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/11.0/sys/netinet/tcp_output.c 300981 2016-05-30 03:31:37Z sephe $");
34
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/hhook.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/sdt.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53
54 #include <net/if.h>
55 #include <net/route.h>
56 #include <net/vnet.h>
57
58 #include <netinet/in.h>
59 #include <netinet/in_kdtrace.h>
60 #include <netinet/in_systm.h>
61 #include <netinet/ip.h>
62 #include <netinet/in_pcb.h>
63 #include <netinet/ip_var.h>
64 #include <netinet/ip_options.h>
65 #ifdef INET6
66 #include <netinet6/in6_pcb.h>
67 #include <netinet/ip6.h>
68 #include <netinet6/ip6_var.h>
69 #endif
70 #ifdef TCP_RFC7413
71 #include <netinet/tcp_fastopen.h>
72 #endif
73 #include <netinet/tcp.h>
74 #define TCPOUTFLAGS
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/tcp_seq.h>
77 #include <netinet/tcp_timer.h>
78 #include <netinet/tcp_var.h>
79 #include <netinet/tcpip.h>
80 #include <netinet/cc/cc.h>
81 #ifdef TCPPCAP
82 #include <netinet/tcp_pcap.h>
83 #endif
84 #ifdef TCPDEBUG
85 #include <netinet/tcp_debug.h>
86 #endif
87 #ifdef TCP_OFFLOAD
88 #include <netinet/tcp_offload.h>
89 #endif
90
91 #ifdef IPSEC
92 #include <netipsec/ipsec.h>
93 #endif /*IPSEC*/
94
95 #include <machine/in_cksum.h>
96
97 #include <security/mac/mac_framework.h>
98
99 VNET_DEFINE(int, path_mtu_discovery) = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
101 &VNET_NAME(path_mtu_discovery), 1,
102 "Enable Path MTU Discovery");
103
104 VNET_DEFINE(int, tcp_do_tso) = 1;
105 #define V_tcp_do_tso VNET(tcp_do_tso)
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
107 &VNET_NAME(tcp_do_tso), 0,
108 "Enable TCP Segmentation Offload");
109
110 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
111 #define V_tcp_sendspace VNET(tcp_sendspace)
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
113 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114
115 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
116 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
118 &VNET_NAME(tcp_do_autosndbuf), 0,
119 "Enable automatic send buffer sizing");
120
121 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
122 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
124 &VNET_NAME(tcp_autosndbuf_inc), 0,
125 "Incrementor step size of automatic send buffer");
126
127 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
128 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
130 &VNET_NAME(tcp_autosndbuf_max), 0,
131 "Max size of automatic send buffer");
132
133 /*
134 * Make sure that either retransmit or persist timer is set for SYN, FIN and
135 * non-ACK.
136 */
137 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
138 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
139 tcp_timer_active((tp), TT_REXMT) || \
140 tcp_timer_active((tp), TT_PERSIST), \
141 ("neither rexmt nor persist timer is set"))
142
143 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
144 struct tcphdr *th, struct tcpopt *to,
145 long len, int tso);
146 static void inline cc_after_idle(struct tcpcb *tp);
147
148 /*
149 * Wrapper for the TCP established output helper hook.
150 */
151 static void inline
152 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
153 struct tcpopt *to, long len, int tso)
154 {
155 struct tcp_hhook_data hhook_data;
156
157 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
158 hhook_data.tp = tp;
159 hhook_data.th = th;
160 hhook_data.to = to;
161 hhook_data.len = len;
162 hhook_data.tso = tso;
163
164 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
165 tp->osd);
166 }
167 }
168
169 /*
170 * CC wrapper hook functions
171 */
172 static void inline
173 cc_after_idle(struct tcpcb *tp)
174 {
175 INP_WLOCK_ASSERT(tp->t_inpcb);
176
177 if (CC_ALGO(tp)->after_idle != NULL)
178 CC_ALGO(tp)->after_idle(tp->ccv);
179 }
180
181 /*
182 * Tcp output routine: figure out what should be sent and send it.
183 */
184 int
185 tcp_output(struct tcpcb *tp)
186 {
187 struct socket *so = tp->t_inpcb->inp_socket;
188 long len, recwin, sendwin;
189 int off, flags, error = 0; /* Keep compiler happy */
190 struct mbuf *m;
191 struct ip *ip = NULL;
192 struct ipovly *ipov = NULL;
193 struct tcphdr *th;
194 u_char opt[TCP_MAXOLEN];
195 unsigned ipoptlen, optlen, hdrlen;
196 #ifdef IPSEC
197 unsigned ipsec_optlen = 0;
198 #endif
199 int idle, sendalot;
200 int sack_rxmit, sack_bytes_rxmt;
201 struct sackhole *p;
202 int tso, mtu;
203 struct tcpopt to;
204 #if 0
205 int maxburst = TCP_MAXBURST;
206 #endif
207 #ifdef INET6
208 struct ip6_hdr *ip6 = NULL;
209 int isipv6;
210
211 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
212 #endif
213
214 INP_WLOCK_ASSERT(tp->t_inpcb);
215
216 #ifdef TCP_OFFLOAD
217 if (tp->t_flags & TF_TOE)
218 return (tcp_offload_output(tp));
219 #endif
220
221 #ifdef TCP_RFC7413
222 /*
223 * For TFO connections in SYN_RECEIVED, only allow the initial
224 * SYN|ACK and those sent by the retransmit timer.
225 */
226 if ((tp->t_flags & TF_FASTOPEN) &&
227 (tp->t_state == TCPS_SYN_RECEIVED) &&
228 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
229 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
230 return (0);
231 #endif
232 /*
233 * Determine length of data that should be transmitted,
234 * and flags that will be used.
235 * If there is some data or critical controls (SYN, RST)
236 * to send, then transmit; otherwise, investigate further.
237 */
238 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
239 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
240 cc_after_idle(tp);
241 tp->t_flags &= ~TF_LASTIDLE;
242 if (idle) {
243 if (tp->t_flags & TF_MORETOCOME) {
244 tp->t_flags |= TF_LASTIDLE;
245 idle = 0;
246 }
247 }
248 again:
249 /*
250 * If we've recently taken a timeout, snd_max will be greater than
251 * snd_nxt. There may be SACK information that allows us to avoid
252 * resending already delivered data. Adjust snd_nxt accordingly.
253 */
254 if ((tp->t_flags & TF_SACK_PERMIT) &&
255 SEQ_LT(tp->snd_nxt, tp->snd_max))
256 tcp_sack_adjust(tp);
257 sendalot = 0;
258 tso = 0;
259 mtu = 0;
260 off = tp->snd_nxt - tp->snd_una;
261 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
262
263 flags = tcp_outflags[tp->t_state];
264 /*
265 * Send any SACK-generated retransmissions. If we're explicitly trying
266 * to send out new data (when sendalot is 1), bypass this function.
267 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
268 * we're replacing a (future) new transmission with a retransmission
269 * now, and we previously incremented snd_cwnd in tcp_input().
270 */
271 /*
272 * Still in sack recovery , reset rxmit flag to zero.
