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
  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/mm.h>
  24 #include <linux/module.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/workqueue.h>
  27 #include <net/tcp.h>
  28 #include <net/inet_common.h>
  29 #include <net/xfrm.h>
  30
  31 #ifdef CONFIG_SYSCTL
  32 #define SYNC_INIT 0 /* let the user enable it */
  33 #else
  34 #define SYNC_INIT 1
  35 #endif
  36
  37 int sysctl_tcp_syncookies __read_mostly = SYNC_INIT;
  38 int sysctl_tcp_abort_on_overflow __read_mostly;
  39
  40 struct inet_timewait_death_row tcp_death_row = {
  41         .sysctl_max_tw_buckets = NR_FILE * 2,
  42         .period         = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS,
  43         .death_lock     = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock),
  44         .hashinfo       = &tcp_hashinfo,
  45         .tw_timer       = TIMER_INITIALIZER(inet_twdr_hangman, 0,
  46                                             (unsigned long)&tcp_death_row),
  47         .twkill_work    = __WORK_INITIALIZER(tcp_death_row.twkill_work,
  48                                              inet_twdr_twkill_work,
  49                                              &tcp_death_row),
  50 /* Short-time timewait calendar */
  51
  52         .twcal_hand     = -1,
  53         .twcal_timer    = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0,
  54                                             (unsigned long)&tcp_death_row),
  55 };
  56
  57 EXPORT_SYMBOL_GPL(tcp_death_row);
  58
  59 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win)
  60 {
  61         if (seq == s_win)
  62                 return 1;
  63         if (after(end_seq, s_win) && before(seq, e_win))
  64                 return 1;
  65         return (seq == e_win && seq == end_seq);
  66 }
  67
  68 /*
  69  * * Main purpose of TIME-WAIT state is to close connection gracefully,
  70  *   when one of ends sits in LAST-ACK or CLOSING retransmitting FIN
  71  *   (and, probably, tail of data) and one or more our ACKs are lost.
  72  * * What is TIME-WAIT timeout? It is associated with maximal packet
  73  *   lifetime in the internet, which results in wrong conclusion, that
  74  *   it is set to catch "old duplicate segments" wandering out of their path.
  75  *   It is not quite correct. This timeout is calculated so that it exceeds
  76  *   maximal retransmission timeout enough to allow to lose one (or more)
  77  *   segments sent by peer and our ACKs. This time may be calculated from RTO.
  78  * * When TIME-WAIT socket receives RST, it means that another end
  79  *   finally closed and we are allowed to kill TIME-WAIT too.
  80  * * Second purpose of TIME-WAIT is catching old duplicate segments.
  81  *   Well, certainly it is pure paranoia, but if we load TIME-WAIT
  82  *   with this semantics, we MUST NOT kill TIME-WAIT state with RSTs.
  83  * * If we invented some more clever way to catch duplicates
  84  *   (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs.
  85  *
  86  * The algorithm below is based on FORMAL INTERPRETATION of RFCs.
  87  * When you compare it to RFCs, please, read section SEGMENT ARRIVES
  88  * from the very beginning.
  89  *
  90  * NOTE. With recycling (and later with fin-wait-2) TW bucket
  91  * is _not_ stateless. It means, that strictly speaking we must
  92  * spinlock it. I do not want! Well, probability of misbehaviour
  93  * is ridiculously low and, seems, we could use some mb() tricks
  94  * to avoid misread sequence numbers, states etc.  --ANK
  95  */
  96 enum tcp_tw_status
  97 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb,
  98                            const struct tcphdr *th)
  99 {
 100         struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
 101         struct tcp_options_received tmp_opt;
 102         int paws_reject = 0;
 103
 104         tmp_opt.saw_tstamp = 0;
 105         if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) {
 106                 tcp_parse_options(skb, &tmp_opt, 0);
 107
 108                 if (tmp_opt.saw_tstamp) {
 109                         tmp_opt.ts_recent       = tcptw->tw_ts_recent;
 110                         tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
 111                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
 112                 }
 113         }
 114
 115         if (tw->tw_substate == TCP_FIN_WAIT2) {
 116                 /* Just repeat all the checks of tcp_rcv_state_process() */
 117
 118                 /* Out of window, send ACK */
 119                 if (paws_reject ||
 120                     !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
 121                                    tcptw->tw_rcv_nxt,
 122                                    tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd))
 123                         return TCP_TW_ACK;
 124
 125                 if (th->rst)
 126                         goto kill;
 127
 128                 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt))
 129                         goto kill_with_rst;
 130
 131                 /* Dup ACK? */
 132                 if (!after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) ||
 133                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) {
 134                         inet_twsk_put(tw);
 135                         return TCP_TW_SUCCESS;
 136                 }
 137
 138                 /* New data or FIN. If new data arrive after half-duplex close,
 139                  * reset.
