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#include <net/tcp.h>
  38
  39#include <linux/compiler.h>
  40#include <linux/gfp.h>
  41#include <linux/module.h>
  42
  43/* People can turn this off for buggy TCP's found in printers etc. */
  44int sysctl_tcp_retrans_collapse __read_mostly = 1;
  45
  46/* People can turn this on to work with those rare, broken TCPs that
  47 * interpret the window field as a signed quantity.
  48 */
  49int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  50
  51/* This limits the percentage of the congestion window which we
  52 * will allow a single TSO frame to consume.  Building TSO frames
  53 * which are too large can cause TCP streams to be bursty.
  54 */
  55int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  56
  57int sysctl_tcp_mtu_probing __read_mostly = 0;
  58int sysctl_tcp_base_mss __read_mostly = 512;
  59
  60/* By default, RFC2861 behavior.  */
  61int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  62
  63int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  64EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  65
  66
  67/* Account for new data that has been sent to the network. */
  68static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  69{
  70        struct tcp_sock *tp = tcp_sk(sk);
  71        unsigned int prior_packets = tp->packets_out;
  72
  73        tcp_advance_send_head(sk, skb);
  74        tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  75
  76        /* Don't override Nagle indefinately with F-RTO */
  77        if (tp->frto_counter == 2)
  78                tp->frto_counter = 3;
  79
  80        tp->packets_out += tcp_skb_pcount(skb);
  81        if (!prior_packets)
  82                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  83                                          inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  84}
  85
  86/* SND.NXT, if window was not shrunk.
  87 * If window has been shrunk, what should we make? It is not clear at all.
  88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  90 * invalid. OK, let's make this for now:
  91 */
  92static inline __u32 tcp_acceptable_seq(struct sock *sk)
  93{
  94        struct tcp_sock *tp = tcp_sk(sk);
  95
  96        if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  97                return tp->snd_nxt;
  98        else
  99                return tcp_wnd_end(tp);
 100}
 101
 102/* Calculate mss to advertise in SYN segment.
 103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 104 *
 105 * 1. It is independent of path mtu.
 106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 108 *    attached devices, because some buggy hosts are confused by
 109 *    large MSS.
 110 * 4. We do not make 3, we advertise MSS, calculated from first
 111 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 112 *    This may be overridden via information stored in routing table.
 113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 114 *    probably even Jumbo".
 115 */
 116static __u16 tcp_advertise_mss(struct sock *sk)
 117{
 118        struct tcp_sock *tp = tcp_sk(sk);
 119        struct dst_entry *dst = __sk_dst_get(sk);
 120        int mss = tp->advmss;
 121
 122        if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
 123                mss = dst_metric(dst, RTAX_ADVMSS);
 124                tp->advmss = mss;
 125        }
 126
 127        return (__u16)mss;
 128}
 129
 130/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 131 * This is the first part of cwnd validation mechanism. */
 132static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
 133{
 134        struct tcp_sock *tp = tcp_sk(sk);
 135        s32 delta = tcp_time_stamp - tp->lsndtime;
 136        u32 restart_cwnd = tcp_init_cwnd(tp, dst);
 137        u32 cwnd = tp->snd_cwnd;
 138
 139        tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 140
 141        tp->snd_ssthresh = tcp_current_ssthresh(sk);
 142        restart_cwnd = min(restart_cwnd, cwnd);
 143
 144        while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 145                cwnd >>= 1;
 146        tp->snd_cwnd = max(cwnd, restart_cwnd);
 147        tp->snd_cwnd_stamp = tcp_time_stamp;
 148        tp->snd_cwnd_used = 0;
 149}
 150
 151/* Congestion state accounting after a packet has been sent. */
 152static void tcp_event_data_sent(struct tcp_sock *tp,
 153                                struct sk_buff *skb, struct sock *sk)
 154{
 155        struct inet_connection_sock *icsk = inet_csk(sk);
 156        const u32 now = tcp_time_stamp;
 157
 158        if (sysctl_tcp_slow_start_after_idle &&
 159            (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
 160                tcp_cwnd_restart(sk, __sk_dst_get(sk));
 161
 162        tp->lsndtime = now;
 163
 164        /* If it is a reply for ato after last received
 165         * packet, enter pingpong mode.
 166         */
 167        if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 168                icsk->icsk_ack.pingpong = 1;
 169}
 170
 171/* Account for an ACK we sent. */
 172static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
 173{
 174        tcp_dec_quickack_mode(sk, pkts);
 175        inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 176}
 177
 178/* Determine a window scaling and initial window to offer.
 179 * Based on the assumption that the given amount of space
 180 * will be offered. Store the results in the tp structure.
 181 * NOTE: for smooth operation initial space offering should
 182 * be a multiple of mss if possible. We assume here that mss >= 1.
 183 * This MUST be enforced by all callers.
 184 */
 185void tcp_select_initial_window(int __space, __u32 mss,
 186                               __u32 *rcv_wnd, __u32 *window_clamp,
 187                               int wscale_ok, __u8 *rcv_wscale,
 188                               __u32 init_rcv_wnd)
 189{
 190        unsigned int space = (__space < 0 ? 0 : __space);
 191
 192        /* If no clamp set the clamp to the max possible scaled window */
 193        if (*window_clamp == 0)
 194                (*window_clamp) = (65535 << 14);
 195        space = min(*window_clamp, space);
 196
 197        /* Quantize space offering to a multiple of mss if possible. */
 198        if (space > mss)
 199                space = (space / mss) * mss;
 200
 201        /* NOTE: offering an initial window larger than 32767
 202         * will break some buggy TCP stacks. If the admin tells us
 203         * it is likely we could be speaking with such a buggy stack
 204         * we will truncate our initial window offering to 32K-1
 205         * unless the remote has sent us a window scaling option,
 206         * which we interpret as a sign the remote TCP is not
 207         * misinterpreting the window field as a signed quantity.
 208         */
 209        if (sysctl_tcp_workaround_signed_windows)
 210                (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 211        else
 212                (*rcv_wnd) = space;
 213
 214        (*rcv_wscale) = 0;
 215        if (wscale_ok) {
 216                /* Set window scaling on max possible window
 217                 * See RFC1323 for an explanation of the limit to 14
 218                 */
 219                space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
 220                space = min_t(u32, space, *window_clamp);
 221                while (space > 65535 && (*rcv_wscale) < 14) {
 222                        space >>= 1;
 223                        (*rcv_wscale)++;
 224                }
 225        }
 226
 227        /* Set initial window to value enough for senders,
 228         * following RFC2414. Senders, not following this RFC,
 229         * will be satisfied with 2.
 230         */
 231        if (mss > (1 << *rcv_wscale)) {
 232                int init_cwnd = 4;
 233                if (mss > 1460 * 3)
 234                        init_cwnd = 2;
 235                else if (mss > 1460)
 236                        init_cwnd = 3;
 237                /* when initializing use the value from init_rcv_wnd
 238                 * rather than the default from above
 239                 */
 240                if (init_rcv_wnd &&
 241                    (*rcv_wnd > init_rcv_wnd * mss))
 242                        *rcv_wnd = init_rcv_wnd * mss;
 243                else if (*rcv_wnd > init_cwnd * mss)
 244                        *rcv_wnd = init_cwnd * mss;
 245        }
 246
 247        /* Set the clamp no higher than max representable value */
 248        (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
 249}
 250
 251/* Chose a new window to advertise, update state in tcp_sock for the
 252 * socket, and return result with RFC1323 scaling applied.  The return
 253 * value can be stuffed directly into th->window for an outgoing
 254 * frame.
 255 */
 256static u16 tcp_select_window(struct sock *sk)
 257{
 258        struct tcp_sock *tp = tcp_sk(sk);
 259        u32 cur_win = tcp_receive_window(tp);
 260        u32 new_win = __tcp_select_window(sk);
 261
 262        /* Never shrink the offered window */
 263        if (new_win < cur_win) {
 264                /* Danger Will Robinson!
 265                 * Don't update rcv_wup/rcv_wnd here or else
 266                 * we will not be able to advertise a zero
 267                 * window in time.  --DaveM
 268                 *
 269                 * Relax Will Robinson.
 270                 */
 271                new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 272        }
 273        tp->rcv_wnd = new_win;
 274        tp->rcv_wup = tp->rcv_nxt;
 275
 276        /* Make sure we do not exceed the maximum possible
 277         * scaled window.
 278         */
 279        if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
 280                new_win = min(new_win, MAX_TCP_WINDOW);
 281        else
 282                new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 283
 284        /* RFC1323 scaling applied */
 285        new_win >>= tp->rx_opt.rcv_wscale;
 286
 287        /* If we advertise zero window, disable fast path. */
 288        if (new_win == 0)
 289                tp->pred_flags = 0;
 290
 291        return new_win;
 292}
 293
 294/* Packet ECN state for a SYN-ACK */
 295static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
 296{
 297        TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
 298        if (!(tp->ecn_flags & TCP_ECN_OK))
 299                TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
 300}
 301
 302/* Packet ECN state for a SYN.  */
 303static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
 304{
 305        struct tcp_sock *tp = tcp_sk(sk);
 306
 307        tp->ecn_flags = 0;
 308        if (sysctl_tcp_ecn == 1) {
 309                TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
 310                tp->ecn_flags = TCP_ECN_OK;
 311        }
 312}
 313
 314static __inline__ void
 315TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
 316{
 317        if (inet_rsk(req)->ecn_ok)
 318                th->ece = 1;
 319}
 320
 321/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 322 * be sent.
 323 */
 324static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
 325                                int tcp_header_len)
 326{
 327        struct tcp_sock *tp = tcp_sk(sk);
 328
 329        if (tp->ecn_flags & TCP_ECN_OK) {
 330                /* Not-retransmitted data segment: set ECT and inject CWR. */
 331                if (skb->len != tcp_header_len &&
 332                    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 333                        INET_ECN_xmit(sk);
 334                        if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 335                                tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 336                                tcp_hdr(skb)->cwr = 1;
 337                                skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 338                        }
 339                } else {
 340                        /* ACK or retransmitted segment: clear ECT|CE */
 341                        INET_ECN_dontxmit(sk);
 342                }
 343                if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 344                        tcp_hdr(skb)->ece = 1;
 345        }
 346}
 347
 348/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 349 * auto increment end seqno.
 350 */
 351static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 352{
 353        skb->ip_summed = CHECKSUM_PARTIAL;
 354        skb->csum = 0;
 355
 356        TCP_SKB_CB(skb)->flags = flags;
 357        TCP_SKB_CB(skb)->sacked = 0;
 358
 359        skb_shinfo(skb)->gso_segs = 1;
 360        skb_shinfo(skb)->gso_size = 0;
 361        skb_shinfo(skb)->gso_type = 0;
 362
 363        TCP_SKB_CB(skb)->seq = seq;
 364        if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
 365                seq++;
 366        TCP_SKB_CB(skb)->end_seq = seq;
 367}
 368
 369static inline int tcp_urg_mode(const struct tcp_sock *tp)
 370{
 371        return tp->snd_una != tp->snd_up;
 372}
 373
 374#define OPTION_SACK_ADVERTISE   (1 << 0)
 375#define OPTION_TS               (1 << 1)
 376#define OPTION_MD5              (1 << 2)
 377#define OPTION_WSCALE           (1 << 3)
 378#define OPTION_COOKIE_EXTENSION (1 << 4)
 379
 380struct tcp_out_options {
 381        u8 options;             /* bit field of OPTION_* */
 382        u8 ws;                  /* window scale, 0 to disable */
 383        u8 num_sack_blocks;     /* number of SACK blocks to include */
 384        u8 hash_size;           /* bytes in hash_location */
 385        u16 mss;                /* 0 to disable */
 386        __u32 tsval, tsecr;     /* need to include OPTION_TS */
 387        __u8 *hash_location;    /* temporary pointer, overloaded */
 388};
 389
 390/* The sysctl int routines are generic, so check consistency here.
