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("$FreeBSD$");
34
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/hhook.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/if.h>
54 #include <net/route.h>
55 #include <net/vnet.h>
56
57 #include <netinet/cc.h>
58 #include <netinet/in.h>
59 #include <netinet/in_systm.h>
60 #include <netinet/ip.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet/ip_var.h>
63 #include <netinet/ip_options.h>
64 #ifdef INET6
65 #include <netinet6/in6_pcb.h>
66 #include <netinet/ip6.h>
67 #include <netinet6/ip6_var.h>
68 #endif
69 #define TCPOUTFLAGS
70 #include <netinet/tcp_fsm.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #ifdef TCPDEBUG
76 #include <netinet/tcp_debug.h>
77 #endif
78
79 #ifdef IPSEC
80 #include <netipsec/ipsec.h>
81 #endif /*IPSEC*/
82
83 #include <machine/in_cksum.h>
84
85 #include <security/mac/mac_framework.h>
86
87 #ifdef notyet
88 extern struct mbuf *m_copypack();
89 #endif
90
91 VNET_DEFINE(int, path_mtu_discovery) = 1;
92 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
93 &VNET_NAME(path_mtu_discovery), 1,
94 "Enable Path MTU Discovery");
95
96 VNET_DEFINE(int, ss_fltsz) = 1;
97 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
98 &VNET_NAME(ss_fltsz), 1,
99 "Slow start flight size");
100
101 VNET_DEFINE(int, ss_fltsz_local) = 4;
102 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
103 CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
104 "Slow start flight size for local networks");
105
106 VNET_DEFINE(int, tcp_do_tso) = 1;
107 #define V_tcp_do_tso VNET(tcp_do_tso)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
109 &VNET_NAME(tcp_do_tso), 0,
110 "Enable TCP Segmentation Offload");
111
112 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
113 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
115 &VNET_NAME(tcp_do_autosndbuf), 0,
116 "Enable automatic send buffer sizing");
117
118 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
119 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
120 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
121 &VNET_NAME(tcp_autosndbuf_inc), 0,
122 "Incrementor step size of automatic send buffer");
123
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024;
125 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
126 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
127 &VNET_NAME(tcp_autosndbuf_max), 0,
128 "Max size of automatic send buffer");
129
130 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
131 struct tcphdr *th, struct tcpopt *to,
132 long len, int tso);
133 static void inline cc_after_idle(struct tcpcb *tp);
134
135 /*
136 * Wrapper for the TCP established ouput helper hook.
137 */
138 static void inline
139 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
140 struct tcpopt *to, long len, int tso)
141 {
142 struct tcp_hhook_data hhook_data;
143
144 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
145 hhook_data.tp = tp;
146 hhook_data.th = th;
147 hhook_data.to = to;
148 hhook_data.len = len;
149 hhook_data.tso = tso;
150
151 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
152 tp->osd);
153 }
154 }
155
156 /*
157 * CC wrapper hook functions
158 */
159 static void inline
160 cc_after_idle(struct tcpcb *tp)
161 {
162 INP_WLOCK_ASSERT(tp->t_inpcb);
163
164 if (CC_ALGO(tp)->after_idle != NULL)
165 CC_ALGO(tp)->after_idle(tp->ccv);
166 }
167
168 /*
169 * Tcp output routine: figure out what should be sent and send it.
170 */
171 int
172 tcp_output(struct tcpcb *tp)
173 {
174 struct socket *so = tp->t_inpcb->inp_socket;
175 long len, recwin, sendwin;
176 int off, flags, error;
177 struct mbuf *m;
178 struct ip *ip = NULL;
179 struct ipovly *ipov = NULL;
180 struct tcphdr *th;
181 u_char opt[TCP_MAXOLEN];
182 unsigned ipoptlen, optlen, hdrlen;
183 #ifdef IPSEC
184 unsigned ipsec_optlen = 0;
185 #endif
186 int idle, sendalot;
187 int sack_rxmit, sack_bytes_rxmt;
188 struct sackhole *p;
189 int tso;
190 struct tcpopt to;
191 #if 0
192 int maxburst = TCP_MAXBURST;
193 #endif
194 #ifdef INET6
195 struct ip6_hdr *ip6 = NULL;
196 int isipv6;
197
198 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
199 #endif
200
201 INP_WLOCK_ASSERT(tp->t_inpcb);
202
203 /*
204 * Determine length of data that should be transmitted,
205 * and flags that will be used.
206 * If there is some data or critical controls (SYN, RST)
207 * to send, then transmit; otherwise, investigate further.
208 */
209 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
210 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
211 cc_after_idle(tp);
212 tp->t_flags &= ~TF_LASTIDLE;
213 if (idle) {
214 if (tp->t_flags & TF_MORETOCOME) {
215 tp->t_flags |= TF_LASTIDLE;
216 idle = 0;
217 }
218 }
219 again:
220 /*
221 * If we've recently taken a timeout, snd_max will be greater than
222 * snd_nxt. There may be SACK information that allows us to avoid
223 * resending already delivered data. Adjust snd_nxt accordingly.
224 */
225 if ((tp->t_flags & TF_SACK_PERMIT) &&
226 SEQ_LT(tp->snd_nxt, tp->snd_max))
227 tcp_sack_adjust(tp);
228 sendalot = 0;
229 tso = 0;
230 off = tp->snd_nxt - tp->snd_una;
231 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
232 sendwin = min(sendwin, tp->snd_bwnd);
233
234 flags = tcp_outflags[tp->t_state];
235 /*
236 * Send any SACK-generated retransmissions. If we're explicitly trying
237 * to send out new data (when sendalot is 1), bypass this function.
238 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
239 * we're replacing a (future) new transmission with a retransmission
240 * now, and we previously incremented snd_cwnd in tcp_input().
241 */
242 /*
243 * Still in sack recovery , reset rxmit flag to zero.
