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/10.1/sys/netinet/tcp_output.c 293894 2016-01-14 09:11:16Z glebius $");
34
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_kdtrace.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/hhook.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/mbuf.h>
48 #include <sys/mutex.h>
49 #include <sys/protosw.h>
50 #include <sys/sdt.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
54
55 #include <net/if.h>
56 #include <net/route.h>
57 #include <net/vnet.h>
58
59 #include <netinet/cc.h>
60 #include <netinet/in.h>
61 #include <netinet/in_kdtrace.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/ip_options.h>
67 #ifdef INET6
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/ip6_var.h>
71 #endif
72 #define TCPOUTFLAGS
73 #include <netinet/tcp_fsm.h>
74 #include <netinet/tcp_seq.h>
75 #include <netinet/tcp_timer.h>
76 #include <netinet/tcp_var.h>
77 #include <netinet/tcpip.h>
78 #ifdef TCPDEBUG
79 #include <netinet/tcp_debug.h>
80 #endif
81 #ifdef TCP_OFFLOAD
82 #include <netinet/tcp_offload.h>
83 #endif
84
85 #ifdef IPSEC
86 #include <netipsec/ipsec.h>
87 #endif /*IPSEC*/
88
89 #include <machine/in_cksum.h>
90
91 #include <security/mac/mac_framework.h>
92
93 VNET_DEFINE(int, path_mtu_discovery) = 1;
94 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
95 &VNET_NAME(path_mtu_discovery), 1,
96 "Enable Path MTU Discovery");
97
98 VNET_DEFINE(int, tcp_do_tso) = 1;
99 #define V_tcp_do_tso VNET(tcp_do_tso)
100 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
101 &VNET_NAME(tcp_do_tso), 0,
102 "Enable TCP Segmentation Offload");
103
104 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
105 #define V_tcp_sendspace VNET(tcp_sendspace)
106 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
107 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
108
109 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
110 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
111 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
112 &VNET_NAME(tcp_do_autosndbuf), 0,
113 "Enable automatic send buffer sizing");
114
115 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
116 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
117 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
118 &VNET_NAME(tcp_autosndbuf_inc), 0,
119 "Incrementor step size of automatic send buffer");
120
121 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
122 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
123 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
124 &VNET_NAME(tcp_autosndbuf_max), 0,
125 "Max size of automatic send buffer");
126
127 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
128 struct tcphdr *th, struct tcpopt *to,
129 long len, int tso);
130 static void inline cc_after_idle(struct tcpcb *tp);
131
132 /*
133 * Wrapper for the TCP established output helper hook.
134 */
135 static void inline
136 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
137 struct tcpopt *to, long len, int tso)
138 {
139 struct tcp_hhook_data hhook_data;
140
141 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
142 hhook_data.tp = tp;
143 hhook_data.th = th;
144 hhook_data.to = to;
145 hhook_data.len = len;
146 hhook_data.tso = tso;
147
148 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
149 tp->osd);
150 }
151 }
152
153 /*
154 * CC wrapper hook functions
155 */
156 static void inline
157 cc_after_idle(struct tcpcb *tp)
158 {
159 INP_WLOCK_ASSERT(tp->t_inpcb);
160
161 if (CC_ALGO(tp)->after_idle != NULL)
162 CC_ALGO(tp)->after_idle(tp->ccv);
163 }
164
165 /*
166 * Tcp output routine: figure out what should be sent and send it.
167 */
168 int
169 tcp_output(struct tcpcb *tp)
170 {
171 struct socket *so = tp->t_inpcb->inp_socket;
172 long len, recwin, sendwin;
173 int off, flags, error = 0; /* Keep compiler happy */
174 struct mbuf *m;
175 struct ip *ip = NULL;
176 struct ipovly *ipov = NULL;
177 struct tcphdr *th;
178 u_char opt[TCP_MAXOLEN];
179 unsigned ipoptlen, optlen, hdrlen;
180 #ifdef IPSEC
181 unsigned ipsec_optlen = 0;
182 #endif
183 int idle, sendalot;
184 int sack_rxmit, sack_bytes_rxmt;
185 struct sackhole *p;
186 int tso, mtu;
187 struct tcpopt to;
188 #if 0
189 int maxburst = TCP_MAXBURST;
190 #endif
191 #ifdef INET6
192 struct ip6_hdr *ip6 = NULL;
193 int isipv6;
194
195 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
196 #endif
197
198 INP_WLOCK_ASSERT(tp->t_inpcb);
199
200 #ifdef TCP_OFFLOAD
201 if (tp->t_flags & TF_TOE)
202 return (tcp_offload_output(tp));
203 #endif
204
205 /*
206 * Determine length of data that should be transmitted,
207 * and flags that will be used.
208 * If there is some data or critical controls (SYN, RST)
209 * to send, then transmit; otherwise, investigate further.
210 */
211 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
212 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
213 cc_after_idle(tp);
214 tp->t_flags &= ~TF_LASTIDLE;
215 if (idle) {
216 if (tp->t_flags & TF_MORETOCOME) {
217 tp->t_flags |= TF_LASTIDLE;
218 idle = 0;
219 }
220 }
221 again:
222 /*
223 * If we've recently taken a timeout, snd_max will be greater than
224 * snd_nxt. There may be SACK information that allows us to avoid
225 * resending already delivered data. Adjust snd_nxt accordingly.
226 */
227 if ((tp->t_flags & TF_SACK_PERMIT) &&
228 SEQ_LT(tp->snd_nxt, tp->snd_max))
229 tcp_sack_adjust(tp);
230 sendalot = 0;
231 tso = 0;
232 mtu = 0;
233 off = tp->snd_nxt - tp->snd_una;
234 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
235
236 flags = tcp_outflags[tp->t_state];
237 /*
238 * Send any SACK-generated retransmissions. If we're explicitly trying
239 * to send out new data (when sendalot is 1), bypass this function.
240 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
241 * we're replacing a (future) new transmission with a retransmission
242 * now, and we previously incremented snd_cwnd in tcp_input().
243 */
244 /*
245 * Still in sack recovery , reset rxmit flag to zero.
