linux/net/ipv4/tcp_input.c
<<
>>
Prefs
   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 * Authors:     Ross Biro
   9 *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10 *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11 *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12 *              Florian La Roche, <flla@stud.uni-sb.de>
  13 *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14 *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15 *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16 *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17 *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18 *              Jorge Cwik, <jorge@laser.satlink.net>
  19 */
  20
  21/*
  22 * Changes:
  23 *              Pedro Roque     :       Fast Retransmit/Recovery.
  24 *                                      Two receive queues.
  25 *                                      Retransmit queue handled by TCP.
  26 *                                      Better retransmit timer handling.
  27 *                                      New congestion avoidance.
  28 *                                      Header prediction.
  29 *                                      Variable renaming.
  30 *
  31 *              Eric            :       Fast Retransmit.
  32 *              Randy Scott     :       MSS option defines.
  33 *              Eric Schenk     :       Fixes to slow start algorithm.
  34 *              Eric Schenk     :       Yet another double ACK bug.
  35 *              Eric Schenk     :       Delayed ACK bug fixes.
  36 *              Eric Schenk     :       Floyd style fast retrans war avoidance.
  37 *              David S. Miller :       Don't allow zero congestion window.
  38 *              Eric Schenk     :       Fix retransmitter so that it sends
  39 *                                      next packet on ack of previous packet.
  40 *              Andi Kleen      :       Moved open_request checking here
  41 *                                      and process RSTs for open_requests.
  42 *              Andi Kleen      :       Better prune_queue, and other fixes.
  43 *              Andrey Savochkin:       Fix RTT measurements in the presence of
  44 *                                      timestamps.
  45 *              Andrey Savochkin:       Check sequence numbers correctly when
  46 *                                      removing SACKs due to in sequence incoming
  47 *                                      data segments.
  48 *              Andi Kleen:             Make sure we never ack data there is not
  49 *                                      enough room for. Also make this condition
  50 *                                      a fatal error if it might still happen.
  51 *              Andi Kleen:             Add tcp_measure_rcv_mss to make
  52 *                                      connections with MSS<min(MTU,ann. MSS)
  53 *                                      work without delayed acks.
  54 *              Andi Kleen:             Process packets with PSH set in the
  55 *                                      fast path.
  56 *              J Hadi Salim:           ECN support
  57 *              Andrei Gurtov,
  58 *              Pasi Sarolahti,
  59 *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
  60 *                                      engine. Lots of bugs are found.
  61 *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
  62 */
  63
  64#include <linux/mm.h>
  65#include <linux/slab.h>
  66#include <linux/module.h>
  67#include <linux/sysctl.h>
  68#include <linux/kernel.h>
  69#include <net/dst.h>
  70#include <net/tcp.h>
  71#include <net/inet_common.h>
  72#include <linux/ipsec.h>
  73#include <asm/unaligned.h>
  74#include <net/netdma.h>
  75
  76int sysctl_tcp_timestamps __read_mostly = 1;
  77int sysctl_tcp_window_scaling __read_mostly = 1;
  78int sysctl_tcp_sack __read_mostly = 1;
  79int sysctl_tcp_fack __read_mostly = 1;
  80int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  81int sysctl_tcp_ecn __read_mostly = 2;
  82int sysctl_tcp_dsack __read_mostly = 1;
  83int sysctl_tcp_app_win __read_mostly = 31;
  84int sysctl_tcp_adv_win_scale __read_mostly = 2;
  85
  86int sysctl_tcp_stdurg __read_mostly;
  87int sysctl_tcp_rfc1337 __read_mostly;
  88int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  89int sysctl_tcp_frto __read_mostly = 2;
  90int sysctl_tcp_frto_response __read_mostly;
  91int sysctl_tcp_nometrics_save __read_mostly;
  92
  93int sysctl_tcp_thin_dupack __read_mostly;
  94
  95int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
  96int sysctl_tcp_abc __read_mostly;
  97
  98#define FLAG_DATA               0x01 /* Incoming frame contained data.          */
  99#define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
 100#define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
 101#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
 102#define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
 103#define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
 104#define FLAG_ECE                0x40 /* ECE in this ACK                         */
 105#define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
 106#define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
 107#define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
 108#define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 109#define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
 110#define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
 111#define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
 112
 113#define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 114#define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
 115#define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
 116#define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
 117#define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
 118
 119#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
 120#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
 121
 122/* Adapt the MSS value used to make delayed ack decision to the
 123 * real world.
 124 */
 125static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
 126{
 127        struct inet_connection_sock *icsk = inet_csk(sk);
 128        const unsigned int lss = icsk->icsk_ack.last_seg_size;
 129        unsigned int len;
 130
 131        icsk->icsk_ack.last_seg_size = 0;
 132
 133        /* skb->len may jitter because of SACKs, even if peer
 134         * sends good full-sized frames.
 135         */
 136        len = skb_shinfo(skb)->gso_size ? : skb->len;
 137        if (len >= icsk->icsk_ack.rcv_mss) {
 138                icsk->icsk_ack.rcv_mss = len;
 139        } else {
 140                /* Otherwise, we make more careful check taking into account,
 141                 * that SACKs block is variable.
 142                 *
 143                 * "len" is invariant segment length, including TCP header.
 144                 */
 145                len += skb->data - skb_transport_header(skb);
 146                if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
 147                    /* If PSH is not set, packet should be
 148                     * full sized, provided peer TCP is not badly broken.
 149                     * This observation (if it is correct 8)) allows
 150                     * to handle super-low mtu links fairly.
 151                     */
 152                    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
 153                     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
 154                        /* Subtract also invariant (if peer is RFC compliant),
 155                         * tcp header plus fixed timestamp option length.
 156                         * Resulting "len" is MSS free of SACK jitter.
 157                         */
 158                        len -= tcp_sk(sk)->tcp_header_len;
 159                        icsk->icsk_ack.last_seg_size = len;
 160                        if (len == lss) {
 161                                icsk->icsk_ack.rcv_mss = len;
 162                                return;
 163                        }
 164                }
 165                if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
 166                        icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
 167                icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
 168        }
 169}
 170
 171static void tcp_incr_quickack(struct sock *sk)
 172{
 173        struct inet_connection_sock *icsk = inet_csk(sk);
 174        unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
 175
 176        if (quickacks == 0)
 177                quickacks = 2;
 178        if (quickacks > icsk->icsk_ack.quick)
 179                icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
 180}
 181
 182void tcp_enter_quickack_mode(struct sock *sk)
 183{
 184        struct inet_connection_sock *icsk = inet_csk(sk);
 185        tcp_incr_quickack(sk);
 186        icsk->icsk_ack.pingpong = 0;
 187        icsk->icsk_ack.ato = TCP_ATO_MIN;
 188}
 189
 190/* Send ACKs quickly, if "quick" count is not exhausted
 191 * and the session is not interactive.
 192 */
 193
 194static inline int tcp_in_quickack_mode(const struct sock *sk)
 195{
 196        const struct inet_connection_sock *icsk = inet_csk(sk);
 197        return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
 198}
 199
 200static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
 201{
 202        if (tp->ecn_flags & TCP_ECN_OK)
 203                tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
 204}
 205
 206static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
 207{
 208        if (tcp_hdr(skb)->cwr)
 209                tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 210}
 211
 212static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
 213{
 214        tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 215}
 216
 217static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
 218{
 219        if (tp->ecn_flags & TCP_ECN_OK) {
 220                if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
 221                        tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
 222                /* Funny extension: if ECT is not set on a segment,
 223                 * it is surely retransmit. It is not in ECN RFC,
 224                 * but Linux follows this rule. */
 225                else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
 226                        tcp_enter_quickack_mode((struct sock *)tp);
 227        }
 228}
 229
 230static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
 231{
 232        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
 233                tp->ecn_flags &= ~TCP_ECN_OK;
 234}
 235
 236static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
 237{
 238        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
 239                tp->ecn_flags &= ~TCP_ECN_OK;
 240}
 241
 242static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
 243{
 244        if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
 245                return 1;
 246        return 0;
 247}
 248
 249/* Buffer size and advertised window tuning.
 250 *
 251 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
 252 */
 253
 254static void tcp_fixup_sndbuf(struct sock *sk)
 255{
 256        int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
 257                     sizeof(struct sk_buff);
 258
 259        if (sk->sk_sndbuf < 3 * sndmem)
 260                sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
 261}
 262
 263/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
 264 *
 265 * All tcp_full_space() is split to two parts: "network" buffer, allocated
 266 * forward and advertised in receiver window (tp->rcv_wnd) and
 267 * "application buffer", required to isolate scheduling/application
 268 * latencies from network.
 269 * window_clamp is maximal advertised window. It can be less than
 270 * tcp_full_space(), in this case tcp_full_space() - window_clamp
 271 * is reserved for "application" buffer. The less window_clamp is
 272 * the smoother our behaviour from viewpoint of network, but the lower
 273 * throughput and the higher sensitivity of the connection to losses. 8)
 274 *
 275 * rcv_ssthresh is more strict window_clamp used at "slow start"
 276 * phase to predict further behaviour of this connection.
 277 * It is used for two goals:
 278 * - to enforce header prediction at sender, even when application
 279 *   requires some significant "application buffer". It is check #1.
 280 * - to prevent pruning of receive queue because of misprediction
 281 *   of receiver window. Check #2.
 282 *
 283 * The scheme does not work when sender sends good segments opening
 284 * window and then starts to feed us spaghetti. But it should work
 285 * in common situations. Otherwise, we have to rely on queue collapsing.
 286 */
 287
 288/* Slow part of check#2. */
 289static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
 290{
 291        struct tcp_sock *tp = tcp_sk(sk);
 292        /* Optimize this! */
 293        int truesize = tcp_win_from_space(skb->truesize) >> 1;
 294        int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
 295
 296        while (tp->rcv_ssthresh <= window) {
 297                if (truesize <= skb->len)
 298                        return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
 299
 300                truesize >>= 1;
 301                window >>= 1;
 302        }
 303        return 0;
 304}
 305
 306static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
 307{
 308        struct tcp_sock *tp = tcp_sk(sk);
 309
 310        /* Check #1 */
 311        if (tp->rcv_ssthresh < tp->window_clamp &&
 312            (int)tp->rcv_ssthresh < tcp_space(sk) &&
 313            !tcp_memory_pressure) {
 314                int incr;
 315
 316                /* Check #2. Increase window, if skb with such overhead
 317                 * will fit to rcvbuf in future.
 318                 */
 319                if (tcp_win_from_space(skb->truesize) <= skb->len)
 320                        incr = 2 * tp->advmss;
 321                else
 322                        incr = __tcp_grow_window(sk, skb);
 323
 324                if (incr) {
 325                        tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
 326                                               tp->window_clamp);
 327                        inet_csk(sk)->icsk_ack.quick |= 1;
 328                }
 329        }
 330}
 331
 332/* 3. Tuning rcvbuf, when connection enters established state. */
 333
 334static void tcp_fixup_rcvbuf(struct sock *sk)
 335{
 336        struct tcp_sock *tp = tcp_sk(sk);
 337        int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
 338
 339        /* Try to select rcvbuf so that 4 mss-sized segments
 340         * will fit to window and corresponding skbs will fit to our rcvbuf.
 341         * (was 3; 4 is minimum to allow fast retransmit to work.)
 342         */
 343        while (tcp_win_from_space(rcvmem) < tp->advmss)
 344                rcvmem += 128;
 345        if (sk->sk_rcvbuf < 4 * rcvmem)
 346                sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
 347}
 348
 349/* 4. Try to fixup all. It is made immediately after connection enters
 350 *    established state.
 351 */
 352static void tcp_init_buffer_space(struct sock *sk)
 353{
 354        struct tcp_sock *tp = tcp_sk(sk);
 355        int maxwin;
 356
 357        if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 358                tcp_fixup_rcvbuf(sk);
 359        if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
 360                tcp_fixup_sndbuf(sk);
 361
 362        tp->rcvq_space.space = tp->rcv_wnd;
 363
 364        maxwin = tcp_full_space(sk);
 365
 366        if (tp->window_clamp >= maxwin) {
 367                tp->window_clamp = maxwin;
 368
 369                if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
 370                        tp->window_clamp = max(maxwin -
 371                                               (maxwin >> sysctl_tcp_app_win),
 372                                               4 * tp->advmss);
 373        }
 374
 375        /* Force reservation of one segment. */
 376        if (sysctl_tcp_app_win &&
 377            tp->window_clamp > 2 * tp->advmss &&
 378            tp->window_clamp + tp->advmss > maxwin)
 379                tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
 380
 381        tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
 382        tp->snd_cwnd_stamp = tcp_time_stamp;
 383}
 384
 385/* 5. Recalculate window clamp after socket hit its memory bounds. */
 386static void tcp_clamp_window(struct sock *sk)
 387{
 388        struct tcp_sock *tp = tcp_sk(sk);
 389        struct inet_connection_sock *icsk = inet_csk(sk);
 390
 391        icsk->icsk_ack.quick = 0;
 392
 393        if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 394            !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
 395            !tcp_memory_pressure &&
 396            atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
 397                sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 398                                    sysctl_tcp_rmem[2]);
 399        }
 400        if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 401                tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
 402}
 403
 404/* Initialize RCV_MSS value.
 405 * RCV_MSS is an our guess about MSS used by the peer.
 406 * We haven't any direct information about the MSS.
 407 * It's better to underestimate the RCV_MSS rather than overestimate.
 408 * Overestimations make us ACKing less frequently than needed.
 409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
 410 */
 411void tcp_initialize_rcv_mss(struct sock *sk)
 412{
 413        struct tcp_sock *tp = tcp_sk(sk);
 414        unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
 415
 416        hint = min(hint, tp->rcv_wnd / 2);
 417        hint = min(hint, TCP_MSS_DEFAULT);
 418        hint = max(hint, TCP_MIN_MSS);
 419
 420        inet_csk(sk)->icsk_ack.rcv_mss = hint;
 421}
 422
 423/* Receiver "autotuning" code.
 424 *
 425 * The algorithm for RTT estimation w/o timestamps is based on
 426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
 427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
 428 *
 429 * More detail on this code can be found at
 430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
 431 * though this reference is out of date.  A new paper
 432 * is pending.
 433 */
 434static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 435{
 436        u32 new_sample = tp->rcv_rtt_est.rtt;
 437        long m = sample;
 438
 439        if (m == 0)
 440                m = 1;
 441
 442        if (new_sample != 0) {
 443                /* If we sample in larger samples in the non-timestamp
 444                 * case, we could grossly overestimate the RTT especially
 445                 * with chatty applications or bulk transfer apps which
 446                 * are stalled on filesystem I/O.
 447                 *
 448                 * Also, since we are only going for a minimum in the
 449                 * non-timestamp case, we do not smooth things out
 450                 * else with timestamps disabled convergence takes too
 451                 * long.
 452                 */
 453                if (!win_dep) {
 454                        m -= (new_sample >> 3);
 455                        new_sample += m;
 456                } else if (m < new_sample)
 457                        new_sample = m << 3;
 458        } else {
 459                /* No previous measure. */
 460                new_sample = m << 3;
 461        }
 462
 463        if (tp->rcv_rtt_est.rtt != new_sample)
 464                tp->rcv_rtt_est.rtt = new_sample;
 465}
 466
 467static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
 468{
 469        if (tp->rcv_rtt_est.time == 0)
 470                goto new_measure;
 471        if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
 472                return;
 473        tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
 474
 475new_measure:
 476        tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
 477        tp->rcv_rtt_est.time = tcp_time_stamp;
 478}
 479
 480static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 481                                          const struct sk_buff *skb)
 482{
 483        struct tcp_sock *tp = tcp_sk(sk);
 484        if (tp->rx_opt.rcv_tsecr &&
 485            (TCP_SKB_CB(skb)->end_seq -
 486             TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
 487                tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
 488}
 489
 490/*
 491 * This function should be called every time data is copied to user space.
 492 * It calculates the appropriate TCP receive buffer space.
 493 */
 494void tcp_rcv_space_adjust(struct sock *sk)
 495{
 496        struct tcp_sock *tp = tcp_sk(sk);
 497        int time;
 498        int space;
 499
 500        if (tp->rcvq_space.time == 0)
 501                goto new_measure;
 502
 503        time = tcp_time_stamp - tp->rcvq_space.time;
 504        if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
 505                return;
 506
 507        space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
 508
 509        space = max(tp->rcvq_space.space, space);
 510
 511        if (tp->rcvq_space.space != space) {
 512                int rcvmem;
 513
 514                tp->rcvq_space.space = space;
 515
 516                if (sysctl_tcp_moderate_rcvbuf &&
 517                    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 518                        int new_clamp = space;
 519
 520                        /* Receive space grows, normalize in order to
 521                         * take into account packet headers and sk_buff
 522                         * structure overhead.
 523                         */
 524                        space /= tp->advmss;
 525                        if (!space)
 526                                space = 1;
 527                        rcvmem = (tp->advmss + MAX_TCP_HEADER +
 528                                  16 + sizeof(struct sk_buff));
 529                        while (tcp_win_from_space(rcvmem) < tp->advmss)
 530                                rcvmem += 128;
 531                        space *= rcvmem;
 532                        space = min(space, sysctl_tcp_rmem[2]);
 533                        if (space > sk->sk_rcvbuf) {
 534                                sk->sk_rcvbuf = space;
 535
 536                                /* Make the window clamp follow along.  */
 537                                tp->window_clamp = new_clamp;
 538                        }
 539                }
 540        }
 541
 542new_measure:
 543        tp->rcvq_space.seq = tp->copied_seq;
 544        tp->rcvq_space.time = tcp_time_stamp;
 545}
 546
 547/* There is something which you must keep in mind when you analyze the
 548 * behavior of the tp->ato delayed ack timeout interval.  When a
 549 * connection starts up, we want to ack as quickly as possible.  The
 550 * problem is that "good" TCP's do slow start at the beginning of data
 551 * transmission.  The means that until we send the first few ACK's the
 552 * sender will sit on his end and only queue most of his data, because
 553 * he can only send snd_cwnd unacked packets at any given time.  For
 554 * each ACK we send, he increments snd_cwnd and transmits more of his
 555 * queue.  -DaveM
 556 */
 557static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
 558{
 559        struct tcp_sock *tp = tcp_sk(sk);
 560        struct inet_connection_sock *icsk = inet_csk(sk);
 561        u32 now;
 562
 563        inet_csk_schedule_ack(sk);
 564
 565        tcp_measure_rcv_mss(sk, skb);
 566
 567        tcp_rcv_rtt_measure(tp);
 568
 569        now = tcp_time_stamp;
 570
 571        if (!icsk->icsk_ack.ato) {
 572                /* The _first_ data packet received, initialize
 573                 * delayed ACK engine.
