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