linux/net/ipv4/tcp_output.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:     Pedro Roque     :       Retransmit queue handled by TCP.
  23 *                              :       Fragmentation on mtu decrease
  24 *                              :       Segment collapse on retransmit
  25 *                              :       AF independence
  26 *
  27 *              Linus Torvalds  :       send_delayed_ack
  28 *              David S. Miller :       Charge memory using the right skb
  29 *                                      during syn/ack processing.
  30 *              David S. Miller :       Output engine completely rewritten.
  31 *              Andrea Arcangeli:       SYNACK carry ts_recent in tsecr.
  32 *              Cacophonix Gaul :       draft-minshall-nagle-01
  33 *              J Hadi Salim    :       ECN support
  34 *
  35 */
  36
  37#define pr_fmt(fmt) "TCP: " fmt
  38
  39#include <net/tcp.h>
  40
  41#include <linux/compiler.h>
  42#include <linux/gfp.h>
  43#include <linux/module.h>
  44
  45/* People can turn this off for buggy TCP's found in printers etc. */
  46int sysctl_tcp_retrans_collapse __read_mostly = 1;
  47
  48/* People can turn this on to work with those rare, broken TCPs that
  49 * interpret the window field as a signed quantity.
  50 */
  51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  52
  53/* Default TSQ limit of two TSO segments */
  54int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
  55
  56/* This limits the percentage of the congestion window which we
  57 * will allow a single TSO frame to consume.  Building TSO frames
  58 * which are too large can cause TCP streams to be bursty.
  59 */
  60int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  61
  62int sysctl_tcp_mtu_probing __read_mostly = 0;
  63int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
  64
  65/* By default, RFC2861 behavior.  */
  66int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  67
  68int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  69EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  70
  71static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  72                           int push_one, gfp_t gfp);
  73
  74/* Account for new data that has been sent to the network. */
  75static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
  76{
  77        struct tcp_sock *tp = tcp_sk(sk);
  78        unsigned int prior_packets = tp->packets_out;
  79
  80        tcp_advance_send_head(sk, skb);
  81        tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  82
  83        /* Don't override Nagle indefinitely with F-RTO */
  84        if (tp->frto_counter == 2)
  85                tp->frto_counter = 3;
  86
  87        tp->packets_out += tcp_skb_pcount(skb);
  88        if (!prior_packets || tp->early_retrans_delayed)
  89                tcp_rearm_rto(sk);
  90}
  91
  92/* SND.NXT, if window was not shrunk.
  93 * If window has been shrunk, what should we make? It is not clear at all.
  94 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  95 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  96 * invalid. OK, let's make this for now:
  97 */
  98static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  99{
 100        const struct tcp_sock *tp = tcp_sk(sk);
 101
 102        if (!before(tcp_wnd_end(tp), tp->snd_nxt))
 103                return tp->snd_nxt;
 104        else
 105                return tcp_wnd_end(tp);
 106}
 107
 108/* Calculate mss to advertise in SYN segment.
 109 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 110 *
 111 * 1. It is independent of path mtu.
 112 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 113 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 114 *    attached devices, because some buggy hosts are confused by
 115 *    large MSS.
 116 * 4. We do not make 3, we advertise MSS, calculated from first
 117 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 118 *    This may be overridden via information stored in routing table.
 119 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 120 *    probably even Jumbo".
 121 */
 122static __u16 tcp_advertise_mss(struct sock *sk)
 123{
 124        struct tcp_sock *tp = tcp_sk(sk);
 125        const struct dst_entry *dst = __sk_dst_get(sk);
 126        int mss = tp->advmss;
 127
 128        if (dst) {
 129                unsigned int metric = dst_metric_advmss(dst);
 130
 131                if (metric < mss) {
 132                        mss = metric;
 133                        tp->advmss = mss;
 134                }
 135        }
 136
 137        return (__u16)mss;
 138}
 139
 140/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 141 * This is the first part of cwnd validation mechanism. */
 142static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
 143{
 144        struct tcp_sock *tp = tcp_sk(sk);
 145        s32 delta = tcp_time_stamp - tp->lsndtime;
 146        u32 restart_cwnd = tcp_init_cwnd(tp, dst);
 147        u32 cwnd = tp->snd_cwnd;
 148
 149        tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 150
 151        tp->snd_ssthresh = tcp_current_ssthresh(sk);
 152        restart_cwnd = min(restart_cwnd, cwnd);
 153
 154        while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 155                cwnd >>= 1;
 156        tp->snd_cwnd = max(cwnd, restart_cwnd);
 157        tp->snd_cwnd_stamp = tcp_time_stamp;
 158        tp->snd_cwnd_used = 0;
 159}
 160
 161/* Congestion state accounting after a packet has been sent. */
 162static void tcp_event_data_sent(struct tcp_sock *tp,
 163                                struct sock *sk)
 164{
 165        struct inet_connection_sock *icsk = inet_csk(sk);
 166        const u32 now = tcp_time_stamp;
 167
 168        if (sysctl_tcp_slow_start_after_idle &&
 169            (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
 170                tcp_cwnd_restart(sk, __sk_dst_get(sk));
 171
 172        tp->lsndtime = now;
 173
 174        /* If it is a reply for ato after last received
 175         * packet, enter pingpong mode.
 176         */
 177        if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 178                icsk->icsk_ack.pingpong = 1;
 179}
 180
 181/* Account for an ACK we sent. */
 182static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
 183{
 184        tcp_dec_quickack_mode(sk, pkts);
 185        inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 186}
 187
 188/* Determine a window scaling and initial window to offer.
 189 * Based on the assumption that the given amount of space
 190 * will be offered. Store the results in the tp structure.
 191 * NOTE: for smooth operation initial space offering should
 192 * be a multiple of mss if possible. We assume here that mss >= 1.
 193 * This MUST be enforced by all callers.
 194 */
 195void tcp_select_initial_window(int __space, __u32 mss,
 196                               __u32 *rcv_wnd, __u32 *window_clamp,
 197                               int wscale_ok, __u8 *rcv_wscale,
 198                               __u32 init_rcv_wnd)
 199{
 200        unsigned int space = (__space < 0 ? 0 : __space);
 201
 202        /* If no clamp set the clamp to the max possible scaled window */
 203        if (*window_clamp == 0)
 204                (*window_clamp) = (65535 << 14);
 205        space = min(*window_clamp, space);
 206
 207        /* Quantize space offering to a multiple of mss if possible. */
 208        if (space > mss)
 209                space = (space / mss) * mss;
 210
 211        /* NOTE: offering an initial window larger than 32767
 212         * will break some buggy TCP stacks. If the admin tells us
 213         * it is likely we could be speaking with such a buggy stack
 214         * we will truncate our initial window offering to 32K-1
 215         * unless the remote has sent us a window scaling option,
 216         * which we interpret as a sign the remote TCP is not
 217         * misinterpreting the window field as a signed quantity.
 218         */
 219        if (sysctl_tcp_workaround_signed_windows)
 220                (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 221        else
 222                (*rcv_wnd) = space;
 223
 224        (*rcv_wscale) = 0;
 225        if (wscale_ok) {
 226                /* Set window scaling on max possible window
 227                 * See RFC1323 for an explanation of the limit to 14
 228                 */
 229                space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
 230                space = min_t(u32, space, *window_clamp);
 231                while (space > 65535 && (*rcv_wscale) < 14) {
 232                        space >>= 1;
 233                        (*rcv_wscale)++;
 234                }
 235        }
 236
 237        /* Set initial window to a value enough for senders starting with
 238         * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
 239         * a limit on the initial window when mss is larger than 1460.
 240         */
 241        if (mss > (1 << *rcv_wscale)) {
 242                int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
 243                if (mss > 1460)
 244                        init_cwnd =
 245                        max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
 246                /* when initializing use the value from init_rcv_wnd
 247                 * rather than the default from above
 248                 */
 249                if (init_rcv_wnd)
 250                        *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 251                else
 252                        *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
 253        }
 254
 255        /* Set the clamp no higher than max representable value */
 256        (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
 257}
 258EXPORT_SYMBOL(tcp_select_initial_window);
 259
 260/* Chose a new window to advertise, update state in tcp_sock for the
 261 * socket, and return result with RFC1323 scaling applied.  The return
 262 * value can be stuffed directly into th->window for an outgoing
 263 * frame.
 264 */
 265static u16 tcp_select_window(struct sock *sk)
 266{
 267        struct tcp_sock *tp = tcp_sk(sk);
 268        u32 cur_win = tcp_receive_window(tp);
 269        u32 new_win = __tcp_select_window(sk);
 270
 271        /* Never shrink the offered window */
 272        if (new_win < cur_win) {
 273                /* Danger Will Robinson!
 274                 * Don't update rcv_wup/rcv_wnd here or else
 275                 * we will not be able to advertise a zero
 276                 * window in time.  --DaveM
 277                 *
 278                 * Relax Will Robinson.
 279                 */
 280                new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 281        }
 282        tp->rcv_wnd = new_win;
 283        tp->rcv_wup = tp->rcv_nxt;
 284
 285        /* Make sure we do not exceed the maximum possible
 286         * scaled window.
 287         */
 288        if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
 289                new_win = min(new_win, MAX_TCP_WINDOW);
 290        else
 291                new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 292
 293        /* RFC1323 scaling applied */
 294        new_win >>= tp->rx_opt.rcv_wscale;
 295
 296        /* If we advertise zero window, disable fast path. */
 297        if (new_win == 0)
 298                tp->pred_flags = 0;
 299
 300        return new_win;
 301}
 302
 303/* Packet ECN state for a SYN-ACK */
 304static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
 305{
 306        TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 307        if (!(tp->ecn_flags & TCP_ECN_OK))
 308                TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 309}
 310
 311/* Packet ECN state for a SYN.  */
 312static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
 313{
 314        struct tcp_sock *tp = tcp_sk(sk);
 315
 316        tp->ecn_flags = 0;
 317        if (sysctl_tcp_ecn == 1) {
 318                TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 319                tp->ecn_flags = TCP_ECN_OK;
 320        }
 321}
 322
 323static __inline__ void
 324TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
 325{
 326        if (inet_rsk(req)->ecn_ok)
 327                th->ece = 1;
 328}
 329
 330/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 331 * be sent.
 332 */
 333static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
 334                                int tcp_header_len)
 335{
 336        struct tcp_sock *tp = tcp_sk(sk);
 337
 338        if (tp->ecn_flags & TCP_ECN_OK) {
 339                /* Not-retransmitted data segment: set ECT and inject CWR. */
 340                if (skb->len != tcp_header_len &&
 341                    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 342                        INET_ECN_xmit(sk);
 343                        if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 344                                tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 345                                tcp_hdr(skb)->cwr = 1;
 346                                skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 347                        }
 348                } else {
 349                        /* ACK or retransmitted segment: clear ECT|CE */
 350                        INET_ECN_dontxmit(sk);
 351                }
 352                if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 353                        tcp_hdr(skb)->ece = 1;
 354        }
 355}
 356
 357/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 358 * auto increment end seqno.
 359 */
 360static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 361{
 362        skb->ip_summed = CHECKSUM_PARTIAL;
 363        skb->csum = 0;
 364
 365        TCP_SKB_CB(skb)->tcp_flags = flags;
 366        TCP_SKB_CB(skb)->sacked = 0;
 367
 368        skb_shinfo(skb)->gso_segs = 1;
 369        skb_shinfo(skb)->gso_size = 0;
 370        skb_shinfo(skb)->gso_type = 0;
 371
 372        TCP_SKB_CB(skb)->seq = seq;
 373        if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 374                seq++;
 375        TCP_SKB_CB(skb)->end_seq = seq;
 376}
 377
 378static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 379{
 380        return tp->snd_una != tp->snd_up;
 381}
 382
 383#define OPTION_SACK_ADVERTISE   (1 << 0)
 384#define OPTION_TS               (1 << 1)
 385#define OPTION_MD5              (1 << 2)
 386#define OPTION_WSCALE           (1 << 3)
 387#define OPTION_COOKIE_EXTENSION (1 << 4)
 388#define OPTION_FAST_OPEN_COOKIE (1 << 8)
 389
 390struct tcp_out_options {
 391        u16 options;            /* bit field of OPTION_* */
 392        u16 mss;                /* 0 to disable */
 393        u8 ws;                  /* window scale, 0 to disable */
 394        u8 num_sack_blocks;     /* number of SACK blocks to include */
 395        u8 hash_size;           /* bytes in hash_location */
 396        __u8 *hash_location;    /* temporary pointer, overloaded */
 397        __u32 tsval, tsecr;     /* need to include OPTION_TS */
 398        struct tcp_fastopen_cookie *fastopen_cookie;    /* Fast open cookie */
 399};
 400
 401/* The sysctl int routines are generic, so check consistency here.
 402 */
 403static u8 tcp_cookie_size_check(u8 desired)
 404{
 405        int cookie_size;
 406
 407        if (desired > 0)
 408                /* previously specified */
 409                return desired;
 410
 411        cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
 412        if (cookie_size <= 0)
 413                /* no default specified */
 414                return 0;
 415
 416        if (cookie_size <= TCP_COOKIE_MIN)
 417                /* value too small, specify minimum */
 418                return TCP_COOKIE_MIN;
 419
 420        if (cookie_size >= TCP_COOKIE_MAX)
 421                /* value too large, specify maximum */
 422                return TCP_COOKIE_MAX;
 423
 424        if (cookie_size & 1)
 425                /* 8-bit multiple, illegal, fix it */
 426                cookie_size++;
 427
 428        return (u8)cookie_size;
 429}
 430
 431/* Write previously computed TCP options to the packet.
