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: releng/7.3/sys/netinet/tcp_output.c 199401 2009-11-17 16:17:39Z jhb $");
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;
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 = 0;
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 * We have been idle for "a while" and no acks are
167 * expected to clock out any data we send --
168 * slow start to get ack "clock" running again.
169 *
170 * Set the slow-start flight size depending on whether
171 * this is a local network or not.
172 */
173 int ss = ss_fltsz;
174 #ifdef INET6
175 if (isipv6) {
176 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
177 ss = ss_fltsz_local;
178 } else
179 #endif /* INET6 */
180 if (in_localaddr(tp->t_inpcb->inp_faddr))
181 ss = ss_fltsz_local;
182 tp->snd_cwnd = tp->t_maxseg * ss;
183 }
184 tp->t_flags &= ~TF_LASTIDLE;
185 if (idle) {
186 if (tp->t_flags & TF_MORETOCOME) {
187 tp->t_flags |= TF_LASTIDLE;
188 idle = 0;
189 }
190 }
191 again:
192 /*
193 * If we've recently taken a timeout, snd_max will be greater than
194 * snd_nxt. There may be SACK information that allows us to avoid
195 * resending already delivered data. Adjust snd_nxt accordingly.
196 */
197 if ((tp->t_flags & TF_SACK_PERMIT) &&
198 SEQ_LT(tp->snd_nxt, tp->snd_max))
199 tcp_sack_adjust(tp);
200 sendalot = 0;
201 off = tp->snd_nxt - tp->snd_una;
202 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
203 sendwin = min(sendwin, tp->snd_bwnd);
204
205 flags = tcp_outflags[tp->t_state];
206 /*
207 * Send any SACK-generated retransmissions. If we're explicitly trying
208 * to send out new data (when sendalot is 1), bypass this function.
209 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
210 * we're replacing a (future) new transmission with a retransmission
211 * now, and we previously incremented snd_cwnd in tcp_input().
212 */
213 /*
214 * Still in sack recovery , reset rxmit flag to zero.
215 */
216 sack_rxmit = 0;
217 sack_bytes_rxmt = 0;
218 len = 0;
219 p = NULL;
220 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
221 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
222 long cwin;
223
224 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
225 if (cwin < 0)
226 cwin = 0;
227 /* Do not retransmit SACK segments beyond snd_recover */
228 if (SEQ_GT(p->end, tp->snd_recover)) {
229 /*
230 * (At least) part of sack hole extends beyond
231 * snd_recover. Check to see if we can rexmit data
232 * for this hole.
233 */
234 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
235 /*
236 * Can't rexmit any more data for this hole.
237 * That data will be rexmitted in the next
238 * sack recovery episode, when snd_recover
239 * moves past p->rxmit.
240 */
241 p = NULL;
242 goto after_sack_rexmit;
243 } else
244 /* Can rexmit part of the current hole */
245 len = ((long)ulmin(cwin,
246 tp->snd_recover - p->rxmit));
247 } else
248 len = ((long)ulmin(cwin, p->end - p->rxmit));
249 off = p->rxmit - tp->snd_una;
250 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
251 __func__, off));
252 if (len > 0) {
253 sack_rxmit = 1;
254 sendalot = 1;
255 tcpstat.tcps_sack_rexmits++;
256 tcpstat.tcps_sack_rexmit_bytes +=
257 min(len, tp->t_maxseg);
258 }
259 }
260 after_sack_rexmit:
261 /*
262 * Get standard flags, and add SYN or FIN if requested by 'hidden'
263 * state flags.
264 */
265 if (tp->t_flags & TF_NEEDFIN)
266 flags |= TH_FIN;
267 if (tp->t_flags & TF_NEEDSYN)
268 flags |= TH_SYN;
269
270 SOCKBUF_LOCK(&so->so_snd);
271 /*
272 * If in persist timeout with window of 0, send 1 byte.
273 * Otherwise, if window is small but nonzero
274 * and timer expired, we will send what we can
275 * and go to transmit state.
276 */
277 if (tp->t_flags & TF_FORCEDATA) {
278 if (sendwin == 0) {
279 /*
280 * If we still have some data to send, then
281 * clear the FIN bit. Usually this would
282 * happen below when it realizes that we
283 * aren't sending all the data. However,
284 * if we have exactly 1 byte of unsent data,
285 * then it won't clear the FIN bit below,
286 * and if we are in persist state, we wind
287 * up sending the packet without recording
288 * that we sent the FIN bit.
289 *
290 * We can't just blindly clear the FIN bit,
291 * because if we don't have any more data
292 * to send then the probe will be the FIN
293 * itself.
294 */
295 if (off < so->so_snd.sb_cc)
296 flags &= ~TH_FIN;
297 sendwin = 1;
298 } else {
299 tcp_timer_activate(tp, TT_PERSIST, 0);
300 tp->t_rxtshift = 0;
301 }
302 }
303
304 /*
305 * If snd_nxt == snd_max and we have transmitted a FIN, the
306 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
307 * a negative length. This can also occur when TCP opens up
308 * its congestion window while receiving additional duplicate
309 * acks after fast-retransmit because TCP will reset snd_nxt
310 * to snd_max after the fast-retransmit.
311 *
312 * In the normal retransmit-FIN-only case, however, snd_nxt will
313 * be set to snd_una, the offset will be 0, and the length may
314 * wind up 0.