244 */
245 sack_rxmit = 0;
246 sack_bytes_rxmt = 0;
247 len = 0;
248 p = NULL;
249 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
250 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
251 long cwin;
252
253 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
254 if (cwin < 0)
255 cwin = 0;
256 /* Do not retransmit SACK segments beyond snd_recover */
257 if (SEQ_GT(p->end, tp->snd_recover)) {
258 /*
259 * (At least) part of sack hole extends beyond
260 * snd_recover. Check to see if we can rexmit data
261 * for this hole.
262 */
263 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
264 /*
265 * Can't rexmit any more data for this hole.
266 * That data will be rexmitted in the next
267 * sack recovery episode, when snd_recover
268 * moves past p->rxmit.
269 */
270 p = NULL;
271 goto after_sack_rexmit;
272 } else
273 /* Can rexmit part of the current hole */
274 len = ((long)ulmin(cwin,
275 tp->snd_recover - p->rxmit));
276 } else
277 len = ((long)ulmin(cwin, p->end - p->rxmit));
278 off = p->rxmit - tp->snd_una;
279 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
280 __func__, off));
281 if (len > 0) {
282 sack_rxmit = 1;
283 sendalot = 1;
284 TCPSTAT_INC(tcps_sack_rexmits);
285 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
286 min(len, tp->t_maxseg));
287 }
288 }
289 after_sack_rexmit:
290 /*
291 * Get standard flags, and add SYN or FIN if requested by 'hidden'
292 * state flags.
293 */
294 if (tp->t_flags & TF_NEEDFIN)
295 flags |= TH_FIN;
296 if (tp->t_flags & TF_NEEDSYN)
297 flags |= TH_SYN;
298
299 SOCKBUF_LOCK(&so->so_snd);
300 /*
301 * If in persist timeout with window of 0, send 1 byte.
302 * Otherwise, if window is small but nonzero
303 * and timer expired, we will send what we can
304 * and go to transmit state.
305 */
306 if (tp->t_flags & TF_FORCEDATA) {
307 if (sendwin == 0) {
308 /*
309 * If we still have some data to send, then
310 * clear the FIN bit. Usually this would
311 * happen below when it realizes that we
312 * aren't sending all the data. However,
313 * if we have exactly 1 byte of unsent data,
314 * then it won't clear the FIN bit below,
315 * and if we are in persist state, we wind
316 * up sending the packet without recording
317 * that we sent the FIN bit.
318 *
319 * We can't just blindly clear the FIN bit,
320 * because if we don't have any more data
321 * to send then the probe will be the FIN
322 * itself.
323 */
324 if (off < so->so_snd.sb_cc)
325 flags &= ~TH_FIN;
326 sendwin = 1;
327 } else {
328 tcp_timer_activate(tp, TT_PERSIST, 0);
329 tp->t_rxtshift = 0;
330 }
331 }
332
333 /*
334 * If snd_nxt == snd_max and we have transmitted a FIN, the
335 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
336 * a negative length. This can also occur when TCP opens up
337 * its congestion window while receiving additional duplicate
338 * acks after fast-retransmit because TCP will reset snd_nxt
339 * to snd_max after the fast-retransmit.
340 *
341 * In the normal retransmit-FIN-only case, however, snd_nxt will
342 * be set to snd_una, the offset will be 0, and the length may
343 * wind up 0.
344 *
345 * If sack_rxmit is true we are retransmitting from the scoreboard
346 * in which case len is already set.
347 */
348 if (sack_rxmit == 0) {
349 if (sack_bytes_rxmt == 0)
350 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
351 else {
352 long cwin;
353
354 /*
355 * We are inside of a SACK recovery episode and are
356 * sending new data, having retransmitted all the
357 * data possible in the scoreboard.
358 */
359 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
360 - off);
361 /*
362 * Don't remove this (len > 0) check !
363 * We explicitly check for len > 0 here (although it
364 * isn't really necessary), to work around a gcc
365 * optimization issue - to force gcc to compute
366 * len above. Without this check, the computation
367 * of len is bungled by the optimizer.
368 */
369 if (len > 0) {
370 cwin = tp->snd_cwnd -
371 (tp->snd_nxt - tp->sack_newdata) -
372 sack_bytes_rxmt;
373 if (cwin < 0)
374 cwin = 0;
375 len = lmin(len, cwin);
376 }
377 }
378 }
379
380 /*
381 * Lop off SYN bit if it has already been sent. However, if this
382 * is SYN-SENT state and if segment contains data and if we don't
383 * know that foreign host supports TAO, suppress sending segment.
384 */
385 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
386 if (tp->t_state != TCPS_SYN_RECEIVED)
387 flags &= ~TH_SYN;
388 off--, len++;
389 }
390
391 /*
392 * Be careful not to send data and/or FIN on SYN segments.
393 * This measure is needed to prevent interoperability problems
394 * with not fully conformant TCP implementations.
395 */
396 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
397 len = 0;
398 flags &= ~TH_FIN;
399 }
400
401 if (len <= 0) {
402 /*
403 * If FIN has been sent but not acked,
404 * but we haven't been called to retransmit,
405 * len will be < 0. Otherwise, window shrank
406 * after we sent into it. If window shrank to 0,
407 * cancel pending retransmit, pull snd_nxt back
408 * to (closed) window, and set the persist timer
409 * if it isn't already going. If the window didn't
410 * close completely, just wait for an ACK.
411 *
412 * We also do a general check here to ensure that
413 * we will set the persist timer when we have data
414 * to send, but a 0-byte window. This makes sure
415 * the persist timer is set even if the packet
416 * hits one of the "goto send" lines below.
417 */
418 len = 0;
419 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
420 (off < (int) so->so_snd.sb_cc)) {
421 tcp_timer_activate(tp, TT_REXMT, 0);
422 tp->t_rxtshift = 0;
423 tp->snd_nxt = tp->snd_una;
424 if (!tcp_timer_active(tp, TT_PERSIST))
425 tcp_setpersist(tp);
426 }
427 }
428
429 /* len will be >= 0 after this point. */
430 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
431
432 /*
433 * Automatic sizing of send socket buffer. Often the send buffer
434 * size is not optimally adjusted to the actual network conditions
435 * at hand (delay bandwidth product). Setting the buffer size too
436 * small limits throughput on links with high bandwidth and high
437 * delay (eg. trans-continental/oceanic links). Setting the
438 * buffer size too big consumes too much real kernel memory,
439 * especially with many connections on busy servers.
