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#define pr_fmt(fmt) "TCP: " fmt
  65
  66#include <linux/mm.h>
  67#include <linux/slab.h>
  68#include <linux/module.h>
  69#include <linux/sysctl.h>
  70#include <linux/kernel.h>
  71#include <net/dst.h>
  72#include <net/tcp.h>
  73#include <net/inet_common.h>
  74#include <linux/ipsec.h>
  75#include <asm/unaligned.h>
  76#include <net/netdma.h>
  77
  78int sysctl_tcp_timestamps __read_mostly = 1;
  79int sysctl_tcp_window_scaling __read_mostly = 1;
  80int sysctl_tcp_sack __read_mostly = 1;
  81int sysctl_tcp_fack __read_mostly = 1;
  82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  83EXPORT_SYMBOL(sysctl_tcp_reordering);
  84int sysctl_tcp_ecn __read_mostly = 2;
  85EXPORT_SYMBOL(sysctl_tcp_ecn);
  86int sysctl_tcp_dsack __read_mostly = 1;
  87int sysctl_tcp_app_win __read_mostly = 31;
  88int sysctl_tcp_adv_win_scale __read_mostly = 1;
  89EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
  90
  91/* rfc5961 challenge ack rate limiting */
  92int sysctl_tcp_challenge_ack_limit = 100;
  93
  94int sysctl_tcp_stdurg __read_mostly;
  95int sysctl_tcp_rfc1337 __read_mostly;
  96int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  97int sysctl_tcp_frto __read_mostly = 2;
  98int sysctl_tcp_frto_response __read_mostly;
  99
 100int sysctl_tcp_thin_dupack __read_mostly;
 101
 102int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 103int sysctl_tcp_abc __read_mostly;
 104int sysctl_tcp_early_retrans __read_mostly = 2;
 105
 106#define FLAG_DATA               0x01 /* Incoming frame contained data.          */
 107#define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
 108#define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
 109#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
 110#define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
 111#define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
 112#define FLAG_ECE                0x40 /* ECE in this ACK                         */
 113#define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
 114#define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
 115#define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 116#define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
 117#define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
 118#define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
 119
 120#define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 121#define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
 122#define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
 123#define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
 124#define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
 125
 126#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
 127#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
 128
 129/* Adapt the MSS value used to make delayed ack decision to the
 130 * real world.
 131 */
 132static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
 133{
 134        struct inet_connection_sock *icsk = inet_csk(sk);
 135        const unsigned int lss = icsk->icsk_ack.last_seg_size;
 136        unsigned int len;
 137
 138        icsk->icsk_ack.last_seg_size = 0;
 139
 140        /* skb->len may jitter because of SACKs, even if peer
 141         * sends good full-sized frames.
 142         */
 143        len = skb_shinfo(skb)->gso_size ? : skb->len;
 144        if (len >= icsk->icsk_ack.rcv_mss) {
 145                icsk->icsk_ack.rcv_mss = len;
 146        } else {
 147                /* Otherwise, we make more careful check taking into account,
 148                 * that SACKs block is variable.
 149                 *
 150                 * "len" is invariant segment length, including TCP header.
 151                 */
 152                len += skb->data - skb_transport_header(skb);
 153                if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
 154                    /* If PSH is not set, packet should be
 155                     * full sized, provided peer TCP is not badly broken.
 156                     * This observation (if it is correct 8)) allows
 157                     * to handle super-low mtu links fairly.
 158                     */
 159                    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
 160                     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
 161                        /* Subtract also invariant (if peer is RFC compliant),
 162                         * tcp header plus fixed timestamp option length.
 163                         * Resulting "len" is MSS free of SACK jitter.
 164                         */
 165                        len -= tcp_sk(sk)->tcp_header_len;
 166                        icsk->icsk_ack.last_seg_size = len;
 167                        if (len == lss) {
 168                                icsk->icsk_ack.rcv_mss = len;
 169                                return;
 170                        }
 171                }
 172                if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
 173                        icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
 174                icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
 175        }
 176}
 177
 178static void tcp_incr_quickack(struct sock *sk)
 179{
 180        struct inet_connection_sock *icsk = inet_csk(sk);
 181        unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
 182
 183        if (quickacks == 0)
 184                quickacks = 2;
 185        if (quickacks > icsk->icsk_ack.quick)
 186                icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
 187}
 188
 189static void tcp_enter_quickack_mode(struct sock *sk)
 190{
 191        struct inet_connection_sock *icsk = inet_csk(sk);
 192        tcp_incr_quickack(sk);
 193        icsk->icsk_ack.pingpong = 0;
 194        icsk->icsk_ack.ato = TCP_ATO_MIN;
 195}
 196
 197/* Send ACKs quickly, if "quick" count is not exhausted
 198 * and the session is not interactive.
 199 */
 200
 201static inline bool tcp_in_quickack_mode(const struct sock *sk)
 202{
 203        const struct inet_connection_sock *icsk = inet_csk(sk);
 204
 205        return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
 206}
 207
 208static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
 209{
 210        if (tp->ecn_flags & TCP_ECN_OK)
 211                tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
 212}
 213
 214static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
 215{
 216        if (tcp_hdr(skb)->cwr)
 217                tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 218}
 219
 220static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
 221{
 222        tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 223}
 224
 225static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
 226{
 227        if (!(tp->ecn_flags & TCP_ECN_OK))
 228                return;
 229
 230        switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
 231        case INET_ECN_NOT_ECT:
 232                /* Funny extension: if ECT is not set on a segment,
 233                 * and we already seen ECT on a previous segment,
 234                 * it is probably a retransmit.
 235                 */
 236                if (tp->ecn_flags & TCP_ECN_SEEN)
 237                        tcp_enter_quickack_mode((struct sock *)tp);
 238                break;
 239        case INET_ECN_CE:
 240                tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
 241                /* fallinto */
 242        default:
 243                tp->ecn_flags |= TCP_ECN_SEEN;
 244        }
 245}
 246
 247static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
 248{
 249        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
 250                tp->ecn_flags &= ~TCP_ECN_OK;
 251}
 252
 253static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
 254{
 255        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
 256                tp->ecn_flags &= ~TCP_ECN_OK;
 257}
 258
 259static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
 260{
 261        if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
 262                return true;
 263        return false;
 264}
 265
 266/* Buffer size and advertised window tuning.
 267 *
 268 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
 269 */
 270
 271static void tcp_fixup_sndbuf(struct sock *sk)
 272{
 273        int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
 274
 275        sndmem *= TCP_INIT_CWND;
 276        if (sk->sk_sndbuf < sndmem)
 277                sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
 278}
 279
 280/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
 281 *
 282 * All tcp_full_space() is split to two parts: "network" buffer, allocated
 283 * forward and advertised in receiver window (tp->rcv_wnd) and
 284 * "application buffer", required to isolate scheduling/application
 285 * latencies from network.
 286 * window_clamp is maximal advertised window. It can be less than
 287 * tcp_full_space(), in this case tcp_full_space() - window_clamp
 288 * is reserved for "application" buffer. The less window_clamp is
 289 * the smoother our behaviour from viewpoint of network, but the lower
 290 * throughput and the higher sensitivity of the connection to losses. 8)
 291 *
 292 * rcv_ssthresh is more strict window_clamp used at "slow start"
 293 * phase to predict further behaviour of this connection.
 294 * It is used for two goals:
 295 * - to enforce header prediction at sender, even when application
 296 *   requires some significant "application buffer". It is check #1.
 297 * - to prevent pruning of receive queue because of misprediction
 298 *   of receiver window. Check #2.
 299 *
 300 * The scheme does not work when sender sends good segments opening
 301 * window and then starts to feed us spaghetti. But it should work
 302 * in common situations. Otherwise, we have to rely on queue collapsing.
 303 */
 304
 305/* Slow part of check#2. */
 306static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
 307{
 308        struct tcp_sock *tp = tcp_sk(sk);
 309        /* Optimize this! */
 310        int truesize = tcp_win_from_space(skb->truesize) >> 1;
 311        int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
 312
 313        while (tp->rcv_ssthresh <= window) {
 314                if (truesize <= skb->len)
 315                        return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
 316
 317                truesize >>= 1;
 318                window >>= 1;
 319        }
 320        return 0;
 321}
 322
 323static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
 324{
 325        struct tcp_sock *tp = tcp_sk(sk);
 326
 327        /* Check #1 */
 328        if (tp->rcv_ssthresh < tp->window_clamp &&
 329            (int)tp->rcv_ssthresh < tcp_space(sk) &&
 330            !sk_under_memory_pressure(sk)) {
 331                int incr;
 332
 333                /* Check #2. Increase window, if skb with such overhead
 334                 * will fit to rcvbuf in future.
 335                 */
 336                if (tcp_win_from_space(skb->truesize) <= skb->len)
 337                        incr = 2 * tp->advmss;
 338                else
 339                        incr = __tcp_grow_window(sk, skb);
 340
 341                if (incr) {
 342                        incr = max_t(int, incr, 2 * skb->len);
 343                        tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
 344                                               tp->window_clamp);
 345                        inet_csk(sk)->icsk_ack.quick |= 1;
 346                }
 347        }
 348}
 349
 350/* 3. Tuning rcvbuf, when connection enters established state. */
 351
 352static void tcp_fixup_rcvbuf(struct sock *sk)
 353{
 354        u32 mss = tcp_sk(sk)->advmss;
 355        u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
 356        int rcvmem;
 357
 358        /* Limit to 10 segments if mss <= 1460,
 359         * or 14600/mss segments, with a minimum of two segments.
 360         */
 361        if (mss > 1460)
 362                icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
 363
 364        rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
 365        while (tcp_win_from_space(rcvmem) < mss)
 366                rcvmem += 128;
 367
 368        rcvmem *= icwnd;
 369
 370        if (sk->sk_rcvbuf < rcvmem)
 371                sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
 372}
 373
 374/* 4. Try to fixup all. It is made immediately after connection enters
 375 *    established state.
 376 */
 377static void tcp_init_buffer_space(struct sock *sk)
 378{
 379        struct tcp_sock *tp = tcp_sk(sk);
 380        int maxwin;
 381
 382        if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 383                tcp_fixup_rcvbuf(sk);
 384        if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
 385                tcp_fixup_sndbuf(sk);
 386
 387        tp->rcvq_space.space = tp->rcv_wnd;
 388
 389        maxwin = tcp_full_space(sk);
 390
 391        if (tp->window_clamp >= maxwin) {
 392                tp->window_clamp = maxwin;
 393
 394                if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
 395                        tp->window_clamp = max(maxwin -
 396                                               (maxwin >> sysctl_tcp_app_win),
 397                                               4 * tp->advmss);
 398        }
 399
 400        /* Force reservation of one segment. */
 401        if (sysctl_tcp_app_win &&
 402            tp->window_clamp > 2 * tp->advmss &&
 403            tp->window_clamp + tp->advmss > maxwin)
 404                tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
 405
 406        tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
 407        tp->snd_cwnd_stamp = tcp_time_stamp;
 408}
 409
 410/* 5. Recalculate window clamp after socket hit its memory bounds. */
 411static void tcp_clamp_window(struct sock *sk)
 412{
 413        struct tcp_sock *tp = tcp_sk(sk);
 414        struct inet_connection_sock *icsk = inet_csk(sk);
 415
 416        icsk->icsk_ack.quick = 0;
 417
 418        if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 419            !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
 420            !sk_under_memory_pressure(sk) &&
 421            sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
 422                sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 423                                    sysctl_tcp_rmem[2]);
 424        }
 425        if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 426                tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
 427}
 428
 429/* Initialize RCV_MSS value.
 430 * RCV_MSS is an our guess about MSS used by the peer.
 431 * We haven't any direct information about the MSS.
 432 * It's better to underestimate the RCV_MSS rather than overestimate.
 433 * Overestimations make us ACKing less frequently than needed.
 434 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
 435 */
 436void tcp_initialize_rcv_mss(struct sock *sk)
 437{
 438        const struct tcp_sock *tp = tcp_sk(sk);
 439        unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
 440
 441        hint = min(hint, tp->rcv_wnd / 2);
 442        hint = min(hint, TCP_MSS_DEFAULT);
 443        hint = max(hint, TCP_MIN_MSS);
 444
 445        inet_csk(sk)->icsk_ack.rcv_mss = hint;
 446}
 447EXPORT_SYMBOL(tcp_initialize_rcv_mss);
 448
 449/* Receiver "autotuning" code.
 450 *
 451 * The algorithm for RTT estimation w/o timestamps is based on
 452 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
 453 * <http://public.lanl.gov/radiant/pubs.html#DRS>
 454 *
 455 * More detail on this code can be found at
 456 * <http://staff.psc.edu/jheffner/>,
 457 * though this reference is out of date.  A new paper
 458 * is pending.
 459 */
 460static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 461{
 462        u32 new_sample = tp->rcv_rtt_est.rtt;
 463        long m = sample;
 464
 465        if (m == 0)
 466                m = 1;
 467
 468        if (new_sample != 0) {
 469                /* If we sample in larger samples in the non-timestamp
 470                 * case, we could grossly overestimate the RTT especially
 471                 * with chatty applications or bulk transfer apps which
 472                 * are stalled on filesystem I/O.
 473                 *
 474                 * Also, since we are only going for a minimum in the
 475                 * non-timestamp case, we do not smooth things out
 476                 * else with timestamps disabled convergence takes too
 477                 * long.
 478                 */
 479                if (!win_dep) {
 480                        m -= (new_sample >> 3);
 481                        new_sample += m;
 482                } else {
 483                        m <<= 3;
 484                        if (m < new_sample)
 485                                new_sample = m;
 486                }
 487        } else {
 488                /* No previous measure. */
 489                new_sample = m << 3;
 490        }
 491
 492        if (tp->rcv_rtt_est.rtt != new_sample)
 493                tp->rcv_rtt_est.rtt = new_sample;
 494}
 495
 496static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
 497{
 498        if (tp->rcv_rtt_est.time == 0)
 499                goto new_measure;
 500        if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
 501                return;
 502        tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
 503
 504new_measure:
 505        tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
 506        tp->rcv_rtt_est.time = tcp_time_stamp;
 507}
 508
 509static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 510                                          const struct sk_buff *skb)
 511{
 512        struct tcp_sock *tp = tcp_sk(sk);
 513        if (tp->rx_opt.rcv_tsecr &&
 514            (TCP_SKB_CB(skb)->end_seq -
 515             TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
 516                tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
 517}
 518
 519/*
 520 * This function should be called every time data is copied to user space.
 521 * It calculates the appropriate TCP receive buffer space.
 522 */
 523void tcp_rcv_space_adjust(struct sock *sk)
 524{
 525        struct tcp_sock *tp = tcp_sk(sk);
 526        int time;
 527        int space;
 528
 529        if (tp->rcvq_space.time == 0)
 530                goto new_measure;
 531
 532        time = tcp_time_stamp - tp->rcvq_space.time;
 533        if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
 534                return;
 535
 536        space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
 537
 538        space = max(tp->rcvq_space.space, space);
 539
 540        if (tp->rcvq_space.space != space) {
 541                int rcvmem;
 542
 543                tp->rcvq_space.space = space;
 544
 545                if (sysctl_tcp_moderate_rcvbuf &&
 546                    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 547                        int new_clamp = space;
 548
 549                        /* Receive space grows, normalize in order to
 550                         * take into account packet headers and sk_buff
 551                         * structure overhead.
 552                         */
 553                        space /= tp->advmss;
 554                        if (!space)
 555                                space = 1;
 556                        rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
 557                        while (tcp_win_from_space(rcvmem) < tp->advmss)
 558                                rcvmem += 128;
 559                        space *= rcvmem;
 560                        space = min(space, sysctl_tcp_rmem[2]);
 561                        if (space > sk->sk_rcvbuf) {
 562                                sk->sk_rcvbuf = space;
 563
 564                                /* Make the window clamp follow along.  */
 565                                tp->window_clamp = new_clamp;
 566                        }
 567                }
 568        }
 569
 570new_measure:
 571        tp->rcvq_space.seq = tp->copied_seq;
 572        tp->rcvq_space.time = tcp_time_stamp;
 573}
 574
 575/* There is something which you must keep in mind when you analyze the
 576 * behavior of the tp->ato delayed ack timeout interval.  When a
 577 * connection starts up, we want to ack as quickly as possible.  The
 578 * problem is that "good" TCP's do slow start at the beginning of data
 579 * transmission.  The means that until we send the first few ACK's the
 580 * sender will sit on his end and only queue most of his data, because
 581 * he can only send snd_cwnd unacked packets at any given time.  For
 582 * each ACK we send, he increments snd_cwnd and transmits more of his
 583 * queue.  -DaveM
 584 */
 585static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
 586{
 587        struct tcp_sock *tp = tcp_sk(sk);
 588        struct inet_connection_sock *icsk = inet_csk(sk);
 589        u32 now;
 590
 591        inet_csk_schedule_ack(sk);
 592
 593        tcp_measure_rcv_mss(sk, skb);
 594
 595        tcp_rcv_rtt_measure(tp);
 596
 597        now = tcp_time_stamp;
 598
 599        if (!icsk->icsk_ack.ato) {
 600                /* The _first_ data packet received, initialize
 601                 * delayed ACK engine.
 602                 */
 603                tcp_incr_quickack(sk);
 604                icsk->icsk_ack.ato = TCP_ATO_MIN;
 605        } else {
 606                int m = now - icsk->icsk_ack.lrcvtime;
 607
 608                if (m <= TCP_ATO_MIN / 2) {
 609                        /* The fastest case is the first. */
 610                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
 611                } else if (m < icsk->icsk_ack.ato) {
 612                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
 613                        if (icsk->icsk_ack.ato > icsk->icsk_rto)
 614                                icsk->icsk_ack.ato = icsk->icsk_rto;
 615                } else if (m > icsk->icsk_rto) {
 616                        /* Too long gap. Apparently sender failed to
 617                         * restart window, so that we send ACKs quickly.
 618                         */
 619                        tcp_incr_quickack(sk);
 620                        sk_mem_reclaim(sk);
 621                }
 622        }
 623        icsk->icsk_ack.lrcvtime = now;
 624
 625        TCP_ECN_check_ce(tp, skb);
 626
 627        if (skb->len >= 128)
 628                tcp_grow_window(sk, skb);
 629}
 630
 631/* Called to compute a smoothed rtt estimate. The data fed to this
 632 * routine either comes from timestamps, or from segments that were
 633 * known _not_ to have been retransmitted [see Karn/Partridge
 634 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
 635 * piece by Van Jacobson.
 636 * NOTE: the next three routines used to be one big routine.
 637 * To save cycles in the RFC 1323 implementation it was better to break
 638 * it up into three procedures. -- erics
 639 */
 640static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
 641{
 642        struct tcp_sock *tp = tcp_sk(sk);
 643        long m = mrtt; /* RTT */
 644
 645        /*      The following amusing code comes from Jacobson's
 646         *      article in SIGCOMM '88.  Note that rtt and mdev
 647         *      are scaled versions of rtt and mean deviation.
 648         *      This is designed to be as fast as possible
 649         *      m stands for "measurement".
 650         *
 651         *      On a 1990 paper the rto value is changed to:
 652         *      RTO = rtt + 4 * mdev
 653         *
 654         * Funny. This algorithm seems to be very broken.
 655         * These formulae increase RTO, when it should be decreased, increase
 656         * too slowly, when it should be increased quickly, decrease too quickly
 657         * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
 658         * does not matter how to _calculate_ it. Seems, it was trap
 659         * that VJ failed to avoid. 8)
 660         */
 661        if (m == 0)
 662                m = 1;
 663        if (tp->srtt != 0) {
 664                m -= (tp->srtt >> 3);   /* m is now error in rtt est */
 665                tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
 666                if (m < 0) {
 667                        m = -m;         /* m is now abs(error) */
 668                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 669                        /* This is similar to one of Eifel findings.
 670                         * Eifel blocks mdev updates when rtt decreases.
 671                         * This solution is a bit different: we use finer gain
 672                         * for mdev in this case (alpha*beta).
 673                         * Like Eifel it also prevents growth of rto,
 674                         * but also it limits too fast rto decreases,
 675                         * happening in pure Eifel.
 676                         */
 677                        if (m > 0)
 678                                m >>= 3;
 679                } else {
 680                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 681                }
 682                tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
 683                if (tp->mdev > tp->mdev_max) {
 684                        tp->mdev_max = tp->mdev;
 685                        if (tp->mdev_max > tp->rttvar)
 686                                tp->rttvar = tp->mdev_max;
 687                }
 688                if (after(tp->snd_una, tp->rtt_seq)) {
 689                        if (tp->mdev_max < tp->rttvar)
 690                                tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
 691                        tp->rtt_seq = tp->snd_nxt;
 692                        tp->mdev_max = tcp_rto_min(sk);
 693                }
 694        } else {
 695                /* no previous measure. */
 696                tp->srtt = m << 3;      /* take the measured time to be rtt */
 697                tp->mdev = m << 1;      /* make sure rto = 3*rtt */
 698                tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 699                tp->rtt_seq = tp->snd_nxt;
 700        }
 701}
 702
 703/* Calculate rto without backoff.  This is the second half of Van Jacobson's
 704 * routine referred to above.
