The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/netinet/tcp_output.c

Version: -  FREEBSD  -  FREEBSD11  -  FREEBSD10  -  FREEBSD9  -  FREEBSD92  -  FREEBSD91  -  FREEBSD90  -  FREEBSD8  -  FREEBSD82  -  FREEBSD81  -  FREEBSD80  -  FREEBSD7  -  FREEBSD74  -  FREEBSD73  -  FREEBSD72  -  FREEBSD71  -  FREEBSD70  -  FREEBSD6  -  FREEBSD64  -  FREEBSD63  -  FREEBSD62  -  FREEBSD61  -  FREEBSD60  -  FREEBSD5  -  FREEBSD55  -  FREEBSD54  -  FREEBSD53  -  FREEBSD52  -  FREEBSD51  -  FREEBSD50  -  FREEBSD4  -  FREEBSD3  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
    3  *      The Regents of the University of California.  All rights reserved.
    4  *
    5  * Redistribution and use in source and binary forms, with or without
    6  * modification, are permitted provided that the following conditions
    7  * are met:
    8  * 1. Redistributions of source code must retain the above copyright
    9  *    notice, this list of conditions and the following disclaimer.
   10  * 2. Redistributions in binary form must reproduce the above copyright
   11  *    notice, this list of conditions and the following disclaimer in the
   12  *    documentation and/or other materials provided with the distribution.
   13  * 4. Neither the name of the University nor the names of its contributors
   14  *    may be used to endorse or promote products derived from this software
   15  *    without specific prior written permission.
   16  *
   17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   27  * SUCH DAMAGE.
   28  *
   29  *      @(#)tcp_output.c        8.4 (Berkeley) 5/24/95
   30  */
   31 
   32 #include <sys/cdefs.h>
   33 __FBSDID("$FreeBSD: stable/7/sys/netinet/tcp_output.c 235659 2012-05-19 18:33:28Z bz $");
   34 
   35 #include "opt_inet.h"
   36 #include "opt_inet6.h"
   37 #include "opt_ipsec.h"
   38 #include "opt_mac.h"
   39 #include "opt_tcpdebug.h"
   40 
   41 #include <sys/param.h>
   42 #include <sys/systm.h>
   43 #include <sys/domain.h>
   44 #include <sys/kernel.h>
   45 #include <sys/lock.h>
   46 #include <sys/mbuf.h>
   47 #include <sys/mutex.h>
   48 #include <sys/protosw.h>
   49 #include <sys/socket.h>
   50 #include <sys/socketvar.h>
   51 #include <sys/sysctl.h>
   52 
   53 #include <net/route.h>
   54 
   55 #include <netinet/in.h>
   56 #include <netinet/in_systm.h>
   57 #include <netinet/ip.h>
   58 #include <netinet/in_pcb.h>
   59 #include <netinet/ip_var.h>
   60 #include <netinet/ip_options.h>
   61 #ifdef INET6
   62 #include <netinet6/in6_pcb.h>
   63 #include <netinet/ip6.h>
   64 #include <netinet6/ip6_var.h>
   65 #endif
   66 #include <netinet/tcp.h>
   67 #define TCPOUTFLAGS
   68 #include <netinet/tcp_fsm.h>
   69 #include <netinet/tcp_seq.h>
   70 #include <netinet/tcp_timer.h>
   71 #include <netinet/tcp_var.h>
   72 #include <netinet/tcpip.h>
   73 #ifdef TCPDEBUG
   74 #include <netinet/tcp_debug.h>
   75 #endif
   76 
   77 #ifdef IPSEC
   78 #include <netipsec/ipsec.h>
   79 #endif /*IPSEC*/
   80 
   81 #include <machine/in_cksum.h>
   82 
   83 #include <security/mac/mac_framework.h>
   84 
   85 #ifdef notyet
   86 extern struct mbuf *m_copypack();
   87 #endif
   88 
   89 int path_mtu_discovery = 1;
   90 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
   91         &path_mtu_discovery, 1, "Enable Path MTU Discovery");
   92 
   93 int ss_fltsz = 1;
   94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
   95         &ss_fltsz, 1, "Slow start flight size");
   96 
   97 int ss_fltsz_local = 4;
   98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
   99         &ss_fltsz_local, 1, "Slow start flight size for local networks");
  100 
  101 int     tcp_do_newreno = 1;
  102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
  103         &tcp_do_newreno, 0, "Enable NewReno Algorithms");
  104 
  105 int     tcp_do_tso = 1;
  106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
  107         &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
  108 
  109 int     tcp_do_autosndbuf = 1;
  110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
  111         &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
  112 
  113 int     tcp_autosndbuf_inc = 8*1024;
  114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
  115         &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
  116 
  117 int     tcp_autosndbuf_max = 256*1024;
  118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
  119         &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
  120 
  121 
  122 /*
  123  * Tcp output routine: figure out what should be sent and send it.
  124  */
  125 int
  126 tcp_output(struct tcpcb *tp)
  127 {
  128         struct socket *so = tp->t_inpcb->inp_socket;
  129         long len, recwin, sendwin;
  130         int off, flags, error, rw;
  131         struct mbuf *m;
  132         struct ip *ip = NULL;
  133         struct ipovly *ipov = NULL;
  134         struct tcphdr *th;
  135         u_char opt[TCP_MAXOLEN];
  136         unsigned ipoptlen, optlen, hdrlen;
  137 #ifdef IPSEC
  138         unsigned ipsec_optlen = 0;
  139 #endif
  140         int idle, sendalot;
  141         int sack_rxmit, sack_bytes_rxmt;
  142         struct sackhole *p;
  143         int tso;
  144         struct tcpopt to;
  145 #if 0
  146         int maxburst = TCP_MAXBURST;
  147 #endif
  148 #ifdef INET6
  149         struct ip6_hdr *ip6 = NULL;
  150         int isipv6;
  151 
  152         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
  153 #endif
  154 
  155         INP_WLOCK_ASSERT(tp->t_inpcb);
  156 
  157         /*
  158          * Determine length of data that should be transmitted,
  159          * and flags that will be used.
  160          * If there is some data or critical controls (SYN, RST)
  161          * to send, then transmit; otherwise, investigate further.
  162          */
  163         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
  164         if (idle && (ticks - (int)tp->t_rcvtime) >= tp->t_rxtcur) {
  165                 /*
  166                  * If we've been idle for more than one retransmit
  167                  * timeout the old congestion window is no longer
  168                  * current and we have to reduce it to the restart
  169                  * window before we can transmit again.
  170                  *
  171                  * The restart window is the initial window or the last
  172                  * CWND, whichever is smaller.
  173                  * 
  174                  * This is done to prevent us from flooding the path with
  175                  * a full CWND at wirespeed, overloading router and switch
  176                  * buffers along the way.
  177                  *
  178                  * See RFC5681 Section 4.1. "Restarting Idle Connections".
