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                if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
 241                        /* Better not delay acks, sender can have a very low cwnd */
 242                        tcp_enter_quickack_mode((struct sock *)tp);
 243                        tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
 244                }
 245                /* fallinto */
 246        default:
 247                tp->ecn_flags |= TCP_ECN_SEEN;
 248        }
 249}
 250
 251static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
 252{
 253        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
 254                tp->ecn_flags &= ~TCP_ECN_OK;
 255}
 256
 257static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
 258{
 259        if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
 260                tp->ecn_flags &= ~TCP_ECN_OK;
 261}
 262
 263static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
 264{
 265        if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
 266                return true;
 267        return false;
 268}
 269
 270/* Buffer size and advertised window tuning.
 271 *
 272 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
 273 */
 274
 275static void tcp_fixup_sndbuf(struct sock *sk)
 276{
 277        int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
 278
 279        sndmem *= TCP_INIT_CWND;
 280        if (sk->sk_sndbuf < sndmem)
 281                sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
 282}
 283
 284/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
 285 *
 286 * All tcp_full_space() is split to two parts: "network" buffer, allocated
 287 * forward and advertised in receiver window (tp->rcv_wnd) and
 288 * "application buffer", required to isolate scheduling/application
 289 * latencies from network.
 290 * window_clamp is maximal advertised window. It can be less than
 291 * tcp_full_space(), in this case tcp_full_space() - window_clamp
 292 * is reserved for "application" buffer. The less window_clamp is
 293 * the smoother our behaviour from viewpoint of network, but the lower
 294 * throughput and the higher sensitivity of the connection to losses. 8)
 295 *
 296 * rcv_ssthresh is more strict window_clamp used at "slow start"
 297 * phase to predict further behaviour of this connection.
 298 * It is used for two goals:
 299 * - to enforce header prediction at sender, even when application
 300 *   requires some significant "application buffer". It is check #1.
 301 * - to prevent pruning of receive queue because of misprediction
 302 *   of receiver window. Check #2.
 303 *
 304 * The scheme does not work when sender sends good segments opening
 305 * window and then starts to feed us spaghetti. But it should work
 306 * in common situations. Otherwise, we have to rely on queue collapsing.
 307 */
 308
 309/* Slow part of check#2. */
 310static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
 311{
 312        struct tcp_sock *tp = tcp_sk(sk);
 313        /* Optimize this! */
 314        int truesize = tcp_win_from_space(skb->truesize) >> 1;
 315        int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
 316
 317        while (tp->rcv_ssthresh <= window) {
 318                if (truesize <= skb->len)
 319                        return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
 320
 321                truesize >>= 1;
 322                window >>= 1;
 323        }
 324        return 0;
 325}
 326
 327static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
 328{
 329        struct tcp_sock *tp = tcp_sk(sk);
 330
 331        /* Check #1 */
 332        if (tp->rcv_ssthresh < tp->window_clamp &&
 333            (int)tp->rcv_ssthresh < tcp_space(sk) &&
 334            !sk_under_memory_pressure(sk)) {
 335                int incr;
 336
 337                /* Check #2. Increase window, if skb with such overhead
 338                 * will fit to rcvbuf in future.
 339                 */
 340                if (tcp_win_from_space(skb->truesize) <= skb->len)
 341                        incr = 2 * tp->advmss;
 342                else
 343                        incr = __tcp_grow_window(sk, skb);
 344
 345                if (incr) {
 346                        incr = max_t(int, incr, 2 * skb->len);
 347                        tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
 348                                               tp->window_clamp);
 349                        inet_csk(sk)->icsk_ack.quick |= 1;
 350                }
 351        }
 352}
 353
 354/* 3. Tuning rcvbuf, when connection enters established state. */
 355
 356static void tcp_fixup_rcvbuf(struct sock *sk)
 357{
 358        u32 mss = tcp_sk(sk)->advmss;
 359        u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
 360        int rcvmem;
 361
 362        /* Limit to 10 segments if mss <= 1460,
 363         * or 14600/mss segments, with a minimum of two segments.
 364         */
 365        if (mss > 1460)
 366                icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
 367
 368        rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
 369        while (tcp_win_from_space(rcvmem) < mss)
 370                rcvmem += 128;
 371
 372        rcvmem *= icwnd;
 373
 374        if (sk->sk_rcvbuf < rcvmem)
 375                sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
 376}
 377
 378/* 4. Try to fixup all. It is made immediately after connection enters
 379 *    established state.
 380 */
 381void tcp_init_buffer_space(struct sock *sk)
 382{
 383        struct tcp_sock *tp = tcp_sk(sk);
 384        int maxwin;
 385
 386        if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 387                tcp_fixup_rcvbuf(sk);
 388        if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
 389                tcp_fixup_sndbuf(sk);
 390
 391        tp->rcvq_space.space = tp->rcv_wnd;
 392
 393        maxwin = tcp_full_space(sk);
 394
 395        if (tp->window_clamp >= maxwin) {
 396                tp->window_clamp = maxwin;
 397
 398                if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
 399                        tp->window_clamp = max(maxwin -
 400                                               (maxwin >> sysctl_tcp_app_win),
 401                                               4 * tp->advmss);
 402        }
 403
 404        /* Force reservation of one segment. */
 405        if (sysctl_tcp_app_win &&
 406            tp->window_clamp > 2 * tp->advmss &&
 407            tp->window_clamp + tp->advmss > maxwin)
 408                tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
 409
 410        tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
 411        tp->snd_cwnd_stamp = tcp_time_stamp;
 412}
 413
 414/* 5. Recalculate window clamp after socket hit its memory bounds. */
 415static void tcp_clamp_window(struct sock *sk)
 416{
 417        struct tcp_sock *tp = tcp_sk(sk);
 418        struct inet_connection_sock *icsk = inet_csk(sk);
 419
 420        icsk->icsk_ack.quick = 0;
 421
 422        if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 423            !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
 424            !sk_under_memory_pressure(sk) &&
 425            sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
 426                sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 427                                    sysctl_tcp_rmem[2]);
 428        }
 429        if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 430                tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
 431}
 432
 433/* Initialize RCV_MSS value.
 434 * RCV_MSS is an our guess about MSS used by the peer.
 435 * We haven't any direct information about the MSS.
 436 * It's better to underestimate the RCV_MSS rather than overestimate.
 437 * Overestimations make us ACKing less frequently than needed.
 438 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
 439 */
 440void tcp_initialize_rcv_mss(struct sock *sk)
 441{
 442        const struct tcp_sock *tp = tcp_sk(sk);
 443        unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
 444
 445        hint = min(hint, tp->rcv_wnd / 2);
 446        hint = min(hint, TCP_MSS_DEFAULT);
 447        hint = max(hint, TCP_MIN_MSS);
 448
 449        inet_csk(sk)->icsk_ack.rcv_mss = hint;
 450}
 451EXPORT_SYMBOL(tcp_initialize_rcv_mss);
 452
 453/* Receiver "autotuning" code.
 454 *
 455 * The algorithm for RTT estimation w/o timestamps is based on
 456 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
 457 * <http://public.lanl.gov/radiant/pubs.html#DRS>
 458 *
 459 * More detail on this code can be found at
 460 * <http://staff.psc.edu/jheffner/>,
 461 * though this reference is out of date.  A new paper
 462 * is pending.
 463 */
 464static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 465{
 466        u32 new_sample = tp->rcv_rtt_est.rtt;
 467        long m = sample;
 468
 469        if (m == 0)
 470                m = 1;
 471
 472        if (new_sample != 0) {
 473                /* If we sample in larger samples in the non-timestamp
 474                 * case, we could grossly overestimate the RTT especially
 475                 * with chatty applications or bulk transfer apps which
 476                 * are stalled on filesystem I/O.
 477                 *
 478                 * Also, since we are only going for a minimum in the
 479                 * non-timestamp case, we do not smooth things out
 480                 * else with timestamps disabled convergence takes too
 481                 * long.
 482                 */
 483                if (!win_dep) {
 484                        m -= (new_sample >> 3);
 485                        new_sample += m;
 486                } else {
 487                        m <<= 3;
 488                        if (m < new_sample)
 489                                new_sample = m;
 490                }
 491        } else {
 492                /* No previous measure. */
 493                new_sample = m << 3;
 494        }
 495
 496        if (tp->rcv_rtt_est.rtt != new_sample)
 497                tp->rcv_rtt_est.rtt = new_sample;
 498}
 499
 500static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
 501{
 502        if (tp->rcv_rtt_est.time == 0)
 503                goto new_measure;
 504        if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
 505                return;
 506        tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
 507
 508new_measure:
 509        tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
 510        tp->rcv_rtt_est.time = tcp_time_stamp;
 511}
 512
 513static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 514                                          const struct sk_buff *skb)
 515{
 516        struct tcp_sock *tp = tcp_sk(sk);
 517        if (tp->rx_opt.rcv_tsecr &&
 518            (TCP_SKB_CB(skb)->end_seq -
 519             TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
 520                tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
 521}
 522
 523/*
 524 * This function should be called every time data is copied to user space.
 525 * It calculates the appropriate TCP receive buffer space.
 526 */
 527void tcp_rcv_space_adjust(struct sock *sk)
 528{
 529        struct tcp_sock *tp = tcp_sk(sk);
 530        int time;
 531        int space;
 532
 533        if (tp->rcvq_space.time == 0)
 534                goto new_measure;
 535
 536        time = tcp_time_stamp - tp->rcvq_space.time;
 537        if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
 538                return;
 539
 540        space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
 541
 542        space = max(tp->rcvq_space.space, space);
 543
 544        if (tp->rcvq_space.space != space) {
 545                int rcvmem;
 546
 547                tp->rcvq_space.space = space;
 548
 549                if (sysctl_tcp_moderate_rcvbuf &&
 550                    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 551                        int new_clamp = space;
 552
 553                        /* Receive space grows, normalize in order to
 554                         * take into account packet headers and sk_buff
 555                         * structure overhead.
 556                         */
 557                        space /= tp->advmss;
 558                        if (!space)
 559                                space = 1;
 560                        rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
 561                        while (tcp_win_from_space(rcvmem) < tp->advmss)
 562                                rcvmem += 128;
 563                        space *= rcvmem;
 564                        space = min(space, sysctl_tcp_rmem[2]);
 565                        if (space > sk->sk_rcvbuf) {
 566                                sk->sk_rcvbuf = space;
 567
 568                                /* Make the window clamp follow along.  */
 569                                tp->window_clamp = new_clamp;
 570                        }
 571                }
 572        }
 573
 574new_measure:
 575        tp->rcvq_space.seq = tp->copied_seq;
 576        tp->rcvq_space.time = tcp_time_stamp;
 577}
 578
 579/* There is something which you must keep in mind when you analyze the
 580 * behavior of the tp->ato delayed ack timeout interval.  When a
 581 * connection starts up, we want to ack as quickly as possible.  The
 582 * problem is that "good" TCP's do slow start at the beginning of data
 583 * transmission.  The means that until we send the first few ACK's the
 584 * sender will sit on his end and only queue most of his data, because
 585 * he can only send snd_cwnd unacked packets at any given time.  For
 586 * each ACK we send, he increments snd_cwnd and transmits more of his
 587 * queue.  -DaveM
 588 */
 589static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
 590{
 591        struct tcp_sock *tp = tcp_sk(sk);
 592        struct inet_connection_sock *icsk = inet_csk(sk);
 593        u32 now;
 594
 595        inet_csk_schedule_ack(sk);
 596
 597        tcp_measure_rcv_mss(sk, skb);
 598
 599        tcp_rcv_rtt_measure(tp);
 600
 601        now = tcp_time_stamp;
 602
 603        if (!icsk->icsk_ack.ato) {
 604                /* The _first_ data packet received, initialize
 605                 * delayed ACK engine.
 606                 */
 607                tcp_incr_quickack(sk);
 608                icsk->icsk_ack.ato = TCP_ATO_MIN;
 609        } else {
 610                int m = now - icsk->icsk_ack.lrcvtime;
 611
 612                if (m <= TCP_ATO_MIN / 2) {
 613                        /* The fastest case is the first. */
 614                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
 615                } else if (m < icsk->icsk_ack.ato) {
 616                        icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
 617                        if (icsk->icsk_ack.ato > icsk->icsk_rto)
 618                                icsk->icsk_ack.ato = icsk->icsk_rto;
 619                } else if (m > icsk->icsk_rto) {
 620                        /* Too long gap. Apparently sender failed to
 621                         * restart window, so that we send ACKs quickly.
 622                         */
 623                        tcp_incr_quickack(sk);
 624                        sk_mem_reclaim(sk);
 625                }
 626        }
 627        icsk->icsk_ack.lrcvtime = now;
 628
 629        TCP_ECN_check_ce(tp, skb);
 630
 631        if (skb->len >= 128)
 632                tcp_grow_window(sk, skb);
 633}
 634
 635/* Called to compute a smoothed rtt estimate. The data fed to this
 636 * routine either comes from timestamps, or from segments that were
 637 * known _not_ to have been retransmitted [see Karn/Partridge
 638 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
 639 * piece by Van Jacobson.
 640 * NOTE: the next three routines used to be one big routine.
 641 * To save cycles in the RFC 1323 implementation it was better to break
 642 * it up into three procedures. -- erics
 643 */
 644static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
 645{
 646        struct tcp_sock *tp = tcp_sk(sk);
 647        long m = mrtt; /* RTT */
 648
 649        /*      The following amusing code comes from Jacobson's
 650         *      article in SIGCOMM '88.  Note that rtt and mdev
 651         *      are scaled versions of rtt and mean deviation.
 652         *      This is designed to be as fast as possible
 653         *      m stands for "measurement".
 654         *
 655         *      On a 1990 paper the rto value is changed to:
 656         *      RTO = rtt + 4 * mdev
 657         *
 658         * Funny. This algorithm seems to be very broken.
 659         * These formulae increase RTO, when it should be decreased, increase
 660         * too slowly, when it should be increased quickly, decrease too quickly
 661         * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
 662         * does not matter how to _calculate_ it. Seems, it was trap
 663         * that VJ failed to avoid. 8)
 664         */
 665        if (m == 0)
 666                m = 1;
 667        if (tp->srtt != 0) {
 668                m -= (tp->srtt >> 3);   /* m is now error in rtt est */
 669                tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
 670                if (m < 0) {
 671                        m = -m;         /* m is now abs(error) */
 672                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 673                        /* This is similar to one of Eifel findings.
 674                         * Eifel blocks mdev updates when rtt decreases.
 675                         * This solution is a bit different: we use finer gain
 676                         * for mdev in this case (alpha*beta).
 677                         * Like Eifel it also prevents growth of rto,
 678                         * but also it limits too fast rto decreases,
 679                         * happening in pure Eifel.
 680                         */
 681                        if (m > 0)
 682                                m >>= 3;
 683                } else {
 684                        m -= (tp->mdev >> 2);   /* similar update on mdev */
 685                }
 686                tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
 687                if (tp->mdev > tp->mdev_max) {
 688                        tp->mdev_max = tp->mdev;
 689                        if (tp->mdev_max > tp->rttvar)
 690                                tp->rttvar = tp->mdev_max;
 691                }
 692                if (after(tp->snd_una, tp->rtt_seq)) {
 693                        if (tp->mdev_max < tp->rttvar)
 694                                tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
 695                        tp->rtt_seq = tp->snd_nxt;
 696                        tp->mdev_max = tcp_rto_min(sk);
 697                }
 698        } else {
 699                /* no previous measure. */
 700                tp->srtt = m << 3;      /* take the measured time to be rtt */
 701                tp->mdev = m << 1;      /* make sure rto = 3*rtt */
 702                tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 703                tp->rtt_seq = tp->snd_nxt;
 704        }
 705}
 706
 707/* Calculate rto without backoff.  This is the second half of Van Jacobson's
 708 * routine referred to above.
 709 */
 710void tcp_set_rto(struct sock *sk)
 711{
 712        const struct tcp_sock *tp = tcp_sk(sk);
 713        /* Old crap is replaced with new one. 8)
 714         *
 715         * More seriously:
 716         * 1. If rtt variance happened to be less 50msec, it is hallucination.
 717         *    It cannot be less due to utterly erratic ACK generation made
 718         *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
 719         *    to do with delayed acks, because at cwnd>2 true delack timeout
 720         *    is invisible. Actually, Linux-2.4 also generates erratic
 721         *    ACKs in some circumstances.
 722         */
 723        inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
 724
 725        /* 2. Fixups made earlier cannot be right.
 726         *    If we do not estimate RTO correctly without them,
 727         *    all the algo is pure shit and should be replaced
 728         *    with correct one. It is exactly, which we pretend to do.
 729         */
 730
 731        /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
 732         * guarantees that rto is higher.
 733         */
 734        tcp_bound_rto(sk);
 735}
 736
 737__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
 738{
 739        __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 740
 741        if (!cwnd)
 742                cwnd = TCP_INIT_CWND;
 743        return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 744}
 745
 746/*
 747 * Packet counting of FACK is based on in-order assumptions, therefore TCP
 748 * disables it when reordering is detected
 749 */
 750void tcp_disable_fack(struct tcp_sock *tp)
 751{
 752        /* RFC3517 uses different metric in lost marker => reset on change */
 753        if (tcp_is_fack(tp))
 754                tp->lost_skb_hint = NULL;
 755        tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
 756}
 757
 758/* Take a notice that peer is sending D-SACKs */
 759static void tcp_dsack_seen(struct tcp_sock *tp)
 760{
 761        tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
 762}
 763
 764static void tcp_update_reordering(struct sock *sk, const int metric,
 765                                  const int ts)
 766{
 767        struct tcp_sock *tp = tcp_sk(sk);
 768        if (metric > tp->reordering) {
 769                int mib_idx;
 770
 771                tp->reordering = min(TCP_MAX_REORDERING, metric);
 772
 773                /* This exciting event is worth to be remembered. 8) */
 774                if (ts)
 775                        mib_idx = LINUX_MIB_TCPTSREORDER;
 776                else if (tcp_is_reno(tp))
 777                        mib_idx = LINUX_MIB_TCPRENOREORDER;
 778                else if (tcp_is_fack(tp))
 779                        mib_idx = LINUX_MIB_TCPFACKREORDER;
 780                else
 781                        mib_idx = LINUX_MIB_TCPSACKREORDER;
 782
 783                NET_INC_STATS_BH(sock_net(sk), mib_idx);
 784#if FASTRETRANS_DEBUG > 1
 785                pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
 786                         tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
 787                         tp->reordering,
 788                         tp->fackets_out,
 789                         tp->sacked_out,
 790                         tp->undo_marker ? tp->undo_retrans : 0);
 791#endif
 792                tcp_disable_fack(tp);
 793        }
 794
 795        if (metric > 0)
 796                tcp_disable_early_retrans(tp);
 797}
 798
 799/* This must be called before lost_out is incremented */
 800static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
 801{
 802        if ((tp->retransmit_skb_hint == NULL) ||
 803            before(TCP_SKB_CB(skb)->seq,
 804                   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
 805                tp->retransmit_skb_hint = skb;
 806
 807        if (!tp->lost_out ||
 808            after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
 809                tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
 810}
 811
 812static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
 813{
 814        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 815                tcp_verify_retransmit_hint(tp, skb);
 816
 817                tp->lost_out += tcp_skb_pcount(skb);
 818                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 819        }
 820}
 821
 822static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
 823                                            struct sk_buff *skb)
 824{
 825        tcp_verify_retransmit_hint(tp, skb);
 826
 827        if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 828                tp->lost_out += tcp_skb_pcount(skb);
 829                TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 830        }
 831}
 832
 833/* This procedure tags the retransmission queue when SACKs arrive.
 834 *
 835 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
 836 * Packets in queue with these bits set are counted in variables
 837 * sacked_out, retrans_out and lost_out, correspondingly.
 838 *
 839 * Valid combinations are:
 840 * Tag  InFlight        Description
 841 * 0    1               - orig segment is in flight.
 842 * S    0               - nothing flies, orig reached receiver.
 843 * L    0               - nothing flies, orig lost by net.
 844 * R    2               - both orig and retransmit are in flight.
 845 * L|R  1               - orig is lost, retransmit is in flight.
 846 * S|R  1               - orig reached receiver, retrans is still in flight.
