net/ipv4/tcp_input.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_input.c,v 1.243 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 /*
  24  * Changes:
  25  *              Pedro Roque     :       Fast Retransmit/Recovery.
  26  *                                      Two receive queues.
  27  *                                      Retransmit queue handled by TCP.
  28  *                                      Better retransmit timer handling.
  29  *                                      New congestion avoidance.
  30  *                                      Header prediction.
  31  *                                      Variable renaming.
  32  *
  33  *              Eric            :       Fast Retransmit.
  34  *              Randy Scott     :       MSS option defines.
  35  *              Eric Schenk     :       Fixes to slow start algorithm.
  36  *              Eric Schenk     :       Yet another double ACK bug.
  37  *              Eric Schenk     :       Delayed ACK bug fixes.
  38  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
  39  *              David S. Miller :       Don't allow zero congestion window.
  40  *              Eric Schenk     :       Fix retransmitter so that it sends
  41  *                                      next packet on ack of previous packet.
  42  *              Andi Kleen      :       Moved open_request checking here
  43  *                                      and process RSTs for open_requests.
  44  *              Andi Kleen      :       Better prune_queue, and other fixes.
  45  *              Andrey Savochkin:       Fix RTT measurements in the presnce of
  46  *                                      timestamps.
  47  *              Andrey Savochkin:       Check sequence numbers correctly when
  48  *                                      removing SACKs due to in sequence incoming
  49  *                                      data segments.
  50  *              Andi Kleen:             Make sure we never ack data there is not
  51  *                                      enough room for. Also make this condition
  52  *                                      a fatal error if it might still happen.
  53  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
  54  *                                      connections with MSS<min(MTU,ann. MSS)
  55  *                                      work without delayed acks.
  56  *              Andi Kleen:             Process packets with PSH set in the
  57  *                                      fast path.
  58  *              J Hadi Salim:           ECN support
  59  *              Andrei Gurtov,
  60  *              Pasi Sarolahti,
  61  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
  62  *                                      engine. Lots of bugs are found.
  63  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
  64  */
  65
  66 #include <linux/config.h>
  67 #include <linux/mm.h>
  68 #include <linux/module.h>
  69 #include <linux/sysctl.h>
  70 #include <net/tcp.h>
  71 #include <net/inet_common.h>
  72 #include <linux/ipsec.h>
  73 #include <asm/unaligned.h>
  74
  75 int sysctl_tcp_timestamps = 1;
  76 int sysctl_tcp_window_scaling = 1;
  77 int sysctl_tcp_sack = 1;
  78 int sysctl_tcp_fack = 1;
  79 int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
  80 int sysctl_tcp_ecn;
  81 int sysctl_tcp_dsack = 1;
  82 int sysctl_tcp_app_win = 31;
  83 int sysctl_tcp_adv_win_scale = 2;
  84
  85 int sysctl_tcp_stdurg;
  86 int sysctl_tcp_rfc1337;
  87 int sysctl_tcp_max_orphans = NR_FILE;
  88 int sysctl_tcp_frto;
  89 int sysctl_tcp_nometrics_save;
  90
  91 int sysctl_tcp_moderate_rcvbuf = 1;
  92 int sysctl_tcp_abc = 1;
  93
  94 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
  95 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
  96 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
  97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
  98 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
  99 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
 100 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
 101 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
 102 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
 103
 104 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 105 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
 106 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
 107 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
 108
 109 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
 110 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
 111 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
 112
 113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
 114
 115 /* Adapt the MSS value used to make delayed ack decision to the
 116  * real world.
 117  */
 118 static inline void tcp_measure_rcv_mss(struct sock *sk,
 119                                        const struct sk_buff *skb)
 120 {
 121         struct inet_connection_sock *icsk = inet_csk(sk);
 122         const unsigned int lss = icsk->icsk_ack.last_seg_size;
 123         unsigned int len;
 124
 125         icsk->icsk_ack.last_seg_size = 0;
 126
 127         /* skb->len may jitter because of SACKs, even if peer
 128          * sends good full-sized frames.
 129          */
 130         len = skb->len;
 131         if (len >= icsk->icsk_ack.rcv_mss) {
 132                 icsk->icsk_ack.rcv_mss = len;
 133         } else {
 134                 /* Otherwise, we make more careful check taking into account,
 135                  * that SACKs block is variable.
 136                  *
 137                  * "len" is invariant segment length, including TCP header.
 138                  */
 139                 len += skb->data - skb->h.raw;
 140                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
 141                     /* If PSH is not set, packet should be
 142                      * full sized, provided peer TCP is not badly broken.
 143                      * This observation (if it is correct 8)) allows
 144                      * to handle super-low mtu links fairly.
 145                      */
 146                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
 147                      !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
 148                         /* Subtract also invariant (if peer is RFC compliant),
 149                          * tcp header plus fixed timestamp option length.
 150                          * Resulting "len" is MSS free of SACK jitter.
 151                          */
 152                         len -= tcp_sk(sk)->tcp_header_len;
 153                         icsk->icsk_ack.last_seg_size = len;
 154                         if (len == lss) {
 155                                 icsk->icsk_ack.rcv_mss = len;
 156                                 return;
 157                         }
 158                 }
 159                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
 160         }
 161 }
 162
 163 static void tcp_incr_quickack(struct sock *sk)
 164 {
 165         struct inet_connection_sock *icsk = inet_csk(sk);
 166         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
 167
 168         if (quickacks==0)
 169                 quickacks=2;
 170         if (quickacks > icsk->icsk_ack.quick)
 171                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
 172 }
 173
 174 void tcp_enter_quickack_mode(struct sock *sk)
 175 {
 176         struct inet_connection_sock *icsk = inet_csk(sk);
 177         tcp_incr_quickack(sk);
 178         icsk->icsk_ack.pingpong = 0;
 179         icsk->icsk_ack.ato = TCP_ATO_MIN;
 180 }
 181
 182 /* Send ACKs quickly, if "quick" count is not exhausted
 183  * and the session is not interactive.
 184  */
 185
 186 static inline int tcp_in_quickack_mode(const struct sock *sk)
 187 {
 188         const struct inet_connection_sock *icsk = inet_csk(sk);
 189         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
 190 }
 191
 192 /* Buffer size and advertised window tuning.
 193  *
 194  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
 195  */
 196
 197 static void tcp_fixup_sndbuf(struct sock *sk)
 198 {
 199         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
 200                      sizeof(struct sk_buff);
 201
 202         if (sk->sk_sndbuf < 3 * sndmem)
 203                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
 204 }
 205
 206 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
 207  *
 208  * All tcp_full_space() is split to two parts: "network" buffer, allocated
 209  * forward and advertised in receiver window (tp->rcv_wnd) and
 210  * "application buffer", required to isolate scheduling/application
 211  * latencies from network.
 212  * window_clamp is maximal advertised window. It can be less than
 213  * tcp_full_space(), in this case tcp_full_space() - window_clamp
 214  * is reserved for "application" buffer. The less window_clamp is
 215  * the smoother our behaviour from viewpoint of network, but the lower
 216  * throughput and the higher sensitivity of the connection to losses. 8)
 217  *
 218  * rcv_ssthresh is more strict window_clamp used at "slow start"
 219  * phase to predict further behaviour of this connection.
 220  * It is used for two goals:
 221  * - to enforce header prediction at sender, even when application
 222  *   requires some significant "application buffer". It is check #1.
 223  * - to prevent pruning of receive queue because of misprediction
 224  *   of receiver window. Check #2.
 225  *
 226  * The scheme does not work when sender sends good segments opening
 227  * window and then starts to feed us spagetti. But it should work
 228  * in common situations. Otherwise, we have to rely on queue collapsing.
 229  */
 230
 231 /* Slow part of check#2. */
 232 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
 233                              const struct sk_buff *skb)
 234 {
 235         /* Optimize this! */
 236         int truesize = tcp_win_from_space(skb->truesize)/2;
 237         int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
 238
 239         while (tp->rcv_ssthresh <= window) {
 240                 if (truesize <= skb->len)
 241                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
 242
 243                 truesize >>= 1;
 244                 window >>= 1;
 245         }
 246         return 0;
 247 }
 248
 249 static inline void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
 250                                    struct sk_buff *skb)
 251 {
 252         /* Check #1 */
 253         if (tp->rcv_ssthresh < tp->window_clamp &&
 254             (int)tp->rcv_ssthresh < tcp_space(sk) &&
 255             !tcp_memory_pressure) {
 256                 int incr;
 257
 258                 /* Check #2. Increase window, if skb with such overhead
 259                  * will fit to rcvbuf in future.
 260                  */
 261                 if (tcp_win_from_space(skb->truesize) <= skb->len)
 262                         incr = 2*tp->advmss;
 263                 else
 264                         incr = __tcp_grow_window(sk, tp, skb);
 265
 266                 if (incr) {
 267                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
 268                         inet_csk(sk)->icsk_ack.quick |= 1;
 269                 }
 270         }
 271 }
 272
 273 /* 3. Tuning rcvbuf, when connection enters established state. */
 274
 275 static void tcp_fixup_rcvbuf(struct sock *sk)
 276 {
 277         struct tcp_sock *tp = tcp_sk(sk);
 278         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
 279
 280         /* Try to select rcvbuf so that 4 mss-sized segments
 281          * will fit to window and correspoding skbs will fit to our rcvbuf.
 282          * (was 3; 4 is minimum to allow fast retransmit to work.)
 283          */
 284         while (tcp_win_from_space(rcvmem) < tp->advmss)
 285                 rcvmem += 128;
 286         if (sk->sk_rcvbuf < 4 * rcvmem)
 287                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
 288 }
 289
 290 /* 4. Try to fixup all. It is made iimediately after connection enters
 291  *    established state.
 292  */
 293 static void tcp_init_buffer_space(struct sock *sk)
 294 {
 295         struct tcp_sock *tp = tcp_sk(sk);
 296         int maxwin;
 297
 298         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 299                 tcp_fixup_rcvbuf(sk);
 300         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
 301                 tcp_fixup_sndbuf(sk);
 302
 303         tp->rcvq_space.space = tp->rcv_wnd;
 304
 305         maxwin = tcp_full_space(sk);
 306
 307         if (tp->window_clamp >= maxwin) {
 308                 tp->window_clamp = maxwin;
 309
 310                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
 311                         tp->window_clamp = max(maxwin -
 312                                                (maxwin >> sysctl_tcp_app_win),
 313                                                4 * tp->advmss);
 314         }
 315
 316         /* Force reservation of one segment. */
 317         if (sysctl_tcp_app_win &&
 318             tp->window_clamp > 2 * tp->advmss &&
 319             tp->window_clamp + tp->advmss > maxwin)
 320                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
 321
 322         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
 323         tp->snd_cwnd_stamp = tcp_time_stamp;
 324 }
 325
 326 /* 5. Recalculate window clamp after socket hit its memory bounds. */
 327 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
 328 {
 329         struct inet_connection_sock *icsk = inet_csk(sk);
 330
 331         icsk->icsk_ack.quick = 0;
 332
 333         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 334             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
 335             !tcp_memory_pressure &&
 336             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
 337                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 338                                     sysctl_tcp_rmem[2]);
 339         }
 340         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 341                 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
 342 }
 343
 344 /* Receiver "autotuning" code.
 345  *
 346  * The algorithm for RTT estimation w/o timestamps is based on
 347  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
 348  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
 349  *
 350  * More detail on this code can be found at
 351  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
 352  * though this reference is out of date.  A new paper
 353  * is pending.
 354  */
 355 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 356 {
 357         u32 new_sample = tp->rcv_rtt_est.rtt;
 358         long m = sample;
 359
 360         if (m == 0)
 361                 m = 1;
 362
 363         if (new_sample != 0) {
 364                 /* If we sample in larger samples in the non-timestamp
 365                  * case, we could grossly overestimate the RTT especially
 366                  * with chatty applications or bulk transfer apps which
 367                  * are stalled on filesystem I/O.
 368                  *
 369                  * Also, since we are only going for a minimum in the
 370                  * non-timestamp case, we do not smoothe things out
 371                  * else with timestamps disabled convergance takes too
 372                  * long.
 373                  */
 374                 if (!win_dep) {
 375                         m -= (new_sample >> 3);
 376                         new_sample += m;
 377                 } else if (m < new_sample)
 378                         new_sample = m << 3;
 379         } else {
 380                 /* No previous mesaure. */
 381                 new_sample = m << 3;
 382         }
 383
 384         if (tp->rcv_rtt_est.rtt != new_sample)
 385                 tp->rcv_rtt_est.rtt = new_sample;
 386 }
 387
 388 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
 389 {
 390         if (tp->rcv_rtt_est.time == 0)
 391                 goto new_measure;
 392         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
 393                 return;
 394         tcp_rcv_rtt_update(tp,
 395                            jiffies - tp->rcv_rtt_est.time,
 396                            1);
 397
 398 new_measure:
 399         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
 400         tp->rcv_rtt_est.time = tcp_time_stamp;
 401 }
 402
 403 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
 404 {
 405         struct tcp_sock *tp = tcp_sk(sk);
 406         if (tp->rx_opt.rcv_tsecr &&
 407             (TCP_SKB_CB(skb)->end_seq -
 408              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
 409                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
 410 }
 411
 412 /*
 413  * This function should be called every time data is copied to user space.
 414  * It calculates the appropriate TCP receive buffer space.
 415  */
 416 void tcp_rcv_space_adjust(struct sock *sk)
 417 {
 418         struct tcp_sock *tp = tcp_sk(sk);
 419         int time;
 420         int space;
 421
 422         if (tp->rcvq_space.time == 0)
 423                 goto new_measure;
 424
 425         time = tcp_time_stamp - tp->rcvq_space.time;
 426         if (time < (tp->rcv_rtt_est.rtt >> 3) ||
 427             tp->rcv_rtt_est.rtt == 0)
 428                 return;
 429
 430         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
 431
 432         space = max(tp->rcvq_space.space, space);
 433
 434         if (tp->rcvq_space.space != space) {
 435                 int rcvmem;
 436
 437                 tp->rcvq_space.space = space;
 438
 439                 if (sysctl_tcp_moderate_rcvbuf) {
 440                         int new_clamp = space;
 441
 442                         /* Receive space grows, normalize in order to
 443                          * take into account packet headers and sk_buff
 444                          * structure overhead.
 445                          */
 446                         space /= tp->advmss;
 447                         if (!space)
 448                                 space = 1;
 449                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
 450                                   16 + sizeof(struct sk_buff));
 451                         while (tcp_win_from_space(rcvmem) < tp->advmss)
 452                                 rcvmem += 128;
 453                         space *= rcvmem;
 454                         space = min(space, sysctl_tcp_rmem[2]);
 455                         if (space > sk->sk_rcvbuf) {
 456                                 sk->sk_rcvbuf = space;
 457
 458                                 /* Make the window clamp follow along.  */
 459                                 tp->window_clamp = new_clamp;
 460                         }
 461                 }
 462         }
 463
 464 new_measure:
 465         tp->rcvq_space.seq = tp->copied_seq;
 466         tp->rcvq_space.time = tcp_time_stamp;
 467 }
 468
 469 /* There is something which you must keep in mind when you analyze the
 470  * behavior of the tp->ato delayed ack timeout interval.  When a
 471  * connection starts up, we want to ack as quickly as possible.  The
 472  * problem is that "good" TCP's do slow start at the beginning of data
 473  * transmission.  The means that until we send the first few ACK's the
 474  * sender will sit on his end and only queue most of his data, because
 475  * he can only send snd_cwnd unacked packets at any given time.  For
 476  * each ACK we send, he increments snd_cwnd and transmits more of his
 477  * queue.  -DaveM
 478  */
 479 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
 480 {
 481         struct inet_connection_sock *icsk = inet_csk(sk);
 482         u32 now;
 483
 484         inet_csk_schedule_ack(sk);
 485
 486         tcp_measure_rcv_mss(sk, skb);
 487
 488         tcp_rcv_rtt_measure(tp);
 489
 490         now = tcp_time_stamp;
 491
 492         if (!icsk->icsk_ack.ato) {
 493                 /* The _first_ data packet received, initialize
 494                  * delayed ACK engine.
 495                  */
 496                 tcp_incr_quickack(sk);
 497                 icsk->icsk_ack.ato = TCP_ATO_MIN;
 498         } else {
 499                 int m = now - icsk->icsk_ack.lrcvtime;
 500
 501                 if (m <= TCP_ATO_MIN/2) {
 502                         /* The fastest case is the first. */
 503                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
 504                 } else if (m < icsk->icsk_ack.ato) {
 505                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
 506                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
 507                                 icsk->icsk_ack.ato = icsk->icsk_rto;
 508                 } else if (m > icsk->icsk_rto) {
 509                         /* Too long gap. Apparently sender falled to
 510                          * restart window, so that we send ACKs quickly.
 511                          */
 512                         tcp_incr_quickack(sk);
 513                         sk_stream_mem_reclaim(sk);
 514                 }
 515         }
 516         icsk->icsk_ack.lrcvtime = now;
 517
 518         TCP_ECN_check_ce(tp, skb);
 519
 520         if (skb->len >= 128)
 521                 tcp_grow_window(sk, tp, skb);
 522 }
 523
 524 /* Called to compute a smoothed rtt estimate. The data fed to this
 525  * routine either comes from timestamps, or from segments that were
 526  * known _not_ to have been retransmitted [see Karn/Partridge
 527  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
 528  * piece by Van Jacobson.
 529  * NOTE: the next three routines used to be one big routine.
 530  * To save cycles in the RFC 1323 implementation it was better to break
 531  * it up into three procedures. -- erics
 532  */
 533 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
 534 {
 535         struct tcp_sock *tp = tcp_sk(sk);
 536         long m = mrtt; /* RTT */
 537
 538         /*      The following amusing code comes from Jacobson's
 539          *      article in SIGCOMM '88.  Note that rtt and mdev
 540          *      are scaled versions of rtt and mean deviation.
 541          *      This is designed to be as fast as possible
 542          *      m stands for "measurement".