 705 */
 706void tcp_set_rto(struct sock *sk)
 707{
 708        const struct tcp_sock *tp = tcp_sk(sk);
 709        /* Old crap is replaced with new one. 8)
 710         *
 711         * More seriously:
 712         * 1. If rtt variance happened to be less 50msec, it is hallucination.
 713         *    It cannot be less due to utterly erratic ACK generation made
 714         *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
 715         *    to do with delayed acks, because at cwnd>2 true delack timeout
 716         *    is invisible. Actually, Linux-2.4 also generates erratic
 717         *    ACKs in some circumstances.
 718         */
 719        inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
 720
 721        /* 2. Fixups made earlier cannot be right.
 722         *    If we do not estimate RTO correctly without them,
 723         *    all the algo is pure shit and should be replaced
 724         *    with correct one. It is exactly, which we pretend to do.
 725         */
 726
 727        /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
 728         * guarantees that rto is higher.
 729         */
 730        tcp_bound_rto(sk);
 731}
 732
 733__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
 734{
 735        __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 736
 737        if (!cwnd)
 738                cwnd = TCP_INIT_CWND;
 739        return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 740}
 741
 742/* Set slow start threshold and cwnd not falling to slow start */
 743void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
 744{
 745        struct tcp_sock *tp = tcp_sk(sk);
 746        const struct inet_connection_sock *icsk = inet_csk(sk);
 747
 748        tp->prior_ssthresh = 0;
 749        tp->bytes_acked = 0;
 750        if (icsk->icsk_ca_state < TCP_CA_CWR) {
 751                tp->undo_marker = 0;
 752                if (set_ssthresh)
 753                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
 754                tp->snd_cwnd = min(tp->snd_cwnd,
 755                                   tcp_packets_in_flight(tp) + 1U);
 756                tp->snd_cwnd_cnt = 0;
 757                tp->high_seq = tp->snd_nxt;
 758                tp->snd_cwnd_stamp = tcp_time_stamp;
 759                TCP_ECN_queue_cwr(tp);
 760
 761                tcp_set_ca_state(sk, TCP_CA_CWR);
 762        }
 763}
 764
 765/*
 766 * Packet counting of FACK is based on in-order assumptions, therefore TCP
 767 * disables it when reordering is detected
 768 */
 769void tcp_disable_fack(struct tcp_sock *tp)
 770{
 771        /* RFC3517 uses different metric in lost marker => reset on change */
 772        if (tcp_is_fack(tp))
 773                tp->lost_skb_hint = NULL;
 774        tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
 775}
 776
 777/* Take a notice that peer is sending D-SACKs */
 778static void tcp_dsack_seen(struct tcp_sock *tp)
 779{
 780        tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
 781}
 782
 783static void tcp_update_reordering(struct sock *sk, const int metric,
 784                                  const int ts)
 785{
 786        struct tcp_sock *tp = tcp_sk(sk);
 787        if (metric > tp->reordering) {
 788                int mib_idx;
 789
 790                tp->reordering = min(TCP_MAX_REORDERING, metric);
 791
 792                /* This exciting event is worth to be remembered. 8) */
 793                if (ts)
 794                        mib_idx = LINUX_MIB_TCPTSREORDER;
 795                else if (tcp_is_reno(tp))
 796                        mib_idx = LINUX_MIB_TCPRENOREORDER;
 797                else if (tcp_is_fack(tp))
 798                        mib_idx = LINUX_MIB_TCPFACKREORDER;
 799                else
 800                        mib_idx = LINUX_MIB_TCPSACKREORDER;
 801
 802                NET_INC_STATS_BH(sock_net(sk), mib_idx);
 803#if FASTRETRANS_DEBUG > 1
 804                pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
 805                         tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
 806                         tp->reordering,
 807                         tp->fackets_out,
 808                         tp->sacked_out,
 809                         tp->undo_marker ? tp->undo_retrans : 0);
 810#endif
 811                tcp_disable_fack(tp);
 812        }
 813
 814        if (metric > 0)
 815                tcp_disable_early_retrans(tp);
 816}
 817
 818/* This must be called before lost_out is incremented */
 819static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
 820{
 821        if ((tp->retransmit_skb_hint == NULL) ||
 822            before(TCP_SKB_CB(skb)->seq,
 823                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
 824                tp->retransmit_skb_hint = skb;
 825
 826        if (!tp->lost_out ||
 827            after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
 828                tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
 829}
 830
 831static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
 832{
 833        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 834                tcp_verify_retransmit_hint(tp, skb);
 835
 836                tp->lost_out += tcp_skb_pcount(skb);
 837                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 838        }
 839}
 840
 841static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
 842                                            struct sk_buff *skb)
 843{
 844        tcp_verify_retransmit_hint(tp, skb);
 845
 846        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 847                tp->lost_out += tcp_skb_pcount(skb);
 848                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 849        }
 850}
 851
 852/* This procedure tags the retransmission queue when SACKs arrive.
 853 *
 854 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
 855 * Packets in queue with these bits set are counted in variables
 856 * sacked_out, retrans_out and lost_out, correspondingly.
 857 *
 858 * Valid combinations are:
 859 * Tag  InFlight        Description
 860 * 0    1               - orig segment is in flight.
 861 * S    0               - nothing flies, orig reached receiver.
 862 * L    0               - nothing flies, orig lost by net.
 863 * R    2               - both orig and retransmit are in flight.
 864 * L|R  1               - orig is lost, retransmit is in flight.
 865 * S|R  1               - orig reached receiver, retrans is still in flight.
 866 * (L|S|R is logically valid, it could occur when L|R is sacked,
 867 *  but it is equivalent to plain S and code short-curcuits it to S.
 868 *  L|S is logically invalid, it would mean -1 packet in flight 8))
 869 *
 870 * These 6 states form finite state machine, controlled by the following events:
 871 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
 872 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
 873 * 3. Loss detection event of two flavors:
 874 *      A. Scoreboard estimator decided the packet is lost.
 875 *         A'. Reno "three dupacks" marks head of queue lost.
 876 *         A''. Its FACK modification, head until snd.fack is lost.
 877 *      B. SACK arrives sacking SND.NXT at the moment, when the
 878 *         segment was retransmitted.
 879 * 4. D-SACK added new rule: D-SACK changes any tag to S.
 880 *
 881 * It is pleasant to note, that state diagram turns out to be commutative,
 882 * so that we are allowed not to be bothered by order of our actions,
 883 * when multiple events arrive simultaneously. (see the function below).
 884 *
 885 * Reordering detection.
 886 * --------------------
 887 * Reordering metric is maximal distance, which a packet can be displaced
 888 * in packet stream. With SACKs we can estimate it:
 889 *
 890 * 1. SACK fills old hole and the corresponding segment was not
 891 *    ever retransmitted -> reordering. Alas, we cannot use it
 892 *    when segment was retransmitted.
 893 * 2. The last flaw is solved with D-SACK. D-SACK arrives
 894 *    for retransmitted and already SACKed segment -> reordering..
 895 * Both of these heuristics are not used in Loss state, when we cannot
 896 * account for retransmits accurately.
 897 *
 898 * SACK block validation.
 899 * ----------------------
 900 *
 901 * SACK block range validation checks that the received SACK block fits to
 902 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
 903 * Note that SND.UNA is not included to the range though being valid because
 904 * it means that the receiver is rather inconsistent with itself reporting
 905 * SACK reneging when it should advance SND.UNA. Such SACK block this is
 906 * perfectly valid, however, in light of RFC2018 which explicitly states
 907 * that "SACK block MUST reflect the newest segment.  Even if the newest
 908 * segment is going to be discarded ...", not that it looks very clever
 909 * in case of head skb. Due to potentional receiver driven attacks, we
 910 * choose to avoid immediate execution of a walk in write queue due to
 911 * reneging and defer head skb's loss recovery to standard loss recovery
 912 * procedure that will eventually trigger (nothing forbids us doing this).
 913 *
 914 * Implements also blockage to start_seq wrap-around. Problem lies in the
 915 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
 916 * there's no guarantee that it will be before snd_nxt (n). The problem
 917 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
 918 * wrap (s_w):
 919 *
 920 *         <- outs wnd ->                          <- wrapzone ->
 921 *         u     e      n                         u_w   e_w  s n_w
 922 *         |     |      |                          |     |   |  |
 923 * |<------------+------+----- TCP seqno space --------------+---------->|
 924 * ...-- <2^31 ->|                                           |<--------...
 925 * ...---- >2^31 ------>|                                    |<--------...
 926 *
 927 * Current code wouldn't be vulnerable but it's better still to discard such
 928 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
 929 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
 930 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
 931 * equal to the ideal case (infinite seqno space without wrap caused issues).
 932 *
 933 * With D-SACK the lower bound is extended to cover sequence space below
 934 * SND.UNA down to undo_marker, which is the last point of interest. Yet
 935 * again, D-SACK block must not to go across snd_una (for the same reason as
 936 * for the normal SACK blocks, explained above). But there all simplicity
 937 * ends, TCP might receive valid D-SACKs below that. As long as they reside
 938 * fully below undo_marker they do not affect behavior in anyway and can
 939 * therefore be safely ignored. In rare cases (which are more or less
 940 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
 941 * fragmentation and packet reordering past skb's retransmission. To consider
 942 * them correctly, the acceptable range must be extended even more though
 943 * the exact amount is rather hard to quantify. However, tp->max_window can
 944 * be used as an exaggerated estimate.
 945 */
 946static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
 947                                   u32 start_seq, u32 end_seq)
 948{
 949        /* Too far in future, or reversed (interpretation is ambiguous) */
 950        if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
 951                return false;
 952
 953        /* Nasty start_seq wrap-around check (see comments above) */
 954        if (!before(start_seq, tp->snd_nxt))
 955                return false;
 956
 957        /* In outstanding window? ...This is valid exit for D-SACKs too.
 958         * start_seq == snd_una is non-sensical (see comments above)
 959         */
 960        if (after(start_seq, tp->snd_una))
 961                return true;
 962
 963        if (!is_dsack || !tp->undo_marker)
 964                return false;
 965
 966        /* ...Then it's D-SACK, and must reside below snd_una completely */
 967        if (after(end_seq, tp->snd_una))
 968                return false;
 969
 970        if (!before(start_seq, tp->undo_marker))
 971                return true;
 972
 973        /* Too old */
 974        if (!after(end_seq, tp->undo_marker))
 975                return false;
 976
 977        /* Undo_marker boundary crossing (overestimates a lot). Known already:
 978         *   start_seq < undo_marker and end_seq >= undo_marker.
 979         */
 980        return !before(start_seq, end_seq - tp->max_window);
 981}
 982
 983/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
 984 * Event "B". Later note: FACK people cheated me again 8), we have to account
 985 * for reordering! Ugly, but should help.
 986 *
 987 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
 988 * less than what is now known to be received by the other end (derived from
 989 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
 990 * retransmitted skbs to avoid some costly processing per ACKs.
 991 */
 992static void tcp_mark_lost_retrans(struct sock *sk)
 993{
 994        const struct inet_connection_sock *icsk = inet_csk(sk);
 995        struct tcp_sock *tp = tcp_sk(sk);
 996        struct sk_buff *skb;
 997        int cnt = 0;
 998        u32 new_low_seq = tp->snd_nxt;
 999        u32 received_upto = tcp_highest_sack_seq(tp);
1000
1001        if (!tcp_is_fack(tp) || !tp->retrans_out ||
1002            !after(received_upto, tp->lost_retrans_low) ||
1003            icsk->icsk_ca_state != TCP_CA_Recovery)
1004                return;
1005
1006        tcp_for_write_queue(skb, sk) {
1007                u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1008
1009                if (skb == tcp_send_head(sk))
1010                        break;
1011                if (cnt == tp->retrans_out)
1012                        break;
1013                if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1014                        continue;
1015
1016                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1017                        continue;
1018
1019                /* TODO: We would like to get rid of tcp_is_fack(tp) only
1020                 * constraint here (see above) but figuring out that at
1021                 * least tp->reordering SACK blocks reside between ack_seq
1022                 * and received_upto is not easy task to do cheaply with
1023                 * the available datastructures.
1024                 *
1025                 * Whether FACK should check here for tp->reordering segs
1026                 * in-between one could argue for either way (it would be
1027                 * rather simple to implement as we could count fack_count
1028                 * during the walk and do tp->fackets_out - fack_count).
1029                 */
1030                if (after(received_upto, ack_seq)) {
1031                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1032                        tp->retrans_out -= tcp_skb_pcount(skb);
1033
1034                        tcp_skb_mark_lost_uncond_verify(tp, skb);
1035                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1036                } else {
1037                        if (before(ack_seq, new_low_seq))
1038                                new_low_seq = ack_seq;
1039                        cnt += tcp_skb_pcount(skb);
1040                }
1041        }
1042
1043        if (tp->retrans_out)
1044                tp->lost_retrans_low = new_low_seq;
1045}
1046
1047static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1048                            struct tcp_sack_block_wire *sp, int num_sacks,
1049                            u32 prior_snd_una)
1050{
1051        struct tcp_sock *tp = tcp_sk(sk);
1052        u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1053        u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1054        bool dup_sack = false;
1055
1056        if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1057                dup_sack = true;
1058                tcp_dsack_seen(tp);
1059                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1060        } else if (num_sacks > 1) {
1061                u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1062                u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1063
1064                if (!after(end_seq_0, end_seq_1) &&
1065                    !before(start_seq_0, start_seq_1)) {
1066                        dup_sack = true;
1067                        tcp_dsack_seen(tp);
1068                        NET_INC_STATS_BH(sock_net(sk),
1069                                        LINUX_MIB_TCPDSACKOFORECV);
1070                }
1071        }
1072
1073        /* D-SACK for already forgotten data... Do dumb counting. */
1074        if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1075            !after(end_seq_0, prior_snd_una) &&
1076            after(end_seq_0, tp->undo_marker))
1077                tp->undo_retrans--;
1078
1079        return dup_sack;
1080}
1081
1082struct tcp_sacktag_state {
1083        int reord;
1084        int fack_count;
1085        int flag;
1086};
1087
1088/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1089 * the incoming SACK may not exactly match but we can find smaller MSS
1090 * aligned portion of it that matches. Therefore we might need to fragment
1091 * which may fail and creates some hassle (caller must handle error case
1092 * returns).
1093 *
1094 * FIXME: this could be merged to shift decision code
1095 */
1096static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1097                                  u32 start_seq, u32 end_seq)
1098{
1099        int err;
1100        bool in_sack;
1101        unsigned int pkt_len;
1102        unsigned int mss;
1103
1104        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1105                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1106
1107        if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1108            after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1109                mss = tcp_skb_mss(skb);
1110                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1111
1112                if (!in_sack) {
1113                        pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1114                        if (pkt_len < mss)
1115                                pkt_len = mss;
1116                } else {
1117                        pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1118                        if (pkt_len < mss)
1119                                return -EINVAL;
1120                }
1121
1122                /* Round if necessary so that SACKs cover only full MSSes
1123                 * and/or the remaining small portion (if present)
1124                 */
1125                if (pkt_len > mss) {
1126                        unsigned int new_len = (pkt_len / mss) * mss;
1127                        if (!in_sack && new_len < pkt_len) {
1128                                new_len += mss;
1129                                if (new_len > skb->len)
1130                                        return 0;
1131                        }
1132                        pkt_len = new_len;
1133                }
1134                err = tcp_fragment(sk, skb, pkt_len, mss);
1135                if (err < 0)
1136                        return err;
1137        }
1138
1139        return in_sack;
1140}
1141
1142/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1143static u8 tcp_sacktag_one(struct sock *sk,
1144                          struct tcp_sacktag_state *state, u8 sacked,
1145                          u32 start_seq, u32 end_seq,
1146                          bool dup_sack, int pcount)
1147{
1148        struct tcp_sock *tp = tcp_sk(sk);
1149        int fack_count = state->fack_count;
1150
1151        /* Account D-SACK for retransmitted packet. */
1152        if (dup_sack && (sacked & TCPCB_RETRANS)) {
1153                if (tp->undo_marker && tp->undo_retrans &&
1154                    after(end_seq, tp->undo_marker))
1155                        tp->undo_retrans--;
1156                if (sacked & TCPCB_SACKED_ACKED)
1157                        state->reord = min(fack_count, state->reord);
1158        }
1159
1160        /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1161        if (!after(end_seq, tp->snd_una))
1162                return sacked;
1163
1164        if (!(sacked & TCPCB_SACKED_ACKED)) {
1165                if (sacked & TCPCB_SACKED_RETRANS) {
1166                        /* If the segment is not tagged as lost,
1167                         * we do not clear RETRANS, believing
1168                         * that retransmission is still in flight.
1169                         */
1170                        if (sacked & TCPCB_LOST) {
1171                                sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1172                                tp->lost_out -= pcount;
1173                                tp->retrans_out -= pcount;
1174                        }
1175                } else {
1176                        if (!(sacked & TCPCB_RETRANS)) {
1177                                /* New sack for not retransmitted frame,
1178                                 * which was in hole. It is reordering.
1179                                 */
1180                                if (before(start_seq,
1181                                           tcp_highest_sack_seq(tp)))
1182                                        state->reord = min(fack_count,
1183                                                           state->reord);
1184
1185                                /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1186                                if (!after(end_seq, tp->frto_highmark))
1187                                        state->flag |= FLAG_ONLY_ORIG_SACKED;
1188                        }
1189
1190                        if (sacked & TCPCB_LOST) {
1191                                sacked &= ~TCPCB_LOST;
1192                                tp->lost_out -= pcount;
1193                        }
1194                }
1195
1196                sacked |= TCPCB_SACKED_ACKED;
1197                state->flag |= FLAG_DATA_SACKED;
1198                tp->sacked_out += pcount;
1199
1200                fack_count += pcount;
1201
1202                /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1203                if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1204                    before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1205                        tp->lost_cnt_hint += pcount;
1206
1207                if (fack_count > tp->fackets_out)
1208                        tp->fackets_out = fack_count;
1209        }
1210
1211        /* D-SACK. We can detect redundant retransmission in S|R and plain R
1212         * frames and clear it. undo_retrans is decreased above, L|R frames
1213         * are accounted above as well.
1214         */
1215        if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1216                sacked &= ~TCPCB_SACKED_RETRANS;
1217                tp->retrans_out -= pcount;
1218        }
1219
1220        return sacked;
1221}
1222
1223/* Shift newly-SACKed bytes from this skb to the immediately previous
1224 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1225 */
1226static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1227                            struct tcp_sacktag_state *state,
1228                            unsigned int pcount, int shifted, int mss,
1229                            bool dup_sack)
1230{
1231        struct tcp_sock *tp = tcp_sk(sk);
1232        struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1233        u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1234        u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1235
1236        BUG_ON(!pcount);
1237
1238        /* Adjust counters and hints for the newly sacked sequence
1239         * range but discard the return value since prev is already
1240         * marked. We must tag the range first because the seq
1241         * advancement below implicitly advances
1242         * tcp_highest_sack_seq() when skb is highest_sack.
1243         */
1244        tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1245                        start_seq, end_seq, dup_sack, pcount);
1246
1247        if (skb == tp->lost_skb_hint)
1248                tp->lost_cnt_hint += pcount;
1249
1250        TCP_SKB_CB(prev)->end_seq += shifted;
1251        TCP_SKB_CB(skb)->seq += shifted;
1252
1253        skb_shinfo(prev)->gso_segs += pcount;
1254        BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1255        skb_shinfo(skb)->gso_segs -= pcount;
1256
1257        /* When we're adding to gso_segs == 1, gso_size will be zero,
1258         * in theory this shouldn't be necessary but as long as DSACK
1259         * code can come after this skb later on it's better to keep
1260         * setting gso_size to something.