 847 * (L|S|R is logically valid, it could occur when L|R is sacked,
 848 *  but it is equivalent to plain S and code short-curcuits it to S.
 849 *  L|S is logically invalid, it would mean -1 packet in flight 8))
 850 *
 851 * These 6 states form finite state machine, controlled by the following events:
 852 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
 853 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
 854 * 3. Loss detection event of two flavors:
 855 *      A. Scoreboard estimator decided the packet is lost.
 856 *         A'. Reno "three dupacks" marks head of queue lost.
 857 *         A''. Its FACK modification, head until snd.fack is lost.
 858 *      B. SACK arrives sacking SND.NXT at the moment, when the
 859 *         segment was retransmitted.
 860 * 4. D-SACK added new rule: D-SACK changes any tag to S.
 861 *
 862 * It is pleasant to note, that state diagram turns out to be commutative,
 863 * so that we are allowed not to be bothered by order of our actions,
 864 * when multiple events arrive simultaneously. (see the function below).
 865 *
 866 * Reordering detection.
 867 * --------------------
 868 * Reordering metric is maximal distance, which a packet can be displaced
 869 * in packet stream. With SACKs we can estimate it:
 870 *
 871 * 1. SACK fills old hole and the corresponding segment was not
 872 *    ever retransmitted -> reordering. Alas, we cannot use it
 873 *    when segment was retransmitted.
 874 * 2. The last flaw is solved with D-SACK. D-SACK arrives
 875 *    for retransmitted and already SACKed segment -> reordering..
 876 * Both of these heuristics are not used in Loss state, when we cannot
 877 * account for retransmits accurately.
 878 *
 879 * SACK block validation.
 880 * ----------------------
 881 *
 882 * SACK block range validation checks that the received SACK block fits to
 883 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
 884 * Note that SND.UNA is not included to the range though being valid because
 885 * it means that the receiver is rather inconsistent with itself reporting
 886 * SACK reneging when it should advance SND.UNA. Such SACK block this is
 887 * perfectly valid, however, in light of RFC2018 which explicitly states
 888 * that "SACK block MUST reflect the newest segment.  Even if the newest
 889 * segment is going to be discarded ...", not that it looks very clever
 890 * in case of head skb. Due to potentional receiver driven attacks, we
 891 * choose to avoid immediate execution of a walk in write queue due to
 892 * reneging and defer head skb's loss recovery to standard loss recovery
 893 * procedure that will eventually trigger (nothing forbids us doing this).
 894 *
 895 * Implements also blockage to start_seq wrap-around. Problem lies in the
 896 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
 897 * there's no guarantee that it will be before snd_nxt (n). The problem
 898 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
 899 * wrap (s_w):
 900 *
 901 *         <- outs wnd ->                          <- wrapzone ->
 902 *         u     e      n                         u_w   e_w  s n_w
 903 *         |     |      |                          |     |   |  |
 904 * |<------------+------+----- TCP seqno space --------------+---------->|
 905 * ...-- <2^31 ->|                                           |<--------...
 906 * ...---- >2^31 ------>|                                    |<--------...
 907 *
 908 * Current code wouldn't be vulnerable but it's better still to discard such
 909 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
 910 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
 911 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
 912 * equal to the ideal case (infinite seqno space without wrap caused issues).
 913 *
 914 * With D-SACK the lower bound is extended to cover sequence space below
 915 * SND.UNA down to undo_marker, which is the last point of interest. Yet
 916 * again, D-SACK block must not to go across snd_una (for the same reason as
 917 * for the normal SACK blocks, explained above). But there all simplicity
 918 * ends, TCP might receive valid D-SACKs below that. As long as they reside
 919 * fully below undo_marker they do not affect behavior in anyway and can
 920 * therefore be safely ignored. In rare cases (which are more or less
 921 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
 922 * fragmentation and packet reordering past skb's retransmission. To consider
 923 * them correctly, the acceptable range must be extended even more though
 924 * the exact amount is rather hard to quantify. However, tp->max_window can
 925 * be used as an exaggerated estimate.
 926 */
 927static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
 928                                   u32 start_seq, u32 end_seq)
 929{
 930        /* Too far in future, or reversed (interpretation is ambiguous) */
 931        if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
 932                return false;
 933
 934        /* Nasty start_seq wrap-around check (see comments above) */
 935        if (!before(start_seq, tp->snd_nxt))
 936                return false;
 937
 938        /* In outstanding window? ...This is valid exit for D-SACKs too.
 939         * start_seq == snd_una is non-sensical (see comments above)
 940         */
 941        if (after(start_seq, tp->snd_una))
 942                return true;
 943
 944        if (!is_dsack || !tp->undo_marker)
 945                return false;
 946
 947        /* ...Then it's D-SACK, and must reside below snd_una completely */
 948        if (after(end_seq, tp->snd_una))
 949                return false;
 950
 951        if (!before(start_seq, tp->undo_marker))
 952                return true;
 953
 954        /* Too old */
 955        if (!after(end_seq, tp->undo_marker))
 956                return false;
 957
 958        /* Undo_marker boundary crossing (overestimates a lot). Known already:
 959         *   start_seq < undo_marker and end_seq >= undo_marker.
 960         */
 961        return !before(start_seq, end_seq - tp->max_window);
 962}
 963
 964/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
 965 * Event "B". Later note: FACK people cheated me again 8), we have to account
 966 * for reordering! Ugly, but should help.
 967 *
 968 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
 969 * less than what is now known to be received by the other end (derived from
 970 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
 971 * retransmitted skbs to avoid some costly processing per ACKs.
 972 */
 973static void tcp_mark_lost_retrans(struct sock *sk)
 974{
 975        const struct inet_connection_sock *icsk = inet_csk(sk);
 976        struct tcp_sock *tp = tcp_sk(sk);
 977        struct sk_buff *skb;
 978        int cnt = 0;
 979        u32 new_low_seq = tp->snd_nxt;
 980        u32 received_upto = tcp_highest_sack_seq(tp);
 981
 982        if (!tcp_is_fack(tp) || !tp->retrans_out ||
 983            !after(received_upto, tp->lost_retrans_low) ||
 984            icsk->icsk_ca_state != TCP_CA_Recovery)
 985                return;
 986
 987        tcp_for_write_queue(skb, sk) {
 988                u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
 989
 990                if (skb == tcp_send_head(sk))
 991                        break;
 992                if (cnt == tp->retrans_out)
 993                        break;
 994                if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
 995                        continue;
 996
 997                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
 998                        continue;
 999
1000                /* TODO: We would like to get rid of tcp_is_fack(tp) only
1001                 * constraint here (see above) but figuring out that at
1002                 * least tp->reordering SACK blocks reside between ack_seq
1003                 * and received_upto is not easy task to do cheaply with
1004                 * the available datastructures.
1005                 *
1006                 * Whether FACK should check here for tp->reordering segs
1007                 * in-between one could argue for either way (it would be
1008                 * rather simple to implement as we could count fack_count
1009                 * during the walk and do tp->fackets_out - fack_count).
1010                 */
1011                if (after(received_upto, ack_seq)) {
1012                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1013                        tp->retrans_out -= tcp_skb_pcount(skb);
1014
1015                        tcp_skb_mark_lost_uncond_verify(tp, skb);
1016                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1017                } else {
1018                        if (before(ack_seq, new_low_seq))
1019                                new_low_seq = ack_seq;
1020                        cnt += tcp_skb_pcount(skb);
1021                }
1022        }
1023
1024        if (tp->retrans_out)
1025                tp->lost_retrans_low = new_low_seq;
1026}
1027
1028static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1029                            struct tcp_sack_block_wire *sp, int num_sacks,
1030                            u32 prior_snd_una)
1031{
1032        struct tcp_sock *tp = tcp_sk(sk);
1033        u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1034        u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1035        bool dup_sack = false;
1036
1037        if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1038                dup_sack = true;
1039                tcp_dsack_seen(tp);
1040                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1041        } else if (num_sacks > 1) {
1042                u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1043                u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1044
1045                if (!after(end_seq_0, end_seq_1) &&
1046                    !before(start_seq_0, start_seq_1)) {
1047                        dup_sack = true;
1048                        tcp_dsack_seen(tp);
1049                        NET_INC_STATS_BH(sock_net(sk),
1050                                        LINUX_MIB_TCPDSACKOFORECV);
1051                }
1052        }
1053
1054        /* D-SACK for already forgotten data... Do dumb counting. */
1055        if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1056            !after(end_seq_0, prior_snd_una) &&
1057            after(end_seq_0, tp->undo_marker))
1058                tp->undo_retrans--;
1059
1060        return dup_sack;
1061}
1062
1063struct tcp_sacktag_state {
1064        int reord;
1065        int fack_count;
1066        int flag;
1067};
1068
1069/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1070 * the incoming SACK may not exactly match but we can find smaller MSS
1071 * aligned portion of it that matches. Therefore we might need to fragment
1072 * which may fail and creates some hassle (caller must handle error case
1073 * returns).
1074 *
1075 * FIXME: this could be merged to shift decision code
1076 */
1077static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1078                                  u32 start_seq, u32 end_seq)
1079{
1080        int err;
1081        bool in_sack;
1082        unsigned int pkt_len;
1083        unsigned int mss;
1084
1085        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1086                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1087
1088        if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1089            after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1090                mss = tcp_skb_mss(skb);
1091                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1092
1093                if (!in_sack) {
1094                        pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1095                        if (pkt_len < mss)
1096                                pkt_len = mss;
1097                } else {
1098                        pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1099                        if (pkt_len < mss)
1100                                return -EINVAL;
1101                }
1102
1103                /* Round if necessary so that SACKs cover only full MSSes
1104                 * and/or the remaining small portion (if present)
1105                 */
1106                if (pkt_len > mss) {
1107                        unsigned int new_len = (pkt_len / mss) * mss;
1108                        if (!in_sack && new_len < pkt_len) {
1109                                new_len += mss;
1110                                if (new_len > skb->len)
1111                                        return 0;
1112                        }
1113                        pkt_len = new_len;
1114                }
1115                err = tcp_fragment(sk, skb, pkt_len, mss);
1116                if (err < 0)
1117                        return err;
1118        }
1119
1120        return in_sack;
1121}
1122
1123/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1124static u8 tcp_sacktag_one(struct sock *sk,
1125                          struct tcp_sacktag_state *state, u8 sacked,
1126                          u32 start_seq, u32 end_seq,
1127                          bool dup_sack, int pcount)
1128{
1129        struct tcp_sock *tp = tcp_sk(sk);
1130        int fack_count = state->fack_count;
1131
1132        /* Account D-SACK for retransmitted packet. */
1133        if (dup_sack && (sacked & TCPCB_RETRANS)) {
1134                if (tp->undo_marker && tp->undo_retrans &&
1135                    after(end_seq, tp->undo_marker))
1136                        tp->undo_retrans--;
1137                if (sacked & TCPCB_SACKED_ACKED)
1138                        state->reord = min(fack_count, state->reord);
1139        }
1140
1141        /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1142        if (!after(end_seq, tp->snd_una))
1143                return sacked;
1144
1145        if (!(sacked & TCPCB_SACKED_ACKED)) {
1146                if (sacked & TCPCB_SACKED_RETRANS) {
1147                        /* If the segment is not tagged as lost,
1148                         * we do not clear RETRANS, believing
1149                         * that retransmission is still in flight.
1150                         */
1151                        if (sacked & TCPCB_LOST) {
1152                                sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1153                                tp->lost_out -= pcount;
1154                                tp->retrans_out -= pcount;
1155                        }
1156                } else {
1157                        if (!(sacked & TCPCB_RETRANS)) {
1158                                /* New sack for not retransmitted frame,
1159                                 * which was in hole. It is reordering.
1160                                 */
1161                                if (before(start_seq,
1162                                           tcp_highest_sack_seq(tp)))
1163                                        state->reord = min(fack_count,
1164                                                           state->reord);
1165
1166                                /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1167                                if (!after(end_seq, tp->frto_highmark))
1168                                        state->flag |= FLAG_ONLY_ORIG_SACKED;
1169                        }
1170
1171                        if (sacked & TCPCB_LOST) {
1172                                sacked &= ~TCPCB_LOST;
1173                                tp->lost_out -= pcount;
1174                        }
1175                }
1176
1177                sacked |= TCPCB_SACKED_ACKED;
1178                state->flag |= FLAG_DATA_SACKED;
1179                tp->sacked_out += pcount;
1180
1181                fack_count += pcount;
1182
1183                /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1184                if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1185                    before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1186                        tp->lost_cnt_hint += pcount;
1187
1188                if (fack_count > tp->fackets_out)
1189                        tp->fackets_out = fack_count;
1190        }
1191
1192        /* D-SACK. We can detect redundant retransmission in S|R and plain R
1193         * frames and clear it. undo_retrans is decreased above, L|R frames
1194         * are accounted above as well.
1195         */
1196        if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1197                sacked &= ~TCPCB_SACKED_RETRANS;
1198                tp->retrans_out -= pcount;
1199        }
1200
1201        return sacked;
1202}
1203
1204/* Shift newly-SACKed bytes from this skb to the immediately previous
1205 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1206 */
1207static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1208                            struct tcp_sacktag_state *state,
1209                            unsigned int pcount, int shifted, int mss,
1210                            bool dup_sack)
1211{
1212        struct tcp_sock *tp = tcp_sk(sk);
1213        struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1214        u32 start_seq = TCP_SKB_CB(skb)->seq;   /* start of newly-SACKed */
1215        u32 end_seq = start_seq + shifted;      /* end of newly-SACKed */
1216
1217        BUG_ON(!pcount);
1218
1219        /* Adjust counters and hints for the newly sacked sequence
1220         * range but discard the return value since prev is already
1221         * marked. We must tag the range first because the seq
1222         * advancement below implicitly advances
1223         * tcp_highest_sack_seq() when skb is highest_sack.
1224         */
1225        tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1226                        start_seq, end_seq, dup_sack, pcount);
1227
1228        if (skb == tp->lost_skb_hint)
1229                tp->lost_cnt_hint += pcount;
1230
1231        TCP_SKB_CB(prev)->end_seq += shifted;
1232        TCP_SKB_CB(skb)->seq += shifted;
1233
1234        skb_shinfo(prev)->gso_segs += pcount;
1235        BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1236        skb_shinfo(skb)->gso_segs -= pcount;
1237
1238        /* When we're adding to gso_segs == 1, gso_size will be zero,
1239         * in theory this shouldn't be necessary but as long as DSACK
1240         * code can come after this skb later on it's better to keep
1241         * setting gso_size to something.
1242         */
1243        if (!skb_shinfo(prev)->gso_size) {
1244                skb_shinfo(prev)->gso_size = mss;
1245                skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1246        }
1247
1248        /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1249        if (skb_shinfo(skb)->gso_segs <= 1) {
1250                skb_shinfo(skb)->gso_size = 0;
1251                skb_shinfo(skb)->gso_type = 0;
1252        }
1253
1254        /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1255        TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1256
1257        if (skb->len > 0) {
1258                BUG_ON(!tcp_skb_pcount(skb));
1259                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1260                return false;
1261        }
1262
1263        /* Whole SKB was eaten :-) */
1264
1265        if (skb == tp->retransmit_skb_hint)
1266                tp->retransmit_skb_hint = prev;
1267        if (skb == tp->scoreboard_skb_hint)
1268                tp->scoreboard_skb_hint = prev;
1269        if (skb == tp->lost_skb_hint) {
1270                tp->lost_skb_hint = prev;
1271                tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1272        }
1273
1274        TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1275        if (skb == tcp_highest_sack(sk))
1276                tcp_advance_highest_sack(sk, skb);
1277
1278        tcp_unlink_write_queue(skb, sk);
1279        sk_wmem_free_skb(sk, skb);
1280
1281        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1282
1283        return true;
1284}
1285
1286/* I wish gso_size would have a bit more sane initialization than
1287 * something-or-zero which complicates things
1288 */
1289static int tcp_skb_seglen(const struct sk_buff *skb)
1290{
1291        return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1292}
1293
1294/* Shifting pages past head area doesn't work */
1295static int skb_can_shift(const struct sk_buff *skb)
1296{
1297        return !skb_headlen(skb) && skb_is_nonlinear(skb);
1298}
1299
1300/* Try collapsing SACK blocks spanning across multiple skbs to a single
1301 * skb.
1302 */
1303static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1304                                          struct tcp_sacktag_state *state,
1305                                          u32 start_seq, u32 end_seq,
1306                                          bool dup_sack)
1307{
1308        struct tcp_sock *tp = tcp_sk(sk);
1309        struct sk_buff *prev;
1310        int mss;
1311        int pcount = 0;
1312        int len;
1313        int in_sack;
1314
1315        if (!sk_can_gso(sk))
1316                goto fallback;
1317
1318        /* Normally R but no L won't result in plain S */
1319        if (!dup_sack &&
1320            (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1321                goto fallback;
1322        if (!skb_can_shift(skb))
1323                goto fallback;
1324        /* This frame is about to be dropped (was ACKed). */
1325        if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1326                goto fallback;
1327
1328        /* Can only happen with delayed DSACK + discard craziness */
1329        if (unlikely(skb == tcp_write_queue_head(sk)))
1330                goto fallback;
1331        prev = tcp_write_queue_prev(sk, skb);
1332
1333        if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1334                goto fallback;
1335
1336        in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1337                  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1338
1339        if (in_sack) {
1340                len = skb->len;
1341                pcount = tcp_skb_pcount(skb);
1342                mss = tcp_skb_seglen(skb);
1343
1344                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1345                 * drop this restriction as unnecessary
1346                 */
1347                if (mss != tcp_skb_seglen(prev))
1348                        goto fallback;
1349        } else {
1350                if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1351                        goto noop;
1352                /* CHECKME: This is non-MSS split case only?, this will
1353                 * cause skipped skbs due to advancing loop btw, original
1354                 * has that feature too
1355                 */
1356                if (tcp_skb_pcount(skb) <= 1)
1357                        goto noop;
1358
1359                in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1360                if (!in_sack) {
1361                        /* TODO: head merge to next could be attempted here
1362                         * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1363                         * though it might not be worth of the additional hassle
1364                         *
1365                         * ...we can probably just fallback to what was done
1366                         * previously. We could try merging non-SACKed ones
1367                         * as well but it probably isn't going to buy off
1368                         * because later SACKs might again split them, and
1369                         * it would make skb timestamp tracking considerably
1370                         * harder problem.
1371                         */
1372                        goto fallback;
1373                }
1374
1375                len = end_seq - TCP_SKB_CB(skb)->seq;
1376                BUG_ON(len < 0);
1377                BUG_ON(len > skb->len);
1378
1379                /* MSS boundaries should be honoured or else pcount will
1380                 * severely break even though it makes things bit trickier.
1381                 * Optimize common case to avoid most of the divides
1382                 */
1383                mss = tcp_skb_mss(skb);
1384
1385                /* TODO: Fix DSACKs to not fragment already SACKed and we can
1386                 * drop this restriction as unnecessary
1387                 */
1388                if (mss != tcp_skb_seglen(prev))
1389                        goto fallback;
1390
1391                if (len == mss) {
1392                        pcount = 1;
1393                } else if (len < mss) {
1394                        goto noop;
1395                } else {
1396                        pcount = len / mss;
1397                        len = pcount * mss;
1398                }
1399        }
1400
1401        /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1402        if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1403                goto fallback;
1404
1405        if (!skb_shift(prev, skb, len))
1406                goto fallback;
1407        if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1408                goto out;
1409
1410        /* Hole filled allows collapsing with the next as well, this is very
1411         * useful when hole on every nth skb pattern happens
1412         */
1413        if (prev == tcp_write_queue_tail(sk))
1414                goto out;
1415        skb = tcp_write_queue_next(sk, prev);
1416
1417        if (!skb_can_shift(skb) ||
1418            (skb == tcp_send_head(sk)) ||
1419            ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1420            (mss != tcp_skb_seglen(skb)))
1421                goto out;
1422
1423        len = skb->len;
1424        if (skb_shift(prev, skb, len)) {
1425                pcount += tcp_skb_pcount(skb);
1426                tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1427        }
1428
1429out:
1430        state->fack_count += pcount;
1431        return prev;
1432
1433noop:
1434        return skb;
1435
1436fallback:
1437        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1438        return NULL;
1439}
1440
1441static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1442                                        struct tcp_sack_block *next_dup,
1443                                        struct tcp_sacktag_state *state,
1444                                        u32 start_seq, u32 end_seq,
1445                                        bool dup_sack_in)
1446{
1447        struct tcp_sock *tp = tcp_sk(sk);
1448        struct sk_buff *tmp;
1449
1450        tcp_for_write_queue_from(skb, sk) {
1451                int in_sack = 0;
1452                bool dup_sack = dup_sack_in;
1453
1454                if (skb == tcp_send_head(sk))
1455                        break;
1456
1457                /* queue is in-order => we can short-circuit the walk early */
1458                if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1459                        break;
1460
1461                if ((next_dup != NULL) &&
1462                    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1463                        in_sack = tcp_match_skb_to_sack(sk, skb,
1464                                                        next_dup->start_seq,
1465                                                        next_dup->end_seq);
1466                        if (in_sack > 0)
1467                                dup_sack = true;
1468                }
1469
1470                /* skb reference here is a bit tricky to get right, since
1471                 * shifting can eat and free both this skb and the next,
1472                 * so not even _safe variant of the loop is enough.