 543          *
 544          *      On a 1990 paper the rto value is changed to:
 545          *      RTO = rtt + 4 * mdev
 546          *
 547          * Funny. This algorithm seems to be very broken.
 548          * These formulae increase RTO, when it should be decreased, increase
 549          * too slowly, when it should be incresed fastly, decrease too fastly
 550          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
 551          * does not matter how to _calculate_ it. Seems, it was trap
 552          * that VJ failed to avoid. 8)
 553          */
 554         if(m == 0)
 555                 m = 1;
 556         if (tp->srtt != 0) {
 557                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
 558                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
 559                 if (m < 0) {
 560                         m = -m;         /* m is now abs(error) */
 561                         m -= (tp->mdev >> 2);   /* similar update on mdev */
 562                         /* This is similar to one of Eifel findings.
 563                          * Eifel blocks mdev updates when rtt decreases.
 564                          * This solution is a bit different: we use finer gain
 565                          * for mdev in this case (alpha*beta).
 566                          * Like Eifel it also prevents growth of rto,
 567                          * but also it limits too fast rto decreases,
 568                          * happening in pure Eifel.
 569                          */
 570                         if (m > 0)
 571                                 m >>= 3;
 572                 } else {
 573                         m -= (tp->mdev >> 2);   /* similar update on mdev */
 574                 }
 575                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
 576                 if (tp->mdev > tp->mdev_max) {
 577                         tp->mdev_max = tp->mdev;
 578                         if (tp->mdev_max > tp->rttvar)
 579                                 tp->rttvar = tp->mdev_max;
 580                 }
 581                 if (after(tp->snd_una, tp->rtt_seq)) {
 582                         if (tp->mdev_max < tp->rttvar)
 583                                 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
 584                         tp->rtt_seq = tp->snd_nxt;
 585                         tp->mdev_max = TCP_RTO_MIN;
 586                 }
 587         } else {
 588                 /* no previous measure. */
 589                 tp->srtt = m<<3;        /* take the measured time to be rtt */
 590                 tp->mdev = m<<1;        /* make sure rto = 3*rtt */
 591                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
 592                 tp->rtt_seq = tp->snd_nxt;
 593         }
 594 }
 595
 596 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
 597  * routine referred to above.
 598  */
 599 static inline void tcp_set_rto(struct sock *sk)
 600 {
 601         const struct tcp_sock *tp = tcp_sk(sk);
 602         /* Old crap is replaced with new one. 8)
 603          *
 604          * More seriously:
 605          * 1. If rtt variance happened to be less 50msec, it is hallucination.
 606          *    It cannot be less due to utterly erratic ACK generation made
 607          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
 608          *    to do with delayed acks, because at cwnd>2 true delack timeout
 609          *    is invisible. Actually, Linux-2.4 also generates erratic
 610          *    ACKs in some curcumstances.
 611          */
 612         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
 613
 614         /* 2. Fixups made earlier cannot be right.
 615          *    If we do not estimate RTO correctly without them,
 616          *    all the algo is pure shit and should be replaced
 617          *    with correct one. It is exaclty, which we pretend to do.
 618          */
 619 }
 620
 621 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
 622  * guarantees that rto is higher.
 623  */
 624 static inline void tcp_bound_rto(struct sock *sk)
 625 {
 626         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
 627                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
 628 }
 629
 630 /* Save metrics learned by this TCP session.
 631    This function is called only, when TCP finishes successfully
 632    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
 633  */
 634 void tcp_update_metrics(struct sock *sk)
 635 {
 636         struct tcp_sock *tp = tcp_sk(sk);
 637         struct dst_entry *dst = __sk_dst_get(sk);
 638
 639         if (sysctl_tcp_nometrics_save)
 640                 return;
 641
 642         dst_confirm(dst);
 643
 644         if (dst && (dst->flags&DST_HOST)) {
 645                 const struct inet_connection_sock *icsk = inet_csk(sk);
 646                 int m;
 647
 648                 if (icsk->icsk_backoff || !tp->srtt) {
 649                         /* This session failed to estimate rtt. Why?
 650                          * Probably, no packets returned in time.
 651                          * Reset our results.
 652                          */
 653                         if (!(dst_metric_locked(dst, RTAX_RTT)))
 654                                 dst->metrics[RTAX_RTT-1] = 0;
 655                         return;
 656                 }
 657
 658                 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
 659
 660                 /* If newly calculated rtt larger than stored one,
 661                  * store new one. Otherwise, use EWMA. Remember,
 662                  * rtt overestimation is always better than underestimation.
 663                  */
 664                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
 665                         if (m <= 0)
 666                                 dst->metrics[RTAX_RTT-1] = tp->srtt;
 667                         else
 668                                 dst->metrics[RTAX_RTT-1] -= (m>>3);
 669                 }
 670
 671                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
 672                         if (m < 0)
 673                                 m = -m;
 674
 675                         /* Scale deviation to rttvar fixed point */
 676                         m >>= 1;
 677                         if (m < tp->mdev)
 678                                 m = tp->mdev;
 679
 680                         if (m >= dst_metric(dst, RTAX_RTTVAR))
 681                                 dst->metrics[RTAX_RTTVAR-1] = m;
 682                         else
 683                                 dst->metrics[RTAX_RTTVAR-1] -=
 684                                         (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
 685                 }
 686
 687                 if (tp->snd_ssthresh >= 0xFFFF) {
 688                         /* Slow start still did not finish. */
 689                         if (dst_metric(dst, RTAX_SSTHRESH) &&
 690                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
 691                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
 692                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
 693                         if (!dst_metric_locked(dst, RTAX_CWND) &&
 694                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
 695                                 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
 696                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 697                            icsk->icsk_ca_state == TCP_CA_Open) {
 698                         /* Cong. avoidance phase, cwnd is reliable. */
 699                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
 700                                 dst->metrics[RTAX_SSTHRESH-1] =
 701                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
 702                         if (!dst_metric_locked(dst, RTAX_CWND))
 703                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
 704                 } else {
 705                         /* Else slow start did not finish, cwnd is non-sense,
 706                            ssthresh may be also invalid.
 707                          */
 708                         if (!dst_metric_locked(dst, RTAX_CWND))
 709                                 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
 710                         if (dst->metrics[RTAX_SSTHRESH-1] &&
 711                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
 712                             tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
 713                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
 714                 }
 715
 716                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
 717                         if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
 718                             tp->reordering != sysctl_tcp_reordering)
 719                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
 720                 }
 721         }
 722 }
 723
 724 /* Numbers are taken from RFC2414.  */
 725 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
 726 {
 727         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 728
 729         if (!cwnd) {
 730                 if (tp->mss_cache > 1460)
 731                         cwnd = 2;
 732                 else
 733                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
 734         }
 735         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 736 }
 737
 738 /* Initialize metrics on socket. */
 739
 740 static void tcp_init_metrics(struct sock *sk)
 741 {
 742         struct tcp_sock *tp = tcp_sk(sk);
 743         struct dst_entry *dst = __sk_dst_get(sk);
 744
 745         if (dst == NULL)
 746                 goto reset;
 747
 748         dst_confirm(dst);
 749
 750         if (dst_metric_locked(dst, RTAX_CWND))
 751                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
 752         if (dst_metric(dst, RTAX_SSTHRESH)) {
 753                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
 754                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 755                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 756         }
 757         if (dst_metric(dst, RTAX_REORDERING) &&
 758             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
 759                 tp->rx_opt.sack_ok &= ~2;
 760                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
 761         }
 762
 763         if (dst_metric(dst, RTAX_RTT) == 0)
 764                 goto reset;
 765
 766         if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
 767                 goto reset;
 768
 769         /* Initial rtt is determined from SYN,SYN-ACK.
 770          * The segment is small and rtt may appear much
 771          * less than real one. Use per-dst memory
 772          * to make it more realistic.
 773          *
 774          * A bit of theory. RTT is time passed after "normal" sized packet
 775          * is sent until it is ACKed. In normal curcumstances sending small
 776          * packets force peer to delay ACKs and calculation is correct too.
 777          * The algorithm is adaptive and, provided we follow specs, it
 778          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 779          * tricks sort of "quick acks" for time long enough to decrease RTT
 780          * to low value, and then abruptly stops to do it and starts to delay
 781          * ACKs, wait for troubles.
 782          */
 783         if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
 784                 tp->srtt = dst_metric(dst, RTAX_RTT);
 785                 tp->rtt_seq = tp->snd_nxt;
 786         }
 787         if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
 788                 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
 789                 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
 790         }
 791         tcp_set_rto(sk);
 792         tcp_bound_rto(sk);
 793         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
 794                 goto reset;
 795         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 796         tp->snd_cwnd_stamp = tcp_time_stamp;
 797         return;
 798
 799 reset:
 800         /* Play conservative. If timestamps are not
 801          * supported, TCP will fail to recalculate correct
 802          * rtt, if initial rto is too small. FORGET ALL AND RESET!
 803          */
 804         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
 805                 tp->srtt = 0;
 806                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
 807                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
 808         }
 809 }
 810
 811 static void tcp_update_reordering(struct sock *sk, const int metric,
 812                                   const int ts)
 813 {
 814         struct tcp_sock *tp = tcp_sk(sk);
 815         if (metric > tp->reordering) {
 816                 tp->reordering = min(TCP_MAX_REORDERING, metric);
 817
 818                 /* This exciting event is worth to be remembered. 8) */
 819                 if (ts)
 820                         NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
 821                 else if (IsReno(tp))
 822                         NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
 823                 else if (IsFack(tp))
 824                         NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
 825                 else
 826                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
 827 #if FASTRETRANS_DEBUG > 1
 828                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
 829                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
 830                        tp->reordering,
 831                        tp->fackets_out,
 832                        tp->sacked_out,
 833                        tp->undo_marker ? tp->undo_retrans : 0);
 834 #endif
 835                 /* Disable FACK yet. */
 836                 tp->rx_opt.sack_ok &= ~2;
 837         }
 838 }
 839
 840 /* This procedure tags the retransmission queue when SACKs arrive.
 841  *
 842  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
 843  * Packets in queue with these bits set are counted in variables
 844  * sacked_out, retrans_out and lost_out, correspondingly.
 845  *
 846  * Valid combinations are:
 847  * Tag  InFlight        Description
 848  * 0    1               - orig segment is in flight.
 849  * S    0               - nothing flies, orig reached receiver.
 850  * L    0               - nothing flies, orig lost by net.
 851  * R    2               - both orig and retransmit are in flight.
 852  * L|R  1               - orig is lost, retransmit is in flight.
 853  * S|R  1               - orig reached receiver, retrans is still in flight.
 854  * (L|S|R is logically valid, it could occur when L|R is sacked,
 855  *  but it is equivalent to plain S and code short-curcuits it to S.
 856  *  L|S is logically invalid, it would mean -1 packet in flight 8))
 857  *
 858  * These 6 states form finite state machine, controlled by the following events:
 859  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
 860  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
 861  * 3. Loss detection event of one of three flavors:
 862  *      A. Scoreboard estimator decided the packet is lost.
 863  *         A'. Reno "three dupacks" marks head of queue lost.
 864  *         A''. Its FACK modfication, head until snd.fack is lost.
 865  *      B. SACK arrives sacking data transmitted after never retransmitted
 866  *         hole was sent out.
 867  *      C. SACK arrives sacking SND.NXT at the moment, when the
 868  *         segment was retransmitted.
 869  * 4. D-SACK added new rule: D-SACK changes any tag to S.
 870  *
 871  * It is pleasant to note, that state diagram turns out to be commutative,
 872  * so that we are allowed not to be bothered by order of our actions,
 873  * when multiple events arrive simultaneously. (see the function below).
 874  *
 875  * Reordering detection.
 876  * --------------------
 877  * Reordering metric is maximal distance, which a packet can be displaced
 878  * in packet stream. With SACKs we can estimate it:
 879  *
 880  * 1. SACK fills old hole and the corresponding segment was not
 881  *    ever retransmitted -> reordering. Alas, we cannot use it
 882  *    when segment was retransmitted.
 883  * 2. The last flaw is solved with D-SACK. D-SACK arrives
 884  *    for retransmitted and already SACKed segment -> reordering..
 885  * Both of these heuristics are not used in Loss state, when we cannot
 886  * account for retransmits accurately.
 887  */
 888 static int
 889 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
 890 {
 891         const struct inet_connection_sock *icsk = inet_csk(sk);
 892         struct tcp_sock *tp = tcp_sk(sk);
 893         unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
 894         struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
 895         int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
 896         int reord = tp->packets_out;
 897         int prior_fackets;
 898         u32 lost_retrans = 0;
 899         int flag = 0;
 900         int i;
 901
 902         if (!tp->sacked_out)
 903                 tp->fackets_out = 0;
 904         prior_fackets = tp->fackets_out;
 905
 906         for (i=0; i<num_sacks; i++, sp++) {
 907                 struct sk_buff *skb;
 908                 __u32 start_seq = ntohl(sp->start_seq);
 909                 __u32 end_seq = ntohl(sp->end_seq);
 910                 int fack_count = 0;
 911                 int dup_sack = 0;
 912
 913                 /* Check for D-SACK. */
 914                 if (i == 0) {
 915                         u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
 916
 917                         if (before(start_seq, ack)) {
 918                                 dup_sack = 1;
 919                                 tp->rx_opt.sack_ok |= 4;
 920                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
 921                         } else if (num_sacks > 1 &&
 922                                    !after(end_seq, ntohl(sp[1].end_seq)) &&
 923                                    !before(start_seq, ntohl(sp[1].start_seq))) {
 924                                 dup_sack = 1;
 925                                 tp->rx_opt.sack_ok |= 4;
 926                                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
 927                         }
 928
 929                         /* D-SACK for already forgotten data...
 930                          * Do dumb counting. */
 931                         if (dup_sack &&
 932                             !after(end_seq, prior_snd_una) &&
 933                             after(end_seq, tp->undo_marker))
 934                                 tp->undo_retrans--;
 935
 936                         /* Eliminate too old ACKs, but take into
 937                          * account more or less fresh ones, they can
 938                          * contain valid SACK info.
 939                          */
 940                         if (before(ack, prior_snd_una - tp->max_window))
 941                                 return 0;
 942                 }
 943
 944                 /* Event "B" in the comment above. */
 945                 if (after(end_seq, tp->high_seq))
 946                         flag |= FLAG_DATA_LOST;
 947
 948                 sk_stream_for_retrans_queue(skb, sk) {
 949                         int in_sack, pcount;
 950                         u8 sacked;
 951
 952                         /* The retransmission queue is always in order, so
 953                          * we can short-circuit the walk early.
 954                          */
 955                         if (!before(TCP_SKB_CB(skb)->seq, end_seq))
 956                                 break;
 957
 958                         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
 959                                 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
 960
 961                         pcount = tcp_skb_pcount(skb);
 962
 963                         if (pcount > 1 && !in_sack &&
 964                             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
 965                                 unsigned int pkt_len;
 966
 967                                 in_sack = !after(start_seq,
 968                                                  TCP_SKB_CB(skb)->seq);
 969
 970                                 if (!in_sack)
 971                                         pkt_len = (start_seq -
 972                                                    TCP_SKB_CB(skb)->seq);
 973                                 else
 974                                         pkt_len = (end_seq -
 975                                                    TCP_SKB_CB(skb)->seq);
 976                                 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->tso_size))
 977                                         break;
 978                                 pcount = tcp_skb_pcount(skb);
 979                         }
 980
 981                         fack_count += pcount;
 982
 983                         sacked = TCP_SKB_CB(skb)->sacked;
 984
 985                         /* Account D-SACK for retransmitted packet. */
 986                         if ((dup_sack && in_sack) &&
 987                             (sacked & TCPCB_RETRANS) &&
 988                             after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
 989                                 tp->undo_retrans--;
 990
 991                         /* The frame is ACKed. */
 992                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
 993                                 if (sacked&TCPCB_RETRANS) {
 994                                         if ((dup_sack && in_sack) &&
 995                                             (sacked&TCPCB_SACKED_ACKED))
 996                                                 reord = min(fack_count, reord);
 997                                 } else {
 998                                         /* If it was in a hole, we detected reordering. */
 999                                         if (fack_count < prior_fackets &&
1000                                             !(sacked&TCPCB_SACKED_ACKED))
1001                                                 reord = min(fack_count, reord);
1002                                 }
1003
1004                                 /* Nothing to do; acked frame is about to be dropped. */
1005                                 continue;
1006                         }
1007
1008                         if ((sacked&TCPCB_SACKED_RETRANS) &&
1009                             after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1010                             (!lost_retrans || after(end_seq, lost_retrans)))
1011                                 lost_retrans = end_seq;
1012
1013                         if (!in_sack)
1014                                 continue;
1015
1016                         if (!(sacked&TCPCB_SACKED_ACKED)) {
1017                                 if (sacked & TCPCB_SACKED_RETRANS) {
1018                                         /* If the segment is not tagged as lost,
1019                                          * we do not clear RETRANS, believing
1020                                          * that retransmission is still in flight.
1021                                          */
1022                                         if (sacked & TCPCB_LOST) {
1023                                                 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1024                                                 tp->lost_out -= tcp_skb_pcount(skb);
1025                                                 tp->retrans_out -= tcp_skb_pcount(skb);
1026                                         }
1027                                 } else {
1028                                         /* New sack for not retransmitted frame,
1029                                          * which was in hole. It is reordering.
1030                                          */
1031                                         if (!(sacked & TCPCB_RETRANS) &&
1032                                             fack_count < prior_fackets)
1033                                                 reord = min(fack_count, reord);
1034
1035                                         if (sacked & TCPCB_LOST) {
1036                                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1037                                                 tp->lost_out -= tcp_skb_pcount(skb);
1038                                         }
1039                                 }
1040
1041                                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1042                                 flag |= FLAG_DATA_SACKED;
1043                                 tp->sacked_out += tcp_skb_pcount(skb);
1044
1045                                 if (fack_count > tp->fackets_out)
1046                                         tp->fackets_out = fack_count;
1047                         } else {
1048                                 if (dup_sack && (sacked&TCPCB_RETRANS))
1049                                         reord = min(fack_count, reord);
1050                         }
1051
1052                         /* D-SACK. We can detect redundant retransmission
1053                          * in S|R and plain R frames and clear it.
