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
32 #include <sys/cdefs.h>
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ipsec.h"
36 #include "opt_kern_tls.h"
37
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/arb.h>
41 #include <sys/domain.h>
42 #ifdef TCP_HHOOK
43 #include <sys/hhook.h>
44 #endif
45 #include <sys/kernel.h>
46 #ifdef KERN_TLS
47 #include <sys/ktls.h>
48 #endif
49 #include <sys/lock.h>
50 #include <sys/mbuf.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
53 #include <sys/qmath.h>
54 #include <sys/sdt.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/stats.h>
59
60 #include <net/if.h>
61 #include <net/route.h>
62 #include <net/route/nhop.h>
63 #include <net/vnet.h>
64
65 #include <netinet/in.h>
66 #include <netinet/in_kdtrace.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
69 #include <netinet/in_pcb.h>
70 #include <netinet/ip_var.h>
71 #include <netinet/ip_options.h>
72 #ifdef INET6
73 #include <netinet6/in6_pcb.h>
74 #include <netinet/ip6.h>
75 #include <netinet6/ip6_var.h>
76 #endif
77 #include <netinet/tcp.h>
78 #define TCPOUTFLAGS
79 #include <netinet/tcp_fsm.h>
80 #include <netinet/tcp_seq.h>
81 #include <netinet/tcp_var.h>
82 #include <netinet/tcp_log_buf.h>
83 #include <netinet/tcp_syncache.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcpip.h>
86 #include <netinet/cc/cc.h>
87 #include <netinet/tcp_fastopen.h>
88 #ifdef TCPPCAP
89 #include <netinet/tcp_pcap.h>
90 #endif
91 #ifdef TCP_OFFLOAD
92 #include <netinet/tcp_offload.h>
93 #endif
94 #include <netinet/tcp_ecn.h>
95
96 #include <netipsec/ipsec_support.h>
97
98 #include <netinet/udp.h>
99 #include <netinet/udp_var.h>
100 #include <machine/in_cksum.h>
101
102 #include <security/mac/mac_framework.h>
103
104 VNET_DEFINE(int, path_mtu_discovery) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
106 &VNET_NAME(path_mtu_discovery), 1,
107 "Enable Path MTU Discovery");
108
109 VNET_DEFINE(int, tcp_do_tso) = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
111 &VNET_NAME(tcp_do_tso), 0,
112 "Enable TCP Segmentation Offload");
113
114 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
115 #define V_tcp_sendspace VNET(tcp_sendspace)
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
117 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
118
119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_do_autosndbuf), 0,
122 "Enable automatic send buffer sizing");
123
124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
126 &VNET_NAME(tcp_autosndbuf_inc), 0,
127 "Incrementor step size of automatic send buffer");
128
129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
131 &VNET_NAME(tcp_autosndbuf_max), 0,
132 "Max size of automatic send buffer");
133
134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
135 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
137 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
138 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
139
140 /*
141 * Make sure that either retransmit or persist timer is set for SYN, FIN and
142 * non-ACK.
143 */
144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
145 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
146 tcp_timer_active((tp), TT_REXMT) || \
147 tcp_timer_active((tp), TT_PERSIST), \
148 ("neither rexmt nor persist timer is set"))
149
150 #ifdef TCP_HHOOK
151 /*
152 * Wrapper for the TCP established output helper hook.
153 */
154 void
hhook_run_tcp_est_out(struct tcpcb * tp,struct tcphdr * th,struct tcpopt * to,uint32_t len,int tso)155 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
156 struct tcpopt *to, uint32_t len, int tso)
157 {
158 struct tcp_hhook_data hhook_data;
159
160 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
161 hhook_data.tp = tp;
162 hhook_data.th = th;
163 hhook_data.to = to;
164 hhook_data.len = len;
165 hhook_data.tso = tso;
166
167 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
168 &tp->t_osd);
169 }
170 }
171 #endif
172
173 /*
174 * CC wrapper hook functions
175 */
176 void
cc_after_idle(struct tcpcb * tp)177 cc_after_idle(struct tcpcb *tp)
178 {
179 INP_WLOCK_ASSERT(tptoinpcb(tp));
180
181 if (CC_ALGO(tp)->after_idle != NULL)
182 CC_ALGO(tp)->after_idle(&tp->t_ccv);
183 }
184
185 /*
186 * Tcp output routine: figure out what should be sent and send it.
187 */
188 int
tcp_default_output(struct tcpcb * tp)189 tcp_default_output(struct tcpcb *tp)
190 {
191 struct socket *so = tptosocket(tp);
192 struct inpcb *inp = tptoinpcb(tp);
193 int32_t len;
194 uint32_t recwin, sendwin;
195 uint16_t flags;
196 int off, error = 0; /* Keep compiler happy */
197 u_int if_hw_tsomaxsegcount = 0;
198 u_int if_hw_tsomaxsegsize = 0;
199 struct mbuf *m;
200 struct ip *ip = NULL;
201 struct tcphdr *th;
202 u_char opt[TCP_MAXOLEN];
203 unsigned ipoptlen, optlen, hdrlen, ulen;
204 unsigned ipsec_optlen = 0;
205 int idle, sendalot, curticks;
206 int sack_rxmit, sack_bytes_rxmt;
207 struct sackhole *p;
208 int tso, mtu;
209 struct tcpopt to;
210 struct udphdr *udp = NULL;
211 struct tcp_log_buffer *lgb;
212 unsigned int wanted_cookie = 0;
213 unsigned int dont_sendalot = 0;
214 #if 0
215 int maxburst = TCP_MAXBURST;
216 #endif
217 #ifdef INET6
218 struct ip6_hdr *ip6 = NULL;
219 const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
220 #endif
221 #ifdef KERN_TLS
222 const bool hw_tls = tp->t_nic_ktls_xmit != 0;
223 #else
224 const bool hw_tls = false;
225 #endif
226
227 NET_EPOCH_ASSERT();
228 INP_WLOCK_ASSERT(inp);
229
230 #ifdef TCP_OFFLOAD
231 if (tp->t_flags & TF_TOE)
232 return (tcp_offload_output(tp));
233 #endif
234
235 /*
236 * For TFO connections in SYN_SENT or SYN_RECEIVED,
237 * only allow the initial SYN or SYN|ACK and those sent
238 * by the retransmit timer.
239 */
240 if ((tp->t_flags & TF_FASTOPEN) &&
241 ((tp->t_state == TCPS_SYN_SENT) ||
242 (tp->t_state == TCPS_SYN_RECEIVED)) &&
243 SEQ_GT(tp->snd_max, tp->snd_una) && /* SYN or SYN|ACK sent */
244 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
245 return (0);
246
247 /*
248 * Determine length of data that should be transmitted,
249 * and flags that will be used.
250 * If there is some data or critical controls (SYN, RST)
251 * to send, then transmit; otherwise, investigate further.
252 */
253 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
254 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
255 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
256 cc_after_idle(tp);
257 tp->t_flags &= ~TF_LASTIDLE;
258 if (idle) {
259 if (tp->t_flags & TF_MORETOCOME) {
260 tp->t_flags |= TF_LASTIDLE;
261 idle = 0;
262 }
263 }
264 again:
265 /*
266 * If we've recently taken a timeout, snd_max will be greater than
267 * snd_nxt. There may be SACK information that allows us to avoid
268 * resending already delivered data. Adjust snd_nxt accordingly.