 391 */
 392static u8 tcp_cookie_size_check(u8 desired)
 393{
 394        if (desired > 0) {
 395                /* previously specified */
 396                return desired;
 397        }
 398        if (sysctl_tcp_cookie_size <= 0) {
 399                /* no default specified */
 400                return 0;
 401        }
 402        if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) {
 403                /* value too small, specify minimum */
 404                return TCP_COOKIE_MIN;
 405        }
 406        if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) {
 407                /* value too large, specify maximum */
 408                return TCP_COOKIE_MAX;
 409        }
 410        if (0x1 & sysctl_tcp_cookie_size) {
 411                /* 8-bit multiple, illegal, fix it */
 412                return (u8)(sysctl_tcp_cookie_size + 0x1);
 413        }
 414        return (u8)sysctl_tcp_cookie_size;
 415}
 416
 417/* Write previously computed TCP options to the packet.
 418 *
 419 * Beware: Something in the Internet is very sensitive to the ordering of
 420 * TCP options, we learned this through the hard way, so be careful here.
 421 * Luckily we can at least blame others for their non-compliance but from
 422 * inter-operatibility perspective it seems that we're somewhat stuck with
 423 * the ordering which we have been using if we want to keep working with
 424 * those broken things (not that it currently hurts anybody as there isn't
 425 * particular reason why the ordering would need to be changed).
 426 *
 427 * At least SACK_PERM as the first option is known to lead to a disaster
 428 * (but it may well be that other scenarios fail similarly).
 429 */
 430static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 431                              struct tcp_out_options *opts)
 432{
 433        u8 options = opts->options;     /* mungable copy */
 434
 435        /* Having both authentication and cookies for security is redundant,
 436         * and there's certainly not enough room.  Instead, the cookie-less
 437         * extension variant is proposed.
 438         *
 439         * Consider the pessimal case with authentication.  The options
 440         * could look like:
 441         *   COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
 442         */
 443        if (unlikely(OPTION_MD5 & options)) {
 444                if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 445                        *ptr++ = htonl((TCPOPT_COOKIE << 24) |
 446                                       (TCPOLEN_COOKIE_BASE << 16) |
 447                                       (TCPOPT_MD5SIG << 8) |
 448                                       TCPOLEN_MD5SIG);
 449                } else {
 450                        *ptr++ = htonl((TCPOPT_NOP << 24) |
 451                                       (TCPOPT_NOP << 16) |
 452                                       (TCPOPT_MD5SIG << 8) |
 453                                       TCPOLEN_MD5SIG);
 454                }
 455                options &= ~OPTION_COOKIE_EXTENSION;
 456                /* overload cookie hash location */
 457                opts->hash_location = (__u8 *)ptr;
 458                ptr += 4;
 459        }
 460
 461        if (unlikely(opts->mss)) {
 462                *ptr++ = htonl((TCPOPT_MSS << 24) |
 463                               (TCPOLEN_MSS << 16) |
 464                               opts->mss);
 465        }
 466
 467        if (likely(OPTION_TS & options)) {
 468                if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 469                        *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 470                                       (TCPOLEN_SACK_PERM << 16) |
 471                                       (TCPOPT_TIMESTAMP << 8) |
 472                                       TCPOLEN_TIMESTAMP);
 473                        options &= ~OPTION_SACK_ADVERTISE;
 474                } else {
 475                        *ptr++ = htonl((TCPOPT_NOP << 24) |
 476                                       (TCPOPT_NOP << 16) |
 477                                       (TCPOPT_TIMESTAMP << 8) |
 478                                       TCPOLEN_TIMESTAMP);
 479                }
 480                *ptr++ = htonl(opts->tsval);
 481                *ptr++ = htonl(opts->tsecr);
 482        }
 483
 484        /* Specification requires after timestamp, so do it now.
 485         *
 486         * Consider the pessimal case without authentication.  The options
 487         * could look like:
 488         *   MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
 489         */
 490        if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 491                __u8 *cookie_copy = opts->hash_location;
 492                u8 cookie_size = opts->hash_size;
 493
 494                /* 8-bit multiple handled in tcp_cookie_size_check() above,
 495                 * and elsewhere.
 496                 */
 497                if (0x2 & cookie_size) {
 498                        __u8 *p = (__u8 *)ptr;
 499
 500                        /* 16-bit multiple */
 501                        *p++ = TCPOPT_COOKIE;
 502                        *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
 503                        *p++ = *cookie_copy++;
 504                        *p++ = *cookie_copy++;
 505                        ptr++;
 506                        cookie_size -= 2;
 507                } else {
 508                        /* 32-bit multiple */
 509                        *ptr++ = htonl(((TCPOPT_NOP << 24) |
 510                                        (TCPOPT_NOP << 16) |
 511                                        (TCPOPT_COOKIE << 8) |
 512                                        TCPOLEN_COOKIE_BASE) +
 513                                       cookie_size);
 514                }
 515
 516                if (cookie_size > 0) {
 517                        memcpy(ptr, cookie_copy, cookie_size);
 518                        ptr += (cookie_size / 4);
 519                }
 520        }
 521
 522        if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 523                *ptr++ = htonl((TCPOPT_NOP << 24) |
 524                               (TCPOPT_NOP << 16) |
 525                               (TCPOPT_SACK_PERM << 8) |
 526                               TCPOLEN_SACK_PERM);
 527        }
 528
 529        if (unlikely(OPTION_WSCALE & options)) {
 530                *ptr++ = htonl((TCPOPT_NOP << 24) |
 531                               (TCPOPT_WINDOW << 16) |
 532                               (TCPOLEN_WINDOW << 8) |
 533                               opts->ws);
 534        }
 535
 536        if (unlikely(opts->num_sack_blocks)) {
 537                struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 538                        tp->duplicate_sack : tp->selective_acks;
 539                int this_sack;
 540
 541                *ptr++ = htonl((TCPOPT_NOP  << 24) |
 542                               (TCPOPT_NOP  << 16) |
 543                               (TCPOPT_SACK <<  8) |
 544                               (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 545                                                     TCPOLEN_SACK_PERBLOCK)));
 546
 547                for (this_sack = 0; this_sack < opts->num_sack_blocks;
 548                     ++this_sack) {
 549                        *ptr++ = htonl(sp[this_sack].start_seq);
 550                        *ptr++ = htonl(sp[this_sack].end_seq);
 551                }
 552
 553                tp->rx_opt.dsack = 0;
 554        }
 555}
 556
 557/* Compute TCP options for SYN packets. This is not the final
 558 * network wire format yet.
 559 */
 560static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 561                                struct tcp_out_options *opts,
 562                                struct tcp_md5sig_key **md5) {
 563        struct tcp_sock *tp = tcp_sk(sk);
 564        struct tcp_cookie_values *cvp = tp->cookie_values;
 565        unsigned remaining = MAX_TCP_OPTION_SPACE;
 566        u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
 567                         tcp_cookie_size_check(cvp->cookie_desired) :
 568                         0;
 569
 570#ifdef CONFIG_TCP_MD5SIG
 571        *md5 = tp->af_specific->md5_lookup(sk, sk);
 572        if (*md5) {
 573                opts->options |= OPTION_MD5;
 574                remaining -= TCPOLEN_MD5SIG_ALIGNED;
 575        }
 576#else
 577        *md5 = NULL;
 578#endif
 579
 580        /* We always get an MSS option.  The option bytes which will be seen in
 581         * normal data packets should timestamps be used, must be in the MSS
 582         * advertised.  But we subtract them from tp->mss_cache so that
 583         * calculations in tcp_sendmsg are simpler etc.  So account for this
 584         * fact here if necessary.  If we don't do this correctly, as a
 585         * receiver we won't recognize data packets as being full sized when we
 586         * should, and thus we won't abide by the delayed ACK rules correctly.
 587         * SACKs don't matter, we never delay an ACK when we have any of those
 588         * going out.  */
 589        opts->mss = tcp_advertise_mss(sk);
 590        remaining -= TCPOLEN_MSS_ALIGNED;
 591
 592        if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
 593                opts->options |= OPTION_TS;
 594                opts->tsval = TCP_SKB_CB(skb)->when;
 595                opts->tsecr = tp->rx_opt.ts_recent;
 596                remaining -= TCPOLEN_TSTAMP_ALIGNED;
 597        }
 598        if (likely(sysctl_tcp_window_scaling)) {
 599                opts->ws = tp->rx_opt.rcv_wscale;
 600                opts->options |= OPTION_WSCALE;
 601                remaining -= TCPOLEN_WSCALE_ALIGNED;
 602        }
 603        if (likely(sysctl_tcp_sack)) {
 604                opts->options |= OPTION_SACK_ADVERTISE;
 605                if (unlikely(!(OPTION_TS & opts->options)))
 606                        remaining -= TCPOLEN_SACKPERM_ALIGNED;
 607        }
 608
 609        /* Note that timestamps are required by the specification.
 610         *
 611         * Odd numbers of bytes are prohibited by the specification, ensuring
 612         * that the cookie is 16-bit aligned, and the resulting cookie pair is
 613         * 32-bit aligned.
 614         */
 615        if (*md5 == NULL &&
 616            (OPTION_TS & opts->options) &&
 617            cookie_size > 0) {
 618                int need = TCPOLEN_COOKIE_BASE + cookie_size;
 619
 620                if (0x2 & need) {
 621                        /* 32-bit multiple */
 622                        need += 2; /* NOPs */
 623
 624                        if (need > remaining) {
 625                                /* try shrinking cookie to fit */
 626                                cookie_size -= 2;
 627                                need -= 4;
 628                        }
 629                }
 630                while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
 631                        cookie_size -= 4;
 632                        need -= 4;
 633                }
 634                if (TCP_COOKIE_MIN <= cookie_size) {
 635                        opts->options |= OPTION_COOKIE_EXTENSION;
 636                        opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
 637                        opts->hash_size = cookie_size;
 638
 639                        /* Remember for future incarnations. */
 640                        cvp->cookie_desired = cookie_size;
 641
 642                        if (cvp->cookie_desired != cvp->cookie_pair_size) {
 643                                /* Currently use random bytes as a nonce,
 644                                 * assuming these are completely unpredictable
 645                                 * by hostile users of the same system.
 646                                 */
 647                                get_random_bytes(&cvp->cookie_pair[0],
 648                                                 cookie_size);
 649                                cvp->cookie_pair_size = cookie_size;
 650                        }
 651
 652                        remaining -= need;
 653                }
 654        }
 655        return MAX_TCP_OPTION_SPACE - remaining;
 656}
 657
 658/* Set up TCP options for SYN-ACKs. */
 659static unsigned tcp_synack_options(struct sock *sk,
 660                                   struct request_sock *req,
 661                                   unsigned mss, struct sk_buff *skb,
 662                                   struct tcp_out_options *opts,
 663                                   struct tcp_md5sig_key **md5,
 664                                   struct tcp_extend_values *xvp)
 665{
 666        struct inet_request_sock *ireq = inet_rsk(req);
 667        unsigned remaining = MAX_TCP_OPTION_SPACE;
 668        u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
 669                         xvp->cookie_plus :
 670                         0;
 671
 672#ifdef CONFIG_TCP_MD5SIG
 673        *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
 674        if (*md5) {
 675                opts->options |= OPTION_MD5;
 676                remaining -= TCPOLEN_MD5SIG_ALIGNED;
 677
 678                /* We can't fit any SACK blocks in a packet with MD5 + TS
 679                 * options. There was discussion about disabling SACK
 680                 * rather than TS in order to fit in better with old,
 681                 * buggy kernels, but that was deemed to be unnecessary.