 432 *
 433 * Beware: Something in the Internet is very sensitive to the ordering of
 434 * TCP options, we learned this through the hard way, so be careful here.
 435 * Luckily we can at least blame others for their non-compliance but from
 436 * inter-operatibility perspective it seems that we're somewhat stuck with
 437 * the ordering which we have been using if we want to keep working with
 438 * those broken things (not that it currently hurts anybody as there isn't
 439 * particular reason why the ordering would need to be changed).
 440 *
 441 * At least SACK_PERM as the first option is known to lead to a disaster
 442 * (but it may well be that other scenarios fail similarly).
 443 */
 444static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 445                              struct tcp_out_options *opts)
 446{
 447        u16 options = opts->options;    /* mungable copy */
 448
 449        /* Having both authentication and cookies for security is redundant,
 450         * and there's certainly not enough room.  Instead, the cookie-less
 451         * extension variant is proposed.
 452         *
 453         * Consider the pessimal case with authentication.  The options
 454         * could look like:
 455         *   COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
 456         */
 457        if (unlikely(OPTION_MD5 & options)) {
 458                if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 459                        *ptr++ = htonl((TCPOPT_COOKIE << 24) |
 460                                       (TCPOLEN_COOKIE_BASE << 16) |
 461                                       (TCPOPT_MD5SIG << 8) |
 462                                       TCPOLEN_MD5SIG);
 463                } else {
 464                        *ptr++ = htonl((TCPOPT_NOP << 24) |
 465                                       (TCPOPT_NOP << 16) |
 466                                       (TCPOPT_MD5SIG << 8) |
 467                                       TCPOLEN_MD5SIG);
 468                }
 469                options &= ~OPTION_COOKIE_EXTENSION;
 470                /* overload cookie hash location */
 471                opts->hash_location = (__u8 *)ptr;
 472                ptr += 4;
 473        }
 474
 475        if (unlikely(opts->mss)) {
 476                *ptr++ = htonl((TCPOPT_MSS << 24) |
 477                               (TCPOLEN_MSS << 16) |
 478                               opts->mss);
 479        }
 480
 481        if (likely(OPTION_TS & options)) {
 482                if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 483                        *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 484                                       (TCPOLEN_SACK_PERM << 16) |
 485                                       (TCPOPT_TIMESTAMP << 8) |
 486                                       TCPOLEN_TIMESTAMP);
 487                        options &= ~OPTION_SACK_ADVERTISE;
 488                } else {
 489                        *ptr++ = htonl((TCPOPT_NOP << 24) |
 490                                       (TCPOPT_NOP << 16) |
 491                                       (TCPOPT_TIMESTAMP << 8) |
 492                                       TCPOLEN_TIMESTAMP);
 493                }
 494                *ptr++ = htonl(opts->tsval);
 495                *ptr++ = htonl(opts->tsecr);
 496        }
 497
 498        /* Specification requires after timestamp, so do it now.
 499         *
 500         * Consider the pessimal case without authentication.  The options
 501         * could look like:
 502         *   MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
 503         */
 504        if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 505                __u8 *cookie_copy = opts->hash_location;
 506                u8 cookie_size = opts->hash_size;
 507
 508                /* 8-bit multiple handled in tcp_cookie_size_check() above,
 509                 * and elsewhere.
 510                 */
 511                if (0x2 & cookie_size) {
 512                        __u8 *p = (__u8 *)ptr;
 513
 514                        /* 16-bit multiple */
 515                        *p++ = TCPOPT_COOKIE;
 516                        *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
 517                        *p++ = *cookie_copy++;
 518                        *p++ = *cookie_copy++;
 519                        ptr++;
 520                        cookie_size -= 2;
 521                } else {
 522                        /* 32-bit multiple */
 523                        *ptr++ = htonl(((TCPOPT_NOP << 24) |
 524                                        (TCPOPT_NOP << 16) |
 525                                        (TCPOPT_COOKIE << 8) |
 526                                        TCPOLEN_COOKIE_BASE) +
 527                                       cookie_size);
 528                }
 529
 530                if (cookie_size > 0) {
 531                        memcpy(ptr, cookie_copy, cookie_size);
 532                        ptr += (cookie_size / 4);
 533                }
 534        }
 535
 536        if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 537                *ptr++ = htonl((TCPOPT_NOP << 24) |
 538                               (TCPOPT_NOP << 16) |
 539                               (TCPOPT_SACK_PERM << 8) |
 540                               TCPOLEN_SACK_PERM);
 541        }
 542
 543        if (unlikely(OPTION_WSCALE & options)) {
 544                *ptr++ = htonl((TCPOPT_NOP << 24) |
 545                               (TCPOPT_WINDOW << 16) |
 546                               (TCPOLEN_WINDOW << 8) |
 547                               opts->ws);
 548        }
 549
 550        if (unlikely(opts->num_sack_blocks)) {
 551                struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 552                        tp->duplicate_sack : tp->selective_acks;
 553                int this_sack;
 554
 555                *ptr++ = htonl((TCPOPT_NOP  << 24) |
 556                               (TCPOPT_NOP  << 16) |
 557                               (TCPOPT_SACK <<  8) |
 558                               (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 559                                                     TCPOLEN_SACK_PERBLOCK)));
 560
 561                for (this_sack = 0; this_sack < opts->num_sack_blocks;
 562                     ++this_sack) {
 563                        *ptr++ = htonl(sp[this_sack].start_seq);
 564                        *ptr++ = htonl(sp[this_sack].end_seq);
 565                }
 566
 567                tp->rx_opt.dsack = 0;
 568        }
 569
 570        if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 571                struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 572
 573                *ptr++ = htonl((TCPOPT_EXP << 24) |
 574                               ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
 575                               TCPOPT_FASTOPEN_MAGIC);
 576
 577                memcpy(ptr, foc->val, foc->len);
 578                if ((foc->len & 3) == 2) {
 579                        u8 *align = ((u8 *)ptr) + foc->len;
 580                        align[0] = align[1] = TCPOPT_NOP;
 581                }
 582                ptr += (foc->len + 3) >> 2;
 583        }
 584}
 585
 586/* Compute TCP options for SYN packets. This is not the final
 587 * network wire format yet.
 588 */
 589static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 590                                struct tcp_out_options *opts,
 591                                struct tcp_md5sig_key **md5)
 592{
 593        struct tcp_sock *tp = tcp_sk(sk);
 594        struct tcp_cookie_values *cvp = tp->cookie_values;
 595        unsigned int remaining = MAX_TCP_OPTION_SPACE;
 596        u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
 597                         tcp_cookie_size_check(cvp->cookie_desired) :
 598                         0;
 599        struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 600
 601#ifdef CONFIG_TCP_MD5SIG
 602        *md5 = tp->af_specific->md5_lookup(sk, sk);
 603        if (*md5) {
 604                opts->options |= OPTION_MD5;
 605                remaining -= TCPOLEN_MD5SIG_ALIGNED;
 606        }
 607#else
 608        *md5 = NULL;
 609#endif
 610
 611        /* We always get an MSS option.  The option bytes which will be seen in
 612         * normal data packets should timestamps be used, must be in the MSS
 613         * advertised.  But we subtract them from tp->mss_cache so that
 614         * calculations in tcp_sendmsg are simpler etc.  So account for this
 615         * fact here if necessary.  If we don't do this correctly, as a
 616         * receiver we won't recognize data packets as being full sized when we
 617         * should, and thus we won't abide by the delayed ACK rules correctly.
 618         * SACKs don't matter, we never delay an ACK when we have any of those
 619         * going out.  */
 620        opts->mss = tcp_advertise_mss(sk);
 621        remaining -= TCPOLEN_MSS_ALIGNED;
 622
 623        if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
 624                opts->options |= OPTION_TS;
 625                opts->tsval = TCP_SKB_CB(skb)->when;
 626                opts->tsecr = tp->rx_opt.ts_recent;
 627                remaining -= TCPOLEN_TSTAMP_ALIGNED;
 628        }
 629        if (likely(sysctl_tcp_window_scaling)) {
 630                opts->ws = tp->rx_opt.rcv_wscale;
 631                opts->options |= OPTION_WSCALE;
 632                remaining -= TCPOLEN_WSCALE_ALIGNED;
 633        }
 634        if (likely(sysctl_tcp_sack)) {
 635                opts->options |= OPTION_SACK_ADVERTISE;
 636                if (unlikely(!(OPTION_TS & opts->options)))
 637                        remaining -= TCPOLEN_SACKPERM_ALIGNED;
 638        }
 639
 640        if (fastopen && fastopen->cookie.len >= 0) {
 641                u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
 642                need = (need + 3) & ~3U;  /* Align to 32 bits */
 643                if (remaining >= need) {
 644                        opts->options |= OPTION_FAST_OPEN_COOKIE;
 645                        opts->fastopen_cookie = &fastopen->cookie;
 646                        remaining -= need;
 647                        tp->syn_fastopen = 1;
 648                }
 649        }
 650        /* Note that timestamps are required by the specification.
 651         *
 652         * Odd numbers of bytes are prohibited by the specification, ensuring
 653         * that the cookie is 16-bit aligned, and the resulting cookie pair is
 654         * 32-bit aligned.
 655         */
 656        if (*md5 == NULL &&
 657            (OPTION_TS & opts->options) &&
 658            cookie_size > 0) {
 659                int need = TCPOLEN_COOKIE_BASE + cookie_size;
 660
 661                if (0x2 & need) {
 662                        /* 32-bit multiple */
 663                        need += 2; /* NOPs */
 664
 665                        if (need > remaining) {
 666                                /* try shrinking cookie to fit */
 667                                cookie_size -= 2;
 668                                need -= 4;
 669                        }
 670                }
 671                while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
 672                        cookie_size -= 4;
 673                        need -= 4;
 674                }
 675                if (TCP_COOKIE_MIN <= cookie_size) {
 676                        opts->options |= OPTION_COOKIE_EXTENSION;
 677                        opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
 678                        opts->hash_size = cookie_size;
 679
 680                        /* Remember for future incarnations. */
 681                        cvp->cookie_desired = cookie_size;
 682
 683                        if (cvp->cookie_desired != cvp->cookie_pair_size) {
 684                                /* Currently use random bytes as a nonce,
 685                                 * assuming these are completely unpredictable
 686                                 * by hostile users of the same system.
 687                                 */
 688                                get_random_bytes(&cvp->cookie_pair[0],
 689                                                 cookie_size);
 690                                cvp->cookie_pair_size = cookie_size;
 691                        }
 692
 693                        remaining -= need;
 694                }
 695        }
 696        return MAX_TCP_OPTION_SPACE - remaining;
 697}
 698
 699/* Set up TCP options for SYN-ACKs. */
 700static unsigned int tcp_synack_options(struct sock *sk,
 701                                   struct request_sock *req,
 702                                   unsigned int mss, struct sk_buff *skb,
 703                                   struct tcp_out_options *opts,
 704                                   struct tcp_md5sig_key **md5,
 705                                   struct tcp_extend_values *xvp,
 706                                   struct tcp_fastopen_cookie *foc)
 707{
 708        struct inet_request_sock *ireq = inet_rsk(req);
 709        unsigned int remaining = MAX_TCP_OPTION_SPACE;
 710        u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
 711                         xvp->cookie_plus :
 712                         0;
 713
 714#ifdef CONFIG_TCP_MD5SIG
 715        *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
 716        if (*md5) {
 717                opts->options |= OPTION_MD5;
 718                remaining -= TCPOLEN_MD5SIG_ALIGNED;
 719
 720                /* We can't fit any SACK blocks in a packet with MD5 + TS
 721                 * options. There was discussion about disabling SACK
 722                 * rather than TS in order to fit in better with old,
 723                 * buggy kernels, but that was deemed to be unnecessary.
 724                 */
 725                ireq->tstamp_ok &= !ireq->sack_ok;
 726        }
 727#else
 728        *md5 = NULL;
 729#endif
 730
 731        /* We always send an MSS option. */
 732        opts->mss = mss;
 733        remaining -= TCPOLEN_MSS_ALIGNED;
 734
 735        if (likely(ireq->wscale_ok)) {
 736                opts->ws = ireq->rcv_wscale;
 737                opts->options |= OPTION_WSCALE;
 738                remaining -= TCPOLEN_WSCALE_ALIGNED;
 739        }
 740        if (likely(ireq->tstamp_ok)) {
 741                opts->options |= OPTION_TS;
 742                opts->tsval = TCP_SKB_CB(skb)->when;
 743                opts->tsecr = req->ts_recent;
 744                remaining -= TCPOLEN_TSTAMP_ALIGNED;
 745        }
 746        if (likely(ireq->sack_ok)) {
 747                opts->options |= OPTION_SACK_ADVERTISE;
 748                if (unlikely(!ireq->tstamp_ok))
 749                        remaining -= TCPOLEN_SACKPERM_ALIGNED;
 750        }
 751        if (foc != NULL) {
 752                u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 753                need = (need + 3) & ~3U;  /* Align to 32 bits */
 754                if (remaining >= need) {
 755                        opts->options |= OPTION_FAST_OPEN_COOKIE;
 756                        opts->fastopen_cookie = foc;
 757                        remaining -= need;
 758                }
 759        }
 760        /* Similar rationale to tcp_syn_options() applies here, too.
