1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
32 */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_kern_tls.h"
41 #include "opt_tcpdebug.h"
42
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/arb.h>
46 #include <sys/domain.h>
47 #ifdef TCP_HHOOK
48 #include <sys/hhook.h>
49 #endif
50 #include <sys/kernel.h>
51 #ifdef KERN_TLS
52 #include <sys/ktls.h>
53 #endif
54 #include <sys/lock.h>
55 #include <sys/mbuf.h>
56 #include <sys/mutex.h>
57 #include <sys/protosw.h>
58 #include <sys/qmath.h>
59 #include <sys/sdt.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/sysctl.h>
63 #include <sys/stats.h>
64
65 #include <net/if.h>
66 #include <net/route.h>
67 #include <net/route/nhop.h>
68 #include <net/vnet.h>
69
70 #include <netinet/in.h>
71 #include <netinet/in_kdtrace.h>
72 #include <netinet/in_systm.h>
73 #include <netinet/ip.h>
74 #include <netinet/in_pcb.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/ip_options.h>
77 #ifdef INET6
78 #include <netinet6/in6_pcb.h>
79 #include <netinet/ip6.h>
80 #include <netinet6/ip6_var.h>
81 #endif
82 #include <netinet/tcp.h>
83 #define TCPOUTFLAGS
84 #include <netinet/tcp_fsm.h>
85 #include <netinet/tcp_log_buf.h>
86 #include <netinet/tcp_seq.h>
87 #include <netinet/tcp_timer.h>
88 #include <netinet/tcp_var.h>
89 #include <netinet/tcpip.h>
90 #include <netinet/cc/cc.h>
91 #include <netinet/tcp_fastopen.h>
92 #ifdef TCPPCAP
93 #include <netinet/tcp_pcap.h>
94 #endif
95 #ifdef TCPDEBUG
96 #include <netinet/tcp_debug.h>
97 #endif
98 #ifdef TCP_OFFLOAD
99 #include <netinet/tcp_offload.h>
100 #endif
101
102 #include <netipsec/ipsec_support.h>
103
104 #include <machine/in_cksum.h>
105
106 #include <security/mac/mac_framework.h>
107
108 VNET_DEFINE(int, path_mtu_discovery) = 1;
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
110 &VNET_NAME(path_mtu_discovery), 1,
111 "Enable Path MTU Discovery");
112
113 VNET_DEFINE(int, tcp_do_tso) = 1;
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_do_tso), 0,
116 "Enable TCP Segmentation Offload");
117
118 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
119 #define V_tcp_sendspace VNET(tcp_sendspace)
120 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
122
123 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
125 &VNET_NAME(tcp_do_autosndbuf), 0,
126 "Enable automatic send buffer sizing");
127
128 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
130 &VNET_NAME(tcp_autosndbuf_inc), 0,
131 "Incrementor step size of automatic send buffer");
132
133 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
134 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
135 &VNET_NAME(tcp_autosndbuf_max), 0,
136 "Max size of automatic send buffer");
137
138 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
139 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
141 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
142 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
143
144 /*
145 * Make sure that either retransmit or persist timer is set for SYN, FIN and
146 * non-ACK.
147 */
148 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
149 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
150 tcp_timer_active((tp), TT_REXMT) || \
151 tcp_timer_active((tp), TT_PERSIST), \
152 ("neither rexmt nor persist timer is set"))
153
154 static void inline cc_after_idle(struct tcpcb *tp);
155
156 #ifdef TCP_HHOOK
157 /*
158 * Wrapper for the TCP established output helper hook.
159 */
160 void
hhook_run_tcp_est_out(struct tcpcb * tp,struct tcphdr * th,struct tcpopt * to,uint32_t len,int tso)161 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
162 struct tcpopt *to, uint32_t len, int tso)
163 {
164 struct tcp_hhook_data hhook_data;
165
166 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
167 hhook_data.tp = tp;
168 hhook_data.th = th;
169 hhook_data.to = to;
170 hhook_data.len = len;
171 hhook_data.tso = tso;
172
173 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
174 tp->osd);
175 }
176 }
177 #endif
178
179 /*
180 * CC wrapper hook functions
181 */
182 static void inline
cc_after_idle(struct tcpcb * tp)183 cc_after_idle(struct tcpcb *tp)
184 {
185 INP_WLOCK_ASSERT(tp->t_inpcb);
186
187 if (CC_ALGO(tp)->after_idle != NULL)
188 CC_ALGO(tp)->after_idle(tp->ccv);
189 }
190
191 /*
192 * Tcp output routine: figure out what should be sent and send it.
193 */
194 int
tcp_output(struct tcpcb * tp)195 tcp_output(struct tcpcb *tp)
196 {
197 struct socket *so = tp->t_inpcb->inp_socket;
198 int32_t len;
199 uint32_t recwin, sendwin;
200 int off, flags, error = 0; /* Keep compiler happy */
201 u_int if_hw_tsomaxsegcount = 0;
202 u_int if_hw_tsomaxsegsize = 0;
203 struct mbuf *m;
204 struct ip *ip = NULL;
205 #ifdef TCPDEBUG
206 struct ipovly *ipov = NULL;
207 #endif
208 struct tcphdr *th;
209 u_char opt[TCP_MAXOLEN];
210 unsigned ipoptlen, optlen, hdrlen;
211 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
212 unsigned ipsec_optlen = 0;
213 #endif
214 int idle, sendalot, curticks;
215 int sack_rxmit, sack_bytes_rxmt;
216 struct sackhole *p;
217 int tso, mtu;
218 struct tcpopt to;
219 unsigned int wanted_cookie = 0;
220 unsigned int dont_sendalot = 0;
221 #if 0
222 int maxburst = TCP_MAXBURST;
223 #endif
224 #ifdef INET6
225 struct ip6_hdr *ip6 = NULL;
226 int isipv6;
227
228 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
229 #endif
230 #ifdef KERN_TLS
231 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
232 #else
233 const bool hw_tls = false;
234 #endif
235
236 NET_EPOCH_ASSERT();
237 INP_WLOCK_ASSERT(tp->t_inpcb);
238
239 #ifdef TCP_OFFLOAD
240 if (tp->t_flags & TF_TOE)
241 return (tcp_offload_output(tp));
242 #endif
243
244 /*
245 * For TFO connections in SYN_SENT or SYN_RECEIVED,
246 * only allow the initial SYN or SYN|ACK and those sent
247 * by the retransmit timer.
248 */
249 if (IS_FASTOPEN(tp->t_flags) &&
250 ((tp->t_state == TCPS_SYN_SENT) ||
251 (tp->t_state == TCPS_SYN_RECEIVED)) &&
252 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
253 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
254 return (0);
255
256 /*
257 * Determine length of data that should be transmitted,
258 * and flags that will be used.
259 * If there is some data or critical controls (SYN, RST)
260 * to send, then transmit; otherwise, investigate further.
261 */
262 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
263 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
264 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
265 cc_after_idle(tp);
266 tp->t_flags &= ~TF_LASTIDLE;
267 if (idle) {
268 if (tp->t_flags & TF_MORETOCOME) {
269 tp->t_flags |= TF_LASTIDLE;
270 idle = 0;
271 }
272 }
273 again:
274 /*
275 * If we've recently taken a timeout, snd_max will be greater than
276 * snd_nxt. There may be SACK information that allows us to avoid
277 * resending already delivered data. Adjust snd_nxt accordingly.
278 */
279 if ((tp->t_flags & TF_SACK_PERMIT) &&
280 SEQ_LT(tp->snd_nxt, tp->snd_max))
281 tcp_sack_adjust(tp);
282 sendalot = 0;
283 tso = 0;
284 mtu = 0;
285 off = tp->snd_nxt - tp->snd_una;
286 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
287
288 flags = tcp_outflags[tp->t_state];
289 /*
290 * Send any SACK-generated retransmissions. If we're explicitly trying
291 * to send out new data (when sendalot is 1), bypass this function.
292 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
293 * we're replacing a (future) new transmission with a retransmission
294 * now, and we previously incremented snd_cwnd in tcp_input().
295 */
296 /*
297 * Still in sack recovery , reset rxmit flag to zero.
298 */
299 sack_rxmit = 0;
300 sack_bytes_rxmt = 0;
301 len = 0;
302 p = NULL;
303 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
304 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
305 uint32_t cwin;
306
307 cwin =
308 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
309 /* Do not retransmit SACK segments beyond snd_recover */
310 if (SEQ_GT(p->end, tp->snd_recover)) {
311 /*
312 * (At least) part of sack hole extends beyond
313 * snd_recover. Check to see if we can rexmit data
314 * for this hole.