 574                 */
 575                tcp_incr_quickack(sk);
 576                icsk->icsk_ack.ato = TCP_ATO_MIN;
 577        } else {
 578                int m = now - icsk->icsk_ack.lrcvtime;
 579
 580                if (m <= TCP_ATO_MIN / 2) {
 581                        /* The fastest case is the first. */
 582                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
 583                } else if (m < icsk->icsk_ack.ato) {
 584                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
 585                        if (icsk->icsk_ack.ato > icsk->icsk_rto)
 586                                icsk->icsk_ack.ato = icsk->icsk_rto;
 587                } else if (m > icsk->icsk_rto) {
 588                        /* Too long gap. Apparently sender failed to
 589                         * restart window, so that we send ACKs quickly.
 590                         */
 591                        tcp_incr_quickack(sk);
 592                        sk_mem_reclaim(sk);
 593                }
 594        }
 595        icsk->icsk_ack.lrcvtime = now;
 596
 597        TCP_ECN_check_ce(tp, skb);
 598
 599        if (skb->len >= 128)
 600                tcp_grow_window(sk, skb);
 601}
 602
 603/* Called to compute a smoothed rtt estimate. The data fed to this
 604 * routine either comes from timestamps, or from segments that were
 605 * known _not_ to have been retransmitted [see Karn/Partridge
 606 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
 607 * piece by Van Jacobson.
 608 * NOTE: the next three routines used to be one big routine.
 609 * To save cycles in the RFC 1323 implementation it was better to break
 610 * it up into three procedures. -- erics
 611 */
 612static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
 613{
 614        struct tcp_sock *tp = tcp_sk(sk);
 615        long m = mrtt; /* RTT */
 616
 617        /*      The following amusing code comes from Jacobson's
 618         *      article in SIGCOMM '88.  Note that rtt and mdev
 619         *      are scaled versions of rtt and mean deviation.
 620         *      This is designed to be as fast as possible
 621         *      m stands for "measurement".
 622         *
 623         *      On a 1990 paper the rto value is changed to:
 624         *      RTO = rtt + 4 * mdev
 625         *
 626         * Funny. This algorithm seems to be very broken.
 627         * These formulae increase RTO, when it should be decreased, increase
 628         * too slowly, when it should be increased quickly, decrease too quickly
 629         * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
 630         * does not matter how to _calculate_ it. Seems, it was trap
 631         * that VJ failed to avoid. 8)
 632         */
 633        if (m == 0)
 634                m = 1;
 635        if (tp->srtt != 0) {
 636                m -= (tp->srtt >> 3);   /* m is now error in rtt est */
 637                tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
 638                if (m < 0) {
 639                        m = -m;         /* m is now abs(error) */
 640                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 641                        /* This is similar to one of Eifel findings.
 642                         * Eifel blocks mdev updates when rtt decreases.
 643                         * This solution is a bit different: we use finer gain
 644                         * for mdev in this case (alpha*beta).
 645                         * Like Eifel it also prevents growth of rto,
 646                         * but also it limits too fast rto decreases,
 647                         * happening in pure Eifel.
 648                         */
 649                        if (m > 0)
 650                                m >>= 3;
 651                } else {
 652                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 653                }
 654                tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
 655                if (tp->mdev > tp->mdev_max) {
 656                        tp->mdev_max = tp->mdev;
 657                        if (tp->mdev_max > tp->rttvar)
 658                                tp->rttvar = tp->mdev_max;
 659                }
 660                if (after(tp->snd_una, tp->rtt_seq)) {
 661                        if (tp->mdev_max < tp->rttvar)
 662                                tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
 663                        tp->rtt_seq = tp->snd_nxt;
 664                        tp->mdev_max = tcp_rto_min(sk);
 665                }
 666        } else {
 667                /* no previous measure. */
 668                tp->srtt = m << 3;      /* take the measured time to be rtt */
 669                tp->mdev = m << 1;      /* make sure rto = 3*rtt */
 670                tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 671                tp->rtt_seq = tp->snd_nxt;
 672        }
 673}
 674
 675/* Calculate rto without backoff.  This is the second half of Van Jacobson's
 676 * routine referred to above.
 677 */
 678static inline void tcp_set_rto(struct sock *sk)
 679{
 680        const struct tcp_sock *tp = tcp_sk(sk);
 681        /* Old crap is replaced with new one. 8)
 682         *
 683         * More seriously:
 684         * 1. If rtt variance happened to be less 50msec, it is hallucination.
 685         *    It cannot be less due to utterly erratic ACK generation made
 686         *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
 687         *    to do with delayed acks, because at cwnd>2 true delack timeout
 688         *    is invisible. Actually, Linux-2.4 also generates erratic
 689         *    ACKs in some circumstances.
 690         */
 691        inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
 692
 693        /* 2. Fixups made earlier cannot be right.
 694         *    If we do not estimate RTO correctly without them,
 695         *    all the algo is pure shit and should be replaced
 696         *    with correct one. It is exactly, which we pretend to do.
 697         */
 698
 699        /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
 700         * guarantees that rto is higher.
 701         */
 702        tcp_bound_rto(sk);
 703}
 704
 705/* Save metrics learned by this TCP session.
 706   This function is called only, when TCP finishes successfully
 707   i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
 708 */
 709void tcp_update_metrics(struct sock *sk)
 710{
 711        struct tcp_sock *tp = tcp_sk(sk);
 712        struct dst_entry *dst = __sk_dst_get(sk);
 713
 714        if (sysctl_tcp_nometrics_save)
 715                return;
 716
 717        dst_confirm(dst);
 718
 719        if (dst && (dst->flags & DST_HOST)) {
 720                const struct inet_connection_sock *icsk = inet_csk(sk);
 721                int m;
 722                unsigned long rtt;
 723
 724                if (icsk->icsk_backoff || !tp->srtt) {
 725                        /* This session failed to estimate rtt. Why?
 726                         * Probably, no packets returned in time.
 727                         * Reset our results.
 728                         */
 729                        if (!(dst_metric_locked(dst, RTAX_RTT)))
 730                                dst->metrics[RTAX_RTT - 1] = 0;
 731                        return;
 732                }
 733
 734                rtt = dst_metric_rtt(dst, RTAX_RTT);
 735                m = rtt - tp->srtt;
 736
 737                /* If newly calculated rtt larger than stored one,
 738                 * store new one. Otherwise, use EWMA. Remember,
 739                 * rtt overestimation is always better than underestimation.
 740                 */
 741                if (!(dst_metric_locked(dst, RTAX_RTT))) {
 742                        if (m <= 0)
 743                                set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
 744                        else
 745                                set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
 746                }
 747
 748                if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
 749                        unsigned long var;
 750                        if (m < 0)
 751                                m = -m;
 752
 753                        /* Scale deviation to rttvar fixed point */
 754                        m >>= 1;
 755                        if (m < tp->mdev)
 756                                m = tp->mdev;
 757
 758                        var = dst_metric_rtt(dst, RTAX_RTTVAR);
 759                        if (m >= var)
 760                                var = m;
 761                        else
 762                                var -= (var - m) >> 2;
 763
 764                        set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
 765                }
 766
 767                if (tcp_in_initial_slowstart(tp)) {
 768                        /* Slow start still did not finish. */
 769                        if (dst_metric(dst, RTAX_SSTHRESH) &&
 770                            !dst_metric_locked(dst, RTAX_SSTHRESH) &&
 771                            (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
 772                                dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
 773                        if (!dst_metric_locked(dst, RTAX_CWND) &&
 774                            tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
 775                                dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
 776                } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 777                           icsk->icsk_ca_state == TCP_CA_Open) {
 778                        /* Cong. avoidance phase, cwnd is reliable. */
 779                        if (!dst_metric_locked(dst, RTAX_SSTHRESH))
 780                                dst->metrics[RTAX_SSTHRESH-1] =
 781                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
 782                        if (!dst_metric_locked(dst, RTAX_CWND))
 783                                dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
 784                } else {
 785                        /* Else slow start did not finish, cwnd is non-sense,
 786                           ssthresh may be also invalid.
 787                         */
 788                        if (!dst_metric_locked(dst, RTAX_CWND))
 789                                dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
 790                        if (dst_metric(dst, RTAX_SSTHRESH) &&
 791                            !dst_metric_locked(dst, RTAX_SSTHRESH) &&
 792                            tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
 793                                dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
 794                }
 795
 796                if (!dst_metric_locked(dst, RTAX_REORDERING)) {
 797                        if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
 798                            tp->reordering != sysctl_tcp_reordering)
 799                                dst->metrics[RTAX_REORDERING-1] = tp->reordering;
 800                }
 801        }
 802}
 803
 804/* Numbers are taken from RFC3390.
 805 *
 806 * John Heffner states:
 807 *
 808 *      The RFC specifies a window of no more than 4380 bytes
 809 *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
 810 *      is a bit misleading because they use a clamp at 4380 bytes
 811 *      rather than use a multiplier in the relevant range.
 812 */
 813__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
 814{
 815        __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 816
 817        if (!cwnd) {
 818                if (tp->mss_cache > 1460)
 819                        cwnd = 2;
 820                else
 821                        cwnd = (tp->mss_cache > 1095) ? 3 : 4;
 822        }
 823        return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 824}
 825
 826/* Set slow start threshold and cwnd not falling to slow start */
 827void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
 828{
 829        struct tcp_sock *tp = tcp_sk(sk);
 830        const struct inet_connection_sock *icsk = inet_csk(sk);
 831
 832        tp->prior_ssthresh = 0;
 833        tp->bytes_acked = 0;
 834        if (icsk->icsk_ca_state < TCP_CA_CWR) {
 835                tp->undo_marker = 0;
 836                if (set_ssthresh)
 837                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
 838                tp->snd_cwnd = min(tp->snd_cwnd,
 839                                   tcp_packets_in_flight(tp) + 1U);
 840                tp->snd_cwnd_cnt = 0;
 841                tp->high_seq = tp->snd_nxt;
 842                tp->snd_cwnd_stamp = tcp_time_stamp;
 843                TCP_ECN_queue_cwr(tp);
 844
 845                tcp_set_ca_state(sk, TCP_CA_CWR);
 846        }
 847}
 848
 849/*
 850 * Packet counting of FACK is based on in-order assumptions, therefore TCP
 851 * disables it when reordering is detected
 852 */
 853static void tcp_disable_fack(struct tcp_sock *tp)
 854{
 855        /* RFC3517 uses different metric in lost marker => reset on change */
 856        if (tcp_is_fack(tp))
 857                tp->lost_skb_hint = NULL;
 858        tp->rx_opt.sack_ok &= ~2;
 859}
 860
 861/* Take a notice that peer is sending D-SACKs */
 862static void tcp_dsack_seen(struct tcp_sock *tp)
 863{
 864        tp->rx_opt.sack_ok |= 4;
 865}
 866
 867/* Initialize metrics on socket. */
 868
 869static void tcp_init_metrics(struct sock *sk)
 870{
 871        struct tcp_sock *tp = tcp_sk(sk);
 872        struct dst_entry *dst = __sk_dst_get(sk);
 873
 874        if (dst == NULL)
 875                goto reset;
 876
 877        dst_confirm(dst);
 878
 879        if (dst_metric_locked(dst, RTAX_CWND))
 880                tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
 881        if (dst_metric(dst, RTAX_SSTHRESH)) {
 882                tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
 883                if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 884                        tp->snd_ssthresh = tp->snd_cwnd_clamp;
 885        }
 886        if (dst_metric(dst, RTAX_REORDERING) &&
 887            tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
 888                tcp_disable_fack(tp);
 889                tp->reordering = dst_metric(dst, RTAX_REORDERING);
 890        }
 891
 892        if (dst_metric(dst, RTAX_RTT) == 0)
 893                goto reset;
 894
 895        if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
 896                goto reset;
 897
 898        /* Initial rtt is determined from SYN,SYN-ACK.
 899         * The segment is small and rtt may appear much
 900         * less than real one. Use per-dst memory
 901         * to make it more realistic.
 902         *
 903         * A bit of theory. RTT is time passed after "normal" sized packet
 904         * is sent until it is ACKed. In normal circumstances sending small
 905         * packets force peer to delay ACKs and calculation is correct too.
 906         * The algorithm is adaptive and, provided we follow specs, it
 907         * NEVER underestimate RTT. BUT! If peer tries to make some clever
 908         * tricks sort of "quick acks" for time long enough to decrease RTT
 909         * to low value, and then abruptly stops to do it and starts to delay
 910         * ACKs, wait for troubles.
 911         */
 912        if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
 913                tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
 914                tp->rtt_seq = tp->snd_nxt;
 915        }
 916        if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
 917                tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
 918                tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 919        }
 920        tcp_set_rto(sk);
 921        if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
 922                goto reset;
 923
 924cwnd:
 925        tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 926        tp->snd_cwnd_stamp = tcp_time_stamp;
 927        return;
 928
 929reset:
 930        /* Play conservative. If timestamps are not
 931         * supported, TCP will fail to recalculate correct
 932         * rtt, if initial rto is too small. FORGET ALL AND RESET!
 933         */
 934        if (!tp->rx_opt.saw_tstamp && tp->srtt) {
 935                tp->srtt = 0;
 936                tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
 937                inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
 938        }
 939        goto cwnd;
 940}
 941
 942static void tcp_update_reordering(struct sock *sk, const int metric,
 943                                  const int ts)
 944{
 945        struct tcp_sock *tp = tcp_sk(sk);
 946        if (metric > tp->reordering) {
 947                int mib_idx;
 948
 949                tp->reordering = min(TCP_MAX_REORDERING, metric);
 950
 951                /* This exciting event is worth to be remembered. 8) */
 952                if (ts)
 953                        mib_idx = LINUX_MIB_TCPTSREORDER;
 954                else if (tcp_is_reno(tp))
 955                        mib_idx = LINUX_MIB_TCPRENOREORDER;
 956                else if (tcp_is_fack(tp))
 957                        mib_idx = LINUX_MIB_TCPFACKREORDER;
 958                else
 959                        mib_idx = LINUX_MIB_TCPSACKREORDER;
 960
 961                NET_INC_STATS_BH(sock_net(sk), mib_idx);
 962#if FASTRETRANS_DEBUG > 1
 963                printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
 964                       tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
 965                       tp->reordering,
 966                       tp->fackets_out,
 967                       tp->sacked_out,
 968                       tp->undo_marker ? tp->undo_retrans : 0);
 969#endif
 970                tcp_disable_fack(tp);
 971        }
 972}
 973
 974/* This must be called before lost_out is incremented */
 975static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
 976{
 977        if ((tp->retransmit_skb_hint == NULL) ||
 978            before(TCP_SKB_CB(skb)->seq,
 979                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
 980                tp->retransmit_skb_hint = skb;
 981
 982        if (!tp->lost_out ||
 983            after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
 984                tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
 985}
 986
 987static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
 988{
 989        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 990                tcp_verify_retransmit_hint(tp, skb);
 991
 992                tp->lost_out += tcp_skb_pcount(skb);
 993                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 994        }
 995}
 996
 997static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
 998                                            struct sk_buff *skb)
 999{
1000        tcp_verify_retransmit_hint(tp, skb);
1001
1002        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1003                tp->lost_out += tcp_skb_pcount(skb);
1004                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1005        }
1006}
1007
1008/* This procedure tags the retransmission queue when SACKs arrive.
1009 *
1010 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1011 * Packets in queue with these bits set are counted in variables
1012 * sacked_out, retrans_out and lost_out, correspondingly.
1013 *
1014 * Valid combinations are:
1015 * Tag  InFlight        Description
1016 * 0    1               - orig segment is in flight.
1017 * S    0               - nothing flies, orig reached receiver.
1018 * L    0               - nothing flies, orig lost by net.
1019 * R    2               - both orig and retransmit are in flight.
1020 * L|R  1               - orig is lost, retransmit is in flight.
1021 * S|R  1               - orig reached receiver, retrans is still in flight.
1022 * (L|S|R is logically valid, it could occur when L|R is sacked,
1023 *  but it is equivalent to plain S and code short-curcuits it to S.
1024 *  L|S is logically invalid, it would mean -1 packet in flight 8))
1025 *
1026 * These 6 states form finite state machine, controlled by the following events:
1027 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1028 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1029 * 3. Loss detection event of one of three flavors:
1030 *      A. Scoreboard estimator decided the packet is lost.
1031 *         A'. Reno "three dupacks" marks head of queue lost.
1032 *         A''. Its FACK modfication, head until snd.fack is lost.
1033 *      B. SACK arrives sacking data transmitted after never retransmitted
1034 *         hole was sent out.
1035 *      C. SACK arrives sacking SND.NXT at the moment, when the
1036 *         segment was retransmitted.
1037 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1038 *
1039 * It is pleasant to note, that state diagram turns out to be commutative,
1040 * so that we are allowed not to be bothered by order of our actions,
1041 * when multiple events arrive simultaneously. (see the function below).
