net/ipv4/tcp_minisocks.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_minisocks.c,v 1.15 2002/02/01 22:01:04 davem Exp $
   9  *
  10  * Authors:     Ross Biro, <bir7@leland.Stanford.Edu>
  11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  13  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  14  *              Florian La Roche, <flla@stud.uni-sb.de>
  15  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  16  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  17  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  18  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  19  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  20  *              Jorge Cwik, <jorge@laser.satlink.net>
  21  */
  22
  23 #include <linux/config.h>
  24 #include <linux/mm.h>
  25 #include <linux/module.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/vs_socket.h>
  29 #include <net/tcp.h>
  30 #include <net/inet_common.h>
  31 #include <net/xfrm.h>
  32
  33 #ifdef CONFIG_SYSCTL
  34 #define SYNC_INIT 0 /* let the user enable it */
  35 #else
  36 #define SYNC_INIT 1
  37 #endif
  38
  39 int sysctl_tcp_tw_recycle;
  40 int sysctl_tcp_max_tw_buckets = NR_FILE*2;
  41
  42 int sysctl_tcp_syncookies = SYNC_INIT;
  43 int sysctl_tcp_abort_on_overflow;
  44
  45 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
  46 {
  47         if (seq == s_win)
  48                 return 1;
  49         if (after(end_seq, s_win) && before(seq, e_win))
  50                 return 1;
  51         return (seq == e_win && seq == end_seq);
  52 }
  53
  54 /* New-style handling of TIME_WAIT sockets. */
  55
  56 int tcp_tw_count;
  57
  58
  59 /* Must be called with locally disabled BHs. */
  60 static void tcp_timewait_kill(struct tcp_tw_bucket *tw)
  61 {
  62         struct tcp_ehash_bucket *ehead;
  63         struct tcp_bind_hashbucket *bhead;
  64         struct tcp_bind_bucket *tb;
  65
  66         /* Unlink from established hashes. */
  67         ehead = &tcp_ehash[tw->tw_hashent];
  68         write_lock(&ehead->lock);
  69         if (hlist_unhashed(&tw->tw_node)) {
  70                 write_unlock(&ehead->lock);
  71                 return;
  72         }
  73         __hlist_del(&tw->tw_node);
  74         sk_node_init(&tw->tw_node);
  75         write_unlock(&ehead->lock);
  76
  77         /* Disassociate with bind bucket. */
  78         bhead = &tcp_bhash[tcp_bhashfn(tw->tw_num)];
  79         spin_lock(&bhead->lock);
  80         tb = tw->tw_tb;
  81         __hlist_del(&tw->tw_bind_node);
  82         tw->tw_tb = NULL;
  83         tcp_bucket_destroy(tb);
  84         spin_unlock(&bhead->lock);
  85
  86 #ifdef INET_REFCNT_DEBUG
  87         if (atomic_read(&tw->tw_refcnt) != 1) {
  88                 printk(KERN_DEBUG "tw_bucket %p refcnt=%d\n", tw,
  89                        atomic_read(&tw->tw_refcnt));
  90         }
  91 #endif
  92         tcp_tw_put(tw);
  93 }
  94
  95 /*
  96  * * Main purpose of TIME-WAIT state is to close connection gracefully,
  97  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
  98  *   (and, probably, tail of data) and one or more our ACKs are lost.
  99  * * What is TIME-WAIT timeout? It is associated with maximal packet
 100  *   lifetime in the internet, which results in wrong conclusion, that
 101  *   it is set to catch "old duplicate segments" wandering out of their path.
 102  *   It is not quite correct. This timeout is calculated so that it exceeds
 103  *   maximal retransmission timeout enough to allow to lose one (or more)
 104  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
 105  * * When TIME-WAIT socket receives RST, it means that another end
 106  *   finally closed and we are allowed to kill TIME-WAIT too.
 107  * * Second purpose of TIME-WAIT is catching old duplicate segments.
 108  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
 109  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
 110  * * If we invented some more clever way to catch duplicates
 111  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
 112  *
 113  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
 114  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
 115  * from the very beginning.
 116  *
 117  * NOTE. With recycling (and later with fin-wait-2) TW bucket
 118  * is _not_ stateless. It means, that strictly speaking we must
 119  * spinlock it. I do not want! Well, probability of misbehaviour
 120  * is ridiculously low and, seems, we could use some mb() tricks
 121  * to avoid misread sequence numbers, states etc.  --ANK
 122  */
 123 enum tcp_tw_status
 124 tcp_timewait_state_process(struct tcp_tw_bucket *tw, struct sk_buff *skb,
 125                            struct tcphdr *th, unsigned len)
 126 {
 127         struct tcp_opt tp;
 128         int paws_reject = 0;
 129
 130         tp.saw_tstamp = 0;
 131         if (th->doff > (sizeof(struct tcphdr) >> 2) && tw->tw_ts_recent_stamp) {
 132                 tcp_parse_options(skb, &tp, 0);
 133
 134                 if (tp.saw_tstamp) {
 135                         tp.ts_recent       = tw->tw_ts_recent;
 136                         tp.ts_recent_stamp = tw->tw_ts_recent_stamp;
 137                         paws_reject = tcp_paws_check(&tp, th->rst);
 138                 }
 139         }
 140
 141         if (tw->tw_substate == TCP_FIN_WAIT2) {
 142                 /* Just repeat all the checks of tcp_rcv_state_process() */
 143
 144                 /* Out of window, send ACK */
 145                 if (paws_reject ||
 146                     !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
 147                                    tw->tw_rcv_nxt,
 148                                    tw->tw_rcv_nxt + tw->tw_rcv_wnd))
 149                         return TCP_TW_ACK;
 150
 151                 if (th->rst)
 152                         goto kill;
 153
 154                 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt))
 155                         goto kill_with_rst;
 156
 157                 /* Dup ACK? */
 158                 if (!after(TCP_SKB_CB(skb)->end_seq, tw->tw_rcv_nxt) ||
 159                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
 160                         tcp_tw_put(tw);
 161                         return TCP_TW_SUCCESS;
 162                 }
 163
 164                 /* New data or FIN. If new data arrive after half-duplex close,
 165                  * reset.