246 */
247 sack_rxmit = 0;
248 sack_bytes_rxmt = 0;
249 len = 0;
250 p = NULL;
251 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
252 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
253 long cwin;
254
255 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
256 if (cwin < 0)
257 cwin = 0;
258 /* Do not retransmit SACK segments beyond snd_recover */
259 if (SEQ_GT(p->end, tp->snd_recover)) {
260 /*
261 * (At least) part of sack hole extends beyond
262 * snd_recover. Check to see if we can rexmit data
263 * for this hole.
264 */
265 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
266 /*
267 * Can't rexmit any more data for this hole.
268 * That data will be rexmitted in the next
269 * sack recovery episode, when snd_recover
270 * moves past p->rxmit.
271 */
272 p = NULL;
273 goto after_sack_rexmit;
274 } else
275 /* Can rexmit part of the current hole */
276 len = ((long)ulmin(cwin,
277 tp->snd_recover - p->rxmit));
278 } else
279 len = ((long)ulmin(cwin, p->end - p->rxmit));
280 off = p->rxmit - tp->snd_una;
281 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
282 __func__, off));
283 if (len > 0) {
284 sack_rxmit = 1;
285 sendalot = 1;
286 TCPSTAT_INC(tcps_sack_rexmits);
287 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
288 min(len, tp->t_maxseg));
289 }
290 }
291 after_sack_rexmit:
292 /*
293 * Get standard flags, and add SYN or FIN if requested by 'hidden'
294 * state flags.
295 */
296 if (tp->t_flags & TF_NEEDFIN)
297 flags |= TH_FIN;
298 if (tp->t_flags & TF_NEEDSYN)
299 flags |= TH_SYN;
300
301 SOCKBUF_LOCK(&so->so_snd);
302 /*
303 * If in persist timeout with window of 0, send 1 byte.
304 * Otherwise, if window is small but nonzero
305 * and timer expired, we will send what we can
306 * and go to transmit state.
307 */
308 if (tp->t_flags & TF_FORCEDATA) {
309 if (sendwin == 0) {
310 /*
311 * If we still have some data to send, then
312 * clear the FIN bit. Usually this would
313 * happen below when it realizes that we
314 * aren't sending all the data. However,
315 * if we have exactly 1 byte of unsent data,
316 * then it won't clear the FIN bit below,
317 * and if we are in persist state, we wind
318 * up sending the packet without recording
319 * that we sent the FIN bit.
320 *
321 * We can't just blindly clear the FIN bit,
322 * because if we don't have any more data
323 * to send then the probe will be the FIN
324 * itself.
325 */
326 if (off < so->so_snd.sb_cc)
327 flags &= ~TH_FIN;
328 sendwin = 1;
329 } else {
330 tcp_timer_activate(tp, TT_PERSIST, 0);
331 tp->t_rxtshift = 0;
332 }
333 }
334
335 /*
336 * If snd_nxt == snd_max and we have transmitted a FIN, the
337 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
338 * a negative length. This can also occur when TCP opens up
339 * its congestion window while receiving additional duplicate
340 * acks after fast-retransmit because TCP will reset snd_nxt
341 * to snd_max after the fast-retransmit.
342 *
343 * In the normal retransmit-FIN-only case, however, snd_nxt will
344 * be set to snd_una, the offset will be 0, and the length may
345 * wind up 0.
346 *
347 * If sack_rxmit is true we are retransmitting from the scoreboard
348 * in which case len is already set.
349 */
350 if (sack_rxmit == 0) {
351 if (sack_bytes_rxmt == 0)
352 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
353 else {
354 long cwin;
355
356 /*
357 * We are inside of a SACK recovery episode and are
358 * sending new data, having retransmitted all the
359 * data possible in the scoreboard.
360 */
361 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
362 - off);
363 /*
364 * Don't remove this (len > 0) check !
365 * We explicitly check for len > 0 here (although it
366 * isn't really necessary), to work around a gcc
367 * optimization issue - to force gcc to compute
368 * len above. Without this check, the computation
369 * of len is bungled by the optimizer.
370 */
371 if (len > 0) {
372 cwin = tp->snd_cwnd -
373 (tp->snd_nxt - tp->sack_newdata) -
374 sack_bytes_rxmt;
375 if (cwin < 0)
376 cwin = 0;
377 len = lmin(len, cwin);
378 }
379 }
380 }
381
382 /*
383 * Lop off SYN bit if it has already been sent. However, if this
384 * is SYN-SENT state and if segment contains data and if we don't
385 * know that foreign host supports TAO, suppress sending segment.
386 */
387 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
388 if (tp->t_state != TCPS_SYN_RECEIVED)
389 flags &= ~TH_SYN;
390 off--, len++;
391 }
392
393 /*
394 * Be careful not to send data and/or FIN on SYN segments.
395 * This measure is needed to prevent interoperability problems
396 * with not fully conformant TCP implementations.
397 */
398 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
399 len = 0;
400 flags &= ~TH_FIN;
401 }
402
403 if (len <= 0) {
404 /*
405 * If FIN has been sent but not acked,
406 * but we haven't been called to retransmit,
407 * len will be < 0. Otherwise, window shrank
408 * after we sent into it. If window shrank to 0,
409 * cancel pending retransmit, pull snd_nxt back
410 * to (closed) window, and set the persist timer
411 * if it isn't already going. If the window didn't
412 * close completely, just wait for an ACK.
413 *
414 * We also do a general check here to ensure that
415 * we will set the persist timer when we have data
416 * to send, but a 0-byte window. This makes sure
417 * the persist timer is set even if the packet
418 * hits one of the "goto send" lines below.
419 */
420 len = 0;
421 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
422 (off < (int) so->so_snd.sb_cc)) {
423 tcp_timer_activate(tp, TT_REXMT, 0);
424 tp->t_rxtshift = 0;
425 tp->snd_nxt = tp->snd_una;
426 if (!tcp_timer_active(tp, TT_PERSIST))
427 tcp_setpersist(tp);
428 }
429 }
430
431 /* len will be >= 0 after this point. */
432 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
433
434 /*
435 * Automatic sizing of send socket buffer. Often the send buffer
436 * size is not optimally adjusted to the actual network conditions
437 * at hand (delay bandwidth product). Setting the buffer size too
438 * small limits throughput on links with high bandwidth and high
439 * delay (eg. trans-continental/oceanic links). Setting the
440 * buffer size too big consumes too much real kernel memory,
441 * especially with many connections on busy servers.