273 */
274 sack_rxmit = 0;
275 sack_bytes_rxmt = 0;
276 len = 0;
277 p = NULL;
278 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
279 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
280 long cwin;
281
282 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
283 if (cwin < 0)
284 cwin = 0;
285 /* Do not retransmit SACK segments beyond snd_recover */
286 if (SEQ_GT(p->end, tp->snd_recover)) {
287 /*
288 * (At least) part of sack hole extends beyond
289 * snd_recover. Check to see if we can rexmit data
290 * for this hole.
291 */
292 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
293 /*
294 * Can't rexmit any more data for this hole.
295 * That data will be rexmitted in the next
296 * sack recovery episode, when snd_recover
297 * moves past p->rxmit.
298 */
299 p = NULL;
300 goto after_sack_rexmit;
301 } else
302 /* Can rexmit part of the current hole */
303 len = ((long)ulmin(cwin,
304 tp->snd_recover - p->rxmit));
305 } else
306 len = ((long)ulmin(cwin, p->end - p->rxmit));
307 off = p->rxmit - tp->snd_una;
308 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
309 __func__, off));
310 if (len > 0) {
311 sack_rxmit = 1;
312 sendalot = 1;
313 TCPSTAT_INC(tcps_sack_rexmits);
314 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
315 min(len, tp->t_maxseg));
316 }
317 }
318 after_sack_rexmit:
319 /*
320 * Get standard flags, and add SYN or FIN if requested by 'hidden'
321 * state flags.
322 */
323 if (tp->t_flags & TF_NEEDFIN)
324 flags |= TH_FIN;
325 if (tp->t_flags & TF_NEEDSYN)
326 flags |= TH_SYN;
327
328 SOCKBUF_LOCK(&so->so_snd);
329 /*
330 * If in persist timeout with window of 0, send 1 byte.
331 * Otherwise, if window is small but nonzero
332 * and timer expired, we will send what we can
333 * and go to transmit state.
334 */
335 if (tp->t_flags & TF_FORCEDATA) {
336 if (sendwin == 0) {
337 /*
338 * If we still have some data to send, then
339 * clear the FIN bit. Usually this would
340 * happen below when it realizes that we
341 * aren't sending all the data. However,
342 * if we have exactly 1 byte of unsent data,
343 * then it won't clear the FIN bit below,
344 * and if we are in persist state, we wind
345 * up sending the packet without recording
346 * that we sent the FIN bit.
347 *
348 * We can't just blindly clear the FIN bit,
349 * because if we don't have any more data
350 * to send then the probe will be the FIN
351 * itself.
352 */
353 if (off < sbused(&so->so_snd))
354 flags &= ~TH_FIN;
355 sendwin = 1;
356 } else {
357 tcp_timer_activate(tp, TT_PERSIST, 0);
358 tp->t_rxtshift = 0;
359 }
360 }
361
362 /*
363 * If snd_nxt == snd_max and we have transmitted a FIN, the
364 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
365 * a negative length. This can also occur when TCP opens up
366 * its congestion window while receiving additional duplicate
367 * acks after fast-retransmit because TCP will reset snd_nxt
368 * to snd_max after the fast-retransmit.
369 *
370 * In the normal retransmit-FIN-only case, however, snd_nxt will
371 * be set to snd_una, the offset will be 0, and the length may
372 * wind up 0.
373 *
374 * If sack_rxmit is true we are retransmitting from the scoreboard
375 * in which case len is already set.
376 */
377 if (sack_rxmit == 0) {
378 if (sack_bytes_rxmt == 0)
379 len = ((long)ulmin(sbavail(&so->so_snd), sendwin) -
380 off);
381 else {
382 long cwin;
383
384 /*
385 * We are inside of a SACK recovery episode and are
386 * sending new data, having retransmitted all the
387 * data possible in the scoreboard.
388 */
389 len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) -
390 off);
391 /*
392 * Don't remove this (len > 0) check !
393 * We explicitly check for len > 0 here (although it
394 * isn't really necessary), to work around a gcc
395 * optimization issue - to force gcc to compute
396 * len above. Without this check, the computation
397 * of len is bungled by the optimizer.
398 */
399 if (len > 0) {
400 cwin = tp->snd_cwnd -
401 (tp->snd_nxt - tp->sack_newdata) -
402 sack_bytes_rxmt;
403 if (cwin < 0)
404 cwin = 0;
405 len = lmin(len, cwin);
406 }
407 }
408 }
409
410 /*
411 * Lop off SYN bit if it has already been sent. However, if this
412 * is SYN-SENT state and if segment contains data and if we don't
413 * know that foreign host supports TAO, suppress sending segment.
414 */
415 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
416 if (tp->t_state != TCPS_SYN_RECEIVED)
417 flags &= ~TH_SYN;
418 #ifdef TCP_RFC7413
419 /*
420 * When sending additional segments following a TFO SYN|ACK,
421 * do not include the SYN bit.
422 */
423 if ((tp->t_flags & TF_FASTOPEN) &&
424 (tp->t_state == TCPS_SYN_RECEIVED))
425 flags &= ~TH_SYN;
426 #endif
427 off--, len++;
428 }
429
430 /*
431 * Be careful not to send data and/or FIN on SYN segments.
432 * This measure is needed to prevent interoperability problems
433 * with not fully conformant TCP implementations.
434 */
435 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
436 len = 0;
437 flags &= ~TH_FIN;
438 }
439
440 #ifdef TCP_RFC7413
441 /*
442 * When retransmitting SYN|ACK on a passively-created TFO socket,
443 * don't include data, as the presence of data may have caused the
444 * original SYN|ACK to have been dropped by a middlebox.
445 */
446 if ((tp->t_flags & TF_FASTOPEN) &&
447 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
448 (flags & TH_RST)))
449 len = 0;
450 #endif
451 if (len <= 0) {
452 /*
453 * If FIN has been sent but not acked,
454 * but we haven't been called to retransmit,
455 * len will be < 0. Otherwise, window shrank
456 * after we sent into it. If window shrank to 0,
457 * cancel pending retransmit, pull snd_nxt back
458 * to (closed) window, and set the persist timer
459 * if it isn't already going. If the window didn't
460 * close completely, just wait for an ACK.
461 *
462 * We also do a general check here to ensure that
463 * we will set the persist timer when we have data
464 * to send, but a 0-byte window. This makes sure
465 * the persist timer is set even if the packet
466 * hits one of the "goto send" lines below.
467 */
468 len = 0;
469 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
470 (off < (int) sbavail(&so->so_snd))) {
471 tcp_timer_activate(tp, TT_REXMT, 0);
472 tp->t_rxtshift = 0;
473 tp->snd_nxt = tp->snd_una;
474 if (!tcp_timer_active(tp, TT_PERSIST))
475 tcp_setpersist(tp);
476 }
477 }
478
479 /* len will be >= 0 after this point. */
480 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
481
482 /*
483 * Automatic sizing of send socket buffer. Often the send buffer
484 * size is not optimally adjusted to the actual network conditions
485 * at hand (delay bandwidth product). Setting the buffer size too
486 * small limits throughput on links with high bandwidth and high
487 * delay (eg. trans-continental/oceanic links). Setting the
488 * buffer size too big consumes too much real kernel memory,
489 * especially with many connections on busy servers.