 166                  */
 167                 if (!th->fin ||
 168                     TCP_SKB_CB(skb)->end_seq != tw->tw_rcv_nxt + 1) {
 169 kill_with_rst:
 170                         tcp_tw_deschedule(tw);
 171                         tcp_tw_put(tw);
 172                         return TCP_TW_RST;
 173                 }
 174
 175                 /* FIN arrived, enter true time-wait state. */
 176                 tw->tw_substate = TCP_TIME_WAIT;
 177                 tw->tw_rcv_nxt  = TCP_SKB_CB(skb)->end_seq;
 178                 if (tp.saw_tstamp) {
 179                         tw->tw_ts_recent_stamp  = xtime.tv_sec;
 180                         tw->tw_ts_recent        = tp.rcv_tsval;
 181                 }
 182
 183                 /* I am shamed, but failed to make it more elegant.
 184                  * Yes, it is direct reference to IP, which is impossible
 185                  * to generalize to IPv6. Taking into account that IPv6
 186                  * do not undertsnad recycling in any case, it not
 187                  * a big problem in practice. --ANK */
 188                 if (tw->tw_family == AF_INET &&
 189                     sysctl_tcp_tw_recycle && tw->tw_ts_recent_stamp &&
 190                     tcp_v4_tw_remember_stamp(tw))
 191                         tcp_tw_schedule(tw, tw->tw_timeout);
 192                 else
 193                         tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
 194                 return TCP_TW_ACK;
 195         }
 196
 197         /*
 198          *      Now real TIME-WAIT state.
 199          *
 200          *      RFC 1122:
 201          *      "When a connection is [...] on TIME-WAIT state [...]
 202          *      [a TCP] MAY accept a new SYN from the remote TCP to
 203          *      reopen the connection directly, if it:
 204          *
 205          *      (1)  assigns its initial sequence number for the new
 206          *      connection to be larger than the largest sequence
 207          *      number it used on the previous connection incarnation,
 208          *      and
 209          *
 210          *      (2)  returns to TIME-WAIT state if the SYN turns out
 211          *      to be an old duplicate".
 212          */
 213
 214         if (!paws_reject &&
 215             (TCP_SKB_CB(skb)->seq == tw->tw_rcv_nxt &&
 216              (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
 217                 /* In window segment, it may be only reset or bare ack. */
 218
 219                 if (th->rst) {
 220                         /* This is TIME_WAIT assasination, in two flavors.
 221                          * Oh well... nobody has a sufficient solution to this
 222                          * protocol bug yet.
 223                          */
 224                         if (sysctl_tcp_rfc1337 == 0) {
 225 kill:
 226                                 tcp_tw_deschedule(tw);
 227                                 tcp_tw_put(tw);
 228                                 return TCP_TW_SUCCESS;
 229                         }
 230                 }
 231                 tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
 232
 233                 if (tp.saw_tstamp) {
 234                         tw->tw_ts_recent        = tp.rcv_tsval;
 235                         tw->tw_ts_recent_stamp  = xtime.tv_sec;
 236                 }
 237
 238                 tcp_tw_put(tw);
 239                 return TCP_TW_SUCCESS;
 240         }
 241
 242         /* Out of window segment.
 243
 244            All the segments are ACKed immediately.
 245
 246            The only exception is new SYN. We accept it, if it is
 247            not old duplicate and we are not in danger to be killed
 248            by delayed old duplicates. RFC check is that it has
 249            newer sequence number works at rates <40Mbit/sec.
 250            However, if paws works, it is reliable AND even more,
 251            we even may relax silly seq space cutoff.
 252
 253            RED-PEN: we violate main RFC requirement, if this SYN will appear
 254            old duplicate (i.e. we receive RST in reply to SYN-ACK),
 255            we must return socket to time-wait state. It is not good,
 256            but not fatal yet.
 257          */
 258
 259         if (th->syn && !th->rst && !th->ack && !paws_reject &&
 260             (after(TCP_SKB_CB(skb)->seq, tw->tw_rcv_nxt) ||
 261              (tp.saw_tstamp && (s32)(tw->tw_ts_recent - tp.rcv_tsval) < 0))) {
 262                 u32 isn = tw->tw_snd_nxt + 65535 + 2;
 263                 if (isn == 0)
 264                         isn++;
 265                 TCP_SKB_CB(skb)->when = isn;
 266                 return TCP_TW_SYN;
 267         }
 268
 269         if (paws_reject)
 270                 NET_INC_STATS_BH(PAWSEstabRejected);
 271
 272         if(!th->rst) {
 273                 /* In this case we must reset the TIMEWAIT timer.
 274                  *
 275                  * If it is ACKless SYN it may be both old duplicate
 276                  * and new good SYN with random sequence number <rcv_nxt.
 277                  * Do not reschedule in the last case.
 278                  */
 279                 if (paws_reject || th->ack)
 280                         tcp_tw_schedule(tw, TCP_TIMEWAIT_LEN);
 281
 282                 /* Send ACK. Note, we do not put the bucket,
 283                  * it will be released by caller.
 284                  */
 285                 return TCP_TW_ACK;
 286         }
 287         tcp_tw_put(tw);
 288         return TCP_TW_SUCCESS;
 289 }
 290
 291 /* Enter the time wait state.  This is called with locally disabled BH.
 292  * Essentially we whip up a timewait bucket, copy the
 293  * relevant info into it from the SK, and mess with hash chains
 294  * and list linkage.