442 *
443 * The criteria to step up the send buffer one notch are:
444 * 1. receive window of remote host is larger than send buffer
445 * (with a fudge factor of 5/4th);
446 * 2. send buffer is filled to 7/8th with data (so we actually
447 * have data to make use of it);
448 * 3. send buffer fill has not hit maximal automatic size;
449 * 4. our send window (slow start and cogestion controlled) is
450 * larger than sent but unacknowledged data in send buffer.
451 *
452 * The remote host receive window scaling factor may limit the
453 * growing of the send buffer before it reaches its allowed
454 * maximum.
455 *
456 * It scales directly with slow start or congestion window
457 * and does at most one step per received ACK. This fast
458 * scaling has the drawback of growing the send buffer beyond
459 * what is strictly necessary to make full use of a given
460 * delay*bandwith product. However testing has shown this not
461 * to be much of an problem. At worst we are trading wasting
462 * of available bandwith (the non-use of it) for wasting some
463 * socket buffer memory.
464 *
465 * TODO: Shrink send buffer during idle periods together
466 * with congestion window. Requires another timer. Has to
467 * wait for upcoming tcp timer rewrite.
468 */
469 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
470 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
471 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
472 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
473 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
474 if (!sbreserve_locked(&so->so_snd,
475 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
476 V_tcp_autosndbuf_max), so, curthread))
477 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
478 }
479 }
480
481 /*
482 * Decide if we can use TCP Segmentation Offloading (if supported by
483 * hardware).
484 *
485 * TSO may only be used if we are in a pure bulk sending state. The
486 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
487 * IP options prevent using TSO. With TSO the TCP header is the same
488 * (except for the sequence number) for all generated packets. This
489 * makes it impossible to transmit any options which vary per generated
490 * segment or packet.
491 */
492 #ifdef IPSEC
493 /*
494 * Pre-calculate here as we save another lookup into the darknesses
495 * of IPsec that way and can actually decide if TSO is ok.
496 */
497 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
498 #endif
499 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
500 ((tp->t_flags & TF_SIGNATURE) == 0) &&
501 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
502 #ifdef IPSEC
503 ipsec_optlen == 0 &&
504 #endif
505 tp->t_inpcb->inp_options == NULL &&
506 tp->t_inpcb->in6p_options == NULL)
507 tso = 1;
508
509 if (sack_rxmit) {
510 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
511 flags &= ~TH_FIN;
512 } else {
513 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
514 flags &= ~TH_FIN;
515 }
516
517 recwin = sbspace(&so->so_rcv);
518
519 /*
520 * Sender silly window avoidance. We transmit under the following
521 * conditions when len is non-zero:
522 *
523 * - We have a full segment (or more with TSO)
524 * - This is the last buffer in a write()/send() and we are
525 * either idle or running NODELAY
526 * - we've timed out (e.g. persist timer)
527 * - we have more then 1/2 the maximum send window's worth of
528 * data (receiver may be limited the window size)
529 * - we need to retransmit
530 */
531 if (len) {
532 if (len >= tp->t_maxseg)
533 goto send;
534 /*
535 * NOTE! on localhost connections an 'ack' from the remote
536 * end may occur synchronously with the output and cause
537 * us to flush a buffer queued with moretocome. XXX
538 *
539 * note: the len + off check is almost certainly unnecessary.
540 */
541 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
542 (idle || (tp->t_flags & TF_NODELAY)) &&
543 len + off >= so->so_snd.sb_cc &&
544 (tp->t_flags & TF_NOPUSH) == 0) {
545 goto send;
546 }
547 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
548 goto send;
549 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
550 goto send;
551 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
552 goto send;
553 if (sack_rxmit)
554 goto send;
555 }
556
557 /*
558 * Sending of standalone window updates.
559 *
560 * Window updates are important when we close our window due to a
561 * full socket buffer and are opening it again after the application
562 * reads data from it. Once the window has opened again and the
563 * remote end starts to send again the ACK clock takes over and
564 * provides the most current window information.
565 *
566 * We must avoid the silly window syndrome whereas every read
567 * from the receive buffer, no matter how small, causes a window
568 * update to be sent. We also should avoid sending a flurry of
569 * window updates when the socket buffer had queued a lot of data
570 * and the application is doing small reads.
571 *
572 * Prevent a flurry of pointless window updates by only sending
573 * an update when we can increase the advertized window by more
574 * than 1/4th of the socket buffer capacity. When the buffer is
575 * getting full or is very small be more aggressive and send an
576 * update whenever we can increase by two mss sized segments.
577 * In all other situations the ACK's to new incoming data will
578 * carry further window increases.
579 *
580 * Don't send an independent window update if a delayed
581 * ACK is pending (it will get piggy-backed on it) or the
582 * remote side already has done a half-close and won't send
583 * more data. Skip this if the connection is in T/TCP
584 * half-open state.
585 */
586 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
587 !(tp->t_flags & TF_DELACK) &&
588 !TCPS_HAVERCVDFIN(tp->t_state)) {
589 /*
590 * "adv" is the amount we could increase the window,
591 * taking into account that we are limited by
592 * TCP_MAXWIN << tp->rcv_scale.
593 */
594 long adv;
595 int oldwin;
596
597 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
598 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
599 oldwin = (tp->rcv_adv - tp->rcv_nxt);
600 adv -= oldwin;
601 } else
602 oldwin = 0;
603
604 /*
605 * If the new window size ends up being the same as the old
606 * size when it is scaled, then don't force a window update.
607 */
608 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
609 goto dontupdate;
610
611 if (adv >= (long)(2 * tp->t_maxseg) &&
612 (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
613 recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
614 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
615 goto send;
616 }
617 dontupdate:
618
619 /*
620 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
621 * is also a catch-all for the retransmit timer timeout case.
622 */
623 if (tp->t_flags & TF_ACKNOW)
624 goto send;
625 if ((flags & TH_RST) ||
626 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
627 goto send;
628 if (SEQ_GT(tp->snd_up, tp->snd_una))
629 goto send;
630 /*
631 * If our state indicates that FIN should be sent
632 * and we have not yet done so, then we need to send.
633 */
634 if (flags & TH_FIN &&
635 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
636 goto send;
637 /*
638 * In SACK, it is possible for tcp_output to fail to send a segment
639 * after the retransmission timer has been turned off. Make sure
640 * that the retransmission timer is set.