 761         * If the <SYN> options fit, the same options should fit now!
 762         */
 763        if (*md5 == NULL &&
 764            ireq->tstamp_ok &&
 765            cookie_plus > TCPOLEN_COOKIE_BASE) {
 766                int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
 767
 768                if (0x2 & need) {
 769                        /* 32-bit multiple */
 770                        need += 2; /* NOPs */
 771                }
 772                if (need <= remaining) {
 773                        opts->options |= OPTION_COOKIE_EXTENSION;
 774                        opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
 775                        remaining -= need;
 776                } else {
 777                        /* There's no error return, so flag it. */
 778                        xvp->cookie_out_never = 1; /* true */
 779                        opts->hash_size = 0;
 780                }
 781        }
 782        return MAX_TCP_OPTION_SPACE - remaining;
 783}
 784
 785/* Compute TCP options for ESTABLISHED sockets. This is not the
 786 * final wire format yet.
 787 */
 788static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 789                                        struct tcp_out_options *opts,
 790                                        struct tcp_md5sig_key **md5)
 791{
 792        struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
 793        struct tcp_sock *tp = tcp_sk(sk);
 794        unsigned int size = 0;
 795        unsigned int eff_sacks;
 796
 797#ifdef CONFIG_TCP_MD5SIG
 798        *md5 = tp->af_specific->md5_lookup(sk, sk);
 799        if (unlikely(*md5)) {
 800                opts->options |= OPTION_MD5;
 801                size += TCPOLEN_MD5SIG_ALIGNED;
 802        }
 803#else
 804        *md5 = NULL;
 805#endif
 806
 807        if (likely(tp->rx_opt.tstamp_ok)) {
 808                opts->options |= OPTION_TS;
 809                opts->tsval = tcb ? tcb->when : 0;
 810                opts->tsecr = tp->rx_opt.ts_recent;
 811                size += TCPOLEN_TSTAMP_ALIGNED;
 812        }
 813
 814        eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 815        if (unlikely(eff_sacks)) {
 816                const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 817                opts->num_sack_blocks =
 818                        min_t(unsigned int, eff_sacks,
 819                              (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 820                              TCPOLEN_SACK_PERBLOCK);
 821                size += TCPOLEN_SACK_BASE_ALIGNED +
 822                        opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 823        }
 824
 825        return size;
 826}
 827
 828
 829/* TCP SMALL QUEUES (TSQ)
 830 *
 831 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 832 * to reduce RTT and bufferbloat.
 833 * We do this using a special skb destructor (tcp_wfree).
 834 *
 835 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 836 * needs to be reallocated in a driver.
 837 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
 838 *
 839 * Since transmit from skb destructor is forbidden, we use a tasklet
 840 * to process all sockets that eventually need to send more skbs.
 841 * We use one tasklet per cpu, with its own queue of sockets.
 842 */
 843struct tsq_tasklet {
 844        struct tasklet_struct   tasklet;
 845        struct list_head        head; /* queue of tcp sockets */
 846};
 847static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
 848
 849static void tcp_tsq_handler(struct sock *sk)
 850{
 851        if ((1 << sk->sk_state) &
 852            (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
 853             TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
 854                tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
 855}
 856/*
 857 * One tasklest per cpu tries to send more skbs.
 858 * We run in tasklet context but need to disable irqs when
 859 * transfering tsq->head because tcp_wfree() might
 860 * interrupt us (non NAPI drivers)
 861 */
 862static void tcp_tasklet_func(unsigned long data)
 863{
 864        struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
 865        LIST_HEAD(list);
 866        unsigned long flags;
 867        struct list_head *q, *n;
 868        struct tcp_sock *tp;
 869        struct sock *sk;
 870
 871        local_irq_save(flags);
 872        list_splice_init(&tsq->head, &list);
 873        local_irq_restore(flags);
 874
 875        list_for_each_safe(q, n, &list) {
 876                tp = list_entry(q, struct tcp_sock, tsq_node);
 877                list_del(&tp->tsq_node);
 878
 879                sk = (struct sock *)tp;
 880                bh_lock_sock(sk);
 881
 882                if (!sock_owned_by_user(sk)) {
 883                        tcp_tsq_handler(sk);
 884                } else {
 885                        /* defer the work to tcp_release_cb() */
 886                        set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
 887                }
 888                bh_unlock_sock(sk);
 889
 890                clear_bit(TSQ_QUEUED, &tp->tsq_flags);
 891                sk_free(sk);
 892        }
 893}
 894
 895#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |           \
 896                          (1UL << TCP_WRITE_TIMER_DEFERRED) |   \
 897                          (1UL << TCP_DELACK_TIMER_DEFERRED) |  \
 898                          (1UL << TCP_MTU_REDUCED_DEFERRED))
 899/**
 900 * tcp_release_cb - tcp release_sock() callback
 901 * @sk: socket
 902 *
 903 * called from release_sock() to perform protocol dependent
 904 * actions before socket release.
 905 */
 906void tcp_release_cb(struct sock *sk)
 907{
 908        struct tcp_sock *tp = tcp_sk(sk);
 909        unsigned long flags, nflags;
 910
 911        /* perform an atomic operation only if at least one flag is set */
 912        do {
 913                flags = tp->tsq_flags;
 914                if (!(flags & TCP_DEFERRED_ALL))
 915                        return;
 916                nflags = flags & ~TCP_DEFERRED_ALL;
 917        } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
 918
 919        if (flags & (1UL << TCP_TSQ_DEFERRED))
 920                tcp_tsq_handler(sk);
 921
 922        if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
 923                tcp_write_timer_handler(sk);
 924                __sock_put(sk);
 925        }
 926        if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
 927                tcp_delack_timer_handler(sk);
 928                __sock_put(sk);
 929        }
 930        if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
 931                sk->sk_prot->mtu_reduced(sk);
 932                __sock_put(sk);
 933        }
 934}
 935EXPORT_SYMBOL(tcp_release_cb);
 936
 937void __init tcp_tasklet_init(void)
 938{
 939        int i;
 940
 941        for_each_possible_cpu(i) {
 942                struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
 943
 944                INIT_LIST_HEAD(&tsq->head);
 945                tasklet_init(&tsq->tasklet,
 946                             tcp_tasklet_func,
 947                             (unsigned long)tsq);
 948        }
 949}
 950
 951/*
 952 * Write buffer destructor automatically called from kfree_skb.
 953 * We cant xmit new skbs from this context, as we might already
 954 * hold qdisc lock.
 955 */
 956static void tcp_wfree(struct sk_buff *skb)
 957{
 958        struct sock *sk = skb->sk;
 959        struct tcp_sock *tp = tcp_sk(sk);
 960
 961        if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
 962            !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
 963                unsigned long flags;
 964                struct tsq_tasklet *tsq;
 965
 966                /* Keep a ref on socket.
 967                 * This last ref will be released in tcp_tasklet_func()
 968                 */
 969                atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
 970
 971                /* queue this socket to tasklet queue */
 972                local_irq_save(flags);
 973                tsq = &__get_cpu_var(tsq_tasklet);
 974                list_add(&tp->tsq_node, &tsq->head);
 975                tasklet_schedule(&tsq->tasklet);
 976                local_irq_restore(flags);
 977        } else {
 978                sock_wfree(skb);
 979        }
 980}
 981
 982/* This routine actually transmits TCP packets queued in by
 983 * tcp_do_sendmsg().  This is used by both the initial
 984 * transmission and possible later retransmissions.
 985 * All SKB's seen here are completely headerless.  It is our
 986 * job to build the TCP header, and pass the packet down to
 987 * IP so it can do the same plus pass the packet off to the
 988 * device.
 989 *
 990 * We are working here with either a clone of the original
 991 * SKB, or a fresh unique copy made by the retransmit engine.
 992 */
 993static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
 994                            gfp_t gfp_mask)
 995{
 996        const struct inet_connection_sock *icsk = inet_csk(sk);
 997        struct inet_sock *inet;
 998        struct tcp_sock *tp;
 999        struct tcp_skb_cb *tcb;
1000        struct tcp_out_options opts;
1001        unsigned int tcp_options_size, tcp_header_size;
1002        struct tcp_md5sig_key *md5;
1003        struct tcphdr *th;
1004        int err;
1005
1006        BUG_ON(!skb || !tcp_skb_pcount(skb));
1007
1008        /* If congestion control is doing timestamping, we must
1009         * take such a timestamp before we potentially clone/copy.
1010         */
1011        if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
1012                __net_timestamp(skb);
1013
1014        if (likely(clone_it)) {
1015                if (unlikely(skb_cloned(skb)))
1016                        skb = pskb_copy(skb, gfp_mask);
1017                else
1018                        skb = skb_clone(skb, gfp_mask);
1019                if (unlikely(!skb))
1020                        return -ENOBUFS;
1021        }
1022
1023        inet = inet_sk(sk);
1024        tp = tcp_sk(sk);
1025        tcb = TCP_SKB_CB(skb);
1026        memset(&opts, 0, sizeof(opts));
1027
1028        if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1029                tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1030        else
1031                tcp_options_size = tcp_established_options(sk, skb, &opts,
1032                                                           &md5);
1033        tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1034
1035        if (tcp_packets_in_flight(tp) == 0) {
1036                tcp_ca_event(sk, CA_EVENT_TX_START);
1037                skb->ooo_okay = 1;
1038        } else
1039                skb->ooo_okay = 0;
1040
1041        skb_push(skb, tcp_header_size);
1042        skb_reset_transport_header(skb);
1043
1044        skb_orphan(skb);
1045        skb->sk = sk;
1046        skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
1047                          tcp_wfree : sock_wfree;
1048        atomic_add(skb->truesize, &sk->sk_wmem_alloc);
1049
1050        /* Build TCP header and checksum it. */
1051        th = tcp_hdr(skb);
1052        th->source              = inet->inet_sport;
1053        th->dest                = inet->inet_dport;
1054        th->seq                 = htonl(tcb->seq);
1055        th->ack_seq             = htonl(tp->rcv_nxt);
1056        *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
1057                                        tcb->tcp_flags);
1058
1059        if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1060                /* RFC1323: The window in SYN & SYN/ACK segments
1061                 * is never scaled.
1062                 */
1063                th->window      = htons(min(tp->rcv_wnd, 65535U));
1064        } else {
1065                th->window      = htons(tcp_select_window(sk));
1066        }
1067        th->check               = 0;
1068        th->urg_ptr             = 0;
1069
1070        /* The urg_mode check is necessary during a below snd_una win probe */
1071        if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1072                if (before(tp->snd_up, tcb->seq + 0x10000)) {
1073                        th->urg_ptr = htons(tp->snd_up - tcb->seq);
1074                        th->urg = 1;
1075                } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1076                        th->urg_ptr = htons(0xFFFF);
1077                        th->urg = 1;
1078                }
1079        }
1080
1081        tcp_options_write((__be32 *)(th + 1), tp, &opts);
1082        if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1083                TCP_ECN_send(sk, skb, tcp_header_size);
1084
1085#ifdef CONFIG_TCP_MD5SIG
1086        /* Calculate the MD5 hash, as we have all we need now */
1087        if (md5) {
1088                sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1089                tp->af_specific->calc_md5_hash(opts.hash_location,
1090                                               md5, sk, NULL, skb);
1091        }
1092#endif
1093
1094        icsk->icsk_af_ops->send_check(sk, skb);
1095
1096        if (likely(tcb->tcp_flags & TCPHDR_ACK))
1097                tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1098
1099        if (skb->len != tcp_header_size)
1100                tcp_event_data_sent(tp, sk);
1101
1102        if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1103                TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1104                              tcp_skb_pcount(skb));
1105
1106        err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1107        if (likely(err <= 0))
1108                return err;
1109
1110        tcp_enter_cwr(sk, 1);
1111
1112        return net_xmit_eval(err);
1113}
1114
1115/* This routine just queues the buffer for sending.
1116 *
1117 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1118 * otherwise socket can stall.
1119 */
1120static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1121{
1122        struct tcp_sock *tp = tcp_sk(sk);
1123
1124        /* Advance write_seq and place onto the write_queue. */
1125        tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1126        skb_header_release(skb);
1127        tcp_add_write_queue_tail(sk, skb);
1128        sk->sk_wmem_queued += skb->truesize;
1129        sk_mem_charge(sk, skb->truesize);
1130}
1131
1132/* Initialize TSO segments for a packet. */
1133static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1134                                 unsigned int mss_now)
1135{
1136        if (skb->len <= mss_now || !sk_can_gso(sk) ||
1137            skb->ip_summed == CHECKSUM_NONE) {
1138                /* Avoid the costly divide in the normal
1139                 * non-TSO case.
1140                 */
1141                skb_shinfo(skb)->gso_segs = 1;
1142                skb_shinfo(skb)->gso_size = 0;
1143                skb_shinfo(skb)->gso_type = 0;
1144        } else {
1145                skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1146                skb_shinfo(skb)->gso_size = mss_now;
1147                skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1148        }
1149}
1150
1151/* When a modification to fackets out becomes necessary, we need to check
1152 * skb is counted to fackets_out or not.