 295  */
 296 static void __tcp_tw_hashdance(struct sock *sk, struct tcp_tw_bucket *tw)
 297 {
 298         struct tcp_ehash_bucket *ehead = &tcp_ehash[sk->sk_hashent];
 299         struct tcp_bind_hashbucket *bhead;
 300
 301         /* Step 1: Put TW into bind hash. Original socket stays there too.
 302            Note, that any socket with inet_sk(sk)->num != 0 MUST be bound in
 303            binding cache, even if it is closed.
 304          */
 305         bhead = &tcp_bhash[tcp_bhashfn(inet_sk(sk)->num)];
 306         spin_lock(&bhead->lock);
 307         tw->tw_tb = tcp_sk(sk)->bind_hash;
 308         BUG_TRAP(tcp_sk(sk)->bind_hash);
 309         tw_add_bind_node(tw, &tw->tw_tb->owners);
 310         spin_unlock(&bhead->lock);
 311
 312         write_lock(&ehead->lock);
 313
 314         /* Step 2: Remove SK from established hash. */
 315         if (__sk_del_node_init(sk))
 316                 sock_prot_dec_use(sk->sk_prot);
 317
 318         /* Step 3: Hash TW into TIMEWAIT half of established hash table. */
 319         tw_add_node(tw, &(ehead + tcp_ehash_size)->chain);
 320         atomic_inc(&tw->tw_refcnt);
 321
 322         write_unlock(&ehead->lock);
 323 }
 324
 325 /*
 326  * Move a socket to time-wait or dead fin-wait-2 state.
 327  */
 328 void tcp_time_wait(struct sock *sk, int state, int timeo)
 329 {
 330         struct tcp_tw_bucket *tw = NULL;
 331         struct tcp_opt *tp = tcp_sk(sk);
 332         int recycle_ok = 0;
 333
 334         if (sysctl_tcp_tw_recycle && tp->ts_recent_stamp)
 335                 recycle_ok = tp->af_specific->remember_stamp(sk);
 336
 337         if (tcp_tw_count < sysctl_tcp_max_tw_buckets)
 338                 tw = kmem_cache_alloc(tcp_timewait_cachep, SLAB_ATOMIC);
 339
 340         if(tw != NULL) {
 341                 struct inet_opt *inet = inet_sk(sk);
 342                 int rto = (tp->rto<<2) - (tp->rto>>1);
 343
 344                 /* Give us an identity. */
 345                 tw->tw_daddr            = inet->daddr;
 346                 tw->tw_rcv_saddr        = inet->rcv_saddr;
 347                 tw->tw_bound_dev_if     = sk->sk_bound_dev_if;
 348                 tw->tw_num              = inet->num;
 349                 tw->tw_state            = TCP_TIME_WAIT;
 350                 tw->tw_substate         = state;
 351                 tw->tw_sport            = inet->sport;
 352                 tw->tw_dport            = inet->dport;
 353                 tw->tw_family           = sk->sk_family;
 354                 tw->tw_reuse            = sk->sk_reuse;
 355                 tw->tw_rcv_wscale       = tp->rcv_wscale;
 356                 atomic_set(&tw->tw_refcnt, 1);
 357
 358                 tw->tw_hashent          = sk->sk_hashent;
 359                 tw->tw_rcv_nxt          = tp->rcv_nxt;
 360                 tw->tw_snd_nxt          = tp->snd_nxt;
 361                 tw->tw_rcv_wnd          = tcp_receive_window(tp);
 362                 tw->tw_ts_recent        = tp->ts_recent;
 363                 tw->tw_ts_recent_stamp  = tp->ts_recent_stamp;
 364                 tw_dead_node_init(tw);
 365
 366                 tw->tw_xid              = sk->sk_xid;
 367                 tw->tw_vx_info          = NULL;
 368                 tw->tw_nid              = sk->sk_nid;
 369                 tw->tw_nx_info          = NULL;
 370
 371 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 372                 if (tw->tw_family == PF_INET6) {
 373                         struct ipv6_pinfo *np = inet6_sk(sk);
 374
 375                         ipv6_addr_copy(&tw->tw_v6_daddr, &np->daddr);
 376                         ipv6_addr_copy(&tw->tw_v6_rcv_saddr, &np->rcv_saddr);
 377                         tw->tw_v6_ipv6only = np->ipv6only;
 378                 } else {
 379                         memset(&tw->tw_v6_daddr, 0, sizeof(tw->tw_v6_daddr));
 380                         memset(&tw->tw_v6_rcv_saddr, 0, sizeof(tw->tw_v6_rcv_saddr));
 381                         tw->tw_v6_ipv6only = 0;
 382                 }
 383 #endif
 384                 /* Linkage updates. */
 385                 __tcp_tw_hashdance(sk, tw);
 386
 387                 /* Get the TIME_WAIT timeout firing. */
 388                 if (timeo < rto)
 389                         timeo = rto;
 390
 391                 if (recycle_ok) {
 392                         tw->tw_timeout = rto;
 393                 } else {
 394                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
 395                         if (state == TCP_TIME_WAIT)
 396                                 timeo = TCP_TIMEWAIT_LEN;
 397                 }
 398
 399                 tcp_tw_schedule(tw, timeo);
 400                 tcp_tw_put(tw);
 401         } else {
 402                 /* Sorry, if we're out of memory, just CLOSE this
 403                  * socket up.  We've got bigger problems than
 404                  * non-graceful socket closings.