315 *
316 * If sack_rxmit is true we are retransmitting from the scoreboard
317 * in which case len is already set.
318 */
319 if (sack_rxmit == 0) {
320 if (sack_bytes_rxmt == 0)
321 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
322 else {
323 long cwin;
324
325 /*
326 * We are inside of a SACK recovery episode and are
327 * sending new data, having retransmitted all the
328 * data possible in the scoreboard.
329 */
330 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
331 - off);
332 /*
333 * Don't remove this (len > 0) check !
334 * We explicitly check for len > 0 here (although it
335 * isn't really necessary), to work around a gcc
336 * optimization issue - to force gcc to compute
337 * len above. Without this check, the computation
338 * of len is bungled by the optimizer.
339 */
340 if (len > 0) {
341 cwin = tp->snd_cwnd -
342 (tp->snd_nxt - tp->sack_newdata) -
343 sack_bytes_rxmt;
344 if (cwin < 0)
345 cwin = 0;
346 len = lmin(len, cwin);
347 }
348 }
349 }
350
351 /*
352 * Lop off SYN bit if it has already been sent. However, if this
353 * is SYN-SENT state and if segment contains data and if we don't
354 * know that foreign host supports TAO, suppress sending segment.
355 */
356 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
357 if (tp->t_state != TCPS_SYN_RECEIVED)
358 flags &= ~TH_SYN;
359 off--, len++;
360 }
361
362 /*
363 * Be careful not to send data and/or FIN on SYN segments.
364 * This measure is needed to prevent interoperability problems
365 * with not fully conformant TCP implementations.
366 */
367 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
368 len = 0;
369 flags &= ~TH_FIN;
370 }
371
372 if (len < 0) {
373 /*
374 * If FIN has been sent but not acked,
375 * but we haven't been called to retransmit,
376 * len will be < 0. Otherwise, window shrank
377 * after we sent into it. If window shrank to 0,
378 * cancel pending retransmit, pull snd_nxt back
379 * to (closed) window, and set the persist timer
380 * if it isn't already going. If the window didn't
381 * close completely, just wait for an ACK.
382 */
383 len = 0;
384 if (sendwin == 0) {
385 tcp_timer_activate(tp, TT_REXMT, 0);
386 tp->t_rxtshift = 0;
387 tp->snd_nxt = tp->snd_una;
388 if (!tcp_timer_active(tp, TT_PERSIST))
389 tcp_setpersist(tp);
390 }
391 }
392
393 /* len will be >= 0 after this point. */
394 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
395
396 /*
397 * Automatic sizing of send socket buffer. Often the send buffer
398 * size is not optimally adjusted to the actual network conditions
399 * at hand (delay bandwidth product). Setting the buffer size too
400 * small limits throughput on links with high bandwidth and high
401 * delay (eg. trans-continental/oceanic links). Setting the
402 * buffer size too big consumes too much real kernel memory,
403 * especially with many connections on busy servers.
404 *
405 * The criteria to step up the send buffer one notch are:
406 * 1. receive window of remote host is larger than send buffer
407 * (with a fudge factor of 5/4th);
408 * 2. send buffer is filled to 7/8th with data (so we actually
409 * have data to make use of it);
410 * 3. send buffer fill has not hit maximal automatic size;
411 * 4. our send window (slow start and cogestion controlled) is
412 * larger than sent but unacknowledged data in send buffer.
413 *
414 * The remote host receive window scaling factor may limit the
415 * growing of the send buffer before it reaches its allowed
416 * maximum.
417 *
418 * It scales directly with slow start or congestion window
419 * and does at most one step per received ACK. This fast
420 * scaling has the drawback of growing the send buffer beyond
421 * what is strictly necessary to make full use of a given
422 * delay*bandwith product. However testing has shown this not
423 * to be much of an problem. At worst we are trading wasting
424 * of available bandwith (the non-use of it) for wasting some
425 * socket buffer memory.
426 *
427 * TODO: Shrink send buffer during idle periods together
428 * with congestion window. Requires another timer. Has to
429 * wait for upcoming tcp timer rewrite.
430 */
431 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
432 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
433 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
434 so->so_snd.sb_cc < tcp_autosndbuf_max &&
435 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
436 if (!sbreserve_locked(&so->so_snd,
437 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
438 tcp_autosndbuf_max), so, curthread))
439 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
440 }
441 }
442
443 /*
444 * Truncate to the maximum segment length or enable TCP Segmentation
445 * Offloading (if supported by hardware) and ensure that FIN is removed
446 * if the length no longer contains the last data byte.
447 *
448 * TSO may only be used if we are in a pure bulk sending state. The
449 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
450 * IP options prevent using TSO. With TSO the TCP header is the same
451 * (except for the sequence number) for all generated packets. This
452 * makes it impossible to transmit any options which vary per generated
453 * segment or packet.
454 *
455 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
456 * removal of FIN (if not already catched here) are handled later after
457 * the exact length of the TCP options are known.
458 */
459 #ifdef IPSEC
460 /*
461 * Pre-calculate here as we save another lookup into the darknesses
462 * of IPsec that way and can actually decide if TSO is ok.
