
2144 {
2145         struct tcp_opt *tp = tcp_sk(sk);
2146
2147         tcp_clear_xmit_timers(sk);
2148
2149         /* Cleanup up the write buffer. */
2150         sk_stream_writequeue_purge(sk);
2151
2152         /* Cleans up our, hopefully empty, out_of_order_queue. */
2153         __skb_queue_purge(&tp->out_of_order_queue);
2154
2155         /* Clean prequeue, it must be empty really */
2156         __skb_queue_purge(&tp->ucopy.prequeue);
2157
2158         /* Clean up a referenced TCP bind bucket. */
2159         if (tp->bind_hash)
2160                 tcp_put_port(sk);
2161
2162         /*
2163          * If sendmsg cached page exists, toss it.
2164          */
2165         if (sk->sk_sndmsg_page) {
2166                 __free_page(sk->sk_sndmsg_page);
2167                 sk->sk_sndmsg_page = NULL;
2168         }
2169
2170         atomic_dec(&tcp_sockets_allocated);
2171
2172         return 0;
2173 }
2174
2175 EXPORT_SYMBOL(tcp_v4_destroy_sock);
2176
2177 #ifdef CONFIG_PROC_FS
2178 /* Proc filesystem TCP sock list dumping. */
2179
2180 static inline struct tcp_tw_bucket *tw_head(struct hlist_head *head)
2181 {
2182         return hlist_empty(head) ? NULL :
2183                 list_entry(head->first, struct tcp_tw_bucket, tw_node);
2184 }
2185
2186 static inline struct tcp_tw_bucket *tw_next(struct tcp_tw_bucket *tw)
2187 {
2188         return tw->tw_node.next ?
2189                 hlist_entry(tw->tw_node.next, typeof(*tw), tw_node) : NULL;
2190 }
2191
2192 static void *listening_get_next(struct seq_file *seq, void *cur)
2193 {
2194         struct tcp_opt *tp;
2195         struct hlist_node *node;
2196         struct sock *sk = cur;
2197         struct tcp_iter_state* st = seq->private;
2198
2199         if (!sk) {
2200                 st->bucket = 0;
2201                 sk = sk_head(&tcp_listening_hash[0]);
2202                 goto get_sk;
2203         }
2204
2205         ++st->num;
2206
2207         if (st->state == TCP_SEQ_STATE_OPENREQ) {
2208                 struct open_request *req = cur;
2209
2210                 tp = tcp_sk(st->syn_wait_sk);
2211                 req = req->dl_next;
2212                 while (1) {
2213                         while (req) {
2214                                 vxdprintk(VXD_CBIT(net, 6),
2215                                         "sk,req: %p [#%d] (from %d)",
2216                                         req->sk, req->sk->sk_xid, current->xid);
2217                                 if (!vx_check(req->sk->sk_xid, VX_IDENT|VX_WATCH))
2218                                         continue;
2219                                 if (req->class->family == st->family) {
2220                                         cur = req;
2221                                         goto out;
2222                                 }
2223                                 req = req->dl_next;
2224                         }
2225                         if (++st->sbucket >= TCP_SYNQ_HSIZE)
2226                                 break;
2227 get_req:
2228                         req = tp->listen_opt->syn_table[st->sbucket];
2229                 }
2230                 sk        = sk_next(st->syn_wait_sk);
2231                 st->state = TCP_SEQ_STATE_LISTENING;
2232                 read_unlock_bh(&tp->syn_wait_lock);
2233         } else
2234                 sk = sk_next(sk);
2235 get_sk:
2236         sk_for_each_from(sk, node) {
2237                 vxdprintk(VXD_CBIT(net, 6), "sk: %p [#%d] (from %d)",
2238                         sk, sk->sk_xid, current->xid);
2239                 if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2240                         continue;
2241                 if (sk->sk_family == st->family) {