1054                          * undo_retrans is decreased above, L|R frames
1055                          * are accounted above as well.
1056                          */
1057                         if (dup_sack &&
1058                             (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1059                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1060                                 tp->retrans_out -= tcp_skb_pcount(skb);
1061                         }
1062                 }
1063         }
1064
1065         /* Check for lost retransmit. This superb idea is
1066          * borrowed from "ratehalving". Event "C".
1067          * Later note: FACK people cheated me again 8),
1068          * we have to account for reordering! Ugly,
1069          * but should help.
1070          */
1071         if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1072                 struct sk_buff *skb;
1073
1074                 sk_stream_for_retrans_queue(skb, sk) {
1075                         if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1076                                 break;
1077                         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1078                                 continue;
1079                         if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1080                             after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1081                             (IsFack(tp) ||
1082                              !before(lost_retrans,
1083                                      TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1084                                      tp->mss_cache))) {
1085                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1086                                 tp->retrans_out -= tcp_skb_pcount(skb);
1087
1088                                 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1089                                         tp->lost_out += tcp_skb_pcount(skb);
1090                                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1091                                         flag |= FLAG_DATA_SACKED;
1092                                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1093                                 }
1094                         }
1095                 }
1096         }
1097
1098         tp->left_out = tp->sacked_out + tp->lost_out;
1099
1100         if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1101                 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1102
1103 #if FASTRETRANS_DEBUG > 0
1104         BUG_TRAP((int)tp->sacked_out >= 0);
1105         BUG_TRAP((int)tp->lost_out >= 0);
1106         BUG_TRAP((int)tp->retrans_out >= 0);
1107         BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1108 #endif
1109         return flag;
1110 }
1111
1112 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1113  * segments to see from the next ACKs whether any data was really missing.
1114  * If the RTO was spurious, new ACKs should arrive.
1115  */
1116 void tcp_enter_frto(struct sock *sk)
1117 {
1118         const struct inet_connection_sock *icsk = inet_csk(sk);
1119         struct tcp_sock *tp = tcp_sk(sk);
1120         struct sk_buff *skb;
1121
1122         tp->frto_counter = 1;
1123
1124         if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1125             tp->snd_una == tp->high_seq ||
1126             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1127                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1128                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1129                 tcp_ca_event(sk, CA_EVENT_FRTO);
1130         }
1131
1132         /* Have to clear retransmission markers here to keep the bookkeeping
1133          * in shape, even though we are not yet in Loss state.
1134          * If something was really lost, it is eventually caught up
1135          * in tcp_enter_frto_loss.
1136          */
1137         tp->retrans_out = 0;
1138         tp->undo_marker = tp->snd_una;
1139         tp->undo_retrans = 0;
1140
1141         sk_stream_for_retrans_queue(skb, sk) {
1142                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1143         }
1144         tcp_sync_left_out(tp);
1145
1146         tcp_set_ca_state(sk, TCP_CA_Open);
1147         tp->frto_highmark = tp->snd_nxt;
1148 }
1149
1150 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1151  * which indicates that we should follow the traditional RTO recovery,
1152  * i.e. mark everything lost and do go-back-N retransmission.
1153  */
1154 static void tcp_enter_frto_loss(struct sock *sk)
1155 {
1156         struct tcp_sock *tp = tcp_sk(sk);
1157         struct sk_buff *skb;
1158         int cnt = 0;
1159
1160         tp->sacked_out = 0;
1161         tp->lost_out = 0;
1162         tp->fackets_out = 0;
1163
1164         sk_stream_for_retrans_queue(skb, sk) {
1165                 cnt += tcp_skb_pcount(skb);
1166                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1167                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1168
1169                         /* Do not mark those segments lost that were
1170                          * forward transmitted after RTO
1171                          */
1172                         if (!after(TCP_SKB_CB(skb)->end_seq,
1173                                    tp->frto_highmark)) {
1174                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1175                                 tp->lost_out += tcp_skb_pcount(skb);
1176                         }
1177                 } else {
1178                         tp->sacked_out += tcp_skb_pcount(skb);
1179                         tp->fackets_out = cnt;
1180                 }
1181         }
1182         tcp_sync_left_out(tp);
1183
1184         tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1185         tp->snd_cwnd_cnt = 0;
1186         tp->snd_cwnd_stamp = tcp_time_stamp;
1187         tp->undo_marker = 0;
1188         tp->frto_counter = 0;
1189
1190         tp->reordering = min_t(unsigned int, tp->reordering,
1191                                              sysctl_tcp_reordering);
1192         tcp_set_ca_state(sk, TCP_CA_Loss);
1193         tp->high_seq = tp->frto_highmark;
1194         TCP_ECN_queue_cwr(tp);
1195 }
1196
1197 void tcp_clear_retrans(struct tcp_sock *tp)
1198 {
1199         tp->left_out = 0;
1200         tp->retrans_out = 0;
1201
1202         tp->fackets_out = 0;
1203         tp->sacked_out = 0;
1204         tp->lost_out = 0;
1205
1206         tp->undo_marker = 0;
1207         tp->undo_retrans = 0;
1208 }
1209
1210 /* Enter Loss state. If "how" is not zero, forget all SACK information
1211  * and reset tags completely, otherwise preserve SACKs. If receiver
1212  * dropped its ofo queue, we will know this due to reneging detection.
1213  */
1214 void tcp_enter_loss(struct sock *sk, int how)
1215 {
1216         const struct inet_connection_sock *icsk = inet_csk(sk);
1217         struct tcp_sock *tp = tcp_sk(sk);
1218         struct sk_buff *skb;
1219         int cnt = 0;
1220
1221         /* Reduce ssthresh if it has not yet been made inside this window. */
1222         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1223             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1224                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1225                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1226                 tcp_ca_event(sk, CA_EVENT_LOSS);
1227         }
1228         tp->snd_cwnd       = 1;
1229         tp->snd_cwnd_cnt   = 0;
1230         tp->snd_cwnd_stamp = tcp_time_stamp;
1231
1232         tp->bytes_acked = 0;
1233         tcp_clear_retrans(tp);
1234
1235         /* Push undo marker, if it was plain RTO and nothing
1236          * was retransmitted. */
1237         if (!how)
1238                 tp->undo_marker = tp->snd_una;
1239
1240         sk_stream_for_retrans_queue(skb, sk) {
1241                 cnt += tcp_skb_pcount(skb);
1242                 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1243                         tp->undo_marker = 0;
1244                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1245                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1246                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1247                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1248                         tp->lost_out += tcp_skb_pcount(skb);
1249                 } else {
1250                         tp->sacked_out += tcp_skb_pcount(skb);
1251                         tp->fackets_out = cnt;
1252                 }
1253         }
1254         tcp_sync_left_out(tp);
1255
1256         tp->reordering = min_t(unsigned int, tp->reordering,
1257                                              sysctl_tcp_reordering);
1258         tcp_set_ca_state(sk, TCP_CA_Loss);
1259         tp->high_seq = tp->snd_nxt;
1260         TCP_ECN_queue_cwr(tp);
1261 }
1262
1263 static int tcp_check_sack_reneging(struct sock *sk)
1264 {
1265         struct sk_buff *skb;
1266
1267         /* If ACK arrived pointing to a remembered SACK,
1268          * it means that our remembered SACKs do not reflect
1269          * real state of receiver i.e.
1270          * receiver _host_ is heavily congested (or buggy).
1271          * Do processing similar to RTO timeout.
1272          */
1273         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1274             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1275                 struct inet_connection_sock *icsk = inet_csk(sk);
1276                 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1277
1278                 tcp_enter_loss(sk, 1);
1279                 icsk->icsk_retransmits++;
1280                 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1281                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1282                                           icsk->icsk_rto, TCP_RTO_MAX);
1283                 return 1;
1284         }
1285         return 0;
1286 }
1287
1288 static inline int tcp_fackets_out(struct tcp_sock *tp)
1289 {
1290         return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1291 }
1292
1293 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1294 {
1295         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1296 }
1297
1298 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1299 {
1300         return tp->packets_out &&
1301                tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1302 }
1303
1304 /* Linux NewReno/SACK/FACK/ECN state machine.
1305  * --------------------------------------
1306  *
1307  * "Open"       Normal state, no dubious events, fast path.
1308  * "Disorder"   In all the respects it is "Open",
1309  *              but requires a bit more attention. It is entered when
1310  *              we see some SACKs or dupacks. It is split of "Open"
1311  *              mainly to move some processing from fast path to slow one.
1312  * "CWR"        CWND was reduced due to some Congestion Notification event.
1313  *              It can be ECN, ICMP source quench, local device congestion.
1314  * "Recovery"   CWND was reduced, we are fast-retransmitting.
1315  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
1316  *
1317  * tcp_fastretrans_alert() is entered:
1318  * - each incoming ACK, if state is not "Open"
1319  * - when arrived ACK is unusual, namely:
1320  *      * SACK
1321  *      * Duplicate ACK.
1322  *      * ECN ECE.
1323  *
1324  * Counting packets in flight is pretty simple.
1325  *
1326  *      in_flight = packets_out - left_out + retrans_out
1327  *
1328  *      packets_out is SND.NXT-SND.UNA counted in packets.
1329  *
1330  *      retrans_out is number of retransmitted segments.
1331  *
1332  *      left_out is number of segments left network, but not ACKed yet.
1333  *
1334  *              left_out = sacked_out + lost_out
1335  *
1336  *     sacked_out: Packets, which arrived to receiver out of order
1337  *                 and hence not ACKed. With SACKs this number is simply
1338  *                 amount of SACKed data. Even without SACKs
1339  *                 it is easy to give pretty reliable estimate of this number,
1340  *                 counting duplicate ACKs.
1341  *
1342  *       lost_out: Packets lost by network. TCP has no explicit
1343  *                 "loss notification" feedback from network (for now).
1344  *                 It means that this number can be only _guessed_.
1345  *                 Actually, it is the heuristics to predict lossage that
1346  *                 distinguishes different algorithms.
1347  *
1348  *      F.e. after RTO, when all the queue is considered as lost,
1349  *      lost_out = packets_out and in_flight = retrans_out.
1350  *
1351  *              Essentially, we have now two algorithms counting
1352  *              lost packets.
1353  *
1354  *              FACK: It is the simplest heuristics. As soon as we decided
1355  *              that something is lost, we decide that _all_ not SACKed
1356  *              packets until the most forward SACK are lost. I.e.
1357  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
1358  *              It is absolutely correct estimate, if network does not reorder
1359  *              packets. And it loses any connection to reality when reordering
1360  *              takes place. We use FACK by default until reordering
1361  *              is suspected on the path to this destination.
1362  *
1363  *              NewReno: when Recovery is entered, we assume that one segment
1364  *              is lost (classic Reno). While we are in Recovery and
1365  *              a partial ACK arrives, we assume that one more packet
1366  *              is lost (NewReno). This heuristics are the same in NewReno
1367  *              and SACK.
1368  *
1369  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
1370  *  deflation etc. CWND is real congestion window, never inflated, changes
1371  *  only according to classic VJ rules.
1372  *
1373  * Really tricky (and requiring careful tuning) part of algorithm
1374  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1375  * The first determines the moment _when_ we should reduce CWND and,
1376  * hence, slow down forward transmission. In fact, it determines the moment
1377  * when we decide that hole is caused by loss, rather than by a reorder.
1378  *
1379  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1380  * holes, caused by lost packets.
1381  *
1382  * And the most logically complicated part of algorithm is undo
1383  * heuristics. We detect false retransmits due to both too early
1384  * fast retransmit (reordering) and underestimated RTO, analyzing
1385  * timestamps and D-SACKs. When we detect that some segments were
1386  * retransmitted by mistake and CWND reduction was wrong, we undo
1387  * window reduction and abort recovery phase. This logic is hidden
1388  * inside several functions named tcp_try_undo_<something>.
1389  */
1390
1391 /* This function decides, when we should leave Disordered state
1392  * and enter Recovery phase, reducing congestion window.
1393  *
1394  * Main question: may we further continue forward transmission
1395  * with the same cwnd?
1396  */
1397 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1398 {
1399         __u32 packets_out;
1400
1401         /* Trick#1: The loss is proven. */
1402         if (tp->lost_out)
1403                 return 1;
1404
1405         /* Not-A-Trick#2 : Classic rule... */
1406         if (tcp_fackets_out(tp) > tp->reordering)
1407                 return 1;
1408
1409         /* Trick#3 : when we use RFC2988 timer restart, fast
1410          * retransmit can be triggered by timeout of queue head.
1411          */
1412         if (tcp_head_timedout(sk, tp))
1413                 return 1;
1414
1415         /* Trick#4: It is still not OK... But will it be useful to delay
1416          * recovery more?
1417          */
1418         packets_out = tp->packets_out;
1419         if (packets_out <= tp->reordering &&
1420             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1421             !tcp_may_send_now(sk, tp)) {
1422                 /* We have nothing to send. This connection is limited
1423                  * either by receiver window or by application.
1424                  */
1425                 return 1;
1426         }
1427
1428         return 0;
1429 }
1430
1431 /* If we receive more dupacks than we expected counting segments
1432  * in assumption of absent reordering, interpret this as reordering.
1433  * The only another reason could be bug in receiver TCP.
1434  */
1435 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1436 {
1437         struct tcp_sock *tp = tcp_sk(sk);
1438         u32 holes;
1439
1440         holes = max(tp->lost_out, 1U);
1441         holes = min(holes, tp->packets_out);
1442
1443         if ((tp->sacked_out + holes) > tp->packets_out) {
1444                 tp->sacked_out = tp->packets_out - holes;
1445                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1446         }
1447 }
1448
1449 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1450
1451 static void tcp_add_reno_sack(struct sock *sk)
1452 {
1453         struct tcp_sock *tp = tcp_sk(sk);
1454         tp->sacked_out++;
1455         tcp_check_reno_reordering(sk, 0);
1456         tcp_sync_left_out(tp);
1457 }
1458
1459 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1460
1461 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1462 {
1463         if (acked > 0) {
1464                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1465                 if (acked-1 >= tp->sacked_out)
1466                         tp->sacked_out = 0;
1467                 else
1468                         tp->sacked_out -= acked-1;
1469         }
1470         tcp_check_reno_reordering(sk, acked);
1471         tcp_sync_left_out(tp);
1472 }
1473
1474 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1475 {
1476         tp->sacked_out = 0;
1477         tp->left_out = tp->lost_out;
1478 }
1479
1480 /* Mark head of queue up as lost. */
1481 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1482                                int packets, u32 high_seq)
1483 {
1484         struct sk_buff *skb;
1485         int cnt = packets;
1486
1487         BUG_TRAP(cnt <= tp->packets_out);
1488
1489         sk_stream_for_retrans_queue(skb, sk) {
1490                 cnt -= tcp_skb_pcount(skb);
1491                 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1492                         break;
1493                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1494                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1495                         tp->lost_out += tcp_skb_pcount(skb);
1496                 }
1497         }
1498         tcp_sync_left_out(tp);
1499 }
1500
1501 /* Account newly detected lost packet(s) */
1502
1503 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1504 {
1505         if (IsFack(tp)) {
1506                 int lost = tp->fackets_out - tp->reordering;
1507                 if (lost <= 0)
1508                         lost = 1;
1509                 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1510         } else {
1511                 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1512         }
1513
1514         /* New heuristics: it is possible only after we switched
1515          * to restart timer each time when something is ACKed.
1516          * Hence, we can detect timed out packets during fast
1517          * retransmit without falling to slow start.
1518          */
1519         if (tcp_head_timedout(sk, tp)) {
1520                 struct sk_buff *skb;
1521
1522                 sk_stream_for_retrans_queue(skb, sk) {
1523                         if (tcp_skb_timedout(sk, skb) &&
1524                             !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1525                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1526                                 tp->lost_out += tcp_skb_pcount(skb);
1527                         }
1528                 }
1529                 tcp_sync_left_out(tp);
1530         }
1531 }
1532
1533 /* CWND moderation, preventing bursts due to too big ACKs
1534  * in dubious situations.
1535  */
1536 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1537 {
1538         tp->snd_cwnd = min(tp->snd_cwnd,
1539                            tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1540         tp->snd_cwnd_stamp = tcp_time_stamp;
1541 }
1542
1543 /* Decrease cwnd each second ack. */
1544 static void tcp_cwnd_down(struct sock *sk)
1545 {
1546         const struct inet_connection_sock *icsk = inet_csk(sk);
1547         struct tcp_sock *tp = tcp_sk(sk);
1548         int decr = tp->snd_cwnd_cnt + 1;
1549
1550         tp->snd_cwnd_cnt = decr&1;
1551         decr >>= 1;
1552
1553         if (decr && tp->snd_cwnd > icsk->icsk_ca_ops->min_cwnd(sk))
1554                 tp->snd_cwnd -= decr;
1555
1556         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1557         tp->snd_cwnd_stamp = tcp_time_stamp;
1558 }
1559
1560 /* Nothing was retransmitted or returned timestamp is less
1561  * than timestamp of the first retransmission.
1562  */
1563 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1564 {
1565         return !tp->retrans_stamp ||
1566                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1567                  (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1568 }
1569
1570 /* Undo procedures. */
1571
1572 #if FASTRETRANS_DEBUG > 1
1573 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1574 {
1575         struct inet_sock *inet = inet_sk(sk);
1576         printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1577                msg,
1578                NIPQUAD(inet->daddr), ntohs(inet->dport),
1579                tp->snd_cwnd, tp->left_out,
1580                tp->snd_ssthresh, tp->prior_ssthresh,
1581                tp->packets_out);
1582 }
1583 #else
1584 #define DBGUNDO(x...) do { } while (0)
1585 #endif
1586
1587 static void tcp_undo_cwr(struct sock *sk, const int undo)
1588 {
1589         struct tcp_sock *tp = tcp_sk(sk);
1590
1591         if (tp->prior_ssthresh) {
1592                 const struct inet_connection_sock *icsk = inet_csk(sk);
1593
1594                 if (icsk->icsk_ca_ops->undo_cwnd)
1595                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1596                 else
1597                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1598
1599                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1600                         tp->snd_ssthresh = tp->prior_ssthresh;
1601                         TCP_ECN_withdraw_cwr(tp);
1602                 }
1603         } else {
1604                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1605         }
1606         tcp_moderate_cwnd(tp);
1607         tp->snd_cwnd_stamp = tcp_time_stamp;
1608 }
1609
1610 static inline int tcp_may_undo(struct tcp_sock *tp)
1611 {
1612         return tp->undo_marker &&
1613                 (!tp->undo_retrans || tcp_packet_delayed(tp));
1614 }
1615
1616 /* People celebrate: "We love our President!" */
1617 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1618 {
1619         if (tcp_may_undo(tp)) {
1620                 /* Happy end! We did not retransmit anything
1621                  * or our original transmission succeeded.