269 */
270 if ((tp->t_flags & TF_SACK_PERMIT) &&
271 SEQ_LT(tp->snd_nxt, tp->snd_max))
272 tcp_sack_adjust(tp);
273 sendalot = 0;
274 tso = 0;
275 mtu = 0;
276 off = tp->snd_nxt - tp->snd_una;
277 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
278
279 flags = tcp_outflags[tp->t_state];
280 /*
281 * Send any SACK-generated retransmissions. If we're explicitly trying
282 * to send out new data (when sendalot is 1), bypass this function.
283 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
284 * we're replacing a (future) new transmission with a retransmission
285 * now, and we previously incremented snd_cwnd in tcp_input().
286 */
287 /*
288 * Still in sack recovery , reset rxmit flag to zero.
289 */
290 sack_rxmit = 0;
291 sack_bytes_rxmt = 0;
292 len = 0;
293 p = NULL;
294 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
295 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
296 uint32_t cwin;
297
298 cwin =
299 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
300 /* Do not retransmit SACK segments beyond snd_recover */
301 if (SEQ_GT(p->end, tp->snd_recover)) {
302 /*
303 * (At least) part of sack hole extends beyond
304 * snd_recover. Check to see if we can rexmit data
305 * for this hole.
306 */
307 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
308 /*
309 * Can't rexmit any more data for this hole.
310 * That data will be rexmitted in the next
311 * sack recovery episode, when snd_recover
312 * moves past p->rxmit.
313 */
314 p = NULL;
315 goto after_sack_rexmit;
316 } else {
317 /* Can rexmit part of the current hole */
318 len = ((int32_t)ulmin(cwin,
319 SEQ_SUB(tp->snd_recover, p->rxmit)));
320 }
321 } else {
322 len = ((int32_t)ulmin(cwin,
323 SEQ_SUB(p->end, p->rxmit)));
324 }
325 if (len > 0) {
326 off = SEQ_SUB(p->rxmit, tp->snd_una);
327 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
328 __func__, off));
329 sack_rxmit = 1;
330 sendalot = 1;
331 }
332 }
333 after_sack_rexmit:
334 /*
335 * Get standard flags, and add SYN or FIN if requested by 'hidden'
336 * state flags.
337 */
338 if (tp->t_flags & TF_NEEDFIN)
339 flags |= TH_FIN;
340 if (tp->t_flags & TF_NEEDSYN)
341 flags |= TH_SYN;
342
343 SOCKBUF_LOCK(&so->so_snd);
344 /*
345 * If in persist timeout with window of 0, send 1 byte.
346 * Otherwise, if window is small but nonzero
347 * and timer expired, we will send what we can
348 * and go to transmit state.
349 */
350 if (tp->t_flags & TF_FORCEDATA) {
351 if (sendwin == 0) {
352 /*
353 * If we still have some data to send, then
354 * clear the FIN bit. Usually this would
355 * happen below when it realizes that we
356 * aren't sending all the data. However,
357 * if we have exactly 1 byte of unsent data,
358 * then it won't clear the FIN bit below,
359 * and if we are in persist state, we wind
360 * up sending the packet without recording
361 * that we sent the FIN bit.
362 *
363 * We can't just blindly clear the FIN bit,
364 * because if we don't have any more data
365 * to send then the probe will be the FIN
366 * itself.
367 */
368 if (off < sbused(&so->so_snd))
369 flags &= ~TH_FIN;
370 sendwin = 1;
371 } else {
372 tcp_timer_activate(tp, TT_PERSIST, 0);
373 tp->t_rxtshift = 0;
374 }
375 }
376
377 /*
378 * If snd_nxt == snd_max and we have transmitted a FIN, the
379 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
380 * a negative length. This can also occur when TCP opens up
381 * its congestion window while receiving additional duplicate
382 * acks after fast-retransmit because TCP will reset snd_nxt
383 * to snd_max after the fast-retransmit.
384 *
385 * In the normal retransmit-FIN-only case, however, snd_nxt will
386 * be set to snd_una, the offset will be 0, and the length may
387 * wind up 0.
388 *
389 * If sack_rxmit is true we are retransmitting from the scoreboard
390 * in which case len is already set.
391 */
392 if (sack_rxmit == 0) {
393 if (sack_bytes_rxmt == 0) {
394 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
395 off);
396 } else {
397 int32_t cwin;
398
399 /*
400 * We are inside of a SACK recovery episode and are
401 * sending new data, having retransmitted all the
402 * data possible in the scoreboard.
403 */
404 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
405 off);
406 /*
407 * Don't remove this (len > 0) check !
408 * We explicitly check for len > 0 here (although it
409 * isn't really necessary), to work around a gcc
410 * optimization issue - to force gcc to compute
411 * len above. Without this check, the computation
412 * of len is bungled by the optimizer.
413 */
414 if (len > 0) {
415 cwin = tp->snd_cwnd - imax(0, (int32_t)
416 (tp->snd_nxt - tp->snd_recover)) -
417 sack_bytes_rxmt;
418 if (cwin < 0)
419 cwin = 0;
420 len = imin(len, cwin);
421 }
422 }
423 }
424
425 /*
426 * Lop off SYN bit if it has already been sent. However, if this
427 * is SYN-SENT state and if segment contains data and if we don't
428 * know that foreign host supports TAO, suppress sending segment.
429 */
430 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
431 if (tp->t_state != TCPS_SYN_RECEIVED)
432 flags &= ~TH_SYN;
433 /*
434 * When sending additional segments following a TFO SYN|ACK,
435 * do not include the SYN bit.
436 */
437 if ((tp->t_flags & TF_FASTOPEN) &&
438 (tp->t_state == TCPS_SYN_RECEIVED))
439 flags &= ~TH_SYN;
440 off--, len++;
441 }
442
443 /*
444 * Be careful not to send data and/or FIN on SYN segments.
445 * This measure is needed to prevent interoperability problems
446 * with not fully conformant TCP implementations.
447 */
448 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
449 len = 0;
450 flags &= ~TH_FIN;
451 }
452
453 /*
454 * On TFO sockets, ensure no data is sent in the following cases:
455 *
456 * - When retransmitting SYN|ACK on a passively-created socket
457 *
458 * - When retransmitting SYN on an actively created socket
459 *
460 * - When sending a zero-length cookie (cookie request) on an
461 * actively created socket
462 *
463 * - When the socket is in the CLOSED state (RST is being sent)
464 */
465 if ((tp->t_flags & TF_FASTOPEN) &&
466 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
467 ((tp->t_state == TCPS_SYN_SENT) &&
468 (tp->t_tfo_client_cookie_len == 0)) ||
469 (flags & TH_RST)))
470 len = 0;
471
472 /* Without fast-open there should never be data sent on a SYN. */
473 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
474 len = 0;
475 }
476
477 if (len <= 0) {
478 /*
479 * If FIN has been sent but not acked,
480 * but we haven't been called to retransmit,
481 * len will be < 0. Otherwise, window shrank
482 * after we sent into it. If window shrank to 0,
483 * cancel pending retransmit, pull snd_nxt back
484 * to (closed) window, and set the persist timer
485 * if it isn't already going. If the window didn't
486 * close completely, just wait for an ACK.