490 *
491 * The criteria to step up the send buffer one notch are:
492 * 1. receive window of remote host is larger than send buffer
493 * (with a fudge factor of 5/4th);
494 * 2. send buffer is filled to 7/8th with data (so we actually
495 * have data to make use of it);
496 * 3. send buffer fill has not hit maximal automatic size;
497 * 4. our send window (slow start and cogestion controlled) is
498 * larger than sent but unacknowledged data in send buffer.
499 *
500 * The remote host receive window scaling factor may limit the
501 * growing of the send buffer before it reaches its allowed
502 * maximum.
503 *
504 * It scales directly with slow start or congestion window
505 * and does at most one step per received ACK. This fast
506 * scaling has the drawback of growing the send buffer beyond
507 * what is strictly necessary to make full use of a given
508 * delay*bandwidth product. However testing has shown this not
509 * to be much of an problem. At worst we are trading wasting
510 * of available bandwidth (the non-use of it) for wasting some
511 * socket buffer memory.
512 *
513 * TODO: Shrink send buffer during idle periods together
514 * with congestion window. Requires another timer. Has to
515 * wait for upcoming tcp timer rewrite.
516 *
517 * XXXGL: should there be used sbused() or sbavail()?
518 */
519 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
520 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
521 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) &&
522 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
523 sendwin >= (sbused(&so->so_snd) -
524 (tp->snd_nxt - tp->snd_una))) {
525 if (!sbreserve_locked(&so->so_snd,
526 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
527 V_tcp_autosndbuf_max), so, curthread))
528 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
529 }
530 }
531
532 /*
533 * Decide if we can use TCP Segmentation Offloading (if supported by
534 * hardware).
535 *
536 * TSO may only be used if we are in a pure bulk sending state. The
537 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
538 * IP options prevent using TSO. With TSO the TCP header is the same
539 * (except for the sequence number) for all generated packets. This
540 * makes it impossible to transmit any options which vary per generated
541 * segment or packet.
542 */
543 #ifdef IPSEC
544 /*
545 * Pre-calculate here as we save another lookup into the darknesses
546 * of IPsec that way and can actually decide if TSO is ok.
547 */
548 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
549 #endif
550 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
551 ((tp->t_flags & TF_SIGNATURE) == 0) &&
552 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
553 #ifdef IPSEC
554 ipsec_optlen == 0 &&
555 #endif
556 tp->t_inpcb->inp_options == NULL &&
557 tp->t_inpcb->in6p_options == NULL)
558 tso = 1;
559
560 if (sack_rxmit) {
561 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
562 flags &= ~TH_FIN;
563 } else {
564 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
565 sbused(&so->so_snd)))
566 flags &= ~TH_FIN;
567 }
568
569 recwin = sbspace(&so->so_rcv);
570
571 /*
572 * Sender silly window avoidance. We transmit under the following
573 * conditions when len is non-zero:
574 *
575 * - We have a full segment (or more with TSO)
576 * - This is the last buffer in a write()/send() and we are
577 * either idle or running NODELAY
578 * - we've timed out (e.g. persist timer)
579 * - we have more then 1/2 the maximum send window's worth of
580 * data (receiver may be limited the window size)
581 * - we need to retransmit
582 */
583 if (len) {
584 if (len >= tp->t_maxseg)
585 goto send;
586 /*
587 * NOTE! on localhost connections an 'ack' from the remote
588 * end may occur synchronously with the output and cause
589 * us to flush a buffer queued with moretocome. XXX
590 *
591 * note: the len + off check is almost certainly unnecessary.
592 */
593 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
594 (idle || (tp->t_flags & TF_NODELAY)) &&
595 len + off >= sbavail(&so->so_snd) &&
596 (tp->t_flags & TF_NOPUSH) == 0) {
597 goto send;
598 }
599 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
600 goto send;
601 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
602 goto send;
603 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
604 goto send;
605 if (sack_rxmit)
606 goto send;
607 }
608
609 /*
610 * Sending of standalone window updates.
611 *
612 * Window updates are important when we close our window due to a
613 * full socket buffer and are opening it again after the application
614 * reads data from it. Once the window has opened again and the
615 * remote end starts to send again the ACK clock takes over and
616 * provides the most current window information.
617 *
618 * We must avoid the silly window syndrome whereas every read
619 * from the receive buffer, no matter how small, causes a window
620 * update to be sent. We also should avoid sending a flurry of
621 * window updates when the socket buffer had queued a lot of data
622 * and the application is doing small reads.
623 *
624 * Prevent a flurry of pointless window updates by only sending
625 * an update when we can increase the advertized window by more
626 * than 1/4th of the socket buffer capacity. When the buffer is
627 * getting full or is very small be more aggressive and send an
628 * update whenever we can increase by two mss sized segments.
629 * In all other situations the ACK's to new incoming data will
630 * carry further window increases.
631 *
632 * Don't send an independent window update if a delayed
633 * ACK is pending (it will get piggy-backed on it) or the
634 * remote side already has done a half-close and won't send
635 * more data. Skip this if the connection is in T/TCP
636 * half-open state.
637 */
638 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
639 !(tp->t_flags & TF_DELACK) &&
640 !TCPS_HAVERCVDFIN(tp->t_state)) {
641 /*
642 * "adv" is the amount we could increase the window,
643 * taking into account that we are limited by
644 * TCP_MAXWIN << tp->rcv_scale.
645 */
646 long adv;
647 int oldwin;
648
649 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
650 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
651 oldwin = (tp->rcv_adv - tp->rcv_nxt);
652 adv -= oldwin;
653 } else
654 oldwin = 0;
655
656 /*
657 * If the new window size ends up being the same as the old
658 * size when it is scaled, then don't force a window update.
659 */
660 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
661 goto dontupdate;
662
663 if (adv >= (long)(2 * tp->t_maxseg) &&
664 (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
665 recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
666 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
667 goto send;
668 }
669 dontupdate:
670
671 /*
672 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
673 * is also a catch-all for the retransmit timer timeout case.
674 */
675 if (tp->t_flags & TF_ACKNOW)
676 goto send;
677 if ((flags & TH_RST) ||
678 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
679 goto send;
680 if (SEQ_GT(tp->snd_up, tp->snd_una))
681 goto send;
682 /*
683 * If our state indicates that FIN should be sent
684 * and we have not yet done so, then we need to send.
685 */
686 if (flags & TH_FIN &&
687 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
688 goto send;
689 /*
690 * In SACK, it is possible for tcp_output to fail to send a segment
691 * after the retransmission timer has been turned off. Make sure
692 * that the retransmission timer is set.
693 */
694 if ((tp->t_flags & TF_SACK_PERMIT) &&
695 SEQ_GT(tp->snd_max, tp->snd_una) &&
696 !tcp_timer_active(tp, TT_REXMT) &&
697 !tcp_timer_active(tp, TT_PERSIST)) {
698 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
699 goto just_return;
700 }
701 /*
702 * TCP window updates are not reliable, rather a polling protocol
703 * using ``persist'' packets is used to insure receipt of window
704 * updates. The three ``states'' for the output side are:
705 * idle not doing retransmits or persists
706 * persisting to move a small or zero window
707 * (re)transmitting and thereby not persisting
708 *
709 * tcp_timer_active(tp, TT_PERSIST)
710 * is true when we are in persist state.