  179                  */
  180                 if (tcp_do_rfc3390)
  181                         rw = min(4 * tp->t_maxseg,
  182                                  max(2 * tp->t_maxseg, 4380));
  183 #ifdef INET6
  184                 else if ((isipv6 ? in6_localaddr(&tp->t_inpcb->in6p_faddr) :
  185                           in_localaddr(tp->t_inpcb->inp_faddr)))
  186 #else
  187                 else if (in_localaddr(tp->t_inpcb->inp_faddr))
  188 #endif
  189                         rw = ss_fltsz_local * tp->t_maxseg;
  190                 else
  191                         rw = ss_fltsz * tp->t_maxseg;
  192 
  193                 tp->snd_cwnd = min(rw, tp->snd_cwnd);
  194         }
  195         tp->t_flags &= ~TF_LASTIDLE;
  196         if (idle) {
  197                 if (tp->t_flags & TF_MORETOCOME) {
  198                         tp->t_flags |= TF_LASTIDLE;
  199                         idle = 0;
  200                 }
  201         }
  202 again:
  203         /*
  204          * If we've recently taken a timeout, snd_max will be greater than
  205          * snd_nxt.  There may be SACK information that allows us to avoid
  206          * resending already delivered data.  Adjust snd_nxt accordingly.
  207          */
  208         if ((tp->t_flags & TF_SACK_PERMIT) &&
  209             SEQ_LT(tp->snd_nxt, tp->snd_max))
  210                 tcp_sack_adjust(tp);
  211         sendalot = 0;
  212         tso = 0;
  213         off = tp->snd_nxt - tp->snd_una;
  214         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
  215         sendwin = min(sendwin, tp->snd_bwnd);
  216 
  217         flags = tcp_outflags[tp->t_state];
  218         /*
  219          * Send any SACK-generated retransmissions.  If we're explicitly trying
  220          * to send out new data (when sendalot is 1), bypass this function.
  221          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
  222          * we're replacing a (future) new transmission with a retransmission
  223          * now, and we previously incremented snd_cwnd in tcp_input().
  224          */
  225         /*
  226          * Still in sack recovery , reset rxmit flag to zero.
  227          */
  228         sack_rxmit = 0;
  229         sack_bytes_rxmt = 0;
  230         len = 0;
  231         p = NULL;
  232         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
  233             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
  234                 long cwin;
  235                 
  236                 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
  237                 if (cwin < 0)
  238                         cwin = 0;
  239                 /* Do not retransmit SACK segments beyond snd_recover */
  240                 if (SEQ_GT(p->end, tp->snd_recover)) {
  241                         /*
  242                          * (At least) part of sack hole extends beyond
  243                          * snd_recover. Check to see if we can rexmit data
  244                          * for this hole.
  245                          */
  246                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
  247                                 /*
  248                                  * Can't rexmit any more data for this hole.
  249                                  * That data will be rexmitted in the next
  250                                  * sack recovery episode, when snd_recover
  251                                  * moves past p->rxmit.
  252                                  */
  253                                 p = NULL;
  254                                 goto after_sack_rexmit;
  255                         } else
  256                                 /* Can rexmit part of the current hole */
  257                                 len = ((long)ulmin(cwin,
  258                                                    tp->snd_recover - p->rxmit));
  259                 } else
  260                         len = ((long)ulmin(cwin, p->end - p->rxmit));
  261                 off = p->rxmit - tp->snd_una;
  262                 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
  263                     __func__, off));
  264                 if (len > 0) {
  265                         sack_rxmit = 1;
  266                         sendalot = 1;
  267                         tcpstat.tcps_sack_rexmits++;
  268                         tcpstat.tcps_sack_rexmit_bytes +=
  269                             min(len, tp->t_maxseg);
  270                 }
  271         }
  272 after_sack_rexmit:
  273         /*
  274          * Get standard flags, and add SYN or FIN if requested by 'hidden'
  275          * state flags.
  276          */
  277         if (tp->t_flags & TF_NEEDFIN)
  278                 flags |= TH_FIN;
  279         if (tp->t_flags & TF_NEEDSYN)
  280                 flags |= TH_SYN;
  281 
  282         SOCKBUF_LOCK(&so->so_snd);
  283         /*
  284          * If in persist timeout with window of 0, send 1 byte.
  285          * Otherwise, if window is small but nonzero
  286          * and timer expired, we will send what we can
  287          * and go to transmit state.
  288          */
  289         if (tp->t_flags & TF_FORCEDATA) {
  290                 if (sendwin == 0) {
  291                         /*
  292                          * If we still have some data to send, then
  293                          * clear the FIN bit.  Usually this would
  294                          * happen below when it realizes that we
  295                          * aren't sending all the data.  However,
  296                          * if we have exactly 1 byte of unsent data,
  297                          * then it won't clear the FIN bit below,
  298                          * and if we are in persist state, we wind
  299                          * up sending the packet without recording
  300                          * that we sent the FIN bit.
  301                          *
  302                          * We can't just blindly clear the FIN bit,
  303                          * because if we don't have any more data
  304                          * to send then the probe will be the FIN
  305                          * itself.
  306                          */
  307                         if (off < so->so_snd.sb_cc)
  308                                 flags &= ~TH_FIN;
  309                         sendwin = 1;
  310                 } else {
  311                         tcp_timer_activate(tp, TT_PERSIST, 0);
  312                         tp->t_rxtshift = 0;
  313                 }
  314         }
  315 
  316         /*
  317          * If snd_nxt == snd_max and we have transmitted a FIN, the
  318          * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
  319          * a negative length.  This can also occur when TCP opens up
  320          * its congestion window while receiving additional duplicate
  321          * acks after fast-retransmit because TCP will reset snd_nxt
  322          * to snd_max after the fast-retransmit.
  323          *
  324          * In the normal retransmit-FIN-only case, however, snd_nxt will
  325          * be set to snd_una, the offset will be 0, and the length may
  326          * wind up 0.
  327          *
  328          * If sack_rxmit is true we are retransmitting from the scoreboard
  329          * in which case len is already set.
  330          */
  331         if (sack_rxmit == 0) {
  332                 if (sack_bytes_rxmt == 0)
  333                         len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
  334                 else {
  335                         long cwin;
  336 
  337                         /*
  338                          * We are inside of a SACK recovery episode and are
  339                          * sending new data, having retransmitted all the
  340                          * data possible in the scoreboard.
  341                          */
  342                         len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd) 
  343                                - off);
  344                         /*
  345                          * Don't remove this (len > 0) check !
  346                          * We explicitly check for len > 0 here (although it 
  347                          * isn't really necessary), to work around a gcc 
  348                          * optimization issue - to force gcc to compute
  349                          * len above. Without this check, the computation
  350                          * of len is bungled by the optimizer.