 405                  */
 406                 if (net_ratelimit())
 407                         printk(KERN_INFO "TCP: time wait bucket table overflow\n");
 408         }
 409
 410         tcp_update_metrics(sk);
 411         tcp_done(sk);
 412 }
 413
 414 /* Kill off TIME_WAIT sockets once their lifetime has expired. */
 415 static int tcp_tw_death_row_slot;
 416
 417 static void tcp_twkill(unsigned long);
 418
 419 /* TIME_WAIT reaping mechanism. */
 420 #define TCP_TWKILL_SLOTS        8       /* Please keep this a power of 2. */
 421 #define TCP_TWKILL_PERIOD       (TCP_TIMEWAIT_LEN/TCP_TWKILL_SLOTS)
 422
 423 #define TCP_TWKILL_QUOTA        100
 424
 425 static struct hlist_head tcp_tw_death_row[TCP_TWKILL_SLOTS];
 426 static spinlock_t tw_death_lock = SPIN_LOCK_UNLOCKED;
 427 static struct timer_list tcp_tw_timer = TIMER_INITIALIZER(tcp_twkill, 0, 0);
 428 static void twkill_work(void *);
 429 static DECLARE_WORK(tcp_twkill_work, twkill_work, NULL);
 430 static u32 twkill_thread_slots;
 431
 432 /* Returns non-zero if quota exceeded.  */
 433 static int tcp_do_twkill_work(int slot, unsigned int quota)
 434 {
 435         struct tcp_tw_bucket *tw;
 436         struct hlist_node *node;
 437         unsigned int killed;
 438         int ret;
 439
 440         /* NOTE: compare this to previous version where lock
 441          * was released after detaching chain. It was racy,
 442          * because tw buckets are scheduled in not serialized context
 443          * in 2.3 (with netfilter), and with softnet it is common, because
 444          * soft irqs are not sequenced.
 445          */
 446         killed = 0;
 447         ret = 0;
 448 rescan:
 449         tw_for_each_inmate(tw, node, &tcp_tw_death_row[slot]) {
 450                 __tw_del_dead_node(tw);
 451                 spin_unlock(&tw_death_lock);
 452                 tcp_timewait_kill(tw);
 453                 tcp_tw_put(tw);
 454                 killed++;
 455                 spin_lock(&tw_death_lock);
 456                 if (killed > quota) {
 457                         ret = 1;
 458                         break;
 459                 }
 460
 461                 /* While we dropped tw_death_lock, another cpu may have
 462                  * killed off the next TW bucket in the list, therefore
 463                  * do a fresh re-read of the hlist head node with the
 464                  * lock reacquired.  We still use the hlist traversal
 465                  * macro in order to get the prefetches.
 466                  */
 467                 goto rescan;
 468         }
 469
 470         tcp_tw_count -= killed;
 471         NET_ADD_STATS_BH(TimeWaited, killed);
 472
 473         return ret;
 474 }
 475
 476 static void tcp_twkill(unsigned long dummy)
 477 {
 478         int need_timer, ret;
 479
 480         spin_lock(&tw_death_lock);
 481
 482         if (tcp_tw_count == 0)
 483                 goto out;
 484
 485         need_timer = 0;
 486         ret = tcp_do_twkill_work(tcp_tw_death_row_slot, TCP_TWKILL_QUOTA);
 487         if (ret) {
 488                 twkill_thread_slots |= (1 << tcp_tw_death_row_slot);
 489                 mb();
 490                 schedule_work(&tcp_twkill_work);
 491                 need_timer = 1;
 492         } else {
 493                 /* We purged the entire slot, anything left?  */
 494                 if (tcp_tw_count)
 495                         need_timer = 1;
 496         }
 497         tcp_tw_death_row_slot =
 498                 ((tcp_tw_death_row_slot + 1) & (TCP_TWKILL_SLOTS - 1));
 499         if (need_timer)
 500                 mod_timer(&tcp_tw_timer, jiffies + TCP_TWKILL_PERIOD);
 501 out:
 502         spin_unlock(&tw_death_lock);
 503 }
 504
 505 extern void twkill_slots_invalid(void);
 506
 507 static void twkill_work(void *dummy)
 508 {
 509         int i;
 510
 511         if ((TCP_TWKILL_SLOTS - 1) > (sizeof(twkill_thread_slots) * 8))
 512                 twkill_slots_invalid();
 513
 514         while (twkill_thread_slots) {
 515                 spin_lock_bh(&tw_death_lock);
 516                 for (i = 0; i < TCP_TWKILL_SLOTS; i++) {
 517                         if (!(twkill_thread_slots & (1 << i)))
 518                                 continue;
 519
 520                         while (tcp_do_twkill_work(i, TCP_TWKILL_QUOTA) != 0) {
 521                                 if (need_resched()) {
 522                                         spin_unlock_bh(&tw_death_lock);
 523                                         schedule();
 524                                         spin_lock_bh(&tw_death_lock);
 525                                 }
 526                         }
 527
 528                         twkill_thread_slots &= ~(1 << i);
 529                 }
 530                 spin_unlock_bh(&tw_death_lock);
 531         }
 532 }
 533
 534 /* These are always called from BH context.  See callers in
 535  * tcp_input.c to verify this.
 536  */
 537
 538 /* This is for handling early-kills of TIME_WAIT sockets. */
 539 void tcp_tw_deschedule(struct tcp_tw_bucket *tw)
 540 {
 541         spin_lock(&tw_death_lock);
 542         if (tw_del_dead_node(tw)) {
 543                 tcp_tw_put(tw);
 544                 if (--tcp_tw_count == 0)
 545                         del_timer(&tcp_tw_timer);
 546         }
 547         spin_unlock(&tw_death_lock);
 548         tcp_timewait_kill(tw);
 549 }
 550
 551 /* Short-time timewait calendar */
 552
 553 static int tcp_twcal_hand = -1;
 554 static int tcp_twcal_jiffie;
 555 static void tcp_twcal_tick(unsigned long);
 556 static struct timer_list tcp_twcal_timer =
 557                 TIMER_INITIALIZER(tcp_twcal_tick, 0, 0);
 558 static struct hlist_head tcp_twcal_row[TCP_TW_RECYCLE_SLOTS];
 559
 560 void tcp_tw_schedule(struct tcp_tw_bucket *tw, int timeo)
 561 {
 562         struct hlist_head *list;
 563         int slot;
 564
 565         /* timeout := RTO * 3.5
 566          *
 567          * 3.5 = 1+2+0.5 to wait for two retransmits.