315 */
316 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
317 /*
318 * Can't rexmit any more data for this hole.
319 * That data will be rexmitted in the next
320 * sack recovery episode, when snd_recover
321 * moves past p->rxmit.
322 */
323 p = NULL;
324 goto after_sack_rexmit;
325 } else
326 /* Can rexmit part of the current hole */
327 len = ((int32_t)ulmin(cwin,
328 tp->snd_recover - p->rxmit));
329 } else
330 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
331 off = p->rxmit - tp->snd_una;
332 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
333 __func__, off));
334 if (len > 0) {
335 sack_rxmit = 1;
336 sendalot = 1;
337 TCPSTAT_INC(tcps_sack_rexmits);
338 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
339 min(len, tcp_maxseg(tp)));
340 }
341 }
342 after_sack_rexmit:
343 /*
344 * Get standard flags, and add SYN or FIN if requested by 'hidden'
345 * state flags.
346 */
347 if (tp->t_flags & TF_NEEDFIN)
348 flags |= TH_FIN;
349 if (tp->t_flags & TF_NEEDSYN)
350 flags |= TH_SYN;
351
352 SOCKBUF_LOCK(&so->so_snd);
353 /*
354 * If in persist timeout with window of 0, send 1 byte.
355 * Otherwise, if window is small but nonzero
356 * and timer expired, we will send what we can
357 * and go to transmit state.
358 */
359 if (tp->t_flags & TF_FORCEDATA) {
360 if (sendwin == 0) {
361 /*
362 * If we still have some data to send, then
363 * clear the FIN bit. Usually this would
364 * happen below when it realizes that we
365 * aren't sending all the data. However,
366 * if we have exactly 1 byte of unsent data,
367 * then it won't clear the FIN bit below,
368 * and if we are in persist state, we wind
369 * up sending the packet without recording
370 * that we sent the FIN bit.
371 *
372 * We can't just blindly clear the FIN bit,
373 * because if we don't have any more data
374 * to send then the probe will be the FIN
375 * itself.
376 */
377 if (off < sbused(&so->so_snd))
378 flags &= ~TH_FIN;
379 sendwin = 1;
380 } else {
381 tcp_timer_activate(tp, TT_PERSIST, 0);
382 tp->t_rxtshift = 0;
383 }
384 }
385
386 /*
387 * If snd_nxt == snd_max and we have transmitted a FIN, the
388 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
389 * a negative length. This can also occur when TCP opens up
390 * its congestion window while receiving additional duplicate
391 * acks after fast-retransmit because TCP will reset snd_nxt
392 * to snd_max after the fast-retransmit.
393 *
394 * In the normal retransmit-FIN-only case, however, snd_nxt will
395 * be set to snd_una, the offset will be 0, and the length may
396 * wind up 0.
397 *
398 * If sack_rxmit is true we are retransmitting from the scoreboard
399 * in which case len is already set.
400 */
401 if (sack_rxmit == 0) {
402 if (sack_bytes_rxmt == 0)
403 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
404 off);
405 else {
406 int32_t cwin;
407
408 /*
409 * We are inside of a SACK recovery episode and are
410 * sending new data, having retransmitted all the
411 * data possible in the scoreboard.
412 */
413 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
414 off);
415 /*
416 * Don't remove this (len > 0) check !
417 * We explicitly check for len > 0 here (although it
418 * isn't really necessary), to work around a gcc
419 * optimization issue - to force gcc to compute
420 * len above. Without this check, the computation
421 * of len is bungled by the optimizer.
422 */
423 if (len > 0) {
424 cwin = tp->snd_cwnd -
425 (tp->snd_nxt - tp->snd_recover) -
426 sack_bytes_rxmt;
427 if (cwin < 0)
428 cwin = 0;
429 len = imin(len, cwin);
430 }
431 }
432 }
433
434 /*
435 * Lop off SYN bit if it has already been sent. However, if this
436 * is SYN-SENT state and if segment contains data and if we don't
437 * know that foreign host supports TAO, suppress sending segment.
438 */
439 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
440 if (tp->t_state != TCPS_SYN_RECEIVED)
441 flags &= ~TH_SYN;
442 /*
443 * When sending additional segments following a TFO SYN|ACK,
444 * do not include the SYN bit.
445 */
446 if (IS_FASTOPEN(tp->t_flags) &&
447 (tp->t_state == TCPS_SYN_RECEIVED))
448 flags &= ~TH_SYN;
449 off--, len++;
450 }
451
452 /*
453 * Be careful not to send data and/or FIN on SYN segments.
454 * This measure is needed to prevent interoperability problems
455 * with not fully conformant TCP implementations.
456 */
457 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
458 len = 0;
459 flags &= ~TH_FIN;
460 }
461
462 /*
463 * On TFO sockets, ensure no data is sent in the following cases:
464 *
465 * - When retransmitting SYN|ACK on a passively-created socket
466 *
467 * - When retransmitting SYN on an actively created socket
468 *
469 * - When sending a zero-length cookie (cookie request) on an
470 * actively created socket
471 *
472 * - When the socket is in the CLOSED state (RST is being sent)
473 */
474 if (IS_FASTOPEN(tp->t_flags) &&
475 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
476 ((tp->t_state == TCPS_SYN_SENT) &&
477 (tp->t_tfo_client_cookie_len == 0)) ||
478 (flags & TH_RST)))
479 len = 0;
480 if (len <= 0) {
481 /*
482 * If FIN has been sent but not acked,
483 * but we haven't been called to retransmit,
484 * len will be < 0. Otherwise, window shrank
485 * after we sent into it. If window shrank to 0,
486 * cancel pending retransmit, pull snd_nxt back
487 * to (closed) window, and set the persist timer
488 * if it isn't already going. If the window didn't
489 * close completely, just wait for an ACK.
490 *
491 * We also do a general check here to ensure that
492 * we will set the persist timer when we have data
493 * to send, but a 0-byte window. This makes sure
494 * the persist timer is set even if the packet
495 * hits one of the "goto send" lines below.
496 */
497 len = 0;
498 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
499 (off < (int) sbavail(&so->so_snd))) {
500 tcp_timer_activate(tp, TT_REXMT, 0);
501 tp->t_rxtshift = 0;
502 tp->snd_nxt = tp->snd_una;
503 if (!tcp_timer_active(tp, TT_PERSIST))
504 tcp_setpersist(tp);
505 }
506 }
507
508 /* len will be >= 0 after this point. */
509 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
510
511 tcp_sndbuf_autoscale(tp, so, sendwin);
512
513 /*
514 * Decide if we can use TCP Segmentation Offloading (if supported by
515 * hardware).
516 *
517 * TSO may only be used if we are in a pure bulk sending state. The
518 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
519 * IP options prevent using TSO. With TSO the TCP header is the same
520 * (except for the sequence number) for all generated packets. This
521 * makes it impossible to transmit any options which vary per generated
522 * segment or packet.
523 *
524 * IPv4 handling has a clear separation of ip options and ip header
525 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
526 * the right thing below to provide length of just ip options and thus
527 * checking for ipoptlen is enough to decide if ip options are present.
528 */
529 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
530 /*
531 * Pre-calculate here as we save another lookup into the darknesses
532 * of IPsec that way and can actually decide if TSO is ok.
533 */
534 #ifdef INET6
535 if (isipv6 && IPSEC_ENABLED(ipv6))
536 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
537 #ifdef INET
538 else
539 #endif
540 #endif /* INET6 */
541 #ifdef INET
542 if (IPSEC_ENABLED(ipv4))
543 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
544 #endif /* INET */
545 #endif /* IPSEC */
546 #ifdef INET6
547 if (isipv6)
548 ipoptlen = ip6_optlen(tp->t_inpcb);
549 else
550 #endif
551 if (tp->t_inpcb->inp_options)
552 ipoptlen = tp->t_inpcb->inp_options->m_len -
553 offsetof(struct ipoption, ipopt_list);
554 else
555 ipoptlen = 0;
556 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
557 ipoptlen += ipsec_optlen;
558 #endif
559
560 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
561 ((tp->t_flags & TF_SIGNATURE) == 0) &&
562 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
563 ipoptlen == 0 && !(flags & TH_SYN))
564 tso = 1;
565
566 if (sack_rxmit) {
567 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
568 flags &= ~TH_FIN;
569 } else {
570 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
571 sbused(&so->so_snd)))
572 flags &= ~TH_FIN;
573 }
574
575 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
576 (long)TCP_MAXWIN << tp->rcv_scale);
577
578 /*
579 * Sender silly window avoidance. We transmit under the following
580 * conditions when len is non-zero:
581 *
582 * - We have a full segment (or more with TSO)
583 * - This is the last buffer in a write()/send() and we are
584 * either idle or running NODELAY
585 * - we've timed out (e.g. persist timer)
586 * - we have more then 1/2 the maximum send window's worth of
587 * data (receiver may be limited the window size)
588 * - we need to retransmit
589 */
590 if (len) {
591 if (len >= tp->t_maxseg)
592 goto send;
593 /*
594 * As the TCP header options are now
595 * considered when setting up the initial
596 * window, we would not send the last segment
597 * if we skip considering the option length here.