440 *
441 * The criteria to step up the send buffer one notch are:
442 * 1. receive window of remote host is larger than send buffer
443 * (with a fudge factor of 5/4th);
444 * 2. send buffer is filled to 7/8th with data (so we actually
445 * have data to make use of it);
446 * 3. send buffer fill has not hit maximal automatic size;
447 * 4. our send window (slow start and cogestion controlled) is
448 * larger than sent but unacknowledged data in send buffer.
449 *
450 * The remote host receive window scaling factor may limit the
451 * growing of the send buffer before it reaches its allowed
452 * maximum.
453 *
454 * It scales directly with slow start or congestion window
455 * and does at most one step per received ACK. This fast
456 * scaling has the drawback of growing the send buffer beyond
457 * what is strictly necessary to make full use of a given
458 * delay*bandwith product. However testing has shown this not
459 * to be much of an problem. At worst we are trading wasting
460 * of available bandwith (the non-use of it) for wasting some
461 * socket buffer memory.
462 *
463 * TODO: Shrink send buffer during idle periods together
464 * with congestion window. Requires another timer. Has to
465 * wait for upcoming tcp timer rewrite.
466 */
467 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
468 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
469 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
470 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
471 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
472 if (!sbreserve_locked(&so->so_snd,
473 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
474 V_tcp_autosndbuf_max), so, curthread))
475 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
476 }
477 }
478
479 /*
480 * Truncate to the maximum segment length or enable TCP Segmentation
481 * Offloading (if supported by hardware) and ensure that FIN is removed
482 * if the length no longer contains the last data byte.
483 *
484 * TSO may only be used if we are in a pure bulk sending state. The
485 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
486 * IP options prevent using TSO. With TSO the TCP header is the same
487 * (except for the sequence number) for all generated packets. This
488 * makes it impossible to transmit any options which vary per generated
489 * segment or packet.
490 *
491 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
492 * removal of FIN (if not already catched here) are handled later after
493 * the exact length of the TCP options are known.
494 */
495 #ifdef IPSEC
496 /*
497 * Pre-calculate here as we save another lookup into the darknesses
498 * of IPsec that way and can actually decide if TSO is ok.
499 */
500 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
501 #endif
502 if (len > tp->t_maxseg) {
503 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
504 ((tp->t_flags & TF_SIGNATURE) == 0) &&
505 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
506 tp->t_inpcb->inp_options == NULL &&
507 tp->t_inpcb->in6p_options == NULL
508 #ifdef IPSEC
509 && ipsec_optlen == 0
510 #endif
511 ) {
512 tso = 1;
513 } else {
514 len = tp->t_maxseg;
515 sendalot = 1;
516 }
517 }
518
519 if (sack_rxmit) {
520 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
521 flags &= ~TH_FIN;
522 } else {
523 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
524 flags &= ~TH_FIN;
525 }
526
527 recwin = sbspace(&so->so_rcv);
528
529 /*
530 * Sender silly window avoidance. We transmit under the following
531 * conditions when len is non-zero:
532 *
533 * - We have a full segment (or more with TSO)
534 * - This is the last buffer in a write()/send() and we are
535 * either idle or running NODELAY
536 * - we've timed out (e.g. persist timer)
537 * - we have more then 1/2 the maximum send window's worth of
538 * data (receiver may be limited the window size)
539 * - we need to retransmit
540 */
541 if (len) {
542 if (len >= tp->t_maxseg)
543 goto send;
544 /*
545 * NOTE! on localhost connections an 'ack' from the remote
546 * end may occur synchronously with the output and cause
547 * us to flush a buffer queued with moretocome. XXX
548 *
549 * note: the len + off check is almost certainly unnecessary.
550 */
551 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
552 (idle || (tp->t_flags & TF_NODELAY)) &&
553 len + off >= so->so_snd.sb_cc &&
554 (tp->t_flags & TF_NOPUSH) == 0) {
555 goto send;
556 }
557 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
558 goto send;
559 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
560 goto send;
561 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
562 goto send;
563 if (sack_rxmit)
564 goto send;
565 }
566
567 /*
568 * Compare available window to amount of window
569 * known to peer (as advertised window less
570 * next expected input). If the difference is at least two
571 * max size segments, or at least 50% of the maximum possible
572 * window, then want to send a window update to peer.
573 * Skip this if the connection is in T/TCP half-open state.
574 * Don't send pure window updates when the peer has closed
575 * the connection and won't ever send more data.
576 */
577 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
578 !TCPS_HAVERCVDFIN(tp->t_state)) {
579 /*
580 * "adv" is the amount we can increase the window,
581 * taking into account that we are limited by
582 * TCP_MAXWIN << tp->rcv_scale.
583 */
584 long adv;
585 int oldwin;
586
587 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
588 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
589 oldwin = (tp->rcv_adv - tp->rcv_nxt);
590 adv -= oldwin;
591 } else
592 oldwin = 0;
593
594 /*
595 * If the new window size ends up being the same as the old
596 * size when it is scaled, then don't force a window update.
597 */
598 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
599 goto dontupdate;
600 if (adv >= (long) (2 * tp->t_maxseg))
601 goto send;
602 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
603 goto send;
604 }
605 dontupdate:
606
607 /*
608 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
609 * is also a catch-all for the retransmit timer timeout case.
610 */
611 if (tp->t_flags & TF_ACKNOW)
612 goto send;
613 if ((flags & TH_RST) ||
614 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
615 goto send;
616 if (SEQ_GT(tp->snd_up, tp->snd_una))
617 goto send;
618 /*
619 * If our state indicates that FIN should be sent
620 * and we have not yet done so, then we need to send.
621 */
622 if (flags & TH_FIN &&
623 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
624 goto send;
625 /*
626 * In SACK, it is possible for tcp_output to fail to send a segment
627 * after the retransmission timer has been turned off. Make sure
628 * that the retransmission timer is set.