  351                          */
  352                         if (len > 0) {
  353                                 cwin = tp->snd_cwnd - 
  354                                         (tp->snd_nxt - tp->sack_newdata) -
  355                                         sack_bytes_rxmt;
  356                                 if (cwin < 0)
  357                                         cwin = 0;
  358                                 len = lmin(len, cwin);
  359                         }
  360                 }
  361         }
  362 
  363         /*
  364          * Lop off SYN bit if it has already been sent.  However, if this
  365          * is SYN-SENT state and if segment contains data and if we don't
  366          * know that foreign host supports TAO, suppress sending segment.
  367          */
  368         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
  369                 if (tp->t_state != TCPS_SYN_RECEIVED)
  370                         flags &= ~TH_SYN;
  371                 off--, len++;
  372         }
  373 
  374         /*
  375          * Be careful not to send data and/or FIN on SYN segments.
  376          * This measure is needed to prevent interoperability problems
  377          * with not fully conformant TCP implementations.
  378          */
  379         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
  380                 len = 0;
  381                 flags &= ~TH_FIN;
  382         }
  383 
  384         if (len < 0) {
  385                 /*
  386                  * If FIN has been sent but not acked,
  387                  * but we haven't been called to retransmit,
  388                  * len will be < 0.  Otherwise, window shrank
  389                  * after we sent into it.  If window shrank to 0,
  390                  * cancel pending retransmit, pull snd_nxt back
  391                  * to (closed) window, and set the persist timer
  392                  * if it isn't already going.  If the window didn't
  393                  * close completely, just wait for an ACK.
  394                  */
  395                 len = 0;
  396                 if (sendwin == 0) {
  397                         tcp_timer_activate(tp, TT_REXMT, 0);
  398                         tp->t_rxtshift = 0;
  399                         tp->snd_nxt = tp->snd_una;
  400                         if (!tcp_timer_active(tp, TT_PERSIST))
  401                                 tcp_setpersist(tp);
  402                 }
  403         }
  404 
  405         /* len will be >= 0 after this point. */
  406         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
  407 
  408         /*
  409          * Automatic sizing of send socket buffer.  Often the send buffer
  410          * size is not optimally adjusted to the actual network conditions
  411          * at hand (delay bandwidth product).  Setting the buffer size too
  412          * small limits throughput on links with high bandwidth and high
  413          * delay (eg. trans-continental/oceanic links).  Setting the
  414          * buffer size too big consumes too much real kernel memory,
  415          * especially with many connections on busy servers.
  416          *
  417          * The criteria to step up the send buffer one notch are:
  418          *  1. receive window of remote host is larger than send buffer
  419          *     (with a fudge factor of 5/4th);
  420          *  2. send buffer is filled to 7/8th with data (so we actually
  421          *     have data to make use of it);
  422          *  3. send buffer fill has not hit maximal automatic size;
  423          *  4. our send window (slow start and cogestion controlled) is
  424          *     larger than sent but unacknowledged data in send buffer.
  425          *
  426          * The remote host receive window scaling factor may limit the
  427          * growing of the send buffer before it reaches its allowed
  428          * maximum.
  429          *
  430          * It scales directly with slow start or congestion window
  431          * and does at most one step per received ACK.  This fast
  432          * scaling has the drawback of growing the send buffer beyond
  433          * what is strictly necessary to make full use of a given
  434          * delay*bandwith product.  However testing has shown this not
  435          * to be much of an problem.  At worst we are trading wasting
  436          * of available bandwith (the non-use of it) for wasting some
  437          * socket buffer memory.
  438          *
  439          * TODO: Shrink send buffer during idle periods together
  440          * with congestion window.  Requires another timer.  Has to
  441          * wait for upcoming tcp timer rewrite.
  442          */
  443         if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
  444                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
  445                     so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
  446                     so->so_snd.sb_cc < tcp_autosndbuf_max &&
  447                     sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
  448                         if (!sbreserve_locked(&so->so_snd,
  449                             min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
  450                              tcp_autosndbuf_max), so, curthread))
  451                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
  452                 }
  453         }
  454 
  455         /*
  456          * Truncate to the maximum segment length or enable TCP Segmentation
  457          * Offloading (if supported by hardware) and ensure that FIN is removed
  458          * if the length no longer contains the last data byte.
  459          *
  460          * TSO may only be used if we are in a pure bulk sending state.  The
  461          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
  462          * IP options prevent using TSO.  With TSO the TCP header is the same
  463          * (except for the sequence number) for all generated packets.  This
  464          * makes it impossible to transmit any options which vary per generated
  465          * segment or packet.
  466          *
  467          * The length of TSO bursts is limited to TCP_MAXWIN.  That limit and
  468          * removal of FIN (if not already catched here) are handled later after
  469          * the exact length of the TCP options are known.
  470          */
  471 #ifdef IPSEC
  472         /*
  473          * Pre-calculate here as we save another lookup into the darknesses
  474          * of IPsec that way and can actually decide if TSO is ok.
  475          */
  476         ipsec_optlen = ipsec_hdrsiz_tcp(tp);
  477 #endif
  478         if (len > tp->t_maxseg) {
  479                 if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
  480                     ((tp->t_flags & TF_SIGNATURE) == 0) &&
  481                     tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
  482                     tp->t_inpcb->inp_options == NULL &&
  483                     tp->t_inpcb->in6p_options == NULL
  484 #ifdef IPSEC
  485                     && ipsec_optlen == 0
  486 #endif
  487                     ) {
  488                         tso = 1;
  489                 } else {
  490                         len = tp->t_maxseg;
  491                         sendalot = 1;
  492                 }
  493         }
  494 
  495         if (sack_rxmit) {
  496                 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
  497                         flags &= ~TH_FIN;
  498         } else {
  499                 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
  500                         flags &= ~TH_FIN;
  501         }
  502 
  503         recwin = sbspace(&so->so_rcv);
  504 
  505         /*
  506          * Sender silly window avoidance.   We transmit under the following
  507          * conditions when len is non-zero:
  508          *
  509          *      - We have a full segment (or more with TSO)
  510          *      - This is the last buffer in a write()/send() and we are
  511          *        either idle or running NODELAY
  512          *      - we've timed out (e.g. persist timer)
  513          *      - we have more then 1/2 the maximum send window's worth of
  514          *        data (receiver may be limited the window size)
  515          *      - we need to retransmit
  516          */
  517         if (len) {
  518                 if (len >= tp->t_maxseg)
  519                         goto send;
  520                 /*
  521                  * NOTE! on localhost connections an 'ack' from the remote
  522                  * end may occur synchronously with the output and cause
  523                  * us to flush a buffer queued with moretocome.  XXX
  524                  *
  525                  * note: the len + off check is almost certainly unnecessary.