598 * Note: this may not work when tcp headers change
599 * very dynamically in the future.
600 */
601 if ((((tp->t_flags & TF_SIGNATURE) ?
602 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
603 ((tp->t_flags & TF_RCVD_TSTMP) ?
604 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
605 len) >= tp->t_maxseg)
606 goto send;
607 /*
608 * NOTE! on localhost connections an 'ack' from the remote
609 * end may occur synchronously with the output and cause
610 * us to flush a buffer queued with moretocome. XXX
611 *
612 * note: the len + off check is almost certainly unnecessary.
613 */
614 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
615 (idle || (tp->t_flags & TF_NODELAY)) &&
616 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
617 (tp->t_flags & TF_NOPUSH) == 0) {
618 goto send;
619 }
620 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
621 goto send;
622 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
623 goto send;
624 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
625 goto send;
626 if (sack_rxmit)
627 goto send;
628 }
629
630 /*
631 * Sending of standalone window updates.
632 *
633 * Window updates are important when we close our window due to a
634 * full socket buffer and are opening it again after the application
635 * reads data from it. Once the window has opened again and the
636 * remote end starts to send again the ACK clock takes over and
637 * provides the most current window information.
638 *
639 * We must avoid the silly window syndrome whereas every read
640 * from the receive buffer, no matter how small, causes a window
641 * update to be sent. We also should avoid sending a flurry of
642 * window updates when the socket buffer had queued a lot of data
643 * and the application is doing small reads.
644 *
645 * Prevent a flurry of pointless window updates by only sending
646 * an update when we can increase the advertized window by more
647 * than 1/4th of the socket buffer capacity. When the buffer is
648 * getting full or is very small be more aggressive and send an
649 * update whenever we can increase by two mss sized segments.
650 * In all other situations the ACK's to new incoming data will
651 * carry further window increases.
652 *
653 * Don't send an independent window update if a delayed
654 * ACK is pending (it will get piggy-backed on it) or the
655 * remote side already has done a half-close and won't send
656 * more data. Skip this if the connection is in T/TCP
657 * half-open state.
658 */
659 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
660 !(tp->t_flags & TF_DELACK) &&
661 !TCPS_HAVERCVDFIN(tp->t_state)) {
662 /*
663 * "adv" is the amount we could increase the window,
664 * taking into account that we are limited by
665 * TCP_MAXWIN << tp->rcv_scale.
666 */
667 int32_t adv;
668 int oldwin;
669
670 adv = recwin;
671 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
672 oldwin = (tp->rcv_adv - tp->rcv_nxt);
673 if (adv > oldwin)
674 adv -= oldwin;
675 else
676 adv = 0;
677 } else
678 oldwin = 0;
679
680 /*
681 * If the new window size ends up being the same as or less
682 * than the old size when it is scaled, then don't force
683 * a window update.
684 */
685 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
686 goto dontupdate;
687
688 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
689 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
690 recwin <= (so->so_rcv.sb_hiwat / 8) ||
691 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
692 adv >= TCP_MAXWIN << tp->rcv_scale))
693 goto send;
694 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
695 goto send;
696 }
697 dontupdate:
698
699 /*
700 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
701 * is also a catch-all for the retransmit timer timeout case.
702 */
703 if (tp->t_flags & TF_ACKNOW)
704 goto send;
705 if ((flags & TH_RST) ||
706 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
707 goto send;
708 if (SEQ_GT(tp->snd_up, tp->snd_una))
709 goto send;
710 /*
711 * If our state indicates that FIN should be sent
712 * and we have not yet done so, then we need to send.
713 */
714 if (flags & TH_FIN &&
715 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
716 goto send;
717 /*
718 * In SACK, it is possible for tcp_output to fail to send a segment
719 * after the retransmission timer has been turned off. Make sure
720 * that the retransmission timer is set.
721 */
722 if ((tp->t_flags & TF_SACK_PERMIT) &&
723 SEQ_GT(tp->snd_max, tp->snd_una) &&
724 !tcp_timer_active(tp, TT_REXMT) &&
725 !tcp_timer_active(tp, TT_PERSIST)) {
726 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
727 goto just_return;
728 }
729 /*
730 * TCP window updates are not reliable, rather a polling protocol
731 * using ``persist'' packets is used to insure receipt of window
732 * updates. The three ``states'' for the output side are:
733 * idle not doing retransmits or persists
734 * persisting to move a small or zero window
735 * (re)transmitting and thereby not persisting
736 *
737 * tcp_timer_active(tp, TT_PERSIST)
738 * is true when we are in persist state.
739 * (tp->t_flags & TF_FORCEDATA)
740 * is set when we are called to send a persist packet.
741 * tcp_timer_active(tp, TT_REXMT)
742 * is set when we are retransmitting
743 * The output side is idle when both timers are zero.
744 *
745 * If send window is too small, there is data to transmit, and no
746 * retransmit or persist is pending, then go to persist state.
747 * If nothing happens soon, send when timer expires:
748 * if window is nonzero, transmit what we can,
749 * otherwise force out a byte.
750 */
751 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
752 !tcp_timer_active(tp, TT_PERSIST)) {
753 tp->t_rxtshift = 0;
754 tcp_setpersist(tp);
755 }
756
757 /*
758 * No reason to send a segment, just return.
759 */
760 just_return:
761 SOCKBUF_UNLOCK(&so->so_snd);
762 return (0);
763
764 send:
765 SOCKBUF_LOCK_ASSERT(&so->so_snd);
766 if (len > 0) {
767 if (len >= tp->t_maxseg)
768 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
769 else
770 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
771 }
772 /*
773 * Before ESTABLISHED, force sending of initial options
774 * unless TCP set not to do any options.
775 * NOTE: we assume that the IP/TCP header plus TCP options
776 * always fit in a single mbuf, leaving room for a maximum
777 * link header, i.e.
778 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
779 */
780 optlen = 0;
781 #ifdef INET6
782 if (isipv6)
783 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
784 else
785 #endif
786 hdrlen = sizeof (struct tcpiphdr);
787
788 if (flags & TH_SYN) {
789 tp->snd_nxt = tp->iss;
790 }
791
792 /*
793 * Compute options for segment.
794 * We only have to care about SYN and established connection
795 * segments. Options for SYN-ACK segments are handled in TCP
796 * syncache.
797 */
798 to.to_flags = 0;
799 if ((tp->t_flags & TF_NOOPT) == 0) {
800 /* Maximum segment size. */
801 if (flags & TH_SYN) {
802 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
803 to.to_flags |= TOF_MSS;
804
805 /*
806 * On SYN or SYN|ACK transmits on TFO connections,
807 * only include the TFO option if it is not a
808 * retransmit, as the presence of the TFO option may
809 * have caused the original SYN or SYN|ACK to have
810 * been dropped by a middlebox.
811 */
812 if (IS_FASTOPEN(tp->t_flags) &&
813 (tp->t_rxtshift == 0)) {
814 if (tp->t_state == TCPS_SYN_RECEIVED) {
815 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
816 to.to_tfo_cookie =
817 (u_int8_t *)&tp->t_tfo_cookie.server;
818 to.to_flags |= TOF_FASTOPEN;
819 wanted_cookie = 1;
820 } else if (tp->t_state == TCPS_SYN_SENT) {
821 to.to_tfo_len =
822 tp->t_tfo_client_cookie_len;
823 to.to_tfo_cookie =
824 tp->t_tfo_cookie.client;
825 to.to_flags |= TOF_FASTOPEN;
826 wanted_cookie = 1;
827 /*
828 * If we wind up having more data to
829 * send with the SYN than can fit in
830 * one segment, don't send any more
831 * until the SYN|ACK comes back from
832 * the other end.