1153 */
1154static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1155                                   int decr)
1156{
1157        struct tcp_sock *tp = tcp_sk(sk);
1158
1159        if (!tp->sacked_out || tcp_is_reno(tp))
1160                return;
1161
1162        if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1163                tp->fackets_out -= decr;
1164}
1165
1166/* Pcount in the middle of the write queue got changed, we need to do various
1167 * tweaks to fix counters
1168 */
1169static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1170{
1171        struct tcp_sock *tp = tcp_sk(sk);
1172
1173        tp->packets_out -= decr;
1174
1175        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1176                tp->sacked_out -= decr;
1177        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1178                tp->retrans_out -= decr;
1179        if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1180                tp->lost_out -= decr;
1181
1182        /* Reno case is special. Sigh... */
1183        if (tcp_is_reno(tp) && decr > 0)
1184                tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1185
1186        tcp_adjust_fackets_out(sk, skb, decr);
1187
1188        if (tp->lost_skb_hint &&
1189            before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1190            (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1191                tp->lost_cnt_hint -= decr;
1192
1193        tcp_verify_left_out(tp);
1194}
1195
1196/* Function to create two new TCP segments.  Shrinks the given segment
1197 * to the specified size and appends a new segment with the rest of the
1198 * packet to the list.  This won't be called frequently, I hope.
1199 * Remember, these are still headerless SKBs at this point.
1200 */
1201int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1202                 unsigned int mss_now)
1203{
1204        struct tcp_sock *tp = tcp_sk(sk);
1205        struct sk_buff *buff;
1206        int nsize, old_factor;
1207        int nlen;
1208        u8 flags;
1209
1210        if (WARN_ON(len > skb->len))
1211                return -EINVAL;
1212
1213        nsize = skb_headlen(skb) - len;
1214        if (nsize < 0)
1215                nsize = 0;
1216
1217        if (skb_cloned(skb) &&
1218            skb_is_nonlinear(skb) &&
1219            pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1220                return -ENOMEM;
1221
1222        /* Get a new skb... force flag on. */
1223        buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1224        if (buff == NULL)
1225                return -ENOMEM; /* We'll just try again later. */
1226
1227        sk->sk_wmem_queued += buff->truesize;
1228        sk_mem_charge(sk, buff->truesize);
1229        nlen = skb->len - len - nsize;
1230        buff->truesize += nlen;
1231        skb->truesize -= nlen;
1232
1233        /* Correct the sequence numbers. */
1234        TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1235        TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1236        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1237
1238        /* PSH and FIN should only be set in the second packet. */
1239        flags = TCP_SKB_CB(skb)->tcp_flags;
1240        TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1241        TCP_SKB_CB(buff)->tcp_flags = flags;
1242        TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1243
1244        if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1245                /* Copy and checksum data tail into the new buffer. */
1246                buff->csum = csum_partial_copy_nocheck(skb->data + len,
1247                                                       skb_put(buff, nsize),
1248                                                       nsize, 0);
1249
1250                skb_trim(skb, len);
1251
1252                skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1253        } else {
1254                skb->ip_summed = CHECKSUM_PARTIAL;
1255                skb_split(skb, buff, len);
1256        }
1257
1258        buff->ip_summed = skb->ip_summed;
1259
1260        /* Looks stupid, but our code really uses when of
1261         * skbs, which it never sent before. --ANK
1262         */
1263        TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1264        buff->tstamp = skb->tstamp;
1265
1266        old_factor = tcp_skb_pcount(skb);
1267
1268        /* Fix up tso_factor for both original and new SKB.  */
1269        tcp_set_skb_tso_segs(sk, skb, mss_now);
1270        tcp_set_skb_tso_segs(sk, buff, mss_now);
1271
1272        /* If this packet has been sent out already, we must
1273         * adjust the various packet counters.
1274         */
1275        if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1276                int diff = old_factor - tcp_skb_pcount(skb) -
1277                        tcp_skb_pcount(buff);
1278
1279                if (diff)
1280                        tcp_adjust_pcount(sk, skb, diff);
1281        }
1282
1283        /* Link BUFF into the send queue. */
1284        skb_header_release(buff);
1285        tcp_insert_write_queue_after(skb, buff, sk);
1286
1287        return 0;
1288}
1289
1290/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1291 * eventually). The difference is that pulled data not copied, but
1292 * immediately discarded.
1293 */
1294static void __pskb_trim_head(struct sk_buff *skb, int len)
1295{
1296        int i, k, eat;
1297
1298        eat = min_t(int, len, skb_headlen(skb));
1299        if (eat) {
1300                __skb_pull(skb, eat);
1301                skb->avail_size -= eat;
1302                len -= eat;
1303                if (!len)
1304                        return;
1305        }
1306        eat = len;
1307        k = 0;
1308        for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1309                int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1310
1311                if (size <= eat) {
1312                        skb_frag_unref(skb, i);
1313                        eat -= size;
1314                } else {
1315                        skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1316                        if (eat) {
1317                                skb_shinfo(skb)->frags[k].page_offset += eat;
1318                                skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1319                                eat = 0;
1320                        }
1321                        k++;
1322                }
1323        }
1324        skb_shinfo(skb)->nr_frags = k;
1325
1326        skb_reset_tail_pointer(skb);
1327        skb->data_len -= len;
1328        skb->len = skb->data_len;
1329}
1330
1331/* Remove acked data from a packet in the transmit queue. */
1332int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1333{
1334        if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1335                return -ENOMEM;
1336
1337        __pskb_trim_head(skb, len);
1338
1339        TCP_SKB_CB(skb)->seq += len;
1340        skb->ip_summed = CHECKSUM_PARTIAL;
1341
1342        skb->truesize        -= len;
1343        sk->sk_wmem_queued   -= len;
1344        sk_mem_uncharge(sk, len);
1345        sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1346
1347        /* Any change of skb->len requires recalculation of tso factor. */
1348        if (tcp_skb_pcount(skb) > 1)
1349                tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1350
1351        return 0;
1352}
1353
1354/* Calculate MSS. Not accounting for SACKs here.  */
1355int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1356{
1357        const struct tcp_sock *tp = tcp_sk(sk);
1358        const struct inet_connection_sock *icsk = inet_csk(sk);
1359        int mss_now;
1360
1361        /* Calculate base mss without TCP options:
1362           It is MMS_S - sizeof(tcphdr) of rfc1122
1363         */
1364        mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1365
1366        /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1367        if (icsk->icsk_af_ops->net_frag_header_len) {
1368                const struct dst_entry *dst = __sk_dst_get(sk);
1369
1370                if (dst && dst_allfrag(dst))
1371                        mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1372        }
1373
1374        /* Clamp it (mss_clamp does not include tcp options) */
1375        if (mss_now > tp->rx_opt.mss_clamp)
1376                mss_now = tp->rx_opt.mss_clamp;
1377
1378        /* Now subtract optional transport overhead */
1379        mss_now -= icsk->icsk_ext_hdr_len;
1380
1381        /* Then reserve room for full set of TCP options and 8 bytes of data */
1382        if (mss_now < 48)
1383                mss_now = 48;
1384
1385        /* Now subtract TCP options size, not including SACKs */
1386        mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1387
1388        return mss_now;
1389}
1390
1391/* Inverse of above */
1392int tcp_mss_to_mtu(struct sock *sk, int mss)
1393{
1394        const struct tcp_sock *tp = tcp_sk(sk);
1395        const struct inet_connection_sock *icsk = inet_csk(sk);
1396        int mtu;
1397
1398        mtu = mss +
1399              tp->tcp_header_len +
1400              icsk->icsk_ext_hdr_len +
1401              icsk->icsk_af_ops->net_header_len;
1402
1403        /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1404        if (icsk->icsk_af_ops->net_frag_header_len) {
1405                const struct dst_entry *dst = __sk_dst_get(sk);
1406
1407                if (dst && dst_allfrag(dst))
1408                        mtu += icsk->icsk_af_ops->net_frag_header_len;
1409        }
1410        return mtu;
1411}
1412
1413/* MTU probing init per socket */
1414void tcp_mtup_init(struct sock *sk)
1415{
1416        struct tcp_sock *tp = tcp_sk(sk);
1417        struct inet_connection_sock *icsk = inet_csk(sk);
1418
1419        icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1420        icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1421                               icsk->icsk_af_ops->net_header_len;
1422        icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1423        icsk->icsk_mtup.probe_size = 0;
1424}
1425EXPORT_SYMBOL(tcp_mtup_init);
1426
1427/* This function synchronize snd mss to current pmtu/exthdr set.
1428
1429   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1430   for TCP options, but includes only bare TCP header.
1431
1432   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1433   It is minimum of user_mss and mss received with SYN.
1434   It also does not include TCP options.
1435
1436   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1437
1438   tp->mss_cache is current effective sending mss, including
1439   all tcp options except for SACKs. It is evaluated,
1440   taking into account current pmtu, but never exceeds
1441   tp->rx_opt.mss_clamp.
1442
1443   NOTE1. rfc1122 clearly states that advertised MSS
1444   DOES NOT include either tcp or ip options.
1445
1446   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1447   are READ ONLY outside this function.         --ANK (980731)
1448 */
1449unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1450{
1451        struct tcp_sock *tp = tcp_sk(sk);
1452        struct inet_connection_sock *icsk = inet_csk(sk);
1453        int mss_now;
1454
1455        if (icsk->icsk_mtup.search_high > pmtu)
1456                icsk->icsk_mtup.search_high = pmtu;
1457
1458        mss_now = tcp_mtu_to_mss(sk, pmtu);
1459        mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1460
1461        /* And store cached results */
1462        icsk->icsk_pmtu_cookie = pmtu;
1463        if (icsk->icsk_mtup.enabled)
1464                mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1465        tp->mss_cache = mss_now;
1466
1467        return mss_now;
1468}
1469EXPORT_SYMBOL(tcp_sync_mss);
1470
1471/* Compute the current effective MSS, taking SACKs and IP options,
1472 * and even PMTU discovery events into account.
1473 */
1474unsigned int tcp_current_mss(struct sock *sk)
1475{
1476        const struct tcp_sock *tp = tcp_sk(sk);
1477        const struct dst_entry *dst = __sk_dst_get(sk);
1478        u32 mss_now;
1479        unsigned int header_len;
1480        struct tcp_out_options opts;
1481        struct tcp_md5sig_key *md5;
1482
1483        mss_now = tp->mss_cache;
1484
1485        if (dst) {
1486                u32 mtu = dst_mtu(dst);
1487                if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1488                        mss_now = tcp_sync_mss(sk, mtu);
1489        }
1490
1491        header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1492                     sizeof(struct tcphdr);
1493        /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1494         * some common options. If this is an odd packet (because we have SACK
1495         * blocks etc) then our calculated header_len will be different, and
1496         * we have to adjust mss_now correspondingly */
1497        if (header_len != tp->tcp_header_len) {
1498                int delta = (int) header_len - tp->tcp_header_len;
1499                mss_now -= delta;
1500        }
1501
1502        return mss_now;
1503}
1504
1505/* Congestion window validation. (RFC2861) */
1506static void tcp_cwnd_validate(struct sock *sk)
1507{
1508        struct tcp_sock *tp = tcp_sk(sk);
1509
1510        if (tp->packets_out >= tp->snd_cwnd) {
1511                /* Network is feed fully. */
1512                tp->snd_cwnd_used = 0;
1513                tp->snd_cwnd_stamp = tcp_time_stamp;
1514        } else {
1515                /* Network starves. */
1516                if (tp->packets_out > tp->snd_cwnd_used)
1517                        tp->snd_cwnd_used = tp->packets_out;
1518
1519                if (sysctl_tcp_slow_start_after_idle &&
1520                    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1521                        tcp_cwnd_application_limited(sk);
1522        }
1523}
1524
1525/* Returns the portion of skb which can be sent right away without
1526 * introducing MSS oddities to segment boundaries. In rare cases where
1527 * mss_now != mss_cache, we will request caller to create a small skb
1528 * per input skb which could be mostly avoided here (if desired).
1529 *
1530 * We explicitly want to create a request for splitting write queue tail
1531 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1532 * thus all the complexity (cwnd_len is always MSS multiple which we
1533 * return whenever allowed by the other factors). Basically we need the
1534 * modulo only when the receiver window alone is the limiting factor or
1535 * when we would be allowed to send the split-due-to-Nagle skb fully.
1536 */
1537static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1538                                        unsigned int mss_now, unsigned int max_segs)
1539{
1540        const struct tcp_sock *tp = tcp_sk(sk);
1541        u32 needed, window, max_len;
1542
1543        window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1544        max_len = mss_now * max_segs;
1545
1546        if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1547                return max_len;
1548
1549        needed = min(skb->len, window);
1550
1551        if (max_len <= needed)
1552                return max_len;
1553
1554        return needed - needed % mss_now;
1555}
1556
1557/* Can at least one segment of SKB be sent right now, according to the
1558 * congestion window rules?  If so, return how many segments are allowed.
1559 */
1560static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1561                                         const struct sk_buff *skb)
1562{
1563        u32 in_flight, cwnd;
1564
1565        /* Don't be strict about the congestion window for the final FIN.  */
1566        if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1567            tcp_skb_pcount(skb) == 1)
1568                return 1;
1569
1570        in_flight = tcp_packets_in_flight(tp);
1571        cwnd = tp->snd_cwnd;
1572        if (in_flight < cwnd)
1573                return (cwnd - in_flight);
1574
1575        return 0;
1576}
1577
1578/* Initialize TSO state of a skb.
1579 * This must be invoked the first time we consider transmitting
1580 * SKB onto the wire.