1261         */
1262        if (!skb_shinfo(prev)->gso_size) {
1263                skb_shinfo(prev)->gso_size = mss;
1264                skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1265        }
1266
1267        /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1268        if (skb_shinfo(skb)->gso_segs <= 1) {
1269                skb_shinfo(skb)->gso_size = 0;
1270                skb_shinfo(skb)->gso_type = 0;
1271        }
1272
1273        /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1274        TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1275
1276        if (skb->len > 0) {
1277                BUG_ON(!tcp_skb_pcount(skb));
1278                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1279                return false;
1280        }
1281
1282        /* Whole SKB was eaten :-) */
1283
1284        if (skb == tp->retransmit_skb_hint)
1285                tp->retransmit_skb_hint = prev;
1286        if (skb == tp->scoreboard_skb_hint)
1287                tp->scoreboard_skb_hint = prev;
1288        if (skb == tp->lost_skb_hint) {
1289                tp->lost_skb_hint = prev;
1290                tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1291        }
1292
1293        TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1294        if (skb == tcp_highest_sack(sk))
1295                tcp_advance_highest_sack(sk, skb);
1296
1297        tcp_unlink_write_queue(skb, sk);
1298        sk_wmem_free_skb(sk, skb);
1299
1300        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1301
1302        return true;
1303}
1304
1305/* I wish gso_size would have a bit more sane initialization than
1306 * something-or-zero which complicates things
1307 */
1308static int tcp_skb_seglen(const struct sk_buff *skb)
1309{
1310        return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1311}
1312
1313/* Shifting pages past head area doesn't work */
1314static int skb_can_shift(const struct sk_buff *skb)
1315{
1316        return !skb_headlen(skb) && skb_is_nonlinear(skb);
1317}
1318
1319/* Try collapsing SACK blocks spanning across multiple skbs to a single
1320 * skb.
1321 */
1322static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1323                                          struct tcp_sacktag_state *state,
1324                                          u32 start_seq, u32 end_seq,
1325                                          bool dup_sack)
1326{
1327        struct tcp_sock *tp = tcp_sk(sk);
1328        struct sk_buff *prev;
1329        int mss;
1330        int pcount = 0;
1331        int len;
1332        int in_sack;
1333
1334        if (!sk_can_gso(sk))
1335                goto fallback;
1336
1337        /* Normally R but no L won't result in plain S */
1338        if (!dup_sack &&
1339            (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1340                goto fallback;
1341        if (!skb_can_shift(skb))
1342                goto fallback;
1343        /* This frame is about to be dropped (was ACKed). */
1344        if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1345                goto fallback;
1346
1347        /* Can only happen with delayed DSACK + discard craziness */
1348        if (unlikely(skb == tcp_write_queue_head(sk)))
1349                goto fallback;
1350        prev = tcp_write_queue_prev(sk, skb);
1351
1352        if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1353                goto fallback;
1354
1355        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1356                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1357
1358        if (in_sack) {
1359                len = skb->len;
1360                pcount = tcp_skb_pcount(skb);
1361                mss = tcp_skb_seglen(skb);
1362
1363                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1364                 * drop this restriction as unnecessary
1365                 */
1366                if (mss != tcp_skb_seglen(prev))
1367                        goto fallback;
1368        } else {
1369                if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1370                        goto noop;
1371                /* CHECKME: This is non-MSS split case only?, this will
1372                 * cause skipped skbs due to advancing loop btw, original
1373                 * has that feature too
1374                 */
1375                if (tcp_skb_pcount(skb) <= 1)
1376                        goto noop;
1377
1378                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1379                if (!in_sack) {
1380                        /* TODO: head merge to next could be attempted here
1381                         * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1382                         * though it might not be worth of the additional hassle
1383                         *
1384                         * ...we can probably just fallback to what was done
1385                         * previously. We could try merging non-SACKed ones
1386                         * as well but it probably isn't going to buy off
1387                         * because later SACKs might again split them, and
1388                         * it would make skb timestamp tracking considerably
1389                         * harder problem.
1390                         */
1391                        goto fallback;
1392                }
1393
1394                len = end_seq - TCP_SKB_CB(skb)->seq;
1395                BUG_ON(len < 0);
1396                BUG_ON(len > skb->len);
1397
1398                /* MSS boundaries should be honoured or else pcount will
1399                 * severely break even though it makes things bit trickier.
1400                 * Optimize common case to avoid most of the divides
1401                 */
1402                mss = tcp_skb_mss(skb);
1403
1404                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1405                 * drop this restriction as unnecessary
1406                 */
1407                if (mss != tcp_skb_seglen(prev))
1408                        goto fallback;
1409
1410                if (len == mss) {
1411                        pcount = 1;
1412                } else if (len < mss) {
1413                        goto noop;
1414                } else {
1415                        pcount = len / mss;
1416                        len = pcount * mss;
1417                }
1418        }
1419
1420        /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1421        if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1422                goto fallback;
1423
1424        if (!skb_shift(prev, skb, len))
1425                goto fallback;
1426        if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1427                goto out;
1428
1429        /* Hole filled allows collapsing with the next as well, this is very
1430         * useful when hole on every nth skb pattern happens
1431         */
1432        if (prev == tcp_write_queue_tail(sk))
1433                goto out;
1434        skb = tcp_write_queue_next(sk, prev);
1435
1436        if (!skb_can_shift(skb) ||
1437            (skb == tcp_send_head(sk)) ||
1438            ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1439            (mss != tcp_skb_seglen(skb)))
1440                goto out;
1441
1442        len = skb->len;
1443        if (skb_shift(prev, skb, len)) {
1444                pcount += tcp_skb_pcount(skb);
1445                tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1446        }
1447
1448out:
1449        state->fack_count += pcount;
1450        return prev;
1451
1452noop:
1453        return skb;
1454
1455fallback:
1456        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1457        return NULL;
1458}
1459
1460static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1461                                        struct tcp_sack_block *next_dup,
1462                                        struct tcp_sacktag_state *state,
1463                                        u32 start_seq, u32 end_seq,
1464                                        bool dup_sack_in)
1465{
1466        struct tcp_sock *tp = tcp_sk(sk);
1467        struct sk_buff *tmp;
1468
1469        tcp_for_write_queue_from(skb, sk) {
1470                int in_sack = 0;
1471                bool dup_sack = dup_sack_in;
1472
1473                if (skb == tcp_send_head(sk))
1474                        break;
1475
1476                /* queue is in-order => we can short-circuit the walk early */
1477                if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1478                        break;
1479
1480                if ((next_dup != NULL) &&
1481                    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1482                        in_sack = tcp_match_skb_to_sack(sk, skb,
1483                                                        next_dup->start_seq,
1484                                                        next_dup->end_seq);
1485                        if (in_sack > 0)
1486                                dup_sack = true;
1487                }
1488
1489                /* skb reference here is a bit tricky to get right, since
1490                 * shifting can eat and free both this skb and the next,
1491                 * so not even _safe variant of the loop is enough.
1492                 */
1493                if (in_sack <= 0) {
1494                        tmp = tcp_shift_skb_data(sk, skb, state,
1495                                                 start_seq, end_seq, dup_sack);
1496                        if (tmp != NULL) {
1497                                if (tmp != skb) {
1498                                        skb = tmp;
1499                                        continue;
1500                                }
1501
1502                                in_sack = 0;
1503                        } else {
1504                                in_sack = tcp_match_skb_to_sack(sk, skb,
1505                                                                start_seq,
1506                                                                end_seq);
1507                        }
1508                }
1509
1510                if (unlikely(in_sack < 0))
1511                        break;
1512
1513                if (in_sack) {
1514                        TCP_SKB_CB(skb)->sacked =
1515                                tcp_sacktag_one(sk,
1516                                                state,
1517                                                TCP_SKB_CB(skb)->sacked,
1518                                                TCP_SKB_CB(skb)->seq,
1519                                                TCP_SKB_CB(skb)->end_seq,
1520                                                dup_sack,
1521                                                tcp_skb_pcount(skb));
1522
1523                        if (!before(TCP_SKB_CB(skb)->seq,
1524                                    tcp_highest_sack_seq(tp)))
1525                                tcp_advance_highest_sack(sk, skb);
1526                }
1527
1528                state->fack_count += tcp_skb_pcount(skb);
1529        }
1530        return skb;
1531}
1532
1533/* Avoid all extra work that is being done by sacktag while walking in
1534 * a normal way
1535 */
1536static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1537                                        struct tcp_sacktag_state *state,
1538                                        u32 skip_to_seq)
1539{
1540        tcp_for_write_queue_from(skb, sk) {
1541                if (skb == tcp_send_head(sk))
1542                        break;
1543
1544                if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1545                        break;
1546
1547                state->fack_count += tcp_skb_pcount(skb);
1548        }
1549        return skb;
1550}
1551
1552static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1553                                                struct sock *sk,
1554                                                struct tcp_sack_block *next_dup,
1555                                                struct tcp_sacktag_state *state,
1556                                                u32 skip_to_seq)
1557{
1558        if (next_dup == NULL)
1559                return skb;
1560
1561        if (before(next_dup->start_seq, skip_to_seq)) {
1562                skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1563                skb = tcp_sacktag_walk(skb, sk, NULL, state,
1564                                       next_dup->start_seq, next_dup->end_seq,
1565                                       1);
1566        }
1567
1568        return skb;
1569}
1570
1571static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1572{
1573        return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1574}
1575
1576static int
1577tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1578                        u32 prior_snd_una)
1579{
1580        const struct inet_connection_sock *icsk = inet_csk(sk);
1581        struct tcp_sock *tp = tcp_sk(sk);
1582        const unsigned char *ptr = (skb_transport_header(ack_skb) +
1583                                    TCP_SKB_CB(ack_skb)->sacked);
1584        struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1585        struct tcp_sack_block sp[TCP_NUM_SACKS];
1586        struct tcp_sack_block *cache;
1587        struct tcp_sacktag_state state;
1588        struct sk_buff *skb;
1589        int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1590        int used_sacks;
1591        bool found_dup_sack = false;
1592        int i, j;
1593        int first_sack_index;
1594
1595        state.flag = 0;
1596        state.reord = tp->packets_out;
1597
1598        if (!tp->sacked_out) {
1599                if (WARN_ON(tp->fackets_out))
1600                        tp->fackets_out = 0;
1601                tcp_highest_sack_reset(sk);
1602        }
1603
1604        found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1605                                         num_sacks, prior_snd_una);
1606        if (found_dup_sack)
1607                state.flag |= FLAG_DSACKING_ACK;
1608
1609        /* Eliminate too old ACKs, but take into
1610         * account more or less fresh ones, they can
1611         * contain valid SACK info.
1612         */
1613        if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1614                return 0;
1615
1616        if (!tp->packets_out)
1617                goto out;
1618
1619        used_sacks = 0;
1620        first_sack_index = 0;
1621        for (i = 0; i < num_sacks; i++) {
1622                bool dup_sack = !i && found_dup_sack;
1623
1624                sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1625                sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1626
1627                if (!tcp_is_sackblock_valid(tp, dup_sack,
1628                                            sp[used_sacks].start_seq,
1629                                            sp[used_sacks].end_seq)) {
1630                        int mib_idx;
1631
1632                        if (dup_sack) {
1633                                if (!tp->undo_marker)
1634                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1635                                else
1636                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1637                        } else {
1638                                /* Don't count olds caused by ACK reordering */
1639                                if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1640                                    !after(sp[used_sacks].end_seq, tp->snd_una))
1641                                        continue;
1642                                mib_idx = LINUX_MIB_TCPSACKDISCARD;
1643                        }
1644
1645                        NET_INC_STATS_BH(sock_net(sk), mib_idx);
1646                        if (i == 0)
1647                                first_sack_index = -1;
1648                        continue;
1649                }
1650
1651                /* Ignore very old stuff early */
1652                if (!after(sp[used_sacks].end_seq, prior_snd_una))
1653                        continue;
1654
1655                used_sacks++;
1656        }
1657
1658        /* order SACK blocks to allow in order walk of the retrans queue */
1659        for (i = used_sacks - 1; i > 0; i--) {
1660                for (j = 0; j < i; j++) {
1661                        if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1662                                swap(sp[j], sp[j + 1]);
1663
1664                                /* Track where the first SACK block goes to */
1665                                if (j == first_sack_index)
1666                                        first_sack_index = j + 1;
1667                        }
1668                }
1669        }
1670
1671        skb = tcp_write_queue_head(sk);
1672        state.fack_count = 0;
1673        i = 0;
1674
1675        if (!tp->sacked_out) {
1676                /* It's already past, so skip checking against it */
1677                cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1678        } else {
1679                cache = tp->recv_sack_cache;
1680                /* Skip empty blocks in at head of the cache */
1681                while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1682                       !cache->end_seq)
1683                        cache++;
1684        }
1685
1686        while (i < used_sacks) {
1687                u32 start_seq = sp[i].start_seq;
1688                u32 end_seq = sp[i].end_seq;
1689                bool dup_sack = (found_dup_sack && (i == first_sack_index));
1690                struct tcp_sack_block *next_dup = NULL;
1691
1692                if (found_dup_sack && ((i + 1) == first_sack_index))
1693                        next_dup = &sp[i + 1];
1694
1695                /* Skip too early cached blocks */
1696                while (tcp_sack_cache_ok(tp, cache) &&
1697                       !before(start_seq, cache->end_seq))
1698                        cache++;
1699
1700                /* Can skip some work by looking recv_sack_cache? */
1701                if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1702                    after(end_seq, cache->start_seq)) {
1703
1704                        /* Head todo? */
1705                        if (before(start_seq, cache->start_seq)) {
1706                                skb = tcp_sacktag_skip(skb, sk, &state,
1707                                                       start_seq);
1708                                skb = tcp_sacktag_walk(skb, sk, next_dup,
1709                                                       &state,
1710                                                       start_seq,
1711                                                       cache->start_seq,
1712                                                       dup_sack);
1713                        }
1714
1715                        /* Rest of the block already fully processed? */
1716                        if (!after(end_seq, cache->end_seq))
1717                                goto advance_sp;
1718
1719                        skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1720                                                       &state,
1721                                                       cache->end_seq);
1722
1723                        /* ...tail remains todo... */
1724                        if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1725                                /* ...but better entrypoint exists! */
1726                                skb = tcp_highest_sack(sk);
1727                                if (skb == NULL)
1728                                        break;
1729                                state.fack_count = tp->fackets_out;
1730                                cache++;
1731                                goto walk;
1732                        }
1733
1734                        skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1735                        /* Check overlap against next cached too (past this one already) */
1736                        cache++;
1737                        continue;
1738                }
1739
1740                if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1741                        skb = tcp_highest_sack(sk);
1742                        if (skb == NULL)
1743                                break;
1744                        state.fack_count = tp->fackets_out;
1745                }
1746                skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1747
1748walk:
1749                skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1750                                       start_seq, end_seq, dup_sack);
1751
1752advance_sp:
1753                /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1754                 * due to in-order walk
1755                 */
1756                if (after(end_seq, tp->frto_highmark))
1757                        state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1758
1759                i++;
1760        }
1761
1762        /* Clear the head of the cache sack blocks so we can skip it next time */
1763        for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1764                tp->recv_sack_cache[i].start_seq = 0;
1765                tp->recv_sack_cache[i].end_seq = 0;
1766        }
1767        for (j = 0; j < used_sacks; j++)
1768                tp->recv_sack_cache[i++] = sp[j];
1769
1770        tcp_mark_lost_retrans(sk);
1771
1772        tcp_verify_left_out(tp);
1773
1774        if ((state.reord < tp->fackets_out) &&
1775            ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1776            (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1777                tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1778
1779out:
1780
1781#if FASTRETRANS_DEBUG > 0
1782        WARN_ON((int)tp->sacked_out < 0);
1783        WARN_ON((int)tp->lost_out < 0);
1784        WARN_ON((int)tp->retrans_out < 0);
1785        WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1786#endif
1787        return state.flag;
1788}
1789
1790/* Limits sacked_out so that sum with lost_out isn't ever larger than
1791 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1792 */
1793static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1794{
1795        u32 holes;
1796
1797        holes = max(tp->lost_out, 1U);
1798        holes = min(holes, tp->packets_out);
1799
1800        if ((tp->sacked_out + holes) > tp->packets_out) {
1801                tp->sacked_out = tp->packets_out - holes;
1802                return true;
1803        }
1804        return false;
1805}
1806
1807/* If we receive more dupacks than we expected counting segments
1808 * in assumption of absent reordering, interpret this as reordering.
1809 * The only another reason could be bug in receiver TCP.
1810 */
1811static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1812{
1813        struct tcp_sock *tp = tcp_sk(sk);
1814        if (tcp_limit_reno_sacked(tp))
1815                tcp_update_reordering(sk, tp->packets_out + addend, 0);
1816}
1817
1818/* Emulate SACKs for SACKless connection: account for a new dupack. */
1819
1820static void tcp_add_reno_sack(struct sock *sk)
1821{
1822        struct tcp_sock *tp = tcp_sk(sk);
1823        tp->sacked_out++;
1824        tcp_check_reno_reordering(sk, 0);
1825        tcp_verify_left_out(tp);
1826}
1827
1828/* Account for ACK, ACKing some data in Reno Recovery phase. */
1829
1830static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1831{
1832        struct tcp_sock *tp = tcp_sk(sk);
1833
1834        if (acked > 0) {
1835                /* One ACK acked hole. The rest eat duplicate ACKs. */
1836                if (acked - 1 >= tp->sacked_out)
1837                        tp->sacked_out = 0;
1838                else
1839                        tp->sacked_out -= acked - 1;
1840        }
1841        tcp_check_reno_reordering(sk, acked);
1842        tcp_verify_left_out(tp);
1843}
1844
1845static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1846{
1847        tp->sacked_out = 0;
1848}
1849
1850static int tcp_is_sackfrto(const struct tcp_sock *tp)
1851{
1852        return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1853}
1854
1855/* F-RTO can only be used if TCP has never retransmitted anything other than
1856 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1857 */
1858bool tcp_use_frto(struct sock *sk)
1859{
1860        const struct tcp_sock *tp = tcp_sk(sk);
1861        const struct inet_connection_sock *icsk = inet_csk(sk);
1862        struct sk_buff *skb;
1863
1864        if (!sysctl_tcp_frto)
1865                return false;
1866
1867        /* MTU probe and F-RTO won't really play nicely along currently */
1868        if (icsk->icsk_mtup.probe_size)
1869                return false;
1870
1871        if (tcp_is_sackfrto(tp))
1872                return true;
1873
1874        /* Avoid expensive walking of rexmit queue if possible */
1875        if (tp->retrans_out > 1)
1876                return false;
1877
1878        skb = tcp_write_queue_head(sk);
1879        if (tcp_skb_is_last(sk, skb))
1880                return true;
1881        skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1882        tcp_for_write_queue_from(skb, sk) {
1883                if (skb == tcp_send_head(sk))
1884                        break;
1885                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1886                        return false;
1887                /* Short-circuit when first non-SACKed skb has been checked */
1888                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1889                        break;
1890        }
1891        return true;
1892}
1893
1894/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1895 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1896 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1897 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1898 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1899 * bits are handled if the Loss state is really to be entered (in
1900 * tcp_enter_frto_loss).
1901 *
1902 * Do like tcp_enter_loss() would; when RTO expires the second time it
1903 * does:
1904 *  "Reduce ssthresh if it has not yet been made inside this window."
1905 */
1906void tcp_enter_frto(struct sock *sk)
1907{
1908        const struct inet_connection_sock *icsk = inet_csk(sk);
1909        struct tcp_sock *tp = tcp_sk(sk);
1910        struct sk_buff *skb;
1911
1912        if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1913            tp->snd_una == tp->high_seq ||
1914            ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1915             !icsk->icsk_retransmits)) {
1916                tp->prior_ssthresh = tcp_current_ssthresh(sk);
1917                /* Our state is too optimistic in ssthresh() call because cwnd
1918                 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1919                 * recovery has not yet completed. Pattern would be this: RTO,
1920                 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1921                 * up here twice).
1922                 * RFC4138 should be more specific on what to do, even though
1923                 * RTO is quite unlikely to occur after the first Cumulative ACK
1924                 * due to back-off and complexity of triggering events ...
1925                 */
1926                if (tp->frto_counter) {
1927                        u32 stored_cwnd;
1928                        stored_cwnd = tp->snd_cwnd;
1929                        tp->snd_cwnd = 2;
1930                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1931                        tp->snd_cwnd = stored_cwnd;
1932                } else {
1933                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1934                }
1935                /* ... in theory, cong.control module could do "any tricks" in
1936                 * ssthresh(), which means that ca_state, lost bits and lost_out
1937                 * counter would have to be faked before the call occurs. We
1938                 * consider that too expensive, unlikely and hacky, so modules
1939                 * using these in ssthresh() must deal these incompatibility
1940                 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1941                 */
1942                tcp_ca_event(sk, CA_EVENT_FRTO);
1943        }
1944
1945        tp->undo_marker = tp->snd_una;
1946        tp->undo_retrans = 0;
1947
1948        skb = tcp_write_queue_head(sk);
1949        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1950                tp->undo_marker = 0;
1951        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1952                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1953                tp->retrans_out -= tcp_skb_pcount(skb);
1954        }
1955        tcp_verify_left_out(tp);
1956
1957        /* Too bad if TCP was application limited */
1958        tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1959
1960        /* Earlier loss recovery underway (see RFC4138; Appendix B).