 682                 */
 683                ireq->tstamp_ok &= !ireq->sack_ok;
 684        }
 685#else
 686        *md5 = NULL;
 687#endif
 688
 689        /* We always send an MSS option. */
 690        opts->mss = mss;
 691        remaining -= TCPOLEN_MSS_ALIGNED;
 692
 693        if (likely(ireq->wscale_ok)) {
 694                opts->ws = ireq->rcv_wscale;
 695                opts->options |= OPTION_WSCALE;
 696                remaining -= TCPOLEN_WSCALE_ALIGNED;
 697        }
 698        if (likely(ireq->tstamp_ok)) {
 699                opts->options |= OPTION_TS;
 700                opts->tsval = TCP_SKB_CB(skb)->when;
 701                opts->tsecr = req->ts_recent;
 702                remaining -= TCPOLEN_TSTAMP_ALIGNED;
 703        }
 704        if (likely(ireq->sack_ok)) {
 705                opts->options |= OPTION_SACK_ADVERTISE;
 706                if (unlikely(!ireq->tstamp_ok))
 707                        remaining -= TCPOLEN_SACKPERM_ALIGNED;
 708        }
 709
 710        /* Similar rationale to tcp_syn_options() applies here, too.
 711         * If the <SYN> options fit, the same options should fit now!
 712         */
 713        if (*md5 == NULL &&
 714            ireq->tstamp_ok &&
 715            cookie_plus > TCPOLEN_COOKIE_BASE) {
 716                int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
 717
 718                if (0x2 & need) {
 719                        /* 32-bit multiple */
 720                        need += 2; /* NOPs */
 721                }
 722                if (need <= remaining) {
 723                        opts->options |= OPTION_COOKIE_EXTENSION;
 724                        opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
 725                        remaining -= need;
 726                } else {
 727                        /* There's no error return, so flag it. */
 728                        xvp->cookie_out_never = 1; /* true */
 729                        opts->hash_size = 0;
 730                }
 731        }
 732        return MAX_TCP_OPTION_SPACE - remaining;
 733}
 734
 735/* Compute TCP options for ESTABLISHED sockets. This is not the
 736 * final wire format yet.
 737 */
 738static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
 739                                        struct tcp_out_options *opts,
 740                                        struct tcp_md5sig_key **md5) {
 741        struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
 742        struct tcp_sock *tp = tcp_sk(sk);
 743        unsigned size = 0;
 744        unsigned int eff_sacks;
 745
 746#ifdef CONFIG_TCP_MD5SIG
 747        *md5 = tp->af_specific->md5_lookup(sk, sk);
 748        if (unlikely(*md5)) {
 749                opts->options |= OPTION_MD5;
 750                size += TCPOLEN_MD5SIG_ALIGNED;
 751        }
 752#else
 753        *md5 = NULL;
 754#endif
 755
 756        if (likely(tp->rx_opt.tstamp_ok)) {
 757                opts->options |= OPTION_TS;
 758                opts->tsval = tcb ? tcb->when : 0;
 759                opts->tsecr = tp->rx_opt.ts_recent;
 760                size += TCPOLEN_TSTAMP_ALIGNED;
 761        }
 762
 763        eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 764        if (unlikely(eff_sacks)) {
 765                const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
 766                opts->num_sack_blocks =
 767                        min_t(unsigned, eff_sacks,
 768                              (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 769                              TCPOLEN_SACK_PERBLOCK);
 770                size += TCPOLEN_SACK_BASE_ALIGNED +
 771                        opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 772        }
 773
 774        return size;
 775}
 776
 777/* This routine actually transmits TCP packets queued in by
 778 * tcp_do_sendmsg().  This is used by both the initial
 779 * transmission and possible later retransmissions.
 780 * All SKB's seen here are completely headerless.  It is our
 781 * job to build the TCP header, and pass the packet down to
 782 * IP so it can do the same plus pass the packet off to the
 783 * device.
 784 *
 785 * We are working here with either a clone of the original
 786 * SKB, or a fresh unique copy made by the retransmit engine.
 787 */
 788static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
 789                            gfp_t gfp_mask)
 790{
 791        const struct inet_connection_sock *icsk = inet_csk(sk);
 792        struct inet_sock *inet;
 793        struct tcp_sock *tp;
 794        struct tcp_skb_cb *tcb;
 795        struct tcp_out_options opts;
 796        unsigned tcp_options_size, tcp_header_size;
 797        struct tcp_md5sig_key *md5;
 798        struct tcphdr *th;
 799        int err;
 800
 801        BUG_ON(!skb || !tcp_skb_pcount(skb));
 802
 803        /* If congestion control is doing timestamping, we must
 804         * take such a timestamp before we potentially clone/copy.
 805         */
 806        if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
 807                __net_timestamp(skb);
 808
 809        if (likely(clone_it)) {
 810                if (unlikely(skb_cloned(skb)))
 811                        skb = pskb_copy(skb, gfp_mask);
 812                else
 813                        skb = skb_clone(skb, gfp_mask);
 814                if (unlikely(!skb))
 815                        return -ENOBUFS;
 816        }
 817
 818        inet = inet_sk(sk);
 819        tp = tcp_sk(sk);
 820        tcb = TCP_SKB_CB(skb);
 821        memset(&opts, 0, sizeof(opts));
 822
 823        if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
 824                tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
 825        else
 826                tcp_options_size = tcp_established_options(sk, skb, &opts,
 827                                                           &md5);
 828        tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
 829
 830        if (tcp_packets_in_flight(tp) == 0)
 831                tcp_ca_event(sk, CA_EVENT_TX_START);
 832
 833        skb_push(skb, tcp_header_size);
 834        skb_reset_transport_header(skb);
 835        skb_set_owner_w(skb, sk);
 836
 837        /* Build TCP header and checksum it. */
 838        th = tcp_hdr(skb);
 839        th->source              = inet->inet_sport;
 840        th->dest                = inet->inet_dport;
 841        th->seq                 = htonl(tcb->seq);
 842        th->ack_seq             = htonl(tp->rcv_nxt);
 843        *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
 844                                        tcb->flags);
 845
 846        if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
 847                /* RFC1323: The window in SYN & SYN/ACK segments
 848                 * is never scaled.
 849                 */
 850                th->window      = htons(min(tp->rcv_wnd, 65535U));
 851        } else {
 852                th->window      = htons(tcp_select_window(sk));
 853        }
 854        th->check               = 0;
 855        th->urg_ptr             = 0;
 856
 857        /* The urg_mode check is necessary during a below snd_una win probe */
 858        if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
 859                if (before(tp->snd_up, tcb->seq + 0x10000)) {
 860                        th->urg_ptr = htons(tp->snd_up - tcb->seq);
 861                        th->urg = 1;
 862                } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
 863                        th->urg_ptr = htons(0xFFFF);
 864                        th->urg = 1;
 865                }
 866        }
 867
 868        tcp_options_write((__be32 *)(th + 1), tp, &opts);
 869        if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
 870                TCP_ECN_send(sk, skb, tcp_header_size);
 871
 872#ifdef CONFIG_TCP_MD5SIG
 873        /* Calculate the MD5 hash, as we have all we need now */
 874        if (md5) {
 875                sk_nocaps_add(sk, NETIF_F_GSO_MASK);
 876                tp->af_specific->calc_md5_hash(opts.hash_location,
 877                                               md5, sk, NULL, skb);
 878        }
 879#endif
 880
 881        icsk->icsk_af_ops->send_check(sk, skb);
 882
 883        if (likely(tcb->flags & TCPCB_FLAG_ACK))
 884                tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
 885
 886        if (skb->len != tcp_header_size)
 887                tcp_event_data_sent(tp, skb, sk);
 888
 889        if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
 890                TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
 891                              tcp_skb_pcount(skb));
 892
 893        err = icsk->icsk_af_ops->queue_xmit(skb);
 894        if (likely(err <= 0))
 895                return err;
 896
 897        tcp_enter_cwr(sk, 1);
 898
 899        return net_xmit_eval(err);
 900}
 901
 902/* This routine just queues the buffer for sending.
 903 *
 904 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
 905 * otherwise socket can stall.
 906 */
 907static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
 908{
 909        struct tcp_sock *tp = tcp_sk(sk);
 910
 911        /* Advance write_seq and place onto the write_queue. */
 912        tp->write_seq = TCP_SKB_CB(skb)->end_seq;
 913        skb_header_release(skb);
 914        tcp_add_write_queue_tail(sk, skb);
 915        sk->sk_wmem_queued += skb->truesize;
 916        sk_mem_charge(sk, skb->truesize);
 917}
 918
 919/* Initialize TSO segments for a packet. */
 920static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
 921                                 unsigned int mss_now)
 922{
 923        if (skb->len <= mss_now || !sk_can_gso(sk) ||
 924            skb->ip_summed == CHECKSUM_NONE) {
 925                /* Avoid the costly divide in the normal
 926                 * non-TSO case.
 927                 */
 928                skb_shinfo(skb)->gso_segs = 1;
 929                skb_shinfo(skb)->gso_size = 0;
 930                skb_shinfo(skb)->gso_type = 0;
 931        } else {
 932                skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
 933                skb_shinfo(skb)->gso_size = mss_now;
 934                skb_shinfo(skb)->gso_type = sk->sk_gso_type;
 935        }
 936}
 937
 938/* When a modification to fackets out becomes necessary, we need to check
 939 * skb is counted to fackets_out or not.
 940 */
 941static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
 942                                   int decr)
 943{
 944        struct tcp_sock *tp = tcp_sk(sk);
 945
 946        if (!tp->sacked_out || tcp_is_reno(tp))
 947                return;
 948
 949        if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
 950                tp->fackets_out -= decr;
 951}
 952
 953/* Pcount in the middle of the write queue got changed, we need to do various
 954 * tweaks to fix counters
 955 */
 956static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
 957{
 958        struct tcp_sock *tp = tcp_sk(sk);
 959
 960        tp->packets_out -= decr;
 961
 962        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
 963                tp->sacked_out -= decr;
 964        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
 965                tp->retrans_out -= decr;
 966        if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
 967                tp->lost_out -= decr;
 968
 969        /* Reno case is special. Sigh... */
 970        if (tcp_is_reno(tp) && decr > 0)
 971                tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
 972
 973        tcp_adjust_fackets_out(sk, skb, decr);
 974
 975        if (tp->lost_skb_hint &&
 976            before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
 977            (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
 978                tp->lost_cnt_hint -= decr;
 979
 980        tcp_verify_left_out(tp);
 981}
 982
 983/* Function to create two new TCP segments.  Shrinks the given segment
 984 * to the specified size and appends a new segment with the rest of the
 985 * packet to the list.  This won't be called frequently, I hope.
 986 * Remember, these are still headerless SKBs at this point.
 987 */
 988int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
 989                 unsigned int mss_now)
 990{
 991        struct tcp_sock *tp = tcp_sk(sk);
 992        struct sk_buff *buff;
 993        int nsize, old_factor;
 994        int nlen;
 995        u8 flags;
 996
 997        BUG_ON(len > skb->len);
 998
 999        nsize = skb_headlen(skb) - len;
1000        if (nsize < 0)
1001                nsize = 0;
1002
1003        if (skb_cloned(skb) &&
1004            skb_is_nonlinear(skb) &&
1005            pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1006                return -ENOMEM;
1007
1008        /* Get a new skb... force flag on. */
1009        buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1010        if (buff == NULL)
1011                return -ENOMEM; /* We'll just try again later. */
1012
1013        sk->sk_wmem_queued += buff->truesize;
1014        sk_mem_charge(sk, buff->truesize);
1015        nlen = skb->len - len - nsize;
1016        buff->truesize += nlen;
1017        skb->truesize -= nlen;
1018
1019        /* Correct the sequence numbers. */
1020        TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1021        TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1022        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1023
1024        /* PSH and FIN should only be set in the second packet. */
1025        flags = TCP_SKB_CB(skb)->flags;
1026        TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1027        TCP_SKB_CB(buff)->flags = flags;
1028        TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1029
1030        if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1031                /* Copy and checksum data tail into the new buffer. */
1032                buff->csum = csum_partial_copy_nocheck(skb->data + len,
1033                                                       skb_put(buff, nsize),
1034                                                       nsize, 0);
1035
1036                skb_trim(skb, len);
1037
1038                skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1039        } else {
1040                skb->ip_summed = CHECKSUM_PARTIAL;
1041                skb_split(skb, buff, len);
1042        }
1043
1044        buff->ip_summed = skb->ip_summed;
1045
1046        /* Looks stupid, but our code really uses when of
1047         * skbs, which it never sent before. --ANK
1048         */
1049        TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1050        buff->tstamp = skb->tstamp;
1051
1052        old_factor = tcp_skb_pcount(skb);
1053
1054        /* Fix up tso_factor for both original and new SKB.  */
1055        tcp_set_skb_tso_segs(sk, skb, mss_now);
1056        tcp_set_skb_tso_segs(sk, buff, mss_now);
1057
1058        /* If this packet has been sent out already, we must
1059         * adjust the various packet counters.