1042 *
1043 * Reordering detection.
1044 * --------------------
1045 * Reordering metric is maximal distance, which a packet can be displaced
1046 * in packet stream. With SACKs we can estimate it:
1047 *
1048 * 1. SACK fills old hole and the corresponding segment was not
1049 *    ever retransmitted -> reordering. Alas, we cannot use it
1050 *    when segment was retransmitted.
1051 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1052 *    for retransmitted and already SACKed segment -> reordering..
1053 * Both of these heuristics are not used in Loss state, when we cannot
1054 * account for retransmits accurately.
1055 *
1056 * SACK block validation.
1057 * ----------------------
1058 *
1059 * SACK block range validation checks that the received SACK block fits to
1060 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1061 * Note that SND.UNA is not included to the range though being valid because
1062 * it means that the receiver is rather inconsistent with itself reporting
1063 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1064 * perfectly valid, however, in light of RFC2018 which explicitly states
1065 * that "SACK block MUST reflect the newest segment.  Even if the newest
1066 * segment is going to be discarded ...", not that it looks very clever
1067 * in case of head skb. Due to potentional receiver driven attacks, we
1068 * choose to avoid immediate execution of a walk in write queue due to
1069 * reneging and defer head skb's loss recovery to standard loss recovery
1070 * procedure that will eventually trigger (nothing forbids us doing this).
1071 *
1072 * Implements also blockage to start_seq wrap-around. Problem lies in the
1073 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1074 * there's no guarantee that it will be before snd_nxt (n). The problem
1075 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1076 * wrap (s_w):
1077 *
1078 *         <- outs wnd ->                          <- wrapzone ->
1079 *         u     e      n                         u_w   e_w  s n_w
1080 *         |     |      |                          |     |   |  |
1081 * |<------------+------+----- TCP seqno space --------------+---------->|
1082 * ...-- <2^31 ->|                                           |<--------...
1083 * ...---- >2^31 ------>|                                    |<--------...
1084 *
1085 * Current code wouldn't be vulnerable but it's better still to discard such
1086 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1087 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1088 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1089 * equal to the ideal case (infinite seqno space without wrap caused issues).
1090 *
1091 * With D-SACK the lower bound is extended to cover sequence space below
1092 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1093 * again, D-SACK block must not to go across snd_una (for the same reason as
1094 * for the normal SACK blocks, explained above). But there all simplicity
1095 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1096 * fully below undo_marker they do not affect behavior in anyway and can
1097 * therefore be safely ignored. In rare cases (which are more or less
1098 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1099 * fragmentation and packet reordering past skb's retransmission. To consider
1100 * them correctly, the acceptable range must be extended even more though
1101 * the exact amount is rather hard to quantify. However, tp->max_window can
1102 * be used as an exaggerated estimate.
1103 */
1104static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1105                                  u32 start_seq, u32 end_seq)
1106{
1107        /* Too far in future, or reversed (interpretation is ambiguous) */
1108        if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1109                return 0;
1110
1111        /* Nasty start_seq wrap-around check (see comments above) */
1112        if (!before(start_seq, tp->snd_nxt))
1113                return 0;
1114
1115        /* In outstanding window? ...This is valid exit for D-SACKs too.
1116         * start_seq == snd_una is non-sensical (see comments above)
1117         */
1118        if (after(start_seq, tp->snd_una))
1119                return 1;
1120
1121        if (!is_dsack || !tp->undo_marker)
1122                return 0;
1123
1124        /* ...Then it's D-SACK, and must reside below snd_una completely */
1125        if (!after(end_seq, tp->snd_una))
1126                return 0;
1127
1128        if (!before(start_seq, tp->undo_marker))
1129                return 1;
1130
1131        /* Too old */
1132        if (!after(end_seq, tp->undo_marker))
1133                return 0;
1134
1135        /* Undo_marker boundary crossing (overestimates a lot). Known already:
1136         *   start_seq < undo_marker and end_seq >= undo_marker.
1137         */
1138        return !before(start_seq, end_seq - tp->max_window);
1139}
1140
1141/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1142 * Event "C". Later note: FACK people cheated me again 8), we have to account
1143 * for reordering! Ugly, but should help.
1144 *
1145 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1146 * less than what is now known to be received by the other end (derived from
1147 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1148 * retransmitted skbs to avoid some costly processing per ACKs.
1149 */
1150static void tcp_mark_lost_retrans(struct sock *sk)
1151{
1152        const struct inet_connection_sock *icsk = inet_csk(sk);
1153        struct tcp_sock *tp = tcp_sk(sk);
1154        struct sk_buff *skb;
1155        int cnt = 0;
1156        u32 new_low_seq = tp->snd_nxt;
1157        u32 received_upto = tcp_highest_sack_seq(tp);
1158
1159        if (!tcp_is_fack(tp) || !tp->retrans_out ||
1160            !after(received_upto, tp->lost_retrans_low) ||
1161            icsk->icsk_ca_state != TCP_CA_Recovery)
1162                return;
1163
1164        tcp_for_write_queue(skb, sk) {
1165                u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1166
1167                if (skb == tcp_send_head(sk))
1168                        break;
1169                if (cnt == tp->retrans_out)
1170                        break;
1171                if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1172                        continue;
1173
1174                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1175                        continue;
1176
1177                /* TODO: We would like to get rid of tcp_is_fack(tp) only
1178                 * constraint here (see above) but figuring out that at
1179                 * least tp->reordering SACK blocks reside between ack_seq
1180                 * and received_upto is not easy task to do cheaply with
1181                 * the available datastructures.
1182                 *
1183                 * Whether FACK should check here for tp->reordering segs
1184                 * in-between one could argue for either way (it would be
1185                 * rather simple to implement as we could count fack_count
1186                 * during the walk and do tp->fackets_out - fack_count).
1187                 */
1188                if (after(received_upto, ack_seq)) {
1189                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1190                        tp->retrans_out -= tcp_skb_pcount(skb);
1191
1192                        tcp_skb_mark_lost_uncond_verify(tp, skb);
1193                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1194                } else {
1195                        if (before(ack_seq, new_low_seq))
1196                                new_low_seq = ack_seq;
1197                        cnt += tcp_skb_pcount(skb);
1198                }
1199        }
1200
1201        if (tp->retrans_out)
1202                tp->lost_retrans_low = new_low_seq;
1203}
1204
1205static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1206                           struct tcp_sack_block_wire *sp, int num_sacks,
1207                           u32 prior_snd_una)
1208{
1209        struct tcp_sock *tp = tcp_sk(sk);
1210        u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1211        u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1212        int dup_sack = 0;
1213
1214        if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1215                dup_sack = 1;
1216                tcp_dsack_seen(tp);
1217                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1218        } else if (num_sacks > 1) {
1219                u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1220                u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1221
1222                if (!after(end_seq_0, end_seq_1) &&
1223                    !before(start_seq_0, start_seq_1)) {
1224                        dup_sack = 1;
1225                        tcp_dsack_seen(tp);
1226                        NET_INC_STATS_BH(sock_net(sk),
1227                                        LINUX_MIB_TCPDSACKOFORECV);
1228                }
1229        }
1230
1231        /* D-SACK for already forgotten data... Do dumb counting. */
1232        if (dup_sack &&
1233            !after(end_seq_0, prior_snd_una) &&
1234            after(end_seq_0, tp->undo_marker))
1235                tp->undo_retrans--;
1236
1237        return dup_sack;
1238}
1239
1240struct tcp_sacktag_state {
1241        int reord;
1242        int fack_count;
1243        int flag;
1244};
1245
1246/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1247 * the incoming SACK may not exactly match but we can find smaller MSS
1248 * aligned portion of it that matches. Therefore we might need to fragment
1249 * which may fail and creates some hassle (caller must handle error case
1250 * returns).
1251 *
1252 * FIXME: this could be merged to shift decision code
1253 */
1254static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1255                                 u32 start_seq, u32 end_seq)
1256{
1257        int in_sack, err;
1258        unsigned int pkt_len;
1259        unsigned int mss;
1260
1261        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1262                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1263
1264        if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1265            after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1266                mss = tcp_skb_mss(skb);
1267                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1268
1269                if (!in_sack) {
1270                        pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1271                        if (pkt_len < mss)
1272                                pkt_len = mss;
1273                } else {
1274                        pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1275                        if (pkt_len < mss)
1276                                return -EINVAL;
1277                }
1278
1279                /* Round if necessary so that SACKs cover only full MSSes
1280                 * and/or the remaining small portion (if present)
1281                 */
1282                if (pkt_len > mss) {
1283                        unsigned int new_len = (pkt_len / mss) * mss;
1284                        if (!in_sack && new_len < pkt_len) {
1285                                new_len += mss;
1286                                if (new_len > skb->len)
1287                                        return 0;
1288                        }
1289                        pkt_len = new_len;
1290                }
1291                err = tcp_fragment(sk, skb, pkt_len, mss);
1292                if (err < 0)
1293                        return err;
1294        }
1295
1296        return in_sack;
1297}
1298
1299static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1300                          struct tcp_sacktag_state *state,
1301                          int dup_sack, int pcount)
1302{
1303        struct tcp_sock *tp = tcp_sk(sk);
1304        u8 sacked = TCP_SKB_CB(skb)->sacked;
1305        int fack_count = state->fack_count;
1306
1307        /* Account D-SACK for retransmitted packet. */
1308        if (dup_sack && (sacked & TCPCB_RETRANS)) {
1309                if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1310                        tp->undo_retrans--;
1311                if (sacked & TCPCB_SACKED_ACKED)
1312                        state->reord = min(fack_count, state->reord);
1313        }
1314
1315        /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1316        if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1317                return sacked;
1318
1319        if (!(sacked & TCPCB_SACKED_ACKED)) {
1320                if (sacked & TCPCB_SACKED_RETRANS) {
1321                        /* If the segment is not tagged as lost,
1322                         * we do not clear RETRANS, believing
1323                         * that retransmission is still in flight.
1324                         */
1325                        if (sacked & TCPCB_LOST) {
1326                                sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1327                                tp->lost_out -= pcount;
1328                                tp->retrans_out -= pcount;
1329                        }
1330                } else {
1331                        if (!(sacked & TCPCB_RETRANS)) {
1332                                /* New sack for not retransmitted frame,
1333                                 * which was in hole. It is reordering.
1334                                 */
1335                                if (before(TCP_SKB_CB(skb)->seq,
1336                                           tcp_highest_sack_seq(tp)))
1337                                        state->reord = min(fack_count,
1338                                                           state->reord);
1339
1340                                /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1341                                if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1342                                        state->flag |= FLAG_ONLY_ORIG_SACKED;
1343                        }
1344
1345                        if (sacked & TCPCB_LOST) {
1346                                sacked &= ~TCPCB_LOST;
1347                                tp->lost_out -= pcount;
1348                        }
1349                }
1350
1351                sacked |= TCPCB_SACKED_ACKED;
1352                state->flag |= FLAG_DATA_SACKED;
1353                tp->sacked_out += pcount;
1354
1355                fack_count += pcount;
1356
1357                /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1358                if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1359                    before(TCP_SKB_CB(skb)->seq,
1360                           TCP_SKB_CB(tp->lost_skb_hint)->seq))
1361                        tp->lost_cnt_hint += pcount;
1362
1363                if (fack_count > tp->fackets_out)
1364                        tp->fackets_out = fack_count;
1365        }
1366
1367        /* D-SACK. We can detect redundant retransmission in S|R and plain R
1368         * frames and clear it. undo_retrans is decreased above, L|R frames
1369         * are accounted above as well.
1370         */
1371        if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1372                sacked &= ~TCPCB_SACKED_RETRANS;
1373                tp->retrans_out -= pcount;
1374        }
1375
1376        return sacked;
1377}
1378
1379static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1380                           struct tcp_sacktag_state *state,
1381                           unsigned int pcount, int shifted, int mss,
1382                           int dup_sack)
1383{
1384        struct tcp_sock *tp = tcp_sk(sk);
1385        struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1386
1387        BUG_ON(!pcount);
1388
1389        /* Tweak before seqno plays */
1390        if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1391            !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1392                tp->lost_cnt_hint += pcount;
1393
1394        TCP_SKB_CB(prev)->end_seq += shifted;
1395        TCP_SKB_CB(skb)->seq += shifted;
1396
1397        skb_shinfo(prev)->gso_segs += pcount;
1398        BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1399        skb_shinfo(skb)->gso_segs -= pcount;
1400
1401        /* When we're adding to gso_segs == 1, gso_size will be zero,
1402         * in theory this shouldn't be necessary but as long as DSACK
1403         * code can come after this skb later on it's better to keep
1404         * setting gso_size to something.
1405         */
1406        if (!skb_shinfo(prev)->gso_size) {
1407                skb_shinfo(prev)->gso_size = mss;
1408                skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1409        }
1410
1411        /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1412        if (skb_shinfo(skb)->gso_segs <= 1) {
1413                skb_shinfo(skb)->gso_size = 0;
1414                skb_shinfo(skb)->gso_type = 0;
1415        }
1416
1417        /* We discard results */
1418        tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1419
1420        /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1421        TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1422
1423        if (skb->len > 0) {
1424                BUG_ON(!tcp_skb_pcount(skb));
1425                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1426                return 0;
1427        }
1428
1429        /* Whole SKB was eaten :-) */
1430
1431        if (skb == tp->retransmit_skb_hint)
1432                tp->retransmit_skb_hint = prev;
1433        if (skb == tp->scoreboard_skb_hint)
1434                tp->scoreboard_skb_hint = prev;
1435        if (skb == tp->lost_skb_hint) {
1436                tp->lost_skb_hint = prev;
1437                tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1438        }
1439
1440        TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1441        if (skb == tcp_highest_sack(sk))
1442                tcp_advance_highest_sack(sk, skb);
1443
1444        tcp_unlink_write_queue(skb, sk);
1445        sk_wmem_free_skb(sk, skb);
1446
1447        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1448
1449        return 1;
1450}
1451
1452/* I wish gso_size would have a bit more sane initialization than
1453 * something-or-zero which complicates things
1454 */
1455static int tcp_skb_seglen(struct sk_buff *skb)
1456{
1457        return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1458}
1459
1460/* Shifting pages past head area doesn't work */
1461static int skb_can_shift(struct sk_buff *skb)
1462{
1463        return !skb_headlen(skb) && skb_is_nonlinear(skb);
1464}
1465
1466/* Try collapsing SACK blocks spanning across multiple skbs to a single
1467 * skb.
1468 */
1469static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1470                                          struct tcp_sacktag_state *state,
1471                                          u32 start_seq, u32 end_seq,
1472                                          int dup_sack)
1473{
1474        struct tcp_sock *tp = tcp_sk(sk);
1475        struct sk_buff *prev;
1476        int mss;
1477        int pcount = 0;
1478        int len;
1479        int in_sack;
1480
1481        if (!sk_can_gso(sk))
1482                goto fallback;
1483
1484        /* Normally R but no L won't result in plain S */
1485        if (!dup_sack &&
1486            (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1487                goto fallback;
1488        if (!skb_can_shift(skb))
1489                goto fallback;
1490        /* This frame is about to be dropped (was ACKed). */
1491        if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1492                goto fallback;
1493
1494        /* Can only happen with delayed DSACK + discard craziness */
1495        if (unlikely(skb == tcp_write_queue_head(sk)))
1496                goto fallback;
1497        prev = tcp_write_queue_prev(sk, skb);
1498
1499        if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1500                goto fallback;
1501
1502        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1503                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1504
1505        if (in_sack) {
1506                len = skb->len;
1507                pcount = tcp_skb_pcount(skb);
1508                mss = tcp_skb_seglen(skb);
1509
1510                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1511                 * drop this restriction as unnecessary
1512                 */
1513                if (mss != tcp_skb_seglen(prev))
1514                        goto fallback;
1515        } else {
1516                if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1517                        goto noop;
1518                /* CHECKME: This is non-MSS split case only?, this will
1519                 * cause skipped skbs due to advancing loop btw, original
1520                 * has that feature too
1521                 */
1522                if (tcp_skb_pcount(skb) <= 1)
1523                        goto noop;
1524
1525                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1526                if (!in_sack) {
1527                        /* TODO: head merge to next could be attempted here
1528                         * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1529                         * though it might not be worth of the additional hassle
1530                         *
1531                         * ...we can probably just fallback to what was done
1532                         * previously. We could try merging non-SACKed ones
1533                         * as well but it probably isn't going to buy off
1534                         * because later SACKs might again split them, and
1535                         * it would make skb timestamp tracking considerably
1536                         * harder problem.
1537                         */
1538                        goto fallback;
1539                }
1540
1541                len = end_seq - TCP_SKB_CB(skb)->seq;
1542                BUG_ON(len < 0);
1543                BUG_ON(len > skb->len);
1544
1545                /* MSS boundaries should be honoured or else pcount will
1546                 * severely break even though it makes things bit trickier.