1961         * The last condition is necessary at least in tp->frto_counter case.
1962         */
1963        if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1964            ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1965            after(tp->high_seq, tp->snd_una)) {
1966                tp->frto_highmark = tp->high_seq;
1967        } else {
1968                tp->frto_highmark = tp->snd_nxt;
1969        }
1970        tcp_set_ca_state(sk, TCP_CA_Disorder);
1971        tp->high_seq = tp->snd_nxt;
1972        tp->frto_counter = 1;
1973}
1974
1975/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1976 * which indicates that we should follow the traditional RTO recovery,
1977 * i.e. mark everything lost and do go-back-N retransmission.
1978 */
1979static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1980{
1981        struct tcp_sock *tp = tcp_sk(sk);
1982        struct sk_buff *skb;
1983
1984        tp->lost_out = 0;
1985        tp->retrans_out = 0;
1986        if (tcp_is_reno(tp))
1987                tcp_reset_reno_sack(tp);
1988
1989        tcp_for_write_queue(skb, sk) {
1990                if (skb == tcp_send_head(sk))
1991                        break;
1992
1993                TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1994                /*
1995                 * Count the retransmission made on RTO correctly (only when
1996                 * waiting for the first ACK and did not get it)...
1997                 */
1998                if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1999                        /* For some reason this R-bit might get cleared? */
2000                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2001                                tp->retrans_out += tcp_skb_pcount(skb);
2002                        /* ...enter this if branch just for the first segment */
2003                        flag |= FLAG_DATA_ACKED;
2004                } else {
2005                        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2006                                tp->undo_marker = 0;
2007                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2008                }
2009
2010                /* Marking forward transmissions that were made after RTO lost
2011                 * can cause unnecessary retransmissions in some scenarios,
2012                 * SACK blocks will mitigate that in some but not in all cases.
2013                 * We used to not mark them but it was causing break-ups with
2014                 * receivers that do only in-order receival.
2015                 *
2016                 * TODO: we could detect presence of such receiver and select
2017                 * different behavior per flow.
2018                 */
2019                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2020                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2021                        tp->lost_out += tcp_skb_pcount(skb);
2022                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2023                }
2024        }
2025        tcp_verify_left_out(tp);
2026
2027        tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2028        tp->snd_cwnd_cnt = 0;
2029        tp->snd_cwnd_stamp = tcp_time_stamp;
2030        tp->frto_counter = 0;
2031        tp->bytes_acked = 0;
2032
2033        tp->reordering = min_t(unsigned int, tp->reordering,
2034                               sysctl_tcp_reordering);
2035        tcp_set_ca_state(sk, TCP_CA_Loss);
2036        tp->high_seq = tp->snd_nxt;
2037        TCP_ECN_queue_cwr(tp);
2038
2039        tcp_clear_all_retrans_hints(tp);
2040}
2041
2042static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2043{
2044        tp->retrans_out = 0;
2045        tp->lost_out = 0;
2046
2047        tp->undo_marker = 0;
2048        tp->undo_retrans = 0;
2049}
2050
2051void tcp_clear_retrans(struct tcp_sock *tp)
2052{
2053        tcp_clear_retrans_partial(tp);
2054
2055        tp->fackets_out = 0;
2056        tp->sacked_out = 0;
2057}
2058
2059/* Enter Loss state. If "how" is not zero, forget all SACK information
2060 * and reset tags completely, otherwise preserve SACKs. If receiver
2061 * dropped its ofo queue, we will know this due to reneging detection.
2062 */
2063void tcp_enter_loss(struct sock *sk, int how)
2064{
2065        const struct inet_connection_sock *icsk = inet_csk(sk);
2066        struct tcp_sock *tp = tcp_sk(sk);
2067        struct sk_buff *skb;
2068
2069        /* Reduce ssthresh if it has not yet been made inside this window. */
2070        if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2071            (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2072                tp->prior_ssthresh = tcp_current_ssthresh(sk);
2073                tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2074                tcp_ca_event(sk, CA_EVENT_LOSS);
2075        }
2076        tp->snd_cwnd       = 1;
2077        tp->snd_cwnd_cnt   = 0;
2078        tp->snd_cwnd_stamp = tcp_time_stamp;
2079
2080        tp->bytes_acked = 0;
2081        tcp_clear_retrans_partial(tp);
2082
2083        if (tcp_is_reno(tp))
2084                tcp_reset_reno_sack(tp);
2085
2086        if (!how) {
2087                /* Push undo marker, if it was plain RTO and nothing
2088                 * was retransmitted. */
2089                tp->undo_marker = tp->snd_una;
2090        } else {
2091                tp->sacked_out = 0;
2092                tp->fackets_out = 0;
2093        }
2094        tcp_clear_all_retrans_hints(tp);
2095
2096        tcp_for_write_queue(skb, sk) {
2097                if (skb == tcp_send_head(sk))
2098                        break;
2099
2100                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2101                        tp->undo_marker = 0;
2102                TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2103                if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2104                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2105                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2106                        tp->lost_out += tcp_skb_pcount(skb);
2107                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2108                }
2109        }
2110        tcp_verify_left_out(tp);
2111
2112        tp->reordering = min_t(unsigned int, tp->reordering,
2113                               sysctl_tcp_reordering);
2114        tcp_set_ca_state(sk, TCP_CA_Loss);
2115        tp->high_seq = tp->snd_nxt;
2116        TCP_ECN_queue_cwr(tp);
2117        /* Abort F-RTO algorithm if one is in progress */
2118        tp->frto_counter = 0;
2119}
2120
2121/* If ACK arrived pointing to a remembered SACK, it means that our
2122 * remembered SACKs do not reflect real state of receiver i.e.
2123 * receiver _host_ is heavily congested (or buggy).
2124 *
2125 * Do processing similar to RTO timeout.
2126 */
2127static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2128{
2129        if (flag & FLAG_SACK_RENEGING) {
2130                struct inet_connection_sock *icsk = inet_csk(sk);
2131                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2132
2133                tcp_enter_loss(sk, 1);
2134                icsk->icsk_retransmits++;
2135                tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2136                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2137                                          icsk->icsk_rto, TCP_RTO_MAX);
2138                return true;
2139        }
2140        return false;
2141}
2142
2143static inline int tcp_fackets_out(const struct tcp_sock *tp)
2144{
2145        return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2146}
2147
2148/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2149 * counter when SACK is enabled (without SACK, sacked_out is used for
2150 * that purpose).
2151 *
2152 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2153 * segments up to the highest received SACK block so far and holes in
2154 * between them.
2155 *
2156 * With reordering, holes may still be in flight, so RFC3517 recovery
2157 * uses pure sacked_out (total number of SACKed segments) even though
2158 * it violates the RFC that uses duplicate ACKs, often these are equal
2159 * but when e.g. out-of-window ACKs or packet duplication occurs,
2160 * they differ. Since neither occurs due to loss, TCP should really
2161 * ignore them.
2162 */
2163static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2164{
2165        return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2166}
2167
2168static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2169{
2170        struct tcp_sock *tp = tcp_sk(sk);
2171        unsigned long delay;
2172
2173        /* Delay early retransmit and entering fast recovery for
2174         * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2175         * available, or RTO is scheduled to fire first.
2176         */
2177        if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2178                return false;
2179
2180        delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2181        if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2182                return false;
2183
2184        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2185        tp->early_retrans_delayed = 1;
2186        return true;
2187}
2188
2189static inline int tcp_skb_timedout(const struct sock *sk,
2190                                   const struct sk_buff *skb)
2191{
2192        return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2193}
2194
2195static inline int tcp_head_timedout(const struct sock *sk)
2196{
2197        const struct tcp_sock *tp = tcp_sk(sk);
2198
2199        return tp->packets_out &&
2200               tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2201}
2202
2203/* Linux NewReno/SACK/FACK/ECN state machine.
2204 * --------------------------------------
2205 *
2206 * "Open"       Normal state, no dubious events, fast path.
2207 * "Disorder"   In all the respects it is "Open",
2208 *              but requires a bit more attention. It is entered when
2209 *              we see some SACKs or dupacks. It is split of "Open"
2210 *              mainly to move some processing from fast path to slow one.
2211 * "CWR"        CWND was reduced due to some Congestion Notification event.
2212 *              It can be ECN, ICMP source quench, local device congestion.
2213 * "Recovery"   CWND was reduced, we are fast-retransmitting.
2214 * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2215 *
2216 * tcp_fastretrans_alert() is entered:
2217 * - each incoming ACK, if state is not "Open"
2218 * - when arrived ACK is unusual, namely:
2219 *      * SACK
2220 *      * Duplicate ACK.
2221 *      * ECN ECE.
2222 *
2223 * Counting packets in flight is pretty simple.
2224 *
2225 *      in_flight = packets_out - left_out + retrans_out
2226 *
2227 *      packets_out is SND.NXT-SND.UNA counted in packets.
2228 *
2229 *      retrans_out is number of retransmitted segments.
2230 *
2231 *      left_out is number of segments left network, but not ACKed yet.
2232 *
2233 *              left_out = sacked_out + lost_out
2234 *
2235 *     sacked_out: Packets, which arrived to receiver out of order
2236 *                 and hence not ACKed. With SACKs this number is simply
2237 *                 amount of SACKed data. Even without SACKs
2238 *                 it is easy to give pretty reliable estimate of this number,
2239 *                 counting duplicate ACKs.
2240 *
2241 *       lost_out: Packets lost by network. TCP has no explicit
2242 *                 "loss notification" feedback from network (for now).
2243 *                 It means that this number can be only _guessed_.
2244 *                 Actually, it is the heuristics to predict lossage that
2245 *                 distinguishes different algorithms.
2246 *
2247 *      F.e. after RTO, when all the queue is considered as lost,
2248 *      lost_out = packets_out and in_flight = retrans_out.
2249 *
2250 *              Essentially, we have now two algorithms counting
2251 *              lost packets.
2252 *
2253 *              FACK: It is the simplest heuristics. As soon as we decided
2254 *              that something is lost, we decide that _all_ not SACKed
2255 *              packets until the most forward SACK are lost. I.e.
2256 *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2257 *              It is absolutely correct estimate, if network does not reorder
2258 *              packets. And it loses any connection to reality when reordering
2259 *              takes place. We use FACK by default until reordering
2260 *              is suspected on the path to this destination.
2261 *
2262 *              NewReno: when Recovery is entered, we assume that one segment
2263 *              is lost (classic Reno). While we are in Recovery and
2264 *              a partial ACK arrives, we assume that one more packet
2265 *              is lost (NewReno). This heuristics are the same in NewReno
2266 *              and SACK.
2267 *
2268 *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2269 *  deflation etc. CWND is real congestion window, never inflated, changes
2270 *  only according to classic VJ rules.
2271 *
2272 * Really tricky (and requiring careful tuning) part of algorithm
2273 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2274 * The first determines the moment _when_ we should reduce CWND and,
2275 * hence, slow down forward transmission. In fact, it determines the moment
2276 * when we decide that hole is caused by loss, rather than by a reorder.
2277 *
2278 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2279 * holes, caused by lost packets.
2280 *
2281 * And the most logically complicated part of algorithm is undo
2282 * heuristics. We detect false retransmits due to both too early
2283 * fast retransmit (reordering) and underestimated RTO, analyzing
2284 * timestamps and D-SACKs. When we detect that some segments were
2285 * retransmitted by mistake and CWND reduction was wrong, we undo
2286 * window reduction and abort recovery phase. This logic is hidden
2287 * inside several functions named tcp_try_undo_<something>.
2288 */
2289
2290/* This function decides, when we should leave Disordered state
2291 * and enter Recovery phase, reducing congestion window.
2292 *
2293 * Main question: may we further continue forward transmission
2294 * with the same cwnd?
2295 */
2296static bool tcp_time_to_recover(struct sock *sk, int flag)
2297{
2298        struct tcp_sock *tp = tcp_sk(sk);
2299        __u32 packets_out;
2300
2301        /* Do not perform any recovery during F-RTO algorithm */
2302        if (tp->frto_counter)
2303                return false;
2304
2305        /* Trick#1: The loss is proven. */
2306        if (tp->lost_out)
2307                return true;
2308
2309        /* Not-A-Trick#2 : Classic rule... */
2310        if (tcp_dupack_heuristics(tp) > tp->reordering)
2311                return true;
2312
2313        /* Trick#3 : when we use RFC2988 timer restart, fast
2314         * retransmit can be triggered by timeout of queue head.
2315         */
2316        if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2317                return true;
2318
2319        /* Trick#4: It is still not OK... But will it be useful to delay
2320         * recovery more?
2321         */
2322        packets_out = tp->packets_out;
2323        if (packets_out <= tp->reordering &&
2324            tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2325            !tcp_may_send_now(sk)) {
2326                /* We have nothing to send. This connection is limited
2327                 * either by receiver window or by application.
2328                 */
2329                return true;
2330        }
2331
2332        /* If a thin stream is detected, retransmit after first
2333         * received dupack. Employ only if SACK is supported in order
2334         * to avoid possible corner-case series of spurious retransmissions
2335         * Use only if there are no unsent data.
2336         */
2337        if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2338            tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2339            tcp_is_sack(tp) && !tcp_send_head(sk))
2340                return true;
2341
2342        /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2343         * retransmissions due to small network reorderings, we implement
2344         * Mitigation A.3 in the RFC and delay the retransmission for a short
2345         * interval if appropriate.
2346         */
2347        if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2348            (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2349            !tcp_may_send_now(sk))
2350                return !tcp_pause_early_retransmit(sk, flag);
2351
2352        return false;
2353}
2354
2355/* New heuristics: it is possible only after we switched to restart timer
2356 * each time when something is ACKed. Hence, we can detect timed out packets
2357 * during fast retransmit without falling to slow start.
2358 *
2359 * Usefulness of this as is very questionable, since we should know which of
2360 * the segments is the next to timeout which is relatively expensive to find
2361 * in general case unless we add some data structure just for that. The
2362 * current approach certainly won't find the right one too often and when it
2363 * finally does find _something_ it usually marks large part of the window
2364 * right away (because a retransmission with a larger timestamp blocks the
2365 * loop from advancing). -ij
2366 */
2367static void tcp_timeout_skbs(struct sock *sk)
2368{
2369        struct tcp_sock *tp = tcp_sk(sk);
2370        struct sk_buff *skb;
2371
2372        if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2373                return;
2374
2375        skb = tp->scoreboard_skb_hint;
2376        if (tp->scoreboard_skb_hint == NULL)
2377                skb = tcp_write_queue_head(sk);
2378
2379        tcp_for_write_queue_from(skb, sk) {
2380                if (skb == tcp_send_head(sk))
2381                        break;
2382                if (!tcp_skb_timedout(sk, skb))
2383                        break;
2384
2385                tcp_skb_mark_lost(tp, skb);
2386        }
2387
2388        tp->scoreboard_skb_hint = skb;
2389
2390        tcp_verify_left_out(tp);
2391}
2392
2393/* Detect loss in event "A" above by marking head of queue up as lost.
2394 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2395 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2396 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2397 * the maximum SACKed segments to pass before reaching this limit.
2398 */
2399static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2400{
2401        struct tcp_sock *tp = tcp_sk(sk);
2402        struct sk_buff *skb;
2403        int cnt, oldcnt;
2404        int err;
2405        unsigned int mss;
2406        /* Use SACK to deduce losses of new sequences sent during recovery */
2407        const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2408
2409        WARN_ON(packets > tp->packets_out);
2410        if (tp->lost_skb_hint) {
2411                skb = tp->lost_skb_hint;
2412                cnt = tp->lost_cnt_hint;
2413                /* Head already handled? */
2414                if (mark_head && skb != tcp_write_queue_head(sk))
2415                        return;
2416        } else {
2417                skb = tcp_write_queue_head(sk);
2418                cnt = 0;
2419        }
2420
2421        tcp_for_write_queue_from(skb, sk) {
2422                if (skb == tcp_send_head(sk))
2423                        break;
2424                /* TODO: do this better */
2425                /* this is not the most efficient way to do this... */
2426                tp->lost_skb_hint = skb;
2427                tp->lost_cnt_hint = cnt;
2428
2429                if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2430                        break;
2431
2432                oldcnt = cnt;
2433                if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2434                    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2435                        cnt += tcp_skb_pcount(skb);
2436
2437                if (cnt > packets) {
2438                        if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2439                            (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2440                            (oldcnt >= packets))
2441                                break;
2442
2443                        mss = skb_shinfo(skb)->gso_size;
2444                        err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2445                        if (err < 0)
2446                                break;
2447                        cnt = packets;
2448                }
2449
2450                tcp_skb_mark_lost(tp, skb);
2451
2452                if (mark_head)
2453                        break;
2454        }
2455        tcp_verify_left_out(tp);
2456}
2457
2458/* Account newly detected lost packet(s) */
2459
2460static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2461{
2462        struct tcp_sock *tp = tcp_sk(sk);
2463
2464        if (tcp_is_reno(tp)) {
2465                tcp_mark_head_lost(sk, 1, 1);
2466        } else if (tcp_is_fack(tp)) {
2467                int lost = tp->fackets_out - tp->reordering;
2468                if (lost <= 0)
2469                        lost = 1;
2470                tcp_mark_head_lost(sk, lost, 0);
2471        } else {
2472                int sacked_upto = tp->sacked_out - tp->reordering;
2473                if (sacked_upto >= 0)
2474                        tcp_mark_head_lost(sk, sacked_upto, 0);
2475                else if (fast_rexmit)
2476                        tcp_mark_head_lost(sk, 1, 1);
2477        }
2478
2479        tcp_timeout_skbs(sk);
2480}
2481
2482/* CWND moderation, preventing bursts due to too big ACKs
2483 * in dubious situations.
2484 */
2485static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2486{
2487        tp->snd_cwnd = min(tp->snd_cwnd,
2488                           tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2489        tp->snd_cwnd_stamp = tcp_time_stamp;
2490}
2491
2492/* Lower bound on congestion window is slow start threshold
2493 * unless congestion avoidance choice decides to overide it.
2494 */
2495static inline u32 tcp_cwnd_min(const struct sock *sk)
2496{
2497        const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2498
2499        return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2500}
2501
2502/* Decrease cwnd each second ack. */
2503static void tcp_cwnd_down(struct sock *sk, int flag)
2504{
2505        struct tcp_sock *tp = tcp_sk(sk);
2506        int decr = tp->snd_cwnd_cnt + 1;
2507
2508        if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2509            (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2510                tp->snd_cwnd_cnt = decr & 1;
2511                decr >>= 1;
2512
2513                if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2514                        tp->snd_cwnd -= decr;
2515
2516                tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2517                tp->snd_cwnd_stamp = tcp_time_stamp;
2518        }
2519}
2520
2521/* Nothing was retransmitted or returned timestamp is less
2522 * than timestamp of the first retransmission.
2523 */
2524static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2525{
2526        return !tp->retrans_stamp ||
2527                (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2528                 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2529}
2530
2531/* Undo procedures. */
2532
2533#if FASTRETRANS_DEBUG > 1
2534static void DBGUNDO(struct sock *sk, const char *msg)
2535{
2536        struct tcp_sock *tp = tcp_sk(sk);
2537        struct inet_sock *inet = inet_sk(sk);
2538
2539        if (sk->sk_family == AF_INET) {
2540                pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2541                         msg,
2542                         &inet->inet_daddr, ntohs(inet->inet_dport),
2543                         tp->snd_cwnd, tcp_left_out(tp),
2544                         tp->snd_ssthresh, tp->prior_ssthresh,
2545                         tp->packets_out);
2546        }
2547#if IS_ENABLED(CONFIG_IPV6)
2548        else if (sk->sk_family == AF_INET6) {
2549                struct ipv6_pinfo *np = inet6_sk(sk);
2550                pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2551                         msg,
2552                         &np->daddr, ntohs(inet->inet_dport),
2553                         tp->snd_cwnd, tcp_left_out(tp),
2554                         tp->snd_ssthresh, tp->prior_ssthresh,
2555                         tp->packets_out);
2556        }
2557#endif
2558}
2559#else
2560#define DBGUNDO(x...) do { } while (0)
2561#endif
2562
2563static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2564{
2565        struct tcp_sock *tp = tcp_sk(sk);
2566
2567        if (tp->prior_ssthresh) {
2568                const struct inet_connection_sock *icsk = inet_csk(sk);
2569
2570                if (icsk->icsk_ca_ops->undo_cwnd)
2571                        tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2572                else
2573                        tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2574
2575                if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2576                        tp->snd_ssthresh = tp->prior_ssthresh;
2577                        TCP_ECN_withdraw_cwr(tp);
2578                }
2579        } else {
2580                tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2581        }
2582        tp->snd_cwnd_stamp = tcp_time_stamp;
2583}
2584
2585static inline bool tcp_may_undo(const struct tcp_sock *tp)
2586{
2587        return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2588}
2589
2590/* People celebrate: "We love our President!" */
2591static bool tcp_try_undo_recovery(struct sock *sk)
2592{
2593        struct tcp_sock *tp = tcp_sk(sk);
2594
2595        if (tcp_may_undo(tp)) {
2596                int mib_idx;
2597
2598                /* Happy end! We did not retransmit anything
2599                 * or our original transmission succeeded.