711 * (tp->t_flags & TF_FORCEDATA)
712 * is set when we are called to send a persist packet.
713 * tcp_timer_active(tp, TT_REXMT)
714 * is set when we are retransmitting
715 * The output side is idle when both timers are zero.
716 *
717 * If send window is too small, there is data to transmit, and no
718 * retransmit or persist is pending, then go to persist state.
719 * If nothing happens soon, send when timer expires:
720 * if window is nonzero, transmit what we can,
721 * otherwise force out a byte.
722 */
723 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
724 !tcp_timer_active(tp, TT_PERSIST)) {
725 tp->t_rxtshift = 0;
726 tcp_setpersist(tp);
727 }
728
729 /*
730 * No reason to send a segment, just return.
731 */
732 just_return:
733 SOCKBUF_UNLOCK(&so->so_snd);
734 return (0);
735
736 send:
737 SOCKBUF_LOCK_ASSERT(&so->so_snd);
738 if (len > 0) {
739 if (len >= tp->t_maxseg)
740 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
741 else
742 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
743 }
744 /*
745 * Before ESTABLISHED, force sending of initial options
746 * unless TCP set not to do any options.
747 * NOTE: we assume that the IP/TCP header plus TCP options
748 * always fit in a single mbuf, leaving room for a maximum
749 * link header, i.e.
750 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
751 */
752 optlen = 0;
753 #ifdef INET6
754 if (isipv6)
755 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
756 else
757 #endif
758 hdrlen = sizeof (struct tcpiphdr);
759
760 /*
761 * Compute options for segment.
762 * We only have to care about SYN and established connection
763 * segments. Options for SYN-ACK segments are handled in TCP
764 * syncache.
765 */
766 to.to_flags = 0;
767 if ((tp->t_flags & TF_NOOPT) == 0) {
768 /* Maximum segment size. */
769 if (flags & TH_SYN) {
770 tp->snd_nxt = tp->iss;
771 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
772 to.to_flags |= TOF_MSS;
773 #ifdef TCP_RFC7413
774 /*
775 * Only include the TFO option on the first
776 * transmission of the SYN|ACK on a
777 * passively-created TFO socket, as the presence of
778 * the TFO option may have caused the original
779 * SYN|ACK to have been dropped by a middlebox.
780 */
781 if ((tp->t_flags & TF_FASTOPEN) &&
782 (tp->t_state == TCPS_SYN_RECEIVED) &&
783 (tp->t_rxtshift == 0)) {
784 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
785 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
786 to.to_flags |= TOF_FASTOPEN;
787 }
788 #endif
789 }
790 /* Window scaling. */
791 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
792 to.to_wscale = tp->request_r_scale;
793 to.to_flags |= TOF_SCALE;
794 }
795 /* Timestamps. */
796 if ((tp->t_flags & TF_RCVD_TSTMP) ||
797 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
798 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
799 to.to_tsecr = tp->ts_recent;
800 to.to_flags |= TOF_TS;
801 /* Set receive buffer autosizing timestamp. */
802 if (tp->rfbuf_ts == 0 &&
803 (so->so_rcv.sb_flags & SB_AUTOSIZE))
804 tp->rfbuf_ts = tcp_ts_getticks();
805 }
806 /* Selective ACK's. */
807 if (tp->t_flags & TF_SACK_PERMIT) {
808 if (flags & TH_SYN)
809 to.to_flags |= TOF_SACKPERM;
810 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
811 (tp->t_flags & TF_SACK_PERMIT) &&
812 tp->rcv_numsacks > 0) {
813 to.to_flags |= TOF_SACK;
814 to.to_nsacks = tp->rcv_numsacks;
815 to.to_sacks = (u_char *)tp->sackblks;
816 }
817 }
818 #ifdef TCP_SIGNATURE
819 /* TCP-MD5 (RFC2385). */
820 if (tp->t_flags & TF_SIGNATURE)
821 to.to_flags |= TOF_SIGNATURE;
822 #endif /* TCP_SIGNATURE */
823
824 /* Processing the options. */
825 hdrlen += optlen = tcp_addoptions(&to, opt);
826 }
827
828 #ifdef INET6
829 if (isipv6)
830 ipoptlen = ip6_optlen(tp->t_inpcb);
831 else
832 #endif
833 if (tp->t_inpcb->inp_options)
834 ipoptlen = tp->t_inpcb->inp_options->m_len -
835 offsetof(struct ipoption, ipopt_list);
836 else
837 ipoptlen = 0;
838 #ifdef IPSEC
839 ipoptlen += ipsec_optlen;
840 #endif
841
842 /*
843 * Adjust data length if insertion of options will
844 * bump the packet length beyond the t_maxseg length.
845 * Clear the FIN bit because we cut off the tail of
846 * the segment.
847 */
848 if (len + optlen + ipoptlen > tp->t_maxseg) {
849 flags &= ~TH_FIN;
850
851 if (tso) {
852 u_int if_hw_tsomax;
853 u_int if_hw_tsomaxsegcount;
854 u_int if_hw_tsomaxsegsize;
855 struct mbuf *mb;
856 u_int moff;
857 int max_len;
858
859 /* extract TSO information */
860 if_hw_tsomax = tp->t_tsomax;
861 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
862 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
863
864 /*
865 * Limit a TSO burst to prevent it from
866 * overflowing or exceeding the maximum length
867 * allowed by the network interface:
868 */
869 KASSERT(ipoptlen == 0,
870 ("%s: TSO can't do IP options", __func__));
871
872 /*
873 * Check if we should limit by maximum payload
874 * length:
875 */
876 if (if_hw_tsomax != 0) {
877 /* compute maximum TSO length */
878 max_len = (if_hw_tsomax - hdrlen -
879 max_linkhdr);
880 if (max_len <= 0) {
881 len = 0;
882 } else if (len > max_len) {
883 sendalot = 1;
884 len = max_len;
885 }
886 }
887
888 /*
889 * Check if we should limit by maximum segment
890 * size and count:
891 */
892 if (if_hw_tsomaxsegcount != 0 &&
893 if_hw_tsomaxsegsize != 0) {
894 /*
895 * Subtract one segment for the LINK
896 * and TCP/IP headers mbuf that will
897 * be prepended to this mbuf chain
898 * after the code in this section
899 * limits the number of mbufs in the
900 * chain to if_hw_tsomaxsegcount.