 140                  */
 141                 if (!th->fin ||
 142                     TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) {
 143 kill_with_rst:
 144                         inet_twsk_deschedule(tw, &tcp_death_row);
 145                         inet_twsk_put(tw);
 146                         return TCP_TW_RST;
 147                 }
 148
 149                 /* FIN arrived, enter true time-wait state. */
 150                 tw->tw_substate   = TCP_TIME_WAIT;
 151                 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq;
 152                 if (tmp_opt.saw_tstamp) {
 153                         tcptw->tw_ts_recent_stamp = xtime.tv_sec;
 154                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
 155                 }
 156
 157                 /* I am shamed, but failed to make it more elegant.
 158                  * Yes, it is direct reference to IP, which is impossible
 159                  * to generalize to IPv6. Taking into account that IPv6
 160                  * do not understand recycling in any case, it not
 161                  * a big problem in practice. --ANK */
 162                 if (tw->tw_family == AF_INET &&
 163                     tcp_death_row.sysctl_tw_recycle && tcptw->tw_ts_recent_stamp &&
 164                     tcp_v4_tw_remember_stamp(tw))
 165                         inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout,
 166                                            TCP_TIMEWAIT_LEN);
 167                 else
 168                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
 169                                            TCP_TIMEWAIT_LEN);
 170                 return TCP_TW_ACK;
 171         }
 172
 173         /*
 174          *      Now real TIME-WAIT state.
 175          *
 176          *      RFC 1122:
 177          *      "When a connection is [...] on TIME-WAIT state [...]
 178          *      [a TCP] MAY accept a new SYN from the remote TCP to
 179          *      reopen the connection directly, if it:
 180          *
 181          *      (1)  assigns its initial sequence number for the new
 182          *      connection to be larger than the largest sequence
 183          *      number it used on the previous connection incarnation,
 184          *      and
 185          *
 186          *      (2)  returns to TIME-WAIT state if the SYN turns out
 187          *      to be an old duplicate".
 188          */
 189
 190         if (!paws_reject &&
 191             (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt &&
 192              (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) {
 193                 /* In window segment, it may be only reset or bare ack. */
 194
 195                 if (th->rst) {
 196                         /* This is TIME_WAIT assassination, in two flavors.
 197                          * Oh well... nobody has a sufficient solution to this
 198                          * protocol bug yet.
 199                          */
 200                         if (sysctl_tcp_rfc1337 == 0) {
 201 kill:
 202                                 inet_twsk_deschedule(tw, &tcp_death_row);
 203                                 inet_twsk_put(tw);
 204                                 return TCP_TW_SUCCESS;
 205                         }
 206                 }
 207                 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
 208                                    TCP_TIMEWAIT_LEN);
 209
 210                 if (tmp_opt.saw_tstamp) {
 211                         tcptw->tw_ts_recent       = tmp_opt.rcv_tsval;
 212                         tcptw->tw_ts_recent_stamp = xtime.tv_sec;
 213                 }
 214
 215                 inet_twsk_put(tw);
 216                 return TCP_TW_SUCCESS;
 217         }
 218
 219         /* Out of window segment.
 220
 221            All the segments are ACKed immediately.
 222
 223            The only exception is new SYN. We accept it, if it is
 224            not old duplicate and we are not in danger to be killed
 225            by delayed old duplicates. RFC check is that it has
 226            newer sequence number works at rates <40Mbit/sec.
 227            However, if paws works, it is reliable AND even more,
 228            we even may relax silly seq space cutoff.
 229
 230            RED-PEN: we violate main RFC requirement, if this SYN will appear
 231            old duplicate (i.e. we receive RST in reply to SYN-ACK),
 232            we must return socket to time-wait state. It is not good,
 233            but not fatal yet.