2600                 */
2601                DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2602                tcp_undo_cwr(sk, true);
2603                if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2604                        mib_idx = LINUX_MIB_TCPLOSSUNDO;
2605                else
2606                        mib_idx = LINUX_MIB_TCPFULLUNDO;
2607
2608                NET_INC_STATS_BH(sock_net(sk), mib_idx);
2609                tp->undo_marker = 0;
2610        }
2611        if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2612                /* Hold old state until something *above* high_seq
2613                 * is ACKed. For Reno it is MUST to prevent false
2614                 * fast retransmits (RFC2582). SACK TCP is safe. */
2615                tcp_moderate_cwnd(tp);
2616                return true;
2617        }
2618        tcp_set_ca_state(sk, TCP_CA_Open);
2619        return false;
2620}
2621
2622/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2623static void tcp_try_undo_dsack(struct sock *sk)
2624{
2625        struct tcp_sock *tp = tcp_sk(sk);
2626
2627        if (tp->undo_marker && !tp->undo_retrans) {
2628                DBGUNDO(sk, "D-SACK");
2629                tcp_undo_cwr(sk, true);
2630                tp->undo_marker = 0;
2631                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2632        }
2633}
2634
2635/* We can clear retrans_stamp when there are no retransmissions in the
2636 * window. It would seem that it is trivially available for us in
2637 * tp->retrans_out, however, that kind of assumptions doesn't consider
2638 * what will happen if errors occur when sending retransmission for the
2639 * second time. ...It could the that such segment has only
2640 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2641 * the head skb is enough except for some reneging corner cases that
2642 * are not worth the effort.
2643 *
2644 * Main reason for all this complexity is the fact that connection dying
2645 * time now depends on the validity of the retrans_stamp, in particular,
2646 * that successive retransmissions of a segment must not advance
2647 * retrans_stamp under any conditions.
2648 */
2649static bool tcp_any_retrans_done(const struct sock *sk)
2650{
2651        const struct tcp_sock *tp = tcp_sk(sk);
2652        struct sk_buff *skb;
2653
2654        if (tp->retrans_out)
2655                return true;
2656
2657        skb = tcp_write_queue_head(sk);
2658        if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2659                return true;
2660
2661        return false;
2662}
2663
2664/* Undo during fast recovery after partial ACK. */
2665
2666static int tcp_try_undo_partial(struct sock *sk, int acked)
2667{
2668        struct tcp_sock *tp = tcp_sk(sk);
2669        /* Partial ACK arrived. Force Hoe's retransmit. */
2670        int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2671
2672        if (tcp_may_undo(tp)) {
2673                /* Plain luck! Hole if filled with delayed
2674                 * packet, rather than with a retransmit.
2675                 */
2676                if (!tcp_any_retrans_done(sk))
2677                        tp->retrans_stamp = 0;
2678
2679                tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2680
2681                DBGUNDO(sk, "Hoe");
2682                tcp_undo_cwr(sk, false);
2683                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2684
2685                /* So... Do not make Hoe's retransmit yet.
2686                 * If the first packet was delayed, the rest
2687                 * ones are most probably delayed as well.
2688                 */
2689                failed = 0;
2690        }
2691        return failed;
2692}
2693
2694/* Undo during loss recovery after partial ACK. */
2695static bool tcp_try_undo_loss(struct sock *sk)
2696{
2697        struct tcp_sock *tp = tcp_sk(sk);
2698
2699        if (tcp_may_undo(tp)) {
2700                struct sk_buff *skb;
2701                tcp_for_write_queue(skb, sk) {
2702                        if (skb == tcp_send_head(sk))
2703                                break;
2704                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2705                }
2706
2707                tcp_clear_all_retrans_hints(tp);
2708
2709                DBGUNDO(sk, "partial loss");
2710                tp->lost_out = 0;
2711                tcp_undo_cwr(sk, true);
2712                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2713                inet_csk(sk)->icsk_retransmits = 0;
2714                tp->undo_marker = 0;
2715                if (tcp_is_sack(tp))
2716                        tcp_set_ca_state(sk, TCP_CA_Open);
2717                return true;
2718        }
2719        return false;
2720}
2721
2722static inline void tcp_complete_cwr(struct sock *sk)
2723{
2724        struct tcp_sock *tp = tcp_sk(sk);
2725
2726        /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2727        if (tp->undo_marker) {
2728                if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2729                        tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2730                        tp->snd_cwnd_stamp = tcp_time_stamp;
2731                } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2732                        /* PRR algorithm. */
2733                        tp->snd_cwnd = tp->snd_ssthresh;
2734                        tp->snd_cwnd_stamp = tcp_time_stamp;
2735                }
2736        }
2737        tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2738}
2739
2740static void tcp_try_keep_open(struct sock *sk)
2741{
2742        struct tcp_sock *tp = tcp_sk(sk);
2743        int state = TCP_CA_Open;
2744
2745        if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2746                state = TCP_CA_Disorder;
2747
2748        if (inet_csk(sk)->icsk_ca_state != state) {
2749                tcp_set_ca_state(sk, state);
2750                tp->high_seq = tp->snd_nxt;
2751        }
2752}
2753
2754static void tcp_try_to_open(struct sock *sk, int flag)
2755{
2756        struct tcp_sock *tp = tcp_sk(sk);
2757
2758        tcp_verify_left_out(tp);
2759
2760        if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2761                tp->retrans_stamp = 0;
2762
2763        if (flag & FLAG_ECE)
2764                tcp_enter_cwr(sk, 1);
2765
2766        if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2767                tcp_try_keep_open(sk);
2768                if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2769                        tcp_moderate_cwnd(tp);
2770        } else {
2771                tcp_cwnd_down(sk, flag);
2772        }
2773}
2774
2775static void tcp_mtup_probe_failed(struct sock *sk)
2776{
2777        struct inet_connection_sock *icsk = inet_csk(sk);
2778
2779        icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2780        icsk->icsk_mtup.probe_size = 0;
2781}
2782
2783static void tcp_mtup_probe_success(struct sock *sk)
2784{
2785        struct tcp_sock *tp = tcp_sk(sk);
2786        struct inet_connection_sock *icsk = inet_csk(sk);
2787
2788        /* FIXME: breaks with very large cwnd */
2789        tp->prior_ssthresh = tcp_current_ssthresh(sk);
2790        tp->snd_cwnd = tp->snd_cwnd *
2791                       tcp_mss_to_mtu(sk, tp->mss_cache) /
2792                       icsk->icsk_mtup.probe_size;
2793        tp->snd_cwnd_cnt = 0;
2794        tp->snd_cwnd_stamp = tcp_time_stamp;
2795        tp->snd_ssthresh = tcp_current_ssthresh(sk);
2796
2797        icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2798        icsk->icsk_mtup.probe_size = 0;
2799        tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2800}
2801
2802/* Do a simple retransmit without using the backoff mechanisms in
2803 * tcp_timer. This is used for path mtu discovery.
2804 * The socket is already locked here.
2805 */
2806void tcp_simple_retransmit(struct sock *sk)
2807{
2808        const struct inet_connection_sock *icsk = inet_csk(sk);
2809        struct tcp_sock *tp = tcp_sk(sk);
2810        struct sk_buff *skb;
2811        unsigned int mss = tcp_current_mss(sk);
2812        u32 prior_lost = tp->lost_out;
2813
2814        tcp_for_write_queue(skb, sk) {
2815                if (skb == tcp_send_head(sk))
2816                        break;
2817                if (tcp_skb_seglen(skb) > mss &&
2818                    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2819                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2820                                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2821                                tp->retrans_out -= tcp_skb_pcount(skb);
2822                        }
2823                        tcp_skb_mark_lost_uncond_verify(tp, skb);
2824                }
2825        }
2826
2827        tcp_clear_retrans_hints_partial(tp);
2828
2829        if (prior_lost == tp->lost_out)
2830                return;
2831
2832        if (tcp_is_reno(tp))
2833                tcp_limit_reno_sacked(tp);
2834
2835        tcp_verify_left_out(tp);
2836
2837        /* Don't muck with the congestion window here.
2838         * Reason is that we do not increase amount of _data_
2839         * in network, but units changed and effective
2840         * cwnd/ssthresh really reduced now.
2841         */
2842        if (icsk->icsk_ca_state != TCP_CA_Loss) {
2843                tp->high_seq = tp->snd_nxt;
2844                tp->snd_ssthresh = tcp_current_ssthresh(sk);
2845                tp->prior_ssthresh = 0;
2846                tp->undo_marker = 0;
2847                tcp_set_ca_state(sk, TCP_CA_Loss);
2848        }
2849        tcp_xmit_retransmit_queue(sk);
2850}
2851EXPORT_SYMBOL(tcp_simple_retransmit);
2852
2853/* This function implements the PRR algorithm, specifcally the PRR-SSRB
2854 * (proportional rate reduction with slow start reduction bound) as described in
2855 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2856 * It computes the number of packets to send (sndcnt) based on packets newly
2857 * delivered:
2858 *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2859 *      cwnd reductions across a full RTT.
2860 *   2) If packets in flight is lower than ssthresh (such as due to excess
2861 *      losses and/or application stalls), do not perform any further cwnd
2862 *      reductions, but instead slow start up to ssthresh.
2863 */
2864static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2865                                        int fast_rexmit, int flag)
2866{
2867        struct tcp_sock *tp = tcp_sk(sk);
2868        int sndcnt = 0;
2869        int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2870
2871        if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2872                u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2873                               tp->prior_cwnd - 1;
2874                sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2875        } else {
2876                sndcnt = min_t(int, delta,
2877                               max_t(int, tp->prr_delivered - tp->prr_out,
2878                                     newly_acked_sacked) + 1);
2879        }
2880
2881        sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2882        tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2883}
2884
2885static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2886{
2887        struct tcp_sock *tp = tcp_sk(sk);
2888        int mib_idx;
2889
2890        if (tcp_is_reno(tp))
2891                mib_idx = LINUX_MIB_TCPRENORECOVERY;
2892        else
2893                mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2894
2895        NET_INC_STATS_BH(sock_net(sk), mib_idx);
2896
2897        tp->high_seq = tp->snd_nxt;
2898        tp->prior_ssthresh = 0;
2899        tp->undo_marker = tp->snd_una;
2900        tp->undo_retrans = tp->retrans_out;
2901
2902        if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2903                if (!ece_ack)
2904                        tp->prior_ssthresh = tcp_current_ssthresh(sk);
2905                tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2906                TCP_ECN_queue_cwr(tp);
2907        }
2908
2909        tp->bytes_acked = 0;
2910        tp->snd_cwnd_cnt = 0;
2911        tp->prior_cwnd = tp->snd_cwnd;
2912        tp->prr_delivered = 0;
2913        tp->prr_out = 0;
2914        tcp_set_ca_state(sk, TCP_CA_Recovery);
2915}
2916
2917/* Process an event, which can update packets-in-flight not trivially.
2918 * Main goal of this function is to calculate new estimate for left_out,
2919 * taking into account both packets sitting in receiver's buffer and
2920 * packets lost by network.
2921 *
2922 * Besides that it does CWND reduction, when packet loss is detected
2923 * and changes state of machine.
2924 *
2925 * It does _not_ decide what to send, it is made in function
2926 * tcp_xmit_retransmit_queue().
2927 */
2928static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2929                                  int prior_sacked, bool is_dupack,
2930                                  int flag)
2931{
2932        struct inet_connection_sock *icsk = inet_csk(sk);
2933        struct tcp_sock *tp = tcp_sk(sk);
2934        int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2935                                    (tcp_fackets_out(tp) > tp->reordering));
2936        int newly_acked_sacked = 0;
2937        int fast_rexmit = 0;
2938
2939        if (WARN_ON(!tp->packets_out && tp->sacked_out))
2940                tp->sacked_out = 0;
2941        if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2942                tp->fackets_out = 0;
2943
2944        /* Now state machine starts.
2945         * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2946        if (flag & FLAG_ECE)
2947                tp->prior_ssthresh = 0;
2948
2949        /* B. In all the states check for reneging SACKs. */
2950        if (tcp_check_sack_reneging(sk, flag))
2951                return;
2952
2953        /* C. Check consistency of the current state. */
2954        tcp_verify_left_out(tp);
2955
2956        /* D. Check state exit conditions. State can be terminated
2957         *    when high_seq is ACKed. */
2958        if (icsk->icsk_ca_state == TCP_CA_Open) {
2959                WARN_ON(tp->retrans_out != 0);
2960                tp->retrans_stamp = 0;
2961        } else if (!before(tp->snd_una, tp->high_seq)) {
2962                switch (icsk->icsk_ca_state) {
2963                case TCP_CA_Loss:
2964                        icsk->icsk_retransmits = 0;
2965                        if (tcp_try_undo_recovery(sk))
2966                                return;
2967                        break;
2968
2969                case TCP_CA_CWR:
2970                        /* CWR is to be held something *above* high_seq
2971                         * is ACKed for CWR bit to reach receiver. */
2972                        if (tp->snd_una != tp->high_seq) {
2973                                tcp_complete_cwr(sk);
2974                                tcp_set_ca_state(sk, TCP_CA_Open);
2975                        }
2976                        break;
2977
2978                case TCP_CA_Recovery:
2979                        if (tcp_is_reno(tp))
2980                                tcp_reset_reno_sack(tp);
2981                        if (tcp_try_undo_recovery(sk))
2982                                return;
2983                        tcp_complete_cwr(sk);
2984                        break;
2985                }
2986        }
2987
2988        /* E. Process state. */
2989        switch (icsk->icsk_ca_state) {
2990        case TCP_CA_Recovery:
2991                if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2992                        if (tcp_is_reno(tp) && is_dupack)
2993                                tcp_add_reno_sack(sk);
2994                } else
2995                        do_lost = tcp_try_undo_partial(sk, pkts_acked);
2996                newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2997                break;
2998        case TCP_CA_Loss:
2999                if (flag & FLAG_DATA_ACKED)
3000                        icsk->icsk_retransmits = 0;
3001                if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3002                        tcp_reset_reno_sack(tp);
3003                if (!tcp_try_undo_loss(sk)) {
3004                        tcp_moderate_cwnd(tp);
3005                        tcp_xmit_retransmit_queue(sk);
3006                        return;
3007                }
3008                if (icsk->icsk_ca_state != TCP_CA_Open)
3009                        return;
3010                /* Loss is undone; fall through to processing in Open state. */
3011        default:
3012                if (tcp_is_reno(tp)) {
3013                        if (flag & FLAG_SND_UNA_ADVANCED)
3014                                tcp_reset_reno_sack(tp);
3015                        if (is_dupack)
3016                                tcp_add_reno_sack(sk);
3017                }
3018                newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
3019
3020                if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3021                        tcp_try_undo_dsack(sk);
3022
3023                if (!tcp_time_to_recover(sk, flag)) {
3024                        tcp_try_to_open(sk, flag);
3025                        return;
3026                }
3027
3028                /* MTU probe failure: don't reduce cwnd */
3029                if (icsk->icsk_ca_state < TCP_CA_CWR &&
3030                    icsk->icsk_mtup.probe_size &&
3031                    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3032                        tcp_mtup_probe_failed(sk);
3033                        /* Restores the reduction we did in tcp_mtup_probe() */
3034                        tp->snd_cwnd++;
3035                        tcp_simple_retransmit(sk);
3036                        return;
3037                }
3038
3039                /* Otherwise enter Recovery state */
3040                tcp_enter_recovery(sk, (flag & FLAG_ECE));
3041                fast_rexmit = 1;
3042        }
3043
3044        if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3045                tcp_update_scoreboard(sk, fast_rexmit);
3046        tp->prr_delivered += newly_acked_sacked;
3047        tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3048        tcp_xmit_retransmit_queue(sk);
3049}
3050
3051void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3052{
3053        tcp_rtt_estimator(sk, seq_rtt);
3054        tcp_set_rto(sk);
3055        inet_csk(sk)->icsk_backoff = 0;
3056}
3057EXPORT_SYMBOL(tcp_valid_rtt_meas);
3058
3059/* Read draft-ietf-tcplw-high-performance before mucking
3060 * with this code. (Supersedes RFC1323)
3061 */
3062static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3063{
3064        /* RTTM Rule: A TSecr value received in a segment is used to
3065         * update the averaged RTT measurement only if the segment
3066         * acknowledges some new data, i.e., only if it advances the
3067         * left edge of the send window.
3068         *
3069         * See draft-ietf-tcplw-high-performance-00, section 3.3.
3070         * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3071         *
3072         * Changed: reset backoff as soon as we see the first valid sample.
3073         * If we do not, we get strongly overestimated rto. With timestamps
3074         * samples are accepted even from very old segments: f.e., when rtt=1
3075         * increases to 8, we retransmit 5 times and after 8 seconds delayed
3076         * answer arrives rto becomes 120 seconds! If at least one of segments
3077         * in window is lost... Voila.                          --ANK (010210)
3078         */
3079        struct tcp_sock *tp = tcp_sk(sk);
3080
3081        tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3082}
3083
3084static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3085{
3086        /* We don't have a timestamp. Can only use
3087         * packets that are not retransmitted to determine
3088         * rtt estimates. Also, we must not reset the
3089         * backoff for rto until we get a non-retransmitted
3090         * packet. This allows us to deal with a situation
3091         * where the network delay has increased suddenly.
3092         * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3093         */
3094
3095        if (flag & FLAG_RETRANS_DATA_ACKED)
3096                return;
3097
3098        tcp_valid_rtt_meas(sk, seq_rtt);
3099}
3100
3101static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3102                                      const s32 seq_rtt)
3103{
3104        const struct tcp_sock *tp = tcp_sk(sk);
3105        /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3106        if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3107                tcp_ack_saw_tstamp(sk, flag);
3108        else if (seq_rtt >= 0)
3109                tcp_ack_no_tstamp(sk, seq_rtt, flag);
3110}
3111
3112static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3113{
3114        const struct inet_connection_sock *icsk = inet_csk(sk);
3115        icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3116        tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3117}
3118
3119/* Restart timer after forward progress on connection.
3120 * RFC2988 recommends to restart timer to now+rto.
3121 */
3122void tcp_rearm_rto(struct sock *sk)
3123{
3124        struct tcp_sock *tp = tcp_sk(sk);
3125
3126        if (!tp->packets_out) {
3127                inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3128        } else {
3129                u32 rto = inet_csk(sk)->icsk_rto;
3130                /* Offset the time elapsed after installing regular RTO */
3131                if (tp->early_retrans_delayed) {
3132                        struct sk_buff *skb = tcp_write_queue_head(sk);
3133                        const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3134                        s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3135                        /* delta may not be positive if the socket is locked
3136                         * when the delayed ER timer fires and is rescheduled.
3137                         */
3138                        if (delta > 0)
3139                                rto = delta;
3140                }
3141                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3142                                          TCP_RTO_MAX);
3143        }
3144        tp->early_retrans_delayed = 0;
3145}
3146
3147/* This function is called when the delayed ER timer fires. TCP enters
3148 * fast recovery and performs fast-retransmit.
3149 */
3150void tcp_resume_early_retransmit(struct sock *sk)
3151{
3152        struct tcp_sock *tp = tcp_sk(sk);
3153
3154        tcp_rearm_rto(sk);
3155
3156        /* Stop if ER is disabled after the delayed ER timer is scheduled */
3157        if (!tp->do_early_retrans)
3158                return;
3159
3160        tcp_enter_recovery(sk, false);
3161        tcp_update_scoreboard(sk, 1);
3162        tcp_xmit_retransmit_queue(sk);
3163}
3164
3165/* If we get here, the whole TSO packet has not been acked. */
3166static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3167{
3168        struct tcp_sock *tp = tcp_sk(sk);
3169        u32 packets_acked;
3170
3171        BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3172
3173        packets_acked = tcp_skb_pcount(skb);
3174        if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3175                return 0;
3176        packets_acked -= tcp_skb_pcount(skb);
3177
3178        if (packets_acked) {
3179                BUG_ON(tcp_skb_pcount(skb) == 0);
3180                BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3181        }
3182
3183        return packets_acked;
3184}
3185
3186/* Remove acknowledged frames from the retransmission queue. If our packet
3187 * is before the ack sequence we can discard it as it's confirmed to have
3188 * arrived at the other end.