1060         */
1061        if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1062                int diff = old_factor - tcp_skb_pcount(skb) -
1063                        tcp_skb_pcount(buff);
1064
1065                if (diff)
1066                        tcp_adjust_pcount(sk, skb, diff);
1067        }
1068
1069        /* Link BUFF into the send queue. */
1070        skb_header_release(buff);
1071        tcp_insert_write_queue_after(skb, buff, sk);
1072
1073        return 0;
1074}
1075
1076/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1077 * eventually). The difference is that pulled data not copied, but
1078 * immediately discarded.
1079 */
1080static void __pskb_trim_head(struct sk_buff *skb, int len)
1081{
1082        int i, k, eat;
1083
1084        eat = len;
1085        k = 0;
1086        for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1087                if (skb_shinfo(skb)->frags[i].size <= eat) {
1088                        put_page(skb_shinfo(skb)->frags[i].page);
1089                        eat -= skb_shinfo(skb)->frags[i].size;
1090                } else {
1091                        skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1092                        if (eat) {
1093                                skb_shinfo(skb)->frags[k].page_offset += eat;
1094                                skb_shinfo(skb)->frags[k].size -= eat;
1095                                eat = 0;
1096                        }
1097                        k++;
1098                }
1099        }
1100        skb_shinfo(skb)->nr_frags = k;
1101
1102        skb_reset_tail_pointer(skb);
1103        skb->data_len -= len;
1104        skb->len = skb->data_len;
1105}
1106
1107/* Remove acked data from a packet in the transmit queue. */
1108int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1109{
1110        if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1111                return -ENOMEM;
1112
1113        /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1114        if (unlikely(len < skb_headlen(skb)))
1115                __skb_pull(skb, len);
1116        else
1117                __pskb_trim_head(skb, len - skb_headlen(skb));
1118
1119        TCP_SKB_CB(skb)->seq += len;
1120        skb->ip_summed = CHECKSUM_PARTIAL;
1121
1122        skb->truesize        -= len;
1123        sk->sk_wmem_queued   -= len;
1124        sk_mem_uncharge(sk, len);
1125        sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1126
1127        /* Any change of skb->len requires recalculation of tso
1128         * factor and mss.
1129         */
1130        if (tcp_skb_pcount(skb) > 1)
1131                tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1132
1133        return 0;
1134}
1135
1136/* Calculate MSS. Not accounting for SACKs here.  */
1137int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1138{
1139        struct tcp_sock *tp = tcp_sk(sk);
1140        struct inet_connection_sock *icsk = inet_csk(sk);
1141        int mss_now;
1142
1143        /* Calculate base mss without TCP options:
1144           It is MMS_S - sizeof(tcphdr) of rfc1122
1145         */
1146        mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1147
1148        /* Clamp it (mss_clamp does not include tcp options) */
1149        if (mss_now > tp->rx_opt.mss_clamp)
1150                mss_now = tp->rx_opt.mss_clamp;
1151
1152        /* Now subtract optional transport overhead */
1153        mss_now -= icsk->icsk_ext_hdr_len;
1154
1155        /* Then reserve room for full set of TCP options and 8 bytes of data */
1156        if (mss_now < 48)
1157                mss_now = 48;
1158
1159        /* Now subtract TCP options size, not including SACKs */
1160        mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1161
1162        return mss_now;
1163}
1164
1165/* Inverse of above */
1166int tcp_mss_to_mtu(struct sock *sk, int mss)
1167{
1168        struct tcp_sock *tp = tcp_sk(sk);
1169        struct inet_connection_sock *icsk = inet_csk(sk);
1170        int mtu;
1171
1172        mtu = mss +
1173              tp->tcp_header_len +
1174              icsk->icsk_ext_hdr_len +
1175              icsk->icsk_af_ops->net_header_len;
1176
1177        return mtu;
1178}
1179
1180/* MTU probing init per socket */
1181void tcp_mtup_init(struct sock *sk)
1182{
1183        struct tcp_sock *tp = tcp_sk(sk);
1184        struct inet_connection_sock *icsk = inet_csk(sk);
1185
1186        icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1187        icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1188                               icsk->icsk_af_ops->net_header_len;
1189        icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1190        icsk->icsk_mtup.probe_size = 0;
1191}
1192
1193/* This function synchronize snd mss to current pmtu/exthdr set.
1194
1195   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1196   for TCP options, but includes only bare TCP header.
1197
1198   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1199   It is minimum of user_mss and mss received with SYN.
1200   It also does not include TCP options.
1201
1202   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1203
1204   tp->mss_cache is current effective sending mss, including
1205   all tcp options except for SACKs. It is evaluated,
1206   taking into account current pmtu, but never exceeds
1207   tp->rx_opt.mss_clamp.
1208
1209   NOTE1. rfc1122 clearly states that advertised MSS
1210   DOES NOT include either tcp or ip options.
1211
1212   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1213   are READ ONLY outside this function.         --ANK (980731)
1214 */
1215unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1216{
1217        struct tcp_sock *tp = tcp_sk(sk);
1218        struct inet_connection_sock *icsk = inet_csk(sk);
1219        int mss_now;
1220
1221        if (icsk->icsk_mtup.search_high > pmtu)
1222                icsk->icsk_mtup.search_high = pmtu;
1223
1224        mss_now = tcp_mtu_to_mss(sk, pmtu);
1225        mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1226
1227        /* And store cached results */
1228        icsk->icsk_pmtu_cookie = pmtu;
1229        if (icsk->icsk_mtup.enabled)
1230                mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1231        tp->mss_cache = mss_now;
1232
1233        return mss_now;
1234}
1235
1236/* Compute the current effective MSS, taking SACKs and IP options,
1237 * and even PMTU discovery events into account.
1238 */
1239unsigned int tcp_current_mss(struct sock *sk)
1240{
1241        struct tcp_sock *tp = tcp_sk(sk);
1242        struct dst_entry *dst = __sk_dst_get(sk);
1243        u32 mss_now;
1244        unsigned header_len;
1245        struct tcp_out_options opts;
1246        struct tcp_md5sig_key *md5;
1247
1248        mss_now = tp->mss_cache;
1249
1250        if (dst) {
1251                u32 mtu = dst_mtu(dst);
1252                if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1253                        mss_now = tcp_sync_mss(sk, mtu);
1254        }
1255
1256        header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1257                     sizeof(struct tcphdr);
1258        /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1259         * some common options. If this is an odd packet (because we have SACK
1260         * blocks etc) then our calculated header_len will be different, and
1261         * we have to adjust mss_now correspondingly */
1262        if (header_len != tp->tcp_header_len) {
1263                int delta = (int) header_len - tp->tcp_header_len;
1264                mss_now -= delta;
1265        }
1266
1267        return mss_now;
1268}
1269
1270/* Congestion window validation. (RFC2861) */
1271static void tcp_cwnd_validate(struct sock *sk)
1272{
1273        struct tcp_sock *tp = tcp_sk(sk);
1274
1275        if (tp->packets_out >= tp->snd_cwnd) {
1276                /* Network is feed fully. */
1277                tp->snd_cwnd_used = 0;
1278                tp->snd_cwnd_stamp = tcp_time_stamp;
1279        } else {
1280                /* Network starves. */
1281                if (tp->packets_out > tp->snd_cwnd_used)
1282                        tp->snd_cwnd_used = tp->packets_out;
1283
1284                if (sysctl_tcp_slow_start_after_idle &&
1285                    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1286                        tcp_cwnd_application_limited(sk);
1287        }
1288}
1289
1290/* Returns the portion of skb which can be sent right away without
1291 * introducing MSS oddities to segment boundaries. In rare cases where
1292 * mss_now != mss_cache, we will request caller to create a small skb
1293 * per input skb which could be mostly avoided here (if desired).
1294 *
1295 * We explicitly want to create a request for splitting write queue tail
1296 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1297 * thus all the complexity (cwnd_len is always MSS multiple which we
1298 * return whenever allowed by the other factors). Basically we need the
1299 * modulo only when the receiver window alone is the limiting factor or
1300 * when we would be allowed to send the split-due-to-Nagle skb fully.
1301 */
1302static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1303                                        unsigned int mss_now, unsigned int cwnd)
1304{
1305        struct tcp_sock *tp = tcp_sk(sk);
1306        u32 needed, window, cwnd_len;
1307
1308        window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1309        cwnd_len = mss_now * cwnd;
1310
1311        if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1312                return cwnd_len;
1313
1314        needed = min(skb->len, window);
1315
1316        if (cwnd_len <= needed)
1317                return cwnd_len;
1318
1319        return needed - needed % mss_now;
1320}
1321
1322/* Can at least one segment of SKB be sent right now, according to the
1323 * congestion window rules?  If so, return how many segments are allowed.
1324 */
1325static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1326                                         struct sk_buff *skb)
1327{
1328        u32 in_flight, cwnd;
1329
1330        /* Don't be strict about the congestion window for the final FIN.  */
1331        if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1332            tcp_skb_pcount(skb) == 1)
1333                return 1;
1334
1335        in_flight = tcp_packets_in_flight(tp);
1336        cwnd = tp->snd_cwnd;
1337        if (in_flight < cwnd)
1338                return (cwnd - in_flight);
1339
1340        return 0;
1341}
1342
1343/* Intialize TSO state of a skb.
1344 * This must be invoked the first time we consider transmitting
1345 * SKB onto the wire.
1346 */
1347static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1348                             unsigned int mss_now)
1349{
1350        int tso_segs = tcp_skb_pcount(skb);
1351
1352        if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1353                tcp_set_skb_tso_segs(sk, skb, mss_now);
1354                tso_segs = tcp_skb_pcount(skb);
1355        }
1356        return tso_segs;
1357}
1358
1359/* Minshall's variant of the Nagle send check. */
1360static inline int tcp_minshall_check(const struct tcp_sock *tp)
1361{
1362        return after(tp->snd_sml, tp->snd_una) &&
1363                !after(tp->snd_sml, tp->snd_nxt);
1364}
1365
1366/* Return 0, if packet can be sent now without violation Nagle's rules:
1367 * 1. It is full sized.
1368 * 2. Or it contains FIN. (already checked by caller)
1369 * 3. Or TCP_NODELAY was set.
1370 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1371 *    With Minshall's modification: all sent small packets are ACKed.
1372 */
1373static inline int tcp_nagle_check(const struct tcp_sock *tp,
1374                                  const struct sk_buff *skb,
1375                                  unsigned mss_now, int nonagle)
1376{
1377        return (skb->len < mss_now &&
1378                ((nonagle & TCP_NAGLE_CORK) ||
1379                 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1380}
1381
1382/* Return non-zero if the Nagle test allows this packet to be
1383 * sent now.