  526                  */
  527                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
  528                     (idle || (tp->t_flags & TF_NODELAY)) &&
  529                     len + off >= so->so_snd.sb_cc &&
  530                     (tp->t_flags & TF_NOPUSH) == 0) {
  531                         goto send;
  532                 }
  533                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
  534                         goto send;
  535                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
  536                         goto send;
  537                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
  538                         goto send;
  539                 if (sack_rxmit)
  540                         goto send;
  541         }
  542 
  543         /*
  544          * Compare available window to amount of window
  545          * known to peer (as advertised window less
  546          * next expected input).  If the difference is at least two
  547          * max size segments, or at least 50% of the maximum possible
  548          * window, then want to send a window update to peer.
  549          * Skip this if the connection is in T/TCP half-open state.
  550          * Don't send pure window updates when the peer has closed
  551          * the connection and won't ever send more data.
  552          */
  553         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
  554             !TCPS_HAVERCVDFIN(tp->t_state)) {
  555                 /*
  556                  * "adv" is the amount we can increase the window,
  557                  * taking into account that we are limited by
  558                  * TCP_MAXWIN << tp->rcv_scale.
  559                  */
  560                 long adv;
  561                 int oldwin;
  562 
  563                 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
  564                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
  565                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
  566                         adv -= oldwin;
  567                 } else
  568                         oldwin = 0;
  569 
  570                 /* 
  571                  * If the new window size ends up being the same as the old
  572                  * size when it is scaled, then don't force a window update.
  573                  */
  574                 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
  575                         goto dontupdate;
  576                 if (adv >= (long) (2 * tp->t_maxseg))
  577                         goto send;
  578                 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
  579                         goto send;
  580         }
  581 dontupdate:
  582 
  583         /*
  584          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
  585          * is also a catch-all for the retransmit timer timeout case.
  586          */
  587         if (tp->t_flags & TF_ACKNOW)
  588                 goto send;
  589         if ((flags & TH_RST) ||
  590             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
  591                 goto send;
  592         if (SEQ_GT(tp->snd_up, tp->snd_una))
  593                 goto send;
  594         /*
  595          * If our state indicates that FIN should be sent
  596          * and we have not yet done so, then we need to send.
  597          */
  598         if (flags & TH_FIN &&
  599             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
  600                 goto send;
  601         /*
  602          * In SACK, it is possible for tcp_output to fail to send a segment
  603          * after the retransmission timer has been turned off.  Make sure
  604          * that the retransmission timer is set.
  605          */
  606         if ((tp->t_flags & TF_SACK_PERMIT) &&
  607             SEQ_GT(tp->snd_max, tp->snd_una) &&
  608             !tcp_timer_active(tp, TT_REXMT) &&
  609             !tcp_timer_active(tp, TT_PERSIST)) {
  610                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
  611                 goto just_return;
  612         } 
  613         /*
  614          * TCP window updates are not reliable, rather a polling protocol
  615          * using ``persist'' packets is used to insure receipt of window
  616          * updates.  The three ``states'' for the output side are:
  617          *      idle                    not doing retransmits or persists
  618          *      persisting              to move a small or zero window
  619          *      (re)transmitting        and thereby not persisting
  620          *
  621          * tcp_timer_active(tp, TT_PERSIST)
  622          *      is true when we are in persist state.
  623          * (tp->t_flags & TF_FORCEDATA)
  624          *      is set when we are called to send a persist packet.
  625          * tcp_timer_active(tp, TT_REXMT)
  626          *      is set when we are retransmitting
  627          * The output side is idle when both timers are zero.
  628          *
  629          * If send window is too small, there is data to transmit, and no
  630          * retransmit or persist is pending, then go to persist state.
  631          * If nothing happens soon, send when timer expires:
  632          * if window is nonzero, transmit what we can,
  633          * otherwise force out a byte.
  634          */
  635         if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
  636             !tcp_timer_active(tp, TT_PERSIST)) {
  637                 tp->t_rxtshift = 0;
  638                 tcp_setpersist(tp);
  639         }
  640 
  641         /*
  642          * No reason to send a segment, just return.
  643          */
  644 just_return:
  645         SOCKBUF_UNLOCK(&so->so_snd);
  646         return (0);
  647 
  648 send:
  649         SOCKBUF_LOCK_ASSERT(&so->so_snd);
  650         /*
  651          * Before ESTABLISHED, force sending of initial options
  652          * unless TCP set not to do any options.
  653          * NOTE: we assume that the IP/TCP header plus TCP options
  654          * always fit in a single mbuf, leaving room for a maximum
  655          * link header, i.e.
  656          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
  657          */
  658         optlen = 0;
  659 #ifdef INET6
  660         if (isipv6)
  661                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
  662         else
  663 #endif
  664         hdrlen = sizeof (struct tcpiphdr);
  665 
  666         /*
  667          * Compute options for segment.
  668          * We only have to care about SYN and established connection
  669          * segments.  Options for SYN-ACK segments are handled in TCP
  670          * syncache.
  671          */
  672         if ((tp->t_flags & TF_NOOPT) == 0) {
  673                 to.to_flags = 0;
  674                 /* Maximum segment size. */
  675                 if (flags & TH_SYN) {
  676                         tp->snd_nxt = tp->iss;
  677                         to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
  678                         to.to_flags |= TOF_MSS;
  679                 }
  680                 /* Window scaling. */
  681                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
  682                         to.to_wscale = tp->request_r_scale;
  683                         to.to_flags |= TOF_SCALE;
  684                 }
  685                 /* Timestamps. */
  686                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
  687                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
  688                         to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
  689                         to.to_tsecr = tp->ts_recent;
  690                         to.to_flags |= TOF_TS;
  691                         /* Set receive buffer autosizing timestamp. */
  692                         if (tp->rfbuf_ts == 0 &&
  693                             (so->so_rcv.sb_flags & SB_AUTOSIZE))
  694                                 tp->rfbuf_ts = tcp_ts_getticks();
  695                 }
  696                 /* Selective ACK's. */
  697                 if (tp->t_flags & TF_SACK_PERMIT) {
  698                         if (flags & TH_SYN)
  699                                 to.to_flags |= TOF_SACKPERM;
  700                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
  701                             (tp->t_flags & TF_SACK_PERMIT) &&
  702                             tp->rcv_numsacks > 0) {
  703                                 to.to_flags |= TOF_SACK;
  704                                 to.to_nsacks = tp->rcv_numsacks;
  705                                 to.to_sacks = (u_char *)tp->sackblks;
  706                         }
  707                 }
  708 #ifdef TCP_SIGNATURE
  709                 /* TCP-MD5 (RFC2385). */
  710                 if (tp->t_flags & TF_SIGNATURE)
  711                         to.to_flags |= TOF_SIGNATURE;
  712 #endif /* TCP_SIGNATURE */
  713 
  714                 /* Processing the options. */
  715                 hdrlen += optlen = tcp_addoptions(&to, opt);
  716         }
  717 
  718 #ifdef INET6
  719         if (isipv6)
  720                 ipoptlen = ip6_optlen(tp->t_inpcb);
  721         else
  722 #endif
  723         if (tp->t_inpcb->inp_options)
  724                 ipoptlen = tp->t_inpcb->inp_options->m_len -
  725                                 offsetof(struct ipoption, ipopt_list);
  726         else
  727                 ipoptlen = 0;
  728 #ifdef IPSEC
  729         ipoptlen += ipsec_optlen;
  730 #endif
  731 
  732         /*
  733          * Adjust data length if insertion of options will
  734          * bump the packet length beyond the t_maxopd length.
