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