641 */
642 if ((tp->t_flags & TF_SACK_PERMIT) &&
643 SEQ_GT(tp->snd_max, tp->snd_una) &&
644 !tcp_timer_active(tp, TT_REXMT) &&
645 !tcp_timer_active(tp, TT_PERSIST)) {
646 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
647 goto just_return;
648 }
649 /*
650 * TCP window updates are not reliable, rather a polling protocol
651 * using ``persist'' packets is used to insure receipt of window
652 * updates. The three ``states'' for the output side are:
653 * idle not doing retransmits or persists
654 * persisting to move a small or zero window
655 * (re)transmitting and thereby not persisting
656 *
657 * tcp_timer_active(tp, TT_PERSIST)
658 * is true when we are in persist state.
659 * (tp->t_flags & TF_FORCEDATA)
660 * is set when we are called to send a persist packet.
661 * tcp_timer_active(tp, TT_REXMT)
662 * is set when we are retransmitting
663 * The output side is idle when both timers are zero.
664 *
665 * If send window is too small, there is data to transmit, and no
666 * retransmit or persist is pending, then go to persist state.
667 * If nothing happens soon, send when timer expires:
668 * if window is nonzero, transmit what we can,
669 * otherwise force out a byte.
670 */
671 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
672 !tcp_timer_active(tp, TT_PERSIST)) {
673 tp->t_rxtshift = 0;
674 tcp_setpersist(tp);
675 }
676
677 /*
678 * No reason to send a segment, just return.
679 */
680 just_return:
681 SOCKBUF_UNLOCK(&so->so_snd);
682 return (0);
683
684 send:
685 SOCKBUF_LOCK_ASSERT(&so->so_snd);
686 /*
687 * Before ESTABLISHED, force sending of initial options
688 * unless TCP set not to do any options.
689 * NOTE: we assume that the IP/TCP header plus TCP options
690 * always fit in a single mbuf, leaving room for a maximum
691 * link header, i.e.
692 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
693 */
694 optlen = 0;
695 #ifdef INET6
696 if (isipv6)
697 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
698 else
699 #endif
700 hdrlen = sizeof (struct tcpiphdr);
701
702 /*
703 * Compute options for segment.
704 * We only have to care about SYN and established connection
705 * segments. Options for SYN-ACK segments are handled in TCP
706 * syncache.
707 */
708 to.to_flags = 0;
709 if ((tp->t_flags & TF_NOOPT) == 0) {
710 /* Maximum segment size. */
711 if (flags & TH_SYN) {
712 tp->snd_nxt = tp->iss;
713 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
714 to.to_flags |= TOF_MSS;
715 }
716 /* Window scaling. */
717 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
718 to.to_wscale = tp->request_r_scale;
719 to.to_flags |= TOF_SCALE;
720 }
721 /* Timestamps. */
722 if ((tp->t_flags & TF_RCVD_TSTMP) ||
723 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
724 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
725 to.to_tsecr = tp->ts_recent;
726 to.to_flags |= TOF_TS;
727 /* Set receive buffer autosizing timestamp. */
728 if (tp->rfbuf_ts == 0 &&
729 (so->so_rcv.sb_flags & SB_AUTOSIZE))
730 tp->rfbuf_ts = tcp_ts_getticks();
731 }
732 /* Selective ACK's. */
733 if (tp->t_flags & TF_SACK_PERMIT) {
734 if (flags & TH_SYN)
735 to.to_flags |= TOF_SACKPERM;
736 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
737 (tp->t_flags & TF_SACK_PERMIT) &&
738 tp->rcv_numsacks > 0) {
739 to.to_flags |= TOF_SACK;
740 to.to_nsacks = tp->rcv_numsacks;
741 to.to_sacks = (u_char *)tp->sackblks;
742 }
743 }
744 #ifdef TCP_SIGNATURE
745 /* TCP-MD5 (RFC2385). */
746 if (tp->t_flags & TF_SIGNATURE)
747 to.to_flags |= TOF_SIGNATURE;
748 #endif /* TCP_SIGNATURE */
749
750 /* Processing the options. */
751 hdrlen += optlen = tcp_addoptions(&to, opt);
752 }
753
754 #ifdef INET6
755 if (isipv6)
756 ipoptlen = ip6_optlen(tp->t_inpcb);
757 else
758 #endif
759 if (tp->t_inpcb->inp_options)
760 ipoptlen = tp->t_inpcb->inp_options->m_len -
761 offsetof(struct ipoption, ipopt_list);
762 else
763 ipoptlen = 0;
764 #ifdef IPSEC
765 ipoptlen += ipsec_optlen;
766 #endif
767
768 /*
769 * Adjust data length if insertion of options will
770 * bump the packet length beyond the t_maxopd length.
771 * Clear the FIN bit because we cut off the tail of
772 * the segment.
773 */
774 if (len + optlen + ipoptlen > tp->t_maxopd) {
775 flags &= ~TH_FIN;
776
777 if (tso) {
778 KASSERT(ipoptlen == 0,
779 ("%s: TSO can't do IP options", __func__));
780
781 /*
782 * Limit a burst to t_tsomax minus IP,
783 * TCP and options length to keep ip->ip_len
784 * from overflowing or exceeding the maximum
785 * length allowed by the network interface.
786 */
787 if (len > tp->t_tsomax - hdrlen) {
788 len = tp->t_tsomax - hdrlen;
789 sendalot = 1;
790 }
791
792 /*
793 * Prevent the last segment from being
794 * fractional unless the send sockbuf can
795 * be emptied.
796 */
797 if (sendalot && off + len < so->so_snd.sb_cc) {
798 len -= len % (tp->t_maxopd - optlen);
799 sendalot = 1;
800 }
801
802 /*
803 * Send the FIN in a separate segment
804 * after the bulk sending is done.
805 * We don't trust the TSO implementations
806 * to clear the FIN flag on all but the
807 * last segment.