487 *
488 * We also do a general check here to ensure that
489 * we will set the persist timer when we have data
490 * to send, but a 0-byte window. This makes sure
491 * the persist timer is set even if the packet
492 * hits one of the "goto send" lines below.
493 */
494 len = 0;
495 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
496 (off < (int) sbavail(&so->so_snd)) &&
497 !tcp_timer_active(tp, TT_PERSIST)) {
498 tcp_timer_activate(tp, TT_REXMT, 0);
499 tp->t_rxtshift = 0;
500 tp->snd_nxt = tp->snd_una;
501 if (!tcp_timer_active(tp, TT_PERSIST))
502 tcp_setpersist(tp);
503 }
504 }
505
506 /* len will be >= 0 after this point. */
507 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
508
509 tcp_sndbuf_autoscale(tp, so, sendwin);
510
511 /*
512 * Decide if we can use TCP Segmentation Offloading (if supported by
513 * hardware).
514 *
515 * TSO may only be used if we are in a pure bulk sending state. The
516 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
517 * IP options prevent using TSO. With TSO the TCP header is the same
518 * (except for the sequence number) for all generated packets. This
519 * makes it impossible to transmit any options which vary per generated
520 * segment or packet.
521 *
522 * IPv4 handling has a clear separation of ip options and ip header
523 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
524 * the right thing below to provide length of just ip options and thus
525 * checking for ipoptlen is enough to decide if ip options are present.
526 */
527 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
528 /*
529 * Pre-calculate here as we save another lookup into the darknesses
530 * of IPsec that way and can actually decide if TSO is ok.
531 */
532 #ifdef INET6
533 if (isipv6 && IPSEC_ENABLED(ipv6))
534 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
535 #ifdef INET
536 else
537 #endif
538 #endif /* INET6 */
539 #ifdef INET
540 if (IPSEC_ENABLED(ipv4))
541 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
542 #endif /* INET */
543 #endif /* IPSEC */
544 #ifdef INET6
545 if (isipv6)
546 ipoptlen = ip6_optlen(inp);
547 else
548 #endif
549 if (inp->inp_options)
550 ipoptlen = inp->inp_options->m_len -
551 offsetof(struct ipoption, ipopt_list);
552 else
553 ipoptlen = 0;
554 ipoptlen += ipsec_optlen;
555
556 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
557 (tp->t_port == 0) &&
558 ((tp->t_flags & TF_SIGNATURE) == 0) &&
559 tp->rcv_numsacks == 0 && ((sack_rxmit == 0) || V_tcp_sack_tso) &&
560 (ipoptlen == 0 || (ipoptlen == ipsec_optlen &&
561 (tp->t_flags2 & TF2_IPSEC_TSO) != 0)) &&
562 !(flags & TH_SYN))
563 tso = 1;
564
565 if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
566 tp->snd_una + sbused(&so->so_snd))) {
567 flags &= ~TH_FIN;
568 }
569
570 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
571 (long)TCP_MAXWIN << tp->rcv_scale);
572
573 /*
574 * Sender silly window avoidance. We transmit under the following
575 * conditions when len is non-zero:
576 *
577 * - We have a full segment (or more with TSO)
578 * - This is the last buffer in a write()/send() and we are
579 * either idle or running NODELAY
580 * - we've timed out (e.g. persist timer)
581 * - we have more then 1/2 the maximum send window's worth of
582 * data (receiver may be limited the window size)
583 * - we need to retransmit
584 */
585 if (len) {
586 if (len >= tp->t_maxseg)
587 goto send;
588 /*
589 * As the TCP header options are now
590 * considered when setting up the initial
591 * window, we would not send the last segment
592 * if we skip considering the option length here.
593 * Note: this may not work when tcp headers change
594 * very dynamically in the future.
595 */
596 if ((((tp->t_flags & TF_SIGNATURE) ?
597 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
598 ((tp->t_flags & TF_RCVD_TSTMP) ?
599 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
600 len) >= tp->t_maxseg)
601 goto send;
602 /*
603 * NOTE! on localhost connections an 'ack' from the remote
604 * end may occur synchronously with the output and cause
605 * us to flush a buffer queued with moretocome. XXX
606 *
607 * note: the len + off check is almost certainly unnecessary.
608 */
609 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
610 (idle || (tp->t_flags & TF_NODELAY)) &&
611 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
612 (tp->t_flags & TF_NOPUSH) == 0) {
613 goto send;
614 }
615 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
616 goto send;
617 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
618 goto send;
619 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
620 goto send;
621 if (sack_rxmit)
622 goto send;
623 }
624
625 /*
626 * Sending of standalone window updates.
627 *
628 * Window updates are important when we close our window due to a
629 * full socket buffer and are opening it again after the application
630 * reads data from it. Once the window has opened again and the
631 * remote end starts to send again the ACK clock takes over and
632 * provides the most current window information.
633 *
634 * We must avoid the silly window syndrome whereas every read
635 * from the receive buffer, no matter how small, causes a window
636 * update to be sent. We also should avoid sending a flurry of
637 * window updates when the socket buffer had queued a lot of data
638 * and the application is doing small reads.
639 *
640 * Prevent a flurry of pointless window updates by only sending
641 * an update when we can increase the advertized window by more
642 * than 1/4th of the socket buffer capacity. When the buffer is
643 * getting full or is very small be more aggressive and send an
644 * update whenever we can increase by two mss sized segments.
645 * In all other situations the ACK's to new incoming data will
646 * carry further window increases.
647 *
648 * Don't send an independent window update if a delayed
649 * ACK is pending (it will get piggy-backed on it) or the
650 * remote side already has done a half-close and won't send
651 * more data.
652 */
653 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
654 !(tp->t_flags & TF_DELACK) &&
655 !TCPS_HAVERCVDFIN(tp->t_state)) {
656 /*
657 * "adv" is the amount we could increase the window,
658 * taking into account that we are limited by
659 * TCP_MAXWIN << tp->rcv_scale.
660 */
661 int32_t adv;
662 int oldwin;
663
664 adv = recwin;
665 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
666 oldwin = (tp->rcv_adv - tp->rcv_nxt);
667 if (adv > oldwin)
668 adv -= oldwin;
669 else
670 adv = 0;
671 } else
672 oldwin = 0;
673
674 /*
675 * If the new window size ends up being the same as or less
676 * than the old size when it is scaled, then don't force
677 * a window update.
678 */
679 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
680 goto dontupdate;
681
682 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
683 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
684 recwin <= (so->so_rcv.sb_hiwat / 8) ||
685 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
686 adv >= TCP_MAXWIN << tp->rcv_scale))
687 goto send;
688 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
689 goto send;
690 }
691 dontupdate:
692
693 /*
694 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
695 * is also a catch-all for the retransmit timer timeout case.