1581 */
1582static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1583                             unsigned int mss_now)
1584{
1585        int tso_segs = tcp_skb_pcount(skb);
1586
1587        if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1588                tcp_set_skb_tso_segs(sk, skb, mss_now);
1589                tso_segs = tcp_skb_pcount(skb);
1590        }
1591        return tso_segs;
1592}
1593
1594/* Minshall's variant of the Nagle send check. */
1595static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1596{
1597        return after(tp->snd_sml, tp->snd_una) &&
1598                !after(tp->snd_sml, tp->snd_nxt);
1599}
1600
1601/* Return false, if packet can be sent now without violation Nagle's rules:
1602 * 1. It is full sized.
1603 * 2. Or it contains FIN. (already checked by caller)
1604 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1605 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1606 *    With Minshall's modification: all sent small packets are ACKed.
1607 */
1608static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1609                                  const struct sk_buff *skb,
1610                                  unsigned int mss_now, int nonagle)
1611{
1612        return skb->len < mss_now &&
1613                ((nonagle & TCP_NAGLE_CORK) ||
1614                 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1615}
1616
1617/* Return true if the Nagle test allows this packet to be
1618 * sent now.
1619 */
1620static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1621                                  unsigned int cur_mss, int nonagle)
1622{
1623        /* Nagle rule does not apply to frames, which sit in the middle of the
1624         * write_queue (they have no chances to get new data).
1625         *
1626         * This is implemented in the callers, where they modify the 'nonagle'
1627         * argument based upon the location of SKB in the send queue.
1628         */
1629        if (nonagle & TCP_NAGLE_PUSH)
1630                return true;
1631
1632        /* Don't use the nagle rule for urgent data (or for the final FIN).
1633         * Nagle can be ignored during F-RTO too (see RFC4138).
1634         */
1635        if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1636            (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1637                return true;
1638
1639        if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1640                return true;
1641
1642        return false;
1643}
1644
1645/* Does at least the first segment of SKB fit into the send window? */
1646static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1647                             const struct sk_buff *skb,
1648                             unsigned int cur_mss)
1649{
1650        u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1651
1652        if (skb->len > cur_mss)
1653                end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1654
1655        return !after(end_seq, tcp_wnd_end(tp));
1656}
1657
1658/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1659 * should be put on the wire right now.  If so, it returns the number of
1660 * packets allowed by the congestion window.
1661 */
1662static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1663                                 unsigned int cur_mss, int nonagle)
1664{
1665        const struct tcp_sock *tp = tcp_sk(sk);
1666        unsigned int cwnd_quota;
1667
1668        tcp_init_tso_segs(sk, skb, cur_mss);
1669
1670        if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1671                return 0;
1672
1673        cwnd_quota = tcp_cwnd_test(tp, skb);
1674        if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1675                cwnd_quota = 0;
1676
1677        return cwnd_quota;
1678}
1679
1680/* Test if sending is allowed right now. */
1681bool tcp_may_send_now(struct sock *sk)
1682{
1683        const struct tcp_sock *tp = tcp_sk(sk);
1684        struct sk_buff *skb = tcp_send_head(sk);
1685
1686        return skb &&
1687                tcp_snd_test(sk, skb, tcp_current_mss(sk),
1688                             (tcp_skb_is_last(sk, skb) ?
1689                              tp->nonagle : TCP_NAGLE_PUSH));
1690}
1691
1692/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1693 * which is put after SKB on the list.  It is very much like
1694 * tcp_fragment() except that it may make several kinds of assumptions
1695 * in order to speed up the splitting operation.  In particular, we
1696 * know that all the data is in scatter-gather pages, and that the
1697 * packet has never been sent out before (and thus is not cloned).
1698 */
1699static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1700                        unsigned int mss_now, gfp_t gfp)
1701{
1702        struct sk_buff *buff;
1703        int nlen = skb->len - len;
1704        u8 flags;
1705
1706        /* All of a TSO frame must be composed of paged data.  */
1707        if (skb->len != skb->data_len)
1708                return tcp_fragment(sk, skb, len, mss_now);
1709
1710        buff = sk_stream_alloc_skb(sk, 0, gfp);
1711        if (unlikely(buff == NULL))
1712                return -ENOMEM;
1713
1714        sk->sk_wmem_queued += buff->truesize;
1715        sk_mem_charge(sk, buff->truesize);
1716        buff->truesize += nlen;
1717        skb->truesize -= nlen;
1718
1719        /* Correct the sequence numbers. */
1720        TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1721        TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1722        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1723
1724        /* PSH and FIN should only be set in the second packet. */
1725        flags = TCP_SKB_CB(skb)->tcp_flags;
1726        TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1727        TCP_SKB_CB(buff)->tcp_flags = flags;
1728
1729        /* This packet was never sent out yet, so no SACK bits. */
1730        TCP_SKB_CB(buff)->sacked = 0;
1731
1732        buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1733        skb_split(skb, buff, len);
1734
1735        /* Fix up tso_factor for both original and new SKB.  */
1736        tcp_set_skb_tso_segs(sk, skb, mss_now);
1737        tcp_set_skb_tso_segs(sk, buff, mss_now);
1738
1739        /* Link BUFF into the send queue. */
1740        skb_header_release(buff);
1741        tcp_insert_write_queue_after(skb, buff, sk);
1742
1743        return 0;
1744}
1745
1746/* Try to defer sending, if possible, in order to minimize the amount
1747 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1748 *
1749 * This algorithm is from John Heffner.
1750 */
1751static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1752{
1753        struct tcp_sock *tp = tcp_sk(sk);
1754        const struct inet_connection_sock *icsk = inet_csk(sk);
1755        u32 send_win, cong_win, limit, in_flight;
1756        int win_divisor;
1757
1758        if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1759                goto send_now;
1760
1761        if (icsk->icsk_ca_state != TCP_CA_Open)
1762                goto send_now;
1763
1764        /* Defer for less than two clock ticks. */
1765        if (tp->tso_deferred &&
1766            (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1767                goto send_now;
1768
1769        in_flight = tcp_packets_in_flight(tp);
1770
1771        BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1772
1773        send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1774
1775        /* From in_flight test above, we know that cwnd > in_flight.  */
1776        cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1777
1778        limit = min(send_win, cong_win);
1779
1780        /* If a full-sized TSO skb can be sent, do it. */
1781        if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1782                           sk->sk_gso_max_segs * tp->mss_cache))
1783                goto send_now;
1784
1785        /* Middle in queue won't get any more data, full sendable already? */
1786        if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1787                goto send_now;
1788
1789        win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1790        if (win_divisor) {
1791                u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1792
1793                /* If at least some fraction of a window is available,
1794                 * just use it.
1795                 */
1796                chunk /= win_divisor;
1797                if (limit >= chunk)
1798                        goto send_now;
1799        } else {
1800                /* Different approach, try not to defer past a single
1801                 * ACK.  Receiver should ACK every other full sized
1802                 * frame, so if we have space for more than 3 frames
1803                 * then send now.
1804                 */
1805                if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1806                        goto send_now;
1807        }
1808
1809        /* Ok, it looks like it is advisable to defer.  */
1810        tp->tso_deferred = 1 | (jiffies << 1);
1811
1812        return true;
1813
1814send_now:
1815        tp->tso_deferred = 0;
1816        return false;
1817}
1818
1819/* Create a new MTU probe if we are ready.
1820 * MTU probe is regularly attempting to increase the path MTU by
1821 * deliberately sending larger packets.  This discovers routing
1822 * changes resulting in larger path MTUs.
1823 *
1824 * Returns 0 if we should wait to probe (no cwnd available),
1825 *         1 if a probe was sent,
1826 *         -1 otherwise
1827 */
1828static int tcp_mtu_probe(struct sock *sk)
1829{
1830        struct tcp_sock *tp = tcp_sk(sk);
1831        struct inet_connection_sock *icsk = inet_csk(sk);
1832        struct sk_buff *skb, *nskb, *next;
1833        int len;
1834        int probe_size;
1835        int size_needed;
1836        int copy;
1837        int mss_now;
1838
1839        /* Not currently probing/verifying,
1840         * not in recovery,
1841         * have enough cwnd, and
1842         * not SACKing (the variable headers throw things off) */
1843        if (!icsk->icsk_mtup.enabled ||
1844            icsk->icsk_mtup.probe_size ||
1845            inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1846            tp->snd_cwnd < 11 ||
1847            tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1848                return -1;
1849
1850        /* Very simple search strategy: just double the MSS. */
1851        mss_now = tcp_current_mss(sk);
1852        probe_size = 2 * tp->mss_cache;
1853        size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1854        if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1855                /* TODO: set timer for probe_converge_event */
1856                return -1;
1857        }
1858
1859        /* Have enough data in the send queue to probe? */
1860        if (tp->write_seq - tp->snd_nxt < size_needed)
1861                return -1;
1862
1863        if (tp->snd_wnd < size_needed)
1864                return -1;
1865        if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1866                return 0;
1867
1868        /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1869        if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1870                if (!tcp_packets_in_flight(tp))
1871                        return -1;
1872                else
1873                        return 0;
1874        }
1875
1876        /* We're allowed to probe.  Build it now. */
1877        if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1878                return -1;
1879        sk->sk_wmem_queued += nskb->truesize;
1880        sk_mem_charge(sk, nskb->truesize);
1881
1882        skb = tcp_send_head(sk);
1883
1884        TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1885        TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1886        TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1887        TCP_SKB_CB(nskb)->sacked = 0;
1888        nskb->csum = 0;
1889        nskb->ip_summed = skb->ip_summed;
1890
1891        tcp_insert_write_queue_before(nskb, skb, sk);
1892
1893        len = 0;
1894        tcp_for_write_queue_from_safe(skb, next, sk) {
1895                copy = min_t(int, skb->len, probe_size - len);
1896                if (nskb->ip_summed)
1897                        skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1898                else
1899                        nskb->csum = skb_copy_and_csum_bits(skb, 0,
1900                                                            skb_put(nskb, copy),
1901                                                            copy, nskb->csum);
1902
1903                if (skb->len <= copy) {
1904                        /* We've eaten all the data from this skb.
1905                         * Throw it away. */
1906                        TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1907                        tcp_unlink_write_queue(skb, sk);
1908                        sk_wmem_free_skb(sk, skb);
1909                } else {
1910                        TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1911                                                   ~(TCPHDR_FIN|TCPHDR_PSH);
1912                        if (!skb_shinfo(skb)->nr_frags) {
1913                                skb_pull(skb, copy);
1914                                if (skb->ip_summed != CHECKSUM_PARTIAL)
1915                                        skb->csum = csum_partial(skb->data,
1916                                                                 skb->len, 0);
1917                        } else {
1918                                __pskb_trim_head(skb, copy);
1919                                tcp_set_skb_tso_segs(sk, skb, mss_now);
1920                        }
1921                        TCP_SKB_CB(skb)->seq += copy;
1922                }
1923
1924                len += copy;
1925
1926                if (len >= probe_size)
1927                        break;
1928        }
1929        tcp_init_tso_segs(sk, nskb, nskb->len);
1930
1931        /* We're ready to send.  If this fails, the probe will
1932         * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1933        TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1934        if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1935                /* Decrement cwnd here because we are sending
1936                 * effectively two packets. */
1937                tp->snd_cwnd--;
1938                tcp_event_new_data_sent(sk, nskb);
1939
1940                icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1941                tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1942                tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1943
1944                return 1;
1945        }
1946
1947        return -1;
1948}
1949
1950/* This routine writes packets to the network.  It advances the
1951 * send_head.  This happens as incoming acks open up the remote
1952 * window for us.
1953 *
1954 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1955 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1956 * account rare use of URG, this is not a big flaw.
1957 *
1958 * Returns true, if no segments are in flight and we have queued segments,
1959 * but cannot send anything now because of SWS or another problem.
1960 */
1961static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1962                           int push_one, gfp_t gfp)
1963{
1964        struct tcp_sock *tp = tcp_sk(sk);
1965        struct sk_buff *skb;
1966        unsigned int tso_segs, sent_pkts;
1967        int cwnd_quota;
1968        int result;
1969
1970        sent_pkts = 0;
1971
1972        if (!push_one) {
1973                /* Do MTU probing. */
1974                result = tcp_mtu_probe(sk);
1975                if (!result) {
1976                        return false;
1977                } else if (result > 0) {
1978                        sent_pkts = 1;
1979                }
1980        }
1981
1982        while ((skb = tcp_send_head(sk))) {
1983                unsigned int limit;
1984
1985
1986                tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1987                BUG_ON(!tso_segs);
1988
1989                if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1990                        goto repair; /* Skip network transmission */
1991
1992                cwnd_quota = tcp_cwnd_test(tp, skb);
1993                if (!cwnd_quota)
1994                        break;
1995
1996                if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1997                        break;
1998
1999                if (tso_segs == 1) {
2000                        if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2001                                                     (tcp_skb_is_last(sk, skb) ?
2002                                                      nonagle : TCP_NAGLE_PUSH))))
2003                                break;
2004                } else {
2005                        if (!push_one && tcp_tso_should_defer(sk, skb))
2006                                break;
2007                }
2008
2009                /* TSQ : sk_wmem_alloc accounts skb truesize,
2010                 * including skb overhead. But thats OK.