629 */
630 if ((tp->t_flags & TF_SACK_PERMIT) &&
631 SEQ_GT(tp->snd_max, tp->snd_una) &&
632 !tcp_timer_active(tp, TT_REXMT) &&
633 !tcp_timer_active(tp, TT_PERSIST)) {
634 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
635 goto just_return;
636 }
637 /*
638 * TCP window updates are not reliable, rather a polling protocol
639 * using ``persist'' packets is used to insure receipt of window
640 * updates. The three ``states'' for the output side are:
641 * idle not doing retransmits or persists
642 * persisting to move a small or zero window
643 * (re)transmitting and thereby not persisting
644 *
645 * tcp_timer_active(tp, TT_PERSIST)
646 * is true when we are in persist state.
647 * (tp->t_flags & TF_FORCEDATA)
648 * is set when we are called to send a persist packet.
649 * tcp_timer_active(tp, TT_REXMT)
650 * is set when we are retransmitting
651 * The output side is idle when both timers are zero.
652 *
653 * If send window is too small, there is data to transmit, and no
654 * retransmit or persist is pending, then go to persist state.
655 * If nothing happens soon, send when timer expires:
656 * if window is nonzero, transmit what we can,
657 * otherwise force out a byte.
658 */
659 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
660 !tcp_timer_active(tp, TT_PERSIST)) {
661 tp->t_rxtshift = 0;
662 tcp_setpersist(tp);
663 }
664
665 /*
666 * No reason to send a segment, just return.
667 */
668 just_return:
669 SOCKBUF_UNLOCK(&so->so_snd);
670 return (0);
671
672 send:
673 SOCKBUF_LOCK_ASSERT(&so->so_snd);
674 /*
675 * Before ESTABLISHED, force sending of initial options
676 * unless TCP set not to do any options.
677 * NOTE: we assume that the IP/TCP header plus TCP options
678 * always fit in a single mbuf, leaving room for a maximum
679 * link header, i.e.
680 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
681 */
682 optlen = 0;
683 #ifdef INET6
684 if (isipv6)
685 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
686 else
687 #endif
688 hdrlen = sizeof (struct tcpiphdr);
689
690 /*
691 * Compute options for segment.
692 * We only have to care about SYN and established connection
693 * segments. Options for SYN-ACK segments are handled in TCP
694 * syncache.
695 */
696 if ((tp->t_flags & TF_NOOPT) == 0) {
697 to.to_flags = 0;
698 /* Maximum segment size. */
699 if (flags & TH_SYN) {
700 tp->snd_nxt = tp->iss;
701 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
702 to.to_flags |= TOF_MSS;
703 }
704 /* Window scaling. */
705 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
706 to.to_wscale = tp->request_r_scale;
707 to.to_flags |= TOF_SCALE;
708 }
709 /* Timestamps. */
710 if ((tp->t_flags & TF_RCVD_TSTMP) ||
711 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
712 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
713 to.to_tsecr = tp->ts_recent;
714 to.to_flags |= TOF_TS;
715 /* Set receive buffer autosizing timestamp. */
716 if (tp->rfbuf_ts == 0 &&
717 (so->so_rcv.sb_flags & SB_AUTOSIZE))
718 tp->rfbuf_ts = tcp_ts_getticks();
719 }
720 /* Selective ACK's. */
721 if (tp->t_flags & TF_SACK_PERMIT) {
722 if (flags & TH_SYN)
723 to.to_flags |= TOF_SACKPERM;
724 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
725 (tp->t_flags & TF_SACK_PERMIT) &&
726 tp->rcv_numsacks > 0) {
727 to.to_flags |= TOF_SACK;
728 to.to_nsacks = tp->rcv_numsacks;
729 to.to_sacks = (u_char *)tp->sackblks;
730 }
731 }
732 #ifdef TCP_SIGNATURE
733 /* TCP-MD5 (RFC2385). */
734 if (tp->t_flags & TF_SIGNATURE)
735 to.to_flags |= TOF_SIGNATURE;
736 #endif /* TCP_SIGNATURE */
737
738 /* Processing the options. */
739 hdrlen += optlen = tcp_addoptions(&to, opt);
740 }
741
742 #ifdef INET6
743 if (isipv6)
744 ipoptlen = ip6_optlen(tp->t_inpcb);
745 else
746 #endif
747 if (tp->t_inpcb->inp_options)
748 ipoptlen = tp->t_inpcb->inp_options->m_len -
749 offsetof(struct ipoption, ipopt_list);
750 else
751 ipoptlen = 0;
752 #ifdef IPSEC
753 ipoptlen += ipsec_optlen;
754 #endif
755
756 /*
757 * Adjust data length if insertion of options will
758 * bump the packet length beyond the t_maxopd length.
759 * Clear the FIN bit because we cut off the tail of
760 * the segment.
761 *
762 * When doing TSO limit a burst to TCP_MAXWIN minus the
763 * IP, TCP and Options length to keep ip->ip_len from
764 * overflowing. Prevent the last segment from being
765 * fractional thus making them all equal sized and set
766 * the flag to continue sending. TSO is disabled when
767 * IP options or IPSEC are present.
768 */
769 if (len + optlen + ipoptlen > tp->t_maxopd) {
770 flags &= ~TH_FIN;
771 if (tso) {
772 if (len > TCP_MAXWIN - hdrlen - optlen) {
773 len = TCP_MAXWIN - hdrlen - optlen;
774 len = len - (len % (tp->t_maxopd - optlen));
775 sendalot = 1;
776 } else if (tp->t_flags & TF_NEEDFIN)
777 sendalot = 1;
778 } else {
779 len = tp->t_maxopd - optlen - ipoptlen;
780 sendalot = 1;
781 }
782 }
783
784 /*#ifdef DIAGNOSTIC*/
785 #ifdef INET6
786 if (max_linkhdr + hdrlen > MCLBYTES)
787 #else
788 if (max_linkhdr + hdrlen > MHLEN)
789 #endif
790 panic("tcphdr too big");
791 /*#endif*/
792
793 /*
794 * This KASSERT is here to catch edge cases at a well defined place.