696 */
697 if (tp->t_flags & TF_ACKNOW)
698 goto send;
699 if ((flags & TH_RST) ||
700 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
701 goto send;
702 if (SEQ_GT(tp->snd_up, tp->snd_una))
703 goto send;
704 /*
705 * If our state indicates that FIN should be sent
706 * and we have not yet done so, then we need to send.
707 */
708 if (flags & TH_FIN &&
709 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
710 goto send;
711 /*
712 * In SACK, it is possible for tcp_output to fail to send a segment
713 * after the retransmission timer has been turned off. Make sure
714 * that the retransmission timer is set.
715 */
716 if ((tp->t_flags & TF_SACK_PERMIT) &&
717 SEQ_GT(tp->snd_max, tp->snd_una) &&
718 !tcp_timer_active(tp, TT_REXMT) &&
719 !tcp_timer_active(tp, TT_PERSIST)) {
720 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
721 goto just_return;
722 }
723 /*
724 * TCP window updates are not reliable, rather a polling protocol
725 * using ``persist'' packets is used to insure receipt of window
726 * updates. The three ``states'' for the output side are:
727 * idle not doing retransmits or persists
728 * persisting to move a small or zero window
729 * (re)transmitting and thereby not persisting
730 *
731 * tcp_timer_active(tp, TT_PERSIST)
732 * is true when we are in persist state.
733 * (tp->t_flags & TF_FORCEDATA)
734 * is set when we are called to send a persist packet.
735 * tcp_timer_active(tp, TT_REXMT)
736 * is set when we are retransmitting
737 * The output side is idle when both timers are zero.
738 *
739 * If send window is too small, there is data to transmit, and no
740 * retransmit or persist is pending, then go to persist state.
741 * If nothing happens soon, send when timer expires:
742 * if window is nonzero, transmit what we can,
743 * otherwise force out a byte.
744 */
745 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
746 !tcp_timer_active(tp, TT_PERSIST)) {
747 tp->t_rxtshift = 0;
748 tcp_setpersist(tp);
749 }
750
751 /*
752 * No reason to send a segment, just return.
753 */
754 just_return:
755 SOCKBUF_UNLOCK(&so->so_snd);
756 return (0);
757
758 send:
759 SOCKBUF_LOCK_ASSERT(&so->so_snd);
760 if (len > 0) {
761 if (len >= tp->t_maxseg)
762 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
763 else
764 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
765 }
766 /*
767 * Before ESTABLISHED, force sending of initial options
768 * unless TCP set not to do any options.
769 * NOTE: we assume that the IP/TCP header plus TCP options
770 * always fit in a single mbuf, leaving room for a maximum
771 * link header, i.e.
772 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
773 */
774 optlen = 0;
775 #ifdef INET6
776 if (isipv6)
777 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
778 else
779 #endif
780 hdrlen = sizeof (struct tcpiphdr);
781
782 if (flags & TH_SYN) {
783 tp->snd_nxt = tp->iss;
784 }
785
786 /*
787 * Compute options for segment.
788 * We only have to care about SYN and established connection
789 * segments. Options for SYN-ACK segments are handled in TCP
790 * syncache.
791 */
792 to.to_flags = 0;
793 if ((tp->t_flags & TF_NOOPT) == 0) {
794 /* Maximum segment size. */
795 if (flags & TH_SYN) {
796 to.to_mss = tcp_mssopt(&inp->inp_inc);
797 if (tp->t_port)
798 to.to_mss -= V_tcp_udp_tunneling_overhead;
799 to.to_flags |= TOF_MSS;
800
801 /*
802 * On SYN or SYN|ACK transmits on TFO connections,
803 * only include the TFO option if it is not a
804 * retransmit, as the presence of the TFO option may
805 * have caused the original SYN or SYN|ACK to have
806 * been dropped by a middlebox.
807 */
808 if ((tp->t_flags & TF_FASTOPEN) &&
809 (tp->t_rxtshift == 0)) {
810 if (tp->t_state == TCPS_SYN_RECEIVED) {
811 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
812 to.to_tfo_cookie =
813 (u_int8_t *)&tp->t_tfo_cookie.server;
814 to.to_flags |= TOF_FASTOPEN;
815 wanted_cookie = 1;
816 } else if (tp->t_state == TCPS_SYN_SENT) {
817 to.to_tfo_len =
818 tp->t_tfo_client_cookie_len;
819 to.to_tfo_cookie =
820 tp->t_tfo_cookie.client;
821 to.to_flags |= TOF_FASTOPEN;
822 wanted_cookie = 1;
823 /*
824 * If we wind up having more data to
825 * send with the SYN than can fit in
826 * one segment, don't send any more
827 * until the SYN|ACK comes back from
828 * the other end.
829 */
830 dont_sendalot = 1;
831 }
832 }
833 }
834 /* Window scaling. */
835 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
836 to.to_wscale = tp->request_r_scale;
837 to.to_flags |= TOF_SCALE;
838 }
839 /* Timestamps. */
840 if ((tp->t_flags & TF_RCVD_TSTMP) ||
841 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
842 curticks = tcp_ts_getticks();
843 to.to_tsval = curticks + tp->ts_offset;
844 to.to_tsecr = tp->ts_recent;
845 to.to_flags |= TOF_TS;
846 if (tp->t_rxtshift == 1)
847 tp->t_badrxtwin = curticks;
848 }
849
850 /* Set receive buffer autosizing timestamp. */
851 if (tp->rfbuf_ts == 0 &&
852 (so->so_rcv.sb_flags & SB_AUTOSIZE))
853 tp->rfbuf_ts = tcp_ts_getticks();
854
855 /* Selective ACK's. */
856 if (tp->t_flags & TF_SACK_PERMIT) {
857 if (flags & TH_SYN)
858 to.to_flags |= TOF_SACKPERM;
859 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
860 tp->rcv_numsacks > 0) {
861 to.to_flags |= TOF_SACK;
862 to.to_nsacks = tp->rcv_numsacks;
863 to.to_sacks = (u_char *)tp->sackblks;
864 }
865 }
866 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
867 /* TCP-MD5 (RFC2385). */
868 /*
869 * Check that TCP_MD5SIG is enabled in tcpcb to
870 * account the size needed to set this TCP option.
871 */
872 if (tp->t_flags & TF_SIGNATURE)
873 to.to_flags |= TOF_SIGNATURE;
874 #endif /* TCP_SIGNATURE */
875
876 /* Processing the options. */
877 hdrlen += optlen = tcp_addoptions(&to, opt);
878 /*
879 * If we wanted a TFO option to be added, but it was unable
880 * to fit, ensure no data is sent.
881 */
882 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
883 !(to.to_flags & TOF_FASTOPEN))
884 len = 0;
885 }
886 if (tp->t_port) {
887 if (V_tcp_udp_tunneling_port == 0) {
888 /* The port was removed?? */
889 SOCKBUF_UNLOCK(&so->so_snd);
890 return (EHOSTUNREACH);
891 }
892 hdrlen += sizeof(struct udphdr);
893 }
894 /*
895 * Adjust data length if insertion of options will
896 * bump the packet length beyond the t_maxseg length.
897 * Clear the FIN bit because we cut off the tail of
898 * the segment.