833 */
834 dont_sendalot = 1;
835 }
836 }
837 }
838 /* Window scaling. */
839 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
840 to.to_wscale = tp->request_r_scale;
841 to.to_flags |= TOF_SCALE;
842 }
843 /* Timestamps. */
844 if ((tp->t_flags & TF_RCVD_TSTMP) ||
845 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
846 curticks = tcp_ts_getticks();
847 to.to_tsval = curticks + tp->ts_offset;
848 to.to_tsecr = tp->ts_recent;
849 to.to_flags |= TOF_TS;
850 if (tp->t_rxtshift == 1)
851 tp->t_badrxtwin = curticks;
852 }
853
854 /* Set receive buffer autosizing timestamp. */
855 if (tp->rfbuf_ts == 0 &&
856 (so->so_rcv.sb_flags & SB_AUTOSIZE))
857 tp->rfbuf_ts = tcp_ts_getticks();
858
859 /* Selective ACK's. */
860 if (tp->t_flags & TF_SACK_PERMIT) {
861 if (flags & TH_SYN)
862 to.to_flags |= TOF_SACKPERM;
863 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
864 tp->rcv_numsacks > 0) {
865 to.to_flags |= TOF_SACK;
866 to.to_nsacks = tp->rcv_numsacks;
867 to.to_sacks = (u_char *)tp->sackblks;
868 }
869 }
870 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
871 /* TCP-MD5 (RFC2385). */
872 /*
873 * Check that TCP_MD5SIG is enabled in tcpcb to
874 * account the size needed to set this TCP option.
875 */
876 if (tp->t_flags & TF_SIGNATURE)
877 to.to_flags |= TOF_SIGNATURE;
878 #endif /* TCP_SIGNATURE */
879
880 /* Processing the options. */
881 hdrlen += optlen = tcp_addoptions(&to, opt);
882 /*
883 * If we wanted a TFO option to be added, but it was unable
884 * to fit, ensure no data is sent.
885 */
886 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
887 !(to.to_flags & TOF_FASTOPEN))
888 len = 0;
889 }
890
891 /*
892 * Adjust data length if insertion of options will
893 * bump the packet length beyond the t_maxseg length.
894 * Clear the FIN bit because we cut off the tail of
895 * the segment.
896 */
897 if (len + optlen + ipoptlen > tp->t_maxseg) {
898 flags &= ~TH_FIN;
899
900 if (tso) {
901 u_int if_hw_tsomax;
902 u_int moff;
903 int max_len;
904
905 /* extract TSO information */
906 if_hw_tsomax = tp->t_tsomax;
907 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
908 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
909
910 /*
911 * Limit a TSO burst to prevent it from
912 * overflowing or exceeding the maximum length
913 * allowed by the network interface:
914 */
915 KASSERT(ipoptlen == 0,
916 ("%s: TSO can't do IP options", __func__));
917
918 /*
919 * Check if we should limit by maximum payload
920 * length:
921 */
922 if (if_hw_tsomax != 0) {
923 /* compute maximum TSO length */
924 max_len = (if_hw_tsomax - hdrlen -
925 max_linkhdr);
926 if (max_len <= 0) {
927 len = 0;
928 } else if (len > max_len) {
929 sendalot = 1;
930 len = max_len;
931 }
932 }
933
934 /*
935 * Prevent the last segment from being
936 * fractional unless the send sockbuf can be
937 * emptied:
938 */
939 max_len = (tp->t_maxseg - optlen);
940 if (((uint32_t)off + (uint32_t)len) <
941 sbavail(&so->so_snd)) {
942 moff = len % max_len;
943 if (moff != 0) {
944 len -= moff;
945 sendalot = 1;
946 }
947 }
948
949 /*
950 * In case there are too many small fragments
951 * don't use TSO:
952 */
953 if (len <= max_len) {
954 len = max_len;
955 sendalot = 1;
956 tso = 0;
957 }
958
959 /*
960 * Send the FIN in a separate segment
961 * after the bulk sending is done.
962 * We don't trust the TSO implementations
963 * to clear the FIN flag on all but the
964 * last segment.
965 */
966 if (tp->t_flags & TF_NEEDFIN)
967 sendalot = 1;
968 } else {
969 if (optlen + ipoptlen >= tp->t_maxseg) {
970 /*
971 * Since we don't have enough space to put
972 * the IP header chain and the TCP header in
973 * one packet as required by RFC 7112, don't
974 * send it. Also ensure that at least one
975 * byte of the payload can be put into the
976 * TCP segment.
977 */
978 SOCKBUF_UNLOCK(&so->so_snd);
979 error = EMSGSIZE;
980 sack_rxmit = 0;
981 goto out;
982 }
983 len = tp->t_maxseg - optlen - ipoptlen;
984 sendalot = 1;
985 if (dont_sendalot)
986 sendalot = 0;
987 }
988 } else
989 tso = 0;
990
991 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
992 ("%s: len > IP_MAXPACKET", __func__));
993
994 /*#ifdef DIAGNOSTIC*/
995 #ifdef INET6
996 if (max_linkhdr + hdrlen > MCLBYTES)
997 #else
998 if (max_linkhdr + hdrlen > MHLEN)
999 #endif
1000 panic("tcphdr too big");
1001 /*#endif*/
1002
1003 /*
1004 * This KASSERT is here to catch edge cases at a well defined place.
1005 * Before, those had triggered (random) panic conditions further down.
1006 */
1007 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1008
1009 /*
1010 * Grab a header mbuf, attaching a copy of data to
1011 * be transmitted, and initialize the header from
1012 * the template for sends on this connection.
1013 */
1014 if (len) {
1015 struct mbuf *mb;
1016 struct sockbuf *msb;
1017 u_int moff;
1018
1019 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1020 TCPSTAT_INC(tcps_sndprobe);
1021 #ifdef STATS
1022 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1023 stats_voi_update_abs_u32(tp->t_stats,
1024 VOI_TCP_RETXPB, len);
1025 else
1026 stats_voi_update_abs_u64(tp->t_stats,
1027 VOI_TCP_TXPB, len);
1028 #endif /* STATS */
1029 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1030 tp->t_sndrexmitpack++;
1031 TCPSTAT_INC(tcps_sndrexmitpack);
1032 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1033 #ifdef STATS
1034 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1035 len);
1036 #endif /* STATS */
1037 } else {
1038 TCPSTAT_INC(tcps_sndpack);
1039 TCPSTAT_ADD(tcps_sndbyte, len);
1040 #ifdef STATS
1041 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1042 len);
1043 #endif /* STATS */
1044 }
1045 #ifdef INET6
1046 if (MHLEN < hdrlen + max_linkhdr)
1047 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1048 else
1049 #endif
1050 m = m_gethdr(M_NOWAIT, MT_DATA);
1051
1052 if (m == NULL) {
1053 SOCKBUF_UNLOCK(&so->so_snd);
1054 error = ENOBUFS;
1055 sack_rxmit = 0;
1056 goto out;
1057 }
1058
1059 m->m_data += max_linkhdr;
1060 m->m_len = hdrlen;
1061
1062 /*
1063 * Start the m_copy functions from the closest mbuf
1064 * to the offset in the socket buffer chain.
1065 */
1066 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1067 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1068 m_copydata(mb, moff, len,
1069 mtod(m, caddr_t) + hdrlen);
1070 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1071 sbsndptr_adv(&so->so_snd, mb, len);
1072 m->m_len += len;
1073 } else {
1074 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1075 msb = NULL;
1076 else
1077 msb = &so->so_snd;
1078 m->m_next = tcp_m_copym(mb, moff,
1079 &len, if_hw_tsomaxsegcount,
1080 if_hw_tsomaxsegsize, msb, hw_tls);
1081 if (len <= (tp->t_maxseg - optlen)) {
1082 /*
1083 * Must have ran out of mbufs for the copy
1084 * shorten it to no longer need tso. Lets
1085 * not put on sendalot since we are low on
1086 * mbufs.
1087 */
1088 tso = 0;
1089 }
1090 if (m->m_next == NULL) {
1091 SOCKBUF_UNLOCK(&so->so_snd);
1092 (void) m_free(m);
1093 error = ENOBUFS;
1094 sack_rxmit = 0;
1095 goto out;
1096 }
1097 }
1098
1099 /*
1100 * If we're sending everything we've got, set PUSH.
1101 * (This will keep happy those implementations which only
1102 * give data to the user when a buffer fills or
1103 * a PUSH comes in.)