3189 */
3190static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3191                               u32 prior_snd_una)
3192{
3193        struct tcp_sock *tp = tcp_sk(sk);
3194        const struct inet_connection_sock *icsk = inet_csk(sk);
3195        struct sk_buff *skb;
3196        u32 now = tcp_time_stamp;
3197        int fully_acked = true;
3198        int flag = 0;
3199        u32 pkts_acked = 0;
3200        u32 reord = tp->packets_out;
3201        u32 prior_sacked = tp->sacked_out;
3202        s32 seq_rtt = -1;
3203        s32 ca_seq_rtt = -1;
3204        ktime_t last_ackt = net_invalid_timestamp();
3205
3206        while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3207                struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3208                u32 acked_pcount;
3209                u8 sacked = scb->sacked;
3210
3211                /* Determine how many packets and what bytes were acked, tso and else */
3212                if (after(scb->end_seq, tp->snd_una)) {
3213                        if (tcp_skb_pcount(skb) == 1 ||
3214                            !after(tp->snd_una, scb->seq))
3215                                break;
3216
3217                        acked_pcount = tcp_tso_acked(sk, skb);
3218                        if (!acked_pcount)
3219                                break;
3220
3221                        fully_acked = false;
3222                } else {
3223                        acked_pcount = tcp_skb_pcount(skb);
3224                }
3225
3226                if (sacked & TCPCB_RETRANS) {
3227                        if (sacked & TCPCB_SACKED_RETRANS)
3228                                tp->retrans_out -= acked_pcount;
3229                        flag |= FLAG_RETRANS_DATA_ACKED;
3230                        ca_seq_rtt = -1;
3231                        seq_rtt = -1;
3232                        if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3233                                flag |= FLAG_NONHEAD_RETRANS_ACKED;
3234                } else {
3235                        ca_seq_rtt = now - scb->when;
3236                        last_ackt = skb->tstamp;
3237                        if (seq_rtt < 0) {
3238                                seq_rtt = ca_seq_rtt;
3239                        }
3240                        if (!(sacked & TCPCB_SACKED_ACKED))
3241                                reord = min(pkts_acked, reord);
3242                }
3243
3244                if (sacked & TCPCB_SACKED_ACKED)
3245                        tp->sacked_out -= acked_pcount;
3246                if (sacked & TCPCB_LOST)
3247                        tp->lost_out -= acked_pcount;
3248
3249                tp->packets_out -= acked_pcount;
3250                pkts_acked += acked_pcount;
3251
3252                /* Initial outgoing SYN's get put onto the write_queue
3253                 * just like anything else we transmit.  It is not
3254                 * true data, and if we misinform our callers that
3255                 * this ACK acks real data, we will erroneously exit
3256                 * connection startup slow start one packet too
3257                 * quickly.  This is severely frowned upon behavior.
3258                 */
3259                if (!(scb->tcp_flags & TCPHDR_SYN)) {
3260                        flag |= FLAG_DATA_ACKED;
3261                } else {
3262                        flag |= FLAG_SYN_ACKED;
3263                        tp->retrans_stamp = 0;
3264                }
3265
3266                if (!fully_acked)
3267                        break;
3268
3269                tcp_unlink_write_queue(skb, sk);
3270                sk_wmem_free_skb(sk, skb);
3271                tp->scoreboard_skb_hint = NULL;
3272                if (skb == tp->retransmit_skb_hint)
3273                        tp->retransmit_skb_hint = NULL;
3274                if (skb == tp->lost_skb_hint)
3275                        tp->lost_skb_hint = NULL;
3276        }
3277
3278        if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3279                tp->snd_up = tp->snd_una;
3280
3281        if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3282                flag |= FLAG_SACK_RENEGING;
3283
3284        if (flag & FLAG_ACKED) {
3285                const struct tcp_congestion_ops *ca_ops
3286                        = inet_csk(sk)->icsk_ca_ops;
3287
3288                if (unlikely(icsk->icsk_mtup.probe_size &&
3289                             !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3290                        tcp_mtup_probe_success(sk);
3291                }
3292
3293                tcp_ack_update_rtt(sk, flag, seq_rtt);
3294                tcp_rearm_rto(sk);
3295
3296                if (tcp_is_reno(tp)) {
3297                        tcp_remove_reno_sacks(sk, pkts_acked);
3298                } else {
3299                        int delta;
3300
3301                        /* Non-retransmitted hole got filled? That's reordering */
3302                        if (reord < prior_fackets)
3303                                tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3304
3305                        delta = tcp_is_fack(tp) ? pkts_acked :
3306                                                  prior_sacked - tp->sacked_out;
3307                        tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3308                }
3309
3310                tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3311
3312                if (ca_ops->pkts_acked) {
3313                        s32 rtt_us = -1;
3314
3315                        /* Is the ACK triggering packet unambiguous? */
3316                        if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3317                                /* High resolution needed and available? */
3318                                if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3319                                    !ktime_equal(last_ackt,
3320                                                 net_invalid_timestamp()))
3321                                        rtt_us = ktime_us_delta(ktime_get_real(),
3322                                                                last_ackt);
3323                                else if (ca_seq_rtt >= 0)
3324                                        rtt_us = jiffies_to_usecs(ca_seq_rtt);
3325                        }
3326
3327                        ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3328                }
3329        }
3330
3331#if FASTRETRANS_DEBUG > 0
3332        WARN_ON((int)tp->sacked_out < 0);
3333        WARN_ON((int)tp->lost_out < 0);
3334        WARN_ON((int)tp->retrans_out < 0);
3335        if (!tp->packets_out && tcp_is_sack(tp)) {
3336                icsk = inet_csk(sk);
3337                if (tp->lost_out) {
3338                        pr_debug("Leak l=%u %d\n",
3339                                 tp->lost_out, icsk->icsk_ca_state);
3340                        tp->lost_out = 0;
3341                }
3342                if (tp->sacked_out) {
3343                        pr_debug("Leak s=%u %d\n",
3344                                 tp->sacked_out, icsk->icsk_ca_state);
3345                        tp->sacked_out = 0;
3346                }
3347                if (tp->retrans_out) {
3348                        pr_debug("Leak r=%u %d\n",
3349                                 tp->retrans_out, icsk->icsk_ca_state);
3350                        tp->retrans_out = 0;
3351                }
3352        }
3353#endif
3354        return flag;
3355}
3356
3357static void tcp_ack_probe(struct sock *sk)
3358{
3359        const struct tcp_sock *tp = tcp_sk(sk);
3360        struct inet_connection_sock *icsk = inet_csk(sk);
3361
3362        /* Was it a usable window open? */
3363
3364        if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3365                icsk->icsk_backoff = 0;
3366                inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3367                /* Socket must be waked up by subsequent tcp_data_snd_check().
3368                 * This function is not for random using!
3369                 */
3370        } else {
3371                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3372                                          min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3373                                          TCP_RTO_MAX);
3374        }
3375}
3376
3377static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3378{
3379        return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3380                inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3381}
3382
3383static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3384{
3385        const struct tcp_sock *tp = tcp_sk(sk);
3386        return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3387                !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3388}
3389
3390/* Check that window update is acceptable.
3391 * The function assumes that snd_una<=ack<=snd_next.
3392 */
3393static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3394                                        const u32 ack, const u32 ack_seq,
3395                                        const u32 nwin)
3396{
3397        return  after(ack, tp->snd_una) ||
3398                after(ack_seq, tp->snd_wl1) ||
3399                (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3400}
3401
3402/* Update our send window.
3403 *
3404 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3405 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3406 */
3407static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3408                                 u32 ack_seq)
3409{
3410        struct tcp_sock *tp = tcp_sk(sk);
3411        int flag = 0;
3412        u32 nwin = ntohs(tcp_hdr(skb)->window);
3413
3414        if (likely(!tcp_hdr(skb)->syn))
3415                nwin <<= tp->rx_opt.snd_wscale;
3416
3417        if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3418                flag |= FLAG_WIN_UPDATE;
3419                tcp_update_wl(tp, ack_seq);
3420
3421                if (tp->snd_wnd != nwin) {
3422                        tp->snd_wnd = nwin;
3423
3424                        /* Note, it is the only place, where
3425                         * fast path is recovered for sending TCP.
3426                         */
3427                        tp->pred_flags = 0;
3428                        tcp_fast_path_check(sk);
3429
3430                        if (nwin > tp->max_window) {
3431                                tp->max_window = nwin;
3432                                tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3433                        }
3434                }
3435        }
3436
3437        tp->snd_una = ack;
3438
3439        return flag;
3440}
3441
3442/* A very conservative spurious RTO response algorithm: reduce cwnd and
3443 * continue in congestion avoidance.
3444 */
3445static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3446{
3447        tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3448        tp->snd_cwnd_cnt = 0;
3449        tp->bytes_acked = 0;
3450        TCP_ECN_queue_cwr(tp);
3451        tcp_moderate_cwnd(tp);
3452}
3453
3454/* A conservative spurious RTO response algorithm: reduce cwnd using
3455 * rate halving and continue in congestion avoidance.
3456 */
3457static void tcp_ratehalving_spur_to_response(struct sock *sk)
3458{
3459        tcp_enter_cwr(sk, 0);
3460}
3461
3462static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3463{
3464        if (flag & FLAG_ECE)
3465                tcp_ratehalving_spur_to_response(sk);
3466        else
3467                tcp_undo_cwr(sk, true);
3468}
3469
3470/* F-RTO spurious RTO detection algorithm (RFC4138)
3471 *
3472 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3473 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3474 * window (but not to or beyond highest sequence sent before RTO):
3475 *   On First ACK,  send two new segments out.
3476 *   On Second ACK, RTO was likely spurious. Do spurious response (response
3477 *                  algorithm is not part of the F-RTO detection algorithm
3478 *                  given in RFC4138 but can be selected separately).
3479 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3480 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3481 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3482 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3483 *
3484 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3485 * original window even after we transmit two new data segments.
3486 *
3487 * SACK version:
3488 *   on first step, wait until first cumulative ACK arrives, then move to
3489 *   the second step. In second step, the next ACK decides.
3490 *
3491 * F-RTO is implemented (mainly) in four functions:
3492 *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3493 *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3494 *     called when tcp_use_frto() showed green light
3495 *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3496 *   - tcp_enter_frto_loss() is called if there is not enough evidence
3497 *     to prove that the RTO is indeed spurious. It transfers the control
3498 *     from F-RTO to the conventional RTO recovery
3499 */
3500static bool tcp_process_frto(struct sock *sk, int flag)
3501{
3502        struct tcp_sock *tp = tcp_sk(sk);
3503
3504        tcp_verify_left_out(tp);
3505
3506        /* Duplicate the behavior from Loss state (fastretrans_alert) */
3507        if (flag & FLAG_DATA_ACKED)
3508                inet_csk(sk)->icsk_retransmits = 0;
3509
3510        if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3511            ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3512                tp->undo_marker = 0;
3513
3514        if (!before(tp->snd_una, tp->frto_highmark)) {
3515                tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3516                return true;
3517        }
3518
3519        if (!tcp_is_sackfrto(tp)) {
3520                /* RFC4138 shortcoming in step 2; should also have case c):
3521                 * ACK isn't duplicate nor advances window, e.g., opposite dir
3522                 * data, winupdate
3523                 */
3524                if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3525                        return true;
3526
3527                if (!(flag & FLAG_DATA_ACKED)) {
3528                        tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3529                                            flag);
3530                        return true;
3531                }
3532        } else {
3533                if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3534                        /* Prevent sending of new data. */
3535                        tp->snd_cwnd = min(tp->snd_cwnd,
3536                                           tcp_packets_in_flight(tp));
3537                        return true;
3538                }
3539
3540                if ((tp->frto_counter >= 2) &&
3541                    (!(flag & FLAG_FORWARD_PROGRESS) ||
3542                     ((flag & FLAG_DATA_SACKED) &&
3543                      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3544                        /* RFC4138 shortcoming (see comment above) */
3545                        if (!(flag & FLAG_FORWARD_PROGRESS) &&
3546                            (flag & FLAG_NOT_DUP))
3547                                return true;
3548
3549                        tcp_enter_frto_loss(sk, 3, flag);
3550                        return true;
3551                }
3552        }
3553
3554        if (tp->frto_counter == 1) {
3555                /* tcp_may_send_now needs to see updated state */
3556                tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3557                tp->frto_counter = 2;
3558
3559                if (!tcp_may_send_now(sk))
3560                        tcp_enter_frto_loss(sk, 2, flag);
3561
3562                return true;
3563        } else {
3564                switch (sysctl_tcp_frto_response) {
3565                case 2:
3566                        tcp_undo_spur_to_response(sk, flag);
3567                        break;
3568                case 1:
3569                        tcp_conservative_spur_to_response(tp);
3570                        break;
3571                default:
3572                        tcp_ratehalving_spur_to_response(sk);
3573                        break;
3574                }
3575                tp->frto_counter = 0;
3576                tp->undo_marker = 0;
3577                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3578        }
3579        return false;
3580}
3581
3582/* This routine deals with incoming acks, but not outgoing ones. */
3583static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3584{
3585        struct inet_connection_sock *icsk = inet_csk(sk);
3586        struct tcp_sock *tp = tcp_sk(sk);
3587        u32 prior_snd_una = tp->snd_una;
3588        u32 ack_seq = TCP_SKB_CB(skb)->seq;
3589        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3590        bool is_dupack = false;
3591        u32 prior_in_flight;
3592        u32 prior_fackets;
3593        int prior_packets;
3594        int prior_sacked = tp->sacked_out;
3595        int pkts_acked = 0;
3596        bool frto_cwnd = false;
3597
3598        /* If the ack is older than previous acks
3599         * then we can probably ignore it.
3600         */
3601        if (before(ack, prior_snd_una))
3602                goto old_ack;
3603
3604        /* If the ack includes data we haven't sent yet, discard
3605         * this segment (RFC793 Section 3.9).
3606         */
3607        if (after(ack, tp->snd_nxt))
3608                goto invalid_ack;
3609
3610        if (tp->early_retrans_delayed)
3611                tcp_rearm_rto(sk);
3612
3613        if (after(ack, prior_snd_una))
3614                flag |= FLAG_SND_UNA_ADVANCED;
3615
3616        if (sysctl_tcp_abc) {
3617                if (icsk->icsk_ca_state < TCP_CA_CWR)
3618                        tp->bytes_acked += ack - prior_snd_una;
3619                else if (icsk->icsk_ca_state == TCP_CA_Loss)
3620                        /* we assume just one segment left network */
3621                        tp->bytes_acked += min(ack - prior_snd_una,
3622                                               tp->mss_cache);
3623        }
3624
3625        prior_fackets = tp->fackets_out;
3626        prior_in_flight = tcp_packets_in_flight(tp);
3627
3628        if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3629                /* Window is constant, pure forward advance.
3630                 * No more checks are required.
3631                 * Note, we use the fact that SND.UNA>=SND.WL2.
3632                 */
3633                tcp_update_wl(tp, ack_seq);
3634                tp->snd_una = ack;
3635                flag |= FLAG_WIN_UPDATE;
3636
3637                tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3638
3639                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3640        } else {
3641                if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3642                        flag |= FLAG_DATA;
3643                else
3644                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3645
3646                flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3647
3648                if (TCP_SKB_CB(skb)->sacked)
3649                        flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3650
3651                if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3652                        flag |= FLAG_ECE;
3653
3654                tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3655        }
3656
3657        /* We passed data and got it acked, remove any soft error
3658         * log. Something worked...
3659         */
3660        sk->sk_err_soft = 0;
3661        icsk->icsk_probes_out = 0;
3662        tp->rcv_tstamp = tcp_time_stamp;
3663        prior_packets = tp->packets_out;
3664        if (!prior_packets)
3665                goto no_queue;
3666
3667        /* See if we can take anything off of the retransmit queue. */
3668        flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3669
3670        pkts_acked = prior_packets - tp->packets_out;
3671
3672        if (tp->frto_counter)
3673                frto_cwnd = tcp_process_frto(sk, flag);
3674        /* Guarantee sacktag reordering detection against wrap-arounds */
3675        if (before(tp->frto_highmark, tp->snd_una))
3676                tp->frto_highmark = 0;
3677
3678        if (tcp_ack_is_dubious(sk, flag)) {
3679                /* Advance CWND, if state allows this. */
3680                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3681                    tcp_may_raise_cwnd(sk, flag))
3682                        tcp_cong_avoid(sk, ack, prior_in_flight);
3683                is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3684                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3685                                      is_dupack, flag);
3686        } else {
3687                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3688                        tcp_cong_avoid(sk, ack, prior_in_flight);
3689        }
3690
3691        if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3692                struct dst_entry *dst = __sk_dst_get(sk);
3693                if (dst)
3694                        dst_confirm(dst);
3695        }
3696        return 1;
3697
3698no_queue:
3699        /* If data was DSACKed, see if we can undo a cwnd reduction. */
3700        if (flag & FLAG_DSACKING_ACK)
3701                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3702                                      is_dupack, flag);
3703        /* If this ack opens up a zero window, clear backoff.  It was
3704         * being used to time the probes, and is probably far higher than
3705         * it needs to be for normal retransmission.
3706         */
3707        if (tcp_send_head(sk))
3708                tcp_ack_probe(sk);
3709        return 1;
3710
3711invalid_ack:
3712        SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3713        return -1;
3714
3715old_ack:
3716        /* If data was SACKed, tag it and see if we should send more data.
3717         * If data was DSACKed, see if we can undo a cwnd reduction.
3718         */
3719        if (TCP_SKB_CB(skb)->sacked) {
3720                flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3721                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3722                                      is_dupack, flag);
3723        }
3724
3725        SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3726        return 0;
3727}
3728
3729/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3730 * But, this can also be called on packets in the established flow when
3731 * the fast version below fails.
3732 */
3733void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3734                       const u8 **hvpp, int estab,
3735                       struct tcp_fastopen_cookie *foc)
3736{
3737        const unsigned char *ptr;
3738        const struct tcphdr *th = tcp_hdr(skb);
3739        int length = (th->doff * 4) - sizeof(struct tcphdr);
3740
3741        ptr = (const unsigned char *)(th + 1);
3742        opt_rx->saw_tstamp = 0;
3743
3744        while (length > 0) {
3745                int opcode = *ptr++;
3746                int opsize;
3747
3748                switch (opcode) {
3749                case TCPOPT_EOL:
3750                        return;
3751                case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3752                        length--;
3753                        continue;
3754                default:
3755                        opsize = *ptr++;
3756                        if (opsize < 2) /* "silly options" */
3757                                return;
3758                        if (opsize > length)
3759                                return; /* don't parse partial options */
3760                        switch (opcode) {
3761                        case TCPOPT_MSS:
3762                                if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3763                                        u16 in_mss = get_unaligned_be16(ptr);
3764                                        if (in_mss) {
3765                                                if (opt_rx->user_mss &&
3766                                                    opt_rx->user_mss < in_mss)
3767                                                        in_mss = opt_rx->user_mss;
3768                                                opt_rx->mss_clamp = in_mss;
3769                                        }
3770                                }
3771                                break;
3772                        case TCPOPT_WINDOW:
3773                                if (opsize == TCPOLEN_WINDOW && th->syn &&
3774                                    !estab && sysctl_tcp_window_scaling) {
3775                                        __u8 snd_wscale = *(__u8 *)ptr;
3776                                        opt_rx->wscale_ok = 1;
3777                                        if (snd_wscale > 14) {
3778                                                net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3779                                                                     __func__,
3780                                                                     snd_wscale);
3781                                                snd_wscale = 14;
3782                                        }
3783                                        opt_rx->snd_wscale = snd_wscale;
3784                                }
3785                                break;
3786                        case TCPOPT_TIMESTAMP:
3787                                if ((opsize == TCPOLEN_TIMESTAMP) &&
3788                                    ((estab && opt_rx->tstamp_ok) ||
3789                                     (!estab && sysctl_tcp_timestamps))) {
3790                                        opt_rx->saw_tstamp = 1;
3791                                        opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3792                                        opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3793                                }
3794                                break;
3795                        case TCPOPT_SACK_PERM:
3796                                if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3797                                    !estab && sysctl_tcp_sack) {
3798                                        opt_rx->sack_ok = TCP_SACK_SEEN;
3799                                        tcp_sack_reset(opt_rx);
3800                                }
3801                                break;
3802
3803                        case TCPOPT_SACK:
3804                                if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3805                                   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3806                                   opt_rx->sack_ok) {
3807                                        TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3808                                }
3809                                break;
3810#ifdef CONFIG_TCP_MD5SIG
3811                        case TCPOPT_MD5SIG:
3812                                /*
3813                                 * The MD5 Hash has already been
3814                                 * checked (see tcp_v{4,6}_do_rcv()).
3815                                 */
3816                                break;
3817#endif
3818                        case TCPOPT_COOKIE:
3819                                /* This option is variable length.