1473                 */
1474                if (in_sack <= 0) {
1475                        tmp = tcp_shift_skb_data(sk, skb, state,
1476                                                 start_seq, end_seq, dup_sack);
1477                        if (tmp != NULL) {
1478                                if (tmp != skb) {
1479                                        skb = tmp;
1480                                        continue;
1481                                }
1482
1483                                in_sack = 0;
1484                        } else {
1485                                in_sack = tcp_match_skb_to_sack(sk, skb,
1486                                                                start_seq,
1487                                                                end_seq);
1488                        }
1489                }
1490
1491                if (unlikely(in_sack < 0))
1492                        break;
1493
1494                if (in_sack) {
1495                        TCP_SKB_CB(skb)->sacked =
1496                                tcp_sacktag_one(sk,
1497                                                state,
1498                                                TCP_SKB_CB(skb)->sacked,
1499                                                TCP_SKB_CB(skb)->seq,
1500                                                TCP_SKB_CB(skb)->end_seq,
1501                                                dup_sack,
1502                                                tcp_skb_pcount(skb));
1503
1504                        if (!before(TCP_SKB_CB(skb)->seq,
1505                                    tcp_highest_sack_seq(tp)))
1506                                tcp_advance_highest_sack(sk, skb);
1507                }
1508
1509                state->fack_count += tcp_skb_pcount(skb);
1510        }
1511        return skb;
1512}
1513
1514/* Avoid all extra work that is being done by sacktag while walking in
1515 * a normal way
1516 */
1517static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1518                                        struct tcp_sacktag_state *state,
1519                                        u32 skip_to_seq)
1520{
1521        tcp_for_write_queue_from(skb, sk) {
1522                if (skb == tcp_send_head(sk))
1523                        break;
1524
1525                if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1526                        break;
1527
1528                state->fack_count += tcp_skb_pcount(skb);
1529        }
1530        return skb;
1531}
1532
1533static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1534                                                struct sock *sk,
1535                                                struct tcp_sack_block *next_dup,
1536                                                struct tcp_sacktag_state *state,
1537                                                u32 skip_to_seq)
1538{
1539        if (next_dup == NULL)
1540                return skb;
1541
1542        if (before(next_dup->start_seq, skip_to_seq)) {
1543                skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1544                skb = tcp_sacktag_walk(skb, sk, NULL, state,
1545                                       next_dup->start_seq, next_dup->end_seq,
1546                                       1);
1547        }
1548
1549        return skb;
1550}
1551
1552static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1553{
1554        return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1555}
1556
1557static int
1558tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1559                        u32 prior_snd_una)
1560{
1561        const struct inet_connection_sock *icsk = inet_csk(sk);
1562        struct tcp_sock *tp = tcp_sk(sk);
1563        const unsigned char *ptr = (skb_transport_header(ack_skb) +
1564                                    TCP_SKB_CB(ack_skb)->sacked);
1565        struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1566        struct tcp_sack_block sp[TCP_NUM_SACKS];
1567        struct tcp_sack_block *cache;
1568        struct tcp_sacktag_state state;
1569        struct sk_buff *skb;
1570        int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1571        int used_sacks;
1572        bool found_dup_sack = false;
1573        int i, j;
1574        int first_sack_index;
1575
1576        state.flag = 0;
1577        state.reord = tp->packets_out;
1578
1579        if (!tp->sacked_out) {
1580                if (WARN_ON(tp->fackets_out))
1581                        tp->fackets_out = 0;
1582                tcp_highest_sack_reset(sk);
1583        }
1584
1585        found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1586                                         num_sacks, prior_snd_una);
1587        if (found_dup_sack)
1588                state.flag |= FLAG_DSACKING_ACK;
1589
1590        /* Eliminate too old ACKs, but take into
1591         * account more or less fresh ones, they can
1592         * contain valid SACK info.
1593         */
1594        if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1595                return 0;
1596
1597        if (!tp->packets_out)
1598                goto out;
1599
1600        used_sacks = 0;
1601        first_sack_index = 0;
1602        for (i = 0; i < num_sacks; i++) {
1603                bool dup_sack = !i && found_dup_sack;
1604
1605                sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1606                sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1607
1608                if (!tcp_is_sackblock_valid(tp, dup_sack,
1609                                            sp[used_sacks].start_seq,
1610                                            sp[used_sacks].end_seq)) {
1611                        int mib_idx;
1612
1613                        if (dup_sack) {
1614                                if (!tp->undo_marker)
1615                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1616                                else
1617                                        mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1618                        } else {
1619                                /* Don't count olds caused by ACK reordering */
1620                                if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1621                                    !after(sp[used_sacks].end_seq, tp->snd_una))
1622                                        continue;
1623                                mib_idx = LINUX_MIB_TCPSACKDISCARD;
1624                        }
1625
1626                        NET_INC_STATS_BH(sock_net(sk), mib_idx);
1627                        if (i == 0)
1628                                first_sack_index = -1;
1629                        continue;
1630                }
1631
1632                /* Ignore very old stuff early */
1633                if (!after(sp[used_sacks].end_seq, prior_snd_una))
1634                        continue;
1635
1636                used_sacks++;
1637        }
1638
1639        /* order SACK blocks to allow in order walk of the retrans queue */
1640        for (i = used_sacks - 1; i > 0; i--) {
1641                for (j = 0; j < i; j++) {
1642                        if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1643                                swap(sp[j], sp[j + 1]);
1644
1645                                /* Track where the first SACK block goes to */
1646                                if (j == first_sack_index)
1647                                        first_sack_index = j + 1;
1648                        }
1649                }
1650        }
1651
1652        skb = tcp_write_queue_head(sk);
1653        state.fack_count = 0;
1654        i = 0;
1655
1656        if (!tp->sacked_out) {
1657                /* It's already past, so skip checking against it */
1658                cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1659        } else {
1660                cache = tp->recv_sack_cache;
1661                /* Skip empty blocks in at head of the cache */
1662                while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1663                       !cache->end_seq)
1664                        cache++;
1665        }
1666
1667        while (i < used_sacks) {
1668                u32 start_seq = sp[i].start_seq;
1669                u32 end_seq = sp[i].end_seq;
1670                bool dup_sack = (found_dup_sack && (i == first_sack_index));
1671                struct tcp_sack_block *next_dup = NULL;
1672
1673                if (found_dup_sack && ((i + 1) == first_sack_index))
1674                        next_dup = &sp[i + 1];
1675
1676                /* Skip too early cached blocks */
1677                while (tcp_sack_cache_ok(tp, cache) &&
1678                       !before(start_seq, cache->end_seq))
1679                        cache++;
1680
1681                /* Can skip some work by looking recv_sack_cache? */
1682                if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1683                    after(end_seq, cache->start_seq)) {
1684
1685                        /* Head todo? */
1686                        if (before(start_seq, cache->start_seq)) {
1687                                skb = tcp_sacktag_skip(skb, sk, &state,
1688                                                       start_seq);
1689                                skb = tcp_sacktag_walk(skb, sk, next_dup,
1690                                                       &state,
1691                                                       start_seq,
1692                                                       cache->start_seq,
1693                                                       dup_sack);
1694                        }
1695
1696                        /* Rest of the block already fully processed? */
1697                        if (!after(end_seq, cache->end_seq))
1698                                goto advance_sp;
1699
1700                        skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1701                                                       &state,
1702                                                       cache->end_seq);
1703
1704                        /* ...tail remains todo... */
1705                        if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1706                                /* ...but better entrypoint exists! */
1707                                skb = tcp_highest_sack(sk);
1708                                if (skb == NULL)
1709                                        break;
1710                                state.fack_count = tp->fackets_out;
1711                                cache++;
1712                                goto walk;
1713                        }
1714
1715                        skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1716                        /* Check overlap against next cached too (past this one already) */
1717                        cache++;
1718                        continue;
1719                }
1720
1721                if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1722                        skb = tcp_highest_sack(sk);
1723                        if (skb == NULL)
1724                                break;
1725                        state.fack_count = tp->fackets_out;
1726                }
1727                skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1728
1729walk:
1730                skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1731                                       start_seq, end_seq, dup_sack);
1732
1733advance_sp:
1734                /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1735                 * due to in-order walk
1736                 */
1737                if (after(end_seq, tp->frto_highmark))
1738                        state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1739
1740                i++;
1741        }
1742
1743        /* Clear the head of the cache sack blocks so we can skip it next time */
1744        for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1745                tp->recv_sack_cache[i].start_seq = 0;
1746                tp->recv_sack_cache[i].end_seq = 0;
1747        }
1748        for (j = 0; j < used_sacks; j++)
1749                tp->recv_sack_cache[i++] = sp[j];
1750
1751        tcp_mark_lost_retrans(sk);
1752
1753        tcp_verify_left_out(tp);
1754
1755        if ((state.reord < tp->fackets_out) &&
1756            ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1757            (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1758                tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1759
1760out:
1761
1762#if FASTRETRANS_DEBUG > 0
1763        WARN_ON((int)tp->sacked_out < 0);
1764        WARN_ON((int)tp->lost_out < 0);
1765        WARN_ON((int)tp->retrans_out < 0);
1766        WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1767#endif
1768        return state.flag;
1769}
1770
1771/* Limits sacked_out so that sum with lost_out isn't ever larger than
1772 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1773 */
1774static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1775{
1776        u32 holes;
1777
1778        holes = max(tp->lost_out, 1U);
1779        holes = min(holes, tp->packets_out);
1780
1781        if ((tp->sacked_out + holes) > tp->packets_out) {
1782                tp->sacked_out = tp->packets_out - holes;
1783                return true;
1784        }
1785        return false;
1786}
1787
1788/* If we receive more dupacks than we expected counting segments
1789 * in assumption of absent reordering, interpret this as reordering.
1790 * The only another reason could be bug in receiver TCP.
1791 */
1792static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1793{
1794        struct tcp_sock *tp = tcp_sk(sk);
1795        if (tcp_limit_reno_sacked(tp))
1796                tcp_update_reordering(sk, tp->packets_out + addend, 0);
1797}
1798
1799/* Emulate SACKs for SACKless connection: account for a new dupack. */
1800
1801static void tcp_add_reno_sack(struct sock *sk)
1802{
1803        struct tcp_sock *tp = tcp_sk(sk);
1804        tp->sacked_out++;
1805        tcp_check_reno_reordering(sk, 0);
1806        tcp_verify_left_out(tp);
1807}
1808
1809/* Account for ACK, ACKing some data in Reno Recovery phase. */
1810
1811static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1812{
1813        struct tcp_sock *tp = tcp_sk(sk);
1814
1815        if (acked > 0) {
1816                /* One ACK acked hole. The rest eat duplicate ACKs. */
1817                if (acked - 1 >= tp->sacked_out)
1818                        tp->sacked_out = 0;
1819                else
1820                        tp->sacked_out -= acked - 1;
1821        }
1822        tcp_check_reno_reordering(sk, acked);
1823        tcp_verify_left_out(tp);
1824}
1825
1826static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1827{
1828        tp->sacked_out = 0;
1829}
1830
1831static int tcp_is_sackfrto(const struct tcp_sock *tp)
1832{
1833        return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1834}
1835
1836/* F-RTO can only be used if TCP has never retransmitted anything other than
1837 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1838 */
1839bool tcp_use_frto(struct sock *sk)
1840{
1841        const struct tcp_sock *tp = tcp_sk(sk);
1842        const struct inet_connection_sock *icsk = inet_csk(sk);
1843        struct sk_buff *skb;
1844
1845        if (!sysctl_tcp_frto)
1846                return false;
1847
1848        /* MTU probe and F-RTO won't really play nicely along currently */
1849        if (icsk->icsk_mtup.probe_size)
1850                return false;
1851
1852        if (tcp_is_sackfrto(tp))
1853                return true;
1854
1855        /* Avoid expensive walking of rexmit queue if possible */
1856        if (tp->retrans_out > 1)
1857                return false;
1858
1859        skb = tcp_write_queue_head(sk);
1860        if (tcp_skb_is_last(sk, skb))
1861                return true;
1862        skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1863        tcp_for_write_queue_from(skb, sk) {
1864                if (skb == tcp_send_head(sk))
1865                        break;
1866                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1867                        return false;
1868                /* Short-circuit when first non-SACKed skb has been checked */
1869                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1870                        break;
1871        }
1872        return true;
1873}
1874
1875/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1876 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1877 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1878 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1879 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1880 * bits are handled if the Loss state is really to be entered (in
1881 * tcp_enter_frto_loss).
1882 *
1883 * Do like tcp_enter_loss() would; when RTO expires the second time it
1884 * does:
1885 *  "Reduce ssthresh if it has not yet been made inside this window."
1886 */
1887void tcp_enter_frto(struct sock *sk)
1888{
1889        const struct inet_connection_sock *icsk = inet_csk(sk);
1890        struct tcp_sock *tp = tcp_sk(sk);
1891        struct sk_buff *skb;
1892
1893        if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1894            tp->snd_una == tp->high_seq ||
1895            ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1896             !icsk->icsk_retransmits)) {
1897                tp->prior_ssthresh = tcp_current_ssthresh(sk);
1898                /* Our state is too optimistic in ssthresh() call because cwnd
1899                 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1900                 * recovery has not yet completed. Pattern would be this: RTO,
1901                 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1902                 * up here twice).
1903                 * RFC4138 should be more specific on what to do, even though
1904                 * RTO is quite unlikely to occur after the first Cumulative ACK
1905                 * due to back-off and complexity of triggering events ...
1906                 */
1907                if (tp->frto_counter) {
1908                        u32 stored_cwnd;
1909                        stored_cwnd = tp->snd_cwnd;
1910                        tp->snd_cwnd = 2;
1911                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1912                        tp->snd_cwnd = stored_cwnd;
1913                } else {
1914                        tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1915                }
1916                /* ... in theory, cong.control module could do "any tricks" in
1917                 * ssthresh(), which means that ca_state, lost bits and lost_out
1918                 * counter would have to be faked before the call occurs. We
1919                 * consider that too expensive, unlikely and hacky, so modules
1920                 * using these in ssthresh() must deal these incompatibility
1921                 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1922                 */
1923                tcp_ca_event(sk, CA_EVENT_FRTO);
1924        }
1925
1926        tp->undo_marker = tp->snd_una;
1927        tp->undo_retrans = 0;
1928
1929        skb = tcp_write_queue_head(sk);
1930        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1931                tp->undo_marker = 0;
1932        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1933                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1934                tp->retrans_out -= tcp_skb_pcount(skb);
1935        }
1936        tcp_verify_left_out(tp);
1937
1938        /* Too bad if TCP was application limited */
1939        tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1940
1941        /* Earlier loss recovery underway (see RFC4138; Appendix B).
1942         * The last condition is necessary at least in tp->frto_counter case.
1943         */
1944        if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1945            ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1946            after(tp->high_seq, tp->snd_una)) {
1947                tp->frto_highmark = tp->high_seq;
1948        } else {
1949                tp->frto_highmark = tp->snd_nxt;
1950        }
1951        tcp_set_ca_state(sk, TCP_CA_Disorder);
1952        tp->high_seq = tp->snd_nxt;
1953        tp->frto_counter = 1;
1954}
1955
1956/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1957 * which indicates that we should follow the traditional RTO recovery,
1958 * i.e. mark everything lost and do go-back-N retransmission.
1959 */
1960static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1961{
1962        struct tcp_sock *tp = tcp_sk(sk);
1963        struct sk_buff *skb;
1964
1965        tp->lost_out = 0;
1966        tp->retrans_out = 0;
1967        if (tcp_is_reno(tp))
1968                tcp_reset_reno_sack(tp);
1969
1970        tcp_for_write_queue(skb, sk) {
1971                if (skb == tcp_send_head(sk))
1972                        break;
1973
1974                TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1975                /*
1976                 * Count the retransmission made on RTO correctly (only when
1977                 * waiting for the first ACK and did not get it)...
1978                 */
1979                if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1980                        /* For some reason this R-bit might get cleared? */
1981                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1982                                tp->retrans_out += tcp_skb_pcount(skb);
1983                        /* ...enter this if branch just for the first segment */
1984                        flag |= FLAG_DATA_ACKED;
1985                } else {
1986                        if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1987                                tp->undo_marker = 0;
1988                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1989                }
1990
1991                /* Marking forward transmissions that were made after RTO lost
1992                 * can cause unnecessary retransmissions in some scenarios,
1993                 * SACK blocks will mitigate that in some but not in all cases.
1994                 * We used to not mark them but it was causing break-ups with
1995                 * receivers that do only in-order receival.
1996                 *
1997                 * TODO: we could detect presence of such receiver and select
1998                 * different behavior per flow.
1999                 */
2000                if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2001                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2002                        tp->lost_out += tcp_skb_pcount(skb);
2003                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2004                }
2005        }
2006        tcp_verify_left_out(tp);
2007
2008        tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2009        tp->snd_cwnd_cnt = 0;
2010        tp->snd_cwnd_stamp = tcp_time_stamp;
2011        tp->frto_counter = 0;
2012        tp->bytes_acked = 0;
2013
2014        tp->reordering = min_t(unsigned int, tp->reordering,
2015                               sysctl_tcp_reordering);
2016        tcp_set_ca_state(sk, TCP_CA_Loss);
2017        tp->high_seq = tp->snd_nxt;
2018        TCP_ECN_queue_cwr(tp);
2019
2020        tcp_clear_all_retrans_hints(tp);
2021}
2022
2023static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2024{
2025        tp->retrans_out = 0;
2026        tp->lost_out = 0;
2027
2028        tp->undo_marker = 0;
2029        tp->undo_retrans = 0;
2030}
2031
2032void tcp_clear_retrans(struct tcp_sock *tp)
2033{
2034        tcp_clear_retrans_partial(tp);
2035
2036        tp->fackets_out = 0;
2037        tp->sacked_out = 0;
2038}
2039
2040/* Enter Loss state. If "how" is not zero, forget all SACK information
2041 * and reset tags completely, otherwise preserve SACKs. If receiver
2042 * dropped its ofo queue, we will know this due to reneging detection.