1622                  */
1623                 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1624                 tcp_undo_cwr(sk, 1);
1625                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1626                         NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1627                 else
1628                         NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1629                 tp->undo_marker = 0;
1630         }
1631         if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1632                 /* Hold old state until something *above* high_seq
1633                  * is ACKed. For Reno it is MUST to prevent false
1634                  * fast retransmits (RFC2582). SACK TCP is safe. */
1635                 tcp_moderate_cwnd(tp);
1636                 return 1;
1637         }
1638         tcp_set_ca_state(sk, TCP_CA_Open);
1639         return 0;
1640 }
1641
1642 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1643 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1644 {
1645         if (tp->undo_marker && !tp->undo_retrans) {
1646                 DBGUNDO(sk, tp, "D-SACK");
1647                 tcp_undo_cwr(sk, 1);
1648                 tp->undo_marker = 0;
1649                 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1650         }
1651 }
1652
1653 /* Undo during fast recovery after partial ACK. */
1654
1655 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1656                                 int acked)
1657 {
1658         /* Partial ACK arrived. Force Hoe's retransmit. */
1659         int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1660
1661         if (tcp_may_undo(tp)) {
1662                 /* Plain luck! Hole if filled with delayed
1663                  * packet, rather than with a retransmit.
1664                  */
1665                 if (tp->retrans_out == 0)
1666                         tp->retrans_stamp = 0;
1667
1668                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1669
1670                 DBGUNDO(sk, tp, "Hoe");
1671                 tcp_undo_cwr(sk, 0);
1672                 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1673
1674                 /* So... Do not make Hoe's retransmit yet.
1675                  * If the first packet was delayed, the rest
1676                  * ones are most probably delayed as well.
1677                  */
1678                 failed = 0;
1679         }
1680         return failed;
1681 }
1682
1683 /* Undo during loss recovery after partial ACK. */
1684 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1685 {
1686         if (tcp_may_undo(tp)) {
1687                 struct sk_buff *skb;
1688                 sk_stream_for_retrans_queue(skb, sk) {
1689                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1690                 }
1691                 DBGUNDO(sk, tp, "partial loss");
1692                 tp->lost_out = 0;
1693                 tp->left_out = tp->sacked_out;
1694                 tcp_undo_cwr(sk, 1);
1695                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1696                 inet_csk(sk)->icsk_retransmits = 0;
1697                 tp->undo_marker = 0;
1698                 if (!IsReno(tp))
1699                         tcp_set_ca_state(sk, TCP_CA_Open);
1700                 return 1;
1701         }
1702         return 0;
1703 }
1704
1705 static inline void tcp_complete_cwr(struct sock *sk)
1706 {
1707         struct tcp_sock *tp = tcp_sk(sk);
1708         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1709         tp->snd_cwnd_stamp = tcp_time_stamp;
1710         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1711 }
1712
1713 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1714 {
1715         tp->left_out = tp->sacked_out;
1716
1717         if (tp->retrans_out == 0)
1718                 tp->retrans_stamp = 0;
1719
1720         if (flag&FLAG_ECE)
1721                 tcp_enter_cwr(sk);
1722
1723         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1724                 int state = TCP_CA_Open;
1725
1726                 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1727                         state = TCP_CA_Disorder;
1728
1729                 if (inet_csk(sk)->icsk_ca_state != state) {
1730                         tcp_set_ca_state(sk, state);
1731                         tp->high_seq = tp->snd_nxt;
1732                 }
1733                 tcp_moderate_cwnd(tp);
1734         } else {
1735                 tcp_cwnd_down(sk);
1736         }
1737 }
1738
1739 /* Process an event, which can update packets-in-flight not trivially.
1740  * Main goal of this function is to calculate new estimate for left_out,
1741  * taking into account both packets sitting in receiver's buffer and
1742  * packets lost by network.
1743  *
1744  * Besides that it does CWND reduction, when packet loss is detected
1745  * and changes state of machine.
1746  *
1747  * It does _not_ decide what to send, it is made in function
1748  * tcp_xmit_retransmit_queue().
1749  */
1750 static void
1751 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1752                       int prior_packets, int flag)
1753 {
1754         struct inet_connection_sock *icsk = inet_csk(sk);
1755         struct tcp_sock *tp = tcp_sk(sk);
1756         int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1757
1758         /* Some technical things:
1759          * 1. Reno does not count dupacks (sacked_out) automatically. */
1760         if (!tp->packets_out)
1761                 tp->sacked_out = 0;
1762         /* 2. SACK counts snd_fack in packets inaccurately. */
1763         if (tp->sacked_out == 0)
1764                 tp->fackets_out = 0;
1765
1766         /* Now state machine starts.
1767          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1768         if (flag&FLAG_ECE)
1769                 tp->prior_ssthresh = 0;
1770
1771         /* B. In all the states check for reneging SACKs. */
1772         if (tp->sacked_out && tcp_check_sack_reneging(sk))
1773                 return;
1774
1775         /* C. Process data loss notification, provided it is valid. */
1776         if ((flag&FLAG_DATA_LOST) &&
1777             before(tp->snd_una, tp->high_seq) &&
1778             icsk->icsk_ca_state != TCP_CA_Open &&
1779             tp->fackets_out > tp->reordering) {
1780                 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1781                 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1782         }
1783
1784         /* D. Synchronize left_out to current state. */
1785         tcp_sync_left_out(tp);
1786
1787         /* E. Check state exit conditions. State can be terminated
1788          *    when high_seq is ACKed. */
1789         if (icsk->icsk_ca_state == TCP_CA_Open) {
1790                 if (!sysctl_tcp_frto)
1791                         BUG_TRAP(tp->retrans_out == 0);
1792                 tp->retrans_stamp = 0;
1793         } else if (!before(tp->snd_una, tp->high_seq)) {
1794                 switch (icsk->icsk_ca_state) {
1795                 case TCP_CA_Loss:
1796                         icsk->icsk_retransmits = 0;
1797                         if (tcp_try_undo_recovery(sk, tp))
1798                                 return;
1799                         break;
1800
1801                 case TCP_CA_CWR:
1802                         /* CWR is to be held something *above* high_seq
1803                          * is ACKed for CWR bit to reach receiver. */
1804                         if (tp->snd_una != tp->high_seq) {
1805                                 tcp_complete_cwr(sk);
1806                                 tcp_set_ca_state(sk, TCP_CA_Open);
1807                         }
1808                         break;
1809
1810                 case TCP_CA_Disorder:
1811                         tcp_try_undo_dsack(sk, tp);
1812                         if (!tp->undo_marker ||
1813                             /* For SACK case do not Open to allow to undo
1814                              * catching for all duplicate ACKs. */
1815                             IsReno(tp) || tp->snd_una != tp->high_seq) {
1816                                 tp->undo_marker = 0;
1817                                 tcp_set_ca_state(sk, TCP_CA_Open);
1818                         }
1819                         break;
1820
1821                 case TCP_CA_Recovery:
1822                         if (IsReno(tp))
1823                                 tcp_reset_reno_sack(tp);
1824                         if (tcp_try_undo_recovery(sk, tp))
1825                                 return;
1826                         tcp_complete_cwr(sk);
1827                         break;
1828                 }
1829         }
1830
1831         /* F. Process state. */
1832         switch (icsk->icsk_ca_state) {
1833         case TCP_CA_Recovery:
1834                 if (prior_snd_una == tp->snd_una) {
1835                         if (IsReno(tp) && is_dupack)
1836                                 tcp_add_reno_sack(sk);
1837                 } else {
1838                         int acked = prior_packets - tp->packets_out;
1839                         if (IsReno(tp))
1840                                 tcp_remove_reno_sacks(sk, tp, acked);
1841                         is_dupack = tcp_try_undo_partial(sk, tp, acked);
1842                 }
1843                 break;
1844         case TCP_CA_Loss:
1845                 if (flag&FLAG_DATA_ACKED)
1846                         icsk->icsk_retransmits = 0;
1847                 if (!tcp_try_undo_loss(sk, tp)) {
1848                         tcp_moderate_cwnd(tp);
1849                         tcp_xmit_retransmit_queue(sk);
1850                         return;
1851                 }
1852                 if (icsk->icsk_ca_state != TCP_CA_Open)
1853                         return;
1854                 /* Loss is undone; fall through to processing in Open state. */
1855         default:
1856                 if (IsReno(tp)) {
1857                         if (tp->snd_una != prior_snd_una)
1858                                 tcp_reset_reno_sack(tp);
1859                         if (is_dupack)
1860                                 tcp_add_reno_sack(sk);
1861                 }
1862
1863                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
1864                         tcp_try_undo_dsack(sk, tp);
1865
1866                 if (!tcp_time_to_recover(sk, tp)) {
1867                         tcp_try_to_open(sk, tp, flag);
1868                         return;
1869                 }
1870
1871                 /* Otherwise enter Recovery state */
1872
1873                 if (IsReno(tp))
1874                         NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1875                 else
1876                         NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1877
1878                 tp->high_seq = tp->snd_nxt;
1879                 tp->prior_ssthresh = 0;
1880                 tp->undo_marker = tp->snd_una;
1881                 tp->undo_retrans = tp->retrans_out;
1882
1883                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
1884                         if (!(flag&FLAG_ECE))
1885                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1886                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1887                         TCP_ECN_queue_cwr(tp);
1888                 }
1889
1890                 tp->bytes_acked = 0;
1891                 tp->snd_cwnd_cnt = 0;
1892                 tcp_set_ca_state(sk, TCP_CA_Recovery);
1893         }
1894
1895         if (is_dupack || tcp_head_timedout(sk, tp))
1896                 tcp_update_scoreboard(sk, tp);
1897         tcp_cwnd_down(sk);
1898         tcp_xmit_retransmit_queue(sk);
1899 }
1900
1901 /* Read draft-ietf-tcplw-high-performance before mucking
1902  * with this code. (Superceeds RFC1323)
1903  */
1904 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
1905 {
1906         /* RTTM Rule: A TSecr value received in a segment is used to
1907          * update the averaged RTT measurement only if the segment
1908          * acknowledges some new data, i.e., only if it advances the
1909          * left edge of the send window.
1910          *
1911          * See draft-ietf-tcplw-high-performance-00, section 3.3.
1912          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1913          *
1914          * Changed: reset backoff as soon as we see the first valid sample.
1915          * If we do not, we get strongly overstimated rto. With timestamps
1916          * samples are accepted even from very old segments: f.e., when rtt=1
1917          * increases to 8, we retransmit 5 times and after 8 seconds delayed
1918          * answer arrives rto becomes 120 seconds! If at least one of segments
1919          * in window is lost... Voila.                          --ANK (010210)
1920          */
1921         struct tcp_sock *tp = tcp_sk(sk);
1922         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
1923         tcp_rtt_estimator(sk, seq_rtt);
1924         tcp_set_rto(sk);
1925         inet_csk(sk)->icsk_backoff = 0;
1926         tcp_bound_rto(sk);
1927 }
1928
1929 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
1930 {
1931         /* We don't have a timestamp. Can only use
1932          * packets that are not retransmitted to determine
1933          * rtt estimates. Also, we must not reset the
1934          * backoff for rto until we get a non-retransmitted
1935          * packet. This allows us to deal with a situation
1936          * where the network delay has increased suddenly.
1937          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1938          */
1939
1940         if (flag & FLAG_RETRANS_DATA_ACKED)
1941                 return;
1942
1943         tcp_rtt_estimator(sk, seq_rtt);
1944         tcp_set_rto(sk);
1945         inet_csk(sk)->icsk_backoff = 0;
1946         tcp_bound_rto(sk);
1947 }
1948
1949 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
1950                                       const s32 seq_rtt)
1951 {
1952         const struct tcp_sock *tp = tcp_sk(sk);
1953         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1954         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
1955                 tcp_ack_saw_tstamp(sk, flag);
1956         else if (seq_rtt >= 0)
1957                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
1958 }
1959
1960 static inline void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
1961                                   u32 in_flight, int good)
1962 {
1963         const struct inet_connection_sock *icsk = inet_csk(sk);
1964         icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
1965         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
1966 }
1967
1968 /* Restart timer after forward progress on connection.
1969  * RFC2988 recommends to restart timer to now+rto.
1970  */
1971
1972 static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
1973 {
1974         if (!tp->packets_out) {
1975                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
1976         } else {
1977                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1978         }
1979 }
1980
1981 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
1982                          __u32 now, __s32 *seq_rtt)
1983 {
1984         struct tcp_sock *tp = tcp_sk(sk);
1985         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
1986         __u32 seq = tp->snd_una;
1987         __u32 packets_acked;
1988         int acked = 0;
1989
1990         /* If we get here, the whole TSO packet has not been
1991          * acked.
1992          */
1993         BUG_ON(!after(scb->end_seq, seq));
1994
1995         packets_acked = tcp_skb_pcount(skb);
1996         if (tcp_trim_head(sk, skb, seq - scb->seq))
1997                 return 0;
1998         packets_acked -= tcp_skb_pcount(skb);
1999
2000         if (packets_acked) {
2001                 __u8 sacked = scb->sacked;
2002
2003                 acked |= FLAG_DATA_ACKED;
2004                 if (sacked) {
2005                         if (sacked & TCPCB_RETRANS) {
2006                                 if (sacked & TCPCB_SACKED_RETRANS)
2007                                         tp->retrans_out -= packets_acked;
2008                                 acked |= FLAG_RETRANS_DATA_ACKED;
2009                                 *seq_rtt = -1;
2010                         } else if (*seq_rtt < 0)
2011                                 *seq_rtt = now - scb->when;
2012                         if (sacked & TCPCB_SACKED_ACKED)
2013                                 tp->sacked_out -= packets_acked;
2014                         if (sacked & TCPCB_LOST)
2015                                 tp->lost_out -= packets_acked;
2016                         if (sacked & TCPCB_URG) {
2017                                 if (tp->urg_mode &&
2018                                     !before(seq, tp->snd_up))
2019                                         tp->urg_mode = 0;
2020                         }
2021                 } else if (*seq_rtt < 0)
2022                         *seq_rtt = now - scb->when;
2023
2024                 if (tp->fackets_out) {
2025                         __u32 dval = min(tp->fackets_out, packets_acked);
2026                         tp->fackets_out -= dval;
2027                 }
2028                 tp->packets_out -= packets_acked;
2029
2030                 BUG_ON(tcp_skb_pcount(skb) == 0);
2031                 BUG_ON(!before(scb->seq, scb->end_seq));
2032         }
2033
2034         return acked;
2035 }
2036
2037 static inline u32 tcp_usrtt(const struct sk_buff *skb)
2038 {
2039         struct timeval tv, now;
2040
2041         do_gettimeofday(&now);
2042         skb_get_timestamp(skb, &tv);
2043         return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
2044 }
2045
2046 /* Remove acknowledged frames from the retransmission queue. */
2047 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2048 {
2049         struct tcp_sock *tp = tcp_sk(sk);
2050         const struct inet_connection_sock *icsk = inet_csk(sk);
2051         struct sk_buff *skb;
2052         __u32 now = tcp_time_stamp;
2053         int acked = 0;
2054         __s32 seq_rtt = -1;
2055         u32 pkts_acked = 0;
2056         void (*rtt_sample)(struct sock *sk, u32 usrtt)
2057                 = icsk->icsk_ca_ops->rtt_sample;
2058
2059         while ((skb = skb_peek(&sk->sk_write_queue)) &&
2060                skb != sk->sk_send_head) {
2061                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2062                 __u8 sacked = scb->sacked;
2063
2064                 /* If our packet is before the ack sequence we can
2065                  * discard it as it's confirmed to have arrived at
2066                  * the other end.
2067                  */
2068                 if (after(scb->end_seq, tp->snd_una)) {
2069                         if (tcp_skb_pcount(skb) > 1 &&
2070                             after(tp->snd_una, scb->seq))
2071                                 acked |= tcp_tso_acked(sk, skb,
2072                                                        now, &seq_rtt);
2073                         break;
2074                 }
2075
2076                 /* Initial outgoing SYN's get put onto the write_queue
2077                  * just like anything else we transmit.  It is not
2078                  * true data, and if we misinform our callers that
2079                  * this ACK acks real data, we will erroneously exit
2080                  * connection startup slow start one packet too
2081                  * quickly.  This is severely frowned upon behavior.