 234          */
 235
 236         if (th->syn && !th->rst && !th->ack && !paws_reject &&
 237             (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) ||
 238              (tmp_opt.saw_tstamp &&
 239               (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) {
 240                 u32 isn = tcptw->tw_snd_nxt + 65535 + 2;
 241                 if (isn == 0)
 242                         isn++;
 243                 TCP_SKB_CB(skb)->when = isn;
 244                 return TCP_TW_SYN;
 245         }
 246
 247         if (paws_reject)
 248                 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
 249
 250         if(!th->rst) {
 251                 /* In this case we must reset the TIMEWAIT timer.
 252                  *
 253                  * If it is ACKless SYN it may be both old duplicate
 254                  * and new good SYN with random sequence number <rcv_nxt.
 255                  * Do not reschedule in the last case.
 256                  */
 257                 if (paws_reject || th->ack)
 258                         inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN,
 259                                            TCP_TIMEWAIT_LEN);
 260
 261                 /* Send ACK. Note, we do not put the bucket,
 262                  * it will be released by caller.
 263                  */
 264                 return TCP_TW_ACK;
 265         }
 266         inet_twsk_put(tw);
 267         return TCP_TW_SUCCESS;
 268 }
 269
 270 /*
 271  * Move a socket to time-wait or dead fin-wait-2 state.
 272  */
 273 void tcp_time_wait(struct sock *sk, int state, int timeo)
 274 {
 275         struct inet_timewait_sock *tw = NULL;
 276         const struct inet_connection_sock *icsk = inet_csk(sk);
 277         const struct tcp_sock *tp = tcp_sk(sk);
 278         int recycle_ok = 0;
 279
 280         if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
 281                 recycle_ok = icsk->icsk_af_ops->remember_stamp(sk);
 282
 283         if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
 284                 tw = inet_twsk_alloc(sk, state);
 285
 286         if (tw != NULL) {
 287                 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
 288                 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
 289
 290                 tw->tw_rcv_wscale       = tp->rx_opt.rcv_wscale;
 291                 tcptw->tw_rcv_nxt       = tp->rcv_nxt;
 292                 tcptw->tw_snd_nxt       = tp->snd_nxt;
 293                 tcptw->tw_rcv_wnd       = tcp_receive_window(tp);
 294                 tcptw->tw_ts_recent     = tp->rx_opt.ts_recent;
 295                 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
 296
 297 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 298                 if (tw->tw_family == PF_INET6) {
 299                         struct ipv6_pinfo *np = inet6_sk(sk);
 300                         struct inet6_timewait_sock *tw6;
 301
 302                         tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
 303                         tw6 = inet6_twsk((struct sock *)tw);
 304                         ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
 305                         ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
 306                         tw->tw_ipv6only = np->ipv6only;
 307                 }
 308 #endif
 309                 /* Linkage updates. */
 310                 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
 311
 312                 /* Get the TIME_WAIT timeout firing. */
 313                 if (timeo < rto)
 314                         timeo = rto;
 315
 316                 if (recycle_ok) {
 317                         tw->tw_timeout = rto;
 318                 } else {
 319                         tw->tw_timeout = TCP_TIMEWAIT_LEN;
 320                         if (state == TCP_TIME_WAIT)
 321                                 timeo = TCP_TIMEWAIT_LEN;
 322                 }
 323
 324                 inet_twsk_schedule(tw, &tcp_death_row, timeo,
 325                                    TCP_TIMEWAIT_LEN);
 326                 inet_twsk_put(tw);
 327         } else {
 328                 /* Sorry, if we're out of memory, just CLOSE this
 329                  * socket up.  We've got bigger problems than
 330                  * non-graceful socket closings.
 331                  */
 332                 if (net_ratelimit())
 333                         printk(KERN_INFO "TCP: time wait bucket table overflow\n");
 334         }
 335
 336         tcp_update_metrics(sk);
 337         tcp_done(sk);
 338 }
 339
 340 /* This is not only more efficient than what we used to do, it eliminates
 341  * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM
 342  *
 343  * Actually, we could lots of memory writes here. tp of listening
 344  * socket contains all necessary default parameters.