  735          * Clear the FIN bit because we cut off the tail of
  736          * the segment.
  737          *
  738          * When doing TSO limit a burst to TCP_MAXWIN minus the
  739          * IP, TCP and Options length to keep ip->ip_len from
  740          * overflowing.  Prevent the last segment from being
  741          * fractional thus making them all equal sized and set
  742          * the flag to continue sending.  TSO is disabled when
  743          * IP options or IPSEC are present.
  744          */
  745         if (len + optlen + ipoptlen > tp->t_maxopd) {
  746                 flags &= ~TH_FIN;
  747                 if (tso) {
  748                         if (len > TCP_MAXWIN - hdrlen - optlen) {
  749                                 len = TCP_MAXWIN - hdrlen - optlen;
  750                                 len = len - (len % (tp->t_maxopd - optlen));
  751                                 sendalot = 1;
  752                         } else if (tp->t_flags & TF_NEEDFIN)
  753                                 sendalot = 1;
  754                 } else {
  755                         len = tp->t_maxopd - optlen - ipoptlen;
  756                         sendalot = 1;
  757                 }
  758         }
  759 
  760 /*#ifdef DIAGNOSTIC*/
  761 #ifdef INET6
  762         if (max_linkhdr + hdrlen > MCLBYTES)
  763 #else
  764         if (max_linkhdr + hdrlen > MHLEN)
  765 #endif
  766                 panic("tcphdr too big");
  767 /*#endif*/
  768 
  769         /*
  770          * This KASSERT is here to catch edge cases at a well defined place.
  771          * Before, those had triggered (random) panic conditions further down.
  772          */
  773         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
  774 
  775         /*
  776          * Grab a header mbuf, attaching a copy of data to
  777          * be transmitted, and initialize the header from
  778          * the template for sends on this connection.
  779          */
  780         if (len) {
  781                 struct mbuf *mb;
  782                 u_int moff;
  783 
  784                 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
  785                         tcpstat.tcps_sndprobe++;
  786                 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
  787                         tcpstat.tcps_sndrexmitpack++;
  788                         tcpstat.tcps_sndrexmitbyte += len;
  789                 } else {
  790                         tcpstat.tcps_sndpack++;
  791                         tcpstat.tcps_sndbyte += len;
  792                 }
  793 #ifdef notyet
  794                 if ((m = m_copypack(so->so_snd.sb_mb, off,
  795                     (int)len, max_linkhdr + hdrlen)) == 0) {
  796                         SOCKBUF_UNLOCK(&so->so_snd);
  797                         error = ENOBUFS;
  798                         goto out;
  799                 }
  800                 /*
  801                  * m_copypack left space for our hdr; use it.
  802                  */
  803                 m->m_len += hdrlen;
  804                 m->m_data -= hdrlen;
  805 #else
  806                 MGETHDR(m, M_DONTWAIT, MT_DATA);
  807                 if (m == NULL) {
  808                         SOCKBUF_UNLOCK(&so->so_snd);
  809                         error = ENOBUFS;
  810                         goto out;
  811                 }
  812 #ifdef INET6
  813                 if (MHLEN < hdrlen + max_linkhdr) {
  814                         MCLGET(m, M_DONTWAIT);
  815                         if ((m->m_flags & M_EXT) == 0) {
  816                                 SOCKBUF_UNLOCK(&so->so_snd);
  817                                 m_freem(m);
  818                                 error = ENOBUFS;
  819                                 goto out;
  820                         }
  821                 }
  822 #endif
  823                 m->m_data += max_linkhdr;
  824                 m->m_len = hdrlen;
  825 
  826                 /*
  827                  * Start the m_copy functions from the closest mbuf
  828                  * to the offset in the socket buffer chain.
  829                  */
  830                 mb = sbsndptr(&so->so_snd, off, len, &moff);
  831 
  832                 if (len <= MHLEN - hdrlen - max_linkhdr) {
  833                         m_copydata(mb, moff, (int)len,
  834                             mtod(m, caddr_t) + hdrlen);
  835                         m->m_len += len;
  836                 } else {
  837                         m->m_next = m_copy(mb, moff, (int)len);
  838                         if (m->m_next == NULL) {
  839                                 SOCKBUF_UNLOCK(&so->so_snd);
  840                                 (void) m_free(m);
  841                                 error = ENOBUFS;
  842                                 goto out;
  843                         }
  844                 }
  845 #endif
  846                 /*
  847                  * If we're sending everything we've got, set PUSH.
  848                  * (This will keep happy those implementations which only
  849                  * give data to the user when a buffer fills or
  850                  * a PUSH comes in.)
  851                  */
  852                 if (off + len == so->so_snd.sb_cc)
  853                         flags |= TH_PUSH;
  854                 SOCKBUF_UNLOCK(&so->so_snd);
  855         } else {
  856                 SOCKBUF_UNLOCK(&so->so_snd);
  857                 if (tp->t_flags & TF_ACKNOW)
  858                         tcpstat.tcps_sndacks++;
  859                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
  860                         tcpstat.tcps_sndctrl++;
  861                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
  862                         tcpstat.tcps_sndurg++;
  863                 else
  864                         tcpstat.tcps_sndwinup++;
  865 
  866                 MGETHDR(m, M_DONTWAIT, MT_DATA);
  867                 if (m == NULL) {
  868                         error = ENOBUFS;
  869                         goto out;
  870                 }
  871 #ifdef INET6
  872                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
  873                     MHLEN >= hdrlen) {
  874                         MH_ALIGN(m, hdrlen);
  875                 } else
  876 #endif
  877                 m->m_data += max_linkhdr;
  878                 m->m_len = hdrlen;
  879         }
  880         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
  881         m->m_pkthdr.rcvif = (struct ifnet *)0;
  882 #ifdef MAC
  883         mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
  884 #endif
  885 #ifdef INET6
  886         if (isipv6) {
  887                 ip6 = mtod(m, struct ip6_hdr *);
  888                 th = (struct tcphdr *)(ip6 + 1);
  889                 tcpip_fillheaders(tp->t_inpcb, ip6, th);
  890         } else
  891 #endif /* INET6 */
  892         {
  893                 ip = mtod(m, struct ip *);
  894                 ipov = (struct ipovly *)ip;
  895                 th = (struct tcphdr *)(ip + 1);
  896                 tcpip_fillheaders(tp->t_inpcb, ip, th);
  897         }
  898 
  899         /*
  900          * Fill in fields, remembering maximum advertised
  901          * window for use in delaying messages about window sizes.