901 */
902 if_hw_tsomaxsegcount -= 1;
903 max_len = 0;
904 mb = sbsndmbuf(&so->so_snd, off, &moff);
905
906 while (mb != NULL && max_len < len) {
907 u_int mlen;
908 u_int frags;
909
910 /*
911 * Get length of mbuf fragment
912 * and how many hardware frags,
913 * rounded up, it would use:
914 */
915 mlen = (mb->m_len - moff);
916 frags = howmany(mlen,
917 if_hw_tsomaxsegsize);
918
919 /* Handle special case: Zero Length Mbuf */
920 if (frags == 0)
921 frags = 1;
922
923 /*
924 * Check if the fragment limit
925 * will be reached or exceeded:
926 */
927 if (frags >= if_hw_tsomaxsegcount) {
928 max_len += min(mlen,
929 if_hw_tsomaxsegcount *
930 if_hw_tsomaxsegsize);
931 break;
932 }
933 max_len += mlen;
934 if_hw_tsomaxsegcount -= frags;
935 moff = 0;
936 mb = mb->m_next;
937 }
938 if (max_len <= 0) {
939 len = 0;
940 } else if (len > max_len) {
941 sendalot = 1;
942 len = max_len;
943 }
944 }
945
946 /*
947 * Prevent the last segment from being
948 * fractional unless the send sockbuf can be
949 * emptied:
950 */
951 max_len = (tp->t_maxseg - optlen);
952 if ((off + len) < sbavail(&so->so_snd)) {
953 moff = len % max_len;
954 if (moff != 0) {
955 len -= moff;
956 sendalot = 1;
957 }
958 }
959
960 /*
961 * In case there are too many small fragments
962 * don't use TSO:
963 */
964 if (len <= max_len) {
965 len = max_len;
966 sendalot = 1;
967 tso = 0;
968 }
969
970 /*
971 * Send the FIN in a separate segment
972 * after the bulk sending is done.
973 * We don't trust the TSO implementations
974 * to clear the FIN flag on all but the
975 * last segment.
976 */
977 if (tp->t_flags & TF_NEEDFIN)
978 sendalot = 1;
979
980 } else {
981 len = tp->t_maxseg - optlen - ipoptlen;
982 sendalot = 1;
983 }
984 } else
985 tso = 0;
986
987 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
988 ("%s: len > IP_MAXPACKET", __func__));
989
990 /*#ifdef DIAGNOSTIC*/
991 #ifdef INET6
992 if (max_linkhdr + hdrlen > MCLBYTES)
993 #else
994 if (max_linkhdr + hdrlen > MHLEN)
995 #endif
996 panic("tcphdr too big");
997 /*#endif*/
998
999 /*
1000 * This KASSERT is here to catch edge cases at a well defined place.
1001 * Before, those had triggered (random) panic conditions further down.
1002 */
1003 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1004
1005 /*
1006 * Grab a header mbuf, attaching a copy of data to
1007 * be transmitted, and initialize the header from
1008 * the template for sends on this connection.
1009 */
1010 if (len) {
1011 struct mbuf *mb;
1012 u_int moff;
1013
1014 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1015 TCPSTAT_INC(tcps_sndprobe);
1016 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1017 tp->t_sndrexmitpack++;
1018 TCPSTAT_INC(tcps_sndrexmitpack);
1019 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1020 } else {
1021 TCPSTAT_INC(tcps_sndpack);
1022 TCPSTAT_ADD(tcps_sndbyte, len);
1023 }
1024 #ifdef INET6
1025 if (MHLEN < hdrlen + max_linkhdr)
1026 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1027 else
1028 #endif
1029 m = m_gethdr(M_NOWAIT, MT_DATA);
1030
1031 if (m == NULL) {
1032 SOCKBUF_UNLOCK(&so->so_snd);
1033 error = ENOBUFS;
1034 sack_rxmit = 0;
1035 goto out;
1036 }
1037
1038 m->m_data += max_linkhdr;
1039 m->m_len = hdrlen;
1040
1041 /*
1042 * Start the m_copy functions from the closest mbuf
1043 * to the offset in the socket buffer chain.
1044 */
1045 mb = sbsndptr(&so->so_snd, off, len, &moff);
1046
1047 if (len <= MHLEN - hdrlen - max_linkhdr) {
1048 m_copydata(mb, moff, (int)len,
1049 mtod(m, caddr_t) + hdrlen);
1050 m->m_len += len;
1051 } else {
1052 m->m_next = m_copy(mb, moff, (int)len);
1053 if (m->m_next == NULL) {
1054 SOCKBUF_UNLOCK(&so->so_snd);
1055 (void) m_free(m);
1056 error = ENOBUFS;
1057 sack_rxmit = 0;
1058 goto out;
1059 }
1060 }
1061
1062 /*
1063 * If we're sending everything we've got, set PUSH.
1064 * (This will keep happy those implementations which only
1065 * give data to the user when a buffer fills or
1066 * a PUSH comes in.)
1067 */
1068 if ((off + len == sbused(&so->so_snd)) && !(flags & TH_SYN))
1069 flags |= TH_PUSH;
1070 SOCKBUF_UNLOCK(&so->so_snd);
1071 } else {
1072 SOCKBUF_UNLOCK(&so->so_snd);
1073 if (tp->t_flags & TF_ACKNOW)
1074 TCPSTAT_INC(tcps_sndacks);
1075 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1076 TCPSTAT_INC(tcps_sndctrl);
1077 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1078 TCPSTAT_INC(tcps_sndurg);
1079 else
1080 TCPSTAT_INC(tcps_sndwinup);
1081
1082 m = m_gethdr(M_NOWAIT, MT_DATA);
1083 if (m == NULL) {
1084 error = ENOBUFS;
1085 sack_rxmit = 0;
1086 goto out;
1087 }
1088 #ifdef INET6
1089 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1090 MHLEN >= hdrlen) {
1091 M_ALIGN(m, hdrlen);
1092 } else
1093 #endif
1094 m->m_data += max_linkhdr;
1095 m->m_len = hdrlen;
1096 }
1097 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1098 m->m_pkthdr.rcvif = (struct ifnet *)0;
1099 #ifdef MAC
1100 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1101 #endif
1102 #ifdef INET6
1103 if (isipv6) {
1104 ip6 = mtod(m, struct ip6_hdr *);
1105 th = (struct tcphdr *)(ip6 + 1);
1106 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1107 } else
1108 #endif /* INET6 */
1109 {
1110 ip = mtod(m, struct ip *);
1111 ipov = (struct ipovly *)ip;
1112 th = (struct tcphdr *)(ip + 1);
1113 tcpip_fillheaders(tp->t_inpcb, ip, th);
1114 }
1115
1116 /*
1117 * Fill in fields, remembering maximum advertised
1118 * window for use in delaying messages about window sizes.
1119 * If resending a FIN, be sure not to use a new sequence number.
1120 */
1121 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1122 tp->snd_nxt == tp->snd_max)
1123 tp->snd_nxt--;
1124 /*
1125 * If we are starting a connection, send ECN setup
1126 * SYN packet. If we are on a retransmit, we may
1127 * resend those bits a number of times as per
1128 * RFC 3168.
1129 */
1130 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1131 if (tp->t_rxtshift >= 1) {
1132 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1133 flags |= TH_ECE|TH_CWR;
1134 } else
1135 flags |= TH_ECE|TH_CWR;
1136 }
1137
1138 if (tp->t_state == TCPS_ESTABLISHED &&
1139 (tp->t_flags & TF_ECN_PERMIT)) {
1140 /*
1141 * If the peer has ECN, mark data packets with
1142 * ECN capable transmission (ECT).
1143 * Ignore pure ack packets, retransmissions and window probes.