2011                 */
2012                if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
2013                        set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2014                        break;
2015                }
2016                limit = mss_now;
2017                if (tso_segs > 1 && !tcp_urg_mode(tp))
2018                        limit = tcp_mss_split_point(sk, skb, mss_now,
2019                                                    min_t(unsigned int,
2020                                                          cwnd_quota,
2021                                                          sk->sk_gso_max_segs));
2022
2023                if (skb->len > limit &&
2024                    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2025                        break;
2026
2027                TCP_SKB_CB(skb)->when = tcp_time_stamp;
2028
2029                if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2030                        break;
2031
2032repair:
2033                /* Advance the send_head.  This one is sent out.
2034                 * This call will increment packets_out.
2035                 */
2036                tcp_event_new_data_sent(sk, skb);
2037
2038                tcp_minshall_update(tp, mss_now, skb);
2039                sent_pkts += tcp_skb_pcount(skb);
2040
2041                if (push_one)
2042                        break;
2043        }
2044
2045        if (likely(sent_pkts)) {
2046                if (tcp_in_cwnd_reduction(sk))
2047                        tp->prr_out += sent_pkts;
2048                tcp_cwnd_validate(sk);
2049                return false;
2050        }
2051        return !tp->packets_out && tcp_send_head(sk);
2052}
2053
2054/* Push out any pending frames which were held back due to
2055 * TCP_CORK or attempt at coalescing tiny packets.
2056 * The socket must be locked by the caller.
2057 */
2058void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2059                               int nonagle)
2060{
2061        /* If we are closed, the bytes will have to remain here.
2062         * In time closedown will finish, we empty the write queue and
2063         * all will be happy.
2064         */
2065        if (unlikely(sk->sk_state == TCP_CLOSE))
2066                return;
2067
2068        if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2069                           sk_gfp_atomic(sk, GFP_ATOMIC)))
2070                tcp_check_probe_timer(sk);
2071}
2072
2073/* Send _single_ skb sitting at the send head. This function requires
2074 * true push pending frames to setup probe timer etc.
2075 */
2076void tcp_push_one(struct sock *sk, unsigned int mss_now)
2077{
2078        struct sk_buff *skb = tcp_send_head(sk);
2079
2080        BUG_ON(!skb || skb->len < mss_now);
2081
2082        tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2083}
2084
2085/* This function returns the amount that we can raise the
2086 * usable window based on the following constraints
2087 *
2088 * 1. The window can never be shrunk once it is offered (RFC 793)
2089 * 2. We limit memory per socket
2090 *
2091 * RFC 1122:
2092 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2093 *  RECV.NEXT + RCV.WIN fixed until:
2094 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2095 *
2096 * i.e. don't raise the right edge of the window until you can raise
2097 * it at least MSS bytes.
2098 *
2099 * Unfortunately, the recommended algorithm breaks header prediction,
2100 * since header prediction assumes th->window stays fixed.
2101 *
2102 * Strictly speaking, keeping th->window fixed violates the receiver
2103 * side SWS prevention criteria. The problem is that under this rule
2104 * a stream of single byte packets will cause the right side of the
2105 * window to always advance by a single byte.
2106 *
2107 * Of course, if the sender implements sender side SWS prevention
2108 * then this will not be a problem.
2109 *
2110 * BSD seems to make the following compromise:
2111 *
2112 *      If the free space is less than the 1/4 of the maximum
2113 *      space available and the free space is less than 1/2 mss,
2114 *      then set the window to 0.
2115 *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2116 *      Otherwise, just prevent the window from shrinking
2117 *      and from being larger than the largest representable value.
2118 *
2119 * This prevents incremental opening of the window in the regime
2120 * where TCP is limited by the speed of the reader side taking
2121 * data out of the TCP receive queue. It does nothing about
2122 * those cases where the window is constrained on the sender side
2123 * because the pipeline is full.
2124 *
2125 * BSD also seems to "accidentally" limit itself to windows that are a
2126 * multiple of MSS, at least until the free space gets quite small.
2127 * This would appear to be a side effect of the mbuf implementation.
2128 * Combining these two algorithms results in the observed behavior
2129 * of having a fixed window size at almost all times.
2130 *
2131 * Below we obtain similar behavior by forcing the offered window to
2132 * a multiple of the mss when it is feasible to do so.
2133 *
2134 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2135 * Regular options like TIMESTAMP are taken into account.
2136 */
2137u32 __tcp_select_window(struct sock *sk)
2138{
2139        struct inet_connection_sock *icsk = inet_csk(sk);
2140        struct tcp_sock *tp = tcp_sk(sk);
2141        /* MSS for the peer's data.  Previous versions used mss_clamp
2142         * here.  I don't know if the value based on our guesses
2143         * of peer's MSS is better for the performance.  It's more correct
2144         * but may be worse for the performance because of rcv_mss
2145         * fluctuations.  --SAW  1998/11/1
2146         */
2147        int mss = icsk->icsk_ack.rcv_mss;
2148        int free_space = tcp_space(sk);
2149        int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2150        int window;
2151
2152        if (mss > full_space)
2153                mss = full_space;
2154
2155        if (free_space < (full_space >> 1)) {
2156                icsk->icsk_ack.quick = 0;
2157
2158                if (sk_under_memory_pressure(sk))
2159                        tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2160                                               4U * tp->advmss);
2161
2162                if (free_space < mss)
2163                        return 0;
2164        }
2165
2166        if (free_space > tp->rcv_ssthresh)
2167                free_space = tp->rcv_ssthresh;
2168
2169        /* Don't do rounding if we are using window scaling, since the
2170         * scaled window will not line up with the MSS boundary anyway.
2171         */
2172        window = tp->rcv_wnd;
2173        if (tp->rx_opt.rcv_wscale) {
2174                window = free_space;
2175
2176                /* Advertise enough space so that it won't get scaled away.
2177                 * Import case: prevent zero window announcement if
2178                 * 1<<rcv_wscale > mss.
2179                 */
2180                if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2181                        window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2182                                  << tp->rx_opt.rcv_wscale);
2183        } else {
2184                /* Get the largest window that is a nice multiple of mss.
2185                 * Window clamp already applied above.
2186                 * If our current window offering is within 1 mss of the
2187                 * free space we just keep it. This prevents the divide
2188                 * and multiply from happening most of the time.
2189                 * We also don't do any window rounding when the free space
2190                 * is too small.
2191                 */
2192                if (window <= free_space - mss || window > free_space)
2193                        window = (free_space / mss) * mss;
2194                else if (mss == full_space &&
2195                         free_space > window + (full_space >> 1))
2196                        window = free_space;
2197        }
2198
2199        return window;
2200}
2201
2202/* Collapses two adjacent SKB's during retransmission. */
2203static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2204{
2205        struct tcp_sock *tp = tcp_sk(sk);
2206        struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2207        int skb_size, next_skb_size;
2208
2209        skb_size = skb->len;
2210        next_skb_size = next_skb->len;
2211
2212        BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2213
2214        tcp_highest_sack_combine(sk, next_skb, skb);
2215
2216        tcp_unlink_write_queue(next_skb, sk);
2217
2218        skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2219                                  next_skb_size);
2220
2221        if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2222                skb->ip_summed = CHECKSUM_PARTIAL;
2223
2224        if (skb->ip_summed != CHECKSUM_PARTIAL)
2225                skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2226
2227        /* Update sequence range on original skb. */
2228        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2229
2230        /* Merge over control information. This moves PSH/FIN etc. over */
2231        TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2232
2233        /* All done, get rid of second SKB and account for it so
2234         * packet counting does not break.
2235         */
2236        TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2237
2238        /* changed transmit queue under us so clear hints */
2239        tcp_clear_retrans_hints_partial(tp);
2240        if (next_skb == tp->retransmit_skb_hint)
2241                tp->retransmit_skb_hint = skb;
2242
2243        tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2244
2245        sk_wmem_free_skb(sk, next_skb);
2246}
2247
2248/* Check if coalescing SKBs is legal. */
2249static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2250{
2251        if (tcp_skb_pcount(skb) > 1)
2252                return false;
2253        /* TODO: SACK collapsing could be used to remove this condition */
2254        if (skb_shinfo(skb)->nr_frags != 0)
2255                return false;
2256        if (skb_cloned(skb))
2257                return false;
2258        if (skb == tcp_send_head(sk))
2259                return false;
2260        /* Some heurestics for collapsing over SACK'd could be invented */
2261        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2262                return false;
2263
2264        return true;
2265}
2266
2267/* Collapse packets in the retransmit queue to make to create
2268 * less packets on the wire. This is only done on retransmission.
2269 */
2270static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2271                                     int space)
2272{
2273        struct tcp_sock *tp = tcp_sk(sk);
2274        struct sk_buff *skb = to, *tmp;
2275        bool first = true;
2276
2277        if (!sysctl_tcp_retrans_collapse)
2278                return;
2279        if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2280                return;
2281
2282        tcp_for_write_queue_from_safe(skb, tmp, sk) {
2283                if (!tcp_can_collapse(sk, skb))
2284                        break;
2285
2286                space -= skb->len;
2287
2288                if (first) {
2289                        first = false;
2290                        continue;
2291                }
2292
2293                if (space < 0)
2294                        break;
2295                /* Punt if not enough space exists in the first SKB for
2296                 * the data in the second
2297                 */
2298                if (skb->len > skb_availroom(to))
2299                        break;
2300
2301                if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2302                        break;
2303
2304                tcp_collapse_retrans(sk, to);
2305        }
2306}
2307
2308/* This retransmits one SKB.  Policy decisions and retransmit queue
2309 * state updates are done by the caller.  Returns non-zero if an
2310 * error occurred which prevented the send.
2311 */
2312int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2313{
2314        struct tcp_sock *tp = tcp_sk(sk);
2315        struct inet_connection_sock *icsk = inet_csk(sk);
2316        unsigned int cur_mss;
2317
2318        /* Inconslusive MTU probe */
2319        if (icsk->icsk_mtup.probe_size) {
2320                icsk->icsk_mtup.probe_size = 0;
2321        }
2322
2323        /* Do not sent more than we queued. 1/4 is reserved for possible
2324         * copying overhead: fragmentation, tunneling, mangling etc.
2325         */
2326        if (atomic_read(&sk->sk_wmem_alloc) >
2327            min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2328                return -EAGAIN;
2329
2330        if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2331                if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2332                        BUG();
2333                if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2334                        return -ENOMEM;
2335        }
2336
2337        if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2338                return -EHOSTUNREACH; /* Routing failure or similar. */
2339
2340        cur_mss = tcp_current_mss(sk);
2341
2342        /* If receiver has shrunk his window, and skb is out of
2343         * new window, do not retransmit it. The exception is the
2344         * case, when window is shrunk to zero. In this case
2345         * our retransmit serves as a zero window probe.
2346         */
2347        if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2348            TCP_SKB_CB(skb)->seq != tp->snd_una)
2349                return -EAGAIN;
2350
2351        if (skb->len > cur_mss) {
2352                if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2353                        return -ENOMEM; /* We'll try again later. */
2354        } else {
2355                int oldpcount = tcp_skb_pcount(skb);
2356
2357                if (unlikely(oldpcount > 1)) {
2358                        tcp_init_tso_segs(sk, skb, cur_mss);
2359                        tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2360                }
2361        }
2362
2363        tcp_retrans_try_collapse(sk, skb, cur_mss);
2364
2365        /* Some Solaris stacks overoptimize and ignore the FIN on a
2366         * retransmit when old data is attached.  So strip it off
2367         * since it is cheap to do so and saves bytes on the network.
2368         */
2369        if (skb->len > 0 &&
2370            (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2371            tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2372                if (!pskb_trim(skb, 0)) {
2373                        /* Reuse, even though it does some unnecessary work */
2374                        tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2375                                             TCP_SKB_CB(skb)->tcp_flags);
2376                        skb->ip_summed = CHECKSUM_NONE;
2377                }
2378        }
2379
2380        /* Make a copy, if the first transmission SKB clone we made
2381         * is still in somebody's hands, else make a clone.
2382         */
2383        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2384
2385        /* make sure skb->data is aligned on arches that require it */
2386        if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2387                struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2388                                                   GFP_ATOMIC);
2389                return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2390                              -ENOBUFS;
2391        } else {
2392                return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2393        }
2394}
2395
2396int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2397{
2398        struct tcp_sock *tp = tcp_sk(sk);
2399        int err = __tcp_retransmit_skb(sk, skb);
2400
2401        if (err == 0) {
2402                /* Update global TCP statistics. */
2403                TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2404
2405                tp->total_retrans++;
2406
2407#if FASTRETRANS_DEBUG > 0
2408                if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2409                        net_dbg_ratelimited("retrans_out leaked\n");
2410                }
2411#endif
2412                if (!tp->retrans_out)
2413                        tp->lost_retrans_low = tp->snd_nxt;
2414                TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2415                tp->retrans_out += tcp_skb_pcount(skb);
2416
2417                /* Save stamp of the first retransmit. */
2418                if (!tp->retrans_stamp)
2419                        tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2420
2421                tp->undo_retrans += tcp_skb_pcount(skb);
2422
2423                /* snd_nxt is stored to detect loss of retransmitted segment,
2424                 * see tcp_input.c tcp_sacktag_write_queue().
2425                 */
2426                TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2427        }
2428        return err;
2429}
2430
2431/* Check if we forward retransmits are possible in the current
2432 * window/congestion state.
2433 */
2434static bool tcp_can_forward_retransmit(struct sock *sk)
2435{
2436        const struct inet_connection_sock *icsk = inet_csk(sk);
2437        const struct tcp_sock *tp = tcp_sk(sk);
2438
2439        /* Forward retransmissions are possible only during Recovery. */
2440        if (icsk->icsk_ca_state != TCP_CA_Recovery)
2441                return false;
2442
2443        /* No forward retransmissions in Reno are possible. */
2444        if (tcp_is_reno(tp))
2445                return false;
2446
2447        /* Yeah, we have to make difficult choice between forward transmission
2448         * and retransmission... Both ways have their merits...