1547                 * Optimize common case to avoid most of the divides
1548                 */
1549                mss = tcp_skb_mss(skb);
1550
1551                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1552                 * drop this restriction as unnecessary
1553                 */
1554                if (mss != tcp_skb_seglen(prev))
1555                        goto fallback;
1556
1557                if (len == mss) {
1558                        pcount = 1;
1559                } else if (len < mss) {
1560                        goto noop;
1561                } else {
1562                        pcount = len / mss;
1563                        len = pcount * mss;
1564                }
1565        }
1566
1567        if (!skb_shift(prev, skb, len))
1568                goto fallback;
1569        if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1570                goto out;
1571
1572        /* Hole filled allows collapsing with the next as well, this is very
1573         * useful when hole on every nth skb pattern happens
1574         */
1575        if (prev == tcp_write_queue_tail(sk))
1576                goto out;
1577        skb = tcp_write_queue_next(sk, prev);
1578
1579        if (!skb_can_shift(skb) ||
1580            (skb == tcp_send_head(sk)) ||
1581            ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1582            (mss != tcp_skb_seglen(skb)))
1583                goto out;
1584
1585        len = skb->len;
1586        if (skb_shift(prev, skb, len)) {
1587                pcount += tcp_skb_pcount(skb);
1588                tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1589        }
1590
1591out:
1592        state->fack_count += pcount;
1593        return prev;
1594
1595noop:
1596        return skb;
1597
1598fallback:
1599        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1600        return NULL;
1601}
1602
1603static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1604                                        struct tcp_sack_block *next_dup,
1605                                        struct tcp_sacktag_state *state,
1606                                        u32 start_seq, u32 end_seq,
1607                                        int dup_sack_in)
1608{
1609        struct tcp_sock *tp = tcp_sk(sk);
1610        struct sk_buff *tmp;
1611
1612        tcp_for_write_queue_from(skb, sk) {
1613                int in_sack = 0;
1614                int dup_sack = dup_sack_in;
1615
1616                if (skb == tcp_send_head(sk))
1617                        break;
1618
1619                /* queue is in-order => we can short-circuit the walk early */
1620                if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1621                        break;
1622
1623                if ((next_dup != NULL) &&
1624                    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1625                        in_sack = tcp_match_skb_to_sack(sk, skb,
1626                                                        next_dup->start_seq,
1627                                                        next_dup->end_seq);
1628                        if (in_sack > 0)
1629                                dup_sack = 1;
1630                }
1631
1632                /* skb reference here is a bit tricky to get right, since
1633                 * shifting can eat and free both this skb and the next,
1634                 * so not even _safe variant of the loop is enough.
1635                 */
1636                if (in_sack <= 0) {
1637                        tmp = tcp_shift_skb_data(sk, skb, state,
1638                                                 start_seq, end_seq, dup_sack);
1639                        if (tmp != NULL) {
1640                                if (tmp != skb) {
1641                                        skb = tmp;
1642                                        continue;
1643                                }
1644
1645                                in_sack = 0;
1646                        } else {
1647                                in_sack = tcp_match_skb_to_sack(sk, skb,
1648                                                                start_seq,
1649                                                                end_seq);
1650                        }
1651                }
1652
1653                if (unlikely(in_sack < 0))
1654                        break;
1655
1656                if (in_sack) {
1657                        TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1658                                                                  state,
1659                                                                  dup_sack,
1660                                                                  tcp_skb_pcount(skb));
1661
1662                        if (!before(TCP_SKB_CB(skb)->seq,
1663                                    tcp_highest_sack_seq(tp)))
1664                                tcp_advance_highest_sack(sk, skb);
1665                }
1666
1667                state->fack_count += tcp_skb_pcount(skb);
1668        }
1669        return skb;
1670}
1671
1672/* Avoid all extra work that is being done by sacktag while walking in
1673 * a normal way
1674 */
1675static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1676                                        struct tcp_sacktag_state *state,
1677                                        u32 skip_to_seq)
1678{
1679        tcp_for_write_queue_from(skb, sk) {
1680                if (skb == tcp_send_head(sk))
1681                        break;
1682
1683                if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1684                        break;
1685
1686                state->fack_count += tcp_skb_pcount(skb);
1687        }
1688        return skb;
1689}
1690
1691static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1692                                                struct sock *sk,
1693                                                struct tcp_sack_block *next_dup,
1694                                                struct tcp_sacktag_state *state,
1695                                                u32 skip_to_seq)
1696{
1697        if (next_dup == NULL)
1698                return skb;
1699
1700        if (before(next_dup->start_seq, skip_to_seq)) {
1701                skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1702                skb = tcp_sacktag_walk(skb, sk, NULL, state,
1703                                       next_dup->start_seq, next_dup->end_seq,
1704                                       1);
1705        }
1706
1707        return skb;
1708}
1709
1710static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1711{
1712        return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1713}
1714
1715static int
1716tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1717                        u32 prior_snd_una)
1718{
1719        const struct inet_connection_sock *icsk = inet_csk(sk);
1720        struct tcp_sock *tp = tcp_sk(sk);
1721        unsigned char *ptr = (skb_transport_header(ack_skb) +
1722                              TCP_SKB_CB(ack_skb)->sacked);
1723        struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1724        struct tcp_sack_block sp[TCP_NUM_SACKS];
1725        struct tcp_sack_block *cache;
1726        struct tcp_sacktag_state state;
1727        struct sk_buff *skb;
1728        int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1729        int used_sacks;
1730        int found_dup_sack = 0;
1731        int i, j;
1732        int first_sack_index;
1733
1734        state.flag = 0;
1735        state.reord = tp->packets_out;
1736
1737        if (!tp->sacked_out) {
1738                if (WARN_ON(tp->fackets_out))
1739                        tp->fackets_out = 0;
1740                tcp_highest_sack_reset(sk);
1741        }
1742
1743        found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1744                                         num_sacks, prior_snd_una);
1745        if (found_dup_sack)
1746                state.flag |= FLAG_DSACKING_ACK;
1747
1748        /* Eliminate too old ACKs, but take into
1749         * account more or less fresh ones, they can
1750         * contain valid SACK info.
1751         */
1752        if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1753                return 0;
1754
1755        if (!tp->packets_out)
1756                goto out;
1757
1758        used_sacks = 0;
1759        first_sack_index = 0;
1760        for (i = 0; i < num_sacks; i++) {
1761                int dup_sack = !i && found_dup_sack;
1762
1763                sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1764                sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1765
1766                if (!tcp_is_sackblock_valid(tp, dup_sack,
1767                                            sp[used_sacks].start_seq,
1768                                            sp[used_sacks].end_seq)) {
1769                        int mib_idx;
1770
1771                        if (dup_sack) {
1772                                if (!tp->undo_marker)
1773                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1774                                else
1775                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1776                        } else {
1777                                /* Don't count olds caused by ACK reordering */
1778                                if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1779                                    !after(sp[used_sacks].end_seq, tp->snd_una))
1780                                        continue;
1781                                mib_idx = LINUX_MIB_TCPSACKDISCARD;
1782                        }
1783
1784                        NET_INC_STATS_BH(sock_net(sk), mib_idx);
1785                        if (i == 0)
1786                                first_sack_index = -1;
1787                        continue;
1788                }
1789
1790                /* Ignore very old stuff early */
1791                if (!after(sp[used_sacks].end_seq, prior_snd_una))
1792                        continue;
1793
1794                used_sacks++;
1795        }
1796
1797        /* order SACK blocks to allow in order walk of the retrans queue */
1798        for (i = used_sacks - 1; i > 0; i--) {
1799                for (j = 0; j < i; j++) {
1800                        if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1801                                swap(sp[j], sp[j + 1]);
1802
1803                                /* Track where the first SACK block goes to */
1804                                if (j == first_sack_index)
1805                                        first_sack_index = j + 1;
1806                        }
1807                }
1808        }
1809
1810        skb = tcp_write_queue_head(sk);
1811        state.fack_count = 0;
1812        i = 0;
1813
1814        if (!tp->sacked_out) {
1815                /* It's already past, so skip checking against it */
1816                cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1817        } else {
1818                cache = tp->recv_sack_cache;
1819                /* Skip empty blocks in at head of the cache */
1820                while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1821                       !cache->end_seq)
1822                        cache++;
1823        }
1824
1825        while (i < used_sacks) {
1826                u32 start_seq = sp[i].start_seq;
1827                u32 end_seq = sp[i].end_seq;
1828                int dup_sack = (found_dup_sack && (i == first_sack_index));
1829                struct tcp_sack_block *next_dup = NULL;
1830
1831                if (found_dup_sack && ((i + 1) == first_sack_index))
1832                        next_dup = &sp[i + 1];
1833
1834                /* Event "B" in the comment above. */
1835                if (after(end_seq, tp->high_seq))
1836                        state.flag |= FLAG_DATA_LOST;
1837
1838                /* Skip too early cached blocks */
1839                while (tcp_sack_cache_ok(tp, cache) &&
1840                       !before(start_seq, cache->end_seq))
1841                        cache++;
1842
1843                /* Can skip some work by looking recv_sack_cache? */
1844                if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1845                    after(end_seq, cache->start_seq)) {
1846
1847                        /* Head todo? */
1848                        if (before(start_seq, cache->start_seq)) {
1849                                skb = tcp_sacktag_skip(skb, sk, &state,
1850                                                       start_seq);
1851                                skb = tcp_sacktag_walk(skb, sk, next_dup,
1852                                                       &state,
1853                                                       start_seq,
1854                                                       cache->start_seq,
1855                                                       dup_sack);
1856                        }
1857
1858                        /* Rest of the block already fully processed? */
1859                        if (!after(end_seq, cache->end_seq))
1860                                goto advance_sp;
1861
1862                        skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1863                                                       &state,
1864                                                       cache->end_seq);
1865
1866                        /* ...tail remains todo... */
1867                        if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1868                                /* ...but better entrypoint exists! */
1869                                skb = tcp_highest_sack(sk);
1870                                if (skb == NULL)
1871                                        break;
1872                                state.fack_count = tp->fackets_out;
1873                                cache++;
1874                                goto walk;
1875                        }
1876
1877                        skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1878                        /* Check overlap against next cached too (past this one already) */
1879                        cache++;
1880                        continue;
1881                }
1882
1883                if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1884                        skb = tcp_highest_sack(sk);
1885                        if (skb == NULL)
1886                                break;
1887                        state.fack_count = tp->fackets_out;
1888                }
1889                skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1890
1891walk:
1892                skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1893                                       start_seq, end_seq, dup_sack);
1894
1895advance_sp:
1896                /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1897                 * due to in-order walk
1898                 */
1899                if (after(end_seq, tp->frto_highmark))
1900                        state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1901
1902                i++;
1903        }
1904
1905        /* Clear the head of the cache sack blocks so we can skip it next time */
1906        for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1907                tp->recv_sack_cache[i].start_seq = 0;
1908                tp->recv_sack_cache[i].end_seq = 0;
1909        }
1910        for (j = 0; j < used_sacks; j++)
1911                tp->recv_sack_cache[i++] = sp[j];
1912
1913        tcp_mark_lost_retrans(sk);
1914
1915        tcp_verify_left_out(tp);
1916
1917        if ((state.reord < tp->fackets_out) &&
1918            ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1919            (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1920                tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1921
1922out:
1923
1924#if FASTRETRANS_DEBUG > 0
1925        WARN_ON((int)tp->sacked_out < 0);
1926        WARN_ON((int)tp->lost_out < 0);
1927        WARN_ON((int)tp->retrans_out < 0);
1928        WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1929#endif
1930        return state.flag;
1931}
1932
1933/* Limits sacked_out so that sum with lost_out isn't ever larger than
1934 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1935 */
1936static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1937{
1938        u32 holes;
1939
1940        holes = max(tp->lost_out, 1U);
1941        holes = min(holes, tp->packets_out);
1942
1943        if ((tp->sacked_out + holes) > tp->packets_out) {
1944                tp->sacked_out = tp->packets_out - holes;
1945                return 1;
1946        }
1947        return 0;
1948}
1949
1950/* If we receive more dupacks than we expected counting segments
1951 * in assumption of absent reordering, interpret this as reordering.
1952 * The only another reason could be bug in receiver TCP.
1953 */
1954static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1955{
1956        struct tcp_sock *tp = tcp_sk(sk);
1957        if (tcp_limit_reno_sacked(tp))
1958                tcp_update_reordering(sk, tp->packets_out + addend, 0);
1959}
1960
1961/* Emulate SACKs for SACKless connection: account for a new dupack. */
1962
1963static void tcp_add_reno_sack(struct sock *sk)
1964{
1965        struct tcp_sock *tp = tcp_sk(sk);
1966        tp->sacked_out++;
1967        tcp_check_reno_reordering(sk, 0);
1968        tcp_verify_left_out(tp);
1969}
1970
1971/* Account for ACK, ACKing some data in Reno Recovery phase. */
1972
1973static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1974{
1975        struct tcp_sock *tp = tcp_sk(sk);
1976
1977        if (acked > 0) {
1978                /* One ACK acked hole. The rest eat duplicate ACKs. */
1979                if (acked - 1 >= tp->sacked_out)
1980                        tp->sacked_out = 0;
1981                else
1982                        tp->sacked_out -= acked - 1;
1983        }
1984        tcp_check_reno_reordering(sk, acked);
1985        tcp_verify_left_out(tp);
1986}
1987
1988static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1989{
1990        tp->sacked_out = 0;
1991}
1992
1993static int tcp_is_sackfrto(const struct tcp_sock *tp)
1994{
1995        return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1996}
1997
1998/* F-RTO can only be used if TCP has never retransmitted anything other than
1999 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2000 */
2001int tcp_use_frto(struct sock *sk)
2002{
2003        const struct tcp_sock *tp = tcp_sk(sk);
2004        const struct inet_connection_sock *icsk = inet_csk(sk);
2005        struct sk_buff *skb;
2006
2007        if (!sysctl_tcp_frto)
2008                return 0;
2009
2010        /* MTU probe and F-RTO won't really play nicely along currently */
2011        if (icsk->icsk_mtup.probe_size)
2012                return 0;
2013
2014        if (tcp_is_sackfrto(tp))
2015                return 1;
2016
2017        /* Avoid expensive walking of rexmit queue if possible */
2018        if (tp->retrans_out > 1)
2019                return 0;
2020
2021        skb = tcp_write_queue_head(sk);
2022        if (tcp_skb_is_last(sk, skb))
2023                return 1;
2024        skb = tcp_write_queue_next(sk, skb);    /* Skips head */
2025        tcp_for_write_queue_from(skb, sk) {
2026                if (skb == tcp_send_head(sk))
2027                        break;
2028                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2029                        return 0;
2030                /* Short-circuit when first non-SACKed skb has been checked */
2031                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2032                        break;
2033        }
2034        return 1;
2035}
2036
2037/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2038 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2039 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2040 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2041 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2042 * bits are handled if the Loss state is really to be entered (in
2043 * tcp_enter_frto_loss).
2044 *
2045 * Do like tcp_enter_loss() would; when RTO expires the second time it
2046 * does:
2047 *  "Reduce ssthresh if it has not yet been made inside this window."
2048 */
2049void tcp_enter_frto(struct sock *sk)
2050{
2051        const struct inet_connection_sock *icsk = inet_csk(sk);
2052        struct tcp_sock *tp = tcp_sk(sk);
2053        struct sk_buff *skb;
2054
2055        if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2056            tp->snd_una == tp->high_seq ||
2057            ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2058             !icsk->icsk_retransmits)) {
2059                tp->prior_ssthresh = tcp_current_ssthresh(sk);
2060                /* Our state is too optimistic in ssthresh() call because cwnd
2061                 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2062                 * recovery has not yet completed. Pattern would be this: RTO,
2063                 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2064                 * up here twice).
2065                 * RFC4138 should be more specific on what to do, even though
2066                 * RTO is quite unlikely to occur after the first Cumulative ACK
2067                 * due to back-off and complexity of triggering events ...
2068                 */
2069                if (tp->frto_counter) {
2070                        u32 stored_cwnd;
2071                        stored_cwnd = tp->snd_cwnd;
2072                        tp->snd_cwnd = 2;
2073                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2074                        tp->snd_cwnd = stored_cwnd;
2075                } else {
2076                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2077                }
2078                /* ... in theory, cong.control module could do "any tricks" in
2079                 * ssthresh(), which means that ca_state, lost bits and lost_out
2080                 * counter would have to be faked before the call occurs. We
2081                 * consider that too expensive, unlikely and hacky, so modules
2082                 * using these in ssthresh() must deal these incompatibility
2083                 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2084                 */
2085                tcp_ca_event(sk, CA_EVENT_FRTO);
2086        }
2087
2088        tp->undo_marker = tp->snd_una;
2089        tp->undo_retrans = 0;
2090
2091        skb = tcp_write_queue_head(sk);
2092        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2093                tp->undo_marker = 0;
2094        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2095                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2096                tp->retrans_out -= tcp_skb_pcount(skb);
2097        }
2098        tcp_verify_left_out(tp);
2099
2100        /* Too bad if TCP was application limited */
2101        tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2102
2103        /* Earlier loss recovery underway (see RFC4138; Appendix B).
2104         * The last condition is necessary at least in tp->frto_counter case.
2105         */
2106        if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2107            ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2108            after(tp->high_seq, tp->snd_una)) {
2109                tp->frto_highmark = tp->high_seq;
2110        } else {
2111                tp->frto_highmark = tp->snd_nxt;
2112        }
2113        tcp_set_ca_state(sk, TCP_CA_Disorder);
2114        tp->high_seq = tp->snd_nxt;
2115        tp->frto_counter = 1;
2116}
2117
2118/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2119 * which indicates that we should follow the traditional RTO recovery,
2120 * i.e. mark everything lost and do go-back-N retransmission.
2121 */
2122static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2123{
2124        struct tcp_sock *tp = tcp_sk(sk);
2125        struct sk_buff *skb;
2126
2127        tp->lost_out = 0;
2128        tp->retrans_out = 0;
2129        if (tcp_is_reno(tp))
2130                tcp_reset_reno_sack(tp);
2131
2132        tcp_for_write_queue(skb, sk) {
2133                if (skb == tcp_send_head(sk))
2134                        break;
2135
2136                TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2137                /*
2138                 * Count the retransmission made on RTO correctly (only when
2139                 * waiting for the first ACK and did not get it)...
2140                 */
2141                if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2142                        /* For some reason this R-bit might get cleared? */
2143                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2144                                tp->retrans_out += tcp_skb_pcount(skb);
2145                        /* ...enter this if branch just for the first segment */
2146                        flag |= FLAG_DATA_ACKED;
2147                } else {
2148                        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2149                                tp->undo_marker = 0;
2150                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2151                }
2152
2153                /* Marking forward transmissions that were made after RTO lost
2154                 * can cause unnecessary retransmissions in some scenarios,
2155                 * SACK blocks will mitigate that in some but not in all cases.