1384 */
1385static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1386                                 unsigned int cur_mss, int nonagle)
1387{
1388        /* Nagle rule does not apply to frames, which sit in the middle of the
1389         * write_queue (they have no chances to get new data).
1390         *
1391         * This is implemented in the callers, where they modify the 'nonagle'
1392         * argument based upon the location of SKB in the send queue.
1393         */
1394        if (nonagle & TCP_NAGLE_PUSH)
1395                return 1;
1396
1397        /* Don't use the nagle rule for urgent data (or for the final FIN).
1398         * Nagle can be ignored during F-RTO too (see RFC4138).
1399         */
1400        if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1401            (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1402                return 1;
1403
1404        if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1405                return 1;
1406
1407        return 0;
1408}
1409
1410/* Does at least the first segment of SKB fit into the send window? */
1411static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1412                                   unsigned int cur_mss)
1413{
1414        u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1415
1416        if (skb->len > cur_mss)
1417                end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1418
1419        return !after(end_seq, tcp_wnd_end(tp));
1420}
1421
1422/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1423 * should be put on the wire right now.  If so, it returns the number of
1424 * packets allowed by the congestion window.
1425 */
1426static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1427                                 unsigned int cur_mss, int nonagle)
1428{
1429        struct tcp_sock *tp = tcp_sk(sk);
1430        unsigned int cwnd_quota;
1431
1432        tcp_init_tso_segs(sk, skb, cur_mss);
1433
1434        if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1435                return 0;
1436
1437        cwnd_quota = tcp_cwnd_test(tp, skb);
1438        if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1439                cwnd_quota = 0;
1440
1441        return cwnd_quota;
1442}
1443
1444/* Test if sending is allowed right now. */
1445int tcp_may_send_now(struct sock *sk)
1446{
1447        struct tcp_sock *tp = tcp_sk(sk);
1448        struct sk_buff *skb = tcp_send_head(sk);
1449
1450        return (skb &&
1451                tcp_snd_test(sk, skb, tcp_current_mss(sk),
1452                             (tcp_skb_is_last(sk, skb) ?
1453                              tp->nonagle : TCP_NAGLE_PUSH)));
1454}
1455
1456/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1457 * which is put after SKB on the list.  It is very much like
1458 * tcp_fragment() except that it may make several kinds of assumptions
1459 * in order to speed up the splitting operation.  In particular, we
1460 * know that all the data is in scatter-gather pages, and that the
1461 * packet has never been sent out before (and thus is not cloned).
1462 */
1463static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1464                        unsigned int mss_now)
1465{
1466        struct sk_buff *buff;
1467        int nlen = skb->len - len;
1468        u8 flags;
1469
1470        /* All of a TSO frame must be composed of paged data.  */
1471        if (skb->len != skb->data_len)
1472                return tcp_fragment(sk, skb, len, mss_now);
1473
1474        buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1475        if (unlikely(buff == NULL))
1476                return -ENOMEM;
1477
1478        sk->sk_wmem_queued += buff->truesize;
1479        sk_mem_charge(sk, buff->truesize);
1480        buff->truesize += nlen;
1481        skb->truesize -= nlen;
1482
1483        /* Correct the sequence numbers. */
1484        TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1485        TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1486        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1487
1488        /* PSH and FIN should only be set in the second packet. */
1489        flags = TCP_SKB_CB(skb)->flags;
1490        TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1491        TCP_SKB_CB(buff)->flags = flags;
1492
1493        /* This packet was never sent out yet, so no SACK bits. */
1494        TCP_SKB_CB(buff)->sacked = 0;
1495
1496        buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1497        skb_split(skb, buff, len);
1498
1499        /* Fix up tso_factor for both original and new SKB.  */
1500        tcp_set_skb_tso_segs(sk, skb, mss_now);
1501        tcp_set_skb_tso_segs(sk, buff, mss_now);
1502
1503        /* Link BUFF into the send queue. */
1504        skb_header_release(buff);
1505        tcp_insert_write_queue_after(skb, buff, sk);
1506
1507        return 0;
1508}
1509
1510/* Try to defer sending, if possible, in order to minimize the amount
1511 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1512 *
1513 * This algorithm is from John Heffner.
1514 */
1515static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1516{
1517        struct tcp_sock *tp = tcp_sk(sk);
1518        const struct inet_connection_sock *icsk = inet_csk(sk);
1519        u32 send_win, cong_win, limit, in_flight;
1520
1521        if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1522                goto send_now;
1523
1524        if (icsk->icsk_ca_state != TCP_CA_Open)
1525                goto send_now;
1526
1527        /* Defer for less than two clock ticks. */
1528        if (tp->tso_deferred &&
1529            (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1530                goto send_now;
1531
1532        in_flight = tcp_packets_in_flight(tp);
1533
1534        BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1535
1536        send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1537
1538        /* From in_flight test above, we know that cwnd > in_flight.  */
1539        cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1540
1541        limit = min(send_win, cong_win);
1542
1543        /* If a full-sized TSO skb can be sent, do it. */
1544        if (limit >= sk->sk_gso_max_size)
1545                goto send_now;
1546
1547        /* Middle in queue won't get any more data, full sendable already? */
1548        if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1549                goto send_now;
1550
1551        if (sysctl_tcp_tso_win_divisor) {
1552                u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1553
1554                /* If at least some fraction of a window is available,
1555                 * just use it.
1556                 */
1557                chunk /= sysctl_tcp_tso_win_divisor;
1558                if (limit >= chunk)
1559                        goto send_now;
1560        } else {
1561                /* Different approach, try not to defer past a single
1562                 * ACK.  Receiver should ACK every other full sized
1563                 * frame, so if we have space for more than 3 frames
1564                 * then send now.
1565                 */
1566                if (limit > tcp_max_burst(tp) * tp->mss_cache)
1567                        goto send_now;
1568        }
1569
1570        /* Ok, it looks like it is advisable to defer.  */
1571        tp->tso_deferred = 1 | (jiffies << 1);
1572
1573        return 1;
1574
1575send_now:
1576        tp->tso_deferred = 0;
1577        return 0;
1578}
1579
1580/* Create a new MTU probe if we are ready.
1581 * MTU probe is regularly attempting to increase the path MTU by
1582 * deliberately sending larger packets.  This discovers routing
1583 * changes resulting in larger path MTUs.
1584 *
1585 * Returns 0 if we should wait to probe (no cwnd available),
1586 *         1 if a probe was sent,
1587 *         -1 otherwise
1588 */
1589static int tcp_mtu_probe(struct sock *sk)
1590{
1591        struct tcp_sock *tp = tcp_sk(sk);
1592        struct inet_connection_sock *icsk = inet_csk(sk);
1593        struct sk_buff *skb, *nskb, *next;
1594        int len;
1595        int probe_size;
1596        int size_needed;
1597        int copy;
1598        int mss_now;
1599
1600        /* Not currently probing/verifying,
1601         * not in recovery,
1602         * have enough cwnd, and
1603         * not SACKing (the variable headers throw things off) */
1604        if (!icsk->icsk_mtup.enabled ||
1605            icsk->icsk_mtup.probe_size ||
1606            inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1607            tp->snd_cwnd < 11 ||
1608            tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1609                return -1;
1610
1611        /* Very simple search strategy: just double the MSS. */
1612        mss_now = tcp_current_mss(sk);
1613        probe_size = 2 * tp->mss_cache;
1614        size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1615        if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1616                /* TODO: set timer for probe_converge_event */
1617                return -1;
1618        }
1619
1620        /* Have enough data in the send queue to probe? */
1621        if (tp->write_seq - tp->snd_nxt < size_needed)
1622                return -1;
1623
1624        if (tp->snd_wnd < size_needed)
1625                return -1;
1626        if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1627                return 0;
1628
1629        /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1630        if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1631                if (!tcp_packets_in_flight(tp))
1632                        return -1;
1633                else
1634                        return 0;
1635        }
1636
1637        /* We're allowed to probe.  Build it now. */
1638        if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1639                return -1;
1640        sk->sk_wmem_queued += nskb->truesize;
1641        sk_mem_charge(sk, nskb->truesize);
1642
1643        skb = tcp_send_head(sk);
1644
1645        TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1646        TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1647        TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1648        TCP_SKB_CB(nskb)->sacked = 0;
1649        nskb->csum = 0;
1650        nskb->ip_summed = skb->ip_summed;
1651
1652        tcp_insert_write_queue_before(nskb, skb, sk);
1653
1654        len = 0;
1655        tcp_for_write_queue_from_safe(skb, next, sk) {
1656                copy = min_t(int, skb->len, probe_size - len);
1657                if (nskb->ip_summed)
1658                        skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1659                else
1660                        nskb->csum = skb_copy_and_csum_bits(skb, 0,
1661                                                            skb_put(nskb, copy),
1662                                                            copy, nskb->csum);
1663
1664                if (skb->len <= copy) {
1665                        /* We've eaten all the data from this skb.
1666                         * Throw it away. */
1667                        TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1668                        tcp_unlink_write_queue(skb, sk);
1669                        sk_wmem_free_skb(sk, skb);
1670                } else {
1671                        TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1672                                                   ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1673                        if (!skb_shinfo(skb)->nr_frags) {
1674                                skb_pull(skb, copy);
1675                                if (skb->ip_summed != CHECKSUM_PARTIAL)
1676                                        skb->csum = csum_partial(skb->data,
1677                                                                 skb->len, 0);
1678                        } else {
1679                                __pskb_trim_head(skb, copy);
1680                                tcp_set_skb_tso_segs(sk, skb, mss_now);
1681                        }
1682                        TCP_SKB_CB(skb)->seq += copy;
1683                }
1684
1685                len += copy;
1686
1687                if (len >= probe_size)
1688                        break;
1689        }
1690        tcp_init_tso_segs(sk, nskb, nskb->len);
1691
1692        /* We're ready to send.  If this fails, the probe will
1693         * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1694        TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1695        if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1696                /* Decrement cwnd here because we are sending
1697                 * effectively two packets. */
1698                tp->snd_cwnd--;
1699                tcp_event_new_data_sent(sk, nskb);
1700
1701                icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1702                tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1703                tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1704
1705                return 1;
1706        }
1707
1708        return -1;
1709}
1710
1711/* This routine writes packets to the network.  It advances the
1712 * send_head.  This happens as incoming acks open up the remote
1713 * window for us.
1714 *
1715 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1716 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1717 * account rare use of URG, this is not a big flaw.
1718 *
1719 * Returns 1, if no segments are in flight and we have queued segments, but
1720 * cannot send anything now because of SWS or another problem.
1721 */
1722static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1723                          int push_one, gfp_t gfp)
1724{
1725        struct tcp_sock *tp = tcp_sk(sk);
1726        struct sk_buff *skb;
1727        unsigned int tso_segs, sent_pkts;
1728        int cwnd_quota;
1729        int result;
1730
1731        sent_pkts = 0;
1732
1733        if (!push_one) {
1734                /* Do MTU probing. */
1735                result = tcp_mtu_probe(sk);
1736                if (!result) {
1737                        return 0;
1738                } else if (result > 0) {
1739                        sent_pkts = 1;
1740                }
1741        }
1742
1743        while ((skb = tcp_send_head(sk))) {
1744                unsigned int limit;
1745
1746                tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1747                BUG_ON(!tso_segs);
1748
1749                cwnd_quota = tcp_cwnd_test(tp, skb);
1750                if (!cwnd_quota)
1751                        break;
1752
1753                if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1754                        break;
1755
1756                if (tso_segs == 1) {
1757                        if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1758                                                     (tcp_skb_is_last(sk, skb) ?