1144 */
1145 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1146 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1147 #ifdef INET6
1148 if (isipv6)
1149 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1150 else
1151 #endif
1152 ip->ip_tos |= IPTOS_ECN_ECT0;
1153 TCPSTAT_INC(tcps_ecn_ect0);
1154 }
1155
1156 /*
1157 * Reply with proper ECN notifications.
1158 */
1159 if (tp->t_flags & TF_ECN_SND_CWR) {
1160 flags |= TH_CWR;
1161 tp->t_flags &= ~TF_ECN_SND_CWR;
1162 }
1163 if (tp->t_flags & TF_ECN_SND_ECE)
1164 flags |= TH_ECE;
1165 }
1166
1167 /*
1168 * If we are doing retransmissions, then snd_nxt will
1169 * not reflect the first unsent octet. For ACK only
1170 * packets, we do not want the sequence number of the
1171 * retransmitted packet, we want the sequence number
1172 * of the next unsent octet. So, if there is no data
1173 * (and no SYN or FIN), use snd_max instead of snd_nxt
1174 * when filling in ti_seq. But if we are in persist
1175 * state, snd_max might reflect one byte beyond the
1176 * right edge of the window, so use snd_nxt in that
1177 * case, since we know we aren't doing a retransmission.
1178 * (retransmit and persist are mutually exclusive...)
1179 */
1180 if (sack_rxmit == 0) {
1181 if (len || (flags & (TH_SYN|TH_FIN)) ||
1182 tcp_timer_active(tp, TT_PERSIST))
1183 th->th_seq = htonl(tp->snd_nxt);
1184 else
1185 th->th_seq = htonl(tp->snd_max);
1186 } else {
1187 th->th_seq = htonl(p->rxmit);
1188 p->rxmit += len;
1189 tp->sackhint.sack_bytes_rexmit += len;
1190 }
1191 th->th_ack = htonl(tp->rcv_nxt);
1192 if (optlen) {
1193 bcopy(opt, th + 1, optlen);
1194 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1195 }
1196 th->th_flags = flags;
1197 /*
1198 * Calculate receive window. Don't shrink window,
1199 * but avoid silly window syndrome.
1200 */
1201 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1202 recwin < (long)tp->t_maxseg)
1203 recwin = 0;
1204 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1205 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1206 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1207 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1208 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1209
1210 /*
1211 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1212 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1213 * case is handled in syncache.
1214 */
1215 if (flags & TH_SYN)
1216 th->th_win = htons((u_short)
1217 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1218 else
1219 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1220
1221 /*
1222 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1223 * a 0 window. This may cause the remote transmitter to stall. This
1224 * flag tells soreceive() to disable delayed acknowledgements when
1225 * draining the buffer. This can occur if the receiver is attempting
1226 * to read more data than can be buffered prior to transmitting on
1227 * the connection.
1228 */
1229 if (th->th_win == 0) {
1230 tp->t_sndzerowin++;
1231 tp->t_flags |= TF_RXWIN0SENT;
1232 } else
1233 tp->t_flags &= ~TF_RXWIN0SENT;
1234 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1235 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1236 th->th_flags |= TH_URG;
1237 } else
1238 /*
1239 * If no urgent pointer to send, then we pull
1240 * the urgent pointer to the left edge of the send window
1241 * so that it doesn't drift into the send window on sequence
1242 * number wraparound.
1243 */
1244 tp->snd_up = tp->snd_una; /* drag it along */
1245
1246 #ifdef TCP_SIGNATURE
1247 if (to.to_flags & TOF_SIGNATURE) {
1248 int sigoff = to.to_signature - opt;
1249 tcp_signature_compute(m, 0, len, optlen,
1250 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1251 }
1252 #endif
1253
1254 /*
1255 * Put TCP length in extended header, and then
1256 * checksum extended header and data.
1257 */
1258 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1259 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1260 #ifdef INET6
1261 if (isipv6) {
1262 /*
1263 * ip6_plen is not need to be filled now, and will be filled
1264 * in ip6_output.
1265 */
1266 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1267 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1268 optlen + len, IPPROTO_TCP, 0);
1269 }
1270 #endif
1271 #if defined(INET6) && defined(INET)
1272 else
1273 #endif
1274 #ifdef INET
1275 {
1276 m->m_pkthdr.csum_flags = CSUM_TCP;
1277 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1278 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1279
1280 /* IP version must be set here for ipv4/ipv6 checking later */
1281 KASSERT(ip->ip_v == IPVERSION,
1282 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1283 }
1284 #endif
1285
1286 /*
1287 * Enable TSO and specify the size of the segments.
1288 * The TCP pseudo header checksum is always provided.
1289 */
1290 if (tso) {
1291 KASSERT(len > tp->t_maxseg - optlen,
1292 ("%s: len <= tso_segsz", __func__));
1293 m->m_pkthdr.csum_flags |= CSUM_TSO;
1294 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1295 }
1296
1297 #ifdef IPSEC
1298 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1299 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1300 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1301 #else
1302 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1303 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1304 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1305 #endif
1306
1307 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1308 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1309
1310 #ifdef TCPDEBUG
1311 /*
1312 * Trace.
1313 */
1314 if (so->so_options & SO_DEBUG) {
1315 u_short save = 0;
1316 #ifdef INET6
1317 if (!isipv6)
1318 #endif
1319 {
1320 save = ipov->ih_len;
1321 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1322 }
1323 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1324 #ifdef INET6
1325 if (!isipv6)
1326 #endif
1327 ipov->ih_len = save;
1328 }
1329 #endif /* TCPDEBUG */
1330 TCP_PROBE3(debug__output, tp, th, mtod(m, const char *));
1331
1332 /*
1333 * Fill in IP length and desired time to live and
1334 * send to IP level. There should be a better way
1335 * to handle ttl and tos; we could keep them in
1336 * the template, but need a way to checksum without them.
1337 */
1338 /*
1339 * m->m_pkthdr.len should have been set before checksum calculation,
1340 * because in6_cksum() need it.
1341 */
1342 #ifdef INET6
1343 if (isipv6) {
1344 struct route_in6 ro;
1345
1346 bzero(&ro, sizeof(ro));
1347 /*
1348 * we separately set hoplimit for every segment, since the
1349 * user might want to change the value via setsockopt.
1350 * Also, desired default hop limit might be changed via
1351 * Neighbor Discovery.
1352 */
1353 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1354
1355 /*
1356 * Set the packet size here for the benefit of DTrace probes.
1357 * ip6_output() will set it properly; it's supposed to include
1358 * the option header lengths as well.
1359 */
1360 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1361
1362 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1363 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1364 else
1365 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1366
1367 if (tp->t_state == TCPS_SYN_SENT)
1368 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1369
1370 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1371
1372 #ifdef TCPPCAP
1373 /* Save packet, if requested. */
1374 tcp_pcap_add(th, m, &(tp->t_outpkts));
1375 #endif
1376
1377 /* TODO: IPv6 IP6TOS_ECT bit on */
1378 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1379 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1380 NULL, NULL, tp->t_inpcb);
1381
1382 if (error == EMSGSIZE && ro.ro_rt != NULL)
1383 mtu = ro.ro_rt->rt_mtu;
1384 RO_RTFREE(&ro);
1385 }
1386 #endif /* INET6 */
1387 #if defined(INET) && defined(INET6)
1388 else
1389 #endif
1390 #ifdef INET
1391 {
1392 ip->ip_len = htons(m->m_pkthdr.len);
1393 #ifdef INET6
1394 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1395 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1396 #endif /* INET6 */
1397 /*
1398 * If we do path MTU discovery, then we set DF on every packet.