 568          *
 569          * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
 570          * our ACK acking that FIN can be lost. If N subsequent retransmitted
 571          * FINs (or previous seqments) are lost (probability of such event
 572          * is p^(N+1), where p is probability to lose single packet and
 573          * time to detect the loss is about RTO*(2^N - 1) with exponential
 574          * backoff). Normal timewait length is calculated so, that we
 575          * waited at least for one retransmitted FIN (maximal RTO is 120sec).
 576          * [ BTW Linux. following BSD, violates this requirement waiting
 577          *   only for 60sec, we should wait at least for 240 secs.
 578          *   Well, 240 consumes too much of resources 8)
 579          * ]
 580          * This interval is not reduced to catch old duplicate and
 581          * responces to our wandering segments living for two MSLs.
 582          * However, if we use PAWS to detect
 583          * old duplicates, we can reduce the interval to bounds required
 584          * by RTO, rather than MSL. So, if peer understands PAWS, we
 585          * kill tw bucket after 3.5*RTO (it is important that this number
 586          * is greater than TS tick!) and detect old duplicates with help
 587          * of PAWS.
 588          */
 589         slot = (timeo + (1<<TCP_TW_RECYCLE_TICK) - 1) >> TCP_TW_RECYCLE_TICK;
 590
 591         spin_lock(&tw_death_lock);
 592
 593         /* Unlink it, if it was scheduled */
 594         if (tw_del_dead_node(tw))
 595                 tcp_tw_count--;
 596         else
 597                 atomic_inc(&tw->tw_refcnt);
 598
 599         if (slot >= TCP_TW_RECYCLE_SLOTS) {
 600                 /* Schedule to slow timer */
 601                 if (timeo >= TCP_TIMEWAIT_LEN) {
 602                         slot = TCP_TWKILL_SLOTS-1;
 603                 } else {
 604                         slot = (timeo + TCP_TWKILL_PERIOD-1) / TCP_TWKILL_PERIOD;
 605                         if (slot >= TCP_TWKILL_SLOTS)
 606                                 slot = TCP_TWKILL_SLOTS-1;
 607                 }
 608                 tw->tw_ttd = jiffies + timeo;
 609                 slot = (tcp_tw_death_row_slot + slot) & (TCP_TWKILL_SLOTS - 1);
 610                 list = &tcp_tw_death_row[slot];
 611         } else {
 612                 tw->tw_ttd = jiffies + (slot << TCP_TW_RECYCLE_TICK);
 613
 614                 if (tcp_twcal_hand < 0) {
 615                         tcp_twcal_hand = 0;
 616                         tcp_twcal_jiffie = jiffies;
 617                         tcp_twcal_timer.expires = tcp_twcal_jiffie + (slot<<TCP_TW_RECYCLE_TICK);
 618                         add_timer(&tcp_twcal_timer);
 619                 } else {
 620                         if (time_after(tcp_twcal_timer.expires, jiffies + (slot<<TCP_TW_RECYCLE_TICK)))
 621                                 mod_timer(&tcp_twcal_timer, jiffies + (slot<<TCP_TW_RECYCLE_TICK));
 622                         slot = (tcp_twcal_hand + slot)&(TCP_TW_RECYCLE_SLOTS-1);
 623                 }
 624                 list = &tcp_twcal_row[slot];
 625         }
 626
 627         hlist_add_head(&tw->tw_death_node, list);
 628
 629         if (tcp_tw_count++ == 0)
 630                 mod_timer(&tcp_tw_timer, jiffies+TCP_TWKILL_PERIOD);
 631         spin_unlock(&tw_death_lock);
 632 }
 633
 634 void tcp_twcal_tick(unsigned long dummy)
 635 {
 636         int n, slot;
 637         unsigned long j;
 638         unsigned long now = jiffies;
 639         int killed = 0;
 640         int adv = 0;
 641
 642         spin_lock(&tw_death_lock);
 643         if (tcp_twcal_hand < 0)
 644                 goto out;
 645
 646         slot = tcp_twcal_hand;
 647         j = tcp_twcal_jiffie;
 648
 649         for (n=0; n<TCP_TW_RECYCLE_SLOTS; n++) {
 650                 if (time_before_eq(j, now)) {
 651                         struct hlist_node *node, *safe;
 652                         struct tcp_tw_bucket *tw;
 653
 654                         tw_for_each_inmate_safe(tw, node, safe,
 655                                            &tcp_twcal_row[slot]) {
 656                                 __tw_del_dead_node(tw);
 657                                 tcp_timewait_kill(tw);
 658                                 tcp_tw_put(tw);
 659                                 killed++;
 660                         }
 661                 } else {
 662                         if (!adv) {
 663                                 adv = 1;
 664                                 tcp_twcal_jiffie = j;
 665                                 tcp_twcal_hand = slot;
 666                         }
 667
 668                         if (!hlist_empty(&tcp_twcal_row[slot])) {
 669                                 mod_timer(&tcp_twcal_timer, j);
 670                                 goto out;
 671                         }
 672                 }
 673                 j += (1<<TCP_TW_RECYCLE_TICK);
 674                 slot = (slot+1)&(TCP_TW_RECYCLE_SLOTS-1);
 675         }
 676         tcp_twcal_hand = -1;
 677
 678 out:
 679         if ((tcp_tw_count -= killed) == 0)
 680                 del_timer(&tcp_tw_timer);
 681         NET_ADD_STATS_BH(TimeWaitKilled, killed);
 682         spin_unlock(&tw_death_lock);
 683 }
 684
 685 /* This is not only more efficient than what we used to do, it eliminates
 686  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
 687  *
 688  * Actually, we could lots of memory writes here. tp of listening
 689  * socket contains all necessary default parameters.