2082                  */
2083                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2084                         acked |= FLAG_DATA_ACKED;
2085                         ++pkts_acked;
2086                 } else {
2087                         acked |= FLAG_SYN_ACKED;
2088                         tp->retrans_stamp = 0;
2089                 }
2090
2091                 if (sacked) {
2092                         if (sacked & TCPCB_RETRANS) {
2093                                 if(sacked & TCPCB_SACKED_RETRANS)
2094                                         tp->retrans_out -= tcp_skb_pcount(skb);
2095                                 acked |= FLAG_RETRANS_DATA_ACKED;
2096                                 seq_rtt = -1;
2097                         } else if (seq_rtt < 0) {
2098                                 seq_rtt = now - scb->when;
2099                                 if (rtt_sample)
2100                                         (*rtt_sample)(sk, tcp_usrtt(skb));
2101                         }
2102                         if (sacked & TCPCB_SACKED_ACKED)
2103                                 tp->sacked_out -= tcp_skb_pcount(skb);
2104                         if (sacked & TCPCB_LOST)
2105                                 tp->lost_out -= tcp_skb_pcount(skb);
2106                         if (sacked & TCPCB_URG) {
2107                                 if (tp->urg_mode &&
2108                                     !before(scb->end_seq, tp->snd_up))
2109                                         tp->urg_mode = 0;
2110                         }
2111                 } else if (seq_rtt < 0) {
2112                         seq_rtt = now - scb->when;
2113                         if (rtt_sample)
2114                                 (*rtt_sample)(sk, tcp_usrtt(skb));
2115                 }
2116                 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2117                 tcp_packets_out_dec(tp, skb);
2118                 __skb_unlink(skb, &sk->sk_write_queue);
2119                 sk_stream_free_skb(sk, skb);
2120         }
2121
2122         if (acked&FLAG_ACKED) {
2123                 tcp_ack_update_rtt(sk, acked, seq_rtt);
2124                 tcp_ack_packets_out(sk, tp);
2125
2126                 if (icsk->icsk_ca_ops->pkts_acked)
2127                         icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2128         }
2129
2130 #if FASTRETRANS_DEBUG > 0
2131         BUG_TRAP((int)tp->sacked_out >= 0);
2132         BUG_TRAP((int)tp->lost_out >= 0);
2133         BUG_TRAP((int)tp->retrans_out >= 0);
2134         if (!tp->packets_out && tp->rx_opt.sack_ok) {
2135                 const struct inet_connection_sock *icsk = inet_csk(sk);
2136                 if (tp->lost_out) {
2137                         printk(KERN_DEBUG "Leak l=%u %d\n",
2138                                tp->lost_out, icsk->icsk_ca_state);
2139                         tp->lost_out = 0;
2140                 }
2141                 if (tp->sacked_out) {
2142                         printk(KERN_DEBUG "Leak s=%u %d\n",
2143                                tp->sacked_out, icsk->icsk_ca_state);
2144                         tp->sacked_out = 0;
2145                 }
2146                 if (tp->retrans_out) {
2147                         printk(KERN_DEBUG "Leak r=%u %d\n",
2148                                tp->retrans_out, icsk->icsk_ca_state);
2149                         tp->retrans_out = 0;
2150                 }
2151         }
2152 #endif
2153         *seq_rtt_p = seq_rtt;
2154         return acked;
2155 }
2156
2157 static void tcp_ack_probe(struct sock *sk)
2158 {
2159         const struct tcp_sock *tp = tcp_sk(sk);
2160         struct inet_connection_sock *icsk = inet_csk(sk);
2161
2162         /* Was it a usable window open? */
2163
2164         if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2165                    tp->snd_una + tp->snd_wnd)) {
2166                 icsk->icsk_backoff = 0;
2167                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2168                 /* Socket must be waked up by subsequent tcp_data_snd_check().
2169                  * This function is not for random using!
2170                  */
2171         } else {
2172                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2173                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2174                                           TCP_RTO_MAX);
2175         }
2176 }
2177
2178 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2179 {
2180         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2181                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2182 }
2183
2184 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2185 {
2186         const struct tcp_sock *tp = tcp_sk(sk);
2187         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2188                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2189 }
2190
2191 /* Check that window update is acceptable.
2192  * The function assumes that snd_una<=ack<=snd_next.
2193  */
2194 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2195                                         const u32 ack_seq, const u32 nwin)
2196 {
2197         return (after(ack, tp->snd_una) ||
2198                 after(ack_seq, tp->snd_wl1) ||
2199                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2200 }
2201
2202 /* Update our send window.
2203  *
2204  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2205  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2206  */
2207 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2208                                  struct sk_buff *skb, u32 ack, u32 ack_seq)
2209 {
2210         int flag = 0;
2211         u32 nwin = ntohs(skb->h.th->window);
2212
2213         if (likely(!skb->h.th->syn))
2214                 nwin <<= tp->rx_opt.snd_wscale;
2215
2216         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2217                 flag |= FLAG_WIN_UPDATE;
2218                 tcp_update_wl(tp, ack, ack_seq);
2219
2220                 if (tp->snd_wnd != nwin) {
2221                         tp->snd_wnd = nwin;
2222
2223                         /* Note, it is the only place, where
2224                          * fast path is recovered for sending TCP.
2225                          */
2226                         tp->pred_flags = 0;
2227                         tcp_fast_path_check(sk, tp);
2228
2229                         if (nwin > tp->max_window) {
2230                                 tp->max_window = nwin;
2231                                 tcp_sync_mss(sk, tp->pmtu_cookie);
2232                         }
2233                 }
2234         }
2235
2236         tp->snd_una = ack;
2237
2238         return flag;
2239 }
2240
2241 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2242 {
2243         struct tcp_sock *tp = tcp_sk(sk);
2244
2245         tcp_sync_left_out(tp);
2246
2247         if (tp->snd_una == prior_snd_una ||
2248             !before(tp->snd_una, tp->frto_highmark)) {
2249                 /* RTO was caused by loss, start retransmitting in
2250                  * go-back-N slow start
2251                  */
2252                 tcp_enter_frto_loss(sk);
2253                 return;
2254         }
2255
2256         if (tp->frto_counter == 1) {
2257                 /* First ACK after RTO advances the window: allow two new
2258                  * segments out.
2259                  */
2260                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2261         } else {
2262                 /* Also the second ACK after RTO advances the window.
2263                  * The RTO was likely spurious. Reduce cwnd and continue
2264                  * in congestion avoidance
2265                  */
2266                 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2267                 tcp_moderate_cwnd(tp);
2268         }
2269
2270         /* F-RTO affects on two new ACKs following RTO.
2271          * At latest on third ACK the TCP behavor is back to normal.
2272          */
2273         tp->frto_counter = (tp->frto_counter + 1) % 3;
2274 }
2275
2276 /* This routine deals with incoming acks, but not outgoing ones. */
2277 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2278 {
2279         struct inet_connection_sock *icsk = inet_csk(sk);
2280         struct tcp_sock *tp = tcp_sk(sk);
2281         u32 prior_snd_una = tp->snd_una;
2282         u32 ack_seq = TCP_SKB_CB(skb)->seq;
2283         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2284         u32 prior_in_flight;
2285         s32 seq_rtt;
2286         int prior_packets;
2287
2288         /* If the ack is newer than sent or older than previous acks
2289          * then we can probably ignore it.
2290          */
2291         if (after(ack, tp->snd_nxt))
2292                 goto uninteresting_ack;
2293
2294         if (before(ack, prior_snd_una))
2295                 goto old_ack;
2296
2297         if (sysctl_tcp_abc && icsk->icsk_ca_state < TCP_CA_CWR)
2298                 tp->bytes_acked += ack - prior_snd_una;
2299
2300         if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2301                 /* Window is constant, pure forward advance.
2302                  * No more checks are required.
2303                  * Note, we use the fact that SND.UNA>=SND.WL2.
2304                  */
2305                 tcp_update_wl(tp, ack, ack_seq);
2306                 tp->snd_una = ack;
2307                 flag |= FLAG_WIN_UPDATE;
2308
2309                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2310
2311                 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2312         } else {
2313                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2314                         flag |= FLAG_DATA;
2315                 else
2316                         NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2317
2318                 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2319
2320                 if (TCP_SKB_CB(skb)->sacked)
2321                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2322
2323                 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2324                         flag |= FLAG_ECE;
2325
2326                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2327         }
2328
2329         /* We passed data and got it acked, remove any soft error
2330          * log. Something worked...
2331          */
2332         sk->sk_err_soft = 0;
2333         tp->rcv_tstamp = tcp_time_stamp;
2334         prior_packets = tp->packets_out;
2335         if (!prior_packets)
2336                 goto no_queue;
2337
2338         prior_in_flight = tcp_packets_in_flight(tp);
2339
2340         /* See if we can take anything off of the retransmit queue. */
2341         flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2342
2343         if (tp->frto_counter)
2344                 tcp_process_frto(sk, prior_snd_una);
2345
2346         if (tcp_ack_is_dubious(sk, flag)) {
2347                 /* Advanve CWND, if state allows this. */
2348                 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2349                         tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
2350                 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2351         } else {
2352                 if ((flag & FLAG_DATA_ACKED))
2353                         tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2354         }
2355
2356         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2357                 dst_confirm(sk->sk_dst_cache);
2358
2359         return 1;
2360
2361 no_queue:
2362         icsk->icsk_probes_out = 0;
2363
2364         /* If this ack opens up a zero window, clear backoff.  It was
2365          * being used to time the probes, and is probably far higher than
2366          * it needs to be for normal retransmission.
2367          */
2368         if (sk->sk_send_head)
2369                 tcp_ack_probe(sk);
2370         return 1;
2371
2372 old_ack:
2373         if (TCP_SKB_CB(skb)->sacked)
2374                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2375
2376 uninteresting_ack:
2377         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2378         return 0;
2379 }
2380
2381
2382 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2383  * But, this can also be called on packets in the established flow when
2384  * the fast version below fails.
2385  */
2386 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2387 {
2388         unsigned char *ptr;
2389         struct tcphdr *th = skb->h.th;
2390         int length=(th->doff*4)-sizeof(struct tcphdr);
2391
2392         ptr = (unsigned char *)(th + 1);
2393         opt_rx->saw_tstamp = 0;
2394
2395         while(length>0) {
2396                 int opcode=*ptr++;
2397                 int opsize;
2398
2399                 switch (opcode) {
2400                         case TCPOPT_EOL:
2401                                 return;
2402                         case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
2403                                 length--;
2404                                 continue;
2405                         default:
2406                                 opsize=*ptr++;
2407                                 if (opsize < 2) /* "silly options" */
2408                                         return;
2409                                 if (opsize > length)
2410                                         return; /* don't parse partial options */
2411                                 switch(opcode) {
2412                                 case TCPOPT_MSS:
2413                                         if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2414                                                 u16 in_mss = ntohs(get_unaligned((__u16 *)ptr));
2415                                                 if (in_mss) {
2416                                                         if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2417                                                                 in_mss = opt_rx->user_mss;
2418                                                         opt_rx->mss_clamp = in_mss;
2419                                                 }
2420                                         }
2421                                         break;
2422                                 case TCPOPT_WINDOW:
2423                                         if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2424                                                 if (sysctl_tcp_window_scaling) {
2425                                                         __u8 snd_wscale = *(__u8 *) ptr;
2426                                                         opt_rx->wscale_ok = 1;
2427                                                         if (snd_wscale > 14) {
2428                                                                 if(net_ratelimit())
2429                                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
2430                                                                                "scaling value %d >14 received.\n",
2431                                                                                snd_wscale);
2432                                                                 snd_wscale = 14;
2433                                                         }
2434                                                         opt_rx->snd_wscale = snd_wscale;
2435                                                 }
2436                                         break;
2437                                 case TCPOPT_TIMESTAMP:
2438                                         if(opsize==TCPOLEN_TIMESTAMP) {
2439                                                 if ((estab && opt_rx->tstamp_ok) ||
2440                                                     (!estab && sysctl_tcp_timestamps)) {
2441                                                         opt_rx->saw_tstamp = 1;
2442                                                         opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr));
2443                                                         opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4)));
2444                                                 }
2445                                         }
2446                                         break;
2447                                 case TCPOPT_SACK_PERM:
2448                                         if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2449                                                 if (sysctl_tcp_sack) {
2450                                                         opt_rx->sack_ok = 1;
2451                                                         tcp_sack_reset(opt_rx);
2452                                                 }
2453                                         }
2454                                         break;
2455
2456                                 case TCPOPT_SACK:
2457                                         if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2458                                            !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2459                                            opt_rx->sack_ok) {
2460                                                 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2461                                         }
2462                                 };
2463                                 ptr+=opsize-2;
2464                                 length-=opsize;
2465                 };
2466         }
2467 }
2468
2469 /* Fast parse options. This hopes to only see timestamps.
2470  * If it is wrong it falls back on tcp_parse_options().
2471  */
2472 static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2473                                          struct tcp_sock *tp)
2474 {
2475         if (th->doff == sizeof(struct tcphdr)>>2) {
2476                 tp->rx_opt.saw_tstamp = 0;
2477                 return 0;
2478         } else if (tp->rx_opt.tstamp_ok &&
2479                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2480                 __u32 *ptr = (__u32 *)(th + 1);
2481                 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2482                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2483                         tp->rx_opt.saw_tstamp = 1;
2484                         ++ptr;
2485                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
2486                         ++ptr;
2487                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2488                         return 1;
2489                 }
2490         }
2491         tcp_parse_options(skb, &tp->rx_opt, 1);
2492         return 1;
2493 }
2494
2495 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2496 {
2497         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2498         tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2499 }
2500
2501 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2502 {
2503         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2504                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2505                  * extra check below makes sure this can only happen
2506                  * for pure ACK frames.  -DaveM
2507                  *
2508                  * Not only, also it occurs for expired timestamps.
2509                  */
2510
2511                 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2512                    xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2513                         tcp_store_ts_recent(tp);
2514         }
2515 }
2516
2517 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2518  *
2519  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2520  * it can pass through stack. So, the following predicate verifies that
2521  * this segment is not used for anything but congestion avoidance or
2522  * fast retransmit. Moreover, we even are able to eliminate most of such
2523  * second order effects, if we apply some small "replay" window (~RTO)
2524  * to timestamp space.
2525  *
2526  * All these measures still do not guarantee that we reject wrapped ACKs
2527  * on networks with high bandwidth, when sequence space is recycled fastly,
2528  * but it guarantees that such events will be very rare and do not affect
2529  * connection seriously. This doesn't look nice, but alas, PAWS is really
2530  * buggy extension.
2531  *
2532  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2533  * states that events when retransmit arrives after original data are rare.
2534  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2535  * the biggest problem on large power networks even with minor reordering.
2536  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2537  * up to bandwidth of 18Gigabit/sec. 8) ]
2538  */
2539
2540 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2541 {
2542         struct tcp_sock *tp = tcp_sk(sk);
2543         struct tcphdr *th = skb->h.th;
2544         u32 seq = TCP_SKB_CB(skb)->seq;
2545         u32 ack = TCP_SKB_CB(skb)->ack_seq;
2546
2547         return (/* 1. Pure ACK with correct sequence number. */
2548                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2549
2550                 /* 2. ... and duplicate ACK. */
2551                 ack == tp->snd_una &&
2552
2553                 /* 3. ... and does not update window. */
2554                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2555
2556                 /* 4. ... and sits in replay window. */
2557                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2558 }
2559
2560 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2561 {
2562         const struct tcp_sock *tp = tcp_sk(sk);
2563         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2564                 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2565                 !tcp_disordered_ack(sk, skb));
2566 }
2567
2568 /* Check segment sequence number for validity.
2569  *
2570  * Segment controls are considered valid, if the segment
2571  * fits to the window after truncation to the window. Acceptability
2572  * of data (and SYN, FIN, of course) is checked separately.
2573  * See tcp_data_queue(), for example.
2574  *
2575  * Also, controls (RST is main one) are accepted using RCV.WUP instead
2576  * of RCV.NXT. Peer still did not advance his SND.UNA when we
2577  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2578  * (borrowed from freebsd)
2579  */
2580
2581 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2582 {
2583         return  !before(end_seq, tp->rcv_wup) &&
2584                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2585 }
2586
2587 /* When we get a reset we do this. */
2588 static void tcp_reset(struct sock *sk)
2589 {
2590         /* We want the right error as BSD sees it (and indeed as we do). */
2591         switch (sk->sk_state) {
2592                 case TCP_SYN_SENT:
2593                         sk->sk_err = ECONNREFUSED;
2594                         break;
2595                 case TCP_CLOSE_WAIT:
2596                         sk->sk_err = EPIPE;
2597                         break;
2598                 case TCP_CLOSE:
2599                         return;
2600                 default:
2601                         sk->sk_err = ECONNRESET;
2602         }
2603
2604         if (!sock_flag(sk, SOCK_DEAD))
2605                 sk->sk_error_report(sk);
2606
2607         tcp_done(sk);
2608 }
2609
2610 /*
2611  *      Process the FIN bit. This now behaves as it is supposed to work
2612  *      and the FIN takes effect when it is validly part of sequence
2613  *      space. Not before when we get holes.
2614  *
2615  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2616  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
2617  *      TIME-WAIT)
2618  *
2619  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
2620  *      close and we go into CLOSING (and later onto TIME-WAIT)
2621  *
2622  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2623  */
2624 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2625 {
2626         struct tcp_sock *tp = tcp_sk(sk);
2627
2628         inet_csk_schedule_ack(sk);
2629
2630         sk->sk_shutdown |= RCV_SHUTDOWN;
2631         sock_set_flag(sk, SOCK_DONE);
2632
2633         switch (sk->sk_state) {
2634                 case TCP_SYN_RECV:
2635                 case TCP_ESTABLISHED:
2636                         /* Move to CLOSE_WAIT */
2637                         tcp_set_state(sk, TCP_CLOSE_WAIT);
2638                         inet_csk(sk)->icsk_ack.pingpong = 1;
2639                         break;
2640
2641                 case TCP_CLOSE_WAIT:
2642                 case TCP_CLOSING:
2643                         /* Received a retransmission of the FIN, do
2644                          * nothing.
2645                          */
2646                         break;
2647                 case TCP_LAST_ACK:
2648                         /* RFC793: Remain in the LAST-ACK state. */
2649                         break;
2650
2651                 case TCP_FIN_WAIT1:
2652                         /* This case occurs when a simultaneous close
2653                          * happens, we must ack the received FIN and
2654                          * enter the CLOSING state.
2655                          */
2656                         tcp_send_ack(sk);
2657                         tcp_set_state(sk, TCP_CLOSING);
2658                         break;
2659                 case TCP_FIN_WAIT2:
2660                         /* Received a FIN -- send ACK and enter TIME_WAIT. */
2661                         tcp_send_ack(sk);
2662                         tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2663                         break;
2664                 default:
2665                         /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2666                          * cases we should never reach this piece of code.