2242                         cur = sk;
2243                         goto out;
2244                 }
2245                 tp = tcp_sk(sk);
2246                 read_lock_bh(&tp->syn_wait_lock);
2247                 if (tp->listen_opt && tp->listen_opt->qlen) {
2248                         st->uid         = sock_i_uid(sk);
2249                         st->syn_wait_sk = sk;
2250                         st->state       = TCP_SEQ_STATE_OPENREQ;
2251                         st->sbucket     = 0;
2252                         goto get_req;
2253                 }
2254                 read_unlock_bh(&tp->syn_wait_lock);
2255         }
2256         if (++st->bucket < TCP_LHTABLE_SIZE) {
2257                 sk = sk_head(&tcp_listening_hash[st->bucket]);
2258                 goto get_sk;
2259         }
2260         cur = NULL;
2261 out:
2262         return cur;
2263 }
2264
2265 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2266 {
2267         void *rc = listening_get_next(seq, NULL);
2268
2269         while (rc && *pos) {
2270                 rc = listening_get_next(seq, rc);
2271                 --*pos;
2272         }
2273         return rc;
2274 }
2275
2276 static void *established_get_first(struct seq_file *seq)
2277 {
2278         struct tcp_iter_state* st = seq->private;
2279         void *rc = NULL;
2280
2281         for (st->bucket = 0; st->bucket < tcp_ehash_size; ++st->bucket) {
2282                 struct sock *sk;
2283                 struct hlist_node *node;
2284                 struct tcp_tw_bucket *tw;
2285
2286                 read_lock(&tcp_ehash[st->bucket].lock);
2287                 sk_for_each(sk, node, &tcp_ehash[st->bucket].chain) {
2288                         vxdprintk(VXD_CBIT(net, 6),
2289                                 "sk,egf: %p [#%d] (from %d)",
2290                                 sk, sk->sk_xid, current->xid);
2291                         if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2292                                 continue;
2293                         if (sk->sk_family != st->family)
2294                                 continue;
2295                         rc = sk;
2296                         goto out;
2297                 }
2298                 st->state = TCP_SEQ_STATE_TIME_WAIT;
2299                 tw_for_each(tw, node,
2300                             &tcp_ehash[st->bucket + tcp_ehash_size].chain) {
2301                         vxdprintk(VXD_CBIT(net, 6),
2302                                 "tw: %p [#%d] (from %d)",
2303                                 tw, tw->tw_xid, current->xid);
2304                         if (!vx_check(tw->tw_xid, VX_IDENT|VX_WATCH))
2305                                 continue;
2306                         if (tw->tw_family != st->family)
2307                                 continue;
2308                         rc = tw;
2309                         goto out;
2310                 }
2311                 read_unlock(&tcp_ehash[st->bucket].lock);
2312                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2313         }
2314 out:
2315         return rc;
2316 }
2317
2318 static void *established_get_next(struct seq_file *seq, void *cur)
2319 {
2320         struct sock *sk = cur;
2321         struct tcp_tw_bucket *tw;
2322         struct hlist_node *node;
2323         struct tcp_iter_state* st = seq->private;
2324
2325         ++st->num;
2326
2327         if (st->state == TCP_SEQ_STATE_TIME_WAIT) {
2328                 tw = cur;
2329                 tw = tw_next(tw);
2330 get_tw:
2331                 while (tw && (tw->tw_family != st->family ||
2332                         !vx_check(tw->tw_xid, VX_IDENT|VX_WATCH))) {
2333                         tw = tw_next(tw);