1759                                                      nonagle : TCP_NAGLE_PUSH))))
1760                                break;
1761                } else {
1762                        if (!push_one && tcp_tso_should_defer(sk, skb))
1763                                break;
1764                }
1765
1766                limit = mss_now;
1767                if (tso_segs > 1 && !tcp_urg_mode(tp))
1768                        limit = tcp_mss_split_point(sk, skb, mss_now,
1769                                                    cwnd_quota);
1770
1771                if (skb->len > limit &&
1772                    unlikely(tso_fragment(sk, skb, limit, mss_now)))
1773                        break;
1774
1775                TCP_SKB_CB(skb)->when = tcp_time_stamp;
1776
1777                if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1778                        break;
1779
1780                /* Advance the send_head.  This one is sent out.
1781                 * This call will increment packets_out.
1782                 */
1783                tcp_event_new_data_sent(sk, skb);
1784
1785                tcp_minshall_update(tp, mss_now, skb);
1786                sent_pkts++;
1787
1788                if (push_one)
1789                        break;
1790        }
1791
1792        if (likely(sent_pkts)) {
1793                tcp_cwnd_validate(sk);
1794                return 0;
1795        }
1796        return !tp->packets_out && tcp_send_head(sk);
1797}
1798
1799/* Push out any pending frames which were held back due to
1800 * TCP_CORK or attempt at coalescing tiny packets.
1801 * The socket must be locked by the caller.
1802 */
1803void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1804                               int nonagle)
1805{
1806        /* If we are closed, the bytes will have to remain here.
1807         * In time closedown will finish, we empty the write queue and
1808         * all will be happy.
1809         */
1810        if (unlikely(sk->sk_state == TCP_CLOSE))
1811                return;
1812
1813        if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1814                tcp_check_probe_timer(sk);
1815}
1816
1817/* Send _single_ skb sitting at the send head. This function requires
1818 * true push pending frames to setup probe timer etc.
1819 */
1820void tcp_push_one(struct sock *sk, unsigned int mss_now)
1821{
1822        struct sk_buff *skb = tcp_send_head(sk);
1823
1824        BUG_ON(!skb || skb->len < mss_now);
1825
1826        tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1827}
1828
1829/* This function returns the amount that we can raise the
1830 * usable window based on the following constraints
1831 *
1832 * 1. The window can never be shrunk once it is offered (RFC 793)
1833 * 2. We limit memory per socket
1834 *
1835 * RFC 1122:
1836 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1837 *  RECV.NEXT + RCV.WIN fixed until:
1838 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1839 *
1840 * i.e. don't raise the right edge of the window until you can raise
1841 * it at least MSS bytes.
1842 *
1843 * Unfortunately, the recommended algorithm breaks header prediction,
1844 * since header prediction assumes th->window stays fixed.
1845 *
1846 * Strictly speaking, keeping th->window fixed violates the receiver
1847 * side SWS prevention criteria. The problem is that under this rule
1848 * a stream of single byte packets will cause the right side of the
1849 * window to always advance by a single byte.
1850 *
1851 * Of course, if the sender implements sender side SWS prevention
1852 * then this will not be a problem.
1853 *
1854 * BSD seems to make the following compromise:
1855 *
1856 *      If the free space is less than the 1/4 of the maximum
1857 *      space available and the free space is less than 1/2 mss,
1858 *      then set the window to 0.
1859 *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1860 *      Otherwise, just prevent the window from shrinking
1861 *      and from being larger than the largest representable value.
1862 *
1863 * This prevents incremental opening of the window in the regime
1864 * where TCP is limited by the speed of the reader side taking
1865 * data out of the TCP receive queue. It does nothing about
1866 * those cases where the window is constrained on the sender side
1867 * because the pipeline is full.
1868 *
1869 * BSD also seems to "accidentally" limit itself to windows that are a
1870 * multiple of MSS, at least until the free space gets quite small.
1871 * This would appear to be a side effect of the mbuf implementation.
1872 * Combining these two algorithms results in the observed behavior
1873 * of having a fixed window size at almost all times.
1874 *
1875 * Below we obtain similar behavior by forcing the offered window to
1876 * a multiple of the mss when it is feasible to do so.
1877 *
1878 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1879 * Regular options like TIMESTAMP are taken into account.
1880 */
1881u32 __tcp_select_window(struct sock *sk)
1882{
1883        struct inet_connection_sock *icsk = inet_csk(sk);
1884        struct tcp_sock *tp = tcp_sk(sk);
1885        /* MSS for the peer's data.  Previous versions used mss_clamp
1886         * here.  I don't know if the value based on our guesses
1887         * of peer's MSS is better for the performance.  It's more correct
1888         * but may be worse for the performance because of rcv_mss
1889         * fluctuations.  --SAW  1998/11/1
1890         */
1891        int mss = icsk->icsk_ack.rcv_mss;
1892        int free_space = tcp_space(sk);
1893        int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1894        int window;
1895
1896        if (mss > full_space)
1897                mss = full_space;
1898
1899        if (free_space < (full_space >> 1)) {
1900                icsk->icsk_ack.quick = 0;
1901
1902                if (tcp_memory_pressure)
1903                        tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1904                                               4U * tp->advmss);
1905
1906                if (free_space < mss)
1907                        return 0;
1908        }
1909
1910        if (free_space > tp->rcv_ssthresh)
1911                free_space = tp->rcv_ssthresh;
1912
1913        /* Don't do rounding if we are using window scaling, since the
1914         * scaled window will not line up with the MSS boundary anyway.
1915         */
1916        window = tp->rcv_wnd;
1917        if (tp->rx_opt.rcv_wscale) {
1918                window = free_space;
1919
1920                /* Advertise enough space so that it won't get scaled away.
1921                 * Import case: prevent zero window announcement if
1922                 * 1<<rcv_wscale > mss.
1923                 */
1924                if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1925                        window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1926                                  << tp->rx_opt.rcv_wscale);
1927        } else {
1928                /* Get the largest window that is a nice multiple of mss.
1929                 * Window clamp already applied above.
1930                 * If our current window offering is within 1 mss of the
1931                 * free space we just keep it. This prevents the divide
1932                 * and multiply from happening most of the time.
1933                 * We also don't do any window rounding when the free space
1934                 * is too small.
1935                 */
1936                if (window <= free_space - mss || window > free_space)
1937                        window = (free_space / mss) * mss;
1938                else if (mss == full_space &&
1939                         free_space > window + (full_space >> 1))
1940                        window = free_space;
1941        }
1942
1943        return window;
1944}
1945
1946/* Collapses two adjacent SKB's during retransmission. */
1947static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1948{
1949        struct tcp_sock *tp = tcp_sk(sk);
1950        struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1951        int skb_size, next_skb_size;
1952
1953        skb_size = skb->len;
1954        next_skb_size = next_skb->len;
1955
1956        BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1957
1958        tcp_highest_sack_combine(sk, next_skb, skb);
1959
1960        tcp_unlink_write_queue(next_skb, sk);
1961
1962        skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1963                                  next_skb_size);
1964
1965        if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1966                skb->ip_summed = CHECKSUM_PARTIAL;
1967
1968        if (skb->ip_summed != CHECKSUM_PARTIAL)
1969                skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1970
1971        /* Update sequence range on original skb. */
1972        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1973
1974        /* Merge over control information. This moves PSH/FIN etc. over */
1975        TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1976
1977        /* All done, get rid of second SKB and account for it so
1978         * packet counting does not break.
1979         */
1980        TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1981
1982        /* changed transmit queue under us so clear hints */
1983        tcp_clear_retrans_hints_partial(tp);
1984        if (next_skb == tp->retransmit_skb_hint)
1985                tp->retransmit_skb_hint = skb;
1986
1987        tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
1988
1989        sk_wmem_free_skb(sk, next_skb);
1990}
1991
1992/* Check if coalescing SKBs is legal. */
1993static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1994{
1995        if (tcp_skb_pcount(skb) > 1)
1996                return 0;
1997        /* TODO: SACK collapsing could be used to remove this condition */
1998        if (skb_shinfo(skb)->nr_frags != 0)
1999                return 0;
2000        if (skb_cloned(skb))
2001                return 0;
2002        if (skb == tcp_send_head(sk))
2003                return 0;
2004        /* Some heurestics for collapsing over SACK'd could be invented */
2005        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2006                return 0;
2007
2008        return 1;
2009}
2010
2011/* Collapse packets in the retransmit queue to make to create
2012 * less packets on the wire. This is only done on retransmission.
2013 */
2014static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2015                                     int space)
2016{
2017        struct tcp_sock *tp = tcp_sk(sk);
2018        struct sk_buff *skb = to, *tmp;
2019        int first = 1;
2020
2021        if (!sysctl_tcp_retrans_collapse)
2022                return;
2023        if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
2024                return;
2025
2026        tcp_for_write_queue_from_safe(skb, tmp, sk) {
2027                if (!tcp_can_collapse(sk, skb))
2028                        break;
2029
2030                space -= skb->len;
2031
2032                if (first) {
2033                        first = 0;
2034                        continue;
2035                }
2036
2037                if (space < 0)
2038                        break;
2039                /* Punt if not enough space exists in the first SKB for
2040                 * the data in the second
2041                 */
2042                if (skb->len > skb_tailroom(to))
2043                        break;
2044
2045                if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2046                        break;
2047
2048                tcp_collapse_retrans(sk, to);
2049        }
2050}
2051
2052/* This retransmits one SKB.  Policy decisions and retransmit queue
2053 * state updates are done by the caller.  Returns non-zero if an
2054 * error occurred which prevented the send.
2055 */
2056int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2057{
2058        struct tcp_sock *tp = tcp_sk(sk);
2059        struct inet_connection_sock *icsk = inet_csk(sk);
2060        unsigned int cur_mss;
2061        int err;
2062
2063        /* Inconslusive MTU probe */
2064        if (icsk->icsk_mtup.probe_size) {
2065                icsk->icsk_mtup.probe_size = 0;
2066        }
2067
2068        /* Do not sent more than we queued. 1/4 is reserved for possible
2069         * copying overhead: fragmentation, tunneling, mangling etc.
2070         */
2071        if (atomic_read(&sk->sk_wmem_alloc) >
2072            min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2073                return -EAGAIN;
2074
2075        if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2076                if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2077                        BUG();
2078                if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2079                        return -ENOMEM;
2080        }
2081
2082        if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2083                return -EHOSTUNREACH; /* Routing failure or similar. */
2084
2085        cur_mss = tcp_current_mss(sk);
2086
2087        /* If receiver has shrunk his window, and skb is out of
2088         * new window, do not retransmit it. The exception is the
2089         * case, when window is shrunk to zero. In this case
2090         * our retransmit serves as a zero window probe.
2091         */
2092        if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2093            TCP_SKB_CB(skb)->seq != tp->snd_una)
2094                return -EAGAIN;
2095
2096        if (skb->len > cur_mss) {
2097                if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2098                        return -ENOMEM; /* We'll try again later. */
2099        } else {
2100                int oldpcount = tcp_skb_pcount(skb);
2101
2102                if (unlikely(oldpcount > 1)) {
2103                        tcp_init_tso_segs(sk, skb, cur_mss);
2104                        tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2105                }
2106        }
2107
2108        tcp_retrans_try_collapse(sk, skb, cur_mss);
2109
2110        /* Some Solaris stacks overoptimize and ignore the FIN on a
2111         * retransmit when old data is attached.  So strip it off
2112         * since it is cheap to do so and saves bytes on the network.
2113         */
2114        if (skb->len > 0 &&
2115            (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
2116            tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2117                if (!pskb_trim(skb, 0)) {
2118                        /* Reuse, even though it does some unnecessary work */
2119                        tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2120                                             TCP_SKB_CB(skb)->flags);
2121                        skb->ip_summed = CHECKSUM_NONE;
2122                }
2123        }
2124
2125        /* Make a copy, if the first transmission SKB clone we made
2126         * is still in somebody's hands, else make a clone.