463 */
464 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
465 #endif
466 if (len > tp->t_maxseg) {
467 if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
468 ((tp->t_flags & TF_SIGNATURE) == 0) &&
469 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
470 tp->t_inpcb->inp_options == NULL &&
471 tp->t_inpcb->in6p_options == NULL
472 #ifdef IPSEC
473 && ipsec_optlen == 0
474 #endif
475 ) {
476 tso = 1;
477 } else {
478 len = tp->t_maxseg;
479 sendalot = 1;
480 tso = 0;
481 }
482 }
483 if (sack_rxmit) {
484 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
485 flags &= ~TH_FIN;
486 } else {
487 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
488 flags &= ~TH_FIN;
489 }
490
491 recwin = sbspace(&so->so_rcv);
492
493 /*
494 * Sender silly window avoidance. We transmit under the following
495 * conditions when len is non-zero:
496 *
497 * - We have a full segment (or more with TSO)
498 * - This is the last buffer in a write()/send() and we are
499 * either idle or running NODELAY
500 * - we've timed out (e.g. persist timer)
501 * - we have more then 1/2 the maximum send window's worth of
502 * data (receiver may be limited the window size)
503 * - we need to retransmit
504 */
505 if (len) {
506 if (len >= tp->t_maxseg)
507 goto send;
508 /*
509 * NOTE! on localhost connections an 'ack' from the remote
510 * end may occur synchronously with the output and cause
511 * us to flush a buffer queued with moretocome. XXX
512 *
513 * note: the len + off check is almost certainly unnecessary.
514 */
515 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
516 (idle || (tp->t_flags & TF_NODELAY)) &&
517 len + off >= so->so_snd.sb_cc &&
518 (tp->t_flags & TF_NOPUSH) == 0) {
519 goto send;
520 }
521 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
522 goto send;
523 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
524 goto send;
525 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
526 goto send;
527 if (sack_rxmit)
528 goto send;
529 }
530
531 /*
532 * Compare available window to amount of window
533 * known to peer (as advertised window less
534 * next expected input). If the difference is at least two
535 * max size segments, or at least 50% of the maximum possible
536 * window, then want to send a window update to peer.
537 * Skip this if the connection is in T/TCP half-open state.
538 * Don't send pure window updates when the peer has closed
539 * the connection and won't ever send more data.
540 */
541 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
542 !TCPS_HAVERCVDFIN(tp->t_state)) {
543 /*
544 * "adv" is the amount we can increase the window,
545 * taking into account that we are limited by
546 * TCP_MAXWIN << tp->rcv_scale.
547 */
548 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
549 (tp->rcv_adv - tp->rcv_nxt);
550
551 if (adv >= (long) (2 * tp->t_maxseg))
552 goto send;
553 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
554 goto send;
555 }
556
557 /*
558 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
559 * is also a catch-all for the retransmit timer timeout case.
560 */
561 if (tp->t_flags & TF_ACKNOW)
562 goto send;
563 if ((flags & TH_RST) ||
564 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
565 goto send;
566 if (SEQ_GT(tp->snd_up, tp->snd_una))
567 goto send;
568 /*
569 * If our state indicates that FIN should be sent
570 * and we have not yet done so, then we need to send.
571 */
572 if (flags & TH_FIN &&
573 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
574 goto send;
575 /*
576 * In SACK, it is possible for tcp_output to fail to send a segment
577 * after the retransmission timer has been turned off. Make sure
578 * that the retransmission timer is set.
579 */
580 if ((tp->t_flags & TF_SACK_PERMIT) &&
581 SEQ_GT(tp->snd_max, tp->snd_una) &&
582 !tcp_timer_active(tp, TT_REXMT) &&
583 !tcp_timer_active(tp, TT_PERSIST)) {
584 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
585 goto just_return;
586 }
587 /*
588 * TCP window updates are not reliable, rather a polling protocol
589 * using ``persist'' packets is used to insure receipt of window
590 * updates. The three ``states'' for the output side are:
591 * idle not doing retransmits or persists
592 * persisting to move a small or zero window
593 * (re)transmitting and thereby not persisting
594 *
595 * tcp_timer_active(tp, TT_PERSIST)
596 * is true when we are in persist state.
597 * (tp->t_flags & TF_FORCEDATA)
598 * is set when we are called to send a persist packet.
599 * tcp_timer_active(tp, TT_REXMT)
600 * is set when we are retransmitting
601 * The output side is idle when both timers are zero.
602 *
603 * If send window is too small, there is data to transmit, and no
604 * retransmit or persist is pending, then go to persist state.
605 * If nothing happens soon, send when timer expires:
606 * if window is nonzero, transmit what we can,
607 * otherwise force out a byte.
608 */
609 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
610 !tcp_timer_active(tp, TT_PERSIST)) {
611 tp->t_rxtshift = 0;
612 tcp_setpersist(tp);
613 }
614
615 /*
616 * No reason to send a segment, just return.
617 */
618 just_return:
619 SOCKBUF_UNLOCK(&so->so_snd);
620 return (0);
621
622 send:
623 SOCKBUF_LOCK_ASSERT(&so->so_snd);
624 /*
625 * Before ESTABLISHED, force sending of initial options
626 * unless TCP set not to do any options.
627 * NOTE: we assume that the IP/TCP header plus TCP options
628 * always fit in a single mbuf, leaving room for a maximum
629 * link header, i.e.