2334                 }
2335                 if (tw) {
2336                         cur = tw;
2337                         goto out;
2338                 }
2339                 read_unlock(&tcp_ehash[st->bucket].lock);
2340                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2341                 if (++st->bucket < tcp_ehash_size) {
2342                         read_lock(&tcp_ehash[st->bucket].lock);
2343                         sk = sk_head(&tcp_ehash[st->bucket].chain);
2344                 } else {
2345                         cur = NULL;
2346                         goto out;
2347                 }
2348         } else
2349                 sk = sk_next(sk);
2350
2351         sk_for_each_from(sk, node) {
2352                 vxdprintk(VXD_CBIT(net, 6),
2353                         "sk,egn: %p [#%d] (from %d)",
2354                         sk, sk->sk_xid, current->xid);
2355                 if (!vx_check(sk->sk_xid, VX_IDENT|VX_WATCH))
2356                         continue;
2357                 if (sk->sk_family == st->family)
2358                         goto found;
2359         }
2360
2361         st->state = TCP_SEQ_STATE_TIME_WAIT;
2362         tw = tw_head(&tcp_ehash[st->bucket + tcp_ehash_size].chain);
2363         goto get_tw;
2364 found:
2365         cur = sk;
2366 out:
2367         return cur;
2368 }
2369
2370 static void *established_get_idx(struct seq_file *seq, loff_t pos)
2371 {
2372         void *rc = established_get_first(seq);
2373
2374         while (rc && pos) {
2375                 rc = established_get_next(seq, rc);
2376                 --pos;
2377         }
2378         return rc;
2379 }
2380
2381 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2382 {
2383         void *rc;
2384         struct tcp_iter_state* st = seq->private;
2385
2386         tcp_listen_lock();
2387         st->state = TCP_SEQ_STATE_LISTENING;
2388         rc        = listening_get_idx(seq, &pos);
2389
2390         if (!rc) {
2391                 tcp_listen_unlock();
2392                 local_bh_disable();
2393                 st->state = TCP_SEQ_STATE_ESTABLISHED;
2394                 rc        = established_get_idx(seq, pos);
2395         }
2396
2397         return rc;
2398 }
2399
2400 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2401 {
2402         struct tcp_iter_state* st = seq->private;
2403         st->state = TCP_SEQ_STATE_LISTENING;
2404         st->num = 0;
2405         return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2406 }
2407
2408 static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2409 {
2410         void *rc = NULL;
2411         struct tcp_iter_state* st;
2412
2413         if (v == SEQ_START_TOKEN) {
2414                 rc = tcp_get_idx(seq, 0);
2415                 goto out;
2416         }
2417         st = seq->private;
2418
2419         switch (st->state) {
2420         case TCP_SEQ_STATE_OPENREQ:
2421         case TCP_SEQ_STATE_LISTENING:
2422                 rc = listening_get_next(seq, v);
2423                 if (!rc) {
2424                         tcp_listen_unlock();
2425                         local_bh_disable();
2426                         st->state = TCP_SEQ_STATE_ESTABLISHED;
2427                         rc        = established_get_first(seq);
2428                 }
2429                 break;
2430         case TCP_SEQ_STATE_ESTABLISHED:
2431         case TCP_SEQ_STATE_TIME_WAIT:
2432                 rc = established_get_next(seq, v);
2433                 break;
2434         }
2435 out:
2436         ++*pos;
2437         return rc;
2438 }
2439
2440 static void tcp_seq_stop(struct seq_file *seq, void *v)
2441 {
2442         struct tcp_iter_state* st = seq->private;
2443