795 * Before, those had triggered (random) panic conditions further down.
796 */
797 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
798
799 /*
800 * Grab a header mbuf, attaching a copy of data to
801 * be transmitted, and initialize the header from
802 * the template for sends on this connection.
803 */
804 if (len) {
805 struct mbuf *mb;
806 u_int moff;
807
808 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
809 TCPSTAT_INC(tcps_sndprobe);
810 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
811 tp->t_sndrexmitpack++;
812 TCPSTAT_INC(tcps_sndrexmitpack);
813 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
814 } else {
815 TCPSTAT_INC(tcps_sndpack);
816 TCPSTAT_ADD(tcps_sndbyte, len);
817 }
818 #ifdef notyet
819 if ((m = m_copypack(so->so_snd.sb_mb, off,
820 (int)len, max_linkhdr + hdrlen)) == 0) {
821 SOCKBUF_UNLOCK(&so->so_snd);
822 error = ENOBUFS;
823 goto out;
824 }
825 /*
826 * m_copypack left space for our hdr; use it.
827 */
828 m->m_len += hdrlen;
829 m->m_data -= hdrlen;
830 #else
831 MGETHDR(m, M_DONTWAIT, MT_DATA);
832 if (m == NULL) {
833 SOCKBUF_UNLOCK(&so->so_snd);
834 error = ENOBUFS;
835 goto out;
836 }
837 #ifdef INET6
838 if (MHLEN < hdrlen + max_linkhdr) {
839 MCLGET(m, M_DONTWAIT);
840 if ((m->m_flags & M_EXT) == 0) {
841 SOCKBUF_UNLOCK(&so->so_snd);
842 m_freem(m);
843 error = ENOBUFS;
844 goto out;
845 }
846 }
847 #endif
848 m->m_data += max_linkhdr;
849 m->m_len = hdrlen;
850
851 /*
852 * Start the m_copy functions from the closest mbuf
853 * to the offset in the socket buffer chain.
854 */
855 mb = sbsndptr(&so->so_snd, off, len, &moff);
856
857 if (len <= MHLEN - hdrlen - max_linkhdr) {
858 m_copydata(mb, moff, (int)len,
859 mtod(m, caddr_t) + hdrlen);
860 m->m_len += len;
861 } else {
862 m->m_next = m_copy(mb, moff, (int)len);
863 if (m->m_next == NULL) {
864 SOCKBUF_UNLOCK(&so->so_snd);
865 (void) m_free(m);
866 error = ENOBUFS;
867 goto out;
868 }
869 }
870 #endif
871 /*
872 * If we're sending everything we've got, set PUSH.
873 * (This will keep happy those implementations which only
874 * give data to the user when a buffer fills or
875 * a PUSH comes in.)
876 */
877 if (off + len == so->so_snd.sb_cc)
878 flags |= TH_PUSH;
879 SOCKBUF_UNLOCK(&so->so_snd);
880 } else {
881 SOCKBUF_UNLOCK(&so->so_snd);
882 if (tp->t_flags & TF_ACKNOW)
883 TCPSTAT_INC(tcps_sndacks);
884 else if (flags & (TH_SYN|TH_FIN|TH_RST))
885 TCPSTAT_INC(tcps_sndctrl);
886 else if (SEQ_GT(tp->snd_up, tp->snd_una))
887 TCPSTAT_INC(tcps_sndurg);
888 else
889 TCPSTAT_INC(tcps_sndwinup);
890
891 MGETHDR(m, M_DONTWAIT, MT_DATA);
892 if (m == NULL) {
893 error = ENOBUFS;
894 goto out;
895 }
896 #ifdef INET6
897 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
898 MHLEN >= hdrlen) {
899 MH_ALIGN(m, hdrlen);
900 } else
901 #endif
902 m->m_data += max_linkhdr;
903 m->m_len = hdrlen;
904 }
905 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
906 m->m_pkthdr.rcvif = (struct ifnet *)0;
907 #ifdef MAC
908 mac_inpcb_create_mbuf(tp->t_inpcb, m);
909 #endif
910 #ifdef INET6
911 if (isipv6) {
912 ip6 = mtod(m, struct ip6_hdr *);
913 th = (struct tcphdr *)(ip6 + 1);
914 tcpip_fillheaders(tp->t_inpcb, ip6, th);
915 } else
916 #endif /* INET6 */
917 {
918 ip = mtod(m, struct ip *);
919 ipov = (struct ipovly *)ip;
920 th = (struct tcphdr *)(ip + 1);
921 tcpip_fillheaders(tp->t_inpcb, ip, th);
922 }
923
924 /*
925 * Fill in fields, remembering maximum advertised
926 * window for use in delaying messages about window sizes.
927 * If resending a FIN, be sure not to use a new sequence number.
928 */
929 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
930 tp->snd_nxt == tp->snd_max)
931 tp->snd_nxt--;
932 /*
933 * If we are starting a connection, send ECN setup
934 * SYN packet. If we are on a retransmit, we may
935 * resend those bits a number of times as per
936 * RFC 3168.
937 */
938 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
939 if (tp->t_rxtshift >= 1) {
940 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
941 flags |= TH_ECE|TH_CWR;
942 } else
943 flags |= TH_ECE|TH_CWR;
944 }
945
946 if (tp->t_state == TCPS_ESTABLISHED &&
947 (tp->t_flags & TF_ECN_PERMIT)) {
948 /*
949 * If the peer has ECN, mark data packets with
950 * ECN capable transmission (ECT).
951 * Ignore pure ack packets, retransmissions and window probes.
952 */
953 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
954 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
955 #ifdef INET6
956 if (isipv6)
957 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
958 else
959 #endif
960 ip->ip_tos |= IPTOS_ECN_ECT0;
961 TCPSTAT_INC(tcps_ecn_ect0);
962 }
963
964 /*
965 * Reply with proper ECN notifications.