1399 * This might not be the best thing to do according to RFC3390
1400 * Section 2. However the tcp hostcache migitates the problem
1401 * so it affects only the first tcp connection with a host.
1402 *
1403 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1404 */
1405 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1406 ip->ip_off |= htons(IP_DF);
1407 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1408 } else {
1409 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1410 }
1411
1412 if (tp->t_state == TCPS_SYN_SENT)
1413 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1414
1415 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1416
1417 #ifdef TCPPCAP
1418 /* Save packet, if requested. */
1419 tcp_pcap_add(th, m, &(tp->t_outpkts));
1420 #endif
1421
1422 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1423 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1424 tp->t_inpcb);
1425
1426 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1427 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1428 }
1429 #endif /* INET */
1430
1431 out:
1432 /*
1433 * In transmit state, time the transmission and arrange for
1434 * the retransmit. In persist state, just set snd_max.
1435 */
1436 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1437 !tcp_timer_active(tp, TT_PERSIST)) {
1438 tcp_seq startseq = tp->snd_nxt;
1439
1440 /*
1441 * Advance snd_nxt over sequence space of this segment.
1442 */
1443 if (flags & (TH_SYN|TH_FIN)) {
1444 if (flags & TH_SYN)
1445 tp->snd_nxt++;
1446 if (flags & TH_FIN) {
1447 tp->snd_nxt++;
1448 tp->t_flags |= TF_SENTFIN;
1449 }
1450 }
1451 if (sack_rxmit)
1452 goto timer;
1453 tp->snd_nxt += len;
1454 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1455 tp->snd_max = tp->snd_nxt;
1456 /*
1457 * Time this transmission if not a retransmission and
1458 * not currently timing anything.
1459 */
1460 if (tp->t_rtttime == 0) {
1461 tp->t_rtttime = ticks;
1462 tp->t_rtseq = startseq;
1463 TCPSTAT_INC(tcps_segstimed);
1464 }
1465 }
1466
1467 /*
1468 * Set retransmit timer if not currently set,
1469 * and not doing a pure ack or a keep-alive probe.
1470 * Initial value for retransmit timer is smoothed
1471 * round-trip time + 2 * round-trip time variance.
1472 * Initialize shift counter which is used for backoff
1473 * of retransmit time.
1474 */
1475 timer:
1476 if (!tcp_timer_active(tp, TT_REXMT) &&
1477 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1478 (tp->snd_nxt != tp->snd_una))) {
1479 if (tcp_timer_active(tp, TT_PERSIST)) {
1480 tcp_timer_activate(tp, TT_PERSIST, 0);
1481 tp->t_rxtshift = 0;
1482 }
1483 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1484 } else if (len == 0 && sbavail(&so->so_snd) &&
1485 !tcp_timer_active(tp, TT_REXMT) &&
1486 !tcp_timer_active(tp, TT_PERSIST)) {
1487 /*
1488 * Avoid a situation where we do not set persist timer
1489 * after a zero window condition. For example:
1490 * 1) A -> B: packet with enough data to fill the window
1491 * 2) B -> A: ACK for #1 + new data (0 window
1492 * advertisement)
1493 * 3) A -> B: ACK for #2, 0 len packet
1494 *
1495 * In this case, A will not activate the persist timer,
1496 * because it chose to send a packet. Unless tcp_output
1497 * is called for some other reason (delayed ack timer,
1498 * another input packet from B, socket syscall), A will
1499 * not send zero window probes.
1500 *
1501 * So, if you send a 0-length packet, but there is data
1502 * in the socket buffer, and neither the rexmt or
1503 * persist timer is already set, then activate the
1504 * persist timer.
1505 */
1506 tp->t_rxtshift = 0;
1507 tcp_setpersist(tp);
1508 }
1509 } else {
1510 /*
1511 * Persist case, update snd_max but since we are in
1512 * persist mode (no window) we do not update snd_nxt.
1513 */
1514 int xlen = len;
1515 if (flags & TH_SYN)
1516 ++xlen;
1517 if (flags & TH_FIN) {
1518 ++xlen;
1519 tp->t_flags |= TF_SENTFIN;
1520 }
1521 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1522 tp->snd_max = tp->snd_nxt + len;
1523 }
1524
1525 if (error) {
1526
1527 /*
1528 * We know that the packet was lost, so back out the
1529 * sequence number advance, if any.
1530 *
1531 * If the error is EPERM the packet got blocked by the
1532 * local firewall. Normally we should terminate the
1533 * connection but the blocking may have been spurious
1534 * due to a firewall reconfiguration cycle. So we treat
1535 * it like a packet loss and let the retransmit timer and
1536 * timeouts do their work over time.
1537 * XXX: It is a POLA question whether calling tcp_drop right
1538 * away would be the really correct behavior instead.
1539 */
1540 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1541 !tcp_timer_active(tp, TT_PERSIST)) &&
1542 ((flags & TH_SYN) == 0) &&
1543 (error != EPERM)) {
1544 if (sack_rxmit) {
1545 p->rxmit -= len;
1546 tp->sackhint.sack_bytes_rexmit -= len;
1547 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1548 ("sackhint bytes rtx >= 0"));
1549 } else
1550 tp->snd_nxt -= len;
1551 }
1552 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1553 switch (error) {
1554 case EPERM:
1555 tp->t_softerror = error;
1556 return (error);
1557 case ENOBUFS:
1558 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1559 tp->snd_cwnd = tp->t_maxseg;
1560 return (0);
1561 case EMSGSIZE:
1562 /*
1563 * For some reason the interface we used initially
1564 * to send segments changed to another or lowered
1565 * its MTU.
1566 * If TSO was active we either got an interface
1567 * without TSO capabilits or TSO was turned off.
1568 * If we obtained mtu from ip_output() then update
1569 * it and try again.
1570 */
1571 if (tso)
1572 tp->t_flags &= ~TF_TSO;
1573 if (mtu != 0) {
1574 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1575 goto again;
1576 }
1577 return (error);
1578 case EHOSTDOWN:
1579 case EHOSTUNREACH:
1580 case ENETDOWN:
1581 case ENETUNREACH:
1582 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1583 tp->t_softerror = error;
1584 return (0);
1585 }
1586 /* FALLTHROUGH */
1587 default:
1588 return (error);
1589 }
1590 }
1591 TCPSTAT_INC(tcps_sndtotal);
1592
1593 /*
1594 * Data sent (as far as we can tell).
1595 * If this advertises a larger window than any other segment,
1596 * then remember the size of the advertised window.
1597 * Any pending ACK has now been sent.