2444         switch (st->state) {
2445         case TCP_SEQ_STATE_OPENREQ:
2446                 if (v) {
2447                         struct tcp_opt *tp = tcp_sk(st->syn_wait_sk);
2448                         read_unlock_bh(&tp->syn_wait_lock);
2449                 }
2450         case TCP_SEQ_STATE_LISTENING:
2451                 if (v != SEQ_START_TOKEN)
2452                         tcp_listen_unlock();
2453                 break;
2454         case TCP_SEQ_STATE_TIME_WAIT:
2455         case TCP_SEQ_STATE_ESTABLISHED:
2456                 if (v)
2457                         read_unlock(&tcp_ehash[st->bucket].lock);
2458                 local_bh_enable();
2459                 break;
2460         }
2461 }
2462
2463 static int tcp_seq_open(struct inode *inode, struct file *file)
2464 {
2465         struct tcp_seq_afinfo *afinfo = PDE(inode)->data;
2466         struct seq_file *seq;
2467         struct tcp_iter_state *s;
2468         int rc;
2469
2470         if (unlikely(afinfo == NULL))
2471                 return -EINVAL;
2472
2473         s = kmalloc(sizeof(*s), GFP_KERNEL);
2474         if (!s)
2475                 return -ENOMEM;
2476         memset(s, 0, sizeof(*s));
2477         s->family               = afinfo->family;
2478         s->seq_ops.start        = tcp_seq_start;
2479         s->seq_ops.next         = tcp_seq_next;
2480         s->seq_ops.show         = afinfo->seq_show;
2481         s->seq_ops.stop         = tcp_seq_stop;
2482
2483         rc = seq_open(file, &s->seq_ops);
2484         if (rc)
2485                 goto out_kfree;
2486         seq          = file->private_data;
2487         seq->private = s;
2488 out:
2489         return rc;
2490 out_kfree:
2491         kfree(s);
2492         goto out;
2493 }
2494
2495 int tcp_proc_register(struct tcp_seq_afinfo *afinfo)
2496 {
2497         int rc = 0;
2498         struct proc_dir_entry *p;
2499
2500         if (!afinfo)
2501                 return -EINVAL;
2502         afinfo->seq_fops->owner         = afinfo->owner;
2503         afinfo->seq_fops->open          = tcp_seq_open;
2504         afinfo->seq_fops->read          = seq_read;
2505         afinfo->seq_fops->llseek        = seq_lseek;
2506         afinfo->seq_fops->release       = seq_release_private;
2507
2508         p = proc_net_fops_create(afinfo->name, S_IRUGO, afinfo->seq_fops);
2509         if (p)
2510                 p->data = afinfo;
2511         else
2512                 rc = -ENOMEM;
2513         return rc;
2514 }
2515
2516 void tcp_proc_unregister(struct tcp_seq_afinfo *afinfo)
2517 {
2518         if (!afinfo)
2519                 return;
2520         proc_net_remove(afinfo->name);
2521         memset(afinfo->seq_fops, 0, sizeof(*afinfo->seq_fops));
2522 }
2523
2524 static void get_openreq4(struct sock *sk, struct open_request *req,
2525                          char *tmpbuf, int i, int uid)
2526 {
2527         int ttd = req->expires - jiffies;
2528
2529         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
2530                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %u %d %p",
2531                 i,
2532                 req->af.v4_req.loc_addr,
2533                 ntohs(inet_sk(sk)->sport),
2534                 req->af.v4_req.rmt_addr,
2535                 ntohs(req->rmt_port),
2536                 TCP_SYN_RECV,
2537                 0, 0, /* could print option size, but that is af dependent. */
2538                 1,    /* timers active (only the expire timer) */
2539                 jiffies_to_clock_t(ttd),
2540                 req->retrans,
2541                 uid,
2542                 0,  /* non standard timer */
2543                 0, /* open_requests have no inode */
2544                 atomic_read(&sk->sk_refcnt),