1104 */
1105 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1106 !(flags & TH_SYN))
1107 flags |= TH_PUSH;
1108 SOCKBUF_UNLOCK(&so->so_snd);
1109 } else {
1110 SOCKBUF_UNLOCK(&so->so_snd);
1111 if (tp->t_flags & TF_ACKNOW)
1112 TCPSTAT_INC(tcps_sndacks);
1113 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1114 TCPSTAT_INC(tcps_sndctrl);
1115 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1116 TCPSTAT_INC(tcps_sndurg);
1117 else
1118 TCPSTAT_INC(tcps_sndwinup);
1119
1120 m = m_gethdr(M_NOWAIT, MT_DATA);
1121 if (m == NULL) {
1122 error = ENOBUFS;
1123 sack_rxmit = 0;
1124 goto out;
1125 }
1126 #ifdef INET6
1127 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1128 MHLEN >= hdrlen) {
1129 M_ALIGN(m, hdrlen);
1130 } else
1131 #endif
1132 m->m_data += max_linkhdr;
1133 m->m_len = hdrlen;
1134 }
1135 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1136 m->m_pkthdr.rcvif = (struct ifnet *)0;
1137 #ifdef MAC
1138 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1139 #endif
1140 #ifdef INET6
1141 if (isipv6) {
1142 ip6 = mtod(m, struct ip6_hdr *);
1143 th = (struct tcphdr *)(ip6 + 1);
1144 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1145 } else
1146 #endif /* INET6 */
1147 {
1148 ip = mtod(m, struct ip *);
1149 #ifdef TCPDEBUG
1150 ipov = (struct ipovly *)ip;
1151 #endif
1152 th = (struct tcphdr *)(ip + 1);
1153 tcpip_fillheaders(tp->t_inpcb, ip, th);
1154 }
1155
1156 /*
1157 * Fill in fields, remembering maximum advertised
1158 * window for use in delaying messages about window sizes.
1159 * If resending a FIN, be sure not to use a new sequence number.
1160 */
1161 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1162 tp->snd_nxt == tp->snd_max)
1163 tp->snd_nxt--;
1164 /*
1165 * If we are starting a connection, send ECN setup
1166 * SYN packet. If we are on a retransmit, we may
1167 * resend those bits a number of times as per
1168 * RFC 3168.
1169 */
1170 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1171 if (tp->t_rxtshift >= 1) {
1172 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1173 flags |= TH_ECE|TH_CWR;
1174 } else
1175 flags |= TH_ECE|TH_CWR;
1176 }
1177 /* Handle parallel SYN for ECN */
1178 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1179 (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1180 flags |= TH_ECE;
1181 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1182 }
1183
1184 if (tp->t_state == TCPS_ESTABLISHED &&
1185 (tp->t_flags2 & TF2_ECN_PERMIT)) {
1186 /*
1187 * If the peer has ECN, mark data packets with
1188 * ECN capable transmission (ECT).
1189 * Ignore pure ack packets, retransmissions and window probes.
1190 */
1191 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1192 (sack_rxmit == 0) &&
1193 !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1194 SEQ_LT(tp->snd_una, tp->snd_max))) {
1195 #ifdef INET6
1196 if (isipv6)
1197 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1198 else
1199 #endif
1200 ip->ip_tos |= IPTOS_ECN_ECT0;
1201 TCPSTAT_INC(tcps_ecn_ect0);
1202 /*
1203 * Reply with proper ECN notifications.
1204 * Only set CWR on new data segments.
1205 */
1206 if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1207 flags |= TH_CWR;
1208 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1209 }
1210 }
1211 if (tp->t_flags2 & TF2_ECN_SND_ECE)
1212 flags |= TH_ECE;
1213 }
1214
1215 /*
1216 * If we are doing retransmissions, then snd_nxt will
1217 * not reflect the first unsent octet. For ACK only
1218 * packets, we do not want the sequence number of the
1219 * retransmitted packet, we want the sequence number
1220 * of the next unsent octet. So, if there is no data
1221 * (and no SYN or FIN), use snd_max instead of snd_nxt
1222 * when filling in ti_seq. But if we are in persist
1223 * state, snd_max might reflect one byte beyond the
1224 * right edge of the window, so use snd_nxt in that
1225 * case, since we know we aren't doing a retransmission.
1226 * (retransmit and persist are mutually exclusive...)
1227 */
1228 if (sack_rxmit == 0) {
1229 if (len || (flags & (TH_SYN|TH_FIN)) ||
1230 tcp_timer_active(tp, TT_PERSIST))
1231 th->th_seq = htonl(tp->snd_nxt);
1232 else
1233 th->th_seq = htonl(tp->snd_max);
1234 } else {
1235 th->th_seq = htonl(p->rxmit);
1236 p->rxmit += len;
1237 tp->sackhint.sack_bytes_rexmit += len;
1238 }
1239 if (IN_RECOVERY(tp->t_flags)) {
1240 /*
1241 * Account all bytes transmitted while
1242 * IN_RECOVERY, simplifying PRR and
1243 * Lost Retransmit Detection
1244 */
1245 tp->sackhint.prr_out += len;
1246 }
1247 th->th_ack = htonl(tp->rcv_nxt);
1248 if (optlen) {
1249 bcopy(opt, th + 1, optlen);
1250 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1251 }
1252 th->th_flags = flags;
1253 /*
1254 * Calculate receive window. Don't shrink window,
1255 * but avoid silly window syndrome.
1256 * If a RST segment is sent, advertise a window of zero.
1257 */
1258 if (flags & TH_RST) {
1259 recwin = 0;
1260 } else {
1261 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1262 recwin < tp->t_maxseg)
1263 recwin = 0;
1264 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1265 recwin < (tp->rcv_adv - tp->rcv_nxt))
1266 recwin = (tp->rcv_adv - tp->rcv_nxt);
1267 }
1268 /*
1269 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1270 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1271 * case is handled in syncache.
1272 */
1273 if (flags & TH_SYN)
1274 th->th_win = htons((u_short)
1275 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1276 else {
1277 /* Avoid shrinking window with window scaling. */
1278 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1279 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1280 }
1281
1282 /*
1283 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1284 * a 0 window. This may cause the remote transmitter to stall. This
1285 * flag tells soreceive() to disable delayed acknowledgements when
1286 * draining the buffer. This can occur if the receiver is attempting
1287 * to read more data than can be buffered prior to transmitting on
1288 * the connection.
1289 */
1290 if (th->th_win == 0) {
1291 tp->t_sndzerowin++;
1292 tp->t_flags |= TF_RXWIN0SENT;
1293 } else
1294 tp->t_flags &= ~TF_RXWIN0SENT;
1295 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1296 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1297 th->th_flags |= TH_URG;
1298 } else
1299 /*
1300 * If no urgent pointer to send, then we pull
1301 * the urgent pointer to the left edge of the send window
1302 * so that it doesn't drift into the send window on sequence
1303 * number wraparound.
1304 */
1305 tp->snd_up = tp->snd_una; /* drag it along */
1306
1307 /*
1308 * Put TCP length in extended header, and then
1309 * checksum extended header and data.
1310 */
1311 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1312 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1313
1314 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1315 if (to.to_flags & TOF_SIGNATURE) {
1316 /*
1317 * Calculate MD5 signature and put it into the place
1318 * determined before.
1319 * NOTE: since TCP options buffer doesn't point into
1320 * mbuf's data, calculate offset and use it.
1321 */
1322 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1323 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1324 /*
1325 * Do not send segment if the calculation of MD5
1326 * digest has failed.
1327 */
1328 m_freem(m);
1329 goto out;
1330 }
1331 }
1332 #endif
1333 #ifdef INET6
1334 if (isipv6) {
1335 /*
1336 * There is no need to fill in ip6_plen right now.
1337 * It will be filled later by ip6_output.
1338 */
1339 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1340 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1341 optlen + len, IPPROTO_TCP, 0);
1342 }
1343 #endif
1344 #if defined(INET6) && defined(INET)
1345 else
1346 #endif
1347 #ifdef INET
1348 {
1349 m->m_pkthdr.csum_flags = CSUM_TCP;
1350 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1351 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1352
1353 /* IP version must be set here for ipv4/ipv6 checking later */
1354 KASSERT(ip->ip_v == IPVERSION,
1355 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1356 }
1357 #endif
1358
1359 /*
1360 * Enable TSO and specify the size of the segments.
1361 * The TCP pseudo header checksum is always provided.