  902          * If resending a FIN, be sure not to use a new sequence number.
  903          */
  904         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
  905             tp->snd_nxt == tp->snd_max)
  906                 tp->snd_nxt--;
  907         /*
  908          * If we are doing retransmissions, then snd_nxt will
  909          * not reflect the first unsent octet.  For ACK only
  910          * packets, we do not want the sequence number of the
  911          * retransmitted packet, we want the sequence number
  912          * of the next unsent octet.  So, if there is no data
  913          * (and no SYN or FIN), use snd_max instead of snd_nxt
  914          * when filling in ti_seq.  But if we are in persist
  915          * state, snd_max might reflect one byte beyond the
  916          * right edge of the window, so use snd_nxt in that
  917          * case, since we know we aren't doing a retransmission.
  918          * (retransmit and persist are mutually exclusive...)
  919          */
  920         if (sack_rxmit == 0) {
  921                 if (len || (flags & (TH_SYN|TH_FIN)) ||
  922                     tcp_timer_active(tp, TT_PERSIST))
  923                         th->th_seq = htonl(tp->snd_nxt);
  924                 else
  925                         th->th_seq = htonl(tp->snd_max);
  926         } else {
  927                 th->th_seq = htonl(p->rxmit);
  928                 p->rxmit += len;
  929                 tp->sackhint.sack_bytes_rexmit += len;
  930         }
  931         th->th_ack = htonl(tp->rcv_nxt);
  932         if (optlen) {
  933                 bcopy(opt, th + 1, optlen);
  934                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
  935         }
  936         th->th_flags = flags;
  937         /*
  938          * Calculate receive window.  Don't shrink window,
  939          * but avoid silly window syndrome.
  940          */
  941         if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
  942             recwin < (long)tp->t_maxseg)
  943                 recwin = 0;
  944         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
  945             recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
  946                 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
  947         if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
  948                 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
  949 
  950         /*
  951          * According to RFC1323 the window field in a SYN (i.e., a <SYN>
  952          * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
  953          * case is handled in syncache.
  954          */
  955         if (flags & TH_SYN)
  956                 th->th_win = htons((u_short)
  957                                 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
  958         else
  959                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
  960 
  961         /*
  962          * Adjust the RXWIN0SENT flag - indicate that we have advertised
  963          * a 0 window.  This may cause the remote transmitter to stall.  This
  964          * flag tells soreceive() to disable delayed acknowledgements when
  965          * draining the buffer.  This can occur if the receiver is attempting
  966          * to read more data then can be buffered prior to transmitting on
  967          * the connection.
  968          */
  969         if (th->th_win == 0)
  970                 tp->t_flags |= TF_RXWIN0SENT;
  971         else
  972                 tp->t_flags &= ~TF_RXWIN0SENT;
  973         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
  974                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
  975                 th->th_flags |= TH_URG;
  976         } else
  977                 /*
  978                  * If no urgent pointer to send, then we pull
  979                  * the urgent pointer to the left edge of the send window
  980                  * so that it doesn't drift into the send window on sequence
  981                  * number wraparound.
  982                  */
  983                 tp->snd_up = tp->snd_una;               /* drag it along */
  984 
  985 #ifdef TCP_SIGNATURE
  986         if (tp->t_flags & TF_SIGNATURE) {
  987                 int sigoff = to.to_signature - opt;
  988                 tcp_signature_compute(m, 0, len, optlen,
  989                     (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
  990         }
  991 #endif
  992 
  993         /*
  994          * Put TCP length in extended header, and then
  995          * checksum extended header and data.
  996          */
  997         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
  998 #ifdef INET6
  999         if (isipv6)
 1000                 /*
 1001                  * ip6_plen is not need to be filled now, and will be filled
 1002                  * in ip6_output.
 1003                  */
 1004                 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
 1005                                        sizeof(struct tcphdr) + optlen + len);
 1006         else
 1007 #endif /* INET6 */
 1008         {
 1009                 m->m_pkthdr.csum_flags = CSUM_TCP;
 1010                 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 1011                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
 1012                     htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
 1013 
 1014                 /* IP version must be set here for ipv4/ipv6 checking later */
 1015                 KASSERT(ip->ip_v == IPVERSION,
 1016                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
 1017         }
 1018 
 1019         /*
 1020          * Enable TSO and specify the size of the segments.
 1021          * The TCP pseudo header checksum is always provided.
 1022          * XXX: Fixme: This is currently not the case for IPv6.
 1023          */
 1024         if (tso) {
 1025                 KASSERT(len > tp->t_maxopd - optlen,
 1026                     ("%s: len <= tso_segsz", __func__));
 1027                 m->m_pkthdr.csum_flags |= CSUM_TSO;
 1028                 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
 1029         }
 1030 
 1031         /*
 1032          * In transmit state, time the transmission and arrange for
 1033          * the retransmit.  In persist state, just set snd_max.
 1034          */
 1035         if ((tp->t_flags & TF_FORCEDATA) == 0 || 
 1036             !tcp_timer_active(tp, TT_PERSIST)) {
 1037                 tcp_seq startseq = tp->snd_nxt;
 1038 
 1039                 /*
 1040                  * Advance snd_nxt over sequence space of this segment.
 1041                  */
 1042                 if (flags & (TH_SYN|TH_FIN)) {
 1043                         if (flags & TH_SYN)
 1044                                 tp->snd_nxt++;
 1045                         if (flags & TH_FIN) {
 1046                                 tp->snd_nxt++;
 1047                                 tp->t_flags |= TF_SENTFIN;
 1048                         }
 1049                 }
 1050                 if (sack_rxmit)
 1051                         goto timer;
 1052                 tp->snd_nxt += len;
 1053                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
 1054                         tp->snd_max = tp->snd_nxt;
 1055                         /*
 1056                          * Time this transmission if not a retransmission and
 1057                          * not currently timing anything.
 1058                          */
 1059                         if (tp->t_rtttime == 0) {
 1060                                 tp->t_rtttime = ticks;
 1061                                 tp->t_rtseq = startseq;
 1062                                 tcpstat.tcps_segstimed++;
 1063                         }
 1064                 }
 1065 
 1066                 /*
 1067                  * Set retransmit timer if not currently set,
 1068                  * and not doing a pure ack or a keep-alive probe.
 1069                  * Initial value for retransmit timer is smoothed
 1070                  * round-trip time + 2 * round-trip time variance.
 1071                  * Initialize shift counter which is used for backoff
 1072                  * of retransmit time.
 1073                  */
 1074 timer:
 1075                 if (!tcp_timer_active(tp, TT_REXMT) &&
 1076                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
 1077                      (tp->snd_nxt != tp->snd_una))) {
 1078                         if (tcp_timer_active(tp, TT_PERSIST)) {
 1079                                 tcp_timer_activate(tp, TT_PERSIST, 0);
 1080                                 tp->t_rxtshift = 0;
 1081                         }
 1082                         tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 1083                 }
 1084         } else {
 1085                 /*
 1086                  * Persist case, update snd_max but since we are in
 1087                  * persist mode (no window) we do not update snd_nxt.