2043 */
2044void tcp_enter_loss(struct sock *sk, int how)
2045{
2046        const struct inet_connection_sock *icsk = inet_csk(sk);
2047        struct tcp_sock *tp = tcp_sk(sk);
2048        struct sk_buff *skb;
2049
2050        /* Reduce ssthresh if it has not yet been made inside this window. */
2051        if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2052            (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2053                tp->prior_ssthresh = tcp_current_ssthresh(sk);
2054                tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2055                tcp_ca_event(sk, CA_EVENT_LOSS);
2056        }
2057        tp->snd_cwnd       = 1;
2058        tp->snd_cwnd_cnt   = 0;
2059        tp->snd_cwnd_stamp = tcp_time_stamp;
2060
2061        tp->bytes_acked = 0;
2062        tcp_clear_retrans_partial(tp);
2063
2064        if (tcp_is_reno(tp))
2065                tcp_reset_reno_sack(tp);
2066
2067        if (!how) {
2068                /* Push undo marker, if it was plain RTO and nothing
2069                 * was retransmitted. */
2070                tp->undo_marker = tp->snd_una;
2071        } else {
2072                tp->sacked_out = 0;
2073                tp->fackets_out = 0;
2074        }
2075        tcp_clear_all_retrans_hints(tp);
2076
2077        tcp_for_write_queue(skb, sk) {
2078                if (skb == tcp_send_head(sk))
2079                        break;
2080
2081                if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2082                        tp->undo_marker = 0;
2083                TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2084                if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2085                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2086                        TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2087                        tp->lost_out += tcp_skb_pcount(skb);
2088                        tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2089                }
2090        }
2091        tcp_verify_left_out(tp);
2092
2093        tp->reordering = min_t(unsigned int, tp->reordering,
2094                               sysctl_tcp_reordering);
2095        tcp_set_ca_state(sk, TCP_CA_Loss);
2096        tp->high_seq = tp->snd_nxt;
2097        TCP_ECN_queue_cwr(tp);
2098        /* Abort F-RTO algorithm if one is in progress */
2099        tp->frto_counter = 0;
2100}
2101
2102/* If ACK arrived pointing to a remembered SACK, it means that our
2103 * remembered SACKs do not reflect real state of receiver i.e.
2104 * receiver _host_ is heavily congested (or buggy).
2105 *
2106 * Do processing similar to RTO timeout.
2107 */
2108static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2109{
2110        if (flag & FLAG_SACK_RENEGING) {
2111                struct inet_connection_sock *icsk = inet_csk(sk);
2112                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2113
2114                tcp_enter_loss(sk, 1);
2115                icsk->icsk_retransmits++;
2116                tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2117                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2118                                          icsk->icsk_rto, TCP_RTO_MAX);
2119                return true;
2120        }
2121        return false;
2122}
2123
2124static inline int tcp_fackets_out(const struct tcp_sock *tp)
2125{
2126        return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2127}
2128
2129/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2130 * counter when SACK is enabled (without SACK, sacked_out is used for
2131 * that purpose).
2132 *
2133 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2134 * segments up to the highest received SACK block so far and holes in
2135 * between them.
2136 *
2137 * With reordering, holes may still be in flight, so RFC3517 recovery
2138 * uses pure sacked_out (total number of SACKed segments) even though
2139 * it violates the RFC that uses duplicate ACKs, often these are equal
2140 * but when e.g. out-of-window ACKs or packet duplication occurs,
2141 * they differ. Since neither occurs due to loss, TCP should really
2142 * ignore them.
2143 */
2144static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2145{
2146        return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2147}
2148
2149static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2150{
2151        struct tcp_sock *tp = tcp_sk(sk);
2152        unsigned long delay;
2153
2154        /* Delay early retransmit and entering fast recovery for
2155         * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2156         * available, or RTO is scheduled to fire first.
2157         */
2158        if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2159                return false;
2160
2161        delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2162        if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2163                return false;
2164
2165        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2166        tp->early_retrans_delayed = 1;
2167        return true;
2168}
2169
2170static inline int tcp_skb_timedout(const struct sock *sk,
2171                                   const struct sk_buff *skb)
2172{
2173        return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2174}
2175
2176static inline int tcp_head_timedout(const struct sock *sk)
2177{
2178        const struct tcp_sock *tp = tcp_sk(sk);
2179
2180        return tp->packets_out &&
2181               tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2182}
2183
2184/* Linux NewReno/SACK/FACK/ECN state machine.
2185 * --------------------------------------
2186 *
2187 * "Open"       Normal state, no dubious events, fast path.
2188 * "Disorder"   In all the respects it is "Open",
2189 *              but requires a bit more attention. It is entered when
2190 *              we see some SACKs or dupacks. It is split of "Open"
2191 *              mainly to move some processing from fast path to slow one.
2192 * "CWR"        CWND was reduced due to some Congestion Notification event.
2193 *              It can be ECN, ICMP source quench, local device congestion.
2194 * "Recovery"   CWND was reduced, we are fast-retransmitting.
2195 * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2196 *
2197 * tcp_fastretrans_alert() is entered:
2198 * - each incoming ACK, if state is not "Open"
2199 * - when arrived ACK is unusual, namely:
2200 *      * SACK
2201 *      * Duplicate ACK.
2202 *      * ECN ECE.
2203 *
2204 * Counting packets in flight is pretty simple.
2205 *
2206 *      in_flight = packets_out - left_out + retrans_out
2207 *
2208 *      packets_out is SND.NXT-SND.UNA counted in packets.
2209 *
2210 *      retrans_out is number of retransmitted segments.
2211 *
2212 *      left_out is number of segments left network, but not ACKed yet.
2213 *
2214 *              left_out = sacked_out + lost_out
2215 *
2216 *     sacked_out: Packets, which arrived to receiver out of order
2217 *                 and hence not ACKed. With SACKs this number is simply
2218 *                 amount of SACKed data. Even without SACKs
2219 *                 it is easy to give pretty reliable estimate of this number,
2220 *                 counting duplicate ACKs.
2221 *
2222 *       lost_out: Packets lost by network. TCP has no explicit
2223 *                 "loss notification" feedback from network (for now).
2224 *                 It means that this number can be only _guessed_.
2225 *                 Actually, it is the heuristics to predict lossage that
2226 *                 distinguishes different algorithms.
2227 *
2228 *      F.e. after RTO, when all the queue is considered as lost,
2229 *      lost_out = packets_out and in_flight = retrans_out.
2230 *
2231 *              Essentially, we have now two algorithms counting
2232 *              lost packets.
2233 *
2234 *              FACK: It is the simplest heuristics. As soon as we decided
2235 *              that something is lost, we decide that _all_ not SACKed
2236 *              packets until the most forward SACK are lost. I.e.
2237 *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2238 *              It is absolutely correct estimate, if network does not reorder
2239 *              packets. And it loses any connection to reality when reordering
2240 *              takes place. We use FACK by default until reordering
2241 *              is suspected on the path to this destination.
2242 *
2243 *              NewReno: when Recovery is entered, we assume that one segment
2244 *              is lost (classic Reno). While we are in Recovery and
2245 *              a partial ACK arrives, we assume that one more packet
2246 *              is lost (NewReno). This heuristics are the same in NewReno
2247 *              and SACK.
2248 *
2249 *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2250 *  deflation etc. CWND is real congestion window, never inflated, changes
2251 *  only according to classic VJ rules.
2252 *
2253 * Really tricky (and requiring careful tuning) part of algorithm
2254 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2255 * The first determines the moment _when_ we should reduce CWND and,
2256 * hence, slow down forward transmission. In fact, it determines the moment
2257 * when we decide that hole is caused by loss, rather than by a reorder.
2258 *
2259 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2260 * holes, caused by lost packets.
2261 *
2262 * And the most logically complicated part of algorithm is undo
2263 * heuristics. We detect false retransmits due to both too early
2264 * fast retransmit (reordering) and underestimated RTO, analyzing
2265 * timestamps and D-SACKs. When we detect that some segments were
2266 * retransmitted by mistake and CWND reduction was wrong, we undo
2267 * window reduction and abort recovery phase. This logic is hidden
2268 * inside several functions named tcp_try_undo_<something>.
2269 */
2270
2271/* This function decides, when we should leave Disordered state
2272 * and enter Recovery phase, reducing congestion window.
2273 *
2274 * Main question: may we further continue forward transmission
2275 * with the same cwnd?
2276 */
2277static bool tcp_time_to_recover(struct sock *sk, int flag)
2278{
2279        struct tcp_sock *tp = tcp_sk(sk);
2280        __u32 packets_out;
2281
2282        /* Do not perform any recovery during F-RTO algorithm */
2283        if (tp->frto_counter)
2284                return false;
2285
2286        /* Trick#1: The loss is proven. */
2287        if (tp->lost_out)
2288                return true;
2289
2290        /* Not-A-Trick#2 : Classic rule... */
2291        if (tcp_dupack_heuristics(tp) > tp->reordering)
2292                return true;
2293
2294        /* Trick#3 : when we use RFC2988 timer restart, fast
2295         * retransmit can be triggered by timeout of queue head.
2296         */
2297        if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2298                return true;
2299
2300        /* Trick#4: It is still not OK... But will it be useful to delay
2301         * recovery more?
2302         */
2303        packets_out = tp->packets_out;
2304        if (packets_out <= tp->reordering &&
2305            tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2306            !tcp_may_send_now(sk)) {
2307                /* We have nothing to send. This connection is limited
2308                 * either by receiver window or by application.
2309                 */
2310                return true;
2311        }
2312
2313        /* If a thin stream is detected, retransmit after first
2314         * received dupack. Employ only if SACK is supported in order
2315         * to avoid possible corner-case series of spurious retransmissions
2316         * Use only if there are no unsent data.
2317         */
2318        if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2319            tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2320            tcp_is_sack(tp) && !tcp_send_head(sk))
2321                return true;
2322
2323        /* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2324         * retransmissions due to small network reorderings, we implement
2325         * Mitigation A.3 in the RFC and delay the retransmission for a short
2326         * interval if appropriate.
2327         */
2328        if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2329            (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2330            !tcp_may_send_now(sk))
2331                return !tcp_pause_early_retransmit(sk, flag);
2332
2333        return false;
2334}
2335
2336/* New heuristics: it is possible only after we switched to restart timer
2337 * each time when something is ACKed. Hence, we can detect timed out packets
2338 * during fast retransmit without falling to slow start.
2339 *
2340 * Usefulness of this as is very questionable, since we should know which of
2341 * the segments is the next to timeout which is relatively expensive to find
2342 * in general case unless we add some data structure just for that. The
2343 * current approach certainly won't find the right one too often and when it
2344 * finally does find _something_ it usually marks large part of the window
2345 * right away (because a retransmission with a larger timestamp blocks the
2346 * loop from advancing). -ij
2347 */
2348static void tcp_timeout_skbs(struct sock *sk)
2349{
2350        struct tcp_sock *tp = tcp_sk(sk);
2351        struct sk_buff *skb;
2352
2353        if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2354                return;
2355
2356        skb = tp->scoreboard_skb_hint;
2357        if (tp->scoreboard_skb_hint == NULL)
2358                skb = tcp_write_queue_head(sk);
2359
2360        tcp_for_write_queue_from(skb, sk) {
2361                if (skb == tcp_send_head(sk))
2362                        break;
2363                if (!tcp_skb_timedout(sk, skb))
2364                        break;
2365
2366                tcp_skb_mark_lost(tp, skb);
2367        }
2368
2369        tp->scoreboard_skb_hint = skb;
2370
2371        tcp_verify_left_out(tp);
2372}
2373
2374/* Detect loss in event "A" above by marking head of queue up as lost.
2375 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2376 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2377 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2378 * the maximum SACKed segments to pass before reaching this limit.
2379 */
2380static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2381{
2382        struct tcp_sock *tp = tcp_sk(sk);
2383        struct sk_buff *skb;
2384        int cnt, oldcnt;
2385        int err;
2386        unsigned int mss;
2387        /* Use SACK to deduce losses of new sequences sent during recovery */
2388        const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2389
2390        WARN_ON(packets > tp->packets_out);
2391        if (tp->lost_skb_hint) {
2392                skb = tp->lost_skb_hint;
2393                cnt = tp->lost_cnt_hint;
2394                /* Head already handled? */
2395                if (mark_head && skb != tcp_write_queue_head(sk))
2396                        return;
2397        } else {
2398                skb = tcp_write_queue_head(sk);
2399                cnt = 0;
2400        }
2401
2402        tcp_for_write_queue_from(skb, sk) {
2403                if (skb == tcp_send_head(sk))
2404                        break;
2405                /* TODO: do this better */
2406                /* this is not the most efficient way to do this... */
2407                tp->lost_skb_hint = skb;
2408                tp->lost_cnt_hint = cnt;
2409
2410                if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2411                        break;
2412
2413                oldcnt = cnt;
2414                if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2415                    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2416                        cnt += tcp_skb_pcount(skb);
2417
2418                if (cnt > packets) {
2419                        if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2420                            (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2421                            (oldcnt >= packets))
2422                                break;
2423
2424                        mss = skb_shinfo(skb)->gso_size;
2425                        err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2426                        if (err < 0)
2427                                break;
2428                        cnt = packets;
2429                }
2430
2431                tcp_skb_mark_lost(tp, skb);
2432
2433                if (mark_head)
2434                        break;
2435        }
2436        tcp_verify_left_out(tp);
2437}
2438
2439/* Account newly detected lost packet(s) */
2440
2441static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2442{
2443        struct tcp_sock *tp = tcp_sk(sk);
2444
2445        if (tcp_is_reno(tp)) {
2446                tcp_mark_head_lost(sk, 1, 1);
2447        } else if (tcp_is_fack(tp)) {
2448                int lost = tp->fackets_out - tp->reordering;
2449                if (lost <= 0)
2450                        lost = 1;
2451                tcp_mark_head_lost(sk, lost, 0);
2452        } else {
2453                int sacked_upto = tp->sacked_out - tp->reordering;
2454                if (sacked_upto >= 0)
2455                        tcp_mark_head_lost(sk, sacked_upto, 0);
2456                else if (fast_rexmit)
2457                        tcp_mark_head_lost(sk, 1, 1);
2458        }
2459
2460        tcp_timeout_skbs(sk);
2461}
2462
2463/* CWND moderation, preventing bursts due to too big ACKs
2464 * in dubious situations.
2465 */
2466static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2467{
2468        tp->snd_cwnd = min(tp->snd_cwnd,
2469                           tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2470        tp->snd_cwnd_stamp = tcp_time_stamp;
2471}
2472
2473/* Nothing was retransmitted or returned timestamp is less
2474 * than timestamp of the first retransmission.
2475 */
2476static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2477{
2478        return !tp->retrans_stamp ||
2479                (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2480                 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2481}
2482
2483/* Undo procedures. */
2484
2485#if FASTRETRANS_DEBUG > 1
2486static void DBGUNDO(struct sock *sk, const char *msg)
2487{
2488        struct tcp_sock *tp = tcp_sk(sk);
2489        struct inet_sock *inet = inet_sk(sk);
2490
2491        if (sk->sk_family == AF_INET) {
2492                pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2493                         msg,
2494                         &inet->inet_daddr, ntohs(inet->inet_dport),
2495                         tp->snd_cwnd, tcp_left_out(tp),
2496                         tp->snd_ssthresh, tp->prior_ssthresh,
2497                         tp->packets_out);
2498        }
2499#if IS_ENABLED(CONFIG_IPV6)
2500        else if (sk->sk_family == AF_INET6) {
2501                struct ipv6_pinfo *np = inet6_sk(sk);
2502                pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2503                         msg,
2504                         &np->daddr, ntohs(inet->inet_dport),
2505                         tp->snd_cwnd, tcp_left_out(tp),
2506                         tp->snd_ssthresh, tp->prior_ssthresh,
2507                         tp->packets_out);
2508        }
2509#endif
2510}
2511#else
2512#define DBGUNDO(x...) do { } while (0)
2513#endif
2514
2515static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2516{
2517        struct tcp_sock *tp = tcp_sk(sk);
2518
2519        if (tp->prior_ssthresh) {
2520                const struct inet_connection_sock *icsk = inet_csk(sk);
2521
2522                if (icsk->icsk_ca_ops->undo_cwnd)
2523                        tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2524                else
2525                        tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2526
2527                if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2528                        tp->snd_ssthresh = tp->prior_ssthresh;
2529                        TCP_ECN_withdraw_cwr(tp);
2530                }
2531        } else {
2532                tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2533        }
2534        tp->snd_cwnd_stamp = tcp_time_stamp;
2535}
2536
2537static inline bool tcp_may_undo(const struct tcp_sock *tp)
2538{
2539        return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2540}
2541
2542/* People celebrate: "We love our President!" */
2543static bool tcp_try_undo_recovery(struct sock *sk)
2544{
2545        struct tcp_sock *tp = tcp_sk(sk);
2546
2547        if (tcp_may_undo(tp)) {
2548                int mib_idx;
2549
2550                /* Happy end! We did not retransmit anything
2551                 * or our original transmission succeeded.
2552                 */
2553                DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2554                tcp_undo_cwr(sk, true);
2555                if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2556                        mib_idx = LINUX_MIB_TCPLOSSUNDO;
2557                else
2558                        mib_idx = LINUX_MIB_TCPFULLUNDO;
2559
2560                NET_INC_STATS_BH(sock_net(sk), mib_idx);
2561                tp->undo_marker = 0;
2562        }
2563        if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2564                /* Hold old state until something *above* high_seq
2565                 * is ACKed. For Reno it is MUST to prevent false
2566                 * fast retransmits (RFC2582). SACK TCP is safe. */
2567                tcp_moderate_cwnd(tp);
2568                return true;
2569        }
2570        tcp_set_ca_state(sk, TCP_CA_Open);
2571        return false;
2572}
2573
2574/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2575static void tcp_try_undo_dsack(struct sock *sk)
2576{
2577        struct tcp_sock *tp = tcp_sk(sk);
2578
2579        if (tp->undo_marker && !tp->undo_retrans) {
2580                DBGUNDO(sk, "D-SACK");
2581                tcp_undo_cwr(sk, true);
2582                tp->undo_marker = 0;
2583                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2584        }
2585}
2586
2587/* We can clear retrans_stamp when there are no retransmissions in the
2588 * window. It would seem that it is trivially available for us in
2589 * tp->retrans_out, however, that kind of assumptions doesn't consider
2590 * what will happen if errors occur when sending retransmission for the
2591 * second time. ...It could the that such segment has only
2592 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2593 * the head skb is enough except for some reneging corner cases that
2594 * are not worth the effort.
2595 *
2596 * Main reason for all this complexity is the fact that connection dying
2597 * time now depends on the validity of the retrans_stamp, in particular,
2598 * that successive retransmissions of a segment must not advance
2599 * retrans_stamp under any conditions.
2600 */
2601static bool tcp_any_retrans_done(const struct sock *sk)
2602{
2603        const struct tcp_sock *tp = tcp_sk(sk);
2604        struct sk_buff *skb;
2605
2606        if (tp->retrans_out)
2607                return true;
2608
2609        skb = tcp_write_queue_head(sk);
2610        if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2611                return true;
2612
2613        return false;
2614}
2615
2616/* Undo during fast recovery after partial ACK. */
2617
2618static int tcp_try_undo_partial(struct sock *sk, int acked)
2619{
2620        struct tcp_sock *tp = tcp_sk(sk);
2621        /* Partial ACK arrived. Force Hoe's retransmit. */
2622        int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2623
2624        if (tcp_may_undo(tp)) {
2625                /* Plain luck! Hole if filled with delayed
2626                 * packet, rather than with a retransmit.
2627                 */
2628                if (!tcp_any_retrans_done(sk))
2629                        tp->retrans_stamp = 0;
2630
2631                tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2632
2633                DBGUNDO(sk, "Hoe");
2634                tcp_undo_cwr(sk, false);
2635                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2636
2637                /* So... Do not make Hoe's retransmit yet.
2638                 * If the first packet was delayed, the rest
2639                 * ones are most probably delayed as well.
2640                 */
2641                failed = 0;
2642        }
2643        return failed;
2644}
2645
2646/* Undo during loss recovery after partial ACK. */
2647static bool tcp_try_undo_loss(struct sock *sk)
2648{
2649        struct tcp_sock *tp = tcp_sk(sk);
2650
2651        if (tcp_may_undo(tp)) {
2652                struct sk_buff *skb;
2653                tcp_for_write_queue(skb, sk) {
2654                        if (skb == tcp_send_head(sk))
2655                                break;
2656                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2657                }
2658
2659                tcp_clear_all_retrans_hints(tp);
2660
2661                DBGUNDO(sk, "partial loss");
2662                tp->lost_out = 0;
2663                tcp_undo_cwr(sk, true);
2664                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2665                inet_csk(sk)->icsk_retransmits = 0;
2666                tp->undo_marker = 0;
2667                if (tcp_is_sack(tp))
2668                        tcp_set_ca_state(sk, TCP_CA_Open);
2669                return true;
2670        }
2671        return false;
2672}
2673
2674/* The cwnd reduction in CWR and Recovery use the PRR algorithm
2675 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2676 * It computes the number of packets to send (sndcnt) based on packets newly
2677 * delivered:
2678 *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2679 *      cwnd reductions across a full RTT.