 345  */
 346 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb)
 347 {
 348         struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC);
 349
 350         if (newsk != NULL) {
 351                 const struct inet_request_sock *ireq = inet_rsk(req);
 352                 struct tcp_request_sock *treq = tcp_rsk(req);
 353                 struct inet_connection_sock *newicsk = inet_csk(sk);
 354                 struct tcp_sock *newtp;
 355
 356                 /* Now setup tcp_sock */
 357                 newtp = tcp_sk(newsk);
 358                 newtp->pred_flags = 0;
 359                 newtp->rcv_nxt = treq->rcv_isn + 1;
 360                 newtp->snd_nxt = newtp->snd_una = newtp->snd_sml = treq->snt_isn + 1;
 361
 362                 tcp_prequeue_init(newtp);
 363
 364                 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn);
 365
 366                 newtp->srtt = 0;
 367                 newtp->mdev = TCP_TIMEOUT_INIT;
 368                 newicsk->icsk_rto = TCP_TIMEOUT_INIT;
 369
 370                 newtp->packets_out = 0;
 371                 newtp->left_out = 0;
 372                 newtp->retrans_out = 0;
 373                 newtp->sacked_out = 0;
 374                 newtp->fackets_out = 0;
 375                 newtp->snd_ssthresh = 0x7fffffff;
 376
 377                 /* So many TCP implementations out there (incorrectly) count the
 378                  * initial SYN frame in their delayed-ACK and congestion control
 379                  * algorithms that we must have the following bandaid to talk
 380                  * efficiently to them.  -DaveM
 381                  */
 382                 newtp->snd_cwnd = 2;
 383                 newtp->snd_cwnd_cnt = 0;
 384                 newtp->bytes_acked = 0;
 385
 386                 newtp->frto_counter = 0;
 387                 newtp->frto_highmark = 0;
 388
 389                 newicsk->icsk_ca_ops = &tcp_init_congestion_ops;
 390
 391                 tcp_set_ca_state(newsk, TCP_CA_Open);
 392                 tcp_init_xmit_timers(newsk);
 393                 skb_queue_head_init(&newtp->out_of_order_queue);
 394                 newtp->rcv_wup = treq->rcv_isn + 1;
 395                 newtp->write_seq = treq->snt_isn + 1;
 396                 newtp->pushed_seq = newtp->write_seq;
 397                 newtp->copied_seq = treq->rcv_isn + 1;
 398
 399                 newtp->rx_opt.saw_tstamp = 0;
 400
 401                 newtp->rx_opt.dsack = 0;
 402                 newtp->rx_opt.eff_sacks = 0;
 403
 404                 newtp->rx_opt.num_sacks = 0;
 405                 newtp->urg_data = 0;
 406
 407                 if (sock_flag(newsk, SOCK_KEEPOPEN))
 408                         inet_csk_reset_keepalive_timer(newsk,
 409                                                        keepalive_time_when(newtp));
 410
 411                 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
 412                 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) {
 413                         if (sysctl_tcp_fack)
 414                                 newtp->rx_opt.sack_ok |= 2;
 415                 }
 416                 newtp->window_clamp = req->window_clamp;
 417                 newtp->rcv_ssthresh = req->rcv_wnd;
 418                 newtp->rcv_wnd = req->rcv_wnd;
 419                 newtp->rx_opt.wscale_ok = ireq->wscale_ok;
 420                 if (newtp->rx_opt.wscale_ok) {
 421                         newtp->rx_opt.snd_wscale = ireq->snd_wscale;
 422                         newtp->rx_opt.rcv_wscale = ireq->rcv_wscale;
 423                 } else {
 424                         newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0;
 425                         newtp->window_clamp = min(newtp->window_clamp, 65535U);
 426                 }
 427                 newtp->snd_wnd = ntohs(skb->h.th->window) << newtp->rx_opt.snd_wscale;
 428                 newtp->max_window = newtp->snd_wnd;
 429
 430                 if (newtp->rx_opt.tstamp_ok) {
 431                         newtp->rx_opt.ts_recent = req->ts_recent;
 432                         newtp->rx_opt.ts_recent_stamp = xtime.tv_sec;
 433                         newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
 434                 } else {
 435                         newtp->rx_opt.ts_recent_stamp = 0;
 436                         newtp->tcp_header_len = sizeof(struct tcphdr);
 437                 }
 438                 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len)
 439                         newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len;
 440                 newtp->rx_opt.mss_clamp = req->mss;
 441                 TCP_ECN_openreq_child(newtp, req);
 442
 443                 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS);
 444         }
 445         return newsk;
 446 }
 447
 448 /*
 449  *      Process an incoming packet for SYN_RECV sockets represented
 450  *      as a request_sock.