 690  */
 691 struct sock *tcp_create_openreq_child(struct sock *sk, struct open_request *req, struct sk_buff *skb)
 692 {
 693         /* allocate the newsk from the same slab of the master sock,
 694          * if not, at sk_free time we'll try to free it from the wrong
 695          * slabcache (i.e. is it TCPv4 or v6?) -acme */
 696         struct sock *newsk = sk_alloc(PF_INET, GFP_ATOMIC, 0, sk->sk_slab);
 697
 698         if(newsk != NULL) {
 699                 struct tcp_opt *newtp;
 700                 struct sk_filter *filter;
 701
 702                 memcpy(newsk, sk, sizeof(struct tcp_sock));
 703                 newsk->sk_state = TCP_SYN_RECV;
 704
 705                 /* SANITY */
 706                 sock_vx_init(newsk);
 707                 sock_nx_init(newsk);
 708                 sk_node_init(&newsk->sk_node);
 709                 tcp_sk(newsk)->bind_hash = NULL;
 710
 711                 /* Clone the TCP header template */
 712                 inet_sk(newsk)->dport = req->rmt_port;
 713
 714                 sock_lock_init(newsk);
 715                 bh_lock_sock(newsk);
 716
 717                 newsk->sk_dst_lock = RW_LOCK_UNLOCKED;
 718                 atomic_set(&newsk->sk_rmem_alloc, 0);
 719                 skb_queue_head_init(&newsk->sk_receive_queue);
 720                 atomic_set(&newsk->sk_wmem_alloc, 0);
 721                 skb_queue_head_init(&newsk->sk_write_queue);
 722                 atomic_set(&newsk->sk_omem_alloc, 0);
 723                 newsk->sk_wmem_queued = 0;
 724                 newsk->sk_forward_alloc = 0;
 725
 726                 sock_reset_flag(newsk, SOCK_DONE);
 727                 newsk->sk_userlocks = sk->sk_userlocks & ~SOCK_BINDPORT_LOCK;
 728                 newsk->sk_backlog.head = newsk->sk_backlog.tail = NULL;
 729                 newsk->sk_send_head = NULL;
 730                 newsk->sk_callback_lock = RW_LOCK_UNLOCKED;
 731                 skb_queue_head_init(&newsk->sk_error_queue);
 732                 newsk->sk_write_space = sk_stream_write_space;
 733
 734                 if ((filter = newsk->sk_filter) != NULL)
 735                         sk_filter_charge(newsk, filter);
 736
 737                 if (unlikely(xfrm_sk_clone_policy(newsk))) {
 738                         /* It is still raw copy of parent, so invalidate
 739                          * destructor and make plain sk_free() */
 740                         newsk->sk_destruct = NULL;
 741                         sk_free(newsk);
 742                         return NULL;
 743                 }
 744
 745                 /* Now setup tcp_opt */
 746                 newtp = tcp_sk(newsk);
 747                 newtp->pred_flags = 0;
 748                 newtp->rcv_nxt = req->rcv_isn + 1;
 749                 newtp->snd_nxt = req->snt_isn + 1;
 750                 newtp->snd_una = req->snt_isn + 1;
 751                 newtp->snd_sml = req->snt_isn + 1;
 752
 753                 tcp_prequeue_init(newtp);
 754
 755                 tcp_init_wl(newtp, req->snt_isn, req->rcv_isn);
 756
 757                 newtp->retransmits = 0;
 758                 newtp->backoff = 0;
 759                 newtp->srtt = 0;
 760                 newtp->mdev = TCP_TIMEOUT_INIT;
 761                 newtp->rto = TCP_TIMEOUT_INIT;
 762
 763                 newtp->packets_out = 0;
 764                 newtp->left_out = 0;
 765                 newtp->retrans_out = 0;
 766                 newtp->sacked_out = 0;
 767                 newtp->fackets_out = 0;
 768                 newtp->snd_ssthresh = 0x7fffffff;
 769
 770                 /* So many TCP implementations out there (incorrectly) count the
 771                  * initial SYN frame in their delayed-ACK and congestion control
 772                  * algorithms that we must have the following bandaid to talk
 773                  * efficiently to them.  -DaveM
 774                  */
 775                 newtp->snd_cwnd = 2;
 776                 newtp->snd_cwnd_cnt = 0;
 777
 778                 newtp->bictcp.cnt = 0;
 779                 newtp->bictcp.last_max_cwnd = newtp->bictcp.last_cwnd = 0;
 780
 781                 newtp->frto_counter = 0;
 782                 newtp->frto_highmark = 0;
 783
 784                 tcp_set_ca_state(newtp, TCP_CA_Open);
 785                 tcp_init_xmit_timers(newsk);
 786                 skb_queue_head_init(&newtp->out_of_order_queue);
 787                 newtp->rcv_wup = req->rcv_isn + 1;
 788                 newtp->write_seq = req->snt_isn + 1;
 789                 newtp->pushed_seq = newtp->write_seq;
 790                 newtp->copied_seq = req->rcv_isn + 1;
 791
 792                 newtp->saw_tstamp = 0;
 793
 794                 newtp->dsack = 0;
 795                 newtp->eff_sacks = 0;
 796
 797                 newtp->probes_out = 0;
 798                 newtp->num_sacks = 0;
 799                 newtp->urg_data = 0;
 800                 newtp->listen_opt = NULL;
 801 #ifdef CONFIG_ACCEPT_QUEUES
 802                 newtp->accept_queue = NULL;
 803                 memset(newtp->acceptq, 0,sizeof(newtp->acceptq));
 804                 newtp->class_index = 0;
 805
 806 #else
 807                 newtp->accept_queue = newtp->accept_queue_tail = NULL;
 808 #endif
 809                 /* Deinitialize syn_wait_lock to trap illegal accesses. */
 810                 memset(&newtp->syn_wait_lock, 0, sizeof(newtp->syn_wait_lock));
 811
 812                 /* Back to base struct sock members. */
 813                 newsk->sk_err = 0;
 814                 newsk->sk_priority = 0;
 815                 atomic_set(&newsk->sk_refcnt, 2);
 816
 817                 set_vx_info(&newsk->sk_vx_info, sk->sk_vx_info);
 818                 newsk->sk_xid = sk->sk_xid;
 819                 set_nx_info(&newsk->sk_nx_info, sk->sk_nx_info);
 820                 newsk->sk_nid = sk->sk_nid;
 821 #ifdef INET_REFCNT_DEBUG
 822                 atomic_inc(&inet_sock_nr);
 823 #endif
 824                 atomic_inc(&tcp_sockets_allocated);
 825
 826                 if (sock_flag(newsk, SOCK_KEEPOPEN))
 827                         tcp_reset_keepalive_timer(newsk,