2680 *   2) If packets in flight is lower than ssthresh (such as due to excess
2681 *      losses and/or application stalls), do not perform any further cwnd
2682 *      reductions, but instead slow start up to ssthresh.
2683 */
2684static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2685{
2686        struct tcp_sock *tp = tcp_sk(sk);
2687
2688        tp->high_seq = tp->snd_nxt;
2689        tp->bytes_acked = 0;
2690        tp->snd_cwnd_cnt = 0;
2691        tp->prior_cwnd = tp->snd_cwnd;
2692        tp->prr_delivered = 0;
2693        tp->prr_out = 0;
2694        if (set_ssthresh)
2695                tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2696        TCP_ECN_queue_cwr(tp);
2697}
2698
2699static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2700                               int fast_rexmit)
2701{
2702        struct tcp_sock *tp = tcp_sk(sk);
2703        int sndcnt = 0;
2704        int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2705
2706        tp->prr_delivered += newly_acked_sacked;
2707        if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2708                u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2709                               tp->prior_cwnd - 1;
2710                sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2711        } else {
2712                sndcnt = min_t(int, delta,
2713                               max_t(int, tp->prr_delivered - tp->prr_out,
2714                                     newly_acked_sacked) + 1);
2715        }
2716
2717        sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2718        tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2719}
2720
2721static inline void tcp_end_cwnd_reduction(struct sock *sk)
2722{
2723        struct tcp_sock *tp = tcp_sk(sk);
2724
2725        /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2726        if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2727            (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2728                tp->snd_cwnd = tp->snd_ssthresh;
2729                tp->snd_cwnd_stamp = tcp_time_stamp;
2730        }
2731        tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2732}
2733
2734/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2735void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2736{
2737        struct tcp_sock *tp = tcp_sk(sk);
2738
2739        tp->prior_ssthresh = 0;
2740        tp->bytes_acked = 0;
2741        if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2742                tp->undo_marker = 0;
2743                tcp_init_cwnd_reduction(sk, set_ssthresh);
2744                tcp_set_ca_state(sk, TCP_CA_CWR);
2745        }
2746}
2747
2748static void tcp_try_keep_open(struct sock *sk)
2749{
2750        struct tcp_sock *tp = tcp_sk(sk);
2751        int state = TCP_CA_Open;
2752
2753        if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2754                state = TCP_CA_Disorder;
2755
2756        if (inet_csk(sk)->icsk_ca_state != state) {
2757                tcp_set_ca_state(sk, state);
2758                tp->high_seq = tp->snd_nxt;
2759        }
2760}
2761
2762static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2763{
2764        struct tcp_sock *tp = tcp_sk(sk);
2765
2766        tcp_verify_left_out(tp);
2767
2768        if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2769                tp->retrans_stamp = 0;
2770
2771        if (flag & FLAG_ECE)
2772                tcp_enter_cwr(sk, 1);
2773
2774        if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2775                tcp_try_keep_open(sk);
2776                if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2777                        tcp_moderate_cwnd(tp);
2778        } else {
2779                tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2780        }
2781}
2782
2783static void tcp_mtup_probe_failed(struct sock *sk)
2784{
2785        struct inet_connection_sock *icsk = inet_csk(sk);
2786
2787        icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2788        icsk->icsk_mtup.probe_size = 0;
2789}
2790
2791static void tcp_mtup_probe_success(struct sock *sk)
2792{
2793        struct tcp_sock *tp = tcp_sk(sk);
2794        struct inet_connection_sock *icsk = inet_csk(sk);
2795
2796        /* FIXME: breaks with very large cwnd */
2797        tp->prior_ssthresh = tcp_current_ssthresh(sk);
2798        tp->snd_cwnd = tp->snd_cwnd *
2799                       tcp_mss_to_mtu(sk, tp->mss_cache) /
2800                       icsk->icsk_mtup.probe_size;
2801        tp->snd_cwnd_cnt = 0;
2802        tp->snd_cwnd_stamp = tcp_time_stamp;
2803        tp->snd_ssthresh = tcp_current_ssthresh(sk);
2804
2805        icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2806        icsk->icsk_mtup.probe_size = 0;
2807        tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2808}
2809
2810/* Do a simple retransmit without using the backoff mechanisms in
2811 * tcp_timer. This is used for path mtu discovery.
2812 * The socket is already locked here.
2813 */
2814void tcp_simple_retransmit(struct sock *sk)
2815{
2816        const struct inet_connection_sock *icsk = inet_csk(sk);
2817        struct tcp_sock *tp = tcp_sk(sk);
2818        struct sk_buff *skb;
2819        unsigned int mss = tcp_current_mss(sk);
2820        u32 prior_lost = tp->lost_out;
2821
2822        tcp_for_write_queue(skb, sk) {
2823                if (skb == tcp_send_head(sk))
2824                        break;
2825                if (tcp_skb_seglen(skb) > mss &&
2826                    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2827                        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2828                                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2829                                tp->retrans_out -= tcp_skb_pcount(skb);
2830                        }
2831                        tcp_skb_mark_lost_uncond_verify(tp, skb);
2832                }
2833        }
2834
2835        tcp_clear_retrans_hints_partial(tp);
2836
2837        if (prior_lost == tp->lost_out)
2838                return;
2839
2840        if (tcp_is_reno(tp))
2841                tcp_limit_reno_sacked(tp);
2842
2843        tcp_verify_left_out(tp);
2844
2845        /* Don't muck with the congestion window here.
2846         * Reason is that we do not increase amount of _data_
2847         * in network, but units changed and effective
2848         * cwnd/ssthresh really reduced now.
2849         */
2850        if (icsk->icsk_ca_state != TCP_CA_Loss) {
2851                tp->high_seq = tp->snd_nxt;
2852                tp->snd_ssthresh = tcp_current_ssthresh(sk);
2853                tp->prior_ssthresh = 0;
2854                tp->undo_marker = 0;
2855                tcp_set_ca_state(sk, TCP_CA_Loss);
2856        }
2857        tcp_xmit_retransmit_queue(sk);
2858}
2859EXPORT_SYMBOL(tcp_simple_retransmit);
2860
2861static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2862{
2863        struct tcp_sock *tp = tcp_sk(sk);
2864        int mib_idx;
2865
2866        if (tcp_is_reno(tp))
2867                mib_idx = LINUX_MIB_TCPRENORECOVERY;
2868        else
2869                mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2870
2871        NET_INC_STATS_BH(sock_net(sk), mib_idx);
2872
2873        tp->prior_ssthresh = 0;
2874        tp->undo_marker = tp->snd_una;
2875        tp->undo_retrans = tp->retrans_out;
2876
2877        if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2878                if (!ece_ack)
2879                        tp->prior_ssthresh = tcp_current_ssthresh(sk);
2880                tcp_init_cwnd_reduction(sk, true);
2881        }
2882        tcp_set_ca_state(sk, TCP_CA_Recovery);
2883}
2884
2885/* Process an event, which can update packets-in-flight not trivially.
2886 * Main goal of this function is to calculate new estimate for left_out,
2887 * taking into account both packets sitting in receiver's buffer and
2888 * packets lost by network.
2889 *
2890 * Besides that it does CWND reduction, when packet loss is detected
2891 * and changes state of machine.
2892 *
2893 * It does _not_ decide what to send, it is made in function
2894 * tcp_xmit_retransmit_queue().
2895 */
2896static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2897                                  int prior_sacked, bool is_dupack,
2898                                  int flag)
2899{
2900        struct inet_connection_sock *icsk = inet_csk(sk);
2901        struct tcp_sock *tp = tcp_sk(sk);
2902        int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2903                                    (tcp_fackets_out(tp) > tp->reordering));
2904        int newly_acked_sacked = 0;
2905        int fast_rexmit = 0;
2906
2907        if (WARN_ON(!tp->packets_out && tp->sacked_out))
2908                tp->sacked_out = 0;
2909        if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2910                tp->fackets_out = 0;
2911
2912        /* Now state machine starts.
2913         * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2914        if (flag & FLAG_ECE)
2915                tp->prior_ssthresh = 0;
2916
2917        /* B. In all the states check for reneging SACKs. */
2918        if (tcp_check_sack_reneging(sk, flag))
2919                return;
2920
2921        /* C. Check consistency of the current state. */
2922        tcp_verify_left_out(tp);
2923
2924        /* D. Check state exit conditions. State can be terminated
2925         *    when high_seq is ACKed. */
2926        if (icsk->icsk_ca_state == TCP_CA_Open) {
2927                WARN_ON(tp->retrans_out != 0);
2928                tp->retrans_stamp = 0;
2929        } else if (!before(tp->snd_una, tp->high_seq)) {
2930                switch (icsk->icsk_ca_state) {
2931                case TCP_CA_Loss:
2932                        icsk->icsk_retransmits = 0;
2933                        if (tcp_try_undo_recovery(sk))
2934                                return;
2935                        break;
2936
2937                case TCP_CA_CWR:
2938                        /* CWR is to be held something *above* high_seq
2939                         * is ACKed for CWR bit to reach receiver. */
2940                        if (tp->snd_una != tp->high_seq) {
2941                                tcp_end_cwnd_reduction(sk);
2942                                tcp_set_ca_state(sk, TCP_CA_Open);
2943                        }
2944                        break;
2945
2946                case TCP_CA_Recovery:
2947                        if (tcp_is_reno(tp))
2948                                tcp_reset_reno_sack(tp);
2949                        if (tcp_try_undo_recovery(sk))
2950                                return;
2951                        tcp_end_cwnd_reduction(sk);
2952                        break;
2953                }
2954        }
2955
2956        /* E. Process state. */
2957        switch (icsk->icsk_ca_state) {
2958        case TCP_CA_Recovery:
2959                if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2960                        if (tcp_is_reno(tp) && is_dupack)
2961                                tcp_add_reno_sack(sk);
2962                } else
2963                        do_lost = tcp_try_undo_partial(sk, pkts_acked);
2964                newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2965                break;
2966        case TCP_CA_Loss:
2967                if (flag & FLAG_DATA_ACKED)
2968                        icsk->icsk_retransmits = 0;
2969                if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2970                        tcp_reset_reno_sack(tp);
2971                if (!tcp_try_undo_loss(sk)) {
2972                        tcp_moderate_cwnd(tp);
2973                        tcp_xmit_retransmit_queue(sk);
2974                        return;
2975                }
2976                if (icsk->icsk_ca_state != TCP_CA_Open)
2977                        return;
2978                /* Loss is undone; fall through to processing in Open state. */
2979        default:
2980                if (tcp_is_reno(tp)) {
2981                        if (flag & FLAG_SND_UNA_ADVANCED)
2982                                tcp_reset_reno_sack(tp);
2983                        if (is_dupack)
2984                                tcp_add_reno_sack(sk);
2985                }
2986                newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2987
2988                if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2989                        tcp_try_undo_dsack(sk);
2990
2991                if (!tcp_time_to_recover(sk, flag)) {
2992                        tcp_try_to_open(sk, flag, newly_acked_sacked);
2993                        return;
2994                }
2995
2996                /* MTU probe failure: don't reduce cwnd */
2997                if (icsk->icsk_ca_state < TCP_CA_CWR &&
2998                    icsk->icsk_mtup.probe_size &&
2999                    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3000                        tcp_mtup_probe_failed(sk);
3001                        /* Restores the reduction we did in tcp_mtup_probe() */
3002                        tp->snd_cwnd++;
3003                        tcp_simple_retransmit(sk);
3004                        return;
3005                }
3006
3007                /* Otherwise enter Recovery state */
3008                tcp_enter_recovery(sk, (flag & FLAG_ECE));
3009                fast_rexmit = 1;
3010        }
3011
3012        if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3013                tcp_update_scoreboard(sk, fast_rexmit);
3014        tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
3015        tcp_xmit_retransmit_queue(sk);
3016}
3017
3018void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3019{
3020        tcp_rtt_estimator(sk, seq_rtt);
3021        tcp_set_rto(sk);
3022        inet_csk(sk)->icsk_backoff = 0;
3023}
3024EXPORT_SYMBOL(tcp_valid_rtt_meas);
3025
3026/* Read draft-ietf-tcplw-high-performance before mucking
3027 * with this code. (Supersedes RFC1323)
3028 */
3029static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3030{
3031        /* RTTM Rule: A TSecr value received in a segment is used to
3032         * update the averaged RTT measurement only if the segment
3033         * acknowledges some new data, i.e., only if it advances the
3034         * left edge of the send window.
3035         *
3036         * See draft-ietf-tcplw-high-performance-00, section 3.3.
3037         * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3038         *
3039         * Changed: reset backoff as soon as we see the first valid sample.
3040         * If we do not, we get strongly overestimated rto. With timestamps
3041         * samples are accepted even from very old segments: f.e., when rtt=1
3042         * increases to 8, we retransmit 5 times and after 8 seconds delayed
3043         * answer arrives rto becomes 120 seconds! If at least one of segments
3044         * in window is lost... Voila.                          --ANK (010210)
3045         */
3046        struct tcp_sock *tp = tcp_sk(sk);
3047
3048        tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3049}
3050
3051static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3052{
3053        /* We don't have a timestamp. Can only use
3054         * packets that are not retransmitted to determine
3055         * rtt estimates. Also, we must not reset the
3056         * backoff for rto until we get a non-retransmitted
3057         * packet. This allows us to deal with a situation
3058         * where the network delay has increased suddenly.
3059         * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3060         */
3061
3062        if (flag & FLAG_RETRANS_DATA_ACKED)
3063                return;
3064
3065        tcp_valid_rtt_meas(sk, seq_rtt);
3066}
3067
3068static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3069                                      const s32 seq_rtt)
3070{
3071        const struct tcp_sock *tp = tcp_sk(sk);
3072        /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3073        if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3074                tcp_ack_saw_tstamp(sk, flag);
3075        else if (seq_rtt >= 0)
3076                tcp_ack_no_tstamp(sk, seq_rtt, flag);
3077}
3078
3079static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3080{
3081        const struct inet_connection_sock *icsk = inet_csk(sk);
3082        icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3083        tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3084}
3085
3086/* Restart timer after forward progress on connection.
3087 * RFC2988 recommends to restart timer to now+rto.
3088 */
3089void tcp_rearm_rto(struct sock *sk)
3090{
3091        struct tcp_sock *tp = tcp_sk(sk);
3092
3093        /* If the retrans timer is currently being used by Fast Open
3094         * for SYN-ACK retrans purpose, stay put.
3095         */
3096        if (tp->fastopen_rsk)
3097                return;
3098
3099        if (!tp->packets_out) {
3100                inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3101        } else {
3102                u32 rto = inet_csk(sk)->icsk_rto;
3103                /* Offset the time elapsed after installing regular RTO */
3104                if (tp->early_retrans_delayed) {
3105                        struct sk_buff *skb = tcp_write_queue_head(sk);
3106                        const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3107                        s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3108                        /* delta may not be positive if the socket is locked
3109                         * when the delayed ER timer fires and is rescheduled.
3110                         */
3111                        if (delta > 0)
3112                                rto = delta;
3113                }
3114                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3115                                          TCP_RTO_MAX);
3116        }
3117        tp->early_retrans_delayed = 0;
3118}
3119
3120/* This function is called when the delayed ER timer fires. TCP enters
3121 * fast recovery and performs fast-retransmit.
3122 */
3123void tcp_resume_early_retransmit(struct sock *sk)
3124{
3125        struct tcp_sock *tp = tcp_sk(sk);
3126
3127        tcp_rearm_rto(sk);
3128
3129        /* Stop if ER is disabled after the delayed ER timer is scheduled */
3130        if (!tp->do_early_retrans)
3131                return;
3132
3133        tcp_enter_recovery(sk, false);
3134        tcp_update_scoreboard(sk, 1);
3135        tcp_xmit_retransmit_queue(sk);
3136}
3137
3138/* If we get here, the whole TSO packet has not been acked. */
3139static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3140{
3141        struct tcp_sock *tp = tcp_sk(sk);
3142        u32 packets_acked;
3143
3144        BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3145
3146        packets_acked = tcp_skb_pcount(skb);
3147        if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3148                return 0;
3149        packets_acked -= tcp_skb_pcount(skb);
3150
3151        if (packets_acked) {
3152                BUG_ON(tcp_skb_pcount(skb) == 0);
3153                BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3154        }
3155
3156        return packets_acked;
3157}
3158
3159/* Remove acknowledged frames from the retransmission queue. If our packet
3160 * is before the ack sequence we can discard it as it's confirmed to have
3161 * arrived at the other end.
3162 */
3163static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3164                               u32 prior_snd_una)
3165{
3166        struct tcp_sock *tp = tcp_sk(sk);
3167        const struct inet_connection_sock *icsk = inet_csk(sk);
3168        struct sk_buff *skb;
3169        u32 now = tcp_time_stamp;
3170        int fully_acked = true;
3171        int flag = 0;
3172        u32 pkts_acked = 0;
3173        u32 reord = tp->packets_out;
3174        u32 prior_sacked = tp->sacked_out;
3175        s32 seq_rtt = -1;
3176        s32 ca_seq_rtt = -1;
3177        ktime_t last_ackt = net_invalid_timestamp();
3178
3179        while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3180                struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3181                u32 acked_pcount;
3182                u8 sacked = scb->sacked;
3183
3184                /* Determine how many packets and what bytes were acked, tso and else */
3185                if (after(scb->end_seq, tp->snd_una)) {
3186                        if (tcp_skb_pcount(skb) == 1 ||
3187                            !after(tp->snd_una, scb->seq))
3188                                break;
3189
3190                        acked_pcount = tcp_tso_acked(sk, skb);
3191                        if (!acked_pcount)
3192                                break;
3193
3194                        fully_acked = false;
3195                } else {
3196                        acked_pcount = tcp_skb_pcount(skb);
3197                }
3198
3199                if (sacked & TCPCB_RETRANS) {
3200                        if (sacked & TCPCB_SACKED_RETRANS)
3201                                tp->retrans_out -= acked_pcount;
3202                        flag |= FLAG_RETRANS_DATA_ACKED;
3203                        ca_seq_rtt = -1;
3204                        seq_rtt = -1;
3205                        if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3206                                flag |= FLAG_NONHEAD_RETRANS_ACKED;
3207                } else {
3208                        ca_seq_rtt = now - scb->when;
3209                        last_ackt = skb->tstamp;
3210                        if (seq_rtt < 0) {
3211                                seq_rtt = ca_seq_rtt;
3212                        }
3213                        if (!(sacked & TCPCB_SACKED_ACKED))
3214                                reord = min(pkts_acked, reord);
3215                }
3216
3217                if (sacked & TCPCB_SACKED_ACKED)
3218                        tp->sacked_out -= acked_pcount;
3219                if (sacked & TCPCB_LOST)
3220                        tp->lost_out -= acked_pcount;
3221
3222                tp->packets_out -= acked_pcount;
3223                pkts_acked += acked_pcount;
3224
3225                /* Initial outgoing SYN's get put onto the write_queue
3226                 * just like anything else we transmit.  It is not
3227                 * true data, and if we misinform our callers that
3228                 * this ACK acks real data, we will erroneously exit
3229                 * connection startup slow start one packet too
3230                 * quickly.  This is severely frowned upon behavior.