1362 */
1363 if (tso) {
1364 KASSERT(len > tp->t_maxseg - optlen,
1365 ("%s: len <= tso_segsz", __func__));
1366 m->m_pkthdr.csum_flags |= CSUM_TSO;
1367 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1368 }
1369
1370 KASSERT(len + hdrlen == m_length(m, NULL),
1371 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1372 __func__, len, hdrlen, m_length(m, NULL)));
1373
1374 #ifdef TCP_HHOOK
1375 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1376 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1377 #endif
1378
1379 #ifdef TCPDEBUG
1380 /*
1381 * Trace.
1382 */
1383 if (so->so_options & SO_DEBUG) {
1384 u_short save = 0;
1385 #ifdef INET6
1386 if (!isipv6)
1387 #endif
1388 {
1389 save = ipov->ih_len;
1390 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1391 }
1392 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1393 #ifdef INET6
1394 if (!isipv6)
1395 #endif
1396 ipov->ih_len = save;
1397 }
1398 #endif /* TCPDEBUG */
1399 TCP_PROBE3(debug__output, tp, th, m);
1400
1401 /* We're getting ready to send; log now. */
1402 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1403 len, NULL, false);
1404
1405 /*
1406 * Fill in IP length and desired time to live and
1407 * send to IP level. There should be a better way
1408 * to handle ttl and tos; we could keep them in
1409 * the template, but need a way to checksum without them.
1410 */
1411 /*
1412 * m->m_pkthdr.len should have been set before checksum calculation,
1413 * because in6_cksum() need it.
1414 */
1415 #ifdef INET6
1416 if (isipv6) {
1417 /*
1418 * we separately set hoplimit for every segment, since the
1419 * user might want to change the value via setsockopt.
1420 * Also, desired default hop limit might be changed via
1421 * Neighbor Discovery.
1422 */
1423 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1424
1425 /*
1426 * Set the packet size here for the benefit of DTrace probes.
1427 * ip6_output() will set it properly; it's supposed to include
1428 * the option header lengths as well.
1429 */
1430 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1431
1432 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1433 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1434 else
1435 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1436
1437 if (tp->t_state == TCPS_SYN_SENT)
1438 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1439
1440 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1441
1442 #ifdef TCPPCAP
1443 /* Save packet, if requested. */
1444 tcp_pcap_add(th, m, &(tp->t_outpkts));
1445 #endif
1446
1447 /* TODO: IPv6 IP6TOS_ECT bit on */
1448 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1449 &tp->t_inpcb->inp_route6,
1450 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1451 NULL, NULL, tp->t_inpcb);
1452
1453 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1454 mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1455 }
1456 #endif /* INET6 */
1457 #if defined(INET) && defined(INET6)
1458 else
1459 #endif
1460 #ifdef INET
1461 {
1462 ip->ip_len = htons(m->m_pkthdr.len);
1463 #ifdef INET6
1464 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1465 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1466 #endif /* INET6 */
1467 /*
1468 * If we do path MTU discovery, then we set DF on every packet.
1469 * This might not be the best thing to do according to RFC3390
1470 * Section 2. However the tcp hostcache migitates the problem
1471 * so it affects only the first tcp connection with a host.
1472 *
1473 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1474 */
1475 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1476 ip->ip_off |= htons(IP_DF);
1477 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1478 } else {
1479 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1480 }
1481
1482 if (tp->t_state == TCPS_SYN_SENT)
1483 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1484
1485 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1486
1487 #ifdef TCPPCAP
1488 /* Save packet, if requested. */
1489 tcp_pcap_add(th, m, &(tp->t_outpkts));
1490 #endif
1491
1492 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1493 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1494 tp->t_inpcb);
1495
1496 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1497 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1498 }
1499 #endif /* INET */
1500
1501 out:
1502 /*
1503 * In transmit state, time the transmission and arrange for
1504 * the retransmit. In persist state, just set snd_max.
1505 */
1506 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1507 !tcp_timer_active(tp, TT_PERSIST)) {
1508 tcp_seq startseq = tp->snd_nxt;
1509
1510 /*
1511 * Advance snd_nxt over sequence space of this segment.
1512 */
1513 if (flags & (TH_SYN|TH_FIN)) {
1514 if (flags & TH_SYN)
1515 tp->snd_nxt++;
1516 if (flags & TH_FIN) {
1517 tp->snd_nxt++;
1518 tp->t_flags |= TF_SENTFIN;
1519 }
1520 }
1521 if (sack_rxmit)
1522 goto timer;
1523 tp->snd_nxt += len;
1524 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1525 tp->snd_max = tp->snd_nxt;
1526 /*
1527 * Time this transmission if not a retransmission and
1528 * not currently timing anything.
1529 */
1530 tp->t_sndtime = ticks;
1531 if (tp->t_rtttime == 0) {
1532 tp->t_rtttime = ticks;
1533 tp->t_rtseq = startseq;
1534 TCPSTAT_INC(tcps_segstimed);
1535 }
1536 #ifdef STATS
1537 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1538 tp->t_flags |= TF_GPUTINPROG;
1539 tp->gput_seq = startseq;
1540 tp->gput_ack = startseq +
1541 ulmin(sbavail(&so->so_snd) - off, sendwin);
1542 tp->gput_ts = tcp_ts_getticks();
1543 }
1544 #endif /* STATS */
1545 }
1546
1547 /*
1548 * Set retransmit timer if not currently set,
1549 * and not doing a pure ack or a keep-alive probe.
1550 * Initial value for retransmit timer is smoothed
1551 * round-trip time + 2 * round-trip time variance.
1552 * Initialize shift counter which is used for backoff
1553 * of retransmit time.
1554 */
1555 timer:
1556 if (!tcp_timer_active(tp, TT_REXMT) &&
1557 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1558 (tp->snd_nxt != tp->snd_una))) {
1559 if (tcp_timer_active(tp, TT_PERSIST)) {
1560 tcp_timer_activate(tp, TT_PERSIST, 0);
1561 tp->t_rxtshift = 0;
1562 }
1563 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1564 } else if (len == 0 && sbavail(&so->so_snd) &&
1565 !tcp_timer_active(tp, TT_REXMT) &&
1566 !tcp_timer_active(tp, TT_PERSIST)) {
1567 /*
1568 * Avoid a situation where we do not set persist timer
1569 * after a zero window condition. For example:
1570 * 1) A -> B: packet with enough data to fill the window
1571 * 2) B -> A: ACK for #1 + new data (0 window
1572 * advertisement)
1573 * 3) A -> B: ACK for #2, 0 len packet
1574 *
1575 * In this case, A will not activate the persist timer,
1576 * because it chose to send a packet. Unless tcp_output
1577 * is called for some other reason (delayed ack timer,
1578 * another input packet from B, socket syscall), A will
1579 * not send zero window probes.
1580 *
1581 * So, if you send a 0-length packet, but there is data
1582 * in the socket buffer, and neither the rexmt or
1583 * persist timer is already set, then activate the
1584 * persist timer.
1585 */
1586 tp->t_rxtshift = 0;
1587 tcp_setpersist(tp);
1588 }
1589 } else {
1590 /*
1591 * Persist case, update snd_max but since we are in
1592 * persist mode (no window) we do not update snd_nxt.
1593 */
1594 int xlen = len;
1595 if (flags & TH_SYN)
1596 ++xlen;
1597 if (flags & TH_FIN) {
1598 ++xlen;
1599 tp->t_flags |= TF_SENTFIN;
1600 }
1601 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1602 tp->snd_max = tp->snd_nxt + xlen;
1603 }
1604 if ((error == 0) &&
1605 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1606 (tp->t_flags & TF_SACK_PERMIT) &&
1607 tp->rcv_numsacks > 0)) {
1608 /* Clean up any DSACK's sent */
1609 tcp_clean_dsack_blocks(tp);
1610 }
1611 if (error) {
1612 /* Record the error. */
1613 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1614 error, 0, NULL, false);
1615
1616 /*
1617 * We know that the packet was lost, so back out the
1618 * sequence number advance, if any.
1619 *
1620 * If the error is EPERM the packet got blocked by the
1621 * local firewall. Normally we should terminate the
1622 * connection but the blocking may have been spurious
1623 * due to a firewall reconfiguration cycle. So we treat
1624 * it like a packet loss and let the retransmit timer and
1625 * timeouts do their work over time.
1626 * XXX: It is a POLA question whether calling tcp_drop right
1627 * away would be the really correct behavior instead.