 828                                                   keepalive_time_when(newtp));
 829                 newsk->sk_socket = NULL;
 830                 newsk->sk_sleep = NULL;
 831                 newsk->sk_owner = NULL;
 832
 833                 newtp->tstamp_ok = req->tstamp_ok;
 834                 if((newtp->sack_ok = req->sack_ok) != 0) {
 835                         if (sysctl_tcp_fack)
 836                                 newtp->sack_ok |= 2;
 837                 }
 838                 newtp->window_clamp = req->window_clamp;
 839                 newtp->rcv_ssthresh = req->rcv_wnd;
 840                 newtp->rcv_wnd = req->rcv_wnd;
 841                 newtp->wscale_ok = req->wscale_ok;
 842                 if (newtp->wscale_ok) {
 843                         newtp->snd_wscale = req->snd_wscale;
 844                         newtp->rcv_wscale = req->rcv_wscale;
 845                 } else {
 846                         newtp->snd_wscale = newtp->rcv_wscale = 0;
 847                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
 848                 }
 849                 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->snd_wscale;
 850                 newtp->max_window = newtp->snd_wnd;
 851
 852                 if (newtp->tstamp_ok) {
 853                         newtp->ts_recent = req->ts_recent;
 854                         newtp->ts_recent_stamp = xtime.tv_sec;
 855                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
 856                 } else {
 857                         newtp->ts_recent_stamp = 0;
 858                         newtp->tcp_header_len = sizeof(struct tcphdr);
 859                 }
 860                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
 861                         newtp->ack.last_seg_size = skb->len-newtp->tcp_header_len;
 862                 newtp->mss_clamp = req->mss;
 863                 TCP_ECN_openreq_child(newtp, req);
 864                 if (newtp->ecn_flags&TCP_ECN_OK)
 865                         newsk->sk_no_largesend = 1;
 866
 867                 tcp_vegas_init(newtp);
 868                 TCP_INC_STATS_BH(TcpPassiveOpens);
 869         }
 870         return newsk;
 871 }
 872
 873 /*
 874  *      Process an incoming packet for SYN_RECV sockets represented
 875  *      as an open_request.
 876  */
 877
 878 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
 879                            struct open_request *req,
 880                            struct open_request **prev)
 881 {
 882         struct tcphdr *th = skb->h.th;
 883         struct tcp_opt *tp = tcp_sk(sk);
 884         u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
 885         int paws_reject = 0;
 886         struct tcp_opt ttp;
 887         struct sock *child;
 888
 889         ttp.saw_tstamp = 0;
 890         if (th->doff > (sizeof(struct tcphdr)>>2)) {
 891                 tcp_parse_options(skb, &ttp, 0);
 892
 893                 if (ttp.saw_tstamp) {
 894                         ttp.ts_recent = req->ts_recent;
 895                         /* We do not store true stamp, but it is not required,
 896                          * it can be estimated (approximately)
 897                          * from another data.
 898                          */
 899                         ttp.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
 900                         paws_reject = tcp_paws_check(&ttp, th->rst);
 901                 }
 902         }
 903
 904         /* Check for pure retransmitted SYN. */
 905         if (TCP_SKB_CB(skb)->seq == req->rcv_isn &&
 906             flg == TCP_FLAG_SYN &&
 907             !paws_reject) {
 908                 /*
 909                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
 910                  * this case on figure 6 and figure 8, but formal
 911                  * protocol description says NOTHING.
 912                  * To be more exact, it says that we should send ACK,
 913                  * because this segment (at least, if it has no data)
 914                  * is out of window.
 915                  *
 916                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
 917                  *  describe SYN-RECV state. All the description
 918                  *  is wrong, we cannot believe to it and should
 919                  *  rely only on common sense and implementation
 920                  *  experience.
 921                  *
 922                  * Enforce "SYN-ACK" according to figure 8, figure 6
 923                  * of RFC793, fixed by RFC1122.
 924                  */
 925                 req->class->rtx_syn_ack(sk, req, NULL);
 926                 return NULL;
 927         }
 928
 929         /* Further reproduces section "SEGMENT ARRIVES"
 930            for state SYN-RECEIVED of RFC793.
 931            It is broken, however, it does not work only
 932            when SYNs are crossed.
 933
 934            You would think that SYN crossing is impossible here, since
 935            we should have a SYN_SENT socket (from connect()) on our end,
 936            but this is not true if the crossed SYNs were sent to both
 937            ends by a malicious third party.  We must defend against this,
 938            and to do that we first verify the ACK (as per RFC793, page
 939            36) and reset if it is invalid.  Is this a true full defense?
 940            To convince ourselves, let us consider a way in which the ACK
 941            test can still pass in this 'malicious crossed SYNs' case.
 942            Malicious sender sends identical SYNs (and thus identical sequence
 943            numbers) to both A and B:
 944
 945                 A: gets SYN, seq=7
 946                 B: gets SYN, seq=7
 947
 948            By our good fortune, both A and B select the same initial
 949            send sequence number of seven :-)
 950
 951                 A: sends SYN|ACK, seq=7, ack_seq=8
 952                 B: sends SYN|ACK, seq=7, ack_seq=8
 953
 954            So we are now A eating this SYN|ACK, ACK test passes.  So
 955            does sequence test, SYN is truncated, and thus we consider
 956            it a bare ACK.