3231                 */
3232                if (!(scb->tcp_flags & TCPHDR_SYN)) {
3233                        flag |= FLAG_DATA_ACKED;
3234                } else {
3235                        flag |= FLAG_SYN_ACKED;
3236                        tp->retrans_stamp = 0;
3237                }
3238
3239                if (!fully_acked)
3240                        break;
3241
3242                tcp_unlink_write_queue(skb, sk);
3243                sk_wmem_free_skb(sk, skb);
3244                tp->scoreboard_skb_hint = NULL;
3245                if (skb == tp->retransmit_skb_hint)
3246                        tp->retransmit_skb_hint = NULL;
3247                if (skb == tp->lost_skb_hint)
3248                        tp->lost_skb_hint = NULL;
3249        }
3250
3251        if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3252                tp->snd_up = tp->snd_una;
3253
3254        if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3255                flag |= FLAG_SACK_RENEGING;
3256
3257        if (flag & FLAG_ACKED) {
3258                const struct tcp_congestion_ops *ca_ops
3259                        = inet_csk(sk)->icsk_ca_ops;
3260
3261                if (unlikely(icsk->icsk_mtup.probe_size &&
3262                             !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3263                        tcp_mtup_probe_success(sk);
3264                }
3265
3266                tcp_ack_update_rtt(sk, flag, seq_rtt);
3267                tcp_rearm_rto(sk);
3268
3269                if (tcp_is_reno(tp)) {
3270                        tcp_remove_reno_sacks(sk, pkts_acked);
3271                } else {
3272                        int delta;
3273
3274                        /* Non-retransmitted hole got filled? That's reordering */
3275                        if (reord < prior_fackets)
3276                                tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3277
3278                        delta = tcp_is_fack(tp) ? pkts_acked :
3279                                                  prior_sacked - tp->sacked_out;
3280                        tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3281                }
3282
3283                tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3284
3285                if (ca_ops->pkts_acked) {
3286                        s32 rtt_us = -1;
3287
3288                        /* Is the ACK triggering packet unambiguous? */
3289                        if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3290                                /* High resolution needed and available? */
3291                                if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3292                                    !ktime_equal(last_ackt,
3293                                                 net_invalid_timestamp()))
3294                                        rtt_us = ktime_us_delta(ktime_get_real(),
3295                                                                last_ackt);
3296                                else if (ca_seq_rtt >= 0)
3297                                        rtt_us = jiffies_to_usecs(ca_seq_rtt);
3298                        }
3299
3300                        ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3301                }
3302        }
3303
3304#if FASTRETRANS_DEBUG > 0
3305        WARN_ON((int)tp->sacked_out < 0);
3306        WARN_ON((int)tp->lost_out < 0);
3307        WARN_ON((int)tp->retrans_out < 0);
3308        if (!tp->packets_out && tcp_is_sack(tp)) {
3309                icsk = inet_csk(sk);
3310                if (tp->lost_out) {
3311                        pr_debug("Leak l=%u %d\n",
3312                                 tp->lost_out, icsk->icsk_ca_state);
3313                        tp->lost_out = 0;
3314                }
3315                if (tp->sacked_out) {
3316                        pr_debug("Leak s=%u %d\n",
3317                                 tp->sacked_out, icsk->icsk_ca_state);
3318                        tp->sacked_out = 0;
3319                }
3320                if (tp->retrans_out) {
3321                        pr_debug("Leak r=%u %d\n",
3322                                 tp->retrans_out, icsk->icsk_ca_state);
3323                        tp->retrans_out = 0;
3324                }
3325        }
3326#endif
3327        return flag;
3328}
3329
3330static void tcp_ack_probe(struct sock *sk)
3331{
3332        const struct tcp_sock *tp = tcp_sk(sk);
3333        struct inet_connection_sock *icsk = inet_csk(sk);
3334
3335        /* Was it a usable window open? */
3336
3337        if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3338                icsk->icsk_backoff = 0;
3339                inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3340                /* Socket must be waked up by subsequent tcp_data_snd_check().
3341                 * This function is not for random using!
3342                 */
3343        } else {
3344                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3345                                          min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3346                                          TCP_RTO_MAX);
3347        }
3348}
3349
3350static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3351{
3352        return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3353                inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3354}
3355
3356static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3357{
3358        const struct tcp_sock *tp = tcp_sk(sk);
3359        return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3360                !tcp_in_cwnd_reduction(sk);
3361}
3362
3363/* Check that window update is acceptable.
3364 * The function assumes that snd_una<=ack<=snd_next.
3365 */
3366static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3367                                        const u32 ack, const u32 ack_seq,
3368                                        const u32 nwin)
3369{
3370        return  after(ack, tp->snd_una) ||
3371                after(ack_seq, tp->snd_wl1) ||
3372                (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3373}
3374
3375/* Update our send window.
3376 *
3377 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3378 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3379 */
3380static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3381                                 u32 ack_seq)
3382{
3383        struct tcp_sock *tp = tcp_sk(sk);
3384        int flag = 0;
3385        u32 nwin = ntohs(tcp_hdr(skb)->window);
3386
3387        if (likely(!tcp_hdr(skb)->syn))
3388                nwin <<= tp->rx_opt.snd_wscale;
3389
3390        if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3391                flag |= FLAG_WIN_UPDATE;
3392                tcp_update_wl(tp, ack_seq);
3393
3394                if (tp->snd_wnd != nwin) {
3395                        tp->snd_wnd = nwin;
3396
3397                        /* Note, it is the only place, where
3398                         * fast path is recovered for sending TCP.
3399                         */
3400                        tp->pred_flags = 0;
3401                        tcp_fast_path_check(sk);
3402
3403                        if (nwin > tp->max_window) {
3404                                tp->max_window = nwin;
3405                                tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3406                        }
3407                }
3408        }
3409
3410        tp->snd_una = ack;
3411
3412        return flag;
3413}
3414
3415/* A very conservative spurious RTO response algorithm: reduce cwnd and
3416 * continue in congestion avoidance.
3417 */
3418static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3419{
3420        tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3421        tp->snd_cwnd_cnt = 0;
3422        tp->bytes_acked = 0;
3423        TCP_ECN_queue_cwr(tp);
3424        tcp_moderate_cwnd(tp);
3425}
3426
3427/* A conservative spurious RTO response algorithm: reduce cwnd using
3428 * PRR and continue in congestion avoidance.
3429 */
3430static void tcp_cwr_spur_to_response(struct sock *sk)
3431{
3432        tcp_enter_cwr(sk, 0);
3433}
3434
3435static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3436{
3437        if (flag & FLAG_ECE)
3438                tcp_cwr_spur_to_response(sk);
3439        else
3440                tcp_undo_cwr(sk, true);
3441}
3442
3443/* F-RTO spurious RTO detection algorithm (RFC4138)
3444 *
3445 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3446 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3447 * window (but not to or beyond highest sequence sent before RTO):
3448 *   On First ACK,  send two new segments out.
3449 *   On Second ACK, RTO was likely spurious. Do spurious response (response
3450 *                  algorithm is not part of the F-RTO detection algorithm
3451 *                  given in RFC4138 but can be selected separately).
3452 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3453 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3454 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3455 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3456 *
3457 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3458 * original window even after we transmit two new data segments.
3459 *
3460 * SACK version:
3461 *   on first step, wait until first cumulative ACK arrives, then move to
3462 *   the second step. In second step, the next ACK decides.
3463 *
3464 * F-RTO is implemented (mainly) in four functions:
3465 *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3466 *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3467 *     called when tcp_use_frto() showed green light
3468 *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3469 *   - tcp_enter_frto_loss() is called if there is not enough evidence
3470 *     to prove that the RTO is indeed spurious. It transfers the control
3471 *     from F-RTO to the conventional RTO recovery
3472 */
3473static bool tcp_process_frto(struct sock *sk, int flag)
3474{
3475        struct tcp_sock *tp = tcp_sk(sk);
3476
3477        tcp_verify_left_out(tp);
3478
3479        /* Duplicate the behavior from Loss state (fastretrans_alert) */
3480        if (flag & FLAG_DATA_ACKED)
3481                inet_csk(sk)->icsk_retransmits = 0;
3482
3483        if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3484            ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3485                tp->undo_marker = 0;
3486
3487        if (!before(tp->snd_una, tp->frto_highmark)) {
3488                tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3489                return true;
3490        }
3491
3492        if (!tcp_is_sackfrto(tp)) {
3493                /* RFC4138 shortcoming in step 2; should also have case c):
3494                 * ACK isn't duplicate nor advances window, e.g., opposite dir
3495                 * data, winupdate
3496                 */
3497                if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3498                        return true;
3499
3500                if (!(flag & FLAG_DATA_ACKED)) {
3501                        tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3502                                            flag);
3503                        return true;
3504                }
3505        } else {
3506                if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3507                        /* Prevent sending of new data. */
3508                        tp->snd_cwnd = min(tp->snd_cwnd,
3509                                           tcp_packets_in_flight(tp));
3510                        return true;
3511                }
3512
3513                if ((tp->frto_counter >= 2) &&
3514                    (!(flag & FLAG_FORWARD_PROGRESS) ||
3515                     ((flag & FLAG_DATA_SACKED) &&
3516                      !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3517                        /* RFC4138 shortcoming (see comment above) */
3518                        if (!(flag & FLAG_FORWARD_PROGRESS) &&
3519                            (flag & FLAG_NOT_DUP))
3520                                return true;
3521
3522                        tcp_enter_frto_loss(sk, 3, flag);
3523                        return true;
3524                }
3525        }
3526
3527        if (tp->frto_counter == 1) {
3528                /* tcp_may_send_now needs to see updated state */
3529                tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3530                tp->frto_counter = 2;
3531
3532                if (!tcp_may_send_now(sk))
3533                        tcp_enter_frto_loss(sk, 2, flag);
3534
3535                return true;
3536        } else {
3537                switch (sysctl_tcp_frto_response) {
3538                case 2:
3539                        tcp_undo_spur_to_response(sk, flag);
3540                        break;
3541                case 1:
3542                        tcp_conservative_spur_to_response(tp);
3543                        break;
3544                default:
3545                        tcp_cwr_spur_to_response(sk);
3546                        break;
3547                }
3548                tp->frto_counter = 0;
3549                tp->undo_marker = 0;
3550                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3551        }
3552        return false;
3553}
3554
3555/* This routine deals with incoming acks, but not outgoing ones. */
3556static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3557{
3558        struct inet_connection_sock *icsk = inet_csk(sk);
3559        struct tcp_sock *tp = tcp_sk(sk);
3560        u32 prior_snd_una = tp->snd_una;
3561        u32 ack_seq = TCP_SKB_CB(skb)->seq;
3562        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3563        bool is_dupack = false;
3564        u32 prior_in_flight;
3565        u32 prior_fackets;
3566        int prior_packets;
3567        int prior_sacked = tp->sacked_out;
3568        int pkts_acked = 0;
3569        bool frto_cwnd = false;
3570
3571        /* If the ack is older than previous acks
3572         * then we can probably ignore it.
3573         */
3574        if (before(ack, prior_snd_una))
3575                goto old_ack;
3576
3577        /* If the ack includes data we haven't sent yet, discard
3578         * this segment (RFC793 Section 3.9).
3579         */
3580        if (after(ack, tp->snd_nxt))
3581                goto invalid_ack;
3582
3583        if (tp->early_retrans_delayed)
3584                tcp_rearm_rto(sk);
3585
3586        if (after(ack, prior_snd_una))
3587                flag |= FLAG_SND_UNA_ADVANCED;
3588
3589        if (sysctl_tcp_abc) {
3590                if (icsk->icsk_ca_state < TCP_CA_CWR)
3591                        tp->bytes_acked += ack - prior_snd_una;
3592                else if (icsk->icsk_ca_state == TCP_CA_Loss)
3593                        /* we assume just one segment left network */
3594                        tp->bytes_acked += min(ack - prior_snd_una,
3595                                               tp->mss_cache);
3596        }
3597
3598        prior_fackets = tp->fackets_out;
3599        prior_in_flight = tcp_packets_in_flight(tp);
3600
3601        if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3602                /* Window is constant, pure forward advance.
3603                 * No more checks are required.
3604                 * Note, we use the fact that SND.UNA>=SND.WL2.
3605                 */
3606                tcp_update_wl(tp, ack_seq);
3607                tp->snd_una = ack;
3608                flag |= FLAG_WIN_UPDATE;
3609
3610                tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3611
3612                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3613        } else {
3614                if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3615                        flag |= FLAG_DATA;
3616                else
3617                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3618
3619                flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3620
3621                if (TCP_SKB_CB(skb)->sacked)
3622                        flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3623
3624                if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3625                        flag |= FLAG_ECE;
3626
3627                tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3628        }
3629
3630        /* We passed data and got it acked, remove any soft error
3631         * log. Something worked...
3632         */
3633        sk->sk_err_soft = 0;
3634        icsk->icsk_probes_out = 0;
3635        tp->rcv_tstamp = tcp_time_stamp;
3636        prior_packets = tp->packets_out;
3637        if (!prior_packets)
3638                goto no_queue;
3639
3640        /* See if we can take anything off of the retransmit queue. */
3641        flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3642
3643        pkts_acked = prior_packets - tp->packets_out;
3644
3645        if (tp->frto_counter)
3646                frto_cwnd = tcp_process_frto(sk, flag);
3647        /* Guarantee sacktag reordering detection against wrap-arounds */
3648        if (before(tp->frto_highmark, tp->snd_una))
3649                tp->frto_highmark = 0;
3650
3651        if (tcp_ack_is_dubious(sk, flag)) {
3652                /* Advance CWND, if state allows this. */
3653                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3654                    tcp_may_raise_cwnd(sk, flag))
3655                        tcp_cong_avoid(sk, ack, prior_in_flight);
3656                is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3657                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3658                                      is_dupack, flag);
3659        } else {
3660                if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3661                        tcp_cong_avoid(sk, ack, prior_in_flight);
3662        }
3663
3664        if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3665                struct dst_entry *dst = __sk_dst_get(sk);
3666                if (dst)
3667                        dst_confirm(dst);
3668        }
3669        return 1;
3670
3671no_queue:
3672        /* If data was DSACKed, see if we can undo a cwnd reduction. */
3673        if (flag & FLAG_DSACKING_ACK)
3674                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3675                                      is_dupack, flag);
3676        /* If this ack opens up a zero window, clear backoff.  It was
3677         * being used to time the probes, and is probably far higher than
3678         * it needs to be for normal retransmission.
3679         */
3680        if (tcp_send_head(sk))
3681                tcp_ack_probe(sk);
3682        return 1;
3683
3684invalid_ack:
3685        SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3686        return -1;
3687
3688old_ack:
3689        /* If data was SACKed, tag it and see if we should send more data.
3690         * If data was DSACKed, see if we can undo a cwnd reduction.
3691         */
3692        if (TCP_SKB_CB(skb)->sacked) {
3693                flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3694                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3695                                      is_dupack, flag);
3696        }
3697
3698        SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3699        return 0;
3700}
3701
3702/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3703 * But, this can also be called on packets in the established flow when
3704 * the fast version below fails.
3705 */
3706void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3707                       const u8 **hvpp, int estab,
3708                       struct tcp_fastopen_cookie *foc)
3709{
3710        const unsigned char *ptr;
3711        const struct tcphdr *th = tcp_hdr(skb);
3712        int length = (th->doff * 4) - sizeof(struct tcphdr);
3713
3714        ptr = (const unsigned char *)(th + 1);
3715        opt_rx->saw_tstamp = 0;
3716
3717        while (length > 0) {
3718                int opcode = *ptr++;
3719                int opsize;
3720
3721                switch (opcode) {
3722                case TCPOPT_EOL:
3723                        return;
3724                case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3725                        length--;
3726                        continue;
3727                default:
3728                        opsize = *ptr++;
3729                        if (opsize < 2) /* "silly options" */
3730                                return;
3731                        if (opsize > length)
3732                                return; /* don't parse partial options */
3733                        switch (opcode) {
3734                        case TCPOPT_MSS:
3735                                if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3736                                        u16 in_mss = get_unaligned_be16(ptr);
3737                                        if (in_mss) {
3738                                                if (opt_rx->user_mss &&
3739                                                    opt_rx->user_mss < in_mss)
3740                                                        in_mss = opt_rx->user_mss;
3741                                                opt_rx->mss_clamp = in_mss;
3742                                        }
3743                                }
3744                                break;
3745                        case TCPOPT_WINDOW:
3746                                if (opsize == TCPOLEN_WINDOW && th->syn &&
3747                                    !estab && sysctl_tcp_window_scaling) {
3748                                        __u8 snd_wscale = *(__u8 *)ptr;
3749                                        opt_rx->wscale_ok = 1;
3750                                        if (snd_wscale > 14) {
3751                                                net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3752                                                                     __func__,
3753                                                                     snd_wscale);
3754                                                snd_wscale = 14;
3755                                        }
3756                                        opt_rx->snd_wscale = snd_wscale;
3757                                }
3758                                break;
3759                        case TCPOPT_TIMESTAMP:
3760                                if ((opsize == TCPOLEN_TIMESTAMP) &&
3761                                    ((estab && opt_rx->tstamp_ok) ||
3762                                     (!estab && sysctl_tcp_timestamps))) {
3763                                        opt_rx->saw_tstamp = 1;
3764                                        opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3765                                        opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3766                                }
3767                                break;
3768                        case TCPOPT_SACK_PERM:
3769                                if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3770                                    !estab && sysctl_tcp_sack) {
3771                                        opt_rx->sack_ok = TCP_SACK_SEEN;
3772                                        tcp_sack_reset(opt_rx);
3773                                }
3774                                break;
3775
3776                        case TCPOPT_SACK:
3777                                if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3778                                   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3779                                   opt_rx->sack_ok) {
3780                                        TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3781                                }
3782                                break;
3783#ifdef CONFIG_TCP_MD5SIG
3784                        case TCPOPT_MD5SIG:
3785                                /*
3786                                 * The MD5 Hash has already been
3787                                 * checked (see tcp_v{4,6}_do_rcv()).
3788                                 */
3789                                break;
3790#endif
3791                        case TCPOPT_COOKIE:
3792                                /* This option is variable length.
3793                                 */
3794                                switch (opsize) {
3795                                case TCPOLEN_COOKIE_BASE:
3796                                        /* not yet implemented */
3797                                        break;
3798                                case TCPOLEN_COOKIE_PAIR:
3799                                        /* not yet implemented */
3800                                        break;
3801                                case TCPOLEN_COOKIE_MIN+0:
3802                                case TCPOLEN_COOKIE_MIN+2:
3803                                case TCPOLEN_COOKIE_MIN+4:
3804                                case TCPOLEN_COOKIE_MIN+6:
3805                                case TCPOLEN_COOKIE_MAX:
3806                                        /* 16-bit multiple */
3807                                        opt_rx->cookie_plus = opsize;
3808                                        *hvpp = ptr;
3809                                        break;
3810                                default:
3811                                        /* ignore option */
3812                                        break;
3813                                }
3814                                break;
3815
3816                        case TCPOPT_EXP:
3817                                /* Fast Open option shares code 254 using a
3818                                 * 16 bits magic number. It's valid only in
3819                                 * SYN or SYN-ACK with an even size.
3820                                 */
3821                                if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3822                                    get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3823                                    foc == NULL || !th->syn || (opsize & 1))
3824                                        break;
3825                                foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3826                                if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3827                                    foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3828                                        memcpy(foc->val, ptr + 2, foc->len);
3829                                else if (foc->len != 0)
3830                                        foc->len = -1;
3831                                break;
3832
3833                        }
3834                        ptr += opsize-2;
3835                        length -= opsize;
3836                }
3837        }
3838}
3839EXPORT_SYMBOL(tcp_parse_options);
3840
3841static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3842{
3843        const __be32 *ptr = (const __be32 *)(th + 1);
3844
3845        if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3846                          | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3847                tp->rx_opt.saw_tstamp = 1;
3848                ++ptr;
3849                tp->rx_opt.rcv_tsval = ntohl(*ptr);
3850                ++ptr;
3851                tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3852                return true;
3853        }
3854        return false;
3855}
3856
3857/* Fast parse options. This hopes to only see timestamps.
3858 * If it is wrong it falls back on tcp_parse_options().
3859 */
3860static bool tcp_fast_parse_options(const struct sk_buff *skb,
3861                                   const struct tcphdr *th,
3862                                   struct tcp_sock *tp, const u8 **hvpp)
3863{
3864        /* In the spirit of fast parsing, compare doff directly to constant
3865         * values.  Because equality is used, short doff can be ignored here.