899 */
900 if (len + optlen + ipoptlen > tp->t_maxseg) {
901 flags &= ~TH_FIN;
902
903 if (tso) {
904 u_int if_hw_tsomax;
905 u_int moff;
906 int max_len;
907
908 /* extract TSO information */
909 if_hw_tsomax = tp->t_tsomax;
910 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
911 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
912
913 /*
914 * Limit a TSO burst to prevent it from
915 * overflowing or exceeding the maximum length
916 * allowed by the network interface:
917 */
918 KASSERT(ipoptlen == ipsec_optlen,
919 ("%s: TSO can't do IP options", __func__));
920
921 /*
922 * Check if we should limit by maximum payload
923 * length:
924 */
925 if (if_hw_tsomax != 0) {
926 /* compute maximum TSO length */
927 max_len = if_hw_tsomax - hdrlen -
928 ipsec_optlen - max_linkhdr;
929 if (max_len <= 0) {
930 len = 0;
931 } else if (len > max_len) {
932 sendalot = 1;
933 len = max_len;
934 }
935 }
936
937 /*
938 * Prevent the last segment from being
939 * fractional unless the send sockbuf can be
940 * emptied:
941 */
942 max_len = tp->t_maxseg - optlen - ipsec_optlen;
943 if (((uint32_t)off + (uint32_t)len) <
944 sbavail(&so->so_snd)) {
945 moff = len % max_len;
946 if (moff != 0) {
947 len -= moff;
948 sendalot = 1;
949 }
950 }
951
952 /*
953 * In case there are too many small fragments
954 * don't use TSO:
955 */
956 if (len <= max_len) {
957 len = max_len;
958 sendalot = 1;
959 tso = 0;
960 }
961
962 /*
963 * Send the FIN in a separate segment
964 * after the bulk sending is done.
965 * We don't trust the TSO implementations
966 * to clear the FIN flag on all but the
967 * last segment.
968 */
969 if (tp->t_flags & TF_NEEDFIN)
970 sendalot = 1;
971 } else {
972 if (optlen + ipoptlen >= tp->t_maxseg) {
973 /*
974 * Since we don't have enough space to put
975 * the IP header chain and the TCP header in
976 * one packet as required by RFC 7112, don't
977 * send it. Also ensure that at least one
978 * byte of the payload can be put into the
979 * TCP segment.
980 */
981 SOCKBUF_UNLOCK(&so->so_snd);
982 error = EMSGSIZE;
983 sack_rxmit = 0;
984 goto out;
985 }
986 len = tp->t_maxseg - optlen - ipoptlen;
987 sendalot = 1;
988 if (dont_sendalot)
989 sendalot = 0;
990 }
991 } else
992 tso = 0;
993
994 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
995 ("%s: len > IP_MAXPACKET", __func__));
996
997 /*#ifdef DIAGNOSTIC*/
998 #ifdef INET6
999 if (max_linkhdr + hdrlen > MCLBYTES)
1000 #else
1001 if (max_linkhdr + hdrlen > MHLEN)
1002 #endif
1003 panic("tcphdr too big");
1004 /*#endif*/
1005
1006 /*
1007 * This KASSERT is here to catch edge cases at a well defined place.
1008 * Before, those had triggered (random) panic conditions further down.
1009 */
1010 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1011
1012 /*
1013 * Grab a header mbuf, attaching a copy of data to
1014 * be transmitted, and initialize the header from
1015 * the template for sends on this connection.
1016 */
1017 if (len) {
1018 struct mbuf *mb;
1019 struct sockbuf *msb;
1020 u_int moff;
1021
1022 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1023 TCPSTAT_INC(tcps_sndprobe);
1024 #ifdef STATS
1025 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1026 stats_voi_update_abs_u32(tp->t_stats,
1027 VOI_TCP_RETXPB, len);
1028 else
1029 stats_voi_update_abs_u64(tp->t_stats,
1030 VOI_TCP_TXPB, len);
1031 #endif /* STATS */
1032 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1033 tp->t_sndrexmitpack++;
1034 TCPSTAT_INC(tcps_sndrexmitpack);
1035 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1036 if (sack_rxmit) {
1037 TCPSTAT_INC(tcps_sack_rexmits);
1038 if (tso) {
1039 TCPSTAT_INC(tcps_sack_rexmits_tso);
1040 }
1041 TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1042 }
1043 #ifdef STATS
1044 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1045 len);
1046 #endif /* STATS */
1047 } else {
1048 TCPSTAT_INC(tcps_sndpack);
1049 TCPSTAT_ADD(tcps_sndbyte, len);
1050 #ifdef STATS
1051 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1052 len);
1053 #endif /* STATS */
1054 }
1055 #ifdef INET6
1056 if (MHLEN < hdrlen + max_linkhdr)
1057 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1058 else
1059 #endif
1060 m = m_gethdr(M_NOWAIT, MT_DATA);
1061
1062 if (m == NULL) {
1063 SOCKBUF_UNLOCK(&so->so_snd);
1064 error = ENOBUFS;
1065 sack_rxmit = 0;
1066 goto out;
1067 }
1068
1069 m->m_data += max_linkhdr;
1070 m->m_len = hdrlen;
1071
1072 /*
1073 * Start the m_copy functions from the closest mbuf
1074 * to the offset in the socket buffer chain.
1075 */
1076 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1077 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1078 m_copydata(mb, moff, len,
1079 mtod(m, caddr_t) + hdrlen);
1080 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1081 sbsndptr_adv(&so->so_snd, mb, len);
1082 m->m_len += len;
1083 } else {
1084 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1085 msb = NULL;
1086 else
1087 msb = &so->so_snd;
1088 m->m_next = tcp_m_copym(mb, moff,
1089 &len, if_hw_tsomaxsegcount,
1090 if_hw_tsomaxsegsize, msb, hw_tls);
1091 if (len <= (tp->t_maxseg - optlen)) {
1092 /*
1093 * Must have ran out of mbufs for the copy
1094 * shorten it to no longer need tso. Lets
1095 * not put on sendalot since we are low on
1096 * mbufs.
1097 */
1098 tso = 0;
1099 }
1100 if (m->m_next == NULL) {
1101 SOCKBUF_UNLOCK(&so->so_snd);
1102 (void) m_free(m);
1103 error = ENOBUFS;
1104 sack_rxmit = 0;
1105 goto out;
1106 }
1107 }
1108
1109 /*
1110 * If we're sending everything we've got, set PUSH.
1111 * (This will keep happy those implementations which only
1112 * give data to the user when a buffer fills or
1113 * a PUSH comes in.)