2156                 * We used to not mark them but it was causing break-ups with
2157                 * receivers that do only in-order receival.
2158                 *
2159                 * TODO: we could detect presence of such receiver and select
2160                 * different behavior per flow.
2161                 */
2162                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2163                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2164                        tp->lost_out += tcp_skb_pcount(skb);
2165                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2166                }
2167        }
2168        tcp_verify_left_out(tp);
2169
2170        tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2171        tp->snd_cwnd_cnt = 0;
2172        tp->snd_cwnd_stamp = tcp_time_stamp;
2173        tp->frto_counter = 0;
2174        tp->bytes_acked = 0;
2175
2176        tp->reordering = min_t(unsigned int, tp->reordering,
2177                               sysctl_tcp_reordering);
2178        tcp_set_ca_state(sk, TCP_CA_Loss);
2179        tp->high_seq = tp->snd_nxt;
2180        TCP_ECN_queue_cwr(tp);
2181
2182        tcp_clear_all_retrans_hints(tp);
2183}
2184
2185static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2186{
2187        tp->retrans_out = 0;
2188        tp->lost_out = 0;
2189
2190        tp->undo_marker = 0;
2191        tp->undo_retrans = 0;
2192}
2193
2194void tcp_clear_retrans(struct tcp_sock *tp)
2195{
2196        tcp_clear_retrans_partial(tp);
2197
2198        tp->fackets_out = 0;
2199        tp->sacked_out = 0;
2200}
2201
2202/* Enter Loss state. If "how" is not zero, forget all SACK information
2203 * and reset tags completely, otherwise preserve SACKs. If receiver
2204 * dropped its ofo queue, we will know this due to reneging detection.
2205 */
2206void tcp_enter_loss(struct sock *sk, int how)
2207{
2208        const struct inet_connection_sock *icsk = inet_csk(sk);
2209        struct tcp_sock *tp = tcp_sk(sk);
2210        struct sk_buff *skb;
2211
2212        /* Reduce ssthresh if it has not yet been made inside this window. */
2213        if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2214            (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2215                tp->prior_ssthresh = tcp_current_ssthresh(sk);
2216                tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2217                tcp_ca_event(sk, CA_EVENT_LOSS);
2218        }
2219        tp->snd_cwnd       = 1;
2220        tp->snd_cwnd_cnt   = 0;
2221        tp->snd_cwnd_stamp = tcp_time_stamp;
2222
2223        tp->bytes_acked = 0;
2224        tcp_clear_retrans_partial(tp);
2225
2226        if (tcp_is_reno(tp))
2227                tcp_reset_reno_sack(tp);
2228
2229        if (!how) {
2230                /* Push undo marker, if it was plain RTO and nothing
2231                 * was retransmitted. */
2232                tp->undo_marker = tp->snd_una;
2233        } else {
2234                tp->sacked_out = 0;
2235                tp->fackets_out = 0;
2236        }
2237        tcp_clear_all_retrans_hints(tp);
2238
2239        tcp_for_write_queue(skb, sk) {
2240                if (skb == tcp_send_head(sk))
2241                        break;
2242
2243                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2244                        tp->undo_marker = 0;
2245                TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2246                if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2247                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2248                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2249                        tp->lost_out += tcp_skb_pcount(skb);
2250                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2251                }
2252        }
2253        tcp_verify_left_out(tp);
2254
2255        tp->reordering = min_t(unsigned int, tp->reordering,
2256                               sysctl_tcp_reordering);
2257        tcp_set_ca_state(sk, TCP_CA_Loss);
2258        tp->high_seq = tp->snd_nxt;
2259        TCP_ECN_queue_cwr(tp);
2260        /* Abort F-RTO algorithm if one is in progress */
2261        tp->frto_counter = 0;
2262}
2263
2264/* If ACK arrived pointing to a remembered SACK, it means that our
2265 * remembered SACKs do not reflect real state of receiver i.e.
2266 * receiver _host_ is heavily congested (or buggy).
2267 *
2268 * Do processing similar to RTO timeout.
2269 */
2270static int tcp_check_sack_reneging(struct sock *sk, int flag)
2271{
2272        if (flag & FLAG_SACK_RENEGING) {
2273                struct inet_connection_sock *icsk = inet_csk(sk);
2274                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2275
2276                tcp_enter_loss(sk, 1);
2277                icsk->icsk_retransmits++;
2278                tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2279                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2280                                          icsk->icsk_rto, TCP_RTO_MAX);
2281                return 1;
2282        }
2283        return 0;
2284}
2285
2286static inline int tcp_fackets_out(struct tcp_sock *tp)
2287{
2288        return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2289}
2290
2291/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2292 * counter when SACK is enabled (without SACK, sacked_out is used for
2293 * that purpose).
2294 *
2295 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2296 * segments up to the highest received SACK block so far and holes in
2297 * between them.
2298 *
2299 * With reordering, holes may still be in flight, so RFC3517 recovery
2300 * uses pure sacked_out (total number of SACKed segments) even though
2301 * it violates the RFC that uses duplicate ACKs, often these are equal
2302 * but when e.g. out-of-window ACKs or packet duplication occurs,
2303 * they differ. Since neither occurs due to loss, TCP should really
2304 * ignore them.
2305 */
2306static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2307{
2308        return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2309}
2310
2311static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2312{
2313        return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2314}
2315
2316static inline int tcp_head_timedout(struct sock *sk)
2317{
2318        struct tcp_sock *tp = tcp_sk(sk);
2319
2320        return tp->packets_out &&
2321               tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2322}
2323
2324/* Linux NewReno/SACK/FACK/ECN state machine.
2325 * --------------------------------------
2326 *
2327 * "Open"       Normal state, no dubious events, fast path.
2328 * "Disorder"   In all the respects it is "Open",
2329 *              but requires a bit more attention. It is entered when
2330 *              we see some SACKs or dupacks. It is split of "Open"
2331 *              mainly to move some processing from fast path to slow one.
2332 * "CWR"        CWND was reduced due to some Congestion Notification event.
2333 *              It can be ECN, ICMP source quench, local device congestion.
2334 * "Recovery"   CWND was reduced, we are fast-retransmitting.
2335 * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2336 *
2337 * tcp_fastretrans_alert() is entered:
2338 * - each incoming ACK, if state is not "Open"
2339 * - when arrived ACK is unusual, namely:
2340 *      * SACK
2341 *      * Duplicate ACK.
2342 *      * ECN ECE.
2343 *
2344 * Counting packets in flight is pretty simple.
2345 *
2346 *      in_flight = packets_out - left_out + retrans_out
2347 *
2348 *      packets_out is SND.NXT-SND.UNA counted in packets.
2349 *
2350 *      retrans_out is number of retransmitted segments.
2351 *
2352 *      left_out is number of segments left network, but not ACKed yet.
2353 *
2354 *              left_out = sacked_out + lost_out
2355 *
2356 *     sacked_out: Packets, which arrived to receiver out of order
2357 *                 and hence not ACKed. With SACKs this number is simply
2358 *                 amount of SACKed data. Even without SACKs
2359 *                 it is easy to give pretty reliable estimate of this number,
2360 *                 counting duplicate ACKs.
2361 *
2362 *       lost_out: Packets lost by network. TCP has no explicit
2363 *                 "loss notification" feedback from network (for now).
2364 *                 It means that this number can be only _guessed_.
2365 *                 Actually, it is the heuristics to predict lossage that
2366 *                 distinguishes different algorithms.
2367 *
2368 *      F.e. after RTO, when all the queue is considered as lost,
2369 *      lost_out = packets_out and in_flight = retrans_out.
2370 *
2371 *              Essentially, we have now two algorithms counting
2372 *              lost packets.
2373 *
2374 *              FACK: It is the simplest heuristics. As soon as we decided
2375 *              that something is lost, we decide that _all_ not SACKed
2376 *              packets until the most forward SACK are lost. I.e.
2377 *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2378 *              It is absolutely correct estimate, if network does not reorder
2379 *              packets. And it loses any connection to reality when reordering
2380 *              takes place. We use FACK by default until reordering
2381 *              is suspected on the path to this destination.
2382 *
2383 *              NewReno: when Recovery is entered, we assume that one segment
2384 *              is lost (classic Reno). While we are in Recovery and
2385 *              a partial ACK arrives, we assume that one more packet
2386 *              is lost (NewReno). This heuristics are the same in NewReno
2387 *              and SACK.
2388 *
2389 *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2390 *  deflation etc. CWND is real congestion window, never inflated, changes
2391 *  only according to classic VJ rules.
2392 *
2393 * Really tricky (and requiring careful tuning) part of algorithm
2394 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2395 * The first determines the moment _when_ we should reduce CWND and,
2396 * hence, slow down forward transmission. In fact, it determines the moment
2397 * when we decide that hole is caused by loss, rather than by a reorder.
2398 *
2399 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2400 * holes, caused by lost packets.
2401 *
2402 * And the most logically complicated part of algorithm is undo
2403 * heuristics. We detect false retransmits due to both too early
2404 * fast retransmit (reordering) and underestimated RTO, analyzing
2405 * timestamps and D-SACKs. When we detect that some segments were
2406 * retransmitted by mistake and CWND reduction was wrong, we undo
2407 * window reduction and abort recovery phase. This logic is hidden
2408 * inside several functions named tcp_try_undo_<something>.
2409 */
2410
2411/* This function decides, when we should leave Disordered state
2412 * and enter Recovery phase, reducing congestion window.
2413 *
2414 * Main question: may we further continue forward transmission
2415 * with the same cwnd?
2416 */
2417static int tcp_time_to_recover(struct sock *sk)
2418{
2419        struct tcp_sock *tp = tcp_sk(sk);
2420        __u32 packets_out;
2421
2422        /* Do not perform any recovery during F-RTO algorithm */
2423        if (tp->frto_counter)
2424                return 0;
2425
2426        /* Trick#1: The loss is proven. */
2427        if (tp->lost_out)
2428                return 1;
2429
2430        /* Not-A-Trick#2 : Classic rule... */
2431        if (tcp_dupack_heuristics(tp) > tp->reordering)
2432                return 1;
2433
2434        /* Trick#3 : when we use RFC2988 timer restart, fast
2435         * retransmit can be triggered by timeout of queue head.
2436         */
2437        if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2438                return 1;
2439
2440        /* Trick#4: It is still not OK... But will it be useful to delay
2441         * recovery more?
2442         */
2443        packets_out = tp->packets_out;
2444        if (packets_out <= tp->reordering &&
2445            tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2446            !tcp_may_send_now(sk)) {
2447                /* We have nothing to send. This connection is limited
2448                 * either by receiver window or by application.
2449                 */
2450                return 1;
2451        }
2452
2453        /* If a thin stream is detected, retransmit after first
2454         * received dupack. Employ only if SACK is supported in order
2455         * to avoid possible corner-case series of spurious retransmissions
2456         * Use only if there are no unsent data.
2457         */
2458        if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2459            tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2460            tcp_is_sack(tp) && !tcp_send_head(sk))
2461                return 1;
2462
2463        return 0;
2464}
2465
2466/* New heuristics: it is possible only after we switched to restart timer
2467 * each time when something is ACKed. Hence, we can detect timed out packets
2468 * during fast retransmit without falling to slow start.
2469 *
2470 * Usefulness of this as is very questionable, since we should know which of
2471 * the segments is the next to timeout which is relatively expensive to find
2472 * in general case unless we add some data structure just for that. The
2473 * current approach certainly won't find the right one too often and when it
2474 * finally does find _something_ it usually marks large part of the window
2475 * right away (because a retransmission with a larger timestamp blocks the
2476 * loop from advancing). -ij
2477 */
2478static void tcp_timeout_skbs(struct sock *sk)
2479{
2480        struct tcp_sock *tp = tcp_sk(sk);
2481        struct sk_buff *skb;
2482
2483        if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2484                return;
2485
2486        skb = tp->scoreboard_skb_hint;
2487        if (tp->scoreboard_skb_hint == NULL)
2488                skb = tcp_write_queue_head(sk);
2489
2490        tcp_for_write_queue_from(skb, sk) {
2491                if (skb == tcp_send_head(sk))
2492                        break;
2493                if (!tcp_skb_timedout(sk, skb))
2494                        break;
2495
2496                tcp_skb_mark_lost(tp, skb);
2497        }
2498
2499        tp->scoreboard_skb_hint = skb;
2500
2501        tcp_verify_left_out(tp);
2502}
2503
2504/* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2505 * is against sacked "cnt", otherwise it's against facked "cnt"
2506 */
2507static void tcp_mark_head_lost(struct sock *sk, int packets)
2508{
2509        struct tcp_sock *tp = tcp_sk(sk);
2510        struct sk_buff *skb;
2511        int cnt, oldcnt;
2512        int err;
2513        unsigned int mss;
2514
2515        if (packets == 0)
2516                return;
2517
2518        WARN_ON(packets > tp->packets_out);
2519        if (tp->lost_skb_hint) {
2520                skb = tp->lost_skb_hint;
2521                cnt = tp->lost_cnt_hint;
2522        } else {
2523                skb = tcp_write_queue_head(sk);
2524                cnt = 0;
2525        }
2526
2527        tcp_for_write_queue_from(skb, sk) {
2528                if (skb == tcp_send_head(sk))
2529                        break;
2530                /* TODO: do this better */
2531                /* this is not the most efficient way to do this... */
2532                tp->lost_skb_hint = skb;
2533                tp->lost_cnt_hint = cnt;
2534
2535                if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2536                        break;
2537
2538                oldcnt = cnt;
2539                if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2540                    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2541                        cnt += tcp_skb_pcount(skb);
2542
2543                if (cnt > packets) {
2544                        if (tcp_is_sack(tp) || (oldcnt >= packets))
2545                                break;
2546
2547                        mss = skb_shinfo(skb)->gso_size;
2548                        err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2549                        if (err < 0)
2550                                break;
2551                        cnt = packets;
2552                }
2553
2554                tcp_skb_mark_lost(tp, skb);
2555        }
2556        tcp_verify_left_out(tp);
2557}
2558
2559/* Account newly detected lost packet(s) */
2560
2561static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2562{
2563        struct tcp_sock *tp = tcp_sk(sk);
2564
2565        if (tcp_is_reno(tp)) {
2566                tcp_mark_head_lost(sk, 1);
2567        } else if (tcp_is_fack(tp)) {
2568                int lost = tp->fackets_out - tp->reordering;
2569                if (lost <= 0)
2570                        lost = 1;
2571                tcp_mark_head_lost(sk, lost);
2572        } else {
2573                int sacked_upto = tp->sacked_out - tp->reordering;
2574                if (sacked_upto < fast_rexmit)
2575                        sacked_upto = fast_rexmit;
2576                tcp_mark_head_lost(sk, sacked_upto);
2577        }
2578
2579        tcp_timeout_skbs(sk);
2580}
2581
2582/* CWND moderation, preventing bursts due to too big ACKs
2583 * in dubious situations.
2584 */
2585static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2586{
2587        tp->snd_cwnd = min(tp->snd_cwnd,
2588                           tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2589        tp->snd_cwnd_stamp = tcp_time_stamp;
2590}
2591
2592/* Lower bound on congestion window is slow start threshold
2593 * unless congestion avoidance choice decides to overide it.
2594 */
2595static inline u32 tcp_cwnd_min(const struct sock *sk)
2596{
2597        const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2598
2599        return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2600}
2601
2602/* Decrease cwnd each second ack. */
2603static void tcp_cwnd_down(struct sock *sk, int flag)
2604{
2605        struct tcp_sock *tp = tcp_sk(sk);
2606        int decr = tp->snd_cwnd_cnt + 1;
2607
2608        if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2609            (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2610                tp->snd_cwnd_cnt = decr & 1;
2611                decr >>= 1;
2612
2613                if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2614                        tp->snd_cwnd -= decr;
2615
2616                tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2617                tp->snd_cwnd_stamp = tcp_time_stamp;
2618        }
2619}
2620
2621/* Nothing was retransmitted or returned timestamp is less
2622 * than timestamp of the first retransmission.
2623 */
2624static inline int tcp_packet_delayed(struct tcp_sock *tp)
2625{
2626        return !tp->retrans_stamp ||
2627                (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2628                 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2629}
2630
2631/* Undo procedures. */
2632
2633#if FASTRETRANS_DEBUG > 1
2634static void DBGUNDO(struct sock *sk, const char *msg)
2635{
2636        struct tcp_sock *tp = tcp_sk(sk);
2637        struct inet_sock *inet = inet_sk(sk);
2638
2639        if (sk->sk_family == AF_INET) {
2640                printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2641                       msg,
2642                       &inet->inet_daddr, ntohs(inet->inet_dport),
2643                       tp->snd_cwnd, tcp_left_out(tp),
2644                       tp->snd_ssthresh, tp->prior_ssthresh,
2645                       tp->packets_out);
2646        }
2647#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2648        else if (sk->sk_family == AF_INET6) {
2649                struct ipv6_pinfo *np = inet6_sk(sk);
2650                printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2651                       msg,
2652                       &np->daddr, ntohs(inet->inet_dport),
2653                       tp->snd_cwnd, tcp_left_out(tp),
2654                       tp->snd_ssthresh, tp->prior_ssthresh,
2655                       tp->packets_out);
2656        }
2657#endif
2658}
2659#else
2660#define DBGUNDO(x...) do { } while (0)
2661#endif
2662
2663static void tcp_undo_cwr(struct sock *sk, const int undo)
2664{
2665        struct tcp_sock *tp = tcp_sk(sk);
2666
2667        if (tp->prior_ssthresh) {
2668                const struct inet_connection_sock *icsk = inet_csk(sk);
2669
2670                if (icsk->icsk_ca_ops->undo_cwnd)
2671                        tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2672                else
2673                        tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2674
2675                if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2676                        tp->snd_ssthresh = tp->prior_ssthresh;
2677                        TCP_ECN_withdraw_cwr(tp);
2678                }
2679        } else {
2680                tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2681        }
2682        tcp_moderate_cwnd(tp);
2683        tp->snd_cwnd_stamp = tcp_time_stamp;
2684}
2685
2686static inline int tcp_may_undo(struct tcp_sock *tp)
2687{
2688        return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2689}
2690
2691/* People celebrate: "We love our President!" */
2692static int tcp_try_undo_recovery(struct sock *sk)
2693{
2694        struct tcp_sock *tp = tcp_sk(sk);
2695
2696        if (tcp_may_undo(tp)) {
2697                int mib_idx;
2698
2699                /* Happy end! We did not retransmit anything
2700                 * or our original transmission succeeded.