808 */
809 if (tp->t_flags & TF_NEEDFIN)
810 sendalot = 1;
811
812 } else {
813 len = tp->t_maxopd - optlen - ipoptlen;
814 sendalot = 1;
815 }
816 } else
817 tso = 0;
818
819 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
820 ("%s: len > IP_MAXPACKET", __func__));
821
822 /*#ifdef DIAGNOSTIC*/
823 #ifdef INET6
824 if (max_linkhdr + hdrlen > MCLBYTES)
825 #else
826 if (max_linkhdr + hdrlen > MHLEN)
827 #endif
828 panic("tcphdr too big");
829 /*#endif*/
830
831 /*
832 * This KASSERT is here to catch edge cases at a well defined place.
833 * Before, those had triggered (random) panic conditions further down.
834 */
835 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
836
837 /*
838 * Grab a header mbuf, attaching a copy of data to
839 * be transmitted, and initialize the header from
840 * the template for sends on this connection.
841 */
842 if (len) {
843 struct mbuf *mb;
844 u_int moff;
845
846 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
847 TCPSTAT_INC(tcps_sndprobe);
848 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
849 tp->t_sndrexmitpack++;
850 TCPSTAT_INC(tcps_sndrexmitpack);
851 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
852 } else {
853 TCPSTAT_INC(tcps_sndpack);
854 TCPSTAT_ADD(tcps_sndbyte, len);
855 }
856 #ifdef INET6
857 if (MHLEN < hdrlen + max_linkhdr)
858 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
859 else
860 #endif
861 m = m_gethdr(M_NOWAIT, MT_DATA);
862
863 if (m == NULL) {
864 SOCKBUF_UNLOCK(&so->so_snd);
865 error = ENOBUFS;
866 sack_rxmit = 0;
867 goto out;
868 }
869
870 m->m_data += max_linkhdr;
871 m->m_len = hdrlen;
872
873 /*
874 * Start the m_copy functions from the closest mbuf
875 * to the offset in the socket buffer chain.
876 */
877 mb = sbsndptr(&so->so_snd, off, len, &moff);
878
879 if (len <= MHLEN - hdrlen - max_linkhdr) {
880 m_copydata(mb, moff, (int)len,
881 mtod(m, caddr_t) + hdrlen);
882 m->m_len += len;
883 } else {
884 m->m_next = m_copy(mb, moff, (int)len);
885 if (m->m_next == NULL) {
886 SOCKBUF_UNLOCK(&so->so_snd);
887 (void) m_free(m);
888 error = ENOBUFS;
889 sack_rxmit = 0;
890 goto out;
891 }
892 }
893
894 /*
895 * If we're sending everything we've got, set PUSH.
896 * (This will keep happy those implementations which only
897 * give data to the user when a buffer fills or
898 * a PUSH comes in.)
899 */
900 if (off + len == so->so_snd.sb_cc)
901 flags |= TH_PUSH;
902 SOCKBUF_UNLOCK(&so->so_snd);
903 } else {
904 SOCKBUF_UNLOCK(&so->so_snd);
905 if (tp->t_flags & TF_ACKNOW)
906 TCPSTAT_INC(tcps_sndacks);
907 else if (flags & (TH_SYN|TH_FIN|TH_RST))
908 TCPSTAT_INC(tcps_sndctrl);
909 else if (SEQ_GT(tp->snd_up, tp->snd_una))
910 TCPSTAT_INC(tcps_sndurg);
911 else
912 TCPSTAT_INC(tcps_sndwinup);
913
914 m = m_gethdr(M_NOWAIT, MT_DATA);
915 if (m == NULL) {
916 error = ENOBUFS;
917 sack_rxmit = 0;
918 goto out;
919 }
920 #ifdef INET6
921 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
922 MHLEN >= hdrlen) {
923 MH_ALIGN(m, hdrlen);
924 } else
925 #endif
926 m->m_data += max_linkhdr;
927 m->m_len = hdrlen;
928 }
929 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
930 m->m_pkthdr.rcvif = (struct ifnet *)0;
931 #ifdef MAC
932 mac_inpcb_create_mbuf(tp->t_inpcb, m);
933 #endif
934 #ifdef INET6
935 if (isipv6) {
936 ip6 = mtod(m, struct ip6_hdr *);
937 th = (struct tcphdr *)(ip6 + 1);
938 tcpip_fillheaders(tp->t_inpcb, ip6, th);
939 } else
940 #endif /* INET6 */
941 {
942 ip = mtod(m, struct ip *);
943 ipov = (struct ipovly *)ip;
944 th = (struct tcphdr *)(ip + 1);
945 tcpip_fillheaders(tp->t_inpcb, ip, th);
946 }
947
948 /*
949 * Fill in fields, remembering maximum advertised
950 * window for use in delaying messages about window sizes.
951 * If resending a FIN, be sure not to use a new sequence number.
952 */
953 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
954 tp->snd_nxt == tp->snd_max)
955 tp->snd_nxt--;
956 /*
957 * If we are starting a connection, send ECN setup
958 * SYN packet. If we are on a retransmit, we may
959 * resend those bits a number of times as per
960 * RFC 3168.
961 */
962 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
963 if (tp->t_rxtshift >= 1) {
964 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
965 flags |= TH_ECE|TH_CWR;
966 } else
967 flags |= TH_ECE|TH_CWR;
968 }
969
970 if (tp->t_state == TCPS_ESTABLISHED &&
971 (tp->t_flags & TF_ECN_PERMIT)) {
972 /*
973 * If the peer has ECN, mark data packets with
974 * ECN capable transmission (ECT).
975 * Ignore pure ack packets, retransmissions and window probes.
976 */
977 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
978 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
979 #ifdef INET6
980 if (isipv6)
981 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
982 else
983 #endif
984 ip->ip_tos |= IPTOS_ECN_ECT0;
985 TCPSTAT_INC(tcps_ecn_ect0);
986 }
987
988 /*
989 * Reply with proper ECN notifications.
990 */
991 if (tp->t_flags & TF_ECN_SND_CWR) {
992 flags |= TH_CWR;
993 tp->t_flags &= ~TF_ECN_SND_CWR;
994 }
995 if (tp->t_flags & TF_ECN_SND_ECE)
996 flags |= TH_ECE;
997 }
998
999 /*
1000 * If we are doing retransmissions, then snd_nxt will
1001 * not reflect the first unsent octet. For ACK only
1002 * packets, we do not want the sequence number of the
1003 * retransmitted packet, we want the sequence number
1004 * of the next unsent octet. So, if there is no data
1005 * (and no SYN or FIN), use snd_max instead of snd_nxt
1006 * when filling in ti_seq. But if we are in persist
1007 * state, snd_max might reflect one byte beyond the
1008 * right edge of the window, so use snd_nxt in that
1009 * case, since we know we aren't doing a retransmission.