2667                          */
2668                         printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2669                                __FUNCTION__, sk->sk_state);
2670                         break;
2671         };
2672
2673         /* It _is_ possible, that we have something out-of-order _after_ FIN.
2674          * Probably, we should reset in this case. For now drop them.
2675          */
2676         __skb_queue_purge(&tp->out_of_order_queue);
2677         if (tp->rx_opt.sack_ok)
2678                 tcp_sack_reset(&tp->rx_opt);
2679         sk_stream_mem_reclaim(sk);
2680
2681         if (!sock_flag(sk, SOCK_DEAD)) {
2682                 sk->sk_state_change(sk);
2683
2684                 /* Do not send POLL_HUP for half duplex close. */
2685                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2686                     sk->sk_state == TCP_CLOSE)
2687                         sk_wake_async(sk, 1, POLL_HUP);
2688                 else
2689                         sk_wake_async(sk, 1, POLL_IN);
2690         }
2691 }
2692
2693 static __inline__ int
2694 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2695 {
2696         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2697                 if (before(seq, sp->start_seq))
2698                         sp->start_seq = seq;
2699                 if (after(end_seq, sp->end_seq))
2700                         sp->end_seq = end_seq;
2701                 return 1;
2702         }
2703         return 0;
2704 }
2705
2706 static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
2707 {
2708         if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2709                 if (before(seq, tp->rcv_nxt))
2710                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
2711                 else
2712                         NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
2713
2714                 tp->rx_opt.dsack = 1;
2715                 tp->duplicate_sack[0].start_seq = seq;
2716                 tp->duplicate_sack[0].end_seq = end_seq;
2717                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
2718         }
2719 }
2720
2721 static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
2722 {
2723         if (!tp->rx_opt.dsack)
2724                 tcp_dsack_set(tp, seq, end_seq);
2725         else
2726                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
2727 }
2728
2729 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
2730 {
2731         struct tcp_sock *tp = tcp_sk(sk);
2732
2733         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
2734             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
2735                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
2736                 tcp_enter_quickack_mode(sk);
2737
2738                 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
2739                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2740
2741                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
2742                                 end_seq = tp->rcv_nxt;
2743                         tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
2744                 }
2745         }
2746
2747         tcp_send_ack(sk);
2748 }
2749
2750 /* These routines update the SACK block as out-of-order packets arrive or
2751  * in-order packets close up the sequence space.
2752  */
2753 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
2754 {
2755         int this_sack;
2756         struct tcp_sack_block *sp = &tp->selective_acks[0];
2757         struct tcp_sack_block *swalk = sp+1;
2758
2759         /* See if the recent change to the first SACK eats into
2760          * or hits the sequence space of other SACK blocks, if so coalesce.
2761          */
2762         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
2763                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
2764                         int i;
2765
2766                         /* Zap SWALK, by moving every further SACK up by one slot.
2767                          * Decrease num_sacks.
2768                          */
2769                         tp->rx_opt.num_sacks--;
2770                         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2771                         for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
2772                                 sp[i] = sp[i+1];
2773                         continue;
2774                 }
2775                 this_sack++, swalk++;
2776         }
2777 }
2778
2779 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
2780 {
2781         __u32 tmp;
2782
2783         tmp = sack1->start_seq;
2784         sack1->start_seq = sack2->start_seq;
2785         sack2->start_seq = tmp;
2786
2787         tmp = sack1->end_seq;
2788         sack1->end_seq = sack2->end_seq;
2789         sack2->end_seq = tmp;
2790 }
2791
2792 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
2793 {
2794         struct tcp_sock *tp = tcp_sk(sk);
2795         struct tcp_sack_block *sp = &tp->selective_acks[0];
2796         int cur_sacks = tp->rx_opt.num_sacks;
2797         int this_sack;
2798
2799         if (!cur_sacks)
2800                 goto new_sack;
2801
2802         for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
2803                 if (tcp_sack_extend(sp, seq, end_seq)) {
2804                         /* Rotate this_sack to the first one. */
2805                         for (; this_sack>0; this_sack--, sp--)
2806                                 tcp_sack_swap(sp, sp-1);
2807                         if (cur_sacks > 1)
2808                                 tcp_sack_maybe_coalesce(tp);
2809                         return;
2810                 }
2811         }
2812
2813         /* Could not find an adjacent existing SACK, build a new one,
2814          * put it at the front, and shift everyone else down.  We
2815          * always know there is at least one SACK present already here.
2816          *
2817          * If the sack array is full, forget about the last one.
2818          */
2819         if (this_sack >= 4) {
2820                 this_sack--;
2821                 tp->rx_opt.num_sacks--;
2822                 sp--;
2823         }
2824         for(; this_sack > 0; this_sack--, sp--)
2825                 *sp = *(sp-1);
2826
2827 new_sack:
2828         /* Build the new head SACK, and we're done. */
2829         sp->start_seq = seq;
2830         sp->end_seq = end_seq;
2831         tp->rx_opt.num_sacks++;
2832         tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2833 }
2834
2835 /* RCV.NXT advances, some SACKs should be eaten. */
2836
2837 static void tcp_sack_remove(struct tcp_sock *tp)
2838 {
2839         struct tcp_sack_block *sp = &tp->selective_acks[0];
2840         int num_sacks = tp->rx_opt.num_sacks;
2841         int this_sack;
2842
2843         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2844         if (skb_queue_empty(&tp->out_of_order_queue)) {
2845                 tp->rx_opt.num_sacks = 0;
2846                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
2847                 return;
2848         }
2849
2850         for(this_sack = 0; this_sack < num_sacks; ) {
2851                 /* Check if the start of the sack is covered by RCV.NXT. */
2852                 if (!before(tp->rcv_nxt, sp->start_seq)) {
2853                         int i;
2854
2855                         /* RCV.NXT must cover all the block! */
2856                         BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
2857
2858                         /* Zap this SACK, by moving forward any other SACKS. */
2859                         for (i=this_sack+1; i < num_sacks; i++)
2860                                 tp->selective_acks[i-1] = tp->selective_acks[i];
2861                         num_sacks--;
2862                         continue;
2863                 }
2864                 this_sack++;
2865                 sp++;
2866         }
2867         if (num_sacks != tp->rx_opt.num_sacks) {
2868                 tp->rx_opt.num_sacks = num_sacks;
2869                 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
2870         }
2871 }
2872
2873 /* This one checks to see if we can put data from the
2874  * out_of_order queue into the receive_queue.
2875  */
2876 static void tcp_ofo_queue(struct sock *sk)
2877 {
2878         struct tcp_sock *tp = tcp_sk(sk);
2879         __u32 dsack_high = tp->rcv_nxt;
2880         struct sk_buff *skb;
2881
2882         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
2883                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
2884                         break;
2885
2886                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
2887                         __u32 dsack = dsack_high;
2888                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
2889                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
2890                         tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
2891                 }
2892
2893                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
2894                         SOCK_DEBUG(sk, "ofo packet was already received \n");
2895                         __skb_unlink(skb, &tp->out_of_order_queue);
2896                         __kfree_skb(skb);
2897                         continue;
2898                 }
2899                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
2900                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
2901                            TCP_SKB_CB(skb)->end_seq);
2902
2903                 __skb_unlink(skb, &tp->out_of_order_queue);
2904                 __skb_queue_tail(&sk->sk_receive_queue, skb);
2905                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2906                 if(skb->h.th->fin)
2907                         tcp_fin(skb, sk, skb->h.th);
2908         }
2909 }
2910
2911 static int tcp_prune_queue(struct sock *sk);
2912
2913 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
2914 {
2915         struct tcphdr *th = skb->h.th;
2916         struct tcp_sock *tp = tcp_sk(sk);
2917         int eaten = -1;
2918
2919         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
2920                 goto drop;
2921
2922         __skb_pull(skb, th->doff*4);
2923
2924         TCP_ECN_accept_cwr(tp, skb);
2925
2926         if (tp->rx_opt.dsack) {
2927                 tp->rx_opt.dsack = 0;
2928                 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
2929                                                     4 - tp->rx_opt.tstamp_ok);
2930         }
2931
2932         /*  Queue data for delivery to the user.
2933          *  Packets in sequence go to the receive queue.
2934          *  Out of sequence packets to the out_of_order_queue.
2935          */
2936         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
2937                 if (tcp_receive_window(tp) == 0)
2938                         goto out_of_window;
2939
2940                 /* Ok. In sequence. In window. */
2941                 if (tp->ucopy.task == current &&
2942                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
2943                     sock_owned_by_user(sk) && !tp->urg_data) {
2944                         int chunk = min_t(unsigned int, skb->len,
2945                                                         tp->ucopy.len);
2946
2947                         __set_current_state(TASK_RUNNING);
2948
2949                         local_bh_enable();
2950                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
2951                                 tp->ucopy.len -= chunk;
2952                                 tp->copied_seq += chunk;
2953                                 eaten = (chunk == skb->len && !th->fin);
2954                                 tcp_rcv_space_adjust(sk);
2955                         }
2956                         local_bh_disable();
2957                 }
2958
2959                 if (eaten <= 0) {
2960 queue_and_out:
2961                         if (eaten < 0 &&
2962                             (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
2963                              !sk_stream_rmem_schedule(sk, skb))) {
2964                                 if (tcp_prune_queue(sk) < 0 ||
2965                                     !sk_stream_rmem_schedule(sk, skb))
2966                                         goto drop;
2967                         }
2968                         sk_stream_set_owner_r(skb, sk);
2969                         __skb_queue_tail(&sk->sk_receive_queue, skb);
2970                 }
2971                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
2972                 if(skb->len)
2973                         tcp_event_data_recv(sk, tp, skb);
2974                 if(th->fin)
2975                         tcp_fin(skb, sk, th);
2976
2977                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
2978                         tcp_ofo_queue(sk);
2979
2980                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
2981                          * gap in queue is filled.
2982                          */
2983                         if (skb_queue_empty(&tp->out_of_order_queue))
2984                                 inet_csk(sk)->icsk_ack.pingpong = 0;
2985                 }
2986
2987                 if (tp->rx_opt.num_sacks)
2988                         tcp_sack_remove(tp);
2989
2990                 tcp_fast_path_check(sk, tp);
2991
2992                 if (eaten > 0)
2993                         __kfree_skb(skb);
2994                 else if (!sock_flag(sk, SOCK_DEAD))
2995                         sk->sk_data_ready(sk, 0);
2996                 return;
2997         }
2998
2999         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3000                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
3001                 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3002                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3003
3004 out_of_window:
3005                 tcp_enter_quickack_mode(sk);
3006                 inet_csk_schedule_ack(sk);
3007 drop:
3008                 __kfree_skb(skb);
3009                 return;
3010         }
3011
3012         /* Out of window. F.e. zero window probe. */
3013         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3014                 goto out_of_window;
3015
3016         tcp_enter_quickack_mode(sk);
3017
3018         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3019                 /* Partial packet, seq < rcv_next < end_seq */
3020                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3021                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3022                            TCP_SKB_CB(skb)->end_seq);
3023
3024                 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3025
3026                 /* If window is closed, drop tail of packet. But after
3027                  * remembering D-SACK for its head made in previous line.
3028                  */
3029                 if (!tcp_receive_window(tp))
3030                         goto out_of_window;
3031                 goto queue_and_out;
3032         }
3033
3034         TCP_ECN_check_ce(tp, skb);
3035
3036         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3037             !sk_stream_rmem_schedule(sk, skb)) {
3038                 if (tcp_prune_queue(sk) < 0 ||
3039                     !sk_stream_rmem_schedule(sk, skb))
3040                         goto drop;
3041         }
3042
3043         /* Disable header prediction. */
3044         tp->pred_flags = 0;
3045         inet_csk_schedule_ack(sk);
3046
3047         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3048                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3049
3050         sk_stream_set_owner_r(skb, sk);
3051
3052         if (!skb_peek(&tp->out_of_order_queue)) {
3053                 /* Initial out of order segment, build 1 SACK. */
3054                 if (tp->rx_opt.sack_ok) {
3055                         tp->rx_opt.num_sacks = 1;
3056                         tp->rx_opt.dsack     = 0;
3057                         tp->rx_opt.eff_sacks = 1;
3058                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3059                         tp->selective_acks[0].end_seq =
3060                                                 TCP_SKB_CB(skb)->end_seq;
3061                 }
3062                 __skb_queue_head(&tp->out_of_order_queue,skb);
3063         } else {
3064                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3065                 u32 seq = TCP_SKB_CB(skb)->seq;
3066                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3067
3068                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3069                         __skb_append(skb1, skb, &tp->out_of_order_queue);
3070
3071                         if (!tp->rx_opt.num_sacks ||
3072                             tp->selective_acks[0].end_seq != seq)
3073                                 goto add_sack;
3074
3075                         /* Common case: data arrive in order after hole. */
3076                         tp->selective_acks[0].end_seq = end_seq;
3077                         return;
3078                 }
3079
3080                 /* Find place to insert this segment. */
3081                 do {
3082                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
3083                                 break;
3084                 } while ((skb1 = skb1->prev) !=
3085                          (struct sk_buff*)&tp->out_of_order_queue);
3086
3087                 /* Do skb overlap to previous one? */
3088                 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3089                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3090                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3091                                 /* All the bits are present. Drop. */
3092                                 __kfree_skb(skb);
3093                                 tcp_dsack_set(tp, seq, end_seq);
3094                                 goto add_sack;
3095                         }
3096                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3097                                 /* Partial overlap. */
3098                                 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3099                         } else {
3100                                 skb1 = skb1->prev;
3101                         }
3102                 }
3103                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3104
3105                 /* And clean segments covered by new one as whole. */
3106                 while ((skb1 = skb->next) !=
3107                        (struct sk_buff*)&tp->out_of_order_queue &&
3108                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3109                        if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3110                                tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3111                                break;
3112                        }
3113                        __skb_unlink(skb1, &tp->out_of_order_queue);
3114                        tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3115                        __kfree_skb(skb1);
3116                 }
3117
3118 add_sack:
3119                 if (tp->rx_opt.sack_ok)
3120                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
3121         }
3122 }
3123
3124 /* Collapse contiguous sequence of skbs head..tail with
3125  * sequence numbers start..end.
3126  * Segments with FIN/SYN are not collapsed (only because this
3127  * simplifies code)
3128  */
3129 static void
3130 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3131              struct sk_buff *head, struct sk_buff *tail,
3132              u32 start, u32 end)
3133 {
3134         struct sk_buff *skb;
3135
3136         /* First, check that queue is collapsable and find
3137          * the point where collapsing can be useful. */
3138         for (skb = head; skb != tail; ) {
3139                 /* No new bits? It is possible on ofo queue. */
3140                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3141                         struct sk_buff *next = skb->next;
3142                         __skb_unlink(skb, list);
3143                         __kfree_skb(skb);
3144                         NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3145                         skb = next;
3146                         continue;
3147                 }
3148
3149                 /* The first skb to collapse is:
3150                  * - not SYN/FIN and
3151                  * - bloated or contains data before "start" or
3152                  *   overlaps to the next one.
3153                  */
3154                 if (!skb->h.th->syn && !skb->h.th->fin &&
3155                     (tcp_win_from_space(skb->truesize) > skb->len ||
3156                      before(TCP_SKB_CB(skb)->seq, start) ||
3157                      (skb->next != tail &&
3158                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3159                         break;
3160
3161                 /* Decided to skip this, advance start seq. */
3162                 start = TCP_SKB_CB(skb)->end_seq;
3163                 skb = skb->next;
3164         }
3165         if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3166                 return;
3167
3168         while (before(start, end)) {
3169                 struct sk_buff *nskb;
3170                 int header = skb_headroom(skb);
3171                 int copy = SKB_MAX_ORDER(header, 0);
3172
3173                 /* Too big header? This can happen with IPv6. */
3174                 if (copy < 0)
3175                         return;
3176                 if (end-start < copy)
3177                         copy = end-start;
3178                 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3179                 if (!nskb)
3180                         return;
3181                 skb_reserve(nskb, header);
3182                 memcpy(nskb->head, skb->head, header);
3183                 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3184                 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3185                 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3186                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3187                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3188                 __skb_insert(nskb, skb->prev, skb, list);
3189                 sk_stream_set_owner_r(nskb, sk);
3190
3191                 /* Copy data, releasing collapsed skbs. */
3192                 while (copy > 0) {
3193                         int offset = start - TCP_SKB_CB(skb)->seq;
3194                         int size = TCP_SKB_CB(skb)->end_seq - start;
3195
3196                         if (offset < 0) BUG();
3197                         if (size > 0) {
3198                                 size = min(copy, size);
3199                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3200                                         BUG();
3201                                 TCP_SKB_CB(nskb)->end_seq += size;
3202                                 copy -= size;
3203                                 start += size;
3204                         }
3205                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3206                                 struct sk_buff *next = skb->next;
3207                                 __skb_unlink(skb, list);
3208                                 __kfree_skb(skb);
3209                                 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3210                                 skb = next;
3211                                 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3212                                         return;
3213                         }
3214                 }
3215         }
3216 }
3217
3218 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3219  * and tcp_collapse() them until all the queue is collapsed.
3220  */
3221 static void tcp_collapse_ofo_queue(struct sock *sk)
3222 {
3223         struct tcp_sock *tp = tcp_sk(sk);
3224         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3225         struct sk_buff *head;
3226         u32 start, end;
3227
3228         if (skb == NULL)
3229                 return;
3230
3231         start = TCP_SKB_CB(skb)->seq;
3232         end = TCP_SKB_CB(skb)->end_seq;
3233         head = skb;
3234
3235         for (;;) {
3236                 skb = skb->next;
3237
3238                 /* Segment is terminated when we see gap or when
3239                  * we are at the end of all the queue. */
3240                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3241                     after(TCP_SKB_CB(skb)->seq, end) ||
3242                     before(TCP_SKB_CB(skb)->end_seq, start)) {
3243                         tcp_collapse(sk, &tp->out_of_order_queue,
3244                                      head, skb, start, end);
3245                         head = skb;
3246                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3247                                 break;
3248                         /* Start new segment */
3249                         start = TCP_SKB_CB(skb)->seq;
3250                         end = TCP_SKB_CB(skb)->end_seq;
3251                 } else {
3252                         if (before(TCP_SKB_CB(skb)->seq, start))
3253                                 start = TCP_SKB_CB(skb)->seq;
3254                         if (after(TCP_SKB_CB(skb)->end_seq, end))
3255                                 end = TCP_SKB_CB(skb)->end_seq;
3256                 }
3257         }
3258 }
3259
3260 /* Reduce allocated memory if we can, trying to get
3261  * the socket within its memory limits again.