2701                 */
2702                DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2703                tcp_undo_cwr(sk, 1);
2704                if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2705                        mib_idx = LINUX_MIB_TCPLOSSUNDO;
2706                else
2707                        mib_idx = LINUX_MIB_TCPFULLUNDO;
2708
2709                NET_INC_STATS_BH(sock_net(sk), mib_idx);
2710                tp->undo_marker = 0;
2711        }
2712        if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2713                /* Hold old state until something *above* high_seq
2714                 * is ACKed. For Reno it is MUST to prevent false
2715                 * fast retransmits (RFC2582). SACK TCP is safe. */
2716                tcp_moderate_cwnd(tp);
2717                return 1;
2718        }
2719        tcp_set_ca_state(sk, TCP_CA_Open);
2720        return 0;
2721}
2722
2723/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2724static void tcp_try_undo_dsack(struct sock *sk)
2725{
2726        struct tcp_sock *tp = tcp_sk(sk);
2727
2728        if (tp->undo_marker && !tp->undo_retrans) {
2729                DBGUNDO(sk, "D-SACK");
2730                tcp_undo_cwr(sk, 1);
2731                tp->undo_marker = 0;
2732                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2733        }
2734}
2735
2736/* We can clear retrans_stamp when there are no retransmissions in the
2737 * window. It would seem that it is trivially available for us in
2738 * tp->retrans_out, however, that kind of assumptions doesn't consider
2739 * what will happen if errors occur when sending retransmission for the
2740 * second time. ...It could the that such segment has only
2741 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2742 * the head skb is enough except for some reneging corner cases that
2743 * are not worth the effort.
2744 *
2745 * Main reason for all this complexity is the fact that connection dying
2746 * time now depends on the validity of the retrans_stamp, in particular,
2747 * that successive retransmissions of a segment must not advance
2748 * retrans_stamp under any conditions.
2749 */
2750static int tcp_any_retrans_done(struct sock *sk)
2751{
2752        struct tcp_sock *tp = tcp_sk(sk);
2753        struct sk_buff *skb;
2754
2755        if (tp->retrans_out)
2756                return 1;
2757
2758        skb = tcp_write_queue_head(sk);
2759        if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2760                return 1;
2761
2762        return 0;
2763}
2764
2765/* Undo during fast recovery after partial ACK. */
2766
2767static int tcp_try_undo_partial(struct sock *sk, int acked)
2768{
2769        struct tcp_sock *tp = tcp_sk(sk);
2770        /* Partial ACK arrived. Force Hoe's retransmit. */
2771        int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2772
2773        if (tcp_may_undo(tp)) {
2774                /* Plain luck! Hole if filled with delayed
2775                 * packet, rather than with a retransmit.
2776                 */
2777                if (!tcp_any_retrans_done(sk))
2778                        tp->retrans_stamp = 0;
2779
2780                tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2781
2782                DBGUNDO(sk, "Hoe");
2783                tcp_undo_cwr(sk, 0);
2784                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2785
2786                /* So... Do not make Hoe's retransmit yet.
2787                 * If the first packet was delayed, the rest
2788                 * ones are most probably delayed as well.
2789                 */
2790                failed = 0;
2791        }
2792        return failed;
2793}
2794
2795/* Undo during loss recovery after partial ACK. */
2796static int tcp_try_undo_loss(struct sock *sk)
2797{
2798        struct tcp_sock *tp = tcp_sk(sk);
2799
2800        if (tcp_may_undo(tp)) {
2801                struct sk_buff *skb;
2802                tcp_for_write_queue(skb, sk) {
2803                        if (skb == tcp_send_head(sk))
2804                                break;
2805                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2806                }
2807
2808                tcp_clear_all_retrans_hints(tp);
2809
2810                DBGUNDO(sk, "partial loss");
2811                tp->lost_out = 0;
2812                tcp_undo_cwr(sk, 1);
2813                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2814                inet_csk(sk)->icsk_retransmits = 0;
2815                tp->undo_marker = 0;
2816                if (tcp_is_sack(tp))
2817                        tcp_set_ca_state(sk, TCP_CA_Open);
2818                return 1;
2819        }
2820        return 0;
2821}
2822
2823static inline void tcp_complete_cwr(struct sock *sk)
2824{
2825        struct tcp_sock *tp = tcp_sk(sk);
2826        tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2827        tp->snd_cwnd_stamp = tcp_time_stamp;
2828        tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2829}
2830
2831static void tcp_try_keep_open(struct sock *sk)
2832{
2833        struct tcp_sock *tp = tcp_sk(sk);
2834        int state = TCP_CA_Open;
2835
2836        if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2837                state = TCP_CA_Disorder;
2838
2839        if (inet_csk(sk)->icsk_ca_state != state) {
2840                tcp_set_ca_state(sk, state);
2841                tp->high_seq = tp->snd_nxt;
2842        }
2843}
2844
2845static void tcp_try_to_open(struct sock *sk, int flag)
2846{
2847        struct tcp_sock *tp = tcp_sk(sk);
2848
2849        tcp_verify_left_out(tp);
2850
2851        if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2852                tp->retrans_stamp = 0;
2853
2854        if (flag & FLAG_ECE)
2855                tcp_enter_cwr(sk, 1);
2856
2857        if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2858                tcp_try_keep_open(sk);
2859                tcp_moderate_cwnd(tp);
2860        } else {
2861                tcp_cwnd_down(sk, flag);
2862        }
2863}
2864
2865static void tcp_mtup_probe_failed(struct sock *sk)
2866{
2867        struct inet_connection_sock *icsk = inet_csk(sk);
2868
2869        icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2870        icsk->icsk_mtup.probe_size = 0;
2871}
2872
2873static void tcp_mtup_probe_success(struct sock *sk)
2874{
2875        struct tcp_sock *tp = tcp_sk(sk);
2876        struct inet_connection_sock *icsk = inet_csk(sk);
2877
2878        /* FIXME: breaks with very large cwnd */
2879        tp->prior_ssthresh = tcp_current_ssthresh(sk);
2880        tp->snd_cwnd = tp->snd_cwnd *
2881                       tcp_mss_to_mtu(sk, tp->mss_cache) /
2882                       icsk->icsk_mtup.probe_size;
2883        tp->snd_cwnd_cnt = 0;
2884        tp->snd_cwnd_stamp = tcp_time_stamp;
2885        tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2886
2887        icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2888        icsk->icsk_mtup.probe_size = 0;
2889        tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2890}
2891
2892/* Do a simple retransmit without using the backoff mechanisms in
2893 * tcp_timer. This is used for path mtu discovery.
2894 * The socket is already locked here.
2895 */
2896void tcp_simple_retransmit(struct sock *sk)
2897{
2898        const struct inet_connection_sock *icsk = inet_csk(sk);
2899        struct tcp_sock *tp = tcp_sk(sk);
2900        struct sk_buff *skb;
2901        unsigned int mss = tcp_current_mss(sk);
2902        u32 prior_lost = tp->lost_out;
2903
2904        tcp_for_write_queue(skb, sk) {
2905                if (skb == tcp_send_head(sk))
2906                        break;
2907                if (tcp_skb_seglen(skb) > mss &&
2908                    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2909                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2910                                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2911                                tp->retrans_out -= tcp_skb_pcount(skb);
2912                        }
2913                        tcp_skb_mark_lost_uncond_verify(tp, skb);
2914                }
2915        }
2916
2917        tcp_clear_retrans_hints_partial(tp);
2918
2919        if (prior_lost == tp->lost_out)
2920                return;
2921
2922        if (tcp_is_reno(tp))
2923                tcp_limit_reno_sacked(tp);
2924
2925        tcp_verify_left_out(tp);
2926
2927        /* Don't muck with the congestion window here.
2928         * Reason is that we do not increase amount of _data_
2929         * in network, but units changed and effective
2930         * cwnd/ssthresh really reduced now.
2931         */
2932        if (icsk->icsk_ca_state != TCP_CA_Loss) {
2933                tp->high_seq = tp->snd_nxt;
2934                tp->snd_ssthresh = tcp_current_ssthresh(sk);
2935                tp->prior_ssthresh = 0;
2936                tp->undo_marker = 0;
2937                tcp_set_ca_state(sk, TCP_CA_Loss);
2938        }
2939        tcp_xmit_retransmit_queue(sk);
2940}
2941
2942/* Process an event, which can update packets-in-flight not trivially.
2943 * Main goal of this function is to calculate new estimate for left_out,
2944 * taking into account both packets sitting in receiver's buffer and
2945 * packets lost by network.
2946 *
2947 * Besides that it does CWND reduction, when packet loss is detected
2948 * and changes state of machine.
2949 *
2950 * It does _not_ decide what to send, it is made in function
2951 * tcp_xmit_retransmit_queue().
2952 */
2953static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2954{
2955        struct inet_connection_sock *icsk = inet_csk(sk);
2956        struct tcp_sock *tp = tcp_sk(sk);
2957        int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2958        int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2959                                    (tcp_fackets_out(tp) > tp->reordering));
2960        int fast_rexmit = 0, mib_idx;
2961
2962        if (WARN_ON(!tp->packets_out && tp->sacked_out))
2963                tp->sacked_out = 0;
2964        if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2965                tp->fackets_out = 0;
2966
2967        /* Now state machine starts.
2968         * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2969        if (flag & FLAG_ECE)
2970                tp->prior_ssthresh = 0;
2971
2972        /* B. In all the states check for reneging SACKs. */
2973        if (tcp_check_sack_reneging(sk, flag))
2974                return;
2975
2976        /* C. Process data loss notification, provided it is valid. */
2977        if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2978            before(tp->snd_una, tp->high_seq) &&
2979            icsk->icsk_ca_state != TCP_CA_Open &&
2980            tp->fackets_out > tp->reordering) {
2981                tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2982                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2983        }
2984
2985        /* D. Check consistency of the current state. */
2986        tcp_verify_left_out(tp);
2987
2988        /* E. Check state exit conditions. State can be terminated
2989         *    when high_seq is ACKed. */
2990        if (icsk->icsk_ca_state == TCP_CA_Open) {
2991                WARN_ON(tp->retrans_out != 0);
2992                tp->retrans_stamp = 0;
2993        } else if (!before(tp->snd_una, tp->high_seq)) {
2994                switch (icsk->icsk_ca_state) {
2995                case TCP_CA_Loss:
2996                        icsk->icsk_retransmits = 0;
2997                        if (tcp_try_undo_recovery(sk))
2998                                return;
2999                        break;
3000
3001                case TCP_CA_CWR:
3002                        /* CWR is to be held something *above* high_seq
3003                         * is ACKed for CWR bit to reach receiver. */
3004                        if (tp->snd_una != tp->high_seq) {
3005                                tcp_complete_cwr(sk);
3006                                tcp_set_ca_state(sk, TCP_CA_Open);
3007                        }
3008                        break;
3009
3010                case TCP_CA_Disorder:
3011                        tcp_try_undo_dsack(sk);
3012                        if (!tp->undo_marker ||
3013                            /* For SACK case do not Open to allow to undo
3014                             * catching for all duplicate ACKs. */
3015                            tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3016                                tp->undo_marker = 0;
3017                                tcp_set_ca_state(sk, TCP_CA_Open);
3018                        }
3019                        break;
3020
3021                case TCP_CA_Recovery:
3022                        if (tcp_is_reno(tp))
3023                                tcp_reset_reno_sack(tp);
3024                        if (tcp_try_undo_recovery(sk))
3025                                return;
3026                        tcp_complete_cwr(sk);
3027                        break;
3028                }
3029        }
3030
3031        /* F. Process state. */
3032        switch (icsk->icsk_ca_state) {
3033        case TCP_CA_Recovery:
3034                if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3035                        if (tcp_is_reno(tp) && is_dupack)
3036                                tcp_add_reno_sack(sk);
3037                } else
3038                        do_lost = tcp_try_undo_partial(sk, pkts_acked);
3039                break;
3040        case TCP_CA_Loss:
3041                if (flag & FLAG_DATA_ACKED)
3042                        icsk->icsk_retransmits = 0;
3043                if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3044                        tcp_reset_reno_sack(tp);
3045                if (!tcp_try_undo_loss(sk)) {
3046                        tcp_moderate_cwnd(tp);
3047                        tcp_xmit_retransmit_queue(sk);
3048                        return;
3049                }
3050                if (icsk->icsk_ca_state != TCP_CA_Open)
3051                        return;
3052                /* Loss is undone; fall through to processing in Open state. */
3053        default:
3054                if (tcp_is_reno(tp)) {
3055                        if (flag & FLAG_SND_UNA_ADVANCED)
3056                                tcp_reset_reno_sack(tp);
3057                        if (is_dupack)
3058                                tcp_add_reno_sack(sk);
3059                }
3060
3061                if (icsk->icsk_ca_state == TCP_CA_Disorder)
3062                        tcp_try_undo_dsack(sk);
3063
3064                if (!tcp_time_to_recover(sk)) {
3065                        tcp_try_to_open(sk, flag);
3066                        return;
3067                }
3068
3069                /* MTU probe failure: don't reduce cwnd */
3070                if (icsk->icsk_ca_state < TCP_CA_CWR &&
3071                    icsk->icsk_mtup.probe_size &&
3072                    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3073                        tcp_mtup_probe_failed(sk);
3074                        /* Restores the reduction we did in tcp_mtup_probe() */
3075                        tp->snd_cwnd++;
3076                        tcp_simple_retransmit(sk);
3077                        return;
3078                }
3079
3080                /* Otherwise enter Recovery state */
3081
3082                if (tcp_is_reno(tp))
3083                        mib_idx = LINUX_MIB_TCPRENORECOVERY;
3084                else
3085                        mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3086
3087                NET_INC_STATS_BH(sock_net(sk), mib_idx);
3088
3089                tp->high_seq = tp->snd_nxt;
3090                tp->prior_ssthresh = 0;
3091                tp->undo_marker = tp->snd_una;
3092                tp->undo_retrans = tp->retrans_out;
3093
3094                if (icsk->icsk_ca_state < TCP_CA_CWR) {
3095                        if (!(flag & FLAG_ECE))
3096                                tp->prior_ssthresh = tcp_current_ssthresh(sk);
3097                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3098                        TCP_ECN_queue_cwr(tp);
3099                }
3100
3101                tp->bytes_acked = 0;
3102                tp->snd_cwnd_cnt = 0;
3103                tcp_set_ca_state(sk, TCP_CA_Recovery);
3104                fast_rexmit = 1;
3105        }
3106
3107        if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3108                tcp_update_scoreboard(sk, fast_rexmit);
3109        tcp_cwnd_down(sk, flag);
3110        tcp_xmit_retransmit_queue(sk);
3111}
3112
3113static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3114{
3115        tcp_rtt_estimator(sk, seq_rtt);
3116        tcp_set_rto(sk);
3117        inet_csk(sk)->icsk_backoff = 0;
3118}
3119
3120/* Read draft-ietf-tcplw-high-performance before mucking
3121 * with this code. (Supersedes RFC1323)
3122 */
3123static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3124{
3125        /* RTTM Rule: A TSecr value received in a segment is used to
3126         * update the averaged RTT measurement only if the segment
3127         * acknowledges some new data, i.e., only if it advances the
3128         * left edge of the send window.
3129         *
3130         * See draft-ietf-tcplw-high-performance-00, section 3.3.
3131         * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3132         *
3133         * Changed: reset backoff as soon as we see the first valid sample.
3134         * If we do not, we get strongly overestimated rto. With timestamps
3135         * samples are accepted even from very old segments: f.e., when rtt=1
3136         * increases to 8, we retransmit 5 times and after 8 seconds delayed
3137         * answer arrives rto becomes 120 seconds! If at least one of segments
3138         * in window is lost... Voila.                          --ANK (010210)
3139         */
3140        struct tcp_sock *tp = tcp_sk(sk);
3141
3142        tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3143}
3144
3145static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3146{
3147        /* We don't have a timestamp. Can only use
3148         * packets that are not retransmitted to determine
3149         * rtt estimates. Also, we must not reset the
3150         * backoff for rto until we get a non-retransmitted
3151         * packet. This allows us to deal with a situation
3152         * where the network delay has increased suddenly.
3153         * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3154         */
3155
3156        if (flag & FLAG_RETRANS_DATA_ACKED)
3157                return;
3158
3159        tcp_valid_rtt_meas(sk, seq_rtt);
3160}
3161
3162static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3163                                      const s32 seq_rtt)
3164{
3165        const struct tcp_sock *tp = tcp_sk(sk);
3166        /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3167        if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3168                tcp_ack_saw_tstamp(sk, flag);
3169        else if (seq_rtt >= 0)
3170                tcp_ack_no_tstamp(sk, seq_rtt, flag);
3171}
3172
3173static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3174{
3175        const struct inet_connection_sock *icsk = inet_csk(sk);
3176        icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3177        tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3178}
3179
3180/* Restart timer after forward progress on connection.