3820                                 */
3821                                switch (opsize) {
3822                                case TCPOLEN_COOKIE_BASE:
3823                                        /* not yet implemented */
3824                                        break;
3825                                case TCPOLEN_COOKIE_PAIR:
3826                                        /* not yet implemented */
3827                                        break;
3828                                case TCPOLEN_COOKIE_MIN+0:
3829                                case TCPOLEN_COOKIE_MIN+2:
3830                                case TCPOLEN_COOKIE_MIN+4:
3831                                case TCPOLEN_COOKIE_MIN+6:
3832                                case TCPOLEN_COOKIE_MAX:
3833                                        /* 16-bit multiple */
3834                                        opt_rx->cookie_plus = opsize;
3835                                        *hvpp = ptr;
3836                                        break;
3837                                default:
3838                                        /* ignore option */
3839                                        break;
3840                                }
3841                                break;
3842
3843                        case TCPOPT_EXP:
3844                                /* Fast Open option shares code 254 using a
3845                                 * 16 bits magic number. It's valid only in
3846                                 * SYN or SYN-ACK with an even size.
3847                                 */
3848                                if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3849                                    get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3850                                    foc == NULL || !th->syn || (opsize & 1))
3851                                        break;
3852                                foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3853                                if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3854                                    foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3855                                        memcpy(foc->val, ptr + 2, foc->len);
3856                                else if (foc->len != 0)
3857                                        foc->len = -1;
3858                                break;
3859
3860                        }
3861                        ptr += opsize-2;
3862                        length -= opsize;
3863                }
3864        }
3865}
3866EXPORT_SYMBOL(tcp_parse_options);
3867
3868static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3869{
3870        const __be32 *ptr = (const __be32 *)(th + 1);
3871
3872        if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3873                          | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3874                tp->rx_opt.saw_tstamp = 1;
3875                ++ptr;
3876                tp->rx_opt.rcv_tsval = ntohl(*ptr);
3877                ++ptr;
3878                tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3879                return true;
3880        }
3881        return false;
3882}
3883
3884/* Fast parse options. This hopes to only see timestamps.
3885 * If it is wrong it falls back on tcp_parse_options().
3886 */
3887static bool tcp_fast_parse_options(const struct sk_buff *skb,
3888                                   const struct tcphdr *th,
3889                                   struct tcp_sock *tp, const u8 **hvpp)
3890{
3891        /* In the spirit of fast parsing, compare doff directly to constant
3892         * values.  Because equality is used, short doff can be ignored here.
3893         */
3894        if (th->doff == (sizeof(*th) / 4)) {
3895                tp->rx_opt.saw_tstamp = 0;
3896                return false;
3897        } else if (tp->rx_opt.tstamp_ok &&
3898                   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3899                if (tcp_parse_aligned_timestamp(tp, th))
3900                        return true;
3901        }
3902        tcp_parse_options(skb, &tp->rx_opt, hvpp, 1, NULL);
3903        return true;
3904}
3905
3906#ifdef CONFIG_TCP_MD5SIG
3907/*
3908 * Parse MD5 Signature option
3909 */
3910const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3911{
3912        int length = (th->doff << 2) - sizeof(*th);
3913        const u8 *ptr = (const u8 *)(th + 1);
3914
3915        /* If the TCP option is too short, we can short cut */
3916        if (length < TCPOLEN_MD5SIG)
3917                return NULL;
3918
3919        while (length > 0) {
3920                int opcode = *ptr++;
3921                int opsize;
3922
3923                switch(opcode) {
3924                case TCPOPT_EOL:
3925                        return NULL;
3926                case TCPOPT_NOP:
3927                        length--;
3928                        continue;
3929                default:
3930                        opsize = *ptr++;
3931                        if (opsize < 2 || opsize > length)
3932                                return NULL;
3933                        if (opcode == TCPOPT_MD5SIG)
3934                                return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3935                }
3936                ptr += opsize - 2;
3937                length -= opsize;
3938        }
3939        return NULL;
3940}
3941EXPORT_SYMBOL(tcp_parse_md5sig_option);
3942#endif
3943
3944static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3945{
3946        tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3947        tp->rx_opt.ts_recent_stamp = get_seconds();
3948}
3949
3950static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3951{
3952        if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3953                /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3954                 * extra check below makes sure this can only happen
3955                 * for pure ACK frames.  -DaveM
3956                 *
3957                 * Not only, also it occurs for expired timestamps.
3958                 */
3959
3960                if (tcp_paws_check(&tp->rx_opt, 0))
3961                        tcp_store_ts_recent(tp);
3962        }
3963}
3964
3965/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3966 *
3967 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3968 * it can pass through stack. So, the following predicate verifies that
3969 * this segment is not used for anything but congestion avoidance or
3970 * fast retransmit. Moreover, we even are able to eliminate most of such
3971 * second order effects, if we apply some small "replay" window (~RTO)
3972 * to timestamp space.
3973 *
3974 * All these measures still do not guarantee that we reject wrapped ACKs
3975 * on networks with high bandwidth, when sequence space is recycled fastly,
3976 * but it guarantees that such events will be very rare and do not affect
3977 * connection seriously. This doesn't look nice, but alas, PAWS is really
3978 * buggy extension.
3979 *
3980 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3981 * states that events when retransmit arrives after original data are rare.
3982 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3983 * the biggest problem on large power networks even with minor reordering.
3984 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3985 * up to bandwidth of 18Gigabit/sec. 8) ]
3986 */
3987
3988static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3989{
3990        const struct tcp_sock *tp = tcp_sk(sk);
3991        const struct tcphdr *th = tcp_hdr(skb);
3992        u32 seq = TCP_SKB_CB(skb)->seq;
3993        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3994
3995        return (/* 1. Pure ACK with correct sequence number. */
3996                (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3997
3998                /* 2. ... and duplicate ACK. */
3999                ack == tp->snd_una &&
4000
4001                /* 3. ... and does not update window. */
4002                !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4003
4004                /* 4. ... and sits in replay window. */
4005                (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4006}
4007
4008static inline bool tcp_paws_discard(const struct sock *sk,
4009                                   const struct sk_buff *skb)
4010{
4011        const struct tcp_sock *tp = tcp_sk(sk);
4012
4013        return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4014               !tcp_disordered_ack(sk, skb);
4015}
4016
4017/* Check segment sequence number for validity.
4018 *
4019 * Segment controls are considered valid, if the segment
4020 * fits to the window after truncation to the window. Acceptability
4021 * of data (and SYN, FIN, of course) is checked separately.
4022 * See tcp_data_queue(), for example.
4023 *
4024 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4025 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4026 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4027 * (borrowed from freebsd)
4028 */
4029
4030static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4031{
4032        return  !before(end_seq, tp->rcv_wup) &&
4033                !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4034}
4035
4036/* When we get a reset we do this. */
4037static void tcp_reset(struct sock *sk)
4038{
4039        /* We want the right error as BSD sees it (and indeed as we do). */
4040        switch (sk->sk_state) {
4041        case TCP_SYN_SENT:
4042                sk->sk_err = ECONNREFUSED;
4043                break;
4044        case TCP_CLOSE_WAIT:
4045                sk->sk_err = EPIPE;
4046                break;
4047        case TCP_CLOSE:
4048                return;
4049        default:
4050                sk->sk_err = ECONNRESET;
4051        }
4052        /* This barrier is coupled with smp_rmb() in tcp_poll() */
4053        smp_wmb();
4054
4055        if (!sock_flag(sk, SOCK_DEAD))
4056                sk->sk_error_report(sk);
4057
4058        tcp_done(sk);
4059}
4060
4061/*
4062 *      Process the FIN bit. This now behaves as it is supposed to work
4063 *      and the FIN takes effect when it is validly part of sequence
4064 *      space. Not before when we get holes.
4065 *
4066 *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4067 *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4068 *      TIME-WAIT)
4069 *
4070 *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4071 *      close and we go into CLOSING (and later onto TIME-WAIT)
4072 *
4073 *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4074 */
4075static void tcp_fin(struct sock *sk)
4076{
4077        struct tcp_sock *tp = tcp_sk(sk);
4078
4079        inet_csk_schedule_ack(sk);
4080
4081        sk->sk_shutdown |= RCV_SHUTDOWN;
4082        sock_set_flag(sk, SOCK_DONE);
4083
4084        switch (sk->sk_state) {
4085        case TCP_SYN_RECV:
4086        case TCP_ESTABLISHED:
4087                /* Move to CLOSE_WAIT */
4088                tcp_set_state(sk, TCP_CLOSE_WAIT);
4089                inet_csk(sk)->icsk_ack.pingpong = 1;
4090                break;
4091
4092        case TCP_CLOSE_WAIT:
4093        case TCP_CLOSING:
4094                /* Received a retransmission of the FIN, do
4095                 * nothing.
4096                 */
4097                break;
4098        case TCP_LAST_ACK:
4099                /* RFC793: Remain in the LAST-ACK state. */
4100                break;
4101
4102        case TCP_FIN_WAIT1:
4103                /* This case occurs when a simultaneous close
4104                 * happens, we must ack the received FIN and
4105                 * enter the CLOSING state.
4106                 */
4107                tcp_send_ack(sk);
4108                tcp_set_state(sk, TCP_CLOSING);
4109                break;
4110        case TCP_FIN_WAIT2:
4111                /* Received a FIN -- send ACK and enter TIME_WAIT. */
4112                tcp_send_ack(sk);
4113                tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4114                break;
4115        default:
4116                /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4117                 * cases we should never reach this piece of code.
4118                 */
4119                pr_err("%s: Impossible, sk->sk_state=%d\n",
4120                       __func__, sk->sk_state);
4121                break;
4122        }
4123
4124        /* It _is_ possible, that we have something out-of-order _after_ FIN.
4125         * Probably, we should reset in this case. For now drop them.
4126         */
4127        __skb_queue_purge(&tp->out_of_order_queue);
4128        if (tcp_is_sack(tp))
4129                tcp_sack_reset(&tp->rx_opt);
4130        sk_mem_reclaim(sk);
4131
4132        if (!sock_flag(sk, SOCK_DEAD)) {
4133                sk->sk_state_change(sk);
4134
4135                /* Do not send POLL_HUP for half duplex close. */
4136                if (sk->sk_shutdown == SHUTDOWN_MASK ||
4137                    sk->sk_state == TCP_CLOSE)
4138                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4139                else
4140                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4141        }
4142}
4143
4144static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4145                                  u32 end_seq)
4146{
4147        if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4148                if (before(seq, sp->start_seq))
4149                        sp->start_seq = seq;
4150                if (after(end_seq, sp->end_seq))
4151                        sp->end_seq = end_seq;
4152                return true;
4153        }
4154        return false;
4155}
4156
4157static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4158{
4159        struct tcp_sock *tp = tcp_sk(sk);
4160
4161        if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4162                int mib_idx;
4163
4164                if (before(seq, tp->rcv_nxt))
4165                        mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4166                else
4167                        mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4168
4169                NET_INC_STATS_BH(sock_net(sk), mib_idx);
4170
4171                tp->rx_opt.dsack = 1;
4172                tp->duplicate_sack[0].start_seq = seq;
4173                tp->duplicate_sack[0].end_seq = end_seq;
4174        }
4175}
4176
4177static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4178{
4179        struct tcp_sock *tp = tcp_sk(sk);
4180
4181        if (!tp->rx_opt.dsack)
4182                tcp_dsack_set(sk, seq, end_seq);
4183        else
4184                tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4185}
4186
4187static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4188{
4189        struct tcp_sock *tp = tcp_sk(sk);
4190
4191        if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4192            before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4193                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4194                tcp_enter_quickack_mode(sk);
4195
4196                if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4197                        u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4198
4199                        if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4200                                end_seq = tp->rcv_nxt;
4201                        tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4202                }
4203        }
4204
4205        tcp_send_ack(sk);
4206}
4207
4208/* These routines update the SACK block as out-of-order packets arrive or
4209 * in-order packets close up the sequence space.
4210 */
4211static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4212{
4213        int this_sack;
4214        struct tcp_sack_block *sp = &tp->selective_acks[0];
4215        struct tcp_sack_block *swalk = sp + 1;
4216
4217        /* See if the recent change to the first SACK eats into
4218         * or hits the sequence space of other SACK blocks, if so coalesce.
4219         */
4220        for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4221                if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4222                        int i;
4223
4224                        /* Zap SWALK, by moving every further SACK up by one slot.
4225                         * Decrease num_sacks.
4226                         */
4227                        tp->rx_opt.num_sacks--;
4228                        for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4229                                sp[i] = sp[i + 1];
4230                        continue;
4231                }
4232                this_sack++, swalk++;
4233        }
4234}
4235
4236static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4237{
4238        struct tcp_sock *tp = tcp_sk(sk);
4239        struct tcp_sack_block *sp = &tp->selective_acks[0];
4240        int cur_sacks = tp->rx_opt.num_sacks;
4241        int this_sack;
4242
4243        if (!cur_sacks)
4244                goto new_sack;
4245
4246        for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4247                if (tcp_sack_extend(sp, seq, end_seq)) {
4248                        /* Rotate this_sack to the first one. */
4249                        for (; this_sack > 0; this_sack--, sp--)
4250                                swap(*sp, *(sp - 1));
4251                        if (cur_sacks > 1)
4252                                tcp_sack_maybe_coalesce(tp);
4253                        return;
4254                }
4255        }
4256
4257        /* Could not find an adjacent existing SACK, build a new one,
4258         * put it at the front, and shift everyone else down.  We
4259         * always know there is at least one SACK present already here.
4260         *
4261         * If the sack array is full, forget about the last one.
4262         */
4263        if (this_sack >= TCP_NUM_SACKS) {
4264                this_sack--;
4265                tp->rx_opt.num_sacks--;
4266                sp--;
4267        }
4268        for (; this_sack > 0; this_sack--, sp--)
4269                *sp = *(sp - 1);
4270
4271new_sack:
4272        /* Build the new head SACK, and we're done. */
4273        sp->start_seq = seq;
4274        sp->end_seq = end_seq;
4275        tp->rx_opt.num_sacks++;
4276}
4277
4278/* RCV.NXT advances, some SACKs should be eaten. */
4279
4280static void tcp_sack_remove(struct tcp_sock *tp)
4281{
4282        struct tcp_sack_block *sp = &tp->selective_acks[0];
4283        int num_sacks = tp->rx_opt.num_sacks;
4284        int this_sack;
4285
4286        /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4287        if (skb_queue_empty(&tp->out_of_order_queue)) {
4288                tp->rx_opt.num_sacks = 0;
4289                return;
4290        }
4291
4292        for (this_sack = 0; this_sack < num_sacks;) {
4293                /* Check if the start of the sack is covered by RCV.NXT. */
4294                if (!before(tp->rcv_nxt, sp->start_seq)) {
4295                        int i;
4296
4297                        /* RCV.NXT must cover all the block! */
4298                        WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4299
4300                        /* Zap this SACK, by moving forward any other SACKS. */
4301                        for (i=this_sack+1; i < num_sacks; i++)
4302                                tp->selective_acks[i-1] = tp->selective_acks[i];
4303                        num_sacks--;
4304                        continue;
4305                }
4306                this_sack++;
4307                sp++;
4308        }
4309        tp->rx_opt.num_sacks = num_sacks;
4310}
4311
4312/* This one checks to see if we can put data from the
4313 * out_of_order queue into the receive_queue.
4314 */
4315static void tcp_ofo_queue(struct sock *sk)
4316{
4317        struct tcp_sock *tp = tcp_sk(sk);
4318        __u32 dsack_high = tp->rcv_nxt;
4319        struct sk_buff *skb;
4320
4321        while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4322                if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4323                        break;
4324
4325                if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4326                        __u32 dsack = dsack_high;
4327                        if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4328                                dsack_high = TCP_SKB_CB(skb)->end_seq;
4329                        tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4330                }
4331
4332                if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4333                        SOCK_DEBUG(sk, "ofo packet was already received\n");
4334                        __skb_unlink(skb, &tp->out_of_order_queue);
4335                        __kfree_skb(skb);
4336                        continue;
4337                }
4338                SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4339                           tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4340                           TCP_SKB_CB(skb)->end_seq);
4341
4342                __skb_unlink(skb, &tp->out_of_order_queue);
4343                __skb_queue_tail(&sk->sk_receive_queue, skb);
4344                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4345                if (tcp_hdr(skb)->fin)
4346                        tcp_fin(sk);
4347        }
4348}
4349
4350static bool tcp_prune_ofo_queue(struct sock *sk);
4351static int tcp_prune_queue(struct sock *sk);
4352
4353static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4354                                 unsigned int size)
4355{
4356        if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4357            !sk_rmem_schedule(sk, skb, size)) {
4358
4359                if (tcp_prune_queue(sk) < 0)
4360                        return -1;
4361
4362                if (!sk_rmem_schedule(sk, skb, size)) {
4363                        if (!tcp_prune_ofo_queue(sk))
4364                                return -1;
4365
4366                        if (!sk_rmem_schedule(sk, skb, size))
4367                                return -1;
4368                }
4369        }
4370        return 0;
4371}
4372
4373/**
4374 * tcp_try_coalesce - try to merge skb to prior one
4375 * @sk: socket
4376 * @to: prior buffer
4377 * @from: buffer to add in queue
4378 * @fragstolen: pointer to boolean
4379 *
4380 * Before queueing skb @from after @to, try to merge them
4381 * to reduce overall memory use and queue lengths, if cost is small.
4382 * Packets in ofo or receive queues can stay a long time.
4383 * Better try to coalesce them right now to avoid future collapses.
4384 * Returns true if caller should free @from instead of queueing it
4385 */
4386static bool tcp_try_coalesce(struct sock *sk,
4387                             struct sk_buff *to,
4388                             struct sk_buff *from,
4389                             bool *fragstolen)
4390{
4391        int delta;
4392
4393        *fragstolen = false;
4394
4395        if (tcp_hdr(from)->fin)
4396                return false;
4397
4398        /* Its possible this segment overlaps with prior segment in queue */
4399        if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4400                return false;
4401
4402        if (!skb_try_coalesce(to, from, fragstolen, &delta))
4403                return false;
4404
4405        atomic_add(delta, &sk->sk_rmem_alloc);
4406        sk_mem_charge(sk, delta);
4407        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4408        TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4409        TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4410        return true;
4411}
4412
4413static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4414{
4415        struct tcp_sock *tp = tcp_sk(sk);
4416        struct sk_buff *skb1;
4417        u32 seq, end_seq;
4418
4419        TCP_ECN_check_ce(tp, skb);
4420
4421        if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4422                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4423                __kfree_skb(skb);
4424                return;
4425        }
4426
4427        /* Disable header prediction. */
4428        tp->pred_flags = 0;
4429        inet_csk_schedule_ack(sk);
4430
4431        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4432        SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4433                   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4434
4435        skb1 = skb_peek_tail(&tp->out_of_order_queue);
4436        if (!skb1) {
4437                /* Initial out of order segment, build 1 SACK. */
4438                if (tcp_is_sack(tp)) {
4439                        tp->rx_opt.num_sacks = 1;
4440                        tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4441                        tp->selective_acks[0].end_seq =
4442                                                TCP_SKB_CB(skb)->end_seq;
4443                }
4444                __skb_queue_head(&tp->out_of_order_queue, skb);
4445                goto end;
4446        }
4447
4448        seq = TCP_SKB_CB(skb)->seq;
4449        end_seq = TCP_SKB_CB(skb)->end_seq;
4450
4451        if (seq == TCP_SKB_CB(skb1)->end_seq) {
4452                bool fragstolen;
4453
4454                if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4455                        __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4456                } else {
4457                        kfree_skb_partial(skb, fragstolen);
4458                        skb = NULL;
4459                }
4460
4461                if (!tp->rx_opt.num_sacks ||
4462                    tp->selective_acks[0].end_seq != seq)
4463                        goto add_sack;
4464
4465                /* Common case: data arrive in order after hole. */
4466                tp->selective_acks[0].end_seq = end_seq;
4467                goto end;
4468        }
4469
4470        /* Find place to insert this segment. */
4471        while (1) {
4472                if (!after(TCP_SKB_CB(skb1)->seq, seq))
4473                        break;
4474                if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4475                        skb1 = NULL;
4476                        break;
4477                }
4478                skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4479        }
4480
4481        /* Do skb overlap to previous one? */
4482        if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4483                if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4484                        /* All the bits are present. Drop. */
4485                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4486                        __kfree_skb(skb);
4487                        skb = NULL;
4488                        tcp_dsack_set(sk, seq, end_seq);
4489                        goto add_sack;
4490                }
4491                if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4492                        /* Partial overlap. */
4493                        tcp_dsack_set(sk, seq,
4494                                      TCP_SKB_CB(skb1)->end_seq);
4495                } else {
4496                        if (skb_queue_is_first(&tp->out_of_order_queue,
4497                                               skb1))
4498                                skb1 = NULL;
4499                        else
4500                                skb1 = skb_queue_prev(
4501                                        &tp->out_of_order_queue,
4502                                        skb1);
4503                }
4504        }
4505        if (!skb1)
4506                __skb_queue_head(&tp->out_of_order_queue, skb);
4507        else
4508                __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4509
4510        /* And clean segments covered by new one as whole. */
4511        while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4512                skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4513
4514                if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4515                        break;
4516                if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4517                        tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4518                                         end_seq);
4519                        break;
4520                }
4521                __skb_unlink(skb1, &tp->out_of_order_queue);
4522                tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4523                                 TCP_SKB_CB(skb1)->end_seq);
4524                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4525                __kfree_skb(skb1);
4526        }
4527
4528add_sack:
4529        if (tcp_is_sack(tp))
4530                tcp_sack_new_ofo_skb(sk, seq, end_seq);
4531end:
4532        if (skb)
4533                skb_set_owner_r(skb, sk);
4534}
4535
4536static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4537                  bool *fragstolen)
4538{
4539        int eaten;
4540        struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4541
4542        __skb_pull(skb, hdrlen);
4543        eaten = (tail &&
4544                 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4545        tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4546        if (!eaten) {
4547                __skb_queue_tail(&sk->sk_receive_queue, skb);
4548                skb_set_owner_r(skb, sk);
4549        }
4550        return eaten;
4551}
4552
4553int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4554{
4555        struct sk_buff *skb = NULL;
4556        struct tcphdr *th;
4557        bool fragstolen;
4558
4559        skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4560        if (!skb)
4561                goto err;
4562
4563        if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4564                goto err_free;
4565
4566        th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4567        skb_reset_transport_header(skb);
4568        memset(th, 0, sizeof(*th));
4569
4570        if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4571                goto err_free;
4572
4573        TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4574        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4575        TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4576
4577        if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4578                WARN_ON_ONCE(fragstolen); /* should not happen */
4579                __kfree_skb(skb);
4580        }
4581        return size;
4582
4583err_free:
4584        kfree_skb(skb);
4585err:
4586        return -ENOMEM;
4587}
4588
4589static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4590{
4591        const struct tcphdr *th = tcp_hdr(skb);
4592        struct tcp_sock *tp = tcp_sk(sk);
4593        int eaten = -1;
4594        bool fragstolen = false;
4595
4596        if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4597                goto drop;
4598
4599        skb_dst_drop(skb);
4600        __skb_pull(skb, th->doff * 4);
4601
4602        TCP_ECN_accept_cwr(tp, skb);
4603
4604        tp->rx_opt.dsack = 0;
4605
4606        /*  Queue data for delivery to the user.