2545                 req);
2546 }
2547
2548 static void get_tcp4_sock(struct sock *sp, char *tmpbuf, int i)
2549 {
2550         int timer_active;
2551         unsigned long timer_expires;
2552         struct tcp_opt *tp = tcp_sk(sp);
2553         struct inet_opt *inet = inet_sk(sp);
2554         unsigned int dest = inet->daddr;
2555         unsigned int src = inet->rcv_saddr;
2556         __u16 destp = ntohs(inet->dport);
2557         __u16 srcp = ntohs(inet->sport);
2558
2559         if (tp->pending == TCP_TIME_RETRANS) {
2560                 timer_active    = 1;
2561                 timer_expires   = tp->timeout;
2562         } else if (tp->pending == TCP_TIME_PROBE0) {
2563                 timer_active    = 4;
2564                 timer_expires   = tp->timeout;
2565         } else if (timer_pending(&sp->sk_timer)) {
2566                 timer_active    = 2;
2567                 timer_expires   = sp->sk_timer.expires;
2568         } else {
2569                 timer_active    = 0;
2570                 timer_expires = jiffies;
2571         }
2572
2573         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2574                         "%08X %5d %8d %lu %d %p %u %u %u %u %d",
2575                 i, src, srcp, dest, destp, sp->sk_state,
2576                 tp->write_seq - tp->snd_una, tp->rcv_nxt - tp->copied_seq,
2577                 timer_active,
2578                 jiffies_to_clock_t(timer_expires - jiffies),
2579                 tp->retransmits,
2580                 sock_i_uid(sp),
2581                 tp->probes_out,
2582                 sock_i_ino(sp),
2583                 atomic_read(&sp->sk_refcnt), sp,
2584                 tp->rto, tp->ack.ato, (tp->ack.quick << 1) | tp->ack.pingpong,
2585                 tp->snd_cwnd,
2586                 tp->snd_ssthresh >= 0xFFFF ? -1 : tp->snd_ssthresh);
2587 }
2588
2589 static void get_timewait4_sock(struct tcp_tw_bucket *tw, char *tmpbuf, int i)
2590 {
2591         unsigned int dest, src;
2592         __u16 destp, srcp;
2593         int ttd = tw->tw_ttd - jiffies;
2594
2595         if (ttd < 0)
2596                 ttd = 0;
2597
2598         dest  = tw->tw_daddr;
2599         src   = tw->tw_rcv_saddr;
2600         destp = ntohs(tw->tw_dport);
2601         srcp  = ntohs(tw->tw_sport);
2602
2603         sprintf(tmpbuf, "%4d: %08X:%04X %08X:%04X"
2604                 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %p",
2605                 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2606                 3, jiffies_to_clock_t(ttd), 0, 0, 0, 0,
2607                 atomic_read(&tw->tw_refcnt), tw);
2608 }
2609
2610 #define TMPSZ 150
2611
2612 static int tcp4_seq_show(struct seq_file *seq, void *v)
2613 {
2614         struct tcp_iter_state* st;
2615         char tmpbuf[TMPSZ + 1];
2616
2617         if (v == SEQ_START_TOKEN) {
2618                 seq_printf(seq, "%-*s\n", TMPSZ - 1,
2619                            "  sl  local_address rem_address   st tx_queue "
2620                            "rx_queue tr tm->when retrnsmt   uid  timeout "
2621                            "inode");
2622                 goto out;
2623         }
2624         st = seq->private;
2625
2626         switch (st->state) {
2627         case TCP_SEQ_STATE_LISTENING:
2628         case TCP_SEQ_STATE_ESTABLISHED:
2629                 get_tcp4_sock(v, tmpbuf, st->num);
2630                 break;
2631         case TCP_SEQ_STATE_OPENREQ:
2632                 get_openreq4(st->syn_wait_sk, v, tmpbuf, st->num, st->uid);
2633                 break;
2634         case TCP_SEQ_STATE_TIME_WAIT:
2635                 get_timewait4_sock(v, tmpbuf, st->num);
2636                 break;
2637         }