3262  *
3263  * Return less than zero if we should start dropping frames
3264  * until the socket owning process reads some of the data
3265  * to stabilize the situation.
3266  */
3267 static int tcp_prune_queue(struct sock *sk)
3268 {
3269         struct tcp_sock *tp = tcp_sk(sk);
3270
3271         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3272
3273         NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3274
3275         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3276                 tcp_clamp_window(sk, tp);
3277         else if (tcp_memory_pressure)
3278                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3279
3280         tcp_collapse_ofo_queue(sk);
3281         tcp_collapse(sk, &sk->sk_receive_queue,
3282                      sk->sk_receive_queue.next,
3283                      (struct sk_buff*)&sk->sk_receive_queue,
3284                      tp->copied_seq, tp->rcv_nxt);
3285         sk_stream_mem_reclaim(sk);
3286
3287         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3288                 return 0;
3289
3290         /* Collapsing did not help, destructive actions follow.
3291          * This must not ever occur. */
3292
3293         /* First, purge the out_of_order queue. */
3294         if (!skb_queue_empty(&tp->out_of_order_queue)) {
3295                 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3296                 __skb_queue_purge(&tp->out_of_order_queue);
3297
3298                 /* Reset SACK state.  A conforming SACK implementation will
3299                  * do the same at a timeout based retransmit.  When a connection
3300                  * is in a sad state like this, we care only about integrity
3301                  * of the connection not performance.
3302                  */
3303                 if (tp->rx_opt.sack_ok)
3304                         tcp_sack_reset(&tp->rx_opt);
3305                 sk_stream_mem_reclaim(sk);
3306         }
3307
3308         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3309                 return 0;
3310
3311         /* If we are really being abused, tell the caller to silently
3312          * drop receive data on the floor.  It will get retransmitted
3313          * and hopefully then we'll have sufficient space.
3314          */
3315         NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3316
3317         /* Massive buffer overcommit. */
3318         tp->pred_flags = 0;
3319         return -1;
3320 }
3321
3322
3323 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3324  * As additional protections, we do not touch cwnd in retransmission phases,
3325  * and if application hit its sndbuf limit recently.
3326  */
3327 void tcp_cwnd_application_limited(struct sock *sk)
3328 {
3329         struct tcp_sock *tp = tcp_sk(sk);
3330
3331         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3332             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3333                 /* Limited by application or receiver window. */
3334                 u32 win_used = max(tp->snd_cwnd_used, 2U);
3335                 if (win_used < tp->snd_cwnd) {
3336                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
3337                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3338                 }
3339                 tp->snd_cwnd_used = 0;
3340         }
3341         tp->snd_cwnd_stamp = tcp_time_stamp;
3342 }
3343
3344 static inline int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3345 {
3346         /* If the user specified a specific send buffer setting, do
3347          * not modify it.
3348          */
3349         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3350                 return 0;
3351
3352         /* If we are under global TCP memory pressure, do not expand.  */
3353         if (tcp_memory_pressure)
3354                 return 0;
3355
3356         /* If we are under soft global TCP memory pressure, do not expand.  */
3357         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3358                 return 0;
3359
3360         /* If we filled the congestion window, do not expand.  */
3361         if (tp->packets_out >= tp->snd_cwnd)
3362                 return 0;
3363
3364         return 1;
3365 }
3366
3367 /* When incoming ACK allowed to free some skb from write_queue,
3368  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3369  * on the exit from tcp input handler.
3370  *
3371  * PROBLEM: sndbuf expansion does not work well with largesend.
3372  */
3373 static void tcp_new_space(struct sock *sk)
3374 {
3375         struct tcp_sock *tp = tcp_sk(sk);
3376
3377         if (tcp_should_expand_sndbuf(sk, tp)) {
3378                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3379                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3380                     demanded = max_t(unsigned int, tp->snd_cwnd,
3381                                                    tp->reordering + 1);
3382                 sndmem *= 2*demanded;
3383                 if (sndmem > sk->sk_sndbuf)
3384                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3385                 tp->snd_cwnd_stamp = tcp_time_stamp;
3386         }
3387
3388         sk->sk_write_space(sk);
3389 }
3390
3391 static inline void tcp_check_space(struct sock *sk)
3392 {
3393         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3394                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3395                 if (sk->sk_socket &&
3396                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3397                         tcp_new_space(sk);
3398         }
3399 }
3400
3401 static __inline__ void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3402 {
3403         tcp_push_pending_frames(sk, tp);
3404         tcp_check_space(sk);
3405 }
3406
3407 /*
3408  * Check if sending an ack is needed.
3409  */
3410 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3411 {
3412         struct tcp_sock *tp = tcp_sk(sk);
3413
3414             /* More than one full frame received... */
3415         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3416              /* ... and right edge of window advances far enough.
3417               * (tcp_recvmsg() will send ACK otherwise). Or...
3418               */
3419              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3420             /* We ACK each frame or... */
3421             tcp_in_quickack_mode(sk) ||
3422             /* We have out of order data. */
3423             (ofo_possible &&
3424              skb_peek(&tp->out_of_order_queue))) {
3425                 /* Then ack it now */
3426                 tcp_send_ack(sk);
3427         } else {
3428                 /* Else, send delayed ack. */
3429                 tcp_send_delayed_ack(sk);
3430         }
3431 }
3432
3433 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3434 {
3435         if (!inet_csk_ack_scheduled(sk)) {
3436                 /* We sent a data segment already. */
3437                 return;
3438         }
3439         __tcp_ack_snd_check(sk, 1);
3440 }
3441
3442 /*
3443  *      This routine is only called when we have urgent data
3444  *      signalled. Its the 'slow' part of tcp_urg. It could be
3445  *      moved inline now as tcp_urg is only called from one
3446  *      place. We handle URGent data wrong. We have to - as
3447  *      BSD still doesn't use the correction from RFC961.
3448  *      For 1003.1g we should support a new option TCP_STDURG to permit
3449  *      either form (or just set the sysctl tcp_stdurg).
3450  */
3451
3452 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3453 {
3454         struct tcp_sock *tp = tcp_sk(sk);
3455         u32 ptr = ntohs(th->urg_ptr);
3456
3457         if (ptr && !sysctl_tcp_stdurg)
3458                 ptr--;
3459         ptr += ntohl(th->seq);
3460
3461         /* Ignore urgent data that we've already seen and read. */
3462         if (after(tp->copied_seq, ptr))
3463                 return;
3464
3465         /* Do not replay urg ptr.
3466          *
3467          * NOTE: interesting situation not covered by specs.
3468          * Misbehaving sender may send urg ptr, pointing to segment,
3469          * which we already have in ofo queue. We are not able to fetch
3470          * such data and will stay in TCP_URG_NOTYET until will be eaten
3471          * by recvmsg(). Seems, we are not obliged to handle such wicked
3472          * situations. But it is worth to think about possibility of some
3473          * DoSes using some hypothetical application level deadlock.
3474          */
3475         if (before(ptr, tp->rcv_nxt))
3476                 return;
3477
3478         /* Do we already have a newer (or duplicate) urgent pointer? */
3479         if (tp->urg_data && !after(ptr, tp->urg_seq))
3480                 return;
3481
3482         /* Tell the world about our new urgent pointer. */
3483         sk_send_sigurg(sk);
3484
3485         /* We may be adding urgent data when the last byte read was
3486          * urgent. To do this requires some care. We cannot just ignore
3487          * tp->copied_seq since we would read the last urgent byte again
3488          * as data, nor can we alter copied_seq until this data arrives
3489          * or we break the sematics of SIOCATMARK (and thus sockatmark())
3490          *
3491          * NOTE. Double Dutch. Rendering to plain English: author of comment
3492          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
3493          * and expect that both A and B disappear from stream. This is _wrong_.
3494          * Though this happens in BSD with high probability, this is occasional.
3495          * Any application relying on this is buggy. Note also, that fix "works"
3496          * only in this artificial test. Insert some normal data between A and B and we will
3497          * decline of BSD again. Verdict: it is better to remove to trap
3498          * buggy users.
3499          */
3500         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3501             !sock_flag(sk, SOCK_URGINLINE) &&
3502             tp->copied_seq != tp->rcv_nxt) {
3503                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3504                 tp->copied_seq++;
3505                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3506                         __skb_unlink(skb, &sk->sk_receive_queue);
3507                         __kfree_skb(skb);
3508                 }
3509         }
3510
3511         tp->urg_data   = TCP_URG_NOTYET;
3512         tp->urg_seq    = ptr;
3513
3514         /* Disable header prediction. */
3515         tp->pred_flags = 0;
3516 }
3517
3518 /* This is the 'fast' part of urgent handling. */
3519 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3520 {
3521         struct tcp_sock *tp = tcp_sk(sk);
3522
3523         /* Check if we get a new urgent pointer - normally not. */
3524         if (th->urg)
3525                 tcp_check_urg(sk,th);
3526
3527         /* Do we wait for any urgent data? - normally not... */
3528         if (tp->urg_data == TCP_URG_NOTYET) {
3529                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3530                           th->syn;
3531
3532                 /* Is the urgent pointer pointing into this packet? */
3533                 if (ptr < skb->len) {
3534                         u8 tmp;
3535                         if (skb_copy_bits(skb, ptr, &tmp, 1))
3536                                 BUG();
3537                         tp->urg_data = TCP_URG_VALID | tmp;
3538                         if (!sock_flag(sk, SOCK_DEAD))
3539                                 sk->sk_data_ready(sk, 0);
3540                 }
3541         }
3542 }
3543
3544 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3545 {
3546         struct tcp_sock *tp = tcp_sk(sk);
3547         int chunk = skb->len - hlen;
3548         int err;
3549
3550         local_bh_enable();
3551         if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3552                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3553         else
3554                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3555                                                        tp->ucopy.iov);
3556
3557         if (!err) {
3558                 tp->ucopy.len -= chunk;
3559                 tp->copied_seq += chunk;
3560                 tcp_rcv_space_adjust(sk);
3561         }
3562
3563         local_bh_disable();
3564         return err;
3565 }
3566
3567 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3568 {
3569         int result;
3570
3571         if (sock_owned_by_user(sk)) {
3572                 local_bh_enable();
3573                 result = __tcp_checksum_complete(skb);
3574                 local_bh_disable();
3575         } else {
3576                 result = __tcp_checksum_complete(skb);
3577         }
3578         return result;
3579 }
3580
3581 static __inline__ int
3582 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3583 {
3584         return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3585                 __tcp_checksum_complete_user(sk, skb);
3586 }
3587
3588 /*
3589  *      TCP receive function for the ESTABLISHED state.
3590  *
3591  *      It is split into a fast path and a slow path. The fast path is
3592  *      disabled when:
3593  *      - A zero window was announced from us - zero window probing
3594  *        is only handled properly in the slow path.
3595  *      - Out of order segments arrived.
3596  *      - Urgent data is expected.
3597  *      - There is no buffer space left
3598  *      - Unexpected TCP flags/window values/header lengths are received
3599  *        (detected by checking the TCP header against pred_flags)
3600  *      - Data is sent in both directions. Fast path only supports pure senders
3601  *        or pure receivers (this means either the sequence number or the ack
3602  *        value must stay constant)
3603  *      - Unexpected TCP option.
3604  *
3605  *      When these conditions are not satisfied it drops into a standard
3606  *      receive procedure patterned after RFC793 to handle all cases.
3607  *      The first three cases are guaranteed by proper pred_flags setting,
3608  *      the rest is checked inline. Fast processing is turned on in
3609  *      tcp_data_queue when everything is OK.
3610  */
3611 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3612                         struct tcphdr *th, unsigned len)
3613 {
3614         struct tcp_sock *tp = tcp_sk(sk);
3615
3616         /*
3617          *      Header prediction.
3618          *      The code loosely follows the one in the famous
3619          *      "30 instruction TCP receive" Van Jacobson mail.
3620          *
3621          *      Van's trick is to deposit buffers into socket queue
3622          *      on a device interrupt, to call tcp_recv function
3623          *      on the receive process context and checksum and copy
3624          *      the buffer to user space. smart...
3625          *
3626          *      Our current scheme is not silly either but we take the
3627          *      extra cost of the net_bh soft interrupt processing...
3628          *      We do checksum and copy also but from device to kernel.
3629          */
3630
3631         tp->rx_opt.saw_tstamp = 0;
3632
3633         /*      pred_flags is 0xS?10 << 16 + snd_wnd
3634          *      if header_predition is to be made
3635          *      'S' will always be tp->tcp_header_len >> 2
3636          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3637          *  turn it off (when there are holes in the receive
3638          *       space for instance)
3639          *      PSH flag is ignored.
3640          */
3641
3642         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3643                 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3644                 int tcp_header_len = tp->tcp_header_len;
3645
3646                 /* Timestamp header prediction: tcp_header_len
3647                  * is automatically equal to th->doff*4 due to pred_flags
3648                  * match.
3649                  */
3650
3651                 /* Check timestamp */
3652                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3653                         __u32 *ptr = (__u32 *)(th + 1);
3654
3655                         /* No? Slow path! */
3656                         if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3657                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
3658                                 goto slow_path;
3659
3660                         tp->rx_opt.saw_tstamp = 1;
3661                         ++ptr;
3662                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3663                         ++ptr;
3664                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3665
3666                         /* If PAWS failed, check it more carefully in slow path */
3667                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
3668                                 goto slow_path;
3669
3670                         /* DO NOT update ts_recent here, if checksum fails
3671                          * and timestamp was corrupted part, it will result
3672                          * in a hung connection since we will drop all
3673                          * future packets due to the PAWS test.
3674                          */
3675                 }
3676
3677                 if (len <= tcp_header_len) {
3678                         /* Bulk data transfer: sender */
3679                         if (len == tcp_header_len) {
3680                                 /* Predicted packet is in window by definition.
3681                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3682                                  * Hence, check seq<=rcv_wup reduces to:
3683                                  */
3684                                 if (tcp_header_len ==
3685                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3686                                     tp->rcv_nxt == tp->rcv_wup)
3687                                         tcp_store_ts_recent(tp);
3688
3689                                 tcp_rcv_rtt_measure_ts(sk, skb);
3690
3691                                 /* We know that such packets are checksummed
3692                                  * on entry.
3693                                  */
3694                                 tcp_ack(sk, skb, 0);
3695                                 __kfree_skb(skb);
3696                                 tcp_data_snd_check(sk, tp);
3697                                 return 0;
3698                         } else { /* Header too small */
3699                                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3700                                 goto discard;
3701                         }
3702                 } else {
3703                         int eaten = 0;
3704
3705                         if (tp->ucopy.task == current &&
3706                             tp->copied_seq == tp->rcv_nxt &&
3707                             len - tcp_header_len <= tp->ucopy.len &&
3708                             sock_owned_by_user(sk)) {
3709                                 __set_current_state(TASK_RUNNING);
3710
3711                                 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
3712                                         /* Predicted packet is in window by definition.
3713                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3714                                          * Hence, check seq<=rcv_wup reduces to:
3715                                          */
3716                                         if (tcp_header_len ==
3717                                             (sizeof(struct tcphdr) +
3718                                              TCPOLEN_TSTAMP_ALIGNED) &&
3719                                             tp->rcv_nxt == tp->rcv_wup)
3720                                                 tcp_store_ts_recent(tp);
3721
3722                                         tcp_rcv_rtt_measure_ts(sk, skb);
3723
3724                                         __skb_pull(skb, tcp_header_len);
3725                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3726                                         NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
3727                                         eaten = 1;
3728                                 }
3729                         }
3730                         if (!eaten) {
3731                                 if (tcp_checksum_complete_user(sk, skb))
3732                                         goto csum_error;
3733
3734                                 /* Predicted packet is in window by definition.
3735                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3736                                  * Hence, check seq<=rcv_wup reduces to:
3737                                  */
3738                                 if (tcp_header_len ==
3739                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
3740                                     tp->rcv_nxt == tp->rcv_wup)
3741                                         tcp_store_ts_recent(tp);
3742
3743                                 tcp_rcv_rtt_measure_ts(sk, skb);
3744
3745                                 if ((int)skb->truesize > sk->sk_forward_alloc)
3746                                         goto step5;
3747
3748                                 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
3749
3750                                 /* Bulk data transfer: receiver */
3751                                 __skb_pull(skb,tcp_header_len);
3752                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3753                                 sk_stream_set_owner_r(skb, sk);
3754                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3755                         }
3756
3757                         tcp_event_data_recv(sk, tp, skb);
3758
3759                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
3760                                 /* Well, only one small jumplet in fast path... */
3761                                 tcp_ack(sk, skb, FLAG_DATA);
3762                                 tcp_data_snd_check(sk, tp);
3763                                 if (!inet_csk_ack_scheduled(sk))
3764                                         goto no_ack;
3765                         }
3766
3767                         __tcp_ack_snd_check(sk, 0);
3768 no_ack:
3769                         if (eaten)
3770                                 __kfree_skb(skb);
3771                         else
3772                                 sk->sk_data_ready(sk, 0);
3773                         return 0;
3774                 }
3775         }
3776
3777 slow_path:
3778         if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
3779                 goto csum_error;
3780
3781         /*
3782          * RFC1323: H1. Apply PAWS check first.
3783          */
3784         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
3785             tcp_paws_discard(sk, skb)) {
3786                 if (!th->rst) {
3787                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
3788                         tcp_send_dupack(sk, skb);
3789                         goto discard;
3790                 }
3791                 /* Resets are accepted even if PAWS failed.
3792
3793                    ts_recent update must be made after we are sure
3794                    that the packet is in window.