2127         */
2128        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2129
2130        err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2131
2132        if (err == 0) {
2133                /* Update global TCP statistics. */
2134                TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2135
2136                tp->total_retrans++;
2137
2138#if FASTRETRANS_DEBUG > 0
2139                if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2140                        if (net_ratelimit())
2141                                printk(KERN_DEBUG "retrans_out leaked.\n");
2142                }
2143#endif
2144                if (!tp->retrans_out)
2145                        tp->lost_retrans_low = tp->snd_nxt;
2146                TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2147                tp->retrans_out += tcp_skb_pcount(skb);
2148
2149                /* Save stamp of the first retransmit. */
2150                if (!tp->retrans_stamp)
2151                        tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2152
2153                tp->undo_retrans++;
2154
2155                /* snd_nxt is stored to detect loss of retransmitted segment,
2156                 * see tcp_input.c tcp_sacktag_write_queue().
2157                 */
2158                TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2159        }
2160        return err;
2161}
2162
2163/* Check if we forward retransmits are possible in the current
2164 * window/congestion state.
2165 */
2166static int tcp_can_forward_retransmit(struct sock *sk)
2167{
2168        const struct inet_connection_sock *icsk = inet_csk(sk);
2169        struct tcp_sock *tp = tcp_sk(sk);
2170
2171        /* Forward retransmissions are possible only during Recovery. */
2172        if (icsk->icsk_ca_state != TCP_CA_Recovery)
2173                return 0;
2174
2175        /* No forward retransmissions in Reno are possible. */
2176        if (tcp_is_reno(tp))
2177                return 0;
2178
2179        /* Yeah, we have to make difficult choice between forward transmission
2180         * and retransmission... Both ways have their merits...
2181         *
2182         * For now we do not retransmit anything, while we have some new
2183         * segments to send. In the other cases, follow rule 3 for
2184         * NextSeg() specified in RFC3517.
2185         */
2186
2187        if (tcp_may_send_now(sk))
2188                return 0;
2189
2190        return 1;
2191}
2192
2193/* This gets called after a retransmit timeout, and the initially
2194 * retransmitted data is acknowledged.  It tries to continue
2195 * resending the rest of the retransmit queue, until either
2196 * we've sent it all or the congestion window limit is reached.
2197 * If doing SACK, the first ACK which comes back for a timeout
2198 * based retransmit packet might feed us FACK information again.
2199 * If so, we use it to avoid unnecessarily retransmissions.
2200 */
2201void tcp_xmit_retransmit_queue(struct sock *sk)
2202{
2203        const struct inet_connection_sock *icsk = inet_csk(sk);
2204        struct tcp_sock *tp = tcp_sk(sk);
2205        struct sk_buff *skb;
2206        struct sk_buff *hole = NULL;
2207        u32 last_lost;
2208        int mib_idx;
2209        int fwd_rexmitting = 0;
2210
2211        if (!tp->packets_out)
2212                return;
2213
2214        if (!tp->lost_out)
2215                tp->retransmit_high = tp->snd_una;
2216
2217        if (tp->retransmit_skb_hint) {
2218                skb = tp->retransmit_skb_hint;
2219                last_lost = TCP_SKB_CB(skb)->end_seq;
2220                if (after(last_lost, tp->retransmit_high))
2221                        last_lost = tp->retransmit_high;
2222        } else {
2223                skb = tcp_write_queue_head(sk);
2224                last_lost = tp->snd_una;
2225        }
2226
2227        tcp_for_write_queue_from(skb, sk) {
2228                __u8 sacked = TCP_SKB_CB(skb)->sacked;
2229
2230                if (skb == tcp_send_head(sk))
2231                        break;
2232                /* we could do better than to assign each time */
2233                if (hole == NULL)
2234                        tp->retransmit_skb_hint = skb;
2235
2236                /* Assume this retransmit will generate
2237                 * only one packet for congestion window
2238                 * calculation purposes.  This works because
2239                 * tcp_retransmit_skb() will chop up the
2240                 * packet to be MSS sized and all the
2241                 * packet counting works out.
2242                 */
2243                if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2244                        return;
2245
2246                if (fwd_rexmitting) {
2247begin_fwd:
2248                        if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2249                                break;
2250                        mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2251
2252                } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2253                        tp->retransmit_high = last_lost;
2254                        if (!tcp_can_forward_retransmit(sk))
2255                                break;
2256                        /* Backtrack if necessary to non-L'ed skb */
2257                        if (hole != NULL) {
2258                                skb = hole;
2259                                hole = NULL;
2260                        }
2261                        fwd_rexmitting = 1;
2262                        goto begin_fwd;
2263
2264                } else if (!(sacked & TCPCB_LOST)) {
2265                        if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2266                                hole = skb;
2267                        continue;
2268
2269                } else {
2270                        last_lost = TCP_SKB_CB(skb)->end_seq;
2271                        if (icsk->icsk_ca_state != TCP_CA_Loss)
2272                                mib_idx = LINUX_MIB_TCPFASTRETRANS;
2273                        else
2274                                mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2275                }
2276
2277                if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2278                        continue;
2279
2280                if (tcp_retransmit_skb(sk, skb))
2281                        return;
2282                NET_INC_STATS_BH(sock_net(sk), mib_idx);
2283
2284                if (skb == tcp_write_queue_head(sk))
2285                        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2286                                                  inet_csk(sk)->icsk_rto,
2287                                                  TCP_RTO_MAX);
2288        }
2289}
2290
2291/* Send a fin.  The caller locks the socket for us.  This cannot be
2292 * allowed to fail queueing a FIN frame under any circumstances.
2293 */
2294void tcp_send_fin(struct sock *sk)
2295{
2296        struct tcp_sock *tp = tcp_sk(sk);
2297        struct sk_buff *skb = tcp_write_queue_tail(sk);
2298        int mss_now;
2299
2300        /* Optimization, tack on the FIN if we have a queue of
2301         * unsent frames.  But be careful about outgoing SACKS
2302         * and IP options.
2303         */
2304        mss_now = tcp_current_mss(sk);
2305
2306        if (tcp_send_head(sk) != NULL) {
2307                TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2308                TCP_SKB_CB(skb)->end_seq++;
2309                tp->write_seq++;
2310        } else {
2311                /* Socket is locked, keep trying until memory is available. */
2312                for (;;) {
2313                        skb = alloc_skb_fclone(MAX_TCP_HEADER,
2314                                               sk->sk_allocation);
2315                        if (skb)
2316                                break;
2317                        yield();
2318                }
2319
2320                /* Reserve space for headers and prepare control bits. */
2321                skb_reserve(skb, MAX_TCP_HEADER);
2322                /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2323                tcp_init_nondata_skb(skb, tp->write_seq,
2324                                     TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2325                tcp_queue_skb(sk, skb);
2326        }
2327        __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2328}
2329
2330/* We get here when a process closes a file descriptor (either due to
2331 * an explicit close() or as a byproduct of exit()'ing) and there
2332 * was unread data in the receive queue.  This behavior is recommended
2333 * by RFC 2525, section 2.17.  -DaveM
2334 */
2335void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2336{
2337        struct sk_buff *skb;
2338
2339        /* NOTE: No TCP options attached and we never retransmit this. */
2340        skb = alloc_skb(MAX_TCP_HEADER, priority);
2341        if (!skb) {
2342                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2343                return;
2344        }
2345
2346        /* Reserve space for headers and prepare control bits. */
2347        skb_reserve(skb, MAX_TCP_HEADER);
2348        tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2349                             TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2350        /* Send it off. */
2351        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2352        if (tcp_transmit_skb(sk, skb, 0, priority))
2353                NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2354
2355        TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2356}
2357
2358/* Send a crossed SYN-ACK during socket establishment.
2359 * WARNING: This routine must only be called when we have already sent
2360 * a SYN packet that crossed the incoming SYN that caused this routine
2361 * to get called. If this assumption fails then the initial rcv_wnd
2362 * and rcv_wscale values will not be correct.
2363 */
2364int tcp_send_synack(struct sock *sk)
2365{
2366        struct sk_buff *skb;
2367
2368        skb = tcp_write_queue_head(sk);
2369        if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2370                printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2371                return -EFAULT;
2372        }
2373        if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2374                if (skb_cloned(skb)) {
2375                        struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2376                        if (nskb == NULL)
2377                                return -ENOMEM;
2378                        tcp_unlink_write_queue(skb, sk);
2379                        skb_header_release(nskb);
2380                        __tcp_add_write_queue_head(sk, nskb);
2381                        sk_wmem_free_skb(sk, skb);
2382                        sk->sk_wmem_queued += nskb->truesize;
2383                        sk_mem_charge(sk, nskb->truesize);
2384                        skb = nskb;
2385                }
2386
2387                TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2388                TCP_ECN_send_synack(tcp_sk(sk), skb);
2389        }
2390        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2391        return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2392}
2393
2394/* Prepare a SYN-ACK. */
2395struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2396                                struct request_sock *req,
2397                                struct request_values *rvp)
2398{
2399        struct tcp_out_options opts;
2400        struct tcp_extend_values *xvp = tcp_xv(rvp);
2401        struct inet_request_sock *ireq = inet_rsk(req);
2402        struct tcp_sock *tp = tcp_sk(sk);
2403        const struct tcp_cookie_values *cvp = tp->cookie_values;
2404        struct tcphdr *th;
2405        struct sk_buff *skb;
2406        struct tcp_md5sig_key *md5;
2407        int tcp_header_size;
2408        int mss;
2409        int s_data_desired = 0;
2410
2411        if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2412                s_data_desired = cvp->s_data_desired;
2413        skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2414        if (skb == NULL)
2415                return NULL;
2416
2417        /* Reserve space for headers. */
2418        skb_reserve(skb, MAX_TCP_HEADER);
2419
2420        skb_dst_set(skb, dst_clone(dst));
2421
2422        mss = dst_metric(dst, RTAX_ADVMSS);
2423        if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2424                mss = tp->rx_opt.user_mss;
2425
2426        if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2427                __u8 rcv_wscale;
2428                /* Set this up on the first call only */
2429                req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2430                /* tcp_full_space because it is guaranteed to be the first packet */
2431                tcp_select_initial_window(tcp_full_space(sk),
2432                        mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2433                        &req->rcv_wnd,
2434                        &req->window_clamp,
2435                        ireq->wscale_ok,
2436                        &rcv_wscale,
2437                        dst_metric(dst, RTAX_INITRWND));
2438                ireq->rcv_wscale = rcv_wscale;
2439        }
2440
2441        memset(&opts, 0, sizeof(opts));
2442#ifdef CONFIG_SYN_COOKIES
2443        if (unlikely(req->cookie_ts))
2444                TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2445        else
2446#endif
2447        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2448        tcp_header_size = tcp_synack_options(sk, req, mss,
2449                                             skb, &opts, &md5, xvp)
2450                        + sizeof(*th);
2451
2452        skb_push(skb, tcp_header_size);
2453        skb_reset_transport_header(skb);
2454
2455        th = tcp_hdr(skb);
2456        memset(th, 0, sizeof(struct tcphdr));
2457        th->syn = 1;
2458        th->ack = 1;
2459        TCP_ECN_make_synack(req, th);
2460        th->source = ireq->loc_port;
2461        th->dest = ireq->rmt_port;
2462        /* Setting of flags are superfluous here for callers (and ECE is
2463         * not even correctly set)
2464         */
2465        tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2466                             TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2467
2468        if (OPTION_COOKIE_EXTENSION & opts.options) {
2469                if (s_data_desired) {
2470                        u8 *buf = skb_put(skb, s_data_desired);
2471
2472                        /* copy data directly from the listening socket. */
2473                        memcpy(buf, cvp->s_data_payload, s_data_desired);
2474                        TCP_SKB_CB(skb)->end_seq += s_data_desired;
2475                }
2476
2477                if (opts.hash_size > 0) {
2478                        __u32 workspace[SHA_WORKSPACE_WORDS];
2479                        u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2480                        u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2481
2482                        /* Secret recipe depends on the Timestamp, (future)
2483                         * Sequence and Acknowledgment Numbers, Initiator
2484                         * Cookie, and others handled by IP variant caller.