630 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
631 */
632 optlen = 0;
633 #ifdef INET6
634 if (isipv6)
635 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
636 else
637 #endif
638 hdrlen = sizeof (struct tcpiphdr);
639
640 /*
641 * Compute options for segment.
642 * We only have to care about SYN and established connection
643 * segments. Options for SYN-ACK segments are handled in TCP
644 * syncache.
645 */
646 if ((tp->t_flags & TF_NOOPT) == 0) {
647 to.to_flags = 0;
648 /* Maximum segment size. */
649 if (flags & TH_SYN) {
650 tp->snd_nxt = tp->iss;
651 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
652 to.to_flags |= TOF_MSS;
653 }
654 /* Window scaling. */
655 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
656 to.to_wscale = tp->request_r_scale;
657 to.to_flags |= TOF_SCALE;
658 }
659 /* Timestamps. */
660 if ((tp->t_flags & TF_RCVD_TSTMP) ||
661 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
662 to.to_tsval = ticks + tp->ts_offset;
663 to.to_tsecr = tp->ts_recent;
664 to.to_flags |= TOF_TS;
665 /* Set receive buffer autosizing timestamp. */
666 if (tp->rfbuf_ts == 0 &&
667 (so->so_rcv.sb_flags & SB_AUTOSIZE))
668 tp->rfbuf_ts = ticks;
669 }
670 /* Selective ACK's. */
671 if (tp->t_flags & TF_SACK_PERMIT) {
672 if (flags & TH_SYN)
673 to.to_flags |= TOF_SACKPERM;
674 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
675 (tp->t_flags & TF_SACK_PERMIT) &&
676 tp->rcv_numsacks > 0) {
677 to.to_flags |= TOF_SACK;
678 to.to_nsacks = tp->rcv_numsacks;
679 to.to_sacks = (u_char *)tp->sackblks;
680 }
681 }
682 #ifdef TCP_SIGNATURE
683 /* TCP-MD5 (RFC2385). */
684 if (tp->t_flags & TF_SIGNATURE)
685 to.to_flags |= TOF_SIGNATURE;
686 #endif /* TCP_SIGNATURE */
687
688 /* Processing the options. */
689 hdrlen += optlen = tcp_addoptions(&to, opt);
690 }
691
692 #ifdef INET6
693 if (isipv6)
694 ipoptlen = ip6_optlen(tp->t_inpcb);
695 else
696 #endif
697 if (tp->t_inpcb->inp_options)
698 ipoptlen = tp->t_inpcb->inp_options->m_len -
699 offsetof(struct ipoption, ipopt_list);
700 else
701 ipoptlen = 0;
702 #ifdef IPSEC
703 ipoptlen += ipsec_optlen;
704 #endif
705
706 /*
707 * Adjust data length if insertion of options will
708 * bump the packet length beyond the t_maxopd length.
709 * Clear the FIN bit because we cut off the tail of
710 * the segment.
711 *
712 * When doing TSO limit a burst to TCP_MAXWIN minus the
713 * IP, TCP and Options length to keep ip->ip_len from
714 * overflowing. Prevent the last segment from being
715 * fractional thus making them all equal sized and set
716 * the flag to continue sending. TSO is disabled when
717 * IP options or IPSEC are present.
718 */
719 if (len + optlen + ipoptlen > tp->t_maxopd) {
720 flags &= ~TH_FIN;
721 if (tso) {
722 if (len > TCP_MAXWIN - hdrlen - optlen) {
723 len = TCP_MAXWIN - hdrlen - optlen;
724 len = len - (len % (tp->t_maxopd - optlen));
725 sendalot = 1;
726 } else if (tp->t_flags & TF_NEEDFIN)
727 sendalot = 1;
728 } else {
729 len = tp->t_maxopd - optlen - ipoptlen;
730 sendalot = 1;
731 }
732 }
733
734 /*#ifdef DIAGNOSTIC*/
735 #ifdef INET6
736 if (max_linkhdr + hdrlen > MCLBYTES)
737 #else
738 if (max_linkhdr + hdrlen > MHLEN)
739 #endif
740 panic("tcphdr too big");
741 /*#endif*/
742
743 /*
744 * This KASSERT is here to catch edge cases at a well defined place.
745 * Before, those had triggered (random) panic conditions further down.
746 */
747 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
748
749 /*
750 * Grab a header mbuf, attaching a copy of data to
751 * be transmitted, and initialize the header from
752 * the template for sends on this connection.
753 */
754 if (len) {
755 struct mbuf *mb;
756 u_int moff;
757
758 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
759 tcpstat.tcps_sndprobe++;
760 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
761 tcpstat.tcps_sndrexmitpack++;
762 tcpstat.tcps_sndrexmitbyte += len;
763 } else {
764 tcpstat.tcps_sndpack++;
765 tcpstat.tcps_sndbyte += len;
766 }
767 #ifdef notyet
768 if ((m = m_copypack(so->so_snd.sb_mb, off,
769 (int)len, max_linkhdr + hdrlen)) == 0) {
770 SOCKBUF_UNLOCK(&so->so_snd);
771 error = ENOBUFS;
772 goto out;
773 }
774 /*
775 * m_copypack left space for our hdr; use it.