1010 * (retransmit and persist are mutually exclusive...)
1011 */
1012 if (sack_rxmit == 0) {
1013 if (len || (flags & (TH_SYN|TH_FIN)) ||
1014 tcp_timer_active(tp, TT_PERSIST))
1015 th->th_seq = htonl(tp->snd_nxt);
1016 else
1017 th->th_seq = htonl(tp->snd_max);
1018 } else {
1019 th->th_seq = htonl(p->rxmit);
1020 p->rxmit += len;
1021 tp->sackhint.sack_bytes_rexmit += len;
1022 }
1023 th->th_ack = htonl(tp->rcv_nxt);
1024 if (optlen) {
1025 bcopy(opt, th + 1, optlen);
1026 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1027 }
1028 th->th_flags = flags;
1029 /*
1030 * Calculate receive window. Don't shrink window,
1031 * but avoid silly window syndrome.
1032 */
1033 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1034 recwin < (long)tp->t_maxseg)
1035 recwin = 0;
1036 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1037 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1038 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1039 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1040 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1041
1042 /*
1043 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1044 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1045 * case is handled in syncache.
1046 */
1047 if (flags & TH_SYN)
1048 th->th_win = htons((u_short)
1049 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1050 else
1051 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1052
1053 /*
1054 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1055 * a 0 window. This may cause the remote transmitter to stall. This
1056 * flag tells soreceive() to disable delayed acknowledgements when
1057 * draining the buffer. This can occur if the receiver is attempting
1058 * to read more data than can be buffered prior to transmitting on
1059 * the connection.
1060 */
1061 if (th->th_win == 0) {
1062 tp->t_sndzerowin++;
1063 tp->t_flags |= TF_RXWIN0SENT;
1064 } else
1065 tp->t_flags &= ~TF_RXWIN0SENT;
1066 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1067 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1068 th->th_flags |= TH_URG;
1069 } else
1070 /*
1071 * If no urgent pointer to send, then we pull
1072 * the urgent pointer to the left edge of the send window
1073 * so that it doesn't drift into the send window on sequence
1074 * number wraparound.
1075 */
1076 tp->snd_up = tp->snd_una; /* drag it along */
1077
1078 #ifdef TCP_SIGNATURE
1079 if (to.to_flags & TOF_SIGNATURE) {
1080 int sigoff = to.to_signature - opt;
1081 tcp_signature_compute(m, 0, len, optlen,
1082 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1083 }
1084 #endif
1085
1086 /*
1087 * Put TCP length in extended header, and then
1088 * checksum extended header and data.
1089 */
1090 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1091 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1092 #ifdef INET6
1093 if (isipv6) {
1094 /*
1095 * ip6_plen is not need to be filled now, and will be filled
1096 * in ip6_output.
1097 */
1098 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1099 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1100 optlen + len, IPPROTO_TCP, 0);
1101 }
1102 #endif
1103 #if defined(INET6) && defined(INET)
1104 else
1105 #endif
1106 #ifdef INET
1107 {
1108 m->m_pkthdr.csum_flags = CSUM_TCP;
1109 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1110 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1111
1112 /* IP version must be set here for ipv4/ipv6 checking later */
1113 KASSERT(ip->ip_v == IPVERSION,
1114 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1115 }
1116 #endif
1117
1118 /*
1119 * Enable TSO and specify the size of the segments.
1120 * The TCP pseudo header checksum is always provided.
1121 * XXX: Fixme: This is currently not the case for IPv6.
1122 */
1123 if (tso) {
1124 KASSERT(len > tp->t_maxopd - optlen,
1125 ("%s: len <= tso_segsz", __func__));
1126 m->m_pkthdr.csum_flags |= CSUM_TSO;
1127 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1128 }
1129
1130 #ifdef IPSEC
1131 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1132 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1133 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1134 #else
1135 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1136 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1137 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1138 #endif
1139
1140 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1141 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1142
1143 #ifdef TCPDEBUG
1144 /*
1145 * Trace.
1146 */
1147 if (so->so_options & SO_DEBUG) {
1148 u_short save = 0;
1149 #ifdef INET6
1150 if (!isipv6)
1151 #endif
1152 {
1153 save = ipov->ih_len;
1154 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1155 }
1156 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1157 #ifdef INET6
1158 if (!isipv6)
1159 #endif
1160 ipov->ih_len = save;
1161 }
1162 #endif /* TCPDEBUG */
1163
1164 /*
1165 * Fill in IP length and desired time to live and
1166 * send to IP level. There should be a better way
1167 * to handle ttl and tos; we could keep them in
1168 * the template, but need a way to checksum without them.
1169 */
1170 /*
1171 * m->m_pkthdr.len should have been set before cksum calcuration,
1172 * because in6_cksum() need it.
1173 */
1174 #ifdef INET6
1175 if (isipv6) {
1176 struct route_in6 ro;
1177
1178 bzero(&ro, sizeof(ro));
1179 /*
1180 * we separately set hoplimit for every segment, since the
1181 * user might want to change the value via setsockopt.
1182 * Also, desired default hop limit might be changed via
1183 * Neighbor Discovery.
1184 */
1185 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1186
1187 /*
1188 * Set the packet size here for the benefit of DTrace probes.
1189 * ip6_output() will set it properly; it's supposed to include
1190 * the option header lengths as well.
1191 */
1192 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1193
1194 if (tp->t_state == TCPS_SYN_SENT)
1195 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1196
1197 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1198
1199 /* TODO: IPv6 IP6TOS_ECT bit on */
1200 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1201 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1202 NULL, NULL, tp->t_inpcb);
1203
1204 if (error == EMSGSIZE && ro.ro_rt != NULL)
1205 mtu = ro.ro_rt->rt_mtu;
1206 RO_RTFREE(&ro);
1207 }
1208 #endif /* INET6 */
1209 #if defined(INET) && defined(INET6)
1210 else
1211 #endif
1212 #ifdef INET
1213 {
1214 struct route ro;
1215
1216 bzero(&ro, sizeof(ro));
1217 ip->ip_len = htons(m->m_pkthdr.len);
1218 #ifdef INET6
1219 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1220 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1221 #endif /* INET6 */
1222 /*
1223 * If we do path MTU discovery, then we set DF on every packet.