1628 */
1629 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1630 !tcp_timer_active(tp, TT_PERSIST)) &&
1631 ((flags & TH_SYN) == 0) &&
1632 (error != EPERM)) {
1633 if (sack_rxmit) {
1634 p->rxmit -= len;
1635 tp->sackhint.sack_bytes_rexmit -= len;
1636 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1637 ("sackhint bytes rtx >= 0"));
1638 } else
1639 tp->snd_nxt -= len;
1640 }
1641 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1642 switch (error) {
1643 case EACCES:
1644 case EPERM:
1645 tp->t_softerror = error;
1646 return (error);
1647 case ENOBUFS:
1648 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1649 tp->snd_cwnd = tp->t_maxseg;
1650 return (0);
1651 case EMSGSIZE:
1652 /*
1653 * For some reason the interface we used initially
1654 * to send segments changed to another or lowered
1655 * its MTU.
1656 * If TSO was active we either got an interface
1657 * without TSO capabilits or TSO was turned off.
1658 * If we obtained mtu from ip_output() then update
1659 * it and try again.
1660 */
1661 if (tso)
1662 tp->t_flags &= ~TF_TSO;
1663 if (mtu != 0) {
1664 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1665 goto again;
1666 }
1667 return (error);
1668 case EHOSTDOWN:
1669 case EHOSTUNREACH:
1670 case ENETDOWN:
1671 case ENETUNREACH:
1672 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1673 tp->t_softerror = error;
1674 return (0);
1675 }
1676 /* FALLTHROUGH */
1677 default:
1678 return (error);
1679 }
1680 }
1681 TCPSTAT_INC(tcps_sndtotal);
1682
1683 /*
1684 * Data sent (as far as we can tell).
1685 * If this advertises a larger window than any other segment,
1686 * then remember the size of the advertised window.
1687 * Any pending ACK has now been sent.
1688 */
1689 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1690 tp->rcv_adv = tp->rcv_nxt + recwin;
1691 tp->last_ack_sent = tp->rcv_nxt;
1692 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1693 if (tcp_timer_active(tp, TT_DELACK))
1694 tcp_timer_activate(tp, TT_DELACK, 0);
1695 #if 0
1696 /*
1697 * This completely breaks TCP if newreno is turned on. What happens
1698 * is that if delayed-acks are turned on on the receiver, this code
1699 * on the transmitter effectively destroys the TCP window, forcing
1700 * it to four packets (1.5Kx4 = 6K window).
1701 */
1702 if (sendalot && --maxburst)
1703 goto again;
1704 #endif
1705 if (sendalot)
1706 goto again;
1707 return (0);
1708 }
1709
1710 void
tcp_setpersist(struct tcpcb * tp)1711 tcp_setpersist(struct tcpcb *tp)
1712 {
1713 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1714 int tt;
1715
1716 tp->t_flags &= ~TF_PREVVALID;
1717 if (tcp_timer_active(tp, TT_REXMT))
1718 panic("tcp_setpersist: retransmit pending");
1719 /*
1720 * Start/restart persistence timer.
1721 */
1722 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1723 tcp_persmin, tcp_persmax);
1724 tcp_timer_activate(tp, TT_PERSIST, tt);
1725 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1726 tp->t_rxtshift++;
1727 }
1728
1729 /*
1730 * Insert TCP options according to the supplied parameters to the place
1731 * optp in a consistent way. Can handle unaligned destinations.
1732 *
1733 * The order of the option processing is crucial for optimal packing and
1734 * alignment for the scarce option space.
1735 *
1736 * The optimal order for a SYN/SYN-ACK segment is:
1737 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1738 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1739 *
1740 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1741 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1742 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1743 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1744 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1745 */
1746 int
tcp_addoptions(struct tcpopt * to,u_char * optp)1747 tcp_addoptions(struct tcpopt *to, u_char *optp)
1748 {
1749 u_int32_t mask, optlen = 0;
1750
1751 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1752 if ((to->to_flags & mask) != mask)
1753 continue;
1754 if (optlen == TCP_MAXOLEN)
1755 break;
1756 switch (to->to_flags & mask) {
1757 case TOF_MSS:
1758 while (optlen % 4) {
1759 optlen += TCPOLEN_NOP;
1760 *optp++ = TCPOPT_NOP;
1761 }
1762 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1763 continue;
1764 optlen += TCPOLEN_MAXSEG;
1765 *optp++ = TCPOPT_MAXSEG;
1766 *optp++ = TCPOLEN_MAXSEG;
1767 to->to_mss = htons(to->to_mss);
1768 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1769 optp += sizeof(to->to_mss);
1770 break;
1771 case TOF_SCALE:
1772 while (!optlen || optlen % 2 != 1) {
1773 optlen += TCPOLEN_NOP;
1774 *optp++ = TCPOPT_NOP;
1775 }
1776 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1777 continue;
1778 optlen += TCPOLEN_WINDOW;
1779 *optp++ = TCPOPT_WINDOW;
1780 *optp++ = TCPOLEN_WINDOW;
1781 *optp++ = to->to_wscale;
1782 break;
1783 case TOF_SACKPERM:
1784 while (optlen % 2) {
1785 optlen += TCPOLEN_NOP;
1786 *optp++ = TCPOPT_NOP;
1787 }
1788 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1789 continue;
1790 optlen += TCPOLEN_SACK_PERMITTED;
1791 *optp++ = TCPOPT_SACK_PERMITTED;
1792 *optp++ = TCPOLEN_SACK_PERMITTED;
1793 break;
1794 case TOF_TS:
1795 while (!optlen || optlen % 4 != 2) {
1796 optlen += TCPOLEN_NOP;
1797 *optp++ = TCPOPT_NOP;
1798 }
1799 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1800 continue;
1801 optlen += TCPOLEN_TIMESTAMP;
1802 *optp++ = TCPOPT_TIMESTAMP;
1803 *optp++ = TCPOLEN_TIMESTAMP;
1804 to->to_tsval = htonl(to->to_tsval);
1805 to->to_tsecr = htonl(to->to_tsecr);
1806 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1807 optp += sizeof(to->to_tsval);
1808 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1809 optp += sizeof(to->to_tsecr);
1810 break;
1811 case TOF_SIGNATURE:
1812 {
1813 int siglen = TCPOLEN_SIGNATURE - 2;
1814
1815 while (!optlen || optlen % 4 != 2) {
1816 optlen += TCPOLEN_NOP;
1817 *optp++ = TCPOPT_NOP;
1818 }
1819 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1820 to->to_flags &= ~TOF_SIGNATURE;
1821 continue;
1822 }
1823 optlen += TCPOLEN_SIGNATURE;
1824 *optp++ = TCPOPT_SIGNATURE;
1825 *optp++ = TCPOLEN_SIGNATURE;
1826 to->to_signature = optp;
1827 while (siglen--)
1828 *optp++ = 0;
1829 break;
1830 }
1831 case TOF_SACK:
1832 {
1833 int sackblks = 0;
1834 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1835 tcp_seq sack_seq;
1836
1837 while (!optlen || optlen % 4 != 2) {
1838 optlen += TCPOLEN_NOP;
1839 *optp++ = TCPOPT_NOP;
1840 }
1841 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1842 continue;
1843 optlen += TCPOLEN_SACKHDR;
1844 *optp++ = TCPOPT_SACK;
1845 sackblks = min(to->to_nsacks,
1846 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1847 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1848 while (sackblks--) {
1849 sack_seq = htonl(sack->start);
1850 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1851 optp += sizeof(sack_seq);
1852 sack_seq = htonl(sack->end);
1853 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1854 optp += sizeof(sack_seq);
1855 optlen += TCPOLEN_SACK;
1856 sack++;
1857 }
1858 TCPSTAT_INC(tcps_sack_send_blocks);
1859 break;
1860 }
1861 case TOF_FASTOPEN:
1862 {
1863 int total_len;
1864
1865 /* XXX is there any point to aligning this option? */
1866 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1867 if (TCP_MAXOLEN - optlen < total_len) {
1868 to->to_flags &= ~TOF_FASTOPEN;
1869 continue;
1870 }
1871 *optp++ = TCPOPT_FAST_OPEN;
1872 *optp++ = total_len;
1873 if (to->to_tfo_len > 0) {
1874 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1875 optp += to->to_tfo_len;
1876 }
1877 optlen += total_len;
1878 break;
1879 }
1880 default:
1881 panic("%s: unknown TCP option type", __func__);
1882 break;
1883 }
1884 }
1885
1886 /* Terminate and pad TCP options to a 4 byte boundary. */
1887 if (optlen % 4) {
1888 optlen += TCPOLEN_EOL;
1889 *optp++ = TCPOPT_EOL;
1890 }
1891 /*
1892 * According to RFC 793 (STD0007):
1893 * "The content of the header beyond the End-of-Option option
1894 * must be header padding (i.e., zero)."