776 */
777 m->m_len += hdrlen;
778 m->m_data -= hdrlen;
779 #else
780 MGETHDR(m, M_DONTWAIT, MT_DATA);
781 if (m == NULL) {
782 SOCKBUF_UNLOCK(&so->so_snd);
783 error = ENOBUFS;
784 goto out;
785 }
786 #ifdef INET6
787 if (MHLEN < hdrlen + max_linkhdr) {
788 MCLGET(m, M_DONTWAIT);
789 if ((m->m_flags & M_EXT) == 0) {
790 SOCKBUF_UNLOCK(&so->so_snd);
791 m_freem(m);
792 error = ENOBUFS;
793 goto out;
794 }
795 }
796 #endif
797 m->m_data += max_linkhdr;
798 m->m_len = hdrlen;
799
800 /*
801 * Start the m_copy functions from the closest mbuf
802 * to the offset in the socket buffer chain.
803 */
804 mb = sbsndptr(&so->so_snd, off, len, &moff);
805
806 if (len <= MHLEN - hdrlen - max_linkhdr) {
807 m_copydata(mb, moff, (int)len,
808 mtod(m, caddr_t) + hdrlen);
809 m->m_len += len;
810 } else {
811 m->m_next = m_copy(mb, moff, (int)len);
812 if (m->m_next == NULL) {
813 SOCKBUF_UNLOCK(&so->so_snd);
814 (void) m_free(m);
815 error = ENOBUFS;
816 goto out;
817 }
818 }
819 #endif
820 /*
821 * If we're sending everything we've got, set PUSH.
822 * (This will keep happy those implementations which only
823 * give data to the user when a buffer fills or
824 * a PUSH comes in.)
825 */
826 if (off + len == so->so_snd.sb_cc)
827 flags |= TH_PUSH;
828 SOCKBUF_UNLOCK(&so->so_snd);
829 } else {
830 SOCKBUF_UNLOCK(&so->so_snd);
831 if (tp->t_flags & TF_ACKNOW)
832 tcpstat.tcps_sndacks++;
833 else if (flags & (TH_SYN|TH_FIN|TH_RST))
834 tcpstat.tcps_sndctrl++;
835 else if (SEQ_GT(tp->snd_up, tp->snd_una))
836 tcpstat.tcps_sndurg++;
837 else
838 tcpstat.tcps_sndwinup++;
839
840 MGETHDR(m, M_DONTWAIT, MT_DATA);
841 if (m == NULL) {
842 error = ENOBUFS;
843 goto out;
844 }
845 #ifdef INET6
846 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
847 MHLEN >= hdrlen) {
848 MH_ALIGN(m, hdrlen);
849 } else
850 #endif
851 m->m_data += max_linkhdr;
852 m->m_len = hdrlen;
853 }
854 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
855 m->m_pkthdr.rcvif = (struct ifnet *)0;
856 #ifdef MAC
857 mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
858 #endif
859 #ifdef INET6
860 if (isipv6) {
861 ip6 = mtod(m, struct ip6_hdr *);
862 th = (struct tcphdr *)(ip6 + 1);
863 tcpip_fillheaders(tp->t_inpcb, ip6, th);
864 } else
865 #endif /* INET6 */
866 {
867 ip = mtod(m, struct ip *);
868 ipov = (struct ipovly *)ip;
869 th = (struct tcphdr *)(ip + 1);
870 tcpip_fillheaders(tp->t_inpcb, ip, th);
871 }
872
873 /*
874 * Fill in fields, remembering maximum advertised
875 * window for use in delaying messages about window sizes.
876 * If resending a FIN, be sure not to use a new sequence number.
877 */
878 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
879 tp->snd_nxt == tp->snd_max)
880 tp->snd_nxt--;
881 /*
882 * If we are doing retransmissions, then snd_nxt will
883 * not reflect the first unsent octet. For ACK only
884 * packets, we do not want the sequence number of the
885 * retransmitted packet, we want the sequence number
886 * of the next unsent octet. So, if there is no data
887 * (and no SYN or FIN), use snd_max instead of snd_nxt
888 * when filling in ti_seq. But if we are in persist
889 * state, snd_max might reflect one byte beyond the
890 * right edge of the window, so use snd_nxt in that
891 * case, since we know we aren't doing a retransmission.
892 * (retransmit and persist are mutually exclusive...)
893 */
894 if (sack_rxmit == 0) {
895 if (len || (flags & (TH_SYN|TH_FIN)) ||
896 tcp_timer_active(tp, TT_PERSIST))
897 th->th_seq = htonl(tp->snd_nxt);
898 else
899 th->th_seq = htonl(tp->snd_max);
900 } else {
901 th->th_seq = htonl(p->rxmit);
902 p->rxmit += len;
903 tp->sackhint.sack_bytes_rexmit += len;
904 }
905 th->th_ack = htonl(tp->rcv_nxt);
906 if (optlen) {
907 bcopy(opt, th + 1, optlen);
908 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
909 }
910 th->th_flags = flags;
911 /*
912 * Calculate receive window. Don't shrink window,
913 * but avoid silly window syndrome.
914 */
915 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
916 recwin < (long)tp->t_maxseg)
917 recwin = 0;
918 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
919 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
920 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
921 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
922
923 /*
924 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
925 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
926 * case is handled in syncache.
