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