2449         *
2450         * For now we do not retransmit anything, while we have some new
2451         * segments to send. In the other cases, follow rule 3 for
2452         * NextSeg() specified in RFC3517.
2453         */
2454
2455        if (tcp_may_send_now(sk))
2456                return false;
2457
2458        return true;
2459}
2460
2461/* This gets called after a retransmit timeout, and the initially
2462 * retransmitted data is acknowledged.  It tries to continue
2463 * resending the rest of the retransmit queue, until either
2464 * we've sent it all or the congestion window limit is reached.
2465 * If doing SACK, the first ACK which comes back for a timeout
2466 * based retransmit packet might feed us FACK information again.
2467 * If so, we use it to avoid unnecessarily retransmissions.
2468 */
2469void tcp_xmit_retransmit_queue(struct sock *sk)
2470{
2471        const struct inet_connection_sock *icsk = inet_csk(sk);
2472        struct tcp_sock *tp = tcp_sk(sk);
2473        struct sk_buff *skb;
2474        struct sk_buff *hole = NULL;
2475        u32 last_lost;
2476        int mib_idx;
2477        int fwd_rexmitting = 0;
2478
2479        if (!tp->packets_out)
2480                return;
2481
2482        if (!tp->lost_out)
2483                tp->retransmit_high = tp->snd_una;
2484
2485        if (tp->retransmit_skb_hint) {
2486                skb = tp->retransmit_skb_hint;
2487                last_lost = TCP_SKB_CB(skb)->end_seq;
2488                if (after(last_lost, tp->retransmit_high))
2489                        last_lost = tp->retransmit_high;
2490        } else {
2491                skb = tcp_write_queue_head(sk);
2492                last_lost = tp->snd_una;
2493        }
2494
2495        tcp_for_write_queue_from(skb, sk) {
2496                __u8 sacked = TCP_SKB_CB(skb)->sacked;
2497
2498                if (skb == tcp_send_head(sk))
2499                        break;
2500                /* we could do better than to assign each time */
2501                if (hole == NULL)
2502                        tp->retransmit_skb_hint = skb;
2503
2504                /* Assume this retransmit will generate
2505                 * only one packet for congestion window
2506                 * calculation purposes.  This works because
2507                 * tcp_retransmit_skb() will chop up the
2508                 * packet to be MSS sized and all the
2509                 * packet counting works out.
2510                 */
2511                if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2512                        return;
2513
2514                if (fwd_rexmitting) {
2515begin_fwd:
2516                        if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2517                                break;
2518                        mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2519
2520                } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2521                        tp->retransmit_high = last_lost;
2522                        if (!tcp_can_forward_retransmit(sk))
2523                                break;
2524                        /* Backtrack if necessary to non-L'ed skb */
2525                        if (hole != NULL) {
2526                                skb = hole;
2527                                hole = NULL;
2528                        }
2529                        fwd_rexmitting = 1;
2530                        goto begin_fwd;
2531
2532                } else if (!(sacked & TCPCB_LOST)) {
2533                        if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2534                                hole = skb;
2535                        continue;
2536
2537                } else {
2538                        last_lost = TCP_SKB_CB(skb)->end_seq;
2539                        if (icsk->icsk_ca_state != TCP_CA_Loss)
2540                                mib_idx = LINUX_MIB_TCPFASTRETRANS;
2541                        else
2542                                mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2543                }
2544
2545                if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2546                        continue;
2547
2548                if (tcp_retransmit_skb(sk, skb)) {
2549                        NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2550                        return;
2551                }
2552                NET_INC_STATS_BH(sock_net(sk), mib_idx);
2553
2554                if (tcp_in_cwnd_reduction(sk))
2555                        tp->prr_out += tcp_skb_pcount(skb);
2556
2557                if (skb == tcp_write_queue_head(sk))
2558                        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2559                                                  inet_csk(sk)->icsk_rto,
2560                                                  TCP_RTO_MAX);
2561        }
2562}
2563
2564/* Send a fin.  The caller locks the socket for us.  This cannot be
2565 * allowed to fail queueing a FIN frame under any circumstances.
2566 */
2567void tcp_send_fin(struct sock *sk)
2568{
2569        struct tcp_sock *tp = tcp_sk(sk);
2570        struct sk_buff *skb = tcp_write_queue_tail(sk);
2571        int mss_now;
2572
2573        /* Optimization, tack on the FIN if we have a queue of
2574         * unsent frames.  But be careful about outgoing SACKS
2575         * and IP options.
2576         */
2577        mss_now = tcp_current_mss(sk);
2578
2579        if (tcp_send_head(sk) != NULL) {
2580                TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2581                TCP_SKB_CB(skb)->end_seq++;
2582                tp->write_seq++;
2583        } else {
2584                /* Socket is locked, keep trying until memory is available. */
2585                for (;;) {
2586                        skb = alloc_skb_fclone(MAX_TCP_HEADER,
2587                                               sk->sk_allocation);
2588                        if (skb)
2589                                break;
2590                        yield();
2591                }
2592
2593                /* Reserve space for headers and prepare control bits. */
2594                skb_reserve(skb, MAX_TCP_HEADER);
2595                /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2596                tcp_init_nondata_skb(skb, tp->write_seq,
2597                                     TCPHDR_ACK | TCPHDR_FIN);
2598                tcp_queue_skb(sk, skb);
2599        }
2600        __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2601}
2602
2603/* We get here when a process closes a file descriptor (either due to
2604 * an explicit close() or as a byproduct of exit()'ing) and there
2605 * was unread data in the receive queue.  This behavior is recommended
2606 * by RFC 2525, section 2.17.  -DaveM
2607 */
2608void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2609{
2610        struct sk_buff *skb;
2611
2612        /* NOTE: No TCP options attached and we never retransmit this. */
2613        skb = alloc_skb(MAX_TCP_HEADER, priority);
2614        if (!skb) {
2615                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2616                return;
2617        }
2618
2619        /* Reserve space for headers and prepare control bits. */
2620        skb_reserve(skb, MAX_TCP_HEADER);
2621        tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2622                             TCPHDR_ACK | TCPHDR_RST);
2623        /* Send it off. */
2624        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2625        if (tcp_transmit_skb(sk, skb, 0, priority))
2626                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2627
2628        TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2629}
2630
2631/* Send a crossed SYN-ACK during socket establishment.
2632 * WARNING: This routine must only be called when we have already sent
2633 * a SYN packet that crossed the incoming SYN that caused this routine
2634 * to get called. If this assumption fails then the initial rcv_wnd
2635 * and rcv_wscale values will not be correct.
2636 */
2637int tcp_send_synack(struct sock *sk)
2638{
2639        struct sk_buff *skb;
2640
2641        skb = tcp_write_queue_head(sk);
2642        if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2643                pr_debug("%s: wrong queue state\n", __func__);
2644                return -EFAULT;
2645        }
2646        if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2647                if (skb_cloned(skb)) {
2648                        struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2649                        if (nskb == NULL)
2650                                return -ENOMEM;
2651                        tcp_unlink_write_queue(skb, sk);
2652                        skb_header_release(nskb);
2653                        __tcp_add_write_queue_head(sk, nskb);
2654                        sk_wmem_free_skb(sk, skb);
2655                        sk->sk_wmem_queued += nskb->truesize;
2656                        sk_mem_charge(sk, nskb->truesize);
2657                        skb = nskb;
2658                }
2659
2660                TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2661                TCP_ECN_send_synack(tcp_sk(sk), skb);
2662        }
2663        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2664        return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2665}
2666
2667/**
2668 * tcp_make_synack - Prepare a SYN-ACK.
2669 * sk: listener socket
2670 * dst: dst entry attached to the SYNACK
2671 * req: request_sock pointer
2672 * rvp: request_values pointer
2673 *
2674 * Allocate one skb and build a SYNACK packet.
2675 * @dst is consumed : Caller should not use it again.
2676 */
2677struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2678                                struct request_sock *req,
2679                                struct request_values *rvp,
2680                                struct tcp_fastopen_cookie *foc)
2681{
2682        struct tcp_out_options opts;
2683        struct tcp_extend_values *xvp = tcp_xv(rvp);
2684        struct inet_request_sock *ireq = inet_rsk(req);
2685        struct tcp_sock *tp = tcp_sk(sk);
2686        const struct tcp_cookie_values *cvp = tp->cookie_values;
2687        struct tcphdr *th;
2688        struct sk_buff *skb;
2689        struct tcp_md5sig_key *md5;
2690        int tcp_header_size;
2691        int mss;
2692        int s_data_desired = 0;
2693
2694        if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2695                s_data_desired = cvp->s_data_desired;
2696        skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired,
2697                        sk_gfp_atomic(sk, GFP_ATOMIC));
2698        if (unlikely(!skb)) {
2699                dst_release(dst);
2700                return NULL;
2701        }
2702        /* Reserve space for headers. */
2703        skb_reserve(skb, MAX_TCP_HEADER);
2704
2705        skb_dst_set(skb, dst);
2706
2707        mss = dst_metric_advmss(dst);
2708        if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2709                mss = tp->rx_opt.user_mss;
2710
2711        if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2712                __u8 rcv_wscale;
2713                /* Set this up on the first call only */
2714                req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2715
2716                /* limit the window selection if the user enforce a smaller rx buffer */
2717                if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2718                    (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2719                        req->window_clamp = tcp_full_space(sk);
2720
2721                /* tcp_full_space because it is guaranteed to be the first packet */
2722                tcp_select_initial_window(tcp_full_space(sk),
2723                        mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2724                        &req->rcv_wnd,
2725                        &req->window_clamp,
2726                        ireq->wscale_ok,
2727                        &rcv_wscale,
2728                        dst_metric(dst, RTAX_INITRWND));
2729                ireq->rcv_wscale = rcv_wscale;
2730        }
2731
2732        memset(&opts, 0, sizeof(opts));
2733#ifdef CONFIG_SYN_COOKIES
2734        if (unlikely(req->cookie_ts))
2735                TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2736        else
2737#endif
2738        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2739        tcp_header_size = tcp_synack_options(sk, req, mss,
2740                                             skb, &opts, &md5, xvp, foc)
2741                        + sizeof(*th);
2742
2743        skb_push(skb, tcp_header_size);
2744        skb_reset_transport_header(skb);
2745
2746        th = tcp_hdr(skb);
2747        memset(th, 0, sizeof(struct tcphdr));
2748        th->syn = 1;
2749        th->ack = 1;
2750        TCP_ECN_make_synack(req, th);
2751        th->source = ireq->loc_port;
2752        th->dest = ireq->rmt_port;
2753        /* Setting of flags are superfluous here for callers (and ECE is
2754         * not even correctly set)
2755         */
2756        tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2757                             TCPHDR_SYN | TCPHDR_ACK);
2758
2759        if (OPTION_COOKIE_EXTENSION & opts.options) {
2760                if (s_data_desired) {
2761                        u8 *buf = skb_put(skb, s_data_desired);
2762
2763                        /* copy data directly from the listening socket. */
2764                        memcpy(buf, cvp->s_data_payload, s_data_desired);
2765                        TCP_SKB_CB(skb)->end_seq += s_data_desired;
2766                }
2767
2768                if (opts.hash_size > 0) {
2769                        __u32 workspace[SHA_WORKSPACE_WORDS];
2770                        u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2771                        u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2772
2773                        /* Secret recipe depends on the Timestamp, (future)
2774                         * Sequence and Acknowledgment Numbers, Initiator
2775                         * Cookie, and others handled by IP variant caller.
2776                         */
2777                        *tail-- ^= opts.tsval;
2778                        *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2779                        *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2780
2781                        /* recommended */
2782                        *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2783                        *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2784
2785                        sha_transform((__u32 *)&xvp->cookie_bakery[0],
2786                                      (char *)mess,
2787                                      &workspace[0]);
2788                        opts.hash_location =
2789                                (__u8 *)&xvp->cookie_bakery[0];
2790                }
2791        }
2792
2793        th->seq = htonl(TCP_SKB_CB(skb)->seq);
2794        /* XXX data is queued and acked as is. No buffer/window check */
2795        th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2796
2797        /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2798        th->window = htons(min(req->rcv_wnd, 65535U));
2799        tcp_options_write((__be32 *)(th + 1), tp, &opts);
2800        th->doff = (tcp_header_size >> 2);
2801        TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2802
2803#ifdef CONFIG_TCP_MD5SIG
2804        /* Okay, we have all we need - do the md5 hash if needed */
2805        if (md5) {
2806                tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2807                                               md5, NULL, req, skb);
2808        }
2809#endif
2810
2811        return skb;
2812}
2813EXPORT_SYMBOL(tcp_make_synack);
2814
2815/* Do all connect socket setups that can be done AF independent. */
2816void tcp_connect_init(struct sock *sk)
2817{
2818        const struct dst_entry *dst = __sk_dst_get(sk);
2819        struct tcp_sock *tp = tcp_sk(sk);
2820        __u8 rcv_wscale;
2821
2822        /* We'll fix this up when we get a response from the other end.