927 */
928 if (flags & TH_SYN)
929 th->th_win = htons((u_short)
930 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
931 else
932 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
933
934 /*
935 * Adjust the RXWIN0SENT flag - indicate that we have advertised
936 * a 0 window. This may cause the remote transmitter to stall. This
937 * flag tells soreceive() to disable delayed acknowledgements when
938 * draining the buffer. This can occur if the receiver is attempting
939 * to read more data then can be buffered prior to transmitting on
940 * the connection.
941 */
942 if (th->th_win == 0)
943 tp->t_flags |= TF_RXWIN0SENT;
944 else
945 tp->t_flags &= ~TF_RXWIN0SENT;
946 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
947 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
948 th->th_flags |= TH_URG;
949 } else
950 /*
951 * If no urgent pointer to send, then we pull
952 * the urgent pointer to the left edge of the send window
953 * so that it doesn't drift into the send window on sequence
954 * number wraparound.
955 */
956 tp->snd_up = tp->snd_una; /* drag it along */
957
958 #ifdef TCP_SIGNATURE
959 if (tp->t_flags & TF_SIGNATURE) {
960 int sigoff = to.to_signature - opt;
961 tcp_signature_compute(m, 0, len, optlen,
962 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
963 }
964 #endif
965
966 /*
967 * Put TCP length in extended header, and then
968 * checksum extended header and data.
969 */
970 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
971 #ifdef INET6
972 if (isipv6)
973 /*
974 * ip6_plen is not need to be filled now, and will be filled
975 * in ip6_output.
976 */
977 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
978 sizeof(struct tcphdr) + optlen + len);
979 else
980 #endif /* INET6 */
981 {
982 m->m_pkthdr.csum_flags = CSUM_TCP;
983 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
984 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
985 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
986
987 /* IP version must be set here for ipv4/ipv6 checking later */
988 KASSERT(ip->ip_v == IPVERSION,
989 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
990 }
991
992 /*
993 * Enable TSO and specify the size of the segments.
994 * The TCP pseudo header checksum is always provided.
995 * XXX: Fixme: This is currently not the case for IPv6.
996 */
997 if (tso) {
998 m->m_pkthdr.csum_flags = CSUM_TSO;
999 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1000 }
1001
1002 /*
1003 * In transmit state, time the transmission and arrange for
1004 * the retransmit. In persist state, just set snd_max.
1005 */
1006 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1007 !tcp_timer_active(tp, TT_PERSIST)) {
1008 tcp_seq startseq = tp->snd_nxt;
1009
1010 /*
1011 * Advance snd_nxt over sequence space of this segment.
1012 */
1013 if (flags & (TH_SYN|TH_FIN)) {
1014 if (flags & TH_SYN)
1015 tp->snd_nxt++;
1016 if (flags & TH_FIN) {
1017 tp->snd_nxt++;
1018 tp->t_flags |= TF_SENTFIN;
1019 }
1020 }
1021 if (sack_rxmit)
1022 goto timer;
1023 tp->snd_nxt += len;
1024 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1025 tp->snd_max = tp->snd_nxt;
1026 /*
1027 * Time this transmission if not a retransmission and
1028 * not currently timing anything.
1029 */
1030 if (tp->t_rtttime == 0) {
1031 tp->t_rtttime = ticks;
1032 tp->t_rtseq = startseq;
1033 tcpstat.tcps_segstimed++;
1034 }
1035 }
1036
1037 /*
1038 * Set retransmit timer if not currently set,
1039 * and not doing a pure ack or a keep-alive probe.
1040 * Initial value for retransmit timer is smoothed
1041 * round-trip time + 2 * round-trip time variance.
1042 * Initialize shift counter which is used for backoff
1043 * of retransmit time.
1044 */
1045 timer:
1046 if (!tcp_timer_active(tp, TT_REXMT) &&
1047 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1048 (tp->snd_nxt != tp->snd_una))) {
1049 if (tcp_timer_active(tp, TT_PERSIST)) {
1050 tcp_timer_activate(tp, TT_PERSIST, 0);
1051 tp->t_rxtshift = 0;
1052 }
1053 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1054 }
1055 } else {
1056 /*
1057 * Persist case, update snd_max but since we are in
1058 * persist mode (no window) we do not update snd_nxt.
1059 */
1060 int xlen = len;
1061 if (flags & TH_SYN)
1062 ++xlen;
1063 if (flags & TH_FIN) {
1064 ++xlen;
1065 tp->t_flags |= TF_SENTFIN;
1066 }
1067 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1068 tp->snd_max = tp->snd_nxt + len;
1069 }
1070
1071 #ifdef TCPDEBUG
1072 /*
1073 * Trace.
1074 */
1075 if (so->so_options & SO_DEBUG) {
1076 u_short save = 0;
1077 #ifdef INET6
1078 if (!isipv6)
1079 #endif
1080 {
1081 save = ipov->ih_len;
1082 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1083 }
1084 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1085 #ifdef INET6
1086 if (!isipv6)
1087 #endif
1088 ipov->ih_len = save;
1089 }
1090 #endif
1091
1092 /*
1093 * Fill in IP length and desired time to live and
1094 * send to IP level. There should be a better way
1095 * to handle ttl and tos; we could keep them in
1096 * the template, but need a way to checksum without them.
1097 */
1098 /*
1099 * m->m_pkthdr.len should have been set before cksum calcuration,
1100 * because in6_cksum() need it.