1224 * This might not be the best thing to do according to RFC3390
1225 * Section 2. However the tcp hostcache migitates the problem
1226 * so it affects only the first tcp connection with a host.
1227 *
1228 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1229 */
1230 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1231 ip->ip_off |= htons(IP_DF);
1232
1233 if (tp->t_state == TCPS_SYN_SENT)
1234 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1235
1236 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1237
1238 error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1239 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1240 tp->t_inpcb);
1241
1242 if (error == EMSGSIZE && ro.ro_rt != NULL)
1243 mtu = ro.ro_rt->rt_mtu;
1244 RO_RTFREE(&ro);
1245 }
1246 #endif /* INET */
1247
1248 out:
1249 /*
1250 * In transmit state, time the transmission and arrange for
1251 * the retransmit. In persist state, just set snd_max.
1252 */
1253 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1254 !tcp_timer_active(tp, TT_PERSIST)) {
1255 tcp_seq startseq = tp->snd_nxt;
1256
1257 /*
1258 * Advance snd_nxt over sequence space of this segment.
1259 */
1260 if (flags & (TH_SYN|TH_FIN)) {
1261 if (flags & TH_SYN)
1262 tp->snd_nxt++;
1263 if (flags & TH_FIN) {
1264 tp->snd_nxt++;
1265 tp->t_flags |= TF_SENTFIN;
1266 }
1267 }
1268 if (sack_rxmit)
1269 goto timer;
1270 tp->snd_nxt += len;
1271 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1272 tp->snd_max = tp->snd_nxt;
1273 /*
1274 * Time this transmission if not a retransmission and
1275 * not currently timing anything.
1276 */
1277 if (tp->t_rtttime == 0) {
1278 tp->t_rtttime = ticks;
1279 tp->t_rtseq = startseq;
1280 TCPSTAT_INC(tcps_segstimed);
1281 }
1282 }
1283
1284 /*
1285 * Set retransmit timer if not currently set,
1286 * and not doing a pure ack or a keep-alive probe.
1287 * Initial value for retransmit timer is smoothed
1288 * round-trip time + 2 * round-trip time variance.
1289 * Initialize shift counter which is used for backoff
1290 * of retransmit time.
1291 */
1292 timer:
1293 if (!tcp_timer_active(tp, TT_REXMT) &&
1294 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1295 (tp->snd_nxt != tp->snd_una))) {
1296 if (tcp_timer_active(tp, TT_PERSIST)) {
1297 tcp_timer_activate(tp, TT_PERSIST, 0);
1298 tp->t_rxtshift = 0;
1299 }
1300 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1301 }
1302 } else {
1303 /*
1304 * Persist case, update snd_max but since we are in
1305 * persist mode (no window) we do not update snd_nxt.
1306 */
1307 int xlen = len;
1308 if (flags & TH_SYN)
1309 ++xlen;
1310 if (flags & TH_FIN) {
1311 ++xlen;
1312 tp->t_flags |= TF_SENTFIN;
1313 }
1314 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1315 tp->snd_max = tp->snd_nxt + len;
1316 }
1317
1318 if (error) {
1319
1320 /*
1321 * We know that the packet was lost, so back out the
1322 * sequence number advance, if any.
1323 *
1324 * If the error is EPERM the packet got blocked by the
1325 * local firewall. Normally we should terminate the
1326 * connection but the blocking may have been spurious
1327 * due to a firewall reconfiguration cycle. So we treat
1328 * it like a packet loss and let the retransmit timer and
1329 * timeouts do their work over time.
1330 * XXX: It is a POLA question whether calling tcp_drop right
1331 * away would be the really correct behavior instead.
1332 */
1333 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1334 !tcp_timer_active(tp, TT_PERSIST)) &&
1335 ((flags & TH_SYN) == 0) &&
1336 (error != EPERM)) {
1337 if (sack_rxmit) {
1338 p->rxmit -= len;
1339 tp->sackhint.sack_bytes_rexmit -= len;
1340 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1341 ("sackhint bytes rtx >= 0"));
1342 } else
1343 tp->snd_nxt -= len;
1344 }
1345 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1346 switch (error) {
1347 case EPERM:
1348 tp->t_softerror = error;
1349 return (error);
1350 case ENOBUFS:
1351 if (!tcp_timer_active(tp, TT_REXMT) &&
1352 !tcp_timer_active(tp, TT_PERSIST))
1353 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1354 tp->snd_cwnd = tp->t_maxseg;
1355 return (0);
1356 case EMSGSIZE:
1357 /*
1358 * For some reason the interface we used initially
1359 * to send segments changed to another or lowered
1360 * its MTU.
1361 * If TSO was active we either got an interface
1362 * without TSO capabilits or TSO was turned off.
1363 * If we obtained mtu from ip_output() then update
1364 * it and try again.
1365 */
1366 if (tso)
1367 tp->t_flags &= ~TF_TSO;
1368 if (mtu != 0) {
1369 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1370 goto again;
1371 }
1372 return (error);
1373 case EHOSTDOWN:
1374 case EHOSTUNREACH:
1375 case ENETDOWN:
1376 case ENETUNREACH:
1377 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1378 tp->t_softerror = error;
1379 return (0);
1380 }
1381 /* FALLTHROUGH */
1382 default:
1383 return (error);
1384 }
1385 }
1386 TCPSTAT_INC(tcps_sndtotal);
1387
1388 /*
1389 * Data sent (as far as we can tell).
1390 * If this advertises a larger window than any other segment,
1391 * then remember the size of the advertised window.
1392 * Any pending ACK has now been sent.
1393 */
1394 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1395 tp->rcv_adv = tp->rcv_nxt + recwin;
1396 tp->last_ack_sent = tp->rcv_nxt;
1397 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1398 if (tcp_timer_active(tp, TT_DELACK))
1399 tcp_timer_activate(tp, TT_DELACK, 0);
1400 #if 0
1401 /*
1402 * This completely breaks TCP if newreno is turned on. What happens
1403 * is that if delayed-acks are turned on on the receiver, this code
1404 * on the transmitter effectively destroys the TCP window, forcing
1405 * it to four packets (1.5Kx4 = 6K window).