1598 */
1599 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1600 tp->rcv_adv = tp->rcv_nxt + recwin;
1601 tp->last_ack_sent = tp->rcv_nxt;
1602 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1603 if (tcp_timer_active(tp, TT_DELACK))
1604 tcp_timer_activate(tp, TT_DELACK, 0);
1605 #if 0
1606 /*
1607 * This completely breaks TCP if newreno is turned on. What happens
1608 * is that if delayed-acks are turned on on the receiver, this code
1609 * on the transmitter effectively destroys the TCP window, forcing
1610 * it to four packets (1.5Kx4 = 6K window).
1611 */
1612 if (sendalot && --maxburst)
1613 goto again;
1614 #endif
1615 if (sendalot)
1616 goto again;
1617 return (0);
1618 }
1619
1620 void
1621 tcp_setpersist(struct tcpcb *tp)
1622 {
1623 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1624 int tt;
1625
1626 tp->t_flags &= ~TF_PREVVALID;
1627 if (tcp_timer_active(tp, TT_REXMT))
1628 panic("tcp_setpersist: retransmit pending");
1629 /*
1630 * Start/restart persistence timer.
1631 */
1632 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1633 tcp_persmin, tcp_persmax);
1634 tcp_timer_activate(tp, TT_PERSIST, tt);
1635 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1636 tp->t_rxtshift++;
1637 }
1638
1639 /*
1640 * Insert TCP options according to the supplied parameters to the place
1641 * optp in a consistent way. Can handle unaligned destinations.
1642 *
1643 * The order of the option processing is crucial for optimal packing and
1644 * alignment for the scarce option space.
1645 *
1646 * The optimal order for a SYN/SYN-ACK segment is:
1647 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1648 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1649 *
1650 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1651 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1652 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1653 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1654 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1655 */
1656 int
1657 tcp_addoptions(struct tcpopt *to, u_char *optp)
1658 {
1659 u_int32_t mask, optlen = 0;
1660
1661 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1662 if ((to->to_flags & mask) != mask)
1663 continue;
1664 if (optlen == TCP_MAXOLEN)
1665 break;
1666 switch (to->to_flags & mask) {
1667 case TOF_MSS:
1668 while (optlen % 4) {
1669 optlen += TCPOLEN_NOP;
1670 *optp++ = TCPOPT_NOP;
1671 }
1672 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1673 continue;
1674 optlen += TCPOLEN_MAXSEG;
1675 *optp++ = TCPOPT_MAXSEG;
1676 *optp++ = TCPOLEN_MAXSEG;
1677 to->to_mss = htons(to->to_mss);
1678 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1679 optp += sizeof(to->to_mss);
1680 break;
1681 case TOF_SCALE:
1682 while (!optlen || optlen % 2 != 1) {
1683 optlen += TCPOLEN_NOP;
1684 *optp++ = TCPOPT_NOP;
1685 }
1686 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1687 continue;
1688 optlen += TCPOLEN_WINDOW;
1689 *optp++ = TCPOPT_WINDOW;
1690 *optp++ = TCPOLEN_WINDOW;
1691 *optp++ = to->to_wscale;
1692 break;
1693 case TOF_SACKPERM:
1694 while (optlen % 2) {
1695 optlen += TCPOLEN_NOP;
1696 *optp++ = TCPOPT_NOP;
1697 }
1698 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1699 continue;
1700 optlen += TCPOLEN_SACK_PERMITTED;
1701 *optp++ = TCPOPT_SACK_PERMITTED;
1702 *optp++ = TCPOLEN_SACK_PERMITTED;
1703 break;
1704 case TOF_TS:
1705 while (!optlen || optlen % 4 != 2) {
1706 optlen += TCPOLEN_NOP;
1707 *optp++ = TCPOPT_NOP;
1708 }
1709 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1710 continue;
1711 optlen += TCPOLEN_TIMESTAMP;
1712 *optp++ = TCPOPT_TIMESTAMP;
1713 *optp++ = TCPOLEN_TIMESTAMP;
1714 to->to_tsval = htonl(to->to_tsval);
1715 to->to_tsecr = htonl(to->to_tsecr);
1716 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1717 optp += sizeof(to->to_tsval);
1718 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1719 optp += sizeof(to->to_tsecr);
1720 break;
1721 #ifdef TCP_SIGNATURE
1722 case TOF_SIGNATURE:
1723 {
1724 int siglen = TCPOLEN_SIGNATURE - 2;
1725
1726 while (!optlen || optlen % 4 != 2) {
1727 optlen += TCPOLEN_NOP;
1728 *optp++ = TCPOPT_NOP;
1729 }
1730 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1731 continue;
1732 optlen += TCPOLEN_SIGNATURE;
1733 *optp++ = TCPOPT_SIGNATURE;
1734 *optp++ = TCPOLEN_SIGNATURE;
1735 to->to_signature = optp;
1736 while (siglen--)
1737 *optp++ = 0;
1738 break;
1739 }
1740 #endif
1741 case TOF_SACK:
1742 {
1743 int sackblks = 0;
1744 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1745 tcp_seq sack_seq;
1746
1747 while (!optlen || optlen % 4 != 2) {
1748 optlen += TCPOLEN_NOP;
1749 *optp++ = TCPOPT_NOP;
1750 }
1751 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1752 continue;
1753 optlen += TCPOLEN_SACKHDR;
1754 *optp++ = TCPOPT_SACK;
1755 sackblks = min(to->to_nsacks,
1756 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1757 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1758 while (sackblks--) {
1759 sack_seq = htonl(sack->start);
1760 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1761 optp += sizeof(sack_seq);
1762 sack_seq = htonl(sack->end);
1763 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1764 optp += sizeof(sack_seq);
1765 optlen += TCPOLEN_SACK;
1766 sack++;
1767 }
1768 TCPSTAT_INC(tcps_sack_send_blocks);
1769 break;
1770 }
1771 #ifdef TCP_RFC7413
1772 case TOF_FASTOPEN:
1773 {
1774 int total_len;
1775
1776 /* XXX is there any point to aligning this option? */
1777 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1778 if (TCP_MAXOLEN - optlen < total_len)
1779 continue;
1780 *optp++ = TCPOPT_FAST_OPEN;
1781 *optp++ = total_len;
1782 if (to->to_tfo_len > 0) {
1783 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1784 optp += to->to_tfo_len;
1785 }
1786 optlen += total_len;
1787 break;
1788 }
1789 #endif
1790 default:
1791 panic("%s: unknown TCP option type", __func__);
1792 break;
1793 }
1794 }
1795
1796 /* Terminate and pad TCP options to a 4 byte boundary. */
1797 if (optlen % 4) {
1798 optlen += TCPOLEN_EOL;
1799 *optp++ = TCPOPT_EOL;
1800 }
1801 /*
1802 * According to RFC 793 (STD0007):
1803 * "The content of the header beyond the End-of-Option option
1804 * must be header padding (i.e., zero)."
1805 * and later: "The padding is composed of zeros."
1806 */
1807 while (optlen % 4) {
1808 optlen += TCPOLEN_PAD;
1809 *optp++ = TCPOPT_PAD;
1810 }
1811
1812 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1813 return (optlen);
1814 }
Cache object: 65b6126b6ff73f21bfbd4368b35362f2
|