1101 */
1102 #ifdef INET6
1103 if (isipv6) {
1104 /*
1105 * we separately set hoplimit for every segment, since the
1106 * user might want to change the value via setsockopt.
1107 * Also, desired default hop limit might be changed via
1108 * Neighbor Discovery.
1109 */
1110 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1111
1112 /* TODO: IPv6 IP6TOS_ECT bit on */
1113 error = ip6_output(m,
1114 tp->t_inpcb->in6p_outputopts, NULL,
1115 ((so->so_options & SO_DONTROUTE) ?
1116 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1117 } else
1118 #endif /* INET6 */
1119 {
1120 ip->ip_len = m->m_pkthdr.len;
1121 #ifdef INET6
1122 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1123 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1124 #endif /* INET6 */
1125 /*
1126 * If we do path MTU discovery, then we set DF on every packet.
1127 * This might not be the best thing to do according to RFC3390
1128 * Section 2. However the tcp hostcache migitates the problem
1129 * so it affects only the first tcp connection with a host.
1130 */
1131 if (path_mtu_discovery)
1132 ip->ip_off |= IP_DF;
1133
1134 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1135 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1136 tp->t_inpcb);
1137 }
1138 if (error) {
1139
1140 /*
1141 * We know that the packet was lost, so back out the
1142 * sequence number advance, if any.
1143 *
1144 * If the error is EPERM the packet got blocked by the
1145 * local firewall. Normally we should terminate the
1146 * connection but the blocking may have been spurious
1147 * due to a firewall reconfiguration cycle. So we treat
1148 * it like a packet loss and let the retransmit timer and
1149 * timeouts do their work over time.
1150 * XXX: It is a POLA question whether calling tcp_drop right
1151 * away would be the really correct behavior instead.
1152 */
1153 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1154 !tcp_timer_active(tp, TT_PERSIST)) &&
1155 ((flags & TH_SYN) == 0) &&
1156 (error != EPERM)) {
1157 if (sack_rxmit) {
1158 p->rxmit -= len;
1159 tp->sackhint.sack_bytes_rexmit -= len;
1160 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1161 ("sackhint bytes rtx >= 0"));
1162 } else
1163 tp->snd_nxt -= len;
1164 }
1165 out:
1166 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1167 switch (error) {
1168 case EPERM:
1169 tp->t_softerror = error;
1170 return (error);
1171 case ENOBUFS:
1172 if (!tcp_timer_active(tp, TT_REXMT) &&
1173 !tcp_timer_active(tp, TT_PERSIST))
1174 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1175 tp->snd_cwnd = tp->t_maxseg;
1176 return (0);
1177 case EMSGSIZE:
1178 /*
1179 * For some reason the interface we used initially
1180 * to send segments changed to another or lowered
1181 * its MTU.
1182 *
1183 * tcp_mtudisc() will find out the new MTU and as
1184 * its last action, initiate retransmission, so it
1185 * is important to not do so here.
1186 *
1187 * If TSO was active we either got an interface
1188 * without TSO capabilits or TSO was turned off.
1189 * Disable it for this connection as too and
1190 * immediatly retry with MSS sized segments generated
1191 * by this function.
1192 */
1193 if (tso)
1194 tp->t_flags &= ~TF_TSO;
1195 tcp_mtudisc(tp->t_inpcb, 0);
1196 return (0);
1197 case EHOSTDOWN:
1198 case EHOSTUNREACH:
1199 case ENETDOWN:
1200 case ENETUNREACH:
1201 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1202 tp->t_softerror = error;
1203 return (0);
1204 }
1205 /* FALLTHROUGH */
1206 default:
1207 return (error);
1208 }
1209 }
1210 tcpstat.tcps_sndtotal++;
1211
1212 /*
1213 * Data sent (as far as we can tell).
1214 * If this advertises a larger window than any other segment,
1215 * then remember the size of the advertised window.
1216 * Any pending ACK has now been sent.
1217 */
1218 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1219 tp->rcv_adv = tp->rcv_nxt + recwin;
1220 tp->last_ack_sent = tp->rcv_nxt;
1221 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1222 if (tcp_timer_active(tp, TT_DELACK))
1223 tcp_timer_activate(tp, TT_DELACK, 0);
1224 #if 0
1225 /*
1226 * This completely breaks TCP if newreno is turned on. What happens
1227 * is that if delayed-acks are turned on on the receiver, this code
1228 * on the transmitter effectively destroys the TCP window, forcing
1229 * it to four packets (1.5Kx4 = 6K window).
1230 */
1231 if (sendalot && (!tcp_do_newreno || --maxburst))
1232 goto again;
1233 #endif
1234 if (sendalot)
1235 goto again;
1236 return (0);
1237 }
1238
1239 void
1240 tcp_setpersist(struct tcpcb *tp)
1241 {
1242 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1243 int tt;
1244
1245 if (tcp_timer_active(tp, TT_REXMT))
1246 panic("tcp_setpersist: retransmit pending");
1247 /*
1248 * Start/restart persistance timer.
1249 */
1250 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1251 TCPTV_PERSMIN, TCPTV_PERSMAX);
1252 tcp_timer_activate(tp, TT_PERSIST, tt);
1253 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1254 tp->t_rxtshift++;
1255 }
1256
1257 /*
1258 * Insert TCP options according to the supplied parameters to the place
1259 * optp in a consistent way. Can handle unaligned destinations.