2638         seq_printf(seq, "%-*s\n", TMPSZ - 1, tmpbuf);
2639 out:
2640         return 0;
2641 }
2642
2643 static struct file_operations tcp4_seq_fops;
2644 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
2645         .owner          = THIS_MODULE,
2646         .name           = "tcp",
2647         .family         = AF_INET,
2648         .seq_show       = tcp4_seq_show,
2649         .seq_fops       = &tcp4_seq_fops,
2650 };
2651
2652 int __init tcp4_proc_init(void)
2653 {
2654         return tcp_proc_register(&tcp4_seq_afinfo);
2655 }
2656
2657 void tcp4_proc_exit(void)
2658 {
2659         tcp_proc_unregister(&tcp4_seq_afinfo);
2660 }
2661 #endif /* CONFIG_PROC_FS */
2662
2663 struct proto tcp_prot = {
2664         .name                   = "TCP",
2665         .close                  = tcp_close,
2666         .connect                = tcp_v4_connect,
2667         .disconnect             = tcp_disconnect,
2668         .accept                 = tcp_accept,
2669         .ioctl                  = tcp_ioctl,
2670         .init                   = tcp_v4_init_sock,
2671         .destroy                = tcp_v4_destroy_sock,
2672         .shutdown               = tcp_shutdown,
2673         .setsockopt             = tcp_setsockopt,
2674         .getsockopt             = tcp_getsockopt,
2675         .sendmsg                = tcp_sendmsg,
2676         .recvmsg                = tcp_recvmsg,
2677         .backlog_rcv            = tcp_v4_do_rcv,
2678         .hash                   = tcp_v4_hash,
2679         .unhash                 = tcp_unhash,
2680         .get_port               = tcp_v4_get_port,
2681         .enter_memory_pressure  = tcp_enter_memory_pressure,
2682         .sockets_allocated      = &tcp_sockets_allocated,
2683         .memory_allocated       = &tcp_memory_allocated,
2684         .memory_pressure        = &tcp_memory_pressure,
2685         .sysctl_mem             = sysctl_tcp_mem,
2686         .sysctl_wmem            = sysctl_tcp_wmem,
2687         .sysctl_rmem            = sysctl_tcp_rmem,
2688         .max_header             = MAX_TCP_HEADER,
2689 };
2690
2691
2692
2693 void __init tcp_v4_init(struct net_proto_family *ops)
2694 {
2695         int err = sock_create_kern(PF_INET, SOCK_RAW, IPPROTO_TCP, &tcp_socket);
2696         if (err < 0)
2697                 panic("Failed to create the TCP control socket.\n");
2698         tcp_socket->sk->sk_allocation   = GFP_ATOMIC;
2699         inet_sk(tcp_socket->sk)->uc_ttl = -1;
2700
2701         /* Unhash it so that IP input processing does not even
2702          * see it, we do not wish this socket to see incoming
2703          * packets.
2704          */
2705         tcp_socket->sk->sk_prot->unhash(tcp_socket->sk);
2706 }
2707
2708 EXPORT_SYMBOL(ipv4_specific);
2709 EXPORT_SYMBOL(tcp_bind_hash);
2710 EXPORT_SYMBOL(tcp_bucket_create);
2711 EXPORT_SYMBOL(tcp_hashinfo);
2712 EXPORT_SYMBOL(tcp_inherit_port);
2713 EXPORT_SYMBOL(tcp_listen_wlock);
2714 EXPORT_SYMBOL(tcp_port_rover);
2715 EXPORT_SYMBOL(tcp_prot);
2716 EXPORT_SYMBOL(tcp_put_port);
2717 EXPORT_SYMBOL(tcp_unhash);
2718 EXPORT_SYMBOL(tcp_v4_conn_request);
2719 EXPORT_SYMBOL(tcp_v4_connect);
2720 EXPORT_SYMBOL(tcp_v4_do_rcv);
2721 EXPORT_SYMBOL(tcp_v4_lookup_listener);
2722 EXPORT_SYMBOL(tcp_v4_rebuild_header);
2723 EXPORT_SYMBOL(tcp_v4_remember_stamp);
2724 EXPORT_SYMBOL(tcp_v4_send_check);
2725 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
2726
2727 #ifdef CONFIG_PROC_FS
2728 EXPORT_SYMBOL(tcp_proc_register);
2729 EXPORT_SYMBOL(tcp_proc_unregister);
2730 #endif
2731 #ifdef CONFIG_SYSCTL
2732 EXPORT_SYMBOL(sysctl_local_port_range);
2733 EXPORT_SYMBOL(sysctl_max_syn_backlog);
2734 EXPORT_SYMBOL(sysctl_tcp_low_latency);
2735 #endif