 451  */
 452
 453 struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb,
 454                            struct request_sock *req,
 455                            struct request_sock **prev)
 456 {
 457         struct tcphdr *th = skb->h.th;
 458         u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK);
 459         int paws_reject = 0;
 460         struct tcp_options_received tmp_opt;
 461         struct sock *child;
 462
 463         tmp_opt.saw_tstamp = 0;
 464         if (th->doff > (sizeof(struct tcphdr)>>2)) {
 465                 tcp_parse_options(skb, &tmp_opt, 0);
 466
 467                 if (tmp_opt.saw_tstamp) {
 468                         tmp_opt.ts_recent = req->ts_recent;
 469                         /* We do not store true stamp, but it is not required,
 470                          * it can be estimated (approximately)
 471                          * from another data.
 472                          */
 473                         tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans);
 474                         paws_reject = tcp_paws_check(&tmp_opt, th->rst);
 475                 }
 476         }
 477
 478         /* Check for pure retransmitted SYN. */
 479         if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn &&
 480             flg == TCP_FLAG_SYN &&
 481             !paws_reject) {
 482                 /*
 483                  * RFC793 draws (Incorrectly! It was fixed in RFC1122)
 484                  * this case on figure 6 and figure 8, but formal
 485                  * protocol description says NOTHING.
 486                  * To be more exact, it says that we should send ACK,
 487                  * because this segment (at least, if it has no data)
 488                  * is out of window.
 489                  *
 490                  *  CONCLUSION: RFC793 (even with RFC1122) DOES NOT
 491                  *  describe SYN-RECV state. All the description
 492                  *  is wrong, we cannot believe to it and should
 493                  *  rely only on common sense and implementation
 494                  *  experience.
 495                  *
 496                  * Enforce "SYN-ACK" according to figure 8, figure 6
 497                  * of RFC793, fixed by RFC1122.
 498                  */
 499                 req->rsk_ops->rtx_syn_ack(sk, req, NULL);
 500                 return NULL;
 501         }
 502
 503         /* Further reproduces section "SEGMENT ARRIVES"
 504            for state SYN-RECEIVED of RFC793.
 505            It is broken, however, it does not work only
 506            when SYNs are crossed.
 507
 508            You would think that SYN crossing is impossible here, since
 509            we should have a SYN_SENT socket (from connect()) on our end,
 510            but this is not true if the crossed SYNs were sent to both
 511            ends by a malicious third party.  We must defend against this,
 512            and to do that we first verify the ACK (as per RFC793, page
 513            36) and reset if it is invalid.  Is this a true full defense?
 514            To convince ourselves, let us consider a way in which the ACK
 515            test can still pass in this 'malicious crossed SYNs' case.
 516            Malicious sender sends identical SYNs (and thus identical sequence
 517            numbers) to both A and B:
 518
 519                 A: gets SYN, seq=7
 520                 B: gets SYN, seq=7
 521
 522            By our good fortune, both A and B select the same initial
 523            send sequence number of seven :-)
 524
 525                 A: sends SYN|ACK, seq=7, ack_seq=8
 526                 B: sends SYN|ACK, seq=7, ack_seq=8
 527
 528            So we are now A eating this SYN|ACK, ACK test passes.  So
 529            does sequence test, SYN is truncated, and thus we consider
 530            it a bare ACK.
 531
 532            If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this
 533            bare ACK.  Otherwise, we create an established connection.  Both
 534            ends (listening sockets) accept the new incoming connection and try
 535            to talk to each other. 8-)
 536
 537            Note: This case is both harmless, and rare.  Possibility is about the
 538            same as us discovering intelligent life on another plant tomorrow.
 539
 540            But generally, we should (RFC lies!) to accept ACK
 541            from SYNACK both here and in tcp_rcv_state_process().
 542            tcp_rcv_state_process() does not, hence, we do not too.
 543
 544            Note that the case is absolutely generic:
 545            we cannot optimize anything here without
 546            violating protocol. All the checks must be made
 547            before attempt to create socket.
 548          */
 549
 550         /* RFC793 page 36: "If the connection is in any non-synchronized state ...
 551          *                  and the incoming segment acknowledges something not yet
 552          *                  sent (the segment carries an unacceptable ACK) ...