1260 *
1261 * The order of the option processing is crucial for optimal packing and
1262 * alignment for the scarce option space.
1263 *
1264 * The optimal order for a SYN/SYN-ACK segment is:
1265 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1266 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1267 *
1268 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1269 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1270 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1271 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1272 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1273 */
1274 int
1275 tcp_addoptions(struct tcpopt *to, u_char *optp)
1276 {
1277 u_int mask, optlen = 0;
1278
1279 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1280 if ((to->to_flags & mask) != mask)
1281 continue;
1282 if (optlen == TCP_MAXOLEN)
1283 break;
1284 switch (to->to_flags & mask) {
1285 case TOF_MSS:
1286 while (optlen % 4) {
1287 optlen += TCPOLEN_NOP;
1288 *optp++ = TCPOPT_NOP;
1289 }
1290 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1291 continue;
1292 optlen += TCPOLEN_MAXSEG;
1293 *optp++ = TCPOPT_MAXSEG;
1294 *optp++ = TCPOLEN_MAXSEG;
1295 to->to_mss = htons(to->to_mss);
1296 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1297 optp += sizeof(to->to_mss);
1298 break;
1299 case TOF_SCALE:
1300 while (!optlen || optlen % 2 != 1) {
1301 optlen += TCPOLEN_NOP;
1302 *optp++ = TCPOPT_NOP;
1303 }
1304 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1305 continue;
1306 optlen += TCPOLEN_WINDOW;
1307 *optp++ = TCPOPT_WINDOW;
1308 *optp++ = TCPOLEN_WINDOW;
1309 *optp++ = to->to_wscale;
1310 break;
1311 case TOF_SACKPERM:
1312 while (optlen % 2) {
1313 optlen += TCPOLEN_NOP;
1314 *optp++ = TCPOPT_NOP;
1315 }
1316 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1317 continue;
1318 optlen += TCPOLEN_SACK_PERMITTED;
1319 *optp++ = TCPOPT_SACK_PERMITTED;
1320 *optp++ = TCPOLEN_SACK_PERMITTED;
1321 break;
1322 case TOF_TS:
1323 while (!optlen || optlen % 4 != 2) {
1324 optlen += TCPOLEN_NOP;
1325 *optp++ = TCPOPT_NOP;
1326 }
1327 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1328 continue;
1329 optlen += TCPOLEN_TIMESTAMP;
1330 *optp++ = TCPOPT_TIMESTAMP;
1331 *optp++ = TCPOLEN_TIMESTAMP;
1332 to->to_tsval = htonl(to->to_tsval);
1333 to->to_tsecr = htonl(to->to_tsecr);
1334 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1335 optp += sizeof(to->to_tsval);
1336 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1337 optp += sizeof(to->to_tsecr);
1338 break;
1339 case TOF_SIGNATURE:
1340 {
1341 int siglen = TCPOLEN_SIGNATURE - 2;
1342
1343 while (!optlen || optlen % 4 != 2) {
1344 optlen += TCPOLEN_NOP;
1345 *optp++ = TCPOPT_NOP;
1346 }
1347 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1348 continue;
1349 optlen += TCPOLEN_SIGNATURE;
1350 *optp++ = TCPOPT_SIGNATURE;
1351 *optp++ = TCPOLEN_SIGNATURE;
1352 to->to_signature = optp;
1353 while (siglen--)
1354 *optp++ = 0;
1355 break;
1356 }
1357 case TOF_SACK:
1358 {
1359 int sackblks = 0;
1360 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1361 tcp_seq sack_seq;
1362
1363 while (!optlen || optlen % 4 != 2) {
1364 optlen += TCPOLEN_NOP;
1365 *optp++ = TCPOPT_NOP;
1366 }
1367 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1368 continue;
1369 optlen += TCPOLEN_SACKHDR;
1370 *optp++ = TCPOPT_SACK;
1371 sackblks = min(to->to_nsacks,
1372 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1373 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1374 while (sackblks--) {
1375 sack_seq = htonl(sack->start);
1376 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1377 optp += sizeof(sack_seq);
1378 sack_seq = htonl(sack->end);
1379 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1380 optp += sizeof(sack_seq);
1381 optlen += TCPOLEN_SACK;
1382 sack++;
1383 }
1384 tcpstat.tcps_sack_send_blocks++;
1385 break;
1386 }
1387 default:
1388 panic("%s: unknown TCP option type", __func__);
1389 break;
1390 }
1391 }
1392
1393 /* Terminate and pad TCP options to a 4 byte boundary. */
1394 if (optlen % 4) {
1395 optlen += TCPOLEN_EOL;
1396 *optp++ = TCPOPT_EOL;
1397 }
1398 /*
1399 * According to RFC 793 (STD0007):
1400 * "The content of the header beyond the End-of-Option option
1401 * must be header padding (i.e., zero)."
1402 * and later: "The padding is composed of zeros."
1403 */
1404 while (optlen % 4) {
1405 optlen += TCPOLEN_PAD;
1406 *optp++ = TCPOPT_PAD;
1407 }
1408
1409 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1410 return (optlen);
1411 }
Cache object: 598b75c23541332edd2579b4d25d9596
|