2823         * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2824         */
2825        tp->tcp_header_len = sizeof(struct tcphdr) +
2826                (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2827
2828#ifdef CONFIG_TCP_MD5SIG
2829        if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2830                tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2831#endif
2832
2833        /* If user gave his TCP_MAXSEG, record it to clamp */
2834        if (tp->rx_opt.user_mss)
2835                tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2836        tp->max_window = 0;
2837        tcp_mtup_init(sk);
2838        tcp_sync_mss(sk, dst_mtu(dst));
2839
2840        if (!tp->window_clamp)
2841                tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2842        tp->advmss = dst_metric_advmss(dst);
2843        if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2844                tp->advmss = tp->rx_opt.user_mss;
2845
2846        tcp_initialize_rcv_mss(sk);
2847
2848        /* limit the window selection if the user enforce a smaller rx buffer */
2849        if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2850            (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2851                tp->window_clamp = tcp_full_space(sk);
2852
2853        tcp_select_initial_window(tcp_full_space(sk),
2854                                  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2855                                  &tp->rcv_wnd,
2856                                  &tp->window_clamp,
2857                                  sysctl_tcp_window_scaling,
2858                                  &rcv_wscale,
2859                                  dst_metric(dst, RTAX_INITRWND));
2860
2861        tp->rx_opt.rcv_wscale = rcv_wscale;
2862        tp->rcv_ssthresh = tp->rcv_wnd;
2863
2864        sk->sk_err = 0;
2865        sock_reset_flag(sk, SOCK_DONE);
2866        tp->snd_wnd = 0;
2867        tcp_init_wl(tp, 0);
2868        tp->snd_una = tp->write_seq;
2869        tp->snd_sml = tp->write_seq;
2870        tp->snd_up = tp->write_seq;
2871        tp->snd_nxt = tp->write_seq;
2872
2873        if (likely(!tp->repair))
2874                tp->rcv_nxt = 0;
2875        tp->rcv_wup = tp->rcv_nxt;
2876        tp->copied_seq = tp->rcv_nxt;
2877
2878        inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2879        inet_csk(sk)->icsk_retransmits = 0;
2880        tcp_clear_retrans(tp);
2881}
2882
2883static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2884{
2885        struct tcp_sock *tp = tcp_sk(sk);
2886        struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2887
2888        tcb->end_seq += skb->len;
2889        skb_header_release(skb);
2890        __tcp_add_write_queue_tail(sk, skb);
2891        sk->sk_wmem_queued += skb->truesize;
2892        sk_mem_charge(sk, skb->truesize);
2893        tp->write_seq = tcb->end_seq;
2894        tp->packets_out += tcp_skb_pcount(skb);
2895}
2896
2897/* Build and send a SYN with data and (cached) Fast Open cookie. However,
2898 * queue a data-only packet after the regular SYN, such that regular SYNs
2899 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2900 * only the SYN sequence, the data are retransmitted in the first ACK.
2901 * If cookie is not cached or other error occurs, falls back to send a
2902 * regular SYN with Fast Open cookie request option.
2903 */
2904static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2905{
2906        struct tcp_sock *tp = tcp_sk(sk);
2907        struct tcp_fastopen_request *fo = tp->fastopen_req;
2908        int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2909        struct sk_buff *syn_data = NULL, *data;
2910        unsigned long last_syn_loss = 0;
2911
2912        tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
2913        tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2914                               &syn_loss, &last_syn_loss);
2915        /* Recurring FO SYN losses: revert to regular handshake temporarily */
2916        if (syn_loss > 1 &&
2917            time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2918                fo->cookie.len = -1;
2919                goto fallback;
2920        }
2921
2922        if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2923                fo->cookie.len = -1;
2924        else if (fo->cookie.len <= 0)
2925                goto fallback;
2926
2927        /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2928         * user-MSS. Reserve maximum option space for middleboxes that add
2929         * private TCP options. The cost is reduced data space in SYN :(
2930         */
2931        if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2932                tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2933        space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2934                MAX_TCP_OPTION_SPACE;
2935
2936        syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2937                                   sk->sk_allocation);
2938        if (syn_data == NULL)
2939                goto fallback;
2940
2941        for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2942                struct iovec *iov = &fo->data->msg_iov[i];
2943                unsigned char __user *from = iov->iov_base;
2944                int len = iov->iov_len;
2945
2946                if (syn_data->len + len > space)
2947                        len = space - syn_data->len;
2948                else if (i + 1 == iovlen)
2949                        /* No more data pending in inet_wait_for_connect() */
2950                        fo->data = NULL;
2951
2952                if (skb_add_data(syn_data, from, len))
2953                        goto fallback;
2954        }
2955
2956        /* Queue a data-only packet after the regular SYN for retransmission */
2957        data = pskb_copy(syn_data, sk->sk_allocation);
2958        if (data == NULL)
2959                goto fallback;
2960        TCP_SKB_CB(data)->seq++;
2961        TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2962        TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2963        tcp_connect_queue_skb(sk, data);
2964        fo->copied = data->len;
2965
2966        if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2967                tp->syn_data = (fo->copied > 0);
2968                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2969                goto done;
2970        }
2971        syn_data = NULL;
2972
2973fallback:
2974        /* Send a regular SYN with Fast Open cookie request option */
2975        if (fo->cookie.len > 0)
2976                fo->cookie.len = 0;
2977        err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2978        if (err)
2979                tp->syn_fastopen = 0;
2980        kfree_skb(syn_data);
2981done:
2982        fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
2983        return err;
2984}
2985
2986/* Build a SYN and send it off. */
2987int tcp_connect(struct sock *sk)
2988{
2989        struct tcp_sock *tp = tcp_sk(sk);
2990        struct sk_buff *buff;
2991        int err;
2992
2993        tcp_connect_init(sk);
2994
2995        buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2996        if (unlikely(buff == NULL))
2997                return -ENOBUFS;
2998
2999        /* Reserve space for headers. */
3000        skb_reserve(buff, MAX_TCP_HEADER);
3001
3002        tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3003        tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
3004        tcp_connect_queue_skb(sk, buff);
3005        TCP_ECN_send_syn(sk, buff);
3006
3007        /* Send off SYN; include data in Fast Open. */
3008        err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3009              tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3010        if (err == -ECONNREFUSED)
3011                return err;
3012
3013        /* We change tp->snd_nxt after the tcp_transmit_skb() call
3014         * in order to make this packet get counted in tcpOutSegs.
3015         */
3016        tp->snd_nxt = tp->write_seq;
3017        tp->pushed_seq = tp->write_seq;
3018        TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3019
3020        /* Timer for repeating the SYN until an answer. */
3021        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3022                                  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3023        return 0;
3024}
3025EXPORT_SYMBOL(tcp_connect);
3026
3027/* Send out a delayed ack, the caller does the policy checking
3028 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3029 * for details.
3030 */
3031void tcp_send_delayed_ack(struct sock *sk)
3032{
3033        struct inet_connection_sock *icsk = inet_csk(sk);
3034        int ato = icsk->icsk_ack.ato;
3035        unsigned long timeout;
3036
3037        if (ato > TCP_DELACK_MIN) {
3038                const struct tcp_sock *tp = tcp_sk(sk);
3039                int max_ato = HZ / 2;
3040
3041                if (icsk->icsk_ack.pingpong ||
3042                    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3043                        max_ato = TCP_DELACK_MAX;
3044
3045                /* Slow path, intersegment interval is "high". */
3046
3047                /* If some rtt estimate is known, use it to bound delayed ack.
3048                 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3049                 * directly.
3050                 */
3051                if (tp->srtt) {
3052                        int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3053
3054                        if (rtt < max_ato)
3055                                max_ato = rtt;
3056                }
3057
3058                ato = min(ato, max_ato);
3059        }
3060
3061        /* Stay within the limit we were given */
3062        timeout = jiffies + ato;
3063
3064        /* Use new timeout only if there wasn't a older one earlier. */
3065        if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3066                /* If delack timer was blocked or is about to expire,
3067                 * send ACK now.
3068                 */
3069                if (icsk->icsk_ack.blocked ||
3070                    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3071                        tcp_send_ack(sk);
3072                        return;
3073                }
3074
3075                if (!time_before(timeout, icsk->icsk_ack.timeout))
3076                        timeout = icsk->icsk_ack.timeout;
3077        }
3078        icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3079        icsk->icsk_ack.timeout = timeout;
3080        sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3081}
3082
3083/* This routine sends an ack and also updates the window. */
3084void tcp_send_ack(struct sock *sk)
3085{
3086        struct sk_buff *buff;
3087
3088        /* If we have been reset, we may not send again. */
3089        if (sk->sk_state == TCP_CLOSE)
3090                return;
3091
3092        /* We are not putting this on the write queue, so
3093         * tcp_transmit_skb() will set the ownership to this
3094         * sock.
3095         */
3096        buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3097        if (buff == NULL) {
3098                inet_csk_schedule_ack(sk);
3099                inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3100                inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3101                                          TCP_DELACK_MAX, TCP_RTO_MAX);
3102                return;
3103        }
3104
3105        /* Reserve space for headers and prepare control bits. */
3106        skb_reserve(buff, MAX_TCP_HEADER);
3107        tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3108
3109        /* Send it off, this clears delayed acks for us. */
3110        TCP_SKB_CB(buff)->when = tcp_time_stamp;
3111        tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3112}
3113
3114/* This routine sends a packet with an out of date sequence
3115 * number. It assumes the other end will try to ack it.
3116 *
3117 * Question: what should we make while urgent mode?
3118 * 4.4BSD forces sending single byte of data. We cannot send
3119 * out of window data, because we have SND.NXT==SND.MAX...
3120 *
3121 * Current solution: to send TWO zero-length segments in urgent mode:
3122 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3123 * out-of-date with SND.UNA-1 to probe window.
3124 */
3125static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3126{
3127        struct tcp_sock *tp = tcp_sk(sk);
3128        struct sk_buff *skb;
3129
3130        /* We don't queue it, tcp_transmit_skb() sets ownership. */
3131        skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3132        if (skb == NULL)
3133                return -1;
3134
3135        /* Reserve space for headers and set control bits. */
3136        skb_reserve(skb, MAX_TCP_HEADER);
3137        /* Use a previous sequence.  This should cause the other
3138         * end to send an ack.  Don't queue or clone SKB, just
3139         * send it.
3140         */
3141        tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3142        TCP_SKB_CB(skb)->when = tcp_time_stamp;
3143        return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3144}
3145
3146void tcp_send_window_probe(struct sock *sk)
3147{
3148        if (sk->sk_state == TCP_ESTABLISHED) {
3149                tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3150                tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3151                tcp_xmit_probe_skb(sk, 0);
3152        }
3153}
3154
3155/* Initiate keepalive or window probe from timer. */
3156int tcp_write_wakeup(struct sock *sk)
3157{
3158        struct tcp_sock *tp = tcp_sk(sk);
3159        struct sk_buff *skb;
3160
3161        if (sk->sk_state == TCP_CLOSE)
3162                return -1;
3163
3164        if ((skb = tcp_send_head(sk)) != NULL &&
3165            before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3166                int err;
3167                unsigned int mss = tcp_current_mss(sk);
3168                unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3169
3170                if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3171                        tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3172
3173                /* We are probing the opening of a window
3174                 * but the window size is != 0
3175                 * must have been a result SWS avoidance ( sender )
3176                 */
3177                if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3178                    skb->len > mss) {
3179                        seg_size = min(seg_size, mss);
3180                        TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3181                        if (tcp_fragment(sk, skb, seg_size, mss))
3182                                return -1;
3183                } else if (!tcp_skb_pcount(skb))
3184                        tcp_set_skb_tso_segs(sk, skb, mss);
3185
3186                TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3187                TCP_SKB_CB(skb)->when = tcp_time_stamp;
3188                err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3189                if (!err)
3190                        tcp_event_new_data_sent(sk, skb);
3191                return err;
3192        } else {
3193                if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3194                        tcp_xmit_probe_skb(sk, 1);
3195                return tcp_xmit_probe_skb(sk, 0);
3196        }
3197}
3198
3199/* A window probe timeout has occurred.  If window is not closed send
3200 * a partial packet else a zero probe.
3201 */
3202void tcp_send_probe0(struct sock *sk)
3203{
3204        struct inet_connection_sock *icsk = inet_csk(sk);
3205        struct tcp_sock *tp = tcp_sk(sk);
3206        int err;
3207
3208        err = tcp_write_wakeup(sk);
3209
3210        if (tp->packets_out || !tcp_send_head(sk)) {
3211                /* Cancel probe timer, if it is not required. */
3212                icsk->icsk_probes_out = 0;
3213                icsk->icsk_backoff = 0;
3214                return;
3215        }
3216
3217        if (err <= 0) {
3218                if (icsk->icsk_backoff < sysctl_tcp_retries2)
3219                        icsk->icsk_backoff++;
3220                icsk->icsk_probes_out++;
3221                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3222                                          min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3223                                          TCP_RTO_MAX);
3224        } else {
3225                /* If packet was not sent due to local congestion,
3226                 * do not backoff and do not remember icsk_probes_out.
3227                 * Let local senders to fight for local resources.
3228                 *
3229                 * Use accumulated backoff yet.
3230                 */
3231                if (!icsk->icsk_probes_out)
3232                        icsk->icsk_probes_out = 1;
3233                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3234                                          min(icsk->icsk_rto << icsk->icsk_backoff,
3235                                              TCP_RESOURCE_PROBE_INTERVAL),
3236                                          TCP_RTO_MAX);
3237        }
3238}
3239
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.