1114 */
1115 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1116 !(flags & TH_SYN))
1117 flags |= TH_PUSH;
1118 SOCKBUF_UNLOCK(&so->so_snd);
1119 } else {
1120 SOCKBUF_UNLOCK(&so->so_snd);
1121 if (tp->t_flags & TF_ACKNOW)
1122 TCPSTAT_INC(tcps_sndacks);
1123 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1124 TCPSTAT_INC(tcps_sndctrl);
1125 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1126 TCPSTAT_INC(tcps_sndurg);
1127 else
1128 TCPSTAT_INC(tcps_sndwinup);
1129
1130 m = m_gethdr(M_NOWAIT, MT_DATA);
1131 if (m == NULL) {
1132 error = ENOBUFS;
1133 sack_rxmit = 0;
1134 goto out;
1135 }
1136 #ifdef INET6
1137 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1138 MHLEN >= hdrlen) {
1139 M_ALIGN(m, hdrlen);
1140 } else
1141 #endif
1142 m->m_data += max_linkhdr;
1143 m->m_len = hdrlen;
1144 }
1145 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1146 m->m_pkthdr.rcvif = (struct ifnet *)0;
1147 #ifdef MAC
1148 mac_inpcb_create_mbuf(inp, m);
1149 #endif
1150 #ifdef INET6
1151 if (isipv6) {
1152 ip6 = mtod(m, struct ip6_hdr *);
1153 if (tp->t_port) {
1154 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1155 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1156 udp->uh_dport = tp->t_port;
1157 ulen = hdrlen + len - sizeof(struct ip6_hdr);
1158 udp->uh_ulen = htons(ulen);
1159 th = (struct tcphdr *)(udp + 1);
1160 } else {
1161 th = (struct tcphdr *)(ip6 + 1);
1162 }
1163 tcpip_fillheaders(inp, tp->t_port, ip6, th);
1164 } else
1165 #endif /* INET6 */
1166 {
1167 ip = mtod(m, struct ip *);
1168 if (tp->t_port) {
1169 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1170 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1171 udp->uh_dport = tp->t_port;
1172 ulen = hdrlen + len - sizeof(struct ip);
1173 udp->uh_ulen = htons(ulen);
1174 th = (struct tcphdr *)(udp + 1);
1175 } else
1176 th = (struct tcphdr *)(ip + 1);
1177 tcpip_fillheaders(inp, tp->t_port, ip, th);
1178 }
1179
1180 /*
1181 * Fill in fields, remembering maximum advertised
1182 * window for use in delaying messages about window sizes.
1183 * If resending a FIN, be sure not to use a new sequence number.
1184 */
1185 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1186 tp->snd_nxt == tp->snd_max)
1187 tp->snd_nxt--;
1188 /*
1189 * If we are starting a connection, send ECN setup
1190 * SYN packet. If we are on a retransmit, we may
1191 * resend those bits a number of times as per
1192 * RFC 3168.
1193 */
1194 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1195 flags |= tcp_ecn_output_syn_sent(tp);
1196 }
1197 /* Also handle parallel SYN for ECN */
1198 if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1199 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1200 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1201 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1202 (tp->t_flags2 & TF2_ECN_SND_ECE))
1203 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1204 #ifdef INET6
1205 if (isipv6) {
1206 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1207 ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1208 }
1209 else
1210 #endif
1211 {
1212 ip->ip_tos &= ~IPTOS_ECN_MASK;
1213 ip->ip_tos |= ect;
1214 }
1215 }
1216
1217 /*
1218 * If we are doing retransmissions, then snd_nxt will
1219 * not reflect the first unsent octet. For ACK only
1220 * packets, we do not want the sequence number of the
1221 * retransmitted packet, we want the sequence number
1222 * of the next unsent octet. So, if there is no data
1223 * (and no SYN or FIN), use snd_max instead of snd_nxt
1224 * when filling in ti_seq. But if we are in persist
1225 * state, snd_max might reflect one byte beyond the
1226 * right edge of the window, so use snd_nxt in that
1227 * case, since we know we aren't doing a retransmission.
1228 * (retransmit and persist are mutually exclusive...)
1229 */
1230 if (sack_rxmit == 0) {
1231 if (len || (flags & (TH_SYN|TH_FIN)) ||
1232 tcp_timer_active(tp, TT_PERSIST))
1233 th->th_seq = htonl(tp->snd_nxt);
1234 else
1235 th->th_seq = htonl(tp->snd_max);
1236 } else {
1237 th->th_seq = htonl(p->rxmit);
1238 p->rxmit += len;
1239 /*
1240 * Lost Retransmission Detection
1241 * trigger resending of a (then
1242 * still existing) hole, when
1243 * fack acks recoverypoint.
1244 */
1245 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1246 p->rxmit = tp->snd_recover;
1247 tp->sackhint.sack_bytes_rexmit += len;
1248 }
1249 if (IN_RECOVERY(tp->t_flags)) {
1250 /*
1251 * Account all bytes transmitted while
1252 * IN_RECOVERY, simplifying PRR and
1253 * Lost Retransmit Detection
1254 */
1255 tp->sackhint.prr_out += len;
1256 }
1257 th->th_ack = htonl(tp->rcv_nxt);
1258 if (optlen) {
1259 bcopy(opt, th + 1, optlen);
1260 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1261 }
1262 tcp_set_flags(th, flags);
1263 /*
1264 * Calculate receive window. Don't shrink window,
1265 * but avoid silly window syndrome.
1266 * If a RST segment is sent, advertise a window of zero.
1267 */
1268 if (flags & TH_RST) {
1269 recwin = 0;
1270 } else {
1271 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1272 recwin < tp->t_maxseg)
1273 recwin = 0;
1274 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1275 recwin < (tp->rcv_adv - tp->rcv_nxt))
1276 recwin = (tp->rcv_adv - tp->rcv_nxt);
1277 }
1278 /*
1279 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1280 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1281 * case is handled in syncache.
1282 */
1283 if (flags & TH_SYN)
1284 th->th_win = htons((u_short)
1285 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1286 else {
1287 /* Avoid shrinking window with window scaling. */
1288 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1289 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1290 }
1291
1292 /*
1293 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1294 * a 0 window. This may cause the remote transmitter to stall. This
1295 * flag tells soreceive() to disable delayed acknowledgements when
1296 * draining the buffer. This can occur if the receiver is attempting
1297 * to read more data than can be buffered prior to transmitting on
1298 * the connection.
1299 */
1300 if (th->th_win == 0) {
1301 tp->t_sndzerowin++;
1302 tp->t_flags |= TF_RXWIN0SENT;
1303 } else
1304 tp->t_flags &= ~TF_RXWIN0SENT;
1305 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1306 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1307 th->th_flags |= TH_URG;
1308 } else
1309 /*
1310 * If no urgent pointer to send, then we pull
1311 * the urgent pointer to the left edge of the send window
1312 * so that it doesn't drift into the send window on sequence
1313 * number wraparound.
1314 */
1315 tp->snd_up = tp->snd_una; /* drag it along */
1316
1317 /*
1318 * Put TCP length in extended header, and then
1319 * checksum extended header and data.
1320 */
1321 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1322
1323 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1324 if (to.to_flags & TOF_SIGNATURE) {
1325 /*
1326 * Calculate MD5 signature and put it into the place
1327 * determined before.
1328 * NOTE: since TCP options buffer doesn't point into
1329 * mbuf's data, calculate offset and use it.
1330 */
1331 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1332 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1333 /*
1334 * Do not send segment if the calculation of MD5
1335 * digest has failed.