966 */
967 if (tp->t_flags & TF_ECN_SND_CWR) {
968 flags |= TH_CWR;
969 tp->t_flags &= ~TF_ECN_SND_CWR;
970 }
971 if (tp->t_flags & TF_ECN_SND_ECE)
972 flags |= TH_ECE;
973 }
974
975 /*
976 * If we are doing retransmissions, then snd_nxt will
977 * not reflect the first unsent octet. For ACK only
978 * packets, we do not want the sequence number of the
979 * retransmitted packet, we want the sequence number
980 * of the next unsent octet. So, if there is no data
981 * (and no SYN or FIN), use snd_max instead of snd_nxt
982 * when filling in ti_seq. But if we are in persist
983 * state, snd_max might reflect one byte beyond the
984 * right edge of the window, so use snd_nxt in that
985 * case, since we know we aren't doing a retransmission.
986 * (retransmit and persist are mutually exclusive...)
987 */
988 if (sack_rxmit == 0) {
989 if (len || (flags & (TH_SYN|TH_FIN)) ||
990 tcp_timer_active(tp, TT_PERSIST))
991 th->th_seq = htonl(tp->snd_nxt);
992 else
993 th->th_seq = htonl(tp->snd_max);
994 } else {
995 th->th_seq = htonl(p->rxmit);
996 p->rxmit += len;
997 tp->sackhint.sack_bytes_rexmit += len;
998 }
999 th->th_ack = htonl(tp->rcv_nxt);
1000 if (optlen) {
1001 bcopy(opt, th + 1, optlen);
1002 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1003 }
1004 th->th_flags = flags;
1005 /*
1006 * Calculate receive window. Don't shrink window,
1007 * but avoid silly window syndrome.
1008 */
1009 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1010 recwin < (long)tp->t_maxseg)
1011 recwin = 0;
1012 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1013 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1014 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1015 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1016 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1017
1018 /*
1019 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1020 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1021 * case is handled in syncache.
1022 */
1023 if (flags & TH_SYN)
1024 th->th_win = htons((u_short)
1025 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1026 else
1027 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1028
1029 /*
1030 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1031 * a 0 window. This may cause the remote transmitter to stall. This
1032 * flag tells soreceive() to disable delayed acknowledgements when
1033 * draining the buffer. This can occur if the receiver is attempting
1034 * to read more data than can be buffered prior to transmitting on
1035 * the connection.
1036 */
1037 if (th->th_win == 0) {
1038 tp->t_sndzerowin++;
1039 tp->t_flags |= TF_RXWIN0SENT;
1040 } else
1041 tp->t_flags &= ~TF_RXWIN0SENT;
1042 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1043 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1044 th->th_flags |= TH_URG;
1045 } else
1046 /*
1047 * If no urgent pointer to send, then we pull
1048 * the urgent pointer to the left edge of the send window
1049 * so that it doesn't drift into the send window on sequence
1050 * number wraparound.
1051 */
1052 tp->snd_up = tp->snd_una; /* drag it along */
1053
1054 #ifdef TCP_SIGNATURE
1055 if (tp->t_flags & TF_SIGNATURE) {
1056 int sigoff = to.to_signature - opt;
1057 tcp_signature_compute(m, 0, len, optlen,
1058 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1059 }
1060 #endif
1061
1062 /*
1063 * Put TCP length in extended header, and then
1064 * checksum extended header and data.
1065 */
1066 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1067 #ifdef INET6
1068 if (isipv6)
1069 /*
1070 * ip6_plen is not need to be filled now, and will be filled
1071 * in ip6_output.
1072 */
1073 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1074 sizeof(struct tcphdr) + optlen + len);
1075 else
1076 #endif /* INET6 */
1077 {
1078 m->m_pkthdr.csum_flags = CSUM_TCP;
1079 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1080 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1081 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1082
1083 /* IP version must be set here for ipv4/ipv6 checking later */
1084 KASSERT(ip->ip_v == IPVERSION,
1085 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1086 }
1087
1088 /*
1089 * Enable TSO and specify the size of the segments.
1090 * The TCP pseudo header checksum is always provided.
1091 * XXX: Fixme: This is currently not the case for IPv6.
1092 */
1093 if (tso) {
1094 KASSERT(len > tp->t_maxopd - optlen,
1095 ("%s: len <= tso_segsz", __func__));
1096 m->m_pkthdr.csum_flags |= CSUM_TSO;
1097 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1098 }
1099
1100 /*
1101 * In transmit state, time the transmission and arrange for
1102 * the retransmit. In persist state, just set snd_max.
1103 */
1104 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1105 !tcp_timer_active(tp, TT_PERSIST)) {
1106 tcp_seq startseq = tp->snd_nxt;
1107
1108 /*
1109 * Advance snd_nxt over sequence space of this segment.
1110 */
1111 if (flags & (TH_SYN|TH_FIN)) {
1112 if (flags & TH_SYN)
1113 tp->snd_nxt++;
1114 if (flags & TH_FIN) {
1115 tp->snd_nxt++;
1116 tp->t_flags |= TF_SENTFIN;
1117 }
1118 }
1119 if (sack_rxmit)
1120 goto timer;
1121 tp->snd_nxt += len;
1122 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1123 tp->snd_max = tp->snd_nxt;
1124 /*
1125 * Time this transmission if not a retransmission and
1126 * not currently timing anything.
1127 */
1128 if (tp->t_rtttime == 0) {
1129 tp->t_rtttime = ticks;
1130 tp->t_rtseq = startseq;
1131 TCPSTAT_INC(tcps_segstimed);
1132 }
1133 }
1134
1135 /*
1136 * Set retransmit timer if not currently set,
1137 * and not doing a pure ack or a keep-alive probe.
1138 * Initial value for retransmit timer is smoothed
1139 * round-trip time + 2 * round-trip time variance.
1140 * Initialize shift counter which is used for backoff
1141 * of retransmit time.