3866         */
3867        if (th->doff == (sizeof(*th) / 4)) {
3868                tp->rx_opt.saw_tstamp = 0;
3869                return false;
3870        } else if (tp->rx_opt.tstamp_ok &&
3871                   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3872                if (tcp_parse_aligned_timestamp(tp, th))
3873                        return true;
3874        }
3875        tcp_parse_options(skb, &tp->rx_opt, hvpp, 1, NULL);
3876        return true;
3877}
3878
3879#ifdef CONFIG_TCP_MD5SIG
3880/*
3881 * Parse MD5 Signature option
3882 */
3883const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3884{
3885        int length = (th->doff << 2) - sizeof(*th);
3886        const u8 *ptr = (const u8 *)(th + 1);
3887
3888        /* If the TCP option is too short, we can short cut */
3889        if (length < TCPOLEN_MD5SIG)
3890                return NULL;
3891
3892        while (length > 0) {
3893                int opcode = *ptr++;
3894                int opsize;
3895
3896                switch(opcode) {
3897                case TCPOPT_EOL:
3898                        return NULL;
3899                case TCPOPT_NOP:
3900                        length--;
3901                        continue;
3902                default:
3903                        opsize = *ptr++;
3904                        if (opsize < 2 || opsize > length)
3905                                return NULL;
3906                        if (opcode == TCPOPT_MD5SIG)
3907                                return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3908                }
3909                ptr += opsize - 2;
3910                length -= opsize;
3911        }
3912        return NULL;
3913}
3914EXPORT_SYMBOL(tcp_parse_md5sig_option);
3915#endif
3916
3917static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3918{
3919        tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3920        tp->rx_opt.ts_recent_stamp = get_seconds();
3921}
3922
3923static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3924{
3925        if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3926                /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3927                 * extra check below makes sure this can only happen
3928                 * for pure ACK frames.  -DaveM
3929                 *
3930                 * Not only, also it occurs for expired timestamps.
3931                 */
3932
3933                if (tcp_paws_check(&tp->rx_opt, 0))
3934                        tcp_store_ts_recent(tp);
3935        }
3936}
3937
3938/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3939 *
3940 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3941 * it can pass through stack. So, the following predicate verifies that
3942 * this segment is not used for anything but congestion avoidance or
3943 * fast retransmit. Moreover, we even are able to eliminate most of such
3944 * second order effects, if we apply some small "replay" window (~RTO)
3945 * to timestamp space.
3946 *
3947 * All these measures still do not guarantee that we reject wrapped ACKs
3948 * on networks with high bandwidth, when sequence space is recycled fastly,
3949 * but it guarantees that such events will be very rare and do not affect
3950 * connection seriously. This doesn't look nice, but alas, PAWS is really
3951 * buggy extension.
3952 *
3953 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3954 * states that events when retransmit arrives after original data are rare.
3955 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3956 * the biggest problem on large power networks even with minor reordering.
3957 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3958 * up to bandwidth of 18Gigabit/sec. 8) ]
3959 */
3960
3961static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3962{
3963        const struct tcp_sock *tp = tcp_sk(sk);
3964        const struct tcphdr *th = tcp_hdr(skb);
3965        u32 seq = TCP_SKB_CB(skb)->seq;
3966        u32 ack = TCP_SKB_CB(skb)->ack_seq;
3967
3968        return (/* 1. Pure ACK with correct sequence number. */
3969                (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3970
3971                /* 2. ... and duplicate ACK. */
3972                ack == tp->snd_una &&
3973
3974                /* 3. ... and does not update window. */
3975                !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3976
3977                /* 4. ... and sits in replay window. */
3978                (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3979}
3980
3981static inline bool tcp_paws_discard(const struct sock *sk,
3982                                   const struct sk_buff *skb)
3983{
3984        const struct tcp_sock *tp = tcp_sk(sk);
3985
3986        return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3987               !tcp_disordered_ack(sk, skb);
3988}
3989
3990/* Check segment sequence number for validity.
3991 *
3992 * Segment controls are considered valid, if the segment
3993 * fits to the window after truncation to the window. Acceptability
3994 * of data (and SYN, FIN, of course) is checked separately.
3995 * See tcp_data_queue(), for example.
3996 *
3997 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3998 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3999 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4000 * (borrowed from freebsd)
4001 */
4002
4003static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4004{
4005        return  !before(end_seq, tp->rcv_wup) &&
4006                !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4007}
4008
4009/* When we get a reset we do this. */
4010void tcp_reset(struct sock *sk)
4011{
4012        /* We want the right error as BSD sees it (and indeed as we do). */
4013        switch (sk->sk_state) {
4014        case TCP_SYN_SENT:
4015                sk->sk_err = ECONNREFUSED;
4016                break;
4017        case TCP_CLOSE_WAIT:
4018                sk->sk_err = EPIPE;
4019                break;
4020        case TCP_CLOSE:
4021                return;
4022        default:
4023                sk->sk_err = ECONNRESET;
4024        }
4025        /* This barrier is coupled with smp_rmb() in tcp_poll() */
4026        smp_wmb();
4027
4028        if (!sock_flag(sk, SOCK_DEAD))
4029                sk->sk_error_report(sk);
4030
4031        tcp_done(sk);
4032}
4033
4034/*
4035 *      Process the FIN bit. This now behaves as it is supposed to work
4036 *      and the FIN takes effect when it is validly part of sequence
4037 *      space. Not before when we get holes.
4038 *
4039 *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4040 *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
4041 *      TIME-WAIT)
4042 *
4043 *      If we are in FINWAIT-1, a received FIN indicates simultaneous
4044 *      close and we go into CLOSING (and later onto TIME-WAIT)
4045 *
4046 *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4047 */
4048static void tcp_fin(struct sock *sk)
4049{
4050        struct tcp_sock *tp = tcp_sk(sk);
4051
4052        inet_csk_schedule_ack(sk);
4053
4054        sk->sk_shutdown |= RCV_SHUTDOWN;
4055        sock_set_flag(sk, SOCK_DONE);
4056
4057        switch (sk->sk_state) {
4058        case TCP_SYN_RECV:
4059        case TCP_ESTABLISHED:
4060                /* Move to CLOSE_WAIT */
4061                tcp_set_state(sk, TCP_CLOSE_WAIT);
4062                inet_csk(sk)->icsk_ack.pingpong = 1;
4063                break;
4064
4065        case TCP_CLOSE_WAIT:
4066        case TCP_CLOSING:
4067                /* Received a retransmission of the FIN, do
4068                 * nothing.
4069                 */
4070                break;
4071        case TCP_LAST_ACK:
4072                /* RFC793: Remain in the LAST-ACK state. */
4073                break;
4074
4075        case TCP_FIN_WAIT1:
4076                /* This case occurs when a simultaneous close
4077                 * happens, we must ack the received FIN and
4078                 * enter the CLOSING state.
4079                 */
4080                tcp_send_ack(sk);
4081                tcp_set_state(sk, TCP_CLOSING);
4082                break;
4083        case TCP_FIN_WAIT2:
4084                /* Received a FIN -- send ACK and enter TIME_WAIT. */
4085                tcp_send_ack(sk);
4086                tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4087                break;
4088        default:
4089                /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4090                 * cases we should never reach this piece of code.
4091                 */
4092                pr_err("%s: Impossible, sk->sk_state=%d\n",
4093                       __func__, sk->sk_state);
4094                break;
4095        }
4096
4097        /* It _is_ possible, that we have something out-of-order _after_ FIN.
4098         * Probably, we should reset in this case. For now drop them.
4099         */
4100        __skb_queue_purge(&tp->out_of_order_queue);
4101        if (tcp_is_sack(tp))
4102                tcp_sack_reset(&tp->rx_opt);
4103        sk_mem_reclaim(sk);
4104
4105        if (!sock_flag(sk, SOCK_DEAD)) {
4106                sk->sk_state_change(sk);
4107
4108                /* Do not send POLL_HUP for half duplex close. */
4109                if (sk->sk_shutdown == SHUTDOWN_MASK ||
4110                    sk->sk_state == TCP_CLOSE)
4111                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4112                else
4113                        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4114        }
4115}
4116
4117static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4118                                  u32 end_seq)
4119{
4120        if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4121                if (before(seq, sp->start_seq))
4122                        sp->start_seq = seq;
4123                if (after(end_seq, sp->end_seq))
4124                        sp->end_seq = end_seq;
4125                return true;
4126        }
4127        return false;
4128}
4129
4130static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4131{
4132        struct tcp_sock *tp = tcp_sk(sk);
4133
4134        if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4135                int mib_idx;
4136
4137                if (before(seq, tp->rcv_nxt))
4138                        mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4139                else
4140                        mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4141
4142                NET_INC_STATS_BH(sock_net(sk), mib_idx);
4143
4144                tp->rx_opt.dsack = 1;
4145                tp->duplicate_sack[0].start_seq = seq;
4146                tp->duplicate_sack[0].end_seq = end_seq;
4147        }
4148}
4149
4150static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4151{
4152        struct tcp_sock *tp = tcp_sk(sk);
4153
4154        if (!tp->rx_opt.dsack)
4155                tcp_dsack_set(sk, seq, end_seq);
4156        else
4157                tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4158}
4159
4160static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4161{
4162        struct tcp_sock *tp = tcp_sk(sk);
4163
4164        if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4165            before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4166                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4167                tcp_enter_quickack_mode(sk);
4168
4169                if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4170                        u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4171
4172                        if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4173                                end_seq = tp->rcv_nxt;
4174                        tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4175                }
4176        }
4177
4178        tcp_send_ack(sk);
4179}
4180
4181/* These routines update the SACK block as out-of-order packets arrive or
4182 * in-order packets close up the sequence space.
4183 */
4184static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4185{
4186        int this_sack;
4187        struct tcp_sack_block *sp = &tp->selective_acks[0];
4188        struct tcp_sack_block *swalk = sp + 1;
4189
4190        /* See if the recent change to the first SACK eats into
4191         * or hits the sequence space of other SACK blocks, if so coalesce.
4192         */
4193        for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4194                if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4195                        int i;
4196
4197                        /* Zap SWALK, by moving every further SACK up by one slot.
4198                         * Decrease num_sacks.
4199                         */
4200                        tp->rx_opt.num_sacks--;
4201                        for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4202                                sp[i] = sp[i + 1];
4203                        continue;
4204                }
4205                this_sack++, swalk++;
4206        }
4207}
4208
4209static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4210{
4211        struct tcp_sock *tp = tcp_sk(sk);
4212        struct tcp_sack_block *sp = &tp->selective_acks[0];
4213        int cur_sacks = tp->rx_opt.num_sacks;
4214        int this_sack;
4215
4216        if (!cur_sacks)
4217                goto new_sack;
4218
4219        for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4220                if (tcp_sack_extend(sp, seq, end_seq)) {
4221                        /* Rotate this_sack to the first one. */
4222                        for (; this_sack > 0; this_sack--, sp--)
4223                                swap(*sp, *(sp - 1));
4224                        if (cur_sacks > 1)
4225                                tcp_sack_maybe_coalesce(tp);
4226                        return;
4227                }
4228        }
4229
4230        /* Could not find an adjacent existing SACK, build a new one,
4231         * put it at the front, and shift everyone else down.  We
4232         * always know there is at least one SACK present already here.
4233         *
4234         * If the sack array is full, forget about the last one.
4235         */
4236        if (this_sack >= TCP_NUM_SACKS) {
4237                this_sack--;
4238                tp->rx_opt.num_sacks--;
4239                sp--;
4240        }
4241        for (; this_sack > 0; this_sack--, sp--)
4242                *sp = *(sp - 1);
4243
4244new_sack:
4245        /* Build the new head SACK, and we're done. */
4246        sp->start_seq = seq;
4247        sp->end_seq = end_seq;
4248        tp->rx_opt.num_sacks++;
4249}
4250
4251/* RCV.NXT advances, some SACKs should be eaten. */
4252
4253static void tcp_sack_remove(struct tcp_sock *tp)
4254{
4255        struct tcp_sack_block *sp = &tp->selective_acks[0];
4256        int num_sacks = tp->rx_opt.num_sacks;
4257        int this_sack;
4258
4259        /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4260        if (skb_queue_empty(&tp->out_of_order_queue)) {
4261                tp->rx_opt.num_sacks = 0;
4262                return;
4263        }
4264
4265        for (this_sack = 0; this_sack < num_sacks;) {
4266                /* Check if the start of the sack is covered by RCV.NXT. */
4267                if (!before(tp->rcv_nxt, sp->start_seq)) {
4268                        int i;
4269
4270                        /* RCV.NXT must cover all the block! */
4271                        WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4272
4273                        /* Zap this SACK, by moving forward any other SACKS. */
4274                        for (i=this_sack+1; i < num_sacks; i++)
4275                                tp->selective_acks[i-1] = tp->selective_acks[i];
4276                        num_sacks--;
4277                        continue;
4278                }
4279                this_sack++;
4280                sp++;
4281        }
4282        tp->rx_opt.num_sacks = num_sacks;
4283}
4284
4285/* This one checks to see if we can put data from the
4286 * out_of_order queue into the receive_queue.
4287 */
4288static void tcp_ofo_queue(struct sock *sk)
4289{
4290        struct tcp_sock *tp = tcp_sk(sk);
4291        __u32 dsack_high = tp->rcv_nxt;
4292        struct sk_buff *skb;
4293
4294        while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4295                if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4296                        break;
4297
4298                if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4299                        __u32 dsack = dsack_high;
4300                        if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4301                                dsack_high = TCP_SKB_CB(skb)->end_seq;
4302                        tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4303                }
4304
4305                if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4306                        SOCK_DEBUG(sk, "ofo packet was already received\n");
4307                        __skb_unlink(skb, &tp->out_of_order_queue);
4308                        __kfree_skb(skb);
4309                        continue;
4310                }
4311                SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4312                           tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4313                           TCP_SKB_CB(skb)->end_seq);
4314
4315                __skb_unlink(skb, &tp->out_of_order_queue);
4316                __skb_queue_tail(&sk->sk_receive_queue, skb);
4317                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4318                if (tcp_hdr(skb)->fin)
4319                        tcp_fin(sk);
4320        }
4321}
4322
4323static bool tcp_prune_ofo_queue(struct sock *sk);
4324static int tcp_prune_queue(struct sock *sk);
4325
4326static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4327                                 unsigned int size)
4328{
4329        if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4330            !sk_rmem_schedule(sk, skb, size)) {
4331
4332                if (tcp_prune_queue(sk) < 0)
4333                        return -1;
4334
4335                if (!sk_rmem_schedule(sk, skb, size)) {
4336                        if (!tcp_prune_ofo_queue(sk))
4337                                return -1;
4338
4339                        if (!sk_rmem_schedule(sk, skb, size))
4340                                return -1;
4341                }
4342        }
4343        return 0;
4344}
4345
4346/**
4347 * tcp_try_coalesce - try to merge skb to prior one
4348 * @sk: socket
4349 * @to: prior buffer
4350 * @from: buffer to add in queue
4351 * @fragstolen: pointer to boolean
4352 *
4353 * Before queueing skb @from after @to, try to merge them
4354 * to reduce overall memory use and queue lengths, if cost is small.
4355 * Packets in ofo or receive queues can stay a long time.
4356 * Better try to coalesce them right now to avoid future collapses.
4357 * Returns true if caller should free @from instead of queueing it
4358 */
4359static bool tcp_try_coalesce(struct sock *sk,
4360                             struct sk_buff *to,
4361                             struct sk_buff *from,
4362                             bool *fragstolen)
4363{
4364        int delta;
4365
4366        *fragstolen = false;
4367
4368        if (tcp_hdr(from)->fin)
4369                return false;
4370
4371        /* Its possible this segment overlaps with prior segment in queue */
4372        if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4373                return false;
4374
4375        if (!skb_try_coalesce(to, from, fragstolen, &delta))
4376                return false;
4377
4378        atomic_add(delta, &sk->sk_rmem_alloc);
4379        sk_mem_charge(sk, delta);
4380        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4381        TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4382        TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4383        return true;
4384}
4385
4386static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4387{
4388        struct tcp_sock *tp = tcp_sk(sk);
4389        struct sk_buff *skb1;
4390        u32 seq, end_seq;
4391
4392        TCP_ECN_check_ce(tp, skb);
4393
4394        if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4395                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4396                __kfree_skb(skb);
4397                return;
4398        }
4399
4400        /* Disable header prediction. */
4401        tp->pred_flags = 0;
4402        inet_csk_schedule_ack(sk);
4403
4404        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4405        SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4406                   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4407
4408        skb1 = skb_peek_tail(&tp->out_of_order_queue);
4409        if (!skb1) {
4410                /* Initial out of order segment, build 1 SACK. */
4411                if (tcp_is_sack(tp)) {
4412                        tp->rx_opt.num_sacks = 1;
4413                        tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4414                        tp->selective_acks[0].end_seq =
4415                                                TCP_SKB_CB(skb)->end_seq;
4416                }
4417                __skb_queue_head(&tp->out_of_order_queue, skb);
4418                goto end;
4419        }
4420
4421        seq = TCP_SKB_CB(skb)->seq;
4422        end_seq = TCP_SKB_CB(skb)->end_seq;
4423
4424        if (seq == TCP_SKB_CB(skb1)->end_seq) {
4425                bool fragstolen;
4426
4427                if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4428                        __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4429                } else {
4430                        kfree_skb_partial(skb, fragstolen);
4431                        skb = NULL;
4432                }
4433
4434                if (!tp->rx_opt.num_sacks ||
4435                    tp->selective_acks[0].end_seq != seq)
4436                        goto add_sack;
4437
4438                /* Common case: data arrive in order after hole. */
4439                tp->selective_acks[0].end_seq = end_seq;
4440                goto end;
4441        }
4442
4443        /* Find place to insert this segment. */
4444        while (1) {
4445                if (!after(TCP_SKB_CB(skb1)->seq, seq))
4446                        break;
4447                if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4448                        skb1 = NULL;
4449                        break;
4450                }
4451                skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4452        }
4453
4454        /* Do skb overlap to previous one? */
4455        if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4456                if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4457                        /* All the bits are present. Drop. */
4458                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4459                        __kfree_skb(skb);
4460                        skb = NULL;
4461                        tcp_dsack_set(sk, seq, end_seq);
4462                        goto add_sack;
4463                }
4464                if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4465                        /* Partial overlap. */
4466                        tcp_dsack_set(sk, seq,
4467                                      TCP_SKB_CB(skb1)->end_seq);
4468                } else {
4469                        if (skb_queue_is_first(&tp->out_of_order_queue,
4470                                               skb1))
4471                                skb1 = NULL;
4472                        else
4473                                skb1 = skb_queue_prev(
4474                                        &tp->out_of_order_queue,
4475                                        skb1);
4476                }
4477        }
4478        if (!skb1)
4479                __skb_queue_head(&tp->out_of_order_queue, skb);
4480        else
4481                __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4482
4483        /* And clean segments covered by new one as whole. */
4484        while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4485                skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4486
4487                if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4488                        break;
4489                if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4490                        tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4491                                         end_seq);
4492                        break;
4493                }
4494                __skb_unlink(skb1, &tp->out_of_order_queue);
4495                tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4496                                 TCP_SKB_CB(skb1)->end_seq);
4497                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4498                __kfree_skb(skb1);
4499        }
4500
4501add_sack:
4502        if (tcp_is_sack(tp))
4503                tcp_sack_new_ofo_skb(sk, seq, end_seq);
4504end:
4505        if (skb)
4506                skb_set_owner_r(skb, sk);
4507}
4508
4509static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4510                  bool *fragstolen)
4511{
4512        int eaten;
4513        struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4514
4515        __skb_pull(skb, hdrlen);
4516        eaten = (tail &&
4517                 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4518        tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4519        if (!eaten) {
4520                __skb_queue_tail(&sk->sk_receive_queue, skb);
4521                skb_set_owner_r(skb, sk);
4522        }
4523        return eaten;
4524}
4525
4526int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4527{
4528        struct sk_buff *skb = NULL;
4529        struct tcphdr *th;
4530        bool fragstolen;
4531
4532        if (size == 0)
4533                return 0;
4534
4535        skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4536        if (!skb)
4537                goto err;
4538
4539        if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4540                goto err_free;
4541
4542        th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4543        skb_reset_transport_header(skb);
4544        memset(th, 0, sizeof(*th));
4545
4546        if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4547                goto err_free;
4548
4549        TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4550        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4551        TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4552
4553        if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4554                WARN_ON_ONCE(fragstolen); /* should not happen */
4555                __kfree_skb(skb);
4556        }
4557        return size;
4558
4559err_free:
4560        kfree_skb(skb);
4561err:
4562        return -ENOMEM;
4563}
4564
4565static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4566{
4567        const struct tcphdr *th = tcp_hdr(skb);
4568        struct tcp_sock *tp = tcp_sk(sk);
4569        int eaten = -1;
4570        bool fragstolen = false;
4571
4572        if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4573                goto drop;
4574
4575        skb_dst_drop(skb);
4576        __skb_pull(skb, th->doff * 4);
4577
4578        TCP_ECN_accept_cwr(tp, skb);
4579
4580        tp->rx_opt.dsack = 0;
4581
4582        /*  Queue data for delivery to the user.