1336 */
1337 m_freem(m);
1338 goto out;
1339 }
1340 }
1341 #endif
1342 #ifdef INET6
1343 if (isipv6) {
1344 /*
1345 * There is no need to fill in ip6_plen right now.
1346 * It will be filled later by ip6_output.
1347 */
1348 if (tp->t_port) {
1349 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1350 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1351 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1352 th->th_sum = htons(0);
1353 UDPSTAT_INC(udps_opackets);
1354 } else {
1355 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1356 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1357 th->th_sum = in6_cksum_pseudo(ip6,
1358 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1359 0);
1360 }
1361 }
1362 #endif
1363 #if defined(INET6) && defined(INET)
1364 else
1365 #endif
1366 #ifdef INET
1367 {
1368 if (tp->t_port) {
1369 m->m_pkthdr.csum_flags = CSUM_UDP;
1370 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1371 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1372 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1373 th->th_sum = htons(0);
1374 UDPSTAT_INC(udps_opackets);
1375 } else {
1376 m->m_pkthdr.csum_flags = CSUM_TCP;
1377 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1378 th->th_sum = in_pseudo(ip->ip_src.s_addr,
1379 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1380 IPPROTO_TCP + len + optlen));
1381 }
1382
1383 /* IP version must be set here for ipv4/ipv6 checking later */
1384 KASSERT(ip->ip_v == IPVERSION,
1385 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1386 }
1387 #endif
1388
1389 /*
1390 * Enable TSO and specify the size of the segments.
1391 * The TCP pseudo header checksum is always provided.
1392 */
1393 if (tso) {
1394 KASSERT(len > tp->t_maxseg - optlen - ipsec_optlen,
1395 ("%s: len <= tso_segsz", __func__));
1396 m->m_pkthdr.csum_flags |= CSUM_TSO;
1397 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen - ipsec_optlen;
1398 }
1399
1400 KASSERT(len + hdrlen == m_length(m, NULL),
1401 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1402 __func__, len, hdrlen, m_length(m, NULL)));
1403
1404 #ifdef TCP_HHOOK
1405 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1406 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1407 #endif
1408
1409 TCP_PROBE3(debug__output, tp, th, m);
1410
1411 /* We're getting ready to send; log now. */
1412 /* XXXMT: We are not honoring verbose logging. */
1413
1414 if (tcp_bblogging_on(tp))
1415 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1416 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1417 NULL);
1418 else
1419 lgb = NULL;
1420
1421 /*
1422 * Fill in IP length and desired time to live and
1423 * send to IP level. There should be a better way
1424 * to handle ttl and tos; we could keep them in
1425 * the template, but need a way to checksum without them.
1426 */
1427 /*
1428 * m->m_pkthdr.len should have been set before checksum calculation,
1429 * because in6_cksum() need it.
1430 */
1431 #ifdef INET6
1432 if (isipv6) {
1433 /*
1434 * we separately set hoplimit for every segment, since the
1435 * user might want to change the value via setsockopt.
1436 * Also, desired default hop limit might be changed via
1437 * Neighbor Discovery.
1438 */
1439 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1440
1441 /*
1442 * Set the packet size here for the benefit of DTrace probes.
1443 * ip6_output() will set it properly; it's supposed to include
1444 * the option header lengths as well.
1445 */
1446 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1447
1448 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1449 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1450 else
1451 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1452
1453 if (tp->t_state == TCPS_SYN_SENT)
1454 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1455
1456 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1457
1458 #ifdef TCPPCAP
1459 /* Save packet, if requested. */
1460 tcp_pcap_add(th, m, &(tp->t_outpkts));
1461 #endif
1462
1463 /* TODO: IPv6 IP6TOS_ECT bit on */
1464 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1465 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1466 NULL, NULL, inp);
1467
1468 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1469 mtu = inp->inp_route6.ro_nh->nh_mtu;
1470 }
1471 #endif /* INET6 */
1472 #if defined(INET) && defined(INET6)
1473 else
1474 #endif
1475 #ifdef INET
1476 {
1477 ip->ip_len = htons(m->m_pkthdr.len);
1478 #ifdef INET6
1479 if (inp->inp_vflag & INP_IPV6PROTO)
1480 ip->ip_ttl = in6_selecthlim(inp, NULL);
1481 #endif /* INET6 */
1482 /*
1483 * If we do path MTU discovery, then we set DF on every packet.
1484 * This might not be the best thing to do according to RFC3390
1485 * Section 2. However the tcp hostcache migitates the problem
1486 * so it affects only the first tcp connection with a host.
1487 *
1488 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1489 */
1490 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1491 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1492 if (tp->t_port == 0 || len < V_tcp_minmss) {
1493 ip->ip_off |= htons(IP_DF);
1494 }
1495 } else {
1496 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1497 }
1498
1499 if (tp->t_state == TCPS_SYN_SENT)
1500 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1501
1502 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1503
1504 #ifdef TCPPCAP
1505 /* Save packet, if requested. */
1506 tcp_pcap_add(th, m, &(tp->t_outpkts));
1507 #endif
1508
1509 error = ip_output(m, inp->inp_options, &inp->inp_route,
1510 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1511
1512 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1513 mtu = inp->inp_route.ro_nh->nh_mtu;
1514 }
1515 #endif /* INET */
1516
1517 if (lgb != NULL) {
1518 lgb->tlb_errno = error;
1519 lgb = NULL;
1520 }
1521 out:
1522 if (error == 0)
1523 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1524 /*
1525 * In transmit state, time the transmission and arrange for
1526 * the retransmit. In persist state, just set snd_max. In a closed
1527 * state just return.
1528 */
1529 if (flags & TH_RST) {
1530 TCPSTAT_INC(tcps_sndtotal);
1531 return (0);
1532 } else if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1533 !tcp_timer_active(tp, TT_PERSIST)) {
1534 tcp_seq startseq = tp->snd_nxt;
1535
1536 /*
1537 * Advance snd_nxt over sequence space of this segment.
1538 */
1539 if (flags & (TH_SYN|TH_FIN)) {
1540 if (flags & TH_SYN)
1541 tp->snd_nxt++;
1542 if (flags & TH_FIN) {
1543 tp->snd_nxt++;
1544 tp->t_flags |= TF_SENTFIN;
1545 }
1546 }
1547 if (sack_rxmit)
1548 goto timer;
1549 tp->snd_nxt += len;
1550 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1551 /*
1552 * Update "made progress" indication if we just
1553 * added new data to an empty socket buffer.
1554 */
1555 if (tp->snd_una == tp->snd_max)
1556 tp->t_acktime = ticks;
1557 tp->snd_max = tp->snd_nxt;
1558 /*
1559 * Time this transmission if not a retransmission and
1560 * not currently timing anything.
1561 */
1562 tp->t_sndtime = ticks;
1563 if (tp->t_rtttime == 0) {
1564 tp->t_rtttime = ticks;
1565 tp->t_rtseq = startseq;
1566 TCPSTAT_INC(tcps_segstimed);
1567 }
1568 #ifdef STATS
1569 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1570 tp->t_flags |= TF_GPUTINPROG;
1571 tp->gput_seq = startseq;
1572 tp->gput_ack = startseq +
1573 ulmin(sbavail(&so->so_snd) - off, sendwin);
1574 tp->gput_ts = tcp_ts_getticks();
1575 }
1576 #endif /* STATS */
1577 }
1578
1579 /*
1580 * Set retransmit timer if not currently set,
1581 * and not doing a pure ack or a keep-alive probe.