 553          *                  a reset is sent."
 554          *
 555          * Invalid ACK: reset will be sent by listening socket
 556          */
 557         if ((flg & TCP_FLAG_ACK) &&
 558             (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1))
 559                 return sk;
 560
 561         /* Also, it would be not so bad idea to check rcv_tsecr, which
 562          * is essentially ACK extension and too early or too late values
 563          * should cause reset in unsynchronized states.
 564          */
 565
 566         /* RFC793: "first check sequence number". */
 567
 568         if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq,
 569                                           tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) {
 570                 /* Out of window: send ACK and drop. */
 571                 if (!(flg & TCP_FLAG_RST))
 572                         req->rsk_ops->send_ack(skb, req);
 573                 if (paws_reject)
 574                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
 575                 return NULL;
 576         }
 577
 578         /* In sequence, PAWS is OK. */
 579
 580         if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1))
 581                         req->ts_recent = tmp_opt.rcv_tsval;
 582
 583                 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) {
 584                         /* Truncate SYN, it is out of window starting
 585                            at tcp_rsk(req)->rcv_isn + 1. */
 586                         flg &= ~TCP_FLAG_SYN;
 587                 }
 588
 589                 /* RFC793: "second check the RST bit" and
 590                  *         "fourth, check the SYN bit"
 591                  */
 592                 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) {
 593                         TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS);
 594                         goto embryonic_reset;
 595                 }
 596
 597                 /* ACK sequence verified above, just make sure ACK is
 598                  * set.  If ACK not set, just silently drop the packet.
 599                  */
 600                 if (!(flg & TCP_FLAG_ACK))
 601                         return NULL;
 602
 603                 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */
 604                 if (inet_csk(sk)->icsk_accept_queue.rskq_defer_accept &&
 605                     TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) {
 606                         inet_rsk(req)->acked = 1;
 607                         return NULL;
 608                 }
 609
 610                 /* OK, ACK is valid, create big socket and
 611                  * feed this segment to it. It will repeat all
 612                  * the tests. THIS SEGMENT MUST MOVE SOCKET TO
 613                  * ESTABLISHED STATE. If it will be dropped after
 614                  * socket is created, wait for troubles.
 615                  */
 616                 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb,
 617                                                                  req, NULL);
 618                 if (child == NULL)
 619                         goto listen_overflow;
 620
 621                 inet_csk_reqsk_queue_unlink(sk, req, prev);
 622                 inet_csk_reqsk_queue_removed(sk, req);
 623
 624                 inet_csk_reqsk_queue_add(sk, req, child);
 625                 return child;
 626
 627         listen_overflow:
 628                 if (!sysctl_tcp_abort_on_overflow) {
 629                         inet_rsk(req)->acked = 1;
 630                         return NULL;
 631                 }
 632
 633         embryonic_reset:
 634                 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS);
 635                 if (!(flg & TCP_FLAG_RST))
 636                         req->rsk_ops->send_reset(skb);
 637
 638                 inet_csk_reqsk_queue_drop(sk, req, prev);
 639                 return NULL;
 640 }
 641
 642 /*
 643  * Queue segment on the new socket if the new socket is active,
 644  * otherwise we just shortcircuit this and continue with
 645  * the new socket.
 646  */
 647
 648 int tcp_child_process(struct sock *parent, struct sock *child,
 649                       struct sk_buff *skb)
 650 {
 651         int ret = 0;
 652         int state = child->sk_state;
 653
 654         if (!sock_owned_by_user(child)) {
 655                 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len);
 656
 657                 /* Wakeup parent, send SIGIO */
 658                 if (state == TCP_SYN_RECV && child->sk_state != state)
 659                         parent->sk_data_ready(parent, 0);
 660         } else {
 661                 /* Alas, it is possible again, because we do lookup
 662                  * in main socket hash table and lock on listening
 663                  * socket does not protect us more.
 664                  */
 665                 sk_add_backlog(child, skb);
 666         }
 667
 668         bh_unlock_sock(child);
 669         sock_put(child);
 670         return ret;
 671 }
 672
 673 EXPORT_SYMBOL(tcp_check_req);
 674 EXPORT_SYMBOL(tcp_child_process);
 675 EXPORT_SYMBOL(tcp_create_openreq_child);
 676 EXPORT_SYMBOL(tcp_timewait_state_process);