1142 */
1143 timer:
1144 if (!tcp_timer_active(tp, TT_REXMT) &&
1145 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1146 (tp->snd_nxt != tp->snd_una))) {
1147 if (tcp_timer_active(tp, TT_PERSIST)) {
1148 tcp_timer_activate(tp, TT_PERSIST, 0);
1149 tp->t_rxtshift = 0;
1150 }
1151 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1152 }
1153 } else {
1154 /*
1155 * Persist case, update snd_max but since we are in
1156 * persist mode (no window) we do not update snd_nxt.
1157 */
1158 int xlen = len;
1159 if (flags & TH_SYN)
1160 ++xlen;
1161 if (flags & TH_FIN) {
1162 ++xlen;
1163 tp->t_flags |= TF_SENTFIN;
1164 }
1165 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1166 tp->snd_max = tp->snd_nxt + len;
1167 }
1168
1169 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1170 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1171
1172 #ifdef TCPDEBUG
1173 /*
1174 * Trace.
1175 */
1176 if (so->so_options & SO_DEBUG) {
1177 u_short save = 0;
1178 #ifdef INET6
1179 if (!isipv6)
1180 #endif
1181 {
1182 save = ipov->ih_len;
1183 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1184 }
1185 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1186 #ifdef INET6
1187 if (!isipv6)
1188 #endif
1189 ipov->ih_len = save;
1190 }
1191 #endif
1192
1193 /*
1194 * Fill in IP length and desired time to live and
1195 * send to IP level. There should be a better way
1196 * to handle ttl and tos; we could keep them in
1197 * the template, but need a way to checksum without them.
1198 */
1199 /*
1200 * m->m_pkthdr.len should have been set before cksum calcuration,
1201 * because in6_cksum() need it.
1202 */
1203 #ifdef INET6
1204 if (isipv6) {
1205 /*
1206 * we separately set hoplimit for every segment, since the
1207 * user might want to change the value via setsockopt.
1208 * Also, desired default hop limit might be changed via
1209 * Neighbor Discovery.
1210 */
1211 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1212
1213 /* TODO: IPv6 IP6TOS_ECT bit on */
1214 error = ip6_output(m,
1215 tp->t_inpcb->in6p_outputopts, NULL,
1216 ((so->so_options & SO_DONTROUTE) ?
1217 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1218 } else
1219 #endif /* INET6 */
1220 {
1221 ip->ip_len = m->m_pkthdr.len;
1222 #ifdef INET6
1223 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1224 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1225 #endif /* INET6 */
1226 /*
1227 * If we do path MTU discovery, then we set DF on every packet.
1228 * This might not be the best thing to do according to RFC3390
1229 * Section 2. However the tcp hostcache migitates the problem
1230 * so it affects only the first tcp connection with a host.
1231 *
1232 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1233 */
1234 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1235 ip->ip_off |= IP_DF;
1236
1237 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1238 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1239 tp->t_inpcb);
1240 }
1241 if (error) {
1242
1243 /*
1244 * We know that the packet was lost, so back out the
1245 * sequence number advance, if any.
1246 *
1247 * If the error is EPERM the packet got blocked by the
1248 * local firewall. Normally we should terminate the
1249 * connection but the blocking may have been spurious
1250 * due to a firewall reconfiguration cycle. So we treat
1251 * it like a packet loss and let the retransmit timer and
1252 * timeouts do their work over time.
1253 * XXX: It is a POLA question whether calling tcp_drop right
1254 * away would be the really correct behavior instead.
1255 */
1256 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1257 !tcp_timer_active(tp, TT_PERSIST)) &&
1258 ((flags & TH_SYN) == 0) &&
1259 (error != EPERM)) {
1260 if (sack_rxmit) {
1261 p->rxmit -= len;
1262 tp->sackhint.sack_bytes_rexmit -= len;
1263 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1264 ("sackhint bytes rtx >= 0"));
1265 } else
1266 tp->snd_nxt -= len;
1267 }
1268 out:
1269 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1270 switch (error) {
1271 case EPERM:
1272 tp->t_softerror = error;
1273 return (error);
1274 case ENOBUFS:
1275 if (!tcp_timer_active(tp, TT_REXMT) &&
1276 !tcp_timer_active(tp, TT_PERSIST))
1277 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1278 tp->snd_cwnd = tp->t_maxseg;
1279 return (0);
1280 case EMSGSIZE:
1281 /*
1282 * For some reason the interface we used initially
1283 * to send segments changed to another or lowered
1284 * its MTU.
1285 *
1286 * tcp_mtudisc() will find out the new MTU and as
1287 * its last action, initiate retransmission, so it
1288 * is important to not do so here.
1289 *
1290 * If TSO was active we either got an interface
1291 * without TSO capabilits or TSO was turned off.
1292 * Disable it for this connection as too and
1293 * immediatly retry with MSS sized segments generated
1294 * by this function.
1295 */
1296 if (tso)
1297 tp->t_flags &= ~TF_TSO;
1298 tcp_mtudisc(tp->t_inpcb, -1);
1299 return (0);
1300 case EHOSTDOWN:
1301 case EHOSTUNREACH:
1302 case ENETDOWN:
1303 case ENETUNREACH:
1304 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1305 tp->t_softerror = error;
1306 return (0);
1307 }
1308 /* FALLTHROUGH */
1309 default:
1310 return (error);
1311 }
1312 }
1313 TCPSTAT_INC(tcps_sndtotal);
1314
1315 /*
1316 * Data sent (as far as we can tell).
1317 * If this advertises a larger window than any other segment,
1318 * then remember the size of the advertised window.
1319 * Any pending ACK has now been sent.
1320 */
1321 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1322 tp->rcv_adv = tp->rcv_nxt + recwin;
1323 tp->last_ack_sent = tp->rcv_nxt;
1324 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1325 if (tcp_timer_active(tp, TT_DELACK))
1326 tcp_timer_activate(tp, TT_DELACK, 0);
1327 #if 0
1328 /*
1329 * This completely breaks TCP if newreno is turned on. What happens
1330 * is that if delayed-acks are turned on on the receiver, this code
1331 * on the transmitter effectively destroys the TCP window, forcing
1332 * it to four packets (1.5Kx4 = 6K window).