4607         *  Packets in sequence go to the receive queue.
4608         *  Out of sequence packets to the out_of_order_queue.
4609         */
4610        if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4611                if (tcp_receive_window(tp) == 0)
4612                        goto out_of_window;
4613
4614                /* Ok. In sequence. In window. */
4615                if (tp->ucopy.task == current &&
4616                    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4617                    sock_owned_by_user(sk) && !tp->urg_data) {
4618                        int chunk = min_t(unsigned int, skb->len,
4619                                          tp->ucopy.len);
4620
4621                        __set_current_state(TASK_RUNNING);
4622
4623                        local_bh_enable();
4624                        if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4625                                tp->ucopy.len -= chunk;
4626                                tp->copied_seq += chunk;
4627                                eaten = (chunk == skb->len);
4628                                tcp_rcv_space_adjust(sk);
4629                        }
4630                        local_bh_disable();
4631                }
4632
4633                if (eaten <= 0) {
4634queue_and_out:
4635                        if (eaten < 0 &&
4636                            tcp_try_rmem_schedule(sk, skb, skb->truesize))
4637                                goto drop;
4638
4639                        eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4640                }
4641                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4642                if (skb->len)
4643                        tcp_event_data_recv(sk, skb);
4644                if (th->fin)
4645                        tcp_fin(sk);
4646
4647                if (!skb_queue_empty(&tp->out_of_order_queue)) {
4648                        tcp_ofo_queue(sk);
4649
4650                        /* RFC2581. 4.2. SHOULD send immediate ACK, when
4651                         * gap in queue is filled.
4652                         */
4653                        if (skb_queue_empty(&tp->out_of_order_queue))
4654                                inet_csk(sk)->icsk_ack.pingpong = 0;
4655                }
4656
4657                if (tp->rx_opt.num_sacks)
4658                        tcp_sack_remove(tp);
4659
4660                tcp_fast_path_check(sk);
4661
4662                if (eaten > 0)
4663                        kfree_skb_partial(skb, fragstolen);
4664                if (!sock_flag(sk, SOCK_DEAD))
4665                        sk->sk_data_ready(sk, 0);
4666                return;
4667        }
4668
4669        if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4670                /* A retransmit, 2nd most common case.  Force an immediate ack. */
4671                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4672                tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4673
4674out_of_window:
4675                tcp_enter_quickack_mode(sk);
4676                inet_csk_schedule_ack(sk);
4677drop:
4678                __kfree_skb(skb);
4679                return;
4680        }
4681
4682        /* Out of window. F.e. zero window probe. */
4683        if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4684                goto out_of_window;
4685
4686        tcp_enter_quickack_mode(sk);
4687
4688        if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4689                /* Partial packet, seq < rcv_next < end_seq */
4690                SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4691                           tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4692                           TCP_SKB_CB(skb)->end_seq);
4693
4694                tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4695
4696                /* If window is closed, drop tail of packet. But after
4697                 * remembering D-SACK for its head made in previous line.
4698                 */
4699                if (!tcp_receive_window(tp))
4700                        goto out_of_window;
4701                goto queue_and_out;
4702        }
4703
4704        tcp_data_queue_ofo(sk, skb);
4705}
4706
4707static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4708                                        struct sk_buff_head *list)
4709{
4710        struct sk_buff *next = NULL;
4711
4712        if (!skb_queue_is_last(list, skb))
4713                next = skb_queue_next(list, skb);
4714
4715        __skb_unlink(skb, list);
4716        __kfree_skb(skb);
4717        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4718
4719        return next;
4720}
4721
4722/* Collapse contiguous sequence of skbs head..tail with
4723 * sequence numbers start..end.
4724 *
4725 * If tail is NULL, this means until the end of the list.
4726 *
4727 * Segments with FIN/SYN are not collapsed (only because this
4728 * simplifies code)
4729 */
4730static void
4731tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4732             struct sk_buff *head, struct sk_buff *tail,
4733             u32 start, u32 end)
4734{
4735        struct sk_buff *skb, *n;
4736        bool end_of_skbs;
4737
4738        /* First, check that queue is collapsible and find
4739         * the point where collapsing can be useful. */
4740        skb = head;
4741restart:
4742        end_of_skbs = true;
4743        skb_queue_walk_from_safe(list, skb, n) {
4744                if (skb == tail)
4745                        break;
4746                /* No new bits? It is possible on ofo queue. */
4747                if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4748                        skb = tcp_collapse_one(sk, skb, list);
4749                        if (!skb)
4750                                break;
4751                        goto restart;
4752                }
4753
4754                /* The first skb to collapse is:
4755                 * - not SYN/FIN and
4756                 * - bloated or contains data before "start" or
4757                 *   overlaps to the next one.
4758                 */
4759                if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4760                    (tcp_win_from_space(skb->truesize) > skb->len ||
4761                     before(TCP_SKB_CB(skb)->seq, start))) {
4762                        end_of_skbs = false;
4763                        break;
4764                }
4765
4766                if (!skb_queue_is_last(list, skb)) {
4767                        struct sk_buff *next = skb_queue_next(list, skb);
4768                        if (next != tail &&
4769                            TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4770                                end_of_skbs = false;
4771                                break;
4772                        }
4773                }
4774
4775                /* Decided to skip this, advance start seq. */
4776                start = TCP_SKB_CB(skb)->end_seq;
4777        }
4778        if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4779                return;
4780
4781        while (before(start, end)) {
4782                struct sk_buff *nskb;
4783                unsigned int header = skb_headroom(skb);
4784                int copy = SKB_MAX_ORDER(header, 0);
4785
4786                /* Too big header? This can happen with IPv6. */
4787                if (copy < 0)
4788                        return;
4789                if (end - start < copy)
4790                        copy = end - start;
4791                nskb = alloc_skb(copy + header, GFP_ATOMIC);
4792                if (!nskb)
4793                        return;
4794
4795                skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4796                skb_set_network_header(nskb, (skb_network_header(skb) -
4797                                              skb->head));
4798                skb_set_transport_header(nskb, (skb_transport_header(skb) -
4799                                                skb->head));
4800                skb_reserve(nskb, header);
4801                memcpy(nskb->head, skb->head, header);
4802                memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4803                TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4804                __skb_queue_before(list, skb, nskb);
4805                skb_set_owner_r(nskb, sk);
4806
4807                /* Copy data, releasing collapsed skbs. */
4808                while (copy > 0) {
4809                        int offset = start - TCP_SKB_CB(skb)->seq;
4810                        int size = TCP_SKB_CB(skb)->end_seq - start;
4811
4812                        BUG_ON(offset < 0);
4813                        if (size > 0) {
4814                                size = min(copy, size);
4815                                if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4816                                        BUG();
4817                                TCP_SKB_CB(nskb)->end_seq += size;
4818                                copy -= size;
4819                                start += size;
4820                        }
4821                        if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4822                                skb = tcp_collapse_one(sk, skb, list);
4823                                if (!skb ||
4824                                    skb == tail ||
4825                                    tcp_hdr(skb)->syn ||
4826                                    tcp_hdr(skb)->fin)
4827                                        return;
4828                        }
4829                }
4830        }
4831}
4832
4833/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4834 * and tcp_collapse() them until all the queue is collapsed.
4835 */
4836static void tcp_collapse_ofo_queue(struct sock *sk)
4837{
4838        struct tcp_sock *tp = tcp_sk(sk);
4839        struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4840        struct sk_buff *head;
4841        u32 start, end;
4842
4843        if (skb == NULL)
4844                return;
4845
4846        start = TCP_SKB_CB(skb)->seq;
4847        end = TCP_SKB_CB(skb)->end_seq;
4848        head = skb;
4849
4850        for (;;) {
4851                struct sk_buff *next = NULL;
4852
4853                if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4854                        next = skb_queue_next(&tp->out_of_order_queue, skb);
4855                skb = next;
4856
4857                /* Segment is terminated when we see gap or when
4858                 * we are at the end of all the queue. */
4859                if (!skb ||
4860                    after(TCP_SKB_CB(skb)->seq, end) ||
4861                    before(TCP_SKB_CB(skb)->end_seq, start)) {
4862                        tcp_collapse(sk, &tp->out_of_order_queue,
4863                                     head, skb, start, end);
4864                        head = skb;
4865                        if (!skb)
4866                                break;
4867                        /* Start new segment */
4868                        start = TCP_SKB_CB(skb)->seq;
4869                        end = TCP_SKB_CB(skb)->end_seq;
4870                } else {
4871                        if (before(TCP_SKB_CB(skb)->seq, start))
4872                                start = TCP_SKB_CB(skb)->seq;
4873                        if (after(TCP_SKB_CB(skb)->end_seq, end))
4874                                end = TCP_SKB_CB(skb)->end_seq;
4875                }
4876        }
4877}
4878
4879/*
4880 * Purge the out-of-order queue.
4881 * Return true if queue was pruned.
4882 */
4883static bool tcp_prune_ofo_queue(struct sock *sk)
4884{
4885        struct tcp_sock *tp = tcp_sk(sk);
4886        bool res = false;
4887
4888        if (!skb_queue_empty(&tp->out_of_order_queue)) {
4889                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4890                __skb_queue_purge(&tp->out_of_order_queue);
4891
4892                /* Reset SACK state.  A conforming SACK implementation will
4893                 * do the same at a timeout based retransmit.  When a connection
4894                 * is in a sad state like this, we care only about integrity
4895                 * of the connection not performance.
4896                 */
4897                if (tp->rx_opt.sack_ok)
4898                        tcp_sack_reset(&tp->rx_opt);
4899                sk_mem_reclaim(sk);
4900                res = true;
4901        }
4902        return res;
4903}
4904
4905/* Reduce allocated memory if we can, trying to get
4906 * the socket within its memory limits again.
4907 *
4908 * Return less than zero if we should start dropping frames
4909 * until the socket owning process reads some of the data
4910 * to stabilize the situation.
4911 */
4912static int tcp_prune_queue(struct sock *sk)
4913{
4914        struct tcp_sock *tp = tcp_sk(sk);
4915
4916        SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4917
4918        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4919
4920        if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4921                tcp_clamp_window(sk);
4922        else if (sk_under_memory_pressure(sk))
4923                tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4924
4925        tcp_collapse_ofo_queue(sk);
4926        if (!skb_queue_empty(&sk->sk_receive_queue))
4927                tcp_collapse(sk, &sk->sk_receive_queue,
4928                             skb_peek(&sk->sk_receive_queue),
4929                             NULL,
4930                             tp->copied_seq, tp->rcv_nxt);
4931        sk_mem_reclaim(sk);
4932
4933        if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4934                return 0;
4935
4936        /* Collapsing did not help, destructive actions follow.
4937         * This must not ever occur. */
4938
4939        tcp_prune_ofo_queue(sk);
4940
4941        if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4942                return 0;
4943
4944        /* If we are really being abused, tell the caller to silently
4945         * drop receive data on the floor.  It will get retransmitted
4946         * and hopefully then we'll have sufficient space.
4947         */
4948        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4949
4950        /* Massive buffer overcommit. */
4951        tp->pred_flags = 0;
4952        return -1;
4953}
4954
4955/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4956 * As additional protections, we do not touch cwnd in retransmission phases,
4957 * and if application hit its sndbuf limit recently.
4958 */
4959void tcp_cwnd_application_limited(struct sock *sk)
4960{
4961        struct tcp_sock *tp = tcp_sk(sk);
4962
4963        if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4964            sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4965                /* Limited by application or receiver window. */
4966                u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4967                u32 win_used = max(tp->snd_cwnd_used, init_win);
4968                if (win_used < tp->snd_cwnd) {
4969                        tp->snd_ssthresh = tcp_current_ssthresh(sk);
4970                        tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4971                }
4972                tp->snd_cwnd_used = 0;
4973        }
4974        tp->snd_cwnd_stamp = tcp_time_stamp;
4975}
4976
4977static bool tcp_should_expand_sndbuf(const struct sock *sk)
4978{
4979        const struct tcp_sock *tp = tcp_sk(sk);
4980
4981        /* If the user specified a specific send buffer setting, do
4982         * not modify it.
4983         */
4984        if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4985                return false;
4986
4987        /* If we are under global TCP memory pressure, do not expand.  */
4988        if (sk_under_memory_pressure(sk))
4989                return false;
4990
4991        /* If we are under soft global TCP memory pressure, do not expand.  */
4992        if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4993                return false;
4994
4995        /* If we filled the congestion window, do not expand.  */
4996        if (tp->packets_out >= tp->snd_cwnd)
4997                return false;
4998
4999        return true;
5000}
5001
5002/* When incoming ACK allowed to free some skb from write_queue,
5003 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5004 * on the exit from tcp input handler.
5005 *
5006 * PROBLEM: sndbuf expansion does not work well with largesend.
5007 */
5008static void tcp_new_space(struct sock *sk)
5009{
5010        struct tcp_sock *tp = tcp_sk(sk);
5011
5012        if (tcp_should_expand_sndbuf(sk)) {
5013                int sndmem = SKB_TRUESIZE(max_t(u32,
5014                                                tp->rx_opt.mss_clamp,
5015                                                tp->mss_cache) +
5016                                          MAX_TCP_HEADER);
5017                int demanded = max_t(unsigned int, tp->snd_cwnd,
5018                                     tp->reordering + 1);
5019                sndmem *= 2 * demanded;
5020                if (sndmem > sk->sk_sndbuf)
5021                        sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5022                tp->snd_cwnd_stamp = tcp_time_stamp;
5023        }
5024
5025        sk->sk_write_space(sk);
5026}
5027
5028static void tcp_check_space(struct sock *sk)
5029{
5030        if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5031                sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5032                if (sk->sk_socket &&
5033                    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5034                        tcp_new_space(sk);
5035        }
5036}
5037
5038static inline void tcp_data_snd_check(struct sock *sk)
5039{
5040        tcp_push_pending_frames(sk);
5041        tcp_check_space(sk);
5042}
5043
5044/*
5045 * Check if sending an ack is needed.
5046 */
5047static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5048{
5049        struct tcp_sock *tp = tcp_sk(sk);
5050
5051            /* More than one full frame received... */
5052        if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5053             /* ... and right edge of window advances far enough.
5054              * (tcp_recvmsg() will send ACK otherwise). Or...
5055              */
5056             __tcp_select_window(sk) >= tp->rcv_wnd) ||
5057            /* We ACK each frame or... */
5058            tcp_in_quickack_mode(sk) ||
5059            /* We have out of order data. */
5060            (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5061                /* Then ack it now */
5062                tcp_send_ack(sk);
5063        } else {
5064                /* Else, send delayed ack. */
5065                tcp_send_delayed_ack(sk);
5066        }
5067}
5068
5069static inline void tcp_ack_snd_check(struct sock *sk)
5070{
5071        if (!inet_csk_ack_scheduled(sk)) {
5072                /* We sent a data segment already. */
5073                return;
5074        }
5075        __tcp_ack_snd_check(sk, 1);
5076}
5077
5078/*
5079 *      This routine is only called when we have urgent data
5080 *      signaled. Its the 'slow' part of tcp_urg. It could be
5081 *      moved inline now as tcp_urg is only called from one
5082 *      place. We handle URGent data wrong. We have to - as
5083 *      BSD still doesn't use the correction from RFC961.
5084 *      For 1003.1g we should support a new option TCP_STDURG to permit
5085 *      either form (or just set the sysctl tcp_stdurg).
5086 */
5087
5088static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5089{
5090        struct tcp_sock *tp = tcp_sk(sk);
5091        u32 ptr = ntohs(th->urg_ptr);
5092
5093        if (ptr && !sysctl_tcp_stdurg)
5094                ptr--;
5095        ptr += ntohl(th->seq);
5096
5097        /* Ignore urgent data that we've already seen and read. */
5098        if (after(tp->copied_seq, ptr))
5099                return;
5100
5101        /* Do not replay urg ptr.
5102         *
5103         * NOTE: interesting situation not covered by specs.
5104         * Misbehaving sender may send urg ptr, pointing to segment,
5105         * which we already have in ofo queue. We are not able to fetch
5106         * such data and will stay in TCP_URG_NOTYET until will be eaten
5107         * by recvmsg(). Seems, we are not obliged to handle such wicked
5108         * situations. But it is worth to think about possibility of some
5109         * DoSes using some hypothetical application level deadlock.
5110         */
5111        if (before(ptr, tp->rcv_nxt))
5112                return;
5113
5114        /* Do we already have a newer (or duplicate) urgent pointer? */
5115        if (tp->urg_data && !after(ptr, tp->urg_seq))
5116                return;
5117
5118        /* Tell the world about our new urgent pointer. */
5119        sk_send_sigurg(sk);
5120
5121        /* We may be adding urgent data when the last byte read was
5122         * urgent. To do this requires some care. We cannot just ignore
5123         * tp->copied_seq since we would read the last urgent byte again
5124         * as data, nor can we alter copied_seq until this data arrives
5125         * or we break the semantics of SIOCATMARK (and thus sockatmark())
5126         *
5127         * NOTE. Double Dutch. Rendering to plain English: author of comment
5128         * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5129         * and expect that both A and B disappear from stream. This is _wrong_.
5130         * Though this happens in BSD with high probability, this is occasional.
5131         * Any application relying on this is buggy. Note also, that fix "works"
5132         * only in this artificial test. Insert some normal data between A and B and we will
5133         * decline of BSD again. Verdict: it is better to remove to trap
5134         * buggy users.
5135         */
5136        if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5137            !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5138                struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5139                tp->copied_seq++;
5140                if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5141                        __skb_unlink(skb, &sk->sk_receive_queue);
5142                        __kfree_skb(skb);
5143                }
5144        }
5145
5146        tp->urg_data = TCP_URG_NOTYET;
5147        tp->urg_seq = ptr;
5148
5149        /* Disable header prediction. */
5150        tp->pred_flags = 0;
5151}
5152
5153/* This is the 'fast' part of urgent handling. */
5154static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5155{
5156        struct tcp_sock *tp = tcp_sk(sk);
5157
5158        /* Check if we get a new urgent pointer - normally not. */
5159<