1895 * and later: "The padding is composed of zeros."
1896 */
1897 while (optlen % 4) {
1898 optlen += TCPOLEN_PAD;
1899 *optp++ = TCPOPT_PAD;
1900 }
1901
1902 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1903 return (optlen);
1904 }
1905
1906 /*
1907 * This is a copy of m_copym(), taking the TSO segment size/limit
1908 * constraints into account, and advancing the sndptr as it goes.
1909 */
1910 struct mbuf *
tcp_m_copym(struct mbuf * m,int32_t off0,int32_t * plen,int32_t seglimit,int32_t segsize,struct sockbuf * sb,bool hw_tls)1911 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1912 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1913 {
1914 #ifdef KERN_TLS
1915 struct ktls_session *tls, *ntls;
1916 struct mbuf *start;
1917 #endif
1918 struct mbuf *n, **np;
1919 struct mbuf *top;
1920 int32_t off = off0;
1921 int32_t len = *plen;
1922 int32_t fragsize;
1923 int32_t len_cp = 0;
1924 int32_t *pkthdrlen;
1925 uint32_t mlen, frags;
1926 bool copyhdr;
1927
1928 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1929 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1930 if (off == 0 && m->m_flags & M_PKTHDR)
1931 copyhdr = true;
1932 else
1933 copyhdr = false;
1934 while (off > 0) {
1935 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1936 if (off < m->m_len)
1937 break;
1938 off -= m->m_len;
1939 if ((sb) && (m == sb->sb_sndptr)) {
1940 sb->sb_sndptroff += m->m_len;
1941 sb->sb_sndptr = m->m_next;
1942 }
1943 m = m->m_next;
1944 }
1945 np = ⊤
1946 top = NULL;
1947 pkthdrlen = NULL;
1948 #ifdef KERN_TLS
1949 if (hw_tls && (m->m_flags & M_EXTPG))
1950 tls = m->m_epg_tls;
1951 else
1952 tls = NULL;
1953 start = m;
1954 #endif
1955 while (len > 0) {
1956 if (m == NULL) {
1957 KASSERT(len == M_COPYALL,
1958 ("tcp_m_copym, length > size of mbuf chain"));
1959 *plen = len_cp;
1960 if (pkthdrlen != NULL)
1961 *pkthdrlen = len_cp;
1962 break;
1963 }
1964 #ifdef KERN_TLS
1965 if (hw_tls) {
1966 if (m->m_flags & M_EXTPG)
1967 ntls = m->m_epg_tls;
1968 else
1969 ntls = NULL;
1970
1971 /*
1972 * Avoid mixing TLS records with handshake
1973 * data or TLS records from different
1974 * sessions.
1975 */
1976 if (tls != ntls) {
1977 MPASS(m != start);
1978 *plen = len_cp;
1979 if (pkthdrlen != NULL)
1980 *pkthdrlen = len_cp;
1981 break;
1982 }
1983 }
1984 #endif
1985 mlen = min(len, m->m_len - off);
1986 if (seglimit) {
1987 /*
1988 * For M_EXTPG mbufs, add 3 segments
1989 * + 1 in case we are crossing page boundaries
1990 * + 2 in case the TLS hdr/trailer are used
1991 * It is cheaper to just add the segments
1992 * than it is to take the cache miss to look
1993 * at the mbuf ext_pgs state in detail.
1994 */
1995 if (m->m_flags & M_EXTPG) {
1996 fragsize = min(segsize, PAGE_SIZE);
1997 frags = 3;
1998 } else {
1999 fragsize = segsize;
2000 frags = 0;
2001 }
2002
2003 /* Break if we really can't fit anymore. */
2004 if ((frags + 1) >= seglimit) {
2005 *plen = len_cp;
2006 if (pkthdrlen != NULL)
2007 *pkthdrlen = len_cp;
2008 break;
2009 }
2010
2011 /*
2012 * Reduce size if you can't copy the whole
2013 * mbuf. If we can't copy the whole mbuf, also
2014 * adjust len so the loop will end after this
2015 * mbuf.
2016 */
2017 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2018 mlen = (seglimit - frags - 1) * fragsize;
2019 len = mlen;
2020 *plen = len_cp + len;
2021 if (pkthdrlen != NULL)
2022 *pkthdrlen = *plen;
2023 }
2024 frags += howmany(mlen, fragsize);
2025 if (frags == 0)
2026 frags++;
2027 seglimit -= frags;
2028 KASSERT(seglimit > 0,
2029 ("%s: seglimit went too low", __func__));
2030 }
2031 if (copyhdr)
2032 n = m_gethdr(M_NOWAIT, m->m_type);
2033 else
2034 n = m_get(M_NOWAIT, m->m_type);
2035 *np = n;
2036 if (n == NULL)
2037 goto nospace;
2038 if (copyhdr) {
2039 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2040 goto nospace;
2041 if (len == M_COPYALL)
2042 n->m_pkthdr.len -= off0;
2043 else
2044 n->m_pkthdr.len = len;
2045 pkthdrlen = &n->m_pkthdr.len;
2046 copyhdr = false;
2047 }
2048 n->m_len = mlen;
2049 len_cp += n->m_len;
2050 if (m->m_flags & (M_EXT|M_EXTPG)) {
2051 n->m_data = m->m_data + off;
2052 mb_dupcl(n, m);
2053 } else
2054 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2055 (u_int)n->m_len);
2056
2057 if (sb && (sb->sb_sndptr == m) &&
2058 ((n->m_len + off) >= m->m_len) && m->m_next) {
2059 sb->sb_sndptroff += m->m_len;
2060 sb->sb_sndptr = m->m_next;
2061 }
2062 off = 0;
2063 if (len != M_COPYALL) {
2064 len -= n->m_len;
2065 }
2066 m = m->m_next;
2067 np = &n->m_next;
2068 }
2069 return (top);
2070 nospace:
2071 m_freem(top);
2072 return (NULL);
2073 }
2074
2075 void
tcp_sndbuf_autoscale(struct tcpcb * tp,struct socket * so,uint32_t sendwin)2076 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2077 {
2078
2079 /*
2080 * Automatic sizing of send socket buffer. Often the send buffer
2081 * size is not optimally adjusted to the actual network conditions
2082 * at hand (delay bandwidth product). Setting the buffer size too
2083 * small limits throughput on links with high bandwidth and high
2084 * delay (eg. trans-continental/oceanic links). Setting the
2085 * buffer size too big consumes too much real kernel memory,
2086 * especially with many connections on busy servers.
2087 *
2088 * The criteria to step up the send buffer one notch are:
2089 * 1. receive window of remote host is larger than send buffer
2090 * (with a fudge factor of 5/4th);
2091 * 2. send buffer is filled to 7/8th with data (so we actually
2092 * have data to make use of it);
2093 * 3. send buffer fill has not hit maximal automatic size;
2094 * 4. our send window (slow start and cogestion controlled) is
2095 * larger than sent but unacknowledged data in send buffer.
2096 *
2097 * The remote host receive window scaling factor may limit the
2098 * growing of the send buffer before it reaches its allowed
2099 * maximum.
2100 *
2101 * It scales directly with slow start or congestion window
2102 * and does at most one step per received ACK. This fast
2103 * scaling has the drawback of growing the send buffer beyond
2104 * what is strictly necessary to make full use of a given
2105 * delay*bandwidth product. However testing has shown this not
2106 * to be much of an problem. At worst we are trading wasting
2107 * of available bandwidth (the non-use of it) for wasting some
2108 * socket buffer memory.
2109 *
2110 * TODO: Shrink send buffer during idle periods together
2111 * with congestion window. Requires another timer. Has to
2112 * wait for upcoming tcp timer rewrite.
2113 *
2114 * XXXGL: should there be used sbused() or sbavail()?
2115 */
2116 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2117 int lowat;
2118
2119 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2120 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2121 sbused(&so->so_snd) >=
2122 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2123 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2124 sendwin >= (sbused(&so->so_snd) -
2125 (tp->snd_nxt - tp->snd_una))) {
2126 if (!sbreserve_locked(&so->so_snd,
2127 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2128 V_tcp_autosndbuf_max), so, curthread))
2129 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2130 }
2131 }
2132 }
2133