 957
 958            If tp->defer_accept, we silently drop this bare ACK.  Otherwise,
 959            we create an established connection.  Both ends (listening sockets)
 960            accept the new incoming connection and try to talk to each other. 8-)
 961
 962            Note: This case is both harmless, and rare.  Possibility is about the
 963            same as us discovering intelligent life on another plant tomorrow.
 964
 965            But generally, we should (RFC lies!) to accept ACK
 966            from SYNACK both here and in tcp_rcv_state_process().
 967            tcp_rcv_state_process() does not, hence, we do not too.
 968
 969            Note that the case is absolutely generic:
 970            we cannot optimize anything here without
 971            violating protocol. All the checks must be made
 972            before attempt to create socket.
 973          */
 974
 975         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
 976          *                  and the incoming segment acknowledges something not yet
 977          *                  sent (the segment carries an unaccaptable ACK) ...
 978          *                  a reset is sent."
 979          *
 980          * Invalid ACK: reset will be sent by listening socket
 981          */
 982         if ((flg & TCP_FLAG_ACK) &&
 983             (TCP_SKB_CB(skb)->ack_seq != req->snt_isn+1))
 984                 return sk;
 985
 986         /* Also, it would be not so bad idea to check rcv_tsecr, which
 987          * is essentially ACK extension and too early or too late values
 988          * should cause reset in unsynchronized states.
 989          */
 990
 991         /* RFC793: "first check sequence number". */
 992
 993         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
 994                                           req->rcv_isn+1, req->rcv_isn+1+req->rcv_wnd)) {
 995                 /* Out of window: send ACK and drop. */
 996                 if (!(flg & TCP_FLAG_RST))
 997                         req->class->send_ack(skb, req);
 998                 if (paws_reject)
 999                         NET_INC_STATS_BH(PAWSEstabRejected);
1000                 return NULL;
1001         }
1002
1003         /* In sequence, PAWS is OK. */
1004
1005         if (ttp.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, req->rcv_isn+1))
1006                 req->ts_recent = ttp.rcv_tsval;
1007
1008         if (TCP_SKB_CB(skb)->seq == req->rcv_isn) {
1009                 /* Truncate SYN, it is out of window starting
1010                    at req->rcv_isn+1. */
1011                 flg &= ~TCP_FLAG_SYN;
1012         }
1013
1014         /* RFC793: "second check the RST bit" and
1015          *         "fourth, check the SYN bit"
1016          */
1017         if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN))
1018                 goto embryonic_reset;
1019
1020         /* ACK sequence verified above, just make sure ACK is
1021          * set.  If ACK not set, just silently drop the packet.
1022          */
1023         if (!(flg & TCP_FLAG_ACK))
1024                 return NULL;
1025
1026         /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
1027         if (tp->defer_accept && TCP_SKB_CB(skb)->end_seq == req->rcv_isn+1) {
1028                 req->acked = 1;
1029                 return NULL;
1030         }
1031
1032         /* OK, ACK is valid, create big socket and
1033          * feed this segment to it. It will repeat all
1034          * the tests. THIS SEGMENT MUST MOVE SOCKET TO
1035          * ESTABLISHED STATE. If it will be dropped after
1036          * socket is created, wait for troubles.
1037          */
1038         child = tp->af_specific->syn_recv_sock(sk, skb, req, NULL);
1039         if (child == NULL)
1040                 goto listen_overflow;
1041
1042         sk_set_owner(child, sk->sk_owner);
1043         tcp_synq_unlink(tp, req, prev);
1044         tcp_synq_removed(sk, req);
1045
1046         tcp_acceptq_queue(sk, req, child);
1047         return child;
1048
1049 listen_overflow:
1050         if (!sysctl_tcp_abort_on_overflow) {
1051                 req->acked = 1;
1052                 return NULL;
1053         }
1054
1055 embryonic_reset:
1056         NET_INC_STATS_BH(EmbryonicRsts);
1057         if (!(flg & TCP_FLAG_RST))
1058                 req->class->send_reset(skb);
1059
1060         tcp_synq_drop(sk, req, prev);
1061         return NULL;
1062 }
1063
1064 /*
1065  * Queue segment on the new socket if the new socket is active,
1066  * otherwise we just shortcircuit this and continue with
1067  * the new socket.
1068  */
1069
1070 int tcp_child_process(struct sock *parent, struct sock *child,
1071                       struct sk_buff *skb)
1072 {
1073         int ret = 0;
1074         int state = child->sk_state;
1075
1076         if (!sock_owned_by_user(child)) {
1077                 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
1078
1079                 /* Wakeup parent, send SIGIO */
1080                 if (state == TCP_SYN_RECV && child->sk_state != state)
1081                         parent->sk_data_ready(parent, 0);
1082         } else {
1083                 /* Alas, it is possible again, because we do lookup
1084                  * in main socket hash table and lock on listening
1085                  * socket does not protect us more.
1086                  */
1087                 sk_add_backlog(child, skb);
1088         }
1089
1090         bh_unlock_sock(child);
1091         sock_put(child);
1092         return ret;
1093 }
1094
1095 EXPORT_SYMBOL(tcp_check_req);
1096 EXPORT_SYMBOL(tcp_child_process);
1097 EXPORT_SYMBOL(tcp_create_openreq_child);
1098 EXPORT_SYMBOL(tcp_timewait_state_process);
1099 EXPORT_SYMBOL(tcp_tw_deschedule);
1100
1101 #ifdef CONFIG_SYSCTL
1102 EXPORT_SYMBOL(sysctl_tcp_tw_recycle);
1103 #endif