 1088                  */
 1089                 int xlen = len;
 1090                 if (flags & TH_SYN)
 1091                         ++xlen;
 1092                 if (flags & TH_FIN) {
 1093                         ++xlen;
 1094                         tp->t_flags |= TF_SENTFIN;
 1095                 }
 1096                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
 1097                         tp->snd_max = tp->snd_nxt + len;
 1098         }
 1099 
 1100 #ifdef TCPDEBUG
 1101         /*
 1102          * Trace.
 1103          */
 1104         if (so->so_options & SO_DEBUG) {
 1105                 u_short save = 0;
 1106 #ifdef INET6
 1107                 if (!isipv6)
 1108 #endif
 1109                 {
 1110                         save = ipov->ih_len;
 1111                         ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
 1112                 }
 1113                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
 1114 #ifdef INET6
 1115                 if (!isipv6)
 1116 #endif
 1117                 ipov->ih_len = save;
 1118         }
 1119 #endif
 1120 
 1121         /*
 1122          * Fill in IP length and desired time to live and
 1123          * send to IP level.  There should be a better way
 1124          * to handle ttl and tos; we could keep them in
 1125          * the template, but need a way to checksum without them.
 1126          */
 1127         /*
 1128          * m->m_pkthdr.len should have been set before cksum calcuration,
 1129          * because in6_cksum() need it.
 1130          */
 1131 #ifdef INET6
 1132         if (isipv6) {
 1133                 /*
 1134                  * we separately set hoplimit for every segment, since the
 1135                  * user might want to change the value via setsockopt.
 1136                  * Also, desired default hop limit might be changed via
 1137                  * Neighbor Discovery.
 1138                  */
 1139                 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
 1140 
 1141                 /* TODO: IPv6 IP6TOS_ECT bit on */
 1142                 error = ip6_output(m,
 1143                             tp->t_inpcb->in6p_outputopts, NULL,
 1144                             ((so->so_options & SO_DONTROUTE) ?
 1145                             IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
 1146         } else
 1147 #endif /* INET6 */
 1148     {
 1149         ip->ip_len = m->m_pkthdr.len;
 1150 #ifdef INET6
 1151         if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
 1152                 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
 1153 #endif /* INET6 */
 1154         /*
 1155          * If we do path MTU discovery, then we set DF on every packet.
 1156          * This might not be the best thing to do according to RFC3390
 1157          * Section 2. However the tcp hostcache migitates the problem
 1158          * so it affects only the first tcp connection with a host.
 1159          *
 1160          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
 1161          */
 1162         if (path_mtu_discovery && tp->t_maxopd > tcp_minmss)
 1163                 ip->ip_off |= IP_DF;
 1164 
 1165         error = ip_output(m, tp->t_inpcb->inp_options, NULL,
 1166             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
 1167             tp->t_inpcb);
 1168     }
 1169         if (error) {
 1170 
 1171                 /*
 1172                  * We know that the packet was lost, so back out the
 1173                  * sequence number advance, if any.
 1174                  *
 1175                  * If the error is EPERM the packet got blocked by the
 1176                  * local firewall.  Normally we should terminate the
 1177                  * connection but the blocking may have been spurious
 1178                  * due to a firewall reconfiguration cycle.  So we treat
 1179                  * it like a packet loss and let the retransmit timer and
 1180                  * timeouts do their work over time.
 1181                  * XXX: It is a POLA question whether calling tcp_drop right
 1182                  * away would be the really correct behavior instead.
 1183                  */
 1184                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
 1185                     !tcp_timer_active(tp, TT_PERSIST)) &&
 1186                     ((flags & TH_SYN) == 0) &&
 1187                     (error != EPERM)) {
 1188                         if (sack_rxmit) {
 1189                                 p->rxmit -= len;
 1190                                 tp->sackhint.sack_bytes_rexmit -= len;
 1191                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
 1192                                     ("sackhint bytes rtx >= 0"));
 1193                         } else
 1194                                 tp->snd_nxt -= len;
 1195                 }
 1196 out:
 1197                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
 1198                 switch (error) {
 1199                 case EPERM:
 1200                         tp->t_softerror = error;
 1201                         return (error);
 1202                 case ENOBUFS:
 1203                         if (!tcp_timer_active(tp, TT_REXMT) &&
 1204                             !tcp_timer_active(tp, TT_PERSIST))
 1205                                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 1206                         tp->snd_cwnd = tp->t_maxseg;
 1207                         return (0);
 1208                 case EMSGSIZE:
 1209                         /*
 1210                          * For some reason the interface we used initially
 1211                          * to send segments changed to another or lowered
 1212                          * its MTU.
 1213                          *
 1214                          * tcp_mtudisc() will find out the new MTU and as
 1215                          * its last action, initiate retransmission, so it
 1216                          * is important to not do so here.
 1217                          *
 1218                          * If TSO was active we either got an interface
 1219                          * without TSO capabilits or TSO was turned off.
 1220                          * Disable it for this connection as too and
 1221                          * immediatly retry with MSS sized segments generated
 1222                          * by this function.
 1223                          */
 1224                         if (tso)
 1225                                 tp->t_flags &= ~TF_TSO;
 1226                         tcp_mtudisc(tp->t_inpcb, 0);
 1227                         return (0);
 1228                 case EHOSTDOWN:
 1229                 case EHOSTUNREACH:
 1230                 case ENETDOWN:
 1231                 case ENETUNREACH:
 1232                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
 1233                                 tp->t_softerror = error;
 1234                                 return (0);
 1235                         }
 1236                         /* FALLTHROUGH */
 1237                 default:
 1238                         return (error);
 1239                 }
 1240         }
 1241         tcpstat.tcps_sndtotal++;
 1242 
 1243         /*
 1244          * Data sent (as far as we can tell).
 1245          * If this advertises a larger window than any other segment,
 1246          * then remember the size of the advertised window.
 1247          * Any pending ACK has now been sent.
 1248          */
 1249         if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
 1250                 tp->rcv_adv = tp->rcv_nxt + recwin;
 1251         tp->last_ack_sent = tp->rcv_nxt;
 1252         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
 1253         if (tcp_timer_active(tp, TT_DELACK))
 1254                 tcp_timer_activate(tp, TT_DELACK, 0);
 1255 #if 0
 1256         /*
 1257          * This completely breaks TCP if newreno is turned on.  What happens
 1258          * is that if delayed-acks are turned on on the receiver, this code
 1259          * on the transmitter effectively destroys the TCP window, forcing
 1260          * it to four packets (1.5Kx4 = 6K window).