1333 */
1334 if (sendalot && --maxburst)
1335 goto again;
1336 #endif
1337 if (sendalot)
1338 goto again;
1339 return (0);
1340 }
1341
1342 void
1343 tcp_setpersist(struct tcpcb *tp)
1344 {
1345 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1346 int tt;
1347
1348 tp->t_flags &= ~TF_PREVVALID;
1349 if (tcp_timer_active(tp, TT_REXMT))
1350 panic("tcp_setpersist: retransmit pending");
1351 /*
1352 * Start/restart persistance timer.
1353 */
1354 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1355 TCPTV_PERSMIN, TCPTV_PERSMAX);
1356 tcp_timer_activate(tp, TT_PERSIST, tt);
1357 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1358 tp->t_rxtshift++;
1359 }
1360
1361 /*
1362 * Insert TCP options according to the supplied parameters to the place
1363 * optp in a consistent way. Can handle unaligned destinations.
1364 *
1365 * The order of the option processing is crucial for optimal packing and
1366 * alignment for the scarce option space.
1367 *
1368 * The optimal order for a SYN/SYN-ACK segment is:
1369 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1370 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1371 *
1372 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1373 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1374 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1375 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1376 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1377 */
1378 int
1379 tcp_addoptions(struct tcpopt *to, u_char *optp)
1380 {
1381 u_int mask, optlen = 0;
1382
1383 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1384 if ((to->to_flags & mask) != mask)
1385 continue;
1386 if (optlen == TCP_MAXOLEN)
1387 break;
1388 switch (to->to_flags & mask) {
1389 case TOF_MSS:
1390 while (optlen % 4) {
1391 optlen += TCPOLEN_NOP;
1392 *optp++ = TCPOPT_NOP;
1393 }
1394 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1395 continue;
1396 optlen += TCPOLEN_MAXSEG;
1397 *optp++ = TCPOPT_MAXSEG;
1398 *optp++ = TCPOLEN_MAXSEG;
1399 to->to_mss = htons(to->to_mss);
1400 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1401 optp += sizeof(to->to_mss);
1402 break;
1403 case TOF_SCALE:
1404 while (!optlen || optlen % 2 != 1) {
1405 optlen += TCPOLEN_NOP;
1406 *optp++ = TCPOPT_NOP;
1407 }
1408 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1409 continue;
1410 optlen += TCPOLEN_WINDOW;
1411 *optp++ = TCPOPT_WINDOW;
1412 *optp++ = TCPOLEN_WINDOW;
1413 *optp++ = to->to_wscale;
1414 break;
1415 case TOF_SACKPERM:
1416 while (optlen % 2) {
1417 optlen += TCPOLEN_NOP;
1418 *optp++ = TCPOPT_NOP;
1419 }
1420 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1421 continue;
1422 optlen += TCPOLEN_SACK_PERMITTED;
1423 *optp++ = TCPOPT_SACK_PERMITTED;
1424 *optp++ = TCPOLEN_SACK_PERMITTED;
1425 break;
1426 case TOF_TS:
1427 while (!optlen || optlen % 4 != 2) {
1428 optlen += TCPOLEN_NOP;
1429 *optp++ = TCPOPT_NOP;
1430 }
1431 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1432 continue;
1433 optlen += TCPOLEN_TIMESTAMP;
1434 *optp++ = TCPOPT_TIMESTAMP;
1435 *optp++ = TCPOLEN_TIMESTAMP;
1436 to->to_tsval = htonl(to->to_tsval);
1437 to->to_tsecr = htonl(to->to_tsecr);
1438 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1439 optp += sizeof(to->to_tsval);
1440 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1441 optp += sizeof(to->to_tsecr);
1442 break;
1443 case TOF_SIGNATURE:
1444 {
1445 int siglen = TCPOLEN_SIGNATURE - 2;
1446
1447 while (!optlen || optlen % 4 != 2) {
1448 optlen += TCPOLEN_NOP;
1449 *optp++ = TCPOPT_NOP;
1450 }
1451 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1452 continue;
1453 optlen += TCPOLEN_SIGNATURE;
1454 *optp++ = TCPOPT_SIGNATURE;
1455 *optp++ = TCPOLEN_SIGNATURE;
1456 to->to_signature = optp;
1457 while (siglen--)
1458 *optp++ = 0;
1459 break;
1460 }
1461 case TOF_SACK:
1462 {
1463 int sackblks = 0;
1464 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1465 tcp_seq sack_seq;
1466
1467 while (!optlen || optlen % 4 != 2) {
1468 optlen += TCPOLEN_NOP;
1469 *optp++ = TCPOPT_NOP;
1470 }
1471 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1472 continue;
1473 optlen += TCPOLEN_SACKHDR;
1474 *optp++ = TCPOPT_SACK;
1475 sackblks = min(to->to_nsacks,
1476 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1477 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1478 while (sackblks--) {
1479 sack_seq = htonl(sack->start);
1480 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1481 optp += sizeof(sack_seq);
1482 sack_seq = htonl(sack->end);
1483 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1484 optp += sizeof(sack_seq);
1485 optlen += TCPOLEN_SACK;
1486 sack++;
1487 }
1488 TCPSTAT_INC(tcps_sack_send_blocks);
1489 break;
1490 }
1491 default:
1492 panic("%s: unknown TCP option type", __func__);
1493 break;
1494 }
1495 }
1496
1497 /* Terminate and pad TCP options to a 4 byte boundary. */
1498 if (optlen % 4) {
1499 optlen += TCPOLEN_EOL;
1500 *optp++ = TCPOPT_EOL;
1501 }
1502 /*
1503 * According to RFC 793 (STD0007):
1504 * "The content of the header beyond the End-of-Option option
1505 * must be header padding (i.e., zero)."
1506 * and later: "The padding is composed of zeros."
1507 */
1508 while (optlen % 4) {
1509 optlen += TCPOLEN_PAD;
1510 *optp++ = TCPOPT_PAD;
1511 }
1512
1513 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1514 return (optlen);
1515 }
Cache object: b12d8c39f1a538bd651aec4a16918896
|