1406 */
1407 if (sendalot && --maxburst)
1408 goto again;
1409 #endif
1410 if (sendalot)
1411 goto again;
1412 return (0);
1413 }
1414
1415 void
1416 tcp_setpersist(struct tcpcb *tp)
1417 {
1418 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1419 int tt;
1420
1421 tp->t_flags &= ~TF_PREVVALID;
1422 if (tcp_timer_active(tp, TT_REXMT))
1423 panic("tcp_setpersist: retransmit pending");
1424 /*
1425 * Start/restart persistance timer.
1426 */
1427 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1428 TCPTV_PERSMIN, TCPTV_PERSMAX);
1429 tcp_timer_activate(tp, TT_PERSIST, tt);
1430 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1431 tp->t_rxtshift++;
1432 }
1433
1434 /*
1435 * Insert TCP options according to the supplied parameters to the place
1436 * optp in a consistent way. Can handle unaligned destinations.
1437 *
1438 * The order of the option processing is crucial for optimal packing and
1439 * alignment for the scarce option space.
1440 *
1441 * The optimal order for a SYN/SYN-ACK segment is:
1442 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1443 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1444 *
1445 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1446 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1447 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1448 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1449 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1450 */
1451 int
1452 tcp_addoptions(struct tcpopt *to, u_char *optp)
1453 {
1454 u_int mask, optlen = 0;
1455
1456 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1457 if ((to->to_flags & mask) != mask)
1458 continue;
1459 if (optlen == TCP_MAXOLEN)
1460 break;
1461 switch (to->to_flags & mask) {
1462 case TOF_MSS:
1463 while (optlen % 4) {
1464 optlen += TCPOLEN_NOP;
1465 *optp++ = TCPOPT_NOP;
1466 }
1467 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1468 continue;
1469 optlen += TCPOLEN_MAXSEG;
1470 *optp++ = TCPOPT_MAXSEG;
1471 *optp++ = TCPOLEN_MAXSEG;
1472 to->to_mss = htons(to->to_mss);
1473 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1474 optp += sizeof(to->to_mss);
1475 break;
1476 case TOF_SCALE:
1477 while (!optlen || optlen % 2 != 1) {
1478 optlen += TCPOLEN_NOP;
1479 *optp++ = TCPOPT_NOP;
1480 }
1481 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1482 continue;
1483 optlen += TCPOLEN_WINDOW;
1484 *optp++ = TCPOPT_WINDOW;
1485 *optp++ = TCPOLEN_WINDOW;
1486 *optp++ = to->to_wscale;
1487 break;
1488 case TOF_SACKPERM:
1489 while (optlen % 2) {
1490 optlen += TCPOLEN_NOP;
1491 *optp++ = TCPOPT_NOP;
1492 }
1493 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1494 continue;
1495 optlen += TCPOLEN_SACK_PERMITTED;
1496 *optp++ = TCPOPT_SACK_PERMITTED;
1497 *optp++ = TCPOLEN_SACK_PERMITTED;
1498 break;
1499 case TOF_TS:
1500 while (!optlen || optlen % 4 != 2) {
1501 optlen += TCPOLEN_NOP;
1502 *optp++ = TCPOPT_NOP;
1503 }
1504 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1505 continue;
1506 optlen += TCPOLEN_TIMESTAMP;
1507 *optp++ = TCPOPT_TIMESTAMP;
1508 *optp++ = TCPOLEN_TIMESTAMP;
1509 to->to_tsval = htonl(to->to_tsval);
1510 to->to_tsecr = htonl(to->to_tsecr);
1511 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1512 optp += sizeof(to->to_tsval);
1513 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1514 optp += sizeof(to->to_tsecr);
1515 break;
1516 #ifdef TCP_SIGNATURE
1517 case TOF_SIGNATURE:
1518 {
1519 int siglen = TCPOLEN_SIGNATURE - 2;
1520
1521 while (!optlen || optlen % 4 != 2) {
1522 optlen += TCPOLEN_NOP;
1523 *optp++ = TCPOPT_NOP;
1524 }
1525 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1526 continue;
1527 optlen += TCPOLEN_SIGNATURE;
1528 *optp++ = TCPOPT_SIGNATURE;
1529 *optp++ = TCPOLEN_SIGNATURE;
1530 to->to_signature = optp;
1531 while (siglen--)
1532 *optp++ = 0;
1533 break;
1534 }
1535 #endif
1536 case TOF_SACK:
1537 {
1538 int sackblks = 0;
1539 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1540 tcp_seq sack_seq;
1541
1542 while (!optlen || optlen % 4 != 2) {
1543 optlen += TCPOLEN_NOP;
1544 *optp++ = TCPOPT_NOP;
1545 }
1546 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1547 continue;
1548 optlen += TCPOLEN_SACKHDR;
1549 *optp++ = TCPOPT_SACK;
1550 sackblks = min(to->to_nsacks,
1551 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1552 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1553 while (sackblks--) {
1554 sack_seq = htonl(sack->start);
1555 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1556 optp += sizeof(sack_seq);
1557 sack_seq = htonl(sack->end);
1558 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1559 optp += sizeof(sack_seq);
1560 optlen += TCPOLEN_SACK;
1561 sack++;
1562 }
1563 TCPSTAT_INC(tcps_sack_send_blocks);
1564 break;
1565 }
1566 default:
1567 panic("%s: unknown TCP option type", __func__);
1568 break;
1569 }
1570 }
1571
1572 /* Terminate and pad TCP options to a 4 byte boundary. */
1573 if (optlen % 4) {
1574 optlen += TCPOLEN_EOL;
1575 *optp++ = TCPOPT_EOL;
1576 }
1577 /*
1578 * According to RFC 793 (STD0007):
1579 * "The content of the header beyond the End-of-Option option
1580 * must be header padding (i.e., zero)."
1581 * and later: "The padding is composed of zeros."
1582 */
1583 while (optlen % 4) {
1584 optlen += TCPOLEN_PAD;
1585 *optp++ = TCPOPT_PAD;
1586 }
1587
1588 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1589 return (optlen);
1590 }
Cache object: 901665d7c280b7879c95af605d9a088d
|