 1261          */
 1262         if (sendalot && (!tcp_do_newreno || --maxburst))
 1263                 goto again;
 1264 #endif
 1265         if (sendalot)
 1266                 goto again;
 1267         return (0);
 1268 }
 1269 
 1270 void
 1271 tcp_setpersist(struct tcpcb *tp)
 1272 {
 1273         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
 1274         int tt;
 1275 
 1276         tp->t_flags &= ~TF_PREVVALID;
 1277         if (tcp_timer_active(tp, TT_REXMT))
 1278                 panic("tcp_setpersist: retransmit pending");
 1279         /*
 1280          * Start/restart persistance timer.
 1281          */
 1282         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
 1283                       TCPTV_PERSMIN, TCPTV_PERSMAX);
 1284         tcp_timer_activate(tp, TT_PERSIST, tt);
 1285         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
 1286                 tp->t_rxtshift++;
 1287 }
 1288 
 1289 /*
 1290  * Insert TCP options according to the supplied parameters to the place
 1291  * optp in a consistent way.  Can handle unaligned destinations.
 1292  *
 1293  * The order of the option processing is crucial for optimal packing and
 1294  * alignment for the scarce option space.
 1295  *
 1296  * The optimal order for a SYN/SYN-ACK segment is:
 1297  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
 1298  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
 1299  *
 1300  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
 1301  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
 1302  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
 1303  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
 1304  * we only have 10 bytes for SACK options (40 - (12 + 18)).
 1305  */
 1306 int
 1307 tcp_addoptions(struct tcpopt *to, u_char *optp)
 1308 {
 1309         u_int mask, optlen = 0;
 1310 
 1311         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
 1312                 if ((to->to_flags & mask) != mask)
 1313                         continue;
 1314                 if (optlen == TCP_MAXOLEN)
 1315                         break;
 1316                 switch (to->to_flags & mask) {
 1317                 case TOF_MSS:
 1318                         while (optlen % 4) {
 1319                                 optlen += TCPOLEN_NOP;
 1320                                 *optp++ = TCPOPT_NOP;
 1321                         }
 1322                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
 1323                                 continue;
 1324                         optlen += TCPOLEN_MAXSEG;
 1325                         *optp++ = TCPOPT_MAXSEG;
 1326                         *optp++ = TCPOLEN_MAXSEG;
 1327                         to->to_mss = htons(to->to_mss);
 1328                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
 1329                         optp += sizeof(to->to_mss);
 1330                         break;
 1331                 case TOF_SCALE:
 1332                         while (!optlen || optlen % 2 != 1) {
 1333                                 optlen += TCPOLEN_NOP;
 1334                                 *optp++ = TCPOPT_NOP;
 1335                         }
 1336                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
 1337                                 continue;
 1338                         optlen += TCPOLEN_WINDOW;
 1339                         *optp++ = TCPOPT_WINDOW;
 1340                         *optp++ = TCPOLEN_WINDOW;
 1341                         *optp++ = to->to_wscale;
 1342                         break;
 1343                 case TOF_SACKPERM:
 1344                         while (optlen % 2) {
 1345                                 optlen += TCPOLEN_NOP;
 1346                                 *optp++ = TCPOPT_NOP;
 1347                         }
 1348                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
 1349                                 continue;
 1350                         optlen += TCPOLEN_SACK_PERMITTED;
 1351                         *optp++ = TCPOPT_SACK_PERMITTED;
 1352                         *optp++ = TCPOLEN_SACK_PERMITTED;
 1353                         break;
 1354                 case TOF_TS:
 1355                         while (!optlen || optlen % 4 != 2) {
 1356                                 optlen += TCPOLEN_NOP;
 1357                                 *optp++ = TCPOPT_NOP;
 1358                         }
 1359                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
 1360                                 continue;
 1361                         optlen += TCPOLEN_TIMESTAMP;
 1362                         *optp++ = TCPOPT_TIMESTAMP;
 1363                         *optp++ = TCPOLEN_TIMESTAMP;
 1364                         to->to_tsval = htonl(to->to_tsval);
 1365                         to->to_tsecr = htonl(to->to_tsecr);
 1366                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
 1367                         optp += sizeof(to->to_tsval);
 1368                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
 1369                         optp += sizeof(to->to_tsecr);
 1370                         break;
 1371                 case TOF_SIGNATURE:
 1372                         {
 1373                         int siglen = TCPOLEN_SIGNATURE - 2;
 1374 
 1375                         while (!optlen || optlen % 4 != 2) {
 1376                                 optlen += TCPOLEN_NOP;
 1377                                 *optp++ = TCPOPT_NOP;
 1378                         }
 1379                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
 1380                                 continue;
 1381                         optlen += TCPOLEN_SIGNATURE;
 1382                         *optp++ = TCPOPT_SIGNATURE;
 1383                         *optp++ = TCPOLEN_SIGNATURE;
 1384                         to->to_signature = optp;
 1385                         while (siglen--)
 1386                                  *optp++ = 0;
 1387                         break;
 1388                         }
 1389                 case TOF_SACK:
 1390                         {
 1391                         int sackblks = 0;
 1392                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
 1393                         tcp_seq sack_seq;
 1394 
 1395                         while (!optlen || optlen % 4 != 2) {
 1396                                 optlen += TCPOLEN_NOP;
 1397                                 *optp++ = TCPOPT_NOP;
 1398                         }
 1399                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
 1400                                 continue;
 1401                         optlen += TCPOLEN_SACKHDR;
 1402                         *optp++ = TCPOPT_SACK;
 1403                         sackblks = min(to->to_nsacks,
 1404                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
 1405                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
 1406                         while (sackblks--) {
 1407                                 sack_seq = htonl(sack->start);
 1408                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
 1409                                 optp += sizeof(sack_seq);
 1410                                 sack_seq = htonl(sack->end);
 1411                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
 1412                                 optp += sizeof(sack_seq);
 1413                                 optlen += TCPOLEN_SACK;
 1414                                 sack++;
 1415                         }
 1416                         tcpstat.tcps_sack_send_blocks++;
 1417                         break;
 1418                         }
 1419                 default:
 1420                         panic("%s: unknown TCP option type", __func__);
 1421                         break;
 1422                 }
 1423         }
 1424 
 1425         /* Terminate and pad TCP options to a 4 byte boundary. */
 1426         if (optlen % 4) {
 1427                 optlen += TCPOLEN_EOL;
 1428                 *optp++ = TCPOPT_EOL;
 1429         }
 1430         /*
 1431          * According to RFC 793 (STD0007):
 1432          *   "The content of the header beyond the End-of-Option option
 1433          *    must be header padding (i.e., zero)."
 1434          *   and later: "The padding is composed of zeros."
 1435          */
 1436         while (optlen % 4) {
 1437                 optlen += TCPOLEN_PAD;
 1438                 *optp++ = TCPOPT_PAD;
 1439         }
 1440 
 1441         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
 1442         return (optlen);
 1443 }

Cache object: b4386c82cff6695d972174b5e3da3fa7


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.