3181 * RFC2988 recommends to restart timer to now+rto.
3182 */
3183static void tcp_rearm_rto(struct sock *sk)
3184{
3185        struct tcp_sock *tp = tcp_sk(sk);
3186
3187        if (!tp->packets_out) {
3188                inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3189        } else {
3190                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3191                                          inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3192        }
3193}
3194
3195/* If we get here, the whole TSO packet has not been acked. */
3196static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3197{
3198        struct tcp_sock *tp = tcp_sk(sk);
3199        u32 packets_acked;
3200
3201        BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3202
3203        packets_acked = tcp_skb_pcount(skb);
3204        if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3205                return 0;
3206        packets_acked -= tcp_skb_pcount(skb);
3207
3208        if (packets_acked) {
3209                BUG_ON(tcp_skb_pcount(skb) == 0);
3210                BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3211        }
3212
3213        return packets_acked;
3214}
3215
3216/* Remove acknowledged frames from the retransmission queue. If our packet
3217 * is before the ack sequence we can discard it as it's confirmed to have
3218 * arrived at the other end.
3219 */
3220static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3221                               u32 prior_snd_una)
3222{
3223        struct tcp_sock *tp = tcp_sk(sk);
3224        const struct inet_connection_sock *icsk = inet_csk(sk);
3225        struct sk_buff *skb;
3226        u32 now = tcp_time_stamp;
3227        int fully_acked = 1;
3228        int flag = 0;
3229        u32 pkts_acked = 0;
3230        u32 reord = tp->packets_out;
3231        u32 prior_sacked = tp->sacked_out;
3232        s32 seq_rtt = -1;
3233        s32 ca_seq_rtt = -1;
3234        ktime_t last_ackt = net_invalid_timestamp();
3235
3236        while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3237                struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3238                u32 acked_pcount;
3239                u8 sacked = scb->sacked;
3240
3241                /* Determine how many packets and what bytes were acked, tso and else */
3242                if (after(scb->end_seq, tp->snd_una)) {
3243                        if (tcp_skb_pcount(skb) == 1 ||
3244                            !after(tp->snd_una, scb->seq))
3245                                break;
3246
3247                        acked_pcount = tcp_tso_acked(sk, skb);
3248                        if (!acked_pcount)
3249                                break;
3250
3251                        fully_acked = 0;
3252                } else {
3253                        acked_pcount = tcp_skb_pcount(skb);
3254                }
3255
3256                if (sacked & TCPCB_RETRANS) {
3257                        if (sacked & TCPCB_SACKED_RETRANS)
3258                                tp->retrans_out -= acked_pcount;
3259                        flag |= FLAG_RETRANS_DATA_ACKED;
3260                        ca_seq_rtt = -1;
3261                        seq_rtt = -1;
3262                        if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3263                                flag |= FLAG_NONHEAD_RETRANS_ACKED;
3264                } else {
3265                        ca_seq_rtt = now - scb->when;
3266                        last_ackt = skb->tstamp;
3267                        if (seq_rtt < 0) {
3268                                seq_rtt = ca_seq_rtt;
3269                        }
3270                        if (!(sacked & TCPCB_SACKED_ACKED))
3271                                reord = min(pkts_acked, reord);
3272                }
3273
3274                if (sacked & TCPCB_SACKED_ACKED)
3275                        tp->sacked_out -= acked_pcount;
3276                if (sacked & TCPCB_LOST)
3277                        tp->lost_out -= acked_pcount;
3278
3279                tp->packets_out -= acked_pcount;
3280                pkts_acked += acked_pcount;
3281
3282                /* Initial outgoing SYN's get put onto the write_queue
3283                 * just like anything else we transmit.  It is not
3284                 * true data, and if we misinform our callers that
3285                 * this ACK acks real data, we will erroneously exit
3286                 * connection startup slow start one packet too
3287                 * quickly.  This is severely frowned upon behavior.
3288                 */
3289                if (!(scb->flags & TCPCB_FLAG_SYN)) {
3290                        flag |= FLAG_DATA_ACKED;
3291                } else {
3292                        flag |= FLAG_SYN_ACKED;
3293                        tp->retrans_stamp = 0;
3294                }
3295
3296                if (!fully_acked)
3297                        break;
3298
3299                tcp_unlink_write_queue(skb, sk);
3300                sk_wmem_free_skb(sk, skb);
3301                tp->scoreboard_skb_hint = NULL;
3302                if (skb == tp->retransmit_skb_hint)
3303                        tp->retransmit_skb_hint = NULL;
3304                if (skb == tp->lost_skb_hint)
3305                        tp->lost_skb_hint = NULL;
3306        }
3307
3308        if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3309                tp->snd_up = tp->snd_una;
3310
3311        if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3312                flag |= FLAG_SACK_RENEGING;
3313
3314        if (flag & FLAG_ACKED) {
3315                const struct tcp_congestion_ops *ca_ops
3316                        = inet_csk(sk)->icsk_ca_ops;
3317
3318                if (unlikely(icsk->icsk_mtup.probe_size &&
3319                             !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3320                        tcp_mtup_probe_success(sk);
3321                }
3322
3323                tcp_ack_update_rtt(sk, flag, seq_rtt);
3324                tcp_rearm_rto(sk);
3325
3326                if (tcp_is_reno(tp)) {
3327                        tcp_remove_reno_sacks(sk, pkts_acked);
3328                } else {
3329                        int delta;
3330
3331                        /* Non-retransmitted hole got filled? That's reordering */
3332                        if (reord < prior_fackets)
3333                                tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3334
3335                        delta = tcp_is_fack(tp) ? pkts_acked :
3336                                                  prior_sacked - tp->sacked_out;
3337                        tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3338                }
3339
3340                tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3341
3342                if (ca_ops->pkts_acked) {
3343                        s32 rtt_us = -1;
3344
3345                        /* Is the ACK triggering packet unambiguous? */
3346                        if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3347                                /* High resolution needed and available? */
3348                                if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3349                                    !ktime_equal(last_ackt,
3350                                                 net_invalid_timestamp()))
3351                                        rtt_us = ktime_us_delta(ktime_get_real(),
3352                                                                last_ackt);
3353                                else if (ca_seq_rtt > 0)
3354                                        rtt_us = jiffies_to_usecs(ca_seq_rtt);
3355                        }
3356
3357                        ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3358                }
3359        }
3360
3361#if FASTRETRANS_DEBUG > 0
3362        WARN_ON((int)tp->sacked_out < 0);
3363        WARN_ON((int)tp->lost_out < 0);
3364        WARN_ON((int)tp->retrans_out < 0);
3365        if (!tp->packets_out && tcp_is_sack(tp)) {
3366                icsk = inet_csk(sk);
3367                if (tp->lost_out) {
3368                        printk(KERN_DEBUG "Leak l=%u %d\n",
3369                               tp->lost_out, icsk->icsk_ca_state);
3370                        tp->lost_out = 0;
3371                }
3372                if (tp->sacked_out) {
3373                        printk(KERN_DEBUG "Leak s=%u %d\n",
3374                               tp->sacked_out, icsk->icsk_ca_state);
3375                        tp->sacked_out = 0;
3376                }
3377                if (tp->retrans_out) {
3378                        printk(KERN_DEBUG "Leak r=%u %d\n",
3379                               tp->retrans_out, icsk->icsk_ca_state);
3380                        tp->retrans_out = 0;
3381                }
3382        }
3383#endif
3384        return flag;
3385}
3386
3387static void tcp_ack_probe(struct sock *sk)
3388{
3389        const struct tcp_sock *tp = tcp_sk(sk);
3390        struct inet_connection_sock *icsk = inet_csk(sk);
3391
3392        /* Was it a usable window open? */
3393
3394        if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3395                icsk->icsk_backoff = 0;
3396                inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3397                /* Socket must be waked up by subsequent tcp_data_snd_check().
3398                 * This function is not for random using!
3399                 */
3400        } else {
3401                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3402                                          min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3403                                          TCP_RTO_MAX);
3404        }
3405}
3406
3407static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3408{
3409        return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3410                inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3411}
3412
3413static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3414{
3415        const struct tcp_sock *tp = tcp_sk(sk);
3416        return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3417                !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3418}
3419
3420/* Check that window update is acceptable.
3421 * The function assumes that snd_una<=ack<=snd_next.
3422 */
3423static inline int tcp_may_update_window(const struct tcp_sock *tp,
3424                                        const u32 ack, const u32 ack_seq,
3425                                        const u32 nwin)
3426{
3427        return (after(ack, tp->snd_una) ||
3428                after(ack_seq, tp->snd_wl1) ||
3429                (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3430}
3431
3432/* Update our send window.
3433 *
3434 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3435 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3436 */
3437static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3438                                 u32 ack_seq)
3439{
3440        struct tcp_sock *tp = tcp_sk(sk);
3441        int flag = 0;
3442        u32 nwin = ntohs(tcp_hdr(skb)->window);
3443
3444        if (likely(!tcp_hdr(skb)->syn))
3445                nwin <<= tp->rx_opt.snd_wscale;
3446
3447        if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3448                flag |= FLAG_WIN_UPDATE;
3449                tcp_update_wl(tp, ack_seq);
3450
3451                if (tp->snd_wnd != nwin) {
3452                        tp->snd_wnd = nwin;
3453
3454                        /* Note, it is the only place, where
3455                         * fast path is recovered for sending TCP.
3456                         */
3457                        tp->pred_flags = 0;
3458                        tcp_fast_path_check(sk);
3459
3460                        if (nwin > tp->max_window) {
3461                                tp->max_window = nwin;
3462                                tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3463                        }
3464                }
3465        }
3466
3467        tp->snd_una = ack;
3468
3469        return flag;
3470}
3471
3472/* A very conservative spurious RTO response algorithm: reduce cwnd and
3473 * continue in congestion avoidance.
3474 */
3475static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3476{
3477        tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3478        tp->snd_cwnd_cnt = 0;
3479        tp->bytes_acked = 0;
3480        TCP_ECN_queue_cwr(tp);
3481        tcp_moderate_cwnd(tp);
3482}
3483
3484/* A conservative spurious RTO response algorithm: reduce cwnd using
3485 * rate halving and continue in congestion avoidance.
3486 */
3487static void tcp_ratehalving_spur_to_response(struct sock *sk)
3488{
3489        tcp_enter_cwr(sk, 0);
3490}
3491
3492static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3493{
3494        if (flag & FLAG_ECE)
3495                tcp_ratehalving_spur_to_response(sk);
3496        else
3497                tcp_undo_cwr(sk, 1);
3498}
3499
3500/* F-RTO spurious RTO detection algorithm (RFC4138)
3501 *
3502 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3503 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3504 * window (but not to or beyond highest sequence sent before RTO):
3505 *   On First ACK,  send two new segments out.
3506 *   On Second ACK, RTO was likely spurious. Do spurious response (response
3507 *                  algorithm is not part of the F-RTO detection algorithm
3508 *                  given in RFC4138 but can be selected separately).
3509 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3510 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3511 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3512 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3513 *
3514 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3515 * original window even after we transmit two new data segments.
3516 *
3517 * SACK version:
3518 *   on first step, wait until first cumulative ACK arrives, then move to
3519 *   the second step. In second step, the next ACK decides.
3520 *
3521 * F-RTO is implemented (mainly) in four functions:
3522 *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3523 *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3524 *     called when tcp_use_frto() showed green light
3525 *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3526 *   - tcp_enter_frto_loss() is called if there is not enough evidence
3527 *     to prove that the RTO is indeed spurious. It transfers the control
3528 *     from F-RTO to the conventional RTO recovery
3529 */
3530static int tcp_process_frto(struct sock *sk, int flag)
3531{
3532        struct tcp_sock *tp = tcp_sk(sk);
3533
3534        tcp_verify_left_out(tp);
3535
3536        /* Duplicate the behavior from Loss state (fastretrans_alert) */
3537        if (flag & FLAG_DATA_ACKED)
3538                inet_csk(sk)->icsk_retransmits = 0;
3539
3540        if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3541            ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3542                tp->undo_marker = 0;
3543
3544        if (!before(tp->snd_una, tp->frto_highmark)) {
3545                tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3546                return 1;
3547        }
3548
3549        if (!tcp_is_sackfrto(tp)) {
3550                /* RFC4138 shortcoming in step 2; should also have case c):
3551                 * ACK isn't duplicate nor advances window, e.g., opposite dir
3552                 * data, winupdate
3553                 */
3554                if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3555                        return 1;
3556
3557                if (!(flag & FLAG_DATA_ACKED)) {
3558                        tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3559                                            flag);
3560                        return 1;
3561                }
3562        } else {
3563                if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3564                        /* Prevent sending of new data. */
3565                        tp->snd_cwnd = min(tp->snd_cwnd,
3566                                           tcp_packets_in_flight(tp));
3567                        return 1;
3568                }
3569
3570                if ((tp->frto_counter >= 2) &&
3571                    (!(flag & FLAG_FORWARD_PROGRESS) ||
3572                     ((flag & FLAG_DATA_SACKED) &&
3573                      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3574                        /* RFC4138 shortcoming (see comment above) */
3575                        if (!(flag & FLAG_FORWARD_PROGRESS) &&
3576                            (flag & FLAG_NOT_DUP))
3577                                return 1;
3578
3579                        tcp_enter_frto_loss(sk, 3, flag);
3580                        return 1;
3581                }
3582        }
3583
3584        if (tp->frto_counter == 1) {
3585                /* tcp_may_send_now needs to see updated state */
3586                tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3587                tp->frto_counter = 2;
3588
3589                if (!tcp_may_send_now(sk))
3590                        tcp_enter_frto_loss(sk, 2, flag);
3591
3592                return 1;
3593        } else {
3594                switch (sysctl_tcp_frto_response) {
3595                case 2:
3596                        tcp_undo_spur_to_response(sk, flag);
3597                        break;
3598                case 1:
3599                        tcp_conservative_spur_to_response(tp);
3600                        break;
3601                default:
3602                        tcp_ratehalving_spur_to_response(sk);
3603                        break;
3604                }
3605                tp->frto_counter = 0;
3606                tp->undo_marker = 0;
3607                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3608        }
3609        return 0;
3610}
3611
3612/* This routine deals with incoming acks, but not outgoing ones. */
3613static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3614{
3615        struct inet_connection_sock *icsk = inet_csk(sk);
3616        struct tcp_sock *tp = tcp_sk(sk);
3617        u32 prior_snd_una = tp->snd_una;
3618        u32 ack_seq = TCP_SKB_CB(skb)->seq;
3619        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3620        u32 prior_in_flight;
3621        u32 prior_fackets;
3622        int prior_packets;
3623        int frto_cwnd = 0;
3624
3625        /* If the ack is older than previous acks
3626         * then we can probably ignore it.
3627         */
3628        if (before(ack, prior_snd_una))
3629                goto old_ack;
3630
3631        /* If the ack includes data we haven't sent yet, discard
3632         * this segment (RFC793 Section 3.9).
3633         */
3634        if (after(ack, tp->snd_nxt))
3635                goto invalid_ack;
3636
3637        if (after(ack, prior_snd_una))
3638                flag |= FLAG_SND_UNA_ADVANCED;
3639
3640        if (sysctl_tcp_abc) {
3641                if (icsk->icsk_ca_state < TCP_CA_CWR)
3642                        tp->bytes_acked += ack - prior_snd_una;
3643                else if (icsk->icsk_ca_state == TCP_CA_Loss)
3644                        /* we assume just one segment left network */
3645                        tp->bytes_acked += min(ack - prior_snd_una,
3646                                               tp->mss_cache);
3647        }
3648
3649        prior_fackets = tp->fackets_out;
3650        prior_in_flight = tcp_packets_in_flight(tp);
3651
3652        if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3653                /* Window is constant, pure forward advance.
3654                 * No more checks are required.
3655                 * Note, we use the fact that SND.UNA>=SND.WL2.
3656                 */
3657                tcp_update_wl(tp, ack_seq);
3658                tp->snd_una = ack;
3659                flag |= FLAG_WIN_UPDATE;
3660
3661                tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3662
3663                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3664        } else {
3665                if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3666                        flag |= FLAG_DATA;
3667                else
3668                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3669
3670                flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3671
3672                if (TCP_SKB_CB(skb)->sacked)
3673                        flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3674
3675                if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3676                        flag |= FLAG_ECE;
3677
3678                tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3679        }
3680
3681        /* We passed data and got it acked, remove any soft error
3682         * log. Something worked...
3683         */
3684        sk->sk_err_soft = 0;
3685        icsk->icsk_probes_out = 0;
3686        tp->rcv_tstamp = tcp_time_stamp;
3687        prior_packets = tp->packets_out;
3688        if (!prior_packets)
3689                goto no_queue;
3690
3691        /* See if we can take anything off of the retransmit queue. */
3692        flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3693
3694        if (tp->frto_counter)
3695                frto_cwnd = tcp_process_frto(sk, flag);
3696        /* Guarantee sacktag reordering detection against wrap-arounds */
3697        if (before(tp->frto_highmark, tp->snd_una))
3698                tp->frto_highmark = 0;
3699
3700        if (tcp_ack_is_dubious(sk, flag)) {
3701                /* Advance CWND, if state allows this. */
3702                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3703                    tcp_may_raise_cwnd(sk, flag))
3704                        tcp_cong_avoid(sk, ack, prior_in_flight);
3705                tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3706                                      flag);
3707        } else {
3708                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3709                        tcp_cong_avoid(sk, ack, prior_in_flight);
3710        }
3711
3712        if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3713                dst_confirm(__sk_dst_get(sk));
3714
3715        return 1;
3716
3717no_queue:
3718        /* If this ack opens up a zero window, clear backoff.  It was
3719         * being used to time the probes, and is probably far higher than
3720         * it needs to be for normal retransmission.