2485                         */
2486                        *tail-- ^= opts.tsval;
2487                        *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2488                        *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2489
2490                        /* recommended */
2491                        *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2492                        *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2493
2494                        sha_transform((__u32 *)&xvp->cookie_bakery[0],
2495                                      (char *)mess,
2496                                      &workspace[0]);
2497                        opts.hash_location =
2498                                (__u8 *)&xvp->cookie_bakery[0];
2499                }
2500        }
2501
2502        th->seq = htonl(TCP_SKB_CB(skb)->seq);
2503        th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2504
2505        /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2506        th->window = htons(min(req->rcv_wnd, 65535U));
2507        tcp_options_write((__be32 *)(th + 1), tp, &opts);
2508        th->doff = (tcp_header_size >> 2);
2509        TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2510
2511#ifdef CONFIG_TCP_MD5SIG
2512        /* Okay, we have all we need - do the md5 hash if needed */
2513        if (md5) {
2514                tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2515                                               md5, NULL, req, skb);
2516        }
2517#endif
2518
2519        return skb;
2520}
2521
2522/* Do all connect socket setups that can be done AF independent. */
2523static void tcp_connect_init(struct sock *sk)
2524{
2525        struct dst_entry *dst = __sk_dst_get(sk);
2526        struct tcp_sock *tp = tcp_sk(sk);
2527        __u8 rcv_wscale;
2528
2529        /* We'll fix this up when we get a response from the other end.
2530         * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2531         */
2532        tp->tcp_header_len = sizeof(struct tcphdr) +
2533                (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2534
2535#ifdef CONFIG_TCP_MD5SIG
2536        if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2537                tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2538#endif
2539
2540        /* If user gave his TCP_MAXSEG, record it to clamp */
2541        if (tp->rx_opt.user_mss)
2542                tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2543        tp->max_window = 0;
2544        tcp_mtup_init(sk);
2545        tcp_sync_mss(sk, dst_mtu(dst));
2546
2547        if (!tp->window_clamp)
2548                tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2549        tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2550        if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2551                tp->advmss = tp->rx_opt.user_mss;
2552
2553        tcp_initialize_rcv_mss(sk);
2554
2555        tcp_select_initial_window(tcp_full_space(sk),
2556                                  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2557                                  &tp->rcv_wnd,
2558                                  &tp->window_clamp,
2559                                  sysctl_tcp_window_scaling,
2560                                  &rcv_wscale,
2561                                  dst_metric(dst, RTAX_INITRWND));
2562
2563        tp->rx_opt.rcv_wscale = rcv_wscale;
2564        tp->rcv_ssthresh = tp->rcv_wnd;
2565
2566        sk->sk_err = 0;
2567        sock_reset_flag(sk, SOCK_DONE);
2568        tp->snd_wnd = 0;
2569        tcp_init_wl(tp, 0);
2570        tp->snd_una = tp->write_seq;
2571        tp->snd_sml = tp->write_seq;
2572        tp->snd_up = tp->write_seq;
2573        tp->rcv_nxt = 0;
2574        tp->rcv_wup = 0;
2575        tp->copied_seq = 0;
2576
2577        inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2578        inet_csk(sk)->icsk_retransmits = 0;
2579        tcp_clear_retrans(tp);
2580}
2581
2582/* Build a SYN and send it off. */
2583int tcp_connect(struct sock *sk)
2584{
2585        struct tcp_sock *tp = tcp_sk(sk);
2586        struct sk_buff *buff;
2587
2588        tcp_connect_init(sk);
2589
2590        buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2591        if (unlikely(buff == NULL))
2592                return -ENOBUFS;
2593
2594        /* Reserve space for headers. */
2595        skb_reserve(buff, MAX_TCP_HEADER);
2596
2597        tp->snd_nxt = tp->write_seq;
2598        tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2599        TCP_ECN_send_syn(sk, buff);
2600
2601        /* Send it off. */
2602        TCP_SKB_CB(buff)->when = tcp_time_stamp;
2603        tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2604        skb_header_release(buff);
2605        __tcp_add_write_queue_tail(sk, buff);
2606        sk->sk_wmem_queued += buff->truesize;
2607        sk_mem_charge(sk, buff->truesize);
2608        tp->packets_out += tcp_skb_pcount(buff);
2609        tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2610
2611        /* We change tp->snd_nxt after the tcp_transmit_skb() call
2612         * in order to make this packet get counted in tcpOutSegs.
2613         */
2614        tp->snd_nxt = tp->write_seq;
2615        tp->pushed_seq = tp->write_seq;
2616        TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2617
2618        /* Timer for repeating the SYN until an answer. */
2619        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2620                                  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2621        return 0;
2622}
2623
2624/* Send out a delayed ack, the caller does the policy checking
2625 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2626 * for details.
2627 */
2628void tcp_send_delayed_ack(struct sock *sk)
2629{
2630        struct inet_connection_sock *icsk = inet_csk(sk);
2631        int ato = icsk->icsk_ack.ato;
2632        unsigned long timeout;
2633
2634        if (ato > TCP_DELACK_MIN) {
2635                const struct tcp_sock *tp = tcp_sk(sk);
2636                int max_ato = HZ / 2;
2637
2638                if (icsk->icsk_ack.pingpong ||
2639                    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2640                        max_ato = TCP_DELACK_MAX;
2641
2642                /* Slow path, intersegment interval is "high". */
2643
2644                /* If some rtt estimate is known, use it to bound delayed ack.
2645                 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2646                 * directly.
2647                 */
2648                if (tp->srtt) {
2649                        int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2650
2651                        if (rtt < max_ato)
2652                                max_ato = rtt;
2653                }
2654
2655                ato = min(ato, max_ato);
2656        }
2657
2658        /* Stay within the limit we were given */
2659        timeout = jiffies + ato;
2660
2661        /* Use new timeout only if there wasn't a older one earlier. */
2662        if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2663                /* If delack timer was blocked or is about to expire,
2664                 * send ACK now.
2665                 */
2666                if (icsk->icsk_ack.blocked ||
2667                    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2668                        tcp_send_ack(sk);
2669                        return;
2670                }
2671
2672                if (!time_before(timeout, icsk->icsk_ack.timeout))
2673                        timeout = icsk->icsk_ack.timeout;
2674        }
2675        icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2676        icsk->icsk_ack.timeout = timeout;
2677        sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2678}
2679
2680/* This routine sends an ack and also updates the window. */
2681void tcp_send_ack(struct sock *sk)
2682{
2683        struct sk_buff *buff;
2684
2685        /* If we have been reset, we may not send again. */
2686        if (sk->sk_state == TCP_CLOSE)
2687                return;
2688
2689        /* We are not putting this on the write queue, so
2690         * tcp_transmit_skb() will set the ownership to this
2691         * sock.
2692         */
2693        buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2694        if (buff == NULL) {
2695                inet_csk_schedule_ack(sk);
2696                inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2697                inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2698                                          TCP_DELACK_MAX, TCP_RTO_MAX);
2699                return;
2700        }
2701
2702        /* Reserve space for headers and prepare control bits. */
2703        skb_reserve(buff, MAX_TCP_HEADER);
2704        tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2705
2706        /* Send it off, this clears delayed acks for us. */
2707        TCP_SKB_CB(buff)->when = tcp_time_stamp;
2708        tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2709}
2710
2711/* This routine sends a packet with an out of date sequence
2712 * number. It assumes the other end will try to ack it.
2713 *
2714 * Question: what should we make while urgent mode?
2715 * 4.4BSD forces sending single byte of data. We cannot send
2716 * out of window data, because we have SND.NXT==SND.MAX...
2717 *
2718 * Current solution: to send TWO zero-length segments in urgent mode:
2719 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2720 * out-of-date with SND.UNA-1 to probe window.
2721 */
2722static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2723{
2724        struct tcp_sock *tp = tcp_sk(sk);
2725        struct sk_buff *skb;
2726
2727        /* We don't queue it, tcp_transmit_skb() sets ownership. */
2728        skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2729        if (skb == NULL)
2730                return -1;
2731
2732        /* Reserve space for headers and set control bits. */
2733        skb_reserve(skb, MAX_TCP_HEADER);
2734        /* Use a previous sequence.  This should cause the other
2735         * end to send an ack.  Don't queue or clone SKB, just
2736         * send it.
2737         */
2738        tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2739        TCP_SKB_CB(skb)->when = tcp_time_stamp;
2740        return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2741}
2742
2743/* Initiate keepalive or window probe from timer. */
2744int tcp_write_wakeup(struct sock *sk)
2745{
2746        struct tcp_sock *tp = tcp_sk(sk);
2747        struct sk_buff *skb;
2748
2749        if (sk->sk_state == TCP_CLOSE)
2750                return -1;
2751
2752        if ((skb = tcp_send_head(sk)) != NULL &&
2753            before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2754                int err;
2755                unsigned int mss = tcp_current_mss(sk);
2756                unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2757
2758                if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2759                        tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2760
2761                /* We are probing the opening of a window
2762                 * but the window size is != 0
2763                 * must have been a result SWS avoidance ( sender )
2764                 */
2765                if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2766                    skb->len > mss) {
2767                        seg_size = min(seg_size, mss);
2768                        TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2769                        if (tcp_fragment(sk, skb, seg_size, mss))
2770                                return -1;
2771                } else if (!tcp_skb_pcount(skb))
2772                        tcp_set_skb_tso_segs(sk, skb, mss);
2773
2774                TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2775                TCP_SKB_CB(skb)->when = tcp_time_stamp;
2776                err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2777                if (!err)
2778                        tcp_event_new_data_sent(sk, skb);
2779                return err;
2780        } else {
2781                if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2782                        tcp_xmit_probe_skb(sk, 1);
2783                return tcp_xmit_probe_skb(sk, 0);
2784        }
2785}
2786
2787/* A window probe timeout has occurred.  If window is not closed send
2788 * a partial packet else a zero probe.
2789 */
2790void tcp_send_probe0(struct sock *sk)
2791{
2792        struct inet_connection_sock *icsk = inet_csk(sk);
2793        struct tcp_sock *tp = tcp_sk(sk);
2794        int err;
2795
2796        err = tcp_write_wakeup(sk);
2797
2798        if (tp->packets_out || !tcp_send_head(sk)) {
2799                /* Cancel probe timer, if it is not required. */
2800                icsk->icsk_probes_out = 0;
2801                icsk->icsk_backoff = 0;
2802                return;
2803        }
2804
2805        if (err <= 0) {
2806                if (icsk->icsk_backoff < sysctl_tcp_retries2)
2807                        icsk->icsk_backoff++;
2808                icsk->icsk_probes_out++;
2809                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2810                                          min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2811                                          TCP_RTO_MAX);
2812        } else {
2813                /* If packet was not sent due to local congestion,
2814                 * do not backoff and do not remember icsk_probes_out.
2815                 * Let local senders to fight for local resources.
2816                 *
2817                 * Use accumulated backoff yet.
2818                 */
2819                if (!icsk->icsk_probes_out)
2820                        icsk->icsk_probes_out = 1;
2821                inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2822                                          min(icsk->icsk_rto << icsk->icsk_backoff,
2823                                              TCP_RESOURCE_PROBE_INTERVAL),
2824                                          TCP_RTO_MAX);
2825        }
2826}
2827
2828EXPORT_SYMBOL(tcp_select_initial_window);
2829EXPORT_SYMBOL(tcp_connect);
2830EXPORT_SYMBOL(tcp_make_synack);
2831EXPORT_SYMBOL(tcp_simple_retransmit);
2832EXPORT_SYMBOL(tcp_sync_mss);
2833EXPORT_SYMBOL(tcp_mtup_init);
2834