4583         *  Packets in sequence go to the receive queue.
4584         *  Out of sequence packets to the out_of_order_queue.
4585         */
4586        if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4587                if (tcp_receive_window(tp) == 0)
4588                        goto out_of_window;
4589
4590                /* Ok. In sequence. In window. */
4591                if (tp->ucopy.task == current &&
4592                    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4593                    sock_owned_by_user(sk) && !tp->urg_data) {
4594                        int chunk = min_t(unsigned int, skb->len,
4595                                          tp->ucopy.len);
4596
4597                        __set_current_state(TASK_RUNNING);
4598
4599                        local_bh_enable();
4600                        if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4601                                tp->ucopy.len -= chunk;
4602                                tp->copied_seq += chunk;
4603                                eaten = (chunk == skb->len);
4604                                tcp_rcv_space_adjust(sk);
4605                        }
4606                        local_bh_disable();
4607                }
4608
4609                if (eaten <= 0) {
4610queue_and_out:
4611                        if (eaten < 0 &&
4612                            tcp_try_rmem_schedule(sk, skb, skb->truesize))
4613                                goto drop;
4614
4615                        eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4616                }
4617                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4618                if (skb->len)
4619                        tcp_event_data_recv(sk, skb);
4620                if (th->fin)
4621                        tcp_fin(sk);
4622
4623                if (!skb_queue_empty(&tp->out_of_order_queue)) {
4624                        tcp_ofo_queue(sk);
4625
4626                        /* RFC2581. 4.2. SHOULD send immediate ACK, when
4627                         * gap in queue is filled.
4628                         */
4629                        if (skb_queue_empty(&tp->out_of_order_queue))
4630                                inet_csk(sk)->icsk_ack.pingpong = 0;
4631                }
4632
4633                if (tp->rx_opt.num_sacks)
4634                        tcp_sack_remove(tp);
4635
4636                tcp_fast_path_check(sk);
4637
4638                if (eaten > 0)
4639                        kfree_skb_partial(skb, fragstolen);
4640                if (!sock_flag(sk, SOCK_DEAD))
4641                        sk->sk_data_ready(sk, 0);
4642                return;
4643        }
4644
4645        if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4646                /* A retransmit, 2nd most common case.  Force an immediate ack. */
4647                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4648                tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4649
4650out_of_window:
4651                tcp_enter_quickack_mode(sk);
4652                inet_csk_schedule_ack(sk);
4653drop:
4654                __kfree_skb(skb);
4655                return;
4656        }
4657
4658        /* Out of window. F.e. zero window probe. */
4659        if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4660                goto out_of_window;
4661
4662        tcp_enter_quickack_mode(sk);
4663
4664        if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4665                /* Partial packet, seq < rcv_next < end_seq */
4666                SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4667                           tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4668                           TCP_SKB_CB(skb)->end_seq);
4669
4670                tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4671
4672                /* If window is closed, drop tail of packet. But after
4673                 * remembering D-SACK for its head made in previous line.
4674                 */
4675                if (!tcp_receive_window(tp))
4676                        goto out_of_window;
4677                goto queue_and_out;
4678        }
4679
4680        tcp_data_queue_ofo(sk, skb);
4681}
4682
4683static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4684                                        struct sk_buff_head *list)
4685{
4686        struct sk_buff *next = NULL;
4687
4688        if (!skb_queue_is_last(list, skb))
4689                next = skb_queue_next(list, skb);
4690
4691        __skb_unlink(skb, list);
4692        __kfree_skb(skb);
4693        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4694
4695        return next;
4696}
4697
4698/* Collapse contiguous sequence of skbs head..tail with
4699 * sequence numbers start..end.
4700 *
4701 * If tail is NULL, this means until the end of the list.
4702 *
4703 * Segments with FIN/SYN are not collapsed (only because this
4704 * simplifies code)
4705 */
4706static void
4707tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4708             struct sk_buff *head, struct sk_buff *tail,
4709             u32 start, u32 end)
4710{
4711        struct sk_buff *skb, *n;
4712        bool end_of_skbs;
4713
4714        /* First, check that queue is collapsible and find
4715         * the point where collapsing can be useful. */
4716        skb = head;
4717restart:
4718        end_of_skbs = true;
4719        skb_queue_walk_from_safe(list, skb, n) {
4720                if (skb == tail)
4721                        break;
4722                /* No new bits? It is possible on ofo queue. */
4723                if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4724                        skb = tcp_collapse_one(sk, skb, list);
4725                        if (!skb)
4726                                break;
4727                        goto restart;
4728                }
4729
4730                /* The first skb to collapse is:
4731                 * - not SYN/FIN and
4732                 * - bloated or contains data before "start" or
4733                 *   overlaps to the next one.
4734                 */
4735                if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4736                    (tcp_win_from_space(skb->truesize) > skb->len ||
4737                     before(TCP_SKB_CB(skb)->seq, start))) {
4738                        end_of_skbs = false;
4739                        break;
4740                }
4741
4742                if (!skb_queue_is_last(list, skb)) {
4743                        struct sk_buff *next = skb_queue_next(list, skb);
4744                        if (next != tail &&
4745                            TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4746                                end_of_skbs = false;
4747                                break;
4748                        }
4749                }
4750
4751                /* Decided to skip this, advance start seq. */
4752                start = TCP_SKB_CB(skb)->end_seq;
4753        }
4754        if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4755                return;
4756
4757        while (before(start, end)) {
4758                struct sk_buff *nskb;
4759                unsigned int header = skb_headroom(skb);
4760                int copy = SKB_MAX_ORDER(header, 0);
4761
4762                /* Too big header? This can happen with IPv6. */
4763                if (copy < 0)
4764                        return;
4765                if (end - start < copy)
4766                        copy = end - start;
4767                nskb = alloc_skb(copy + header, GFP_ATOMIC);
4768                if (!nskb)
4769                        return;
4770
4771                skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4772                skb_set_network_header(nskb, (skb_network_header(skb) -
4773                                              skb->head));
4774                skb_set_transport_header(nskb, (skb_transport_header(skb) -
4775                                                skb->head));
4776                skb_reserve(nskb, header);
4777                memcpy(nskb->head, skb->head, header);
4778                memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4779                TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4780                __skb_queue_before(list, skb, nskb);
4781                skb_set_owner_r(nskb, sk);
4782
4783                /* Copy data, releasing collapsed skbs. */
4784                while (copy > 0) {
4785                        int offset = start - TCP_SKB_CB(skb)->seq;
4786                        int size = TCP_SKB_CB(skb)->end_seq - start;
4787
4788                        BUG_ON(offset < 0);
4789                        if (size > 0) {
4790                                size = min(copy, size);
4791                                if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4792                                        BUG();
4793                                TCP_SKB_CB(nskb)->end_seq += size;
4794                                copy -= size;
4795                                start += size;
4796                        }
4797                        if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4798                                skb = tcp_collapse_one(sk, skb, list);
4799                                if (!skb ||
4800                                    skb == tail ||
4801                                    tcp_hdr(skb)->syn ||
4802                                    tcp_hdr(skb)->fin)
4803                                        return;
4804                        }
4805                }
4806        }
4807}
4808
4809/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4810 * and tcp_collapse() them until all the queue is collapsed.
4811 */
4812static void tcp_collapse_ofo_queue(struct sock *sk)
4813{
4814        struct tcp_sock *tp = tcp_sk(sk);
4815        struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4816        struct sk_buff *head;
4817        u32 start, end;
4818
4819        if (skb == NULL)
4820                return;
4821
4822        start = TCP_SKB_CB(skb)->seq;
4823        end = TCP_SKB_CB(skb)->end_seq;
4824        head = skb;
4825
4826        for (;;) {
4827                struct sk_buff *next = NULL;
4828
4829                if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4830                        next = skb_queue_next(&tp->out_of_order_queue, skb);
4831                skb = next;
4832
4833                /* Segment is terminated when we see gap or when
4834                 * we are at the end of all the queue. */
4835                if (!skb ||
4836                    after(TCP_SKB_CB(skb)->seq, end) ||
4837                    before(TCP_SKB_CB(skb)->end_seq, start)) {
4838                        tcp_collapse(sk, &tp->out_of_order_queue,
4839                                     head, skb, start, end);
4840                        head = skb;
4841                        if (!skb)
4842                                break;
4843                        /* Start new segment */
4844                        start = TCP_SKB_CB(skb)->seq;
4845                        end = TCP_SKB_CB(skb)->end_seq;
4846                } else {
4847                        if (before(TCP_SKB_CB(skb)->seq, start))
4848                                start = TCP_SKB_CB(skb)->seq;
4849                        if (after(TCP_SKB_CB(skb)->end_seq, end))
4850                                end = TCP_SKB_CB(skb)->end_seq;
4851                }
4852        }
4853}
4854
4855/*
4856 * Purge the out-of-order queue.
4857 * Return true if queue was pruned.
4858 */
4859static bool tcp_prune_ofo_queue(struct sock *sk)
4860{
4861        struct tcp_sock *tp = tcp_sk(sk);
4862        bool res = false;
4863
4864        if (!skb_queue_empty(&tp->out_of_order_queue)) {
4865                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4866                __skb_queue_purge(&tp->out_of_order_queue);
4867
4868                /* Reset SACK state.  A conforming SACK implementation will
4869                 * do the same at a timeout based retransmit.  When a connection
4870                 * is in a sad state like this, we care only about integrity
4871                 * of the connection not performance.
4872                 */
4873                if (tp->rx_opt.sack_ok)
4874                        tcp_sack_reset(&tp->rx_opt);
4875                sk_mem_reclaim(sk);
4876                res = true;
4877        }
4878        return res;
4879}
4880
4881/* Reduce allocated memory if we can, trying to get
4882 * the socket within its memory limits again.
4883 *
4884 * Return less than zero if we should start dropping frames
4885 * until the socket owning process reads some of the data
4886 * to stabilize the situation.
4887 */
4888static int tcp_prune_queue(struct sock *sk)
4889{
4890        struct tcp_sock *tp = tcp_sk(sk);
4891
4892        SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4893
4894        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4895
4896        if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4897                tcp_clamp_window(sk);
4898        else if (sk_under_memory_pressure(sk))
4899                tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4900
4901        tcp_collapse_ofo_queue(sk);
4902        if (!skb_queue_empty(&sk->sk_receive_queue))
4903                tcp_collapse(sk, &sk->sk_receive_queue,
4904                             skb_peek(&sk->sk_receive_queue),
4905                             NULL,
4906                             tp->copied_seq, tp->rcv_nxt);
4907        sk_mem_reclaim(sk);
4908
4909        if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4910                return 0;
4911
4912        /* Collapsing did not help, destructive actions follow.
4913         * This must not ever occur. */
4914
4915        tcp_prune_ofo_queue(sk);
4916
4917        if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4918                return 0;
4919
4920        /* If we are really being abused, tell the caller to silently
4921         * drop receive data on the floor.  It will get retransmitted
4922         * and hopefully then we'll have sufficient space.
4923         */
4924        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4925
4926        /* Massive buffer overcommit. */
4927        tp->pred_flags = 0;
4928        return -1;
4929}
4930
4931/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4932 * As additional protections, we do not touch cwnd in retransmission phases,
4933 * and if application hit its sndbuf limit recently.
4934 */
4935void tcp_cwnd_application_limited(struct sock *sk)
4936{
4937        struct tcp_sock *tp = tcp_sk(sk);
4938
4939        if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4940            sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4941                /* Limited by application or receiver window. */
4942                u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4943                u32 win_used = max(tp->snd_cwnd_used, init_win);
4944                if (win_used < tp->snd_cwnd) {
4945                        tp->snd_ssthresh = tcp_current_ssthresh(sk);
4946                        tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4947                }
4948                tp->snd_cwnd_used = 0;
4949        }
4950        tp->snd_cwnd_stamp = tcp_time_stamp;
4951}
4952
4953static bool tcp_should_expand_sndbuf(const struct sock *sk)
4954{
4955        const struct tcp_sock *tp = tcp_sk(sk);
4956
4957        /* If the user specified a specific send buffer setting, do
4958         * not modify it.
4959         */
4960        if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4961                return false;
4962
4963        /* If we are under global TCP memory pressure, do not expand.  */
4964        if (sk_under_memory_pressure(sk))
4965                return false;
4966
4967        /* If we are under soft global TCP memory pressure, do not expand.  */
4968        if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4969                return false;
4970
4971        /* If we filled the congestion window, do not expand.  */
4972        if (tp->packets_out >= tp->snd_cwnd)
4973                return false;
4974
4975        return true;
4976}
4977
4978/* When incoming ACK allowed to free some skb from write_queue,
4979 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4980 * on the exit from tcp input handler.
4981 *
4982 * PROBLEM: sndbuf expansion does not work well with largesend.
4983 */
4984static void tcp_new_space(struct sock *sk)
4985{
4986        struct tcp_sock *tp = tcp_sk(sk);
4987
4988        if (tcp_should_expand_sndbuf(sk)) {
4989                int sndmem = SKB_TRUESIZE(max_t(u32,
4990                                                tp->rx_opt.mss_clamp,
4991                                                tp->mss_cache) +
4992                                          MAX_TCP_HEADER);
4993                int demanded = max_t(unsigned int, tp->snd_cwnd,
4994                                     tp->reordering + 1);
4995                sndmem *= 2 * demanded;
4996                if (sndmem > sk->sk_sndbuf)
4997                        sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4998                tp->snd_cwnd_stamp = tcp_time_stamp;
4999        }
5000
5001        sk->sk_write_space(sk);
5002}
5003
5004static void tcp_check_space(struct sock *sk)
5005{
5006        if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5007                sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5008                if (sk->sk_socket &&
5009                    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5010                        tcp_new_space(sk);
5011        }
5012}
5013
5014static inline void tcp_data_snd_check(struct sock *sk)
5015{
5016        tcp_push_pending_frames(sk);
5017        tcp_check_space(sk);
5018}
5019
5020/*
5021 * Check if sending an ack is needed.
5022 */
5023static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5024{
5025        struct tcp_sock *tp = tcp_sk(sk);
5026
5027            /* More than one full frame received... */
5028        if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5029             /* ... and right edge of window advances far enough.
5030              * (tcp_recvmsg() will send ACK otherwise). Or...
5031              */
5032             __tcp_select_window(sk) >= tp->rcv_wnd) ||
5033            /* We ACK each frame or... */
5034            tcp_in_quickack_mode(sk) ||
5035            /* We have out of order data. */
5036            (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5037                /* Then ack it now */
5038                tcp_send_ack(sk);
5039        } else {
5040                /* Else, send delayed ack. */
5041                tcp_send_delayed_ack(sk);
5042        }
5043}
5044
5045static inline void tcp_ack_snd_check(struct sock *sk)
5046{
5047        if (!inet_csk_ack_scheduled(sk)) {
5048                /* We sent a data segment already. */
5049                return;
5050        }
5051        __tcp_ack_snd_check(sk, 1);
5052}
5053
5054/*
5055 *      This routine is only called when we have urgent data
5056 *      signaled. Its the 'slow' part of tcp_urg. It could be
5057 *      moved inline now as tcp_urg is only called from one
5058 *      place. We handle URGent data wrong. We have to - as
5059 *      BSD still doesn't use the correction from RFC961.
5060 *      For 1003.1g we should support a new option TCP_STDURG to permit
5061 *      either form (or just set the sysctl tcp_stdurg).
5062 */
5063
5064static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5065{
5066        struct tcp_sock *tp = tcp_sk(sk);
5067        u32 ptr = ntohs(th->urg_ptr);
5068
5069        if (ptr && !sysctl_tcp_stdurg)
5070                ptr--;
5071        ptr += ntohl(th->seq);
5072
5073        /* Ignore urgent data that we've already seen and read. */
5074        if (after(tp->copied_seq, ptr))
5075                return;
5076
5077        /* Do not replay urg ptr.
5078         *
5079         * NOTE: interesting situation not covered by specs.
5080         * Misbehaving sender may send urg ptr, pointing to segment,
5081         * which we already have in ofo queue. We are not able to fetch
5082         * such data and will stay in TCP_URG_NOTYET until will be eaten
5083         * by recvmsg(). Seems, we are not obliged to handle such wicked
5084         * situations. But it is worth to think about possibility of some
5085         * DoSes using some hypothetical application level deadlock.
5086         */
5087        if (before(ptr, tp->rcv_nxt))
5088                return;
5089
5090        /* Do we already have a newer (or duplicate) urgent pointer? */
5091        if (tp->urg_data && !after(ptr, tp->urg_seq))
5092                return;
5093
5094        /* Tell the world about our new urgent pointer. */
5095        sk_send_sigurg(sk);
5096
5097        /* We may be adding urgent data when the last byte read was
5098         * urgent. To do this requires some care. We cannot just ignore
5099         * tp->copied_seq since we would read the last urgent byte again
5100         * as data, nor can we alter copied_seq until this data arrives
5101         * or we break the semantics of SIOCATMARK (and thus sockatmark())
5102         *
5103         * NOTE. Double Dutch. Rendering to plain English: author of comment
5104         * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
5105         * and expect that both A and B disappear from stream. This is _wrong_.
5106         * Though this happens in BSD with high probability, this is occasional.
5107         * Any application relying on this is buggy. Note also, that fix "works"
5108         * only in this artificial test. Insert some normal data between A and B and we will
5109         * decline of BSD again. Verdict: it is better to remove to trap
5110         * buggy users.
5111         */
5112        if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5113            !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5114                struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5115                tp->copied_seq++;
5116                if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5117                        __skb_unlink(skb, &sk->sk_receive_queue);
5118                        __kfree_skb(skb);
5119                }
5120        }
5121
5122        tp->urg_data = TCP_URG_NOTYET;
5123        tp->urg_seq = ptr;
5124
5125        /* Disable header prediction. */
5126        tp->pred_flags = 0;
5127}
5128
5129/* This is the 'fast' part of urgent handling. */
5130static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5131{
5132        struct tcp_sock *tp = tcp_sk(sk);
5133
5134        /* Check if we get a new urgent pointer - normally not. */
5135        if (th->urg)
5136                tcp_check_urg(sk, th);
5137
5138        /* Do we wait for any urgent data? - normally not... */
5139        if (tp->urg_data == TCP_URG_NOTYET) {
5140                u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5141                          th->syn;
5142
5143                /* Is the urgent pointer pointing into this packet? */
5144                if (ptr < skb->len) {
5145                        u8 tmp;
5146                        if (skb_copy_bits(skb, ptr, &tmp, 1))
5147                                BUG();
5148                        tp->urg_data = TCP_URG_VALID | tmp;
5149                        if (!sock_flag(sk, SOCK_DEAD))
5150                                sk->sk_data_ready(sk, 0);
5151                }
5152        }
5153}
5154
5155static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5156{
5157        struct tcp_sock *tp = tcp_sk(sk);
5158        int chunk = skb->len - hlen;
5159        int err;
5160
5161        local_bh_enable();
5162        if (skb_csum_unnecessary(skb))
5163                err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5164        else
5165                err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5166                                                       tp->ucopy.iov);
5167
5168        if (!err) {
5169                tp->ucopy.len -= chunk;
5170                tp->copied_seq += chunk;
5171                tcp_rcv_space_adjust(sk);
5172        }
5173
5174        local_bh_disable();
5175        return err;
5176}
5177
5178static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5179                                            struct sk_buff *skb)
5180{
5181        __sum16 result;
5182
5183        if (sock_owned_by_user(sk)) {
5184                local_bh_enable();
5185                result = __tcp_checksum_complete(skb);
5186                local_bh_disable();
5187        } else {
5188                result = __tcp_checksum_complete(skb);
5189        }
5190        return result;
5191}
5192
5193static