3721         */
3722        if (tcp_send_head(sk))
3723                tcp_ack_probe(sk);
3724        return 1;
3725
3726invalid_ack:
3727        SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3728        return -1;
3729
3730old_ack:
3731        if (TCP_SKB_CB(skb)->sacked) {
3732                tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3733                if (icsk->icsk_ca_state == TCP_CA_Open)
3734                        tcp_try_keep_open(sk);
3735        }
3736
3737        SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3738        return 0;
3739}
3740
3741/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3742 * But, this can also be called on packets in the established flow when
3743 * the fast version below fails.
3744 */
3745void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3746                       u8 **hvpp, int estab)
3747{
3748        unsigned char *ptr;
3749        struct tcphdr *th = tcp_hdr(skb);
3750        int length = (th->doff * 4) - sizeof(struct tcphdr);
3751
3752        ptr = (unsigned char *)(th + 1);
3753        opt_rx->saw_tstamp = 0;
3754
3755        while (length > 0) {
3756                int opcode = *ptr++;
3757                int opsize;
3758
3759                switch (opcode) {
3760                case TCPOPT_EOL:
3761                        return;
3762                case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3763                        length--;
3764                        continue;
3765                default:
3766                        opsize = *ptr++;
3767                        if (opsize < 2) /* "silly options" */
3768                                return;
3769                        if (opsize > length)
3770                                return; /* don't parse partial options */
3771                        switch (opcode) {
3772                        case TCPOPT_MSS:
3773                                if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3774                                        u16 in_mss = get_unaligned_be16(ptr);
3775                                        if (in_mss) {
3776                                                if (opt_rx->user_mss &&
3777                                                    opt_rx->user_mss < in_mss)
3778                                                        in_mss = opt_rx->user_mss;
3779                                                opt_rx->mss_clamp = in_mss;
3780                                        }
3781                                }
3782                                break;
3783                        case TCPOPT_WINDOW:
3784                                if (opsize == TCPOLEN_WINDOW && th->syn &&
3785                                    !estab && sysctl_tcp_window_scaling) {
3786                                        __u8 snd_wscale = *(__u8 *)ptr;
3787                                        opt_rx->wscale_ok = 1;
3788                                        if (snd_wscale > 14) {
3789                                                if (net_ratelimit())
3790                                                        printk(KERN_INFO "tcp_parse_options: Illegal window "
3791                                                               "scaling value %d >14 received.\n",
3792                                                               snd_wscale);
3793                                                snd_wscale = 14;
3794                                        }
3795                                        opt_rx->snd_wscale = snd_wscale;
3796                                }
3797                                break;
3798                        case TCPOPT_TIMESTAMP:
3799                                if ((opsize == TCPOLEN_TIMESTAMP) &&
3800                                    ((estab && opt_rx->tstamp_ok) ||
3801                                     (!estab && sysctl_tcp_timestamps))) {
3802                                        opt_rx->saw_tstamp = 1;
3803                                        opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3804                                        opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3805                                }
3806                                break;
3807                        case TCPOPT_SACK_PERM:
3808                                if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3809                                    !estab && sysctl_tcp_sack) {
3810                                        opt_rx->sack_ok = 1;
3811                                        tcp_sack_reset(opt_rx);
3812                                }
3813                                break;
3814
3815                        case TCPOPT_SACK:
3816                                if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3817                                   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3818                                   opt_rx->sack_ok) {
3819                                        TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3820                                }
3821                                break;
3822#ifdef CONFIG_TCP_MD5SIG
3823                        case TCPOPT_MD5SIG:
3824                                /*
3825                                 * The MD5 Hash has already been
3826                                 * checked (see tcp_v{4,6}_do_rcv()).
3827                                 */
3828                                break;
3829#endif
3830                        case TCPOPT_COOKIE:
3831                                /* This option is variable length.
3832                                 */
3833                                switch (opsize) {
3834                                case TCPOLEN_COOKIE_BASE:
3835                                        /* not yet implemented */
3836                                        break;
3837                                case TCPOLEN_COOKIE_PAIR:
3838                                        /* not yet implemented */
3839                                        break;
3840                                case TCPOLEN_COOKIE_MIN+0:
3841                                case TCPOLEN_COOKIE_MIN+2:
3842                                case TCPOLEN_COOKIE_MIN+4:
3843                                case TCPOLEN_COOKIE_MIN+6:
3844                                case TCPOLEN_COOKIE_MAX:
3845                                        /* 16-bit multiple */
3846                                        opt_rx->cookie_plus = opsize;
3847                                        *hvpp = ptr;
3848                                        break;
3849                                default:
3850                                        /* ignore option */
3851                                        break;
3852                                }
3853                                break;
3854                        }
3855
3856                        ptr += opsize-2;
3857                        length -= opsize;
3858                }
3859        }
3860}
3861
3862static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3863{
3864        __be32 *ptr = (__be32 *)(th + 1);
3865
3866        if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3867                          | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3868                tp->rx_opt.saw_tstamp = 1;
3869                ++ptr;
3870                tp->rx_opt.rcv_tsval = ntohl(*ptr);
3871                ++ptr;
3872                tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3873                return 1;
3874        }
3875        return 0;
3876}
3877
3878/* Fast parse options. This hopes to only see timestamps.
3879 * If it is wrong it falls back on tcp_parse_options().
3880 */
3881static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3882                                  struct tcp_sock *tp, u8 **hvpp)
3883{
3884        /* In the spirit of fast parsing, compare doff directly to constant
3885         * values.  Because equality is used, short doff can be ignored here.
3886         */
3887        if (th->doff == (sizeof(*th) / 4)) {
3888                tp->rx_opt.saw_tstamp = 0;
3889                return 0;
3890        } else if (tp->rx_opt.tstamp_ok &&
3891                   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3892                if (tcp_parse_aligned_timestamp(tp, th))
3893                        return 1;
3894        }
3895        tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3896        return 1;
3897}
3898
3899#ifdef CONFIG_TCP_MD5SIG
3900/*
3901 * Parse MD5 Signature option
3902 */
3903u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3904{
3905        int length = (th->doff << 2) - sizeof (*th);
3906        u8 *ptr = (u8*)(th + 1);
3907
3908        /* If the TCP option is too short, we can short cut */
3909        if (length < TCPOLEN_MD5SIG)
3910                return NULL;
3911
3912        while (length > 0) {
3913                int opcode = *ptr++;
3914                int opsize;
3915
3916                switch(opcode) {
3917                case TCPOPT_EOL:
3918                        return NULL;
3919                case TCPOPT_NOP:
3920                        length--;
3921                        continue;
3922                default:
3923                        opsize = *ptr++;
3924                        if (opsize < 2 || opsize > length)
3925                                return NULL;
3926                        if (opcode == TCPOPT_MD5SIG)
3927                                return ptr;
3928                }
3929                ptr += opsize - 2;
3930                length -= opsize;
3931        }
3932        return NULL;
3933}
3934#endif
3935
3936static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3937{
3938        tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3939        tp->rx_opt.ts_recent_stamp = get_seconds();
3940}
3941
3942static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3943{
3944        if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3945                /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3946                 * extra check below makes sure this can only happen
3947                 * for pure ACK frames.  -DaveM
3948                 *
3949                 * Not only, also it occurs for expired timestamps.
3950                 */
3951
3952                if (tcp_paws_check(&tp->rx_opt, 0))
3953                        tcp_store_ts_recent(tp);
3954        }
3955}
3956
3957/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3958 *
3959 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3960 * it can pass through stack. So, the following predicate verifies that
3961 * this segment is not used for anything but congestion avoidance or
3962 * fast retransmit. Moreover, we even are able to eliminate most of such
3963 * second order effects, if we apply some small "replay" window (~RTO)
3964 * to timestamp space.
3965 *
3966 * All these measures still do not guarantee that we reject wrapped ACKs
3967 * on networks with high bandwidth, when sequence space is recycled fastly,
3968 * but it guarantees that such events will be very rare and do not affect
3969 * connection seriously. This doesn't look nice, but alas, PAWS is really
3970 * buggy extension.
3971 *
3972 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3973 * states that events when retransmit arrives after original data are rare.
3974 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3975 * the biggest problem on large power networks even with minor reordering.
3976 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3977 * up to bandwidth of 18Gigabit/sec. 8) ]
3978 */
3979
3980static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3981{
3982        struct tcp_sock *tp = tcp_sk(sk);
3983        struct tcphdr *th = tcp_hdr(skb);
3984        u32 seq = TCP_SKB_CB(skb)->seq;
3985        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3986
3987        return (/* 1. Pure ACK with correct sequence number. */
3988                (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3989
3990                /* 2. ... and duplicate ACK. */
3991                ack == tp->snd_una &&
3992
3993                /* 3. ... and does not update window. */
3994                !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3995
3996                /* 4. ... and sits in replay window. */
3997                (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3998}
3999
4000static inline int tcp_paws_discard(const struct sock *sk,
4001                                   const struct sk_buff *skb)
4002{
4003        const struct tcp_sock *tp = tcp_sk(sk);
4004
4005        return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4006               !tcp_disordered_ack(sk, skb);
4007}
4008
4009/* Check segment sequence number for validity.
4010 *
4011 * Segment controls are considered valid, if the segment
4012 * fits to the window after truncation to the window. Acceptability
4013 * of data (and SYN, FIN, of course) is checked separately.
4014 * See tcp_data_queue(), for example.
4015 *
4016 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4017 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4018 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4019 * (borrowed from freebsd)
4020 */
4021
4022static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4023{
4024        return  !before(end_seq, tp->rcv_wup) &&
4025                !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4026}
4027
4028/* When we get a reset we do this. */
4029static void tcp_reset(struct sock *sk)
4030{
4031        /* We want the right error as BSD sees it (and indeed as we do). */
4032        switch (sk->sk_state) {
4033        case TCP_SYN_SENT:
4034                sk->sk_err = ECONNREFUSED;
4035                break;
4036        case TCP_CLOSE_WAIT:
4037                sk->sk_err = EPIPE;
4038                break;
4039        case TCP_CLOSE:
4040                return;
4041        default:
4042                sk->sk_err = ECONNRESET;
4043        }
4044
4045        if (!sock_flag(sk, SOCK_DEAD))
4046                sk->sk_error_report(sk);
4047
4048        tcp_done(sk);
4049}
4050
4051/*
4052 *      Process the FIN bit. This now behaves as it is supposed to work
4053 *      and the FIN takes effect when it is validly part of sequence
4054 *      space. Not before when we get holes.
4055 *
4056 *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4057 *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4058 *      TIME-WAIT)
4059 *
4060 *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4061 *      close and we go into CLOSING (and later onto TIME-WAIT)
4062 *
4063 *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4064 */
4065static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4066{
4067        struct tcp_sock *tp = tcp_sk(sk);
4068
4069        inet_csk_schedule_ack(sk);
4070
4071        sk->sk_shutdown |= RCV_SHUTDOWN;
4072        sock_set_flag(sk, SOCK_DONE);
4073
4074        switch (sk->sk_state) {
4075        case TCP_SYN_RECV:
4076        case TCP_ESTABLISHED:
4077                /* Move to CLOSE_WAIT */
4078                tcp_set_state(sk, TCP_CLOSE_WAIT);
4079                inet_csk(sk)->icsk_ack.pingpong = 1;
4080                break;
4081
4082        case TCP_CLOSE_WAIT:
4083        case TCP_CLOSING:
4084                /* Received a retransmission of the FIN, do
4085                 * nothing.
4086                 */
4087                break;
4088        case TCP_LAST_ACK:
4089                /* RFC793: Remain in the LAST-ACK state. */
4090                break;
4091
4092        case TCP_FIN_WAIT1:
4093                /* This case occurs when a simultaneous close
4094                 * happens, we must ack the received FIN and
4095                 * enter the CLOSING state.
4096                 */
4097                tcp_send_ack(sk);
4098                tcp_set_state(sk, TCP_CLOSING);
4099                break;
4100        case TCP_FIN_WAIT2:
4101                /* Received a FIN -- send ACK and enter TIME_WAIT. */
4102                tcp_send_ack(sk);
4103                tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4104                break;
4105        default:
4106                /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4107                 * cases we should never reach this piece of code.
4108                 */
4109                printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4110                       __func__, sk->sk_state);
4111                break;
4112        }
4113
4114        /* It _is_ possible, that we have something out-of-order _after_ FIN.
4115         * Probably, we should reset in this case. For now drop them.
4116         */
4117        __skb_queue_purge(&tp->out_of_order_queue);
4118        if (tcp_is_sack(tp))
4119                tcp_sack_reset(&tp->rx_opt);
4120        sk_mem_reclaim(sk);
4121
4122        if (!sock_flag(sk, SOCK_DEAD)) {
4123                sk->sk_state_change(sk);
4124
4125                /* Do not send POLL_HUP for half duplex close. */
4126                if (sk->sk_shutdown == SHUTDOWN_MASK ||
4127                    sk->sk_state == TCP_CLOSE)
4128                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4129                else
4130                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4131        }
4132}
4133
4134static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4135                                  u32 end_seq)
4136{
4137        if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4138                if (before(seq, sp->start_seq))
4139                        sp->start_seq = seq;
4140                if (after(end_seq, sp->end_seq))
4141                        sp->end_seq = end_seq;
4142                return 1;
4143        }
4144        return 0;
4145}
4146
4147static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4148{
4149        struct tcp_sock *tp = tcp_sk(sk);
4150
4151        if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4152                int mib_idx;
4153
4154                if (before(seq, tp->rcv_nxt))
4155                        mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4156                else
4157                        mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4158
4159                NET_INC_STATS_BH(sock_net(sk), mib_idx);
4160
4161                tp->rx_opt.dsack = 1;
4162                tp->duplicate_sack[0].start_seq = seq;
4163                tp->duplicate_sack[0].end_seq = end_seq;
4164        }
4165}
4166
4167static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4168{
4169        struct tcp_sock *tp = tcp_sk(sk);
4170
4171        if (!tp->rx_opt.dsack)
4172                tcp_dsack_set(sk, seq, end_seq);
4173        else
4174                tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4175}
4176
4177static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4178{
4179        struct tcp_sock *tp = tcp_sk(sk);
4180
4181        if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4182            before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4183                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4184                tcp_enter_quickack_mode(sk);
4185
4186                if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4187                        u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4188
4189                        if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4190                                end_seq = tp->rcv_nxt;
4191                        tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4192                }
4193        }
4194
4195        tcp_send_ack(sk);
4196}
4197
4198/* These routines update the SACK block as out-of-order packets arrive or
4199 * in-order packets close up the sequence space.
4200 */
4201static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4202{
4203        int this_sack;
4204        struct tcp_sack_block *sp = &tp->selective_acks[0];
4205        struct tcp_sack_block *swalk = sp + 1;
4206
4207        /* See if the recent change to the first SACK eats into
4208         * or hits the sequence space of other SACK blocks, if so coalesce.
4209         */
4210        for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4211                if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4212                        int i;
4213
4214                        /* Zap SWALK, by moving every further SACK up by one slot.
4215                         * Decrease num_sacks.
4216                         */
4217                        tp->rx_opt.num_sacks--;
4218                        for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4219                                sp[i] = sp[i + 1];
4220                        continue;
4221                }
4222                this_sack++, swalk++;
4223        }
4224}
4225
4226static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4227{
4228        struct tcp_sock *tp = tcp_sk(sk);
4229        struct tcp_sack_block *sp = &tp->selective_acks[0];
4230        int cur_sacks = tp->rx_opt.num_sacks;
4231        int this_sack;
4232
4233        if (!cur_sacks)
4234                goto new_sack;
4235
4236        for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4237                if (tcp_sack_extend(sp, seq, end_seq)) {
4238                        /* Rotate this_sack to the first one. */
4239                        for (; this_sack > 0; this_sack--, sp--)
4240                                swap(*sp, *(sp - 1));
4241                        if (cur_sacks > 1)
4242                                tcp_sack_maybe_coalesce(tp);
4243                        return;
4244                }
4245        }
4246
4247        /* Could not find an adjacent existing SACK, build a new one,
4248         * put it at the front, and shift everyone else down.  We
4249         * always know there is at least one SACK present already here.
4250         *
4251         * If the sack array is full, forget about the last one.
4252         */
4253        if (this_sack >= TCP_NUM_SACKS) {
4254                this_sack--;
4255                tp->rx_opt.num_sacks--;
4256                sp--;
4257        }
4258        for (; this_sack > 0; this_sack--, sp--)
4259                *sp = *(sp - 1);
4260
4261new_sack:
4262        /* Build the new head SACK, and we're done. */
4263        sp->start_seq = seq;
4264        sp->end_seq = end_seq;
4265        tp->rx_opt.num_sacks++;
4266}
4267
4268/* RCV.NXT advances, some SACKs should be eaten. */
4269
4270static void tcp_sack_remove(struct tcp_sock *tp)
4271{
4272        struct tcp_sack_block *sp = &tp->selective_acks[0];
4273        int num_sacks = tp->rx_opt.num_sacks;
4274        int this_sack;
4275
4276        /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4277        if (skb_queue_empty(&tp->out_of_order_queue)) {
4278                tp->rx_opt.num_sacks = 0;
4279                return;
4280        }
4281
4282        for (this_sack = 0; this_sack < num_sacks;) {
4283                /* Check if the start of the sack is covered by RCV.NXT. */
4284                if (!before(tp->rcv_nxt, sp->start_seq)) {
4285                        int i;
4286
4287                        /* RCV.NXT must cover all the block! */
4288                        WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4289
4290                        /* Zap this SACK, by moving forward any other SACKS. */
4291                        for (i=this_sack+1; i < num_sacks; i++)
4292                                tp->selective_acks[i-1] = tp->selective_acks[i];
4293                        num_sacks--;
4294                        continue;
4295                }
4296                this_sack++;
4297                sp++;
4298        }
4299