3795                  */
3796         }
3797
3798         /*
3799          *      Standard slow path.
3800          */
3801
3802         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
3803                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3804                  * (RST) segments are validated by checking their SEQ-fields."
3805                  * And page 69: "If an incoming segment is not acceptable,
3806                  * an acknowledgment should be sent in reply (unless the RST bit
3807                  * is set, if so drop the segment and return)".
3808                  */
3809                 if (!th->rst)
3810                         tcp_send_dupack(sk, skb);
3811                 goto discard;
3812         }
3813
3814         if(th->rst) {
3815                 tcp_reset(sk);
3816                 goto discard;
3817         }
3818
3819         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3820
3821         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3822                 TCP_INC_STATS_BH(TCP_MIB_INERRS);
3823                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
3824                 tcp_reset(sk);
3825                 return 1;
3826         }
3827
3828 step5:
3829         if(th->ack)
3830                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3831
3832         tcp_rcv_rtt_measure_ts(sk, skb);
3833
3834         /* Process urgent data. */
3835         tcp_urg(sk, skb, th);
3836
3837         /* step 7: process the segment text */
3838         tcp_data_queue(sk, skb);
3839
3840         tcp_data_snd_check(sk, tp);
3841         tcp_ack_snd_check(sk);
3842         return 0;
3843
3844 csum_error:
3845         TCP_INC_STATS_BH(TCP_MIB_INERRS);
3846
3847 discard:
3848         __kfree_skb(skb);
3849         return 0;
3850 }
3851
3852 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
3853                                          struct tcphdr *th, unsigned len)
3854 {
3855         struct tcp_sock *tp = tcp_sk(sk);
3856         int saved_clamp = tp->rx_opt.mss_clamp;
3857
3858         tcp_parse_options(skb, &tp->rx_opt, 0);
3859
3860         if (th->ack) {
3861                 struct inet_connection_sock *icsk;
3862                 /* rfc793:
3863                  * "If the state is SYN-SENT then
3864                  *    first check the ACK bit
3865                  *      If the ACK bit is set
3866                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3867                  *        a reset (unless the RST bit is set, if so drop
3868                  *        the segment and return)"
3869                  *
3870                  *  We do not send data with SYN, so that RFC-correct
3871                  *  test reduces to:
3872                  */
3873                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
3874                         goto reset_and_undo;
3875
3876                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3877                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
3878                              tcp_time_stamp)) {
3879                         NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
3880                         goto reset_and_undo;
3881                 }
3882
3883                 /* Now ACK is acceptable.
3884                  *
3885                  * "If the RST bit is set
3886                  *    If the ACK was acceptable then signal the user "error:
3887                  *    connection reset", drop the segment, enter CLOSED state,
3888                  *    delete TCB, and return."
3889                  */
3890
3891                 if (th->rst) {
3892                         tcp_reset(sk);
3893                         goto discard;
3894                 }
3895
3896                 /* rfc793:
3897                  *   "fifth, if neither of the SYN or RST bits is set then
3898                  *    drop the segment and return."
3899                  *
3900                  *    See note below!
3901                  *                                        --ANK(990513)
3902                  */
3903                 if (!th->syn)
3904                         goto discard_and_undo;
3905
3906                 /* rfc793:
3907                  *   "If the SYN bit is on ...
3908                  *    are acceptable then ...
3909                  *    (our SYN has been ACKed), change the connection
3910                  *    state to ESTABLISHED..."
3911                  */
3912
3913                 TCP_ECN_rcv_synack(tp, th);
3914                 if (tp->ecn_flags&TCP_ECN_OK)
3915                         sock_set_flag(sk, SOCK_NO_LARGESEND);
3916
3917                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
3918                 tcp_ack(sk, skb, FLAG_SLOWPATH);
3919
3920                 /* Ok.. it's good. Set up sequence numbers and
3921                  * move to established.
3922                  */
3923                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
3924                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
3925
3926                 /* RFC1323: The window in SYN & SYN/ACK segments is
3927                  * never scaled.
3928                  */
3929                 tp->snd_wnd = ntohs(th->window);
3930                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
3931
3932                 if (!tp->rx_opt.wscale_ok) {
3933                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
3934                         tp->window_clamp = min(tp->window_clamp, 65535U);
3935                 }
3936
3937                 if (tp->rx_opt.saw_tstamp) {
3938                         tp->rx_opt.tstamp_ok       = 1;
3939                         tp->tcp_header_len =
3940                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
3941                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
3942                         tcp_store_ts_recent(tp);
3943                 } else {
3944                         tp->tcp_header_len = sizeof(struct tcphdr);
3945                 }
3946
3947                 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
3948                         tp->rx_opt.sack_ok |= 2;
3949
3950                 tcp_sync_mss(sk, tp->pmtu_cookie);
3951                 tcp_initialize_rcv_mss(sk);
3952
3953                 /* Remember, tcp_poll() does not lock socket!
3954                  * Change state from SYN-SENT only after copied_seq
3955                  * is initialized. */
3956                 tp->copied_seq = tp->rcv_nxt;
3957                 mb();
3958                 tcp_set_state(sk, TCP_ESTABLISHED);
3959
3960                 /* Make sure socket is routed, for correct metrics.  */
3961                 tp->af_specific->rebuild_header(sk);
3962
3963                 tcp_init_metrics(sk);
3964
3965                 tcp_init_congestion_control(sk);
3966
3967                 /* Prevent spurious tcp_cwnd_restart() on first data
3968                  * packet.
3969                  */
3970                 tp->lsndtime = tcp_time_stamp;
3971
3972                 tcp_init_buffer_space(sk);
3973
3974                 if (sock_flag(sk, SOCK_KEEPOPEN))
3975                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
3976
3977                 if (!tp->rx_opt.snd_wscale)
3978                         __tcp_fast_path_on(tp, tp->snd_wnd);
3979                 else
3980                         tp->pred_flags = 0;
3981
3982                 if (!sock_flag(sk, SOCK_DEAD)) {
3983                         sk->sk_state_change(sk);
3984                         sk_wake_async(sk, 0, POLL_OUT);
3985                 }
3986
3987                 icsk = inet_csk(sk);
3988
3989                 if (sk->sk_write_pending ||
3990                     icsk->icsk_accept_queue.rskq_defer_accept ||
3991                     icsk->icsk_ack.pingpong) {
3992                         /* Save one ACK. Data will be ready after
3993                          * several ticks, if write_pending is set.
3994                          *
3995                          * It may be deleted, but with this feature tcpdumps
3996                          * look so _wonderfully_ clever, that I was not able
3997                          * to stand against the temptation 8)     --ANK
3998                          */
3999                         inet_csk_schedule_ack(sk);
4000                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4001                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
4002                         tcp_incr_quickack(sk);
4003                         tcp_enter_quickack_mode(sk);
4004                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4005                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
4006
4007 discard:
4008                         __kfree_skb(skb);
4009                         return 0;
4010                 } else {
4011                         tcp_send_ack(sk);
4012                 }
4013                 return -1;
4014         }
4015
4016         /* No ACK in the segment */
4017
4018         if (th->rst) {
4019                 /* rfc793:
4020                  * "If the RST bit is set
4021                  *
4022                  *      Otherwise (no ACK) drop the segment and return."
4023                  */
4024
4025                 goto discard_and_undo;
4026         }
4027
4028         /* PAWS check. */
4029         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4030                 goto discard_and_undo;
4031
4032         if (th->syn) {
4033                 /* We see SYN without ACK. It is attempt of
4034                  * simultaneous connect with crossed SYNs.
4035                  * Particularly, it can be connect to self.
4036                  */
4037                 tcp_set_state(sk, TCP_SYN_RECV);
4038
4039                 if (tp->rx_opt.saw_tstamp) {
4040                         tp->rx_opt.tstamp_ok = 1;
4041                         tcp_store_ts_recent(tp);
4042                         tp->tcp_header_len =
4043                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4044                 } else {
4045                         tp->tcp_header_len = sizeof(struct tcphdr);
4046                 }
4047
4048                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4049                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4050
4051                 /* RFC1323: The window in SYN & SYN/ACK segments is
4052                  * never scaled.
4053                  */
4054                 tp->snd_wnd    = ntohs(th->window);
4055                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
4056                 tp->max_window = tp->snd_wnd;
4057
4058                 TCP_ECN_rcv_syn(tp, th);
4059                 if (tp->ecn_flags&TCP_ECN_OK)
4060                         sock_set_flag(sk, SOCK_NO_LARGESEND);
4061
4062                 tcp_sync_mss(sk, tp->pmtu_cookie);
4063                 tcp_initialize_rcv_mss(sk);
4064
4065
4066                 tcp_send_synack(sk);
4067 #if 0
4068                 /* Note, we could accept data and URG from this segment.
4069                  * There are no obstacles to make this.
4070                  *
4071                  * However, if we ignore data in ACKless segments sometimes,
4072                  * we have no reasons to accept it sometimes.
4073                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
4074                  * is not flawless. So, discard packet for sanity.
4075                  * Uncomment this return to process the data.
4076                  */
4077                 return -1;
4078 #else
4079                 goto discard;
4080 #endif
4081         }
4082         /* "fifth, if neither of the SYN or RST bits is set then
4083          * drop the segment and return."
4084          */
4085
4086 discard_and_undo:
4087         tcp_clear_options(&tp->rx_opt);
4088         tp->rx_opt.mss_clamp = saved_clamp;
4089         goto discard;
4090
4091 reset_and_undo:
4092         tcp_clear_options(&tp->rx_opt);
4093         tp->rx_opt.mss_clamp = saved_clamp;
4094         return 1;
4095 }
4096
4097
4098 /*
4099  *      This function implements the receiving procedure of RFC 793 for
4100  *      all states except ESTABLISHED and TIME_WAIT.
4101  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4102  *      address independent.
4103  */
4104
4105 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4106                           struct tcphdr *th, unsigned len)
4107 {
4108         struct tcp_sock *tp = tcp_sk(sk);
4109         int queued = 0;
4110
4111         tp->rx_opt.saw_tstamp = 0;
4112
4113         switch (sk->sk_state) {
4114         case TCP_CLOSE:
4115                 goto discard;
4116
4117         case TCP_LISTEN:
4118                 if(th->ack)
4119                         return 1;
4120
4121                 if(th->rst)
4122                         goto discard;
4123
4124                 if(th->syn) {
4125                         if(tp->af_specific->conn_request(sk, skb) < 0)
4126                                 return 1;
4127
4128                         /* Now we have several options: In theory there is
4129                          * nothing else in the frame. KA9Q has an option to
4130                          * send data with the syn, BSD accepts data with the
4131                          * syn up to the [to be] advertised window and
4132                          * Solaris 2.1 gives you a protocol error. For now
4133                          * we just ignore it, that fits the spec precisely
4134                          * and avoids incompatibilities. It would be nice in
4135                          * future to drop through and process the data.
4136                          *
4137                          * Now that TTCP is starting to be used we ought to
4138                          * queue this data.
4139                          * But, this leaves one open to an easy denial of
4140                          * service attack, and SYN cookies can't defend
4141                          * against this problem. So, we drop the data
4142                          * in the interest of security over speed.
4143                          */
4144                         goto discard;
4145                 }
4146                 goto discard;
4147
4148         case TCP_SYN_SENT:
4149                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4150                 if (queued >= 0)
4151                         return queued;
4152
4153                 /* Do step6 onward by hand. */
4154                 tcp_urg(sk, skb, th);
4155                 __kfree_skb(skb);
4156                 tcp_data_snd_check(sk, tp);
4157                 return 0;
4158         }
4159
4160         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4161             tcp_paws_discard(sk, skb)) {
4162                 if (!th->rst) {
4163                         NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4164                         tcp_send_dupack(sk, skb);
4165                         goto discard;
4166                 }
4167                 /* Reset is accepted even if it did not pass PAWS. */
4168         }
4169
4170         /* step 1: check sequence number */
4171         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4172                 if (!th->rst)
4173                         tcp_send_dupack(sk, skb);
4174                 goto discard;
4175         }
4176
4177         /* step 2: check RST bit */
4178         if(th->rst) {
4179                 tcp_reset(sk);
4180                 goto discard;
4181         }
4182
4183         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4184
4185         /* step 3: check security and precedence [ignored] */
4186
4187         /*      step 4:
4188          *
4189          *      Check for a SYN in window.
4190          */
4191         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4192                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4193                 tcp_reset(sk);
4194                 return 1;
4195         }
4196
4197         /* step 5: check the ACK field */
4198         if (th->ack) {
4199                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4200
4201                 switch(sk->sk_state) {
4202                 case TCP_SYN_RECV:
4203                         if (acceptable) {
4204                                 tp->copied_seq = tp->rcv_nxt;
4205                                 mb();
4206                                 tcp_set_state(sk, TCP_ESTABLISHED);
4207                                 sk->sk_state_change(sk);
4208
4209                                 /* Note, that this wakeup is only for marginal
4210                                  * crossed SYN case. Passively open sockets
4211                                  * are not waked up, because sk->sk_sleep ==
4212                                  * NULL and sk->sk_socket == NULL.
4213                                  */
4214                                 if (sk->sk_socket) {
4215                                         sk_wake_async(sk,0,POLL_OUT);
4216                                 }
4217
4218                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4219                                 tp->snd_wnd = ntohs(th->window) <<
4220                                               tp->rx_opt.snd_wscale;
4221                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4222                                             TCP_SKB_CB(skb)->seq);
4223
4224                                 /* tcp_ack considers this ACK as duplicate
4225                                  * and does not calculate rtt.
4226                                  * Fix it at least with timestamps.
4227                                  */
4228                                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4229                                     !tp->srtt)
4230                                         tcp_ack_saw_tstamp(sk, 0);
4231
4232                                 if (tp->rx_opt.tstamp_ok)
4233                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4234
4235                                 /* Make sure socket is routed, for
4236                                  * correct metrics.
4237                                  */
4238                                 tp->af_specific->rebuild_header(sk);
4239
4240                                 tcp_init_metrics(sk);
4241
4242                                 tcp_init_congestion_control(sk);
4243
4244                                 /* Prevent spurious tcp_cwnd_restart() on
4245                                  * first data packet.
4246                                  */
4247                                 tp->lsndtime = tcp_time_stamp;
4248
4249                                 tcp_initialize_rcv_mss(sk);
4250                                 tcp_init_buffer_space(sk);
4251                                 tcp_fast_path_on(tp);
4252                         } else {
4253                                 return 1;
4254                         }
4255                         break;
4256
4257                 case TCP_FIN_WAIT1:
4258                         if (tp->snd_una == tp->write_seq) {
4259                                 tcp_set_state(sk, TCP_FIN_WAIT2);
4260                                 sk->sk_shutdown |= SEND_SHUTDOWN;
4261                                 dst_confirm(sk->sk_dst_cache);
4262
4263                                 if (!sock_flag(sk, SOCK_DEAD))
4264                                         /* Wake up lingering close() */
4265                                         sk->sk_state_change(sk);
4266                                 else {
4267                                         int tmo;
4268
4269                                         if (tp->linger2 < 0 ||
4270                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4271                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4272                                                 tcp_done(sk);
4273                                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4274                                                 return 1;
4275                                         }
4276
4277                                         tmo = tcp_fin_time(sk);
4278                                         if (tmo > TCP_TIMEWAIT_LEN) {
4279                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4280                                         } else if (th->fin || sock_owned_by_user(sk)) {
4281                                                 /* Bad case. We could lose such FIN otherwise.
4282                                                  * It is not a big problem, but it looks confusing
4283                                                  * and not so rare event. We still can lose it now,
4284                                                  * if it spins in bh_lock_sock(), but it is really
4285                                                  * marginal case.
4286                                                  */
4287                                                 inet_csk_reset_keepalive_timer(sk, tmo);
4288                                         } else {
4289                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4290                                                 goto discard;
4291                                         }
4292                                 }
4293                         }
4294                         break;
4295
4296                 case TCP_CLOSING:
4297                         if (tp->snd_una == tp->write_seq) {
4298                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4299                                 goto discard;
4300                         }
4301                         break;
4302
4303                 case TCP_LAST_ACK:
4304                         if (tp->snd_una == tp->write_seq) {
4305                                 tcp_update_metrics(sk);
4306                                 tcp_done(sk);
4307                                 goto discard;
4308                         }
4309                         break;
4310                 }
4311         } else
4312                 goto discard;
4313
4314         /* step 6: check the URG bit */
4315         tcp_urg(sk, skb, th);
4316
4317         /* step 7: process the segment text */
4318         switch (sk->sk_state) {
4319         case TCP_CLOSE_WAIT:
4320         case TCP_CLOSING:
4321         case TCP_LAST_ACK:
4322                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4323                         break;
4324         case TCP_FIN_WAIT1:
4325         case TCP_FIN_WAIT2:
4326                 /* RFC 793 says to queue data in these states,
4327                  * RFC 1122 says we MUST send a reset.
4328                  * BSD 4.4 also does reset.
4329                  */
4330                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4331                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4332                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4333                                 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4334                                 tcp_reset(sk);
4335                                 return 1;
4336                         }
4337                 }
4338                 /* Fall through */
4339         case TCP_ESTABLISHED:
4340                 tcp_data_queue(sk, skb);
4341                 queued = 1;
4342                 break;
4343         }
4344
4345         /* tcp_data could move socket to TIME-WAIT */
4346         if (sk->sk_state != TCP_CLOSE) {
4347                 tcp_data_snd_check(sk, tp);
4348                 tcp_ack_snd_check(sk);
4349         }
4350
4351         if (!queued) {
4352 discard:
4353                 __kfree_skb(skb);
4354         }
4355         return 0;
4356 }
4357
4358 EXPORT_SYMBOL(sysctl_tcp_ecn);
4359 EXPORT_SYMBOL(sysctl_tcp_reordering);
4360 EXPORT_SYMBOL(sysctl_tcp_abc);
4361 EXPORT_SYMBOL(tcp_parse_options);
4362 EXPORT_SYMBOL(tcp_rcv_established);
4363 EXPORT_SYMBOL(tcp_rcv_state_process);