1582 * Initial value for retransmit timer is smoothed
1583 * round-trip time + 2 * round-trip time variance.
1584 * Initialize shift counter which is used for backoff
1585 * of retransmit time.
1586 */
1587 timer:
1588 if (!tcp_timer_active(tp, TT_REXMT) &&
1589 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1590 (tp->snd_nxt != tp->snd_una))) {
1591 if (tcp_timer_active(tp, TT_PERSIST)) {
1592 tcp_timer_activate(tp, TT_PERSIST, 0);
1593 tp->t_rxtshift = 0;
1594 }
1595 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1596 } else if (len == 0 && sbavail(&so->so_snd) &&
1597 !tcp_timer_active(tp, TT_REXMT) &&
1598 !tcp_timer_active(tp, TT_PERSIST)) {
1599 /*
1600 * Avoid a situation where we do not set persist timer
1601 * after a zero window condition. For example:
1602 * 1) A -> B: packet with enough data to fill the window
1603 * 2) B -> A: ACK for #1 + new data (0 window
1604 * advertisement)
1605 * 3) A -> B: ACK for #2, 0 len packet
1606 *
1607 * In this case, A will not activate the persist timer,
1608 * because it chose to send a packet. Unless tcp_output
1609 * is called for some other reason (delayed ack timer,
1610 * another input packet from B, socket syscall), A will
1611 * not send zero window probes.
1612 *
1613 * So, if you send a 0-length packet, but there is data
1614 * in the socket buffer, and neither the rexmt or
1615 * persist timer is already set, then activate the
1616 * persist timer.
1617 */
1618 tp->t_rxtshift = 0;
1619 tcp_setpersist(tp);
1620 }
1621 } else {
1622 /*
1623 * Persist case, update snd_max but since we are in
1624 * persist mode (no window) we do not update snd_nxt.
1625 */
1626 int xlen = len;
1627 if (flags & TH_SYN)
1628 ++xlen;
1629 if (flags & TH_FIN) {
1630 ++xlen;
1631 tp->t_flags |= TF_SENTFIN;
1632 }
1633 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1634 tp->snd_max = tp->snd_nxt + xlen;
1635 }
1636 if ((error == 0) &&
1637 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1638 (tp->t_flags & TF_SACK_PERMIT) &&
1639 tp->rcv_numsacks > 0)) {
1640 /* Clean up any DSACK's sent */
1641 tcp_clean_dsack_blocks(tp);
1642 }
1643 if (error) {
1644 /*
1645 * We know that the packet was lost, so back out the
1646 * sequence number advance, if any.
1647 *
1648 * If the error is EPERM the packet got blocked by the
1649 * local firewall. Normally we should terminate the
1650 * connection but the blocking may have been spurious
1651 * due to a firewall reconfiguration cycle. So we treat
1652 * it like a packet loss and let the retransmit timer and
1653 * timeouts do their work over time.
1654 * XXX: It is a POLA question whether calling tcp_drop right
1655 * away would be the really correct behavior instead.
1656 */
1657 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1658 !tcp_timer_active(tp, TT_PERSIST)) &&
1659 ((flags & TH_SYN) == 0) &&
1660 (error != EPERM)) {
1661 if (sack_rxmit) {
1662 p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1663 tp->sackhint.sack_bytes_rexmit -= len;
1664 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1665 ("sackhint bytes rtx >= 0"));
1666 KASSERT((flags & TH_FIN) == 0,
1667 ("error while FIN with SACK rxmit"));
1668 } else {
1669 tp->snd_nxt -= len;
1670 if (flags & TH_FIN)
1671 tp->snd_nxt--;
1672 }
1673 if (IN_RECOVERY(tp->t_flags))
1674 tp->sackhint.prr_out -= len;
1675 }
1676 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1677 switch (error) {
1678 case EACCES:
1679 case EPERM:
1680 tp->t_softerror = error;
1681 return (error);
1682 case ENOBUFS:
1683 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1684 tp->snd_cwnd = tp->t_maxseg;
1685 return (0);
1686 case EMSGSIZE:
1687 /*
1688 * For some reason the interface we used initially
1689 * to send segments changed to another or lowered
1690 * its MTU.
1691 * If TSO was active we either got an interface
1692 * without TSO capabilits or TSO was turned off.
1693 * If we obtained mtu from ip_output() then update
1694 * it and try again.
1695 */
1696 if (tso)
1697 tp->t_flags &= ~TF_TSO;
1698 if (mtu != 0) {
1699 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1700 goto again;
1701 }
1702 return (error);
1703 case EHOSTDOWN:
1704 case EHOSTUNREACH:
1705 case ENETDOWN:
1706 case ENETUNREACH:
1707 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1708 tp->t_softerror = error;
1709 return (0);
1710 }
1711 /* FALLTHROUGH */
1712 default:
1713 return (error);
1714 }
1715 }
1716 TCPSTAT_INC(tcps_sndtotal);
1717
1718 /*
1719 * Data sent (as far as we can tell).
1720 * If this advertises a larger window than any other segment,
1721 * then remember the size of the advertised window.
1722 * Any pending ACK has now been sent.
1723 */
1724 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1725 tp->rcv_adv = tp->rcv_nxt + recwin;
1726 tp->last_ack_sent = tp->rcv_nxt;
1727 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1728 if (tcp_timer_active(tp, TT_DELACK))
1729 tcp_timer_activate(tp, TT_DELACK, 0);
1730 if (sendalot)
1731 goto again;
1732 return (0);
1733 }
1734
1735 void
tcp_setpersist(struct tcpcb * tp)1736 tcp_setpersist(struct tcpcb *tp)
1737 {
1738 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1739 int tt;
1740 int maxunacktime;
1741
1742 tp->t_flags &= ~TF_PREVVALID;
1743 if (tcp_timer_active(tp, TT_REXMT))
1744 panic("tcp_setpersist: retransmit pending");
1745 /*
1746 * If the state is already closed, don't bother.
1747 */
1748 if (tp->t_state == TCPS_CLOSED)
1749 return;
1750
1751 /*
1752 * Start/restart persistence timer.
1753 */
1754 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1755 tcp_persmin, tcp_persmax);
1756 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1757 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1758 if (maxunacktime < 1)
1759 maxunacktime = 1;
1760 if (maxunacktime < tt)
1761 tt = maxunacktime;
1762 }
1763 tcp_timer_activate(tp, TT_PERSIST, tt);
1764 if (tp->t_rxtshift < V_tcp_retries)
1765 tp->t_rxtshift++;
1766 }
1767
1768 /*
1769 * Insert TCP options according to the supplied parameters to the place
1770 * optp in a consistent way. Can handle unaligned destinations.
1771 *
1772 * The order of the option processing is crucial for optimal packing and
1773 * alignment for the scarce option space.
1774 *
1775 * The optimal order for a SYN/SYN-ACK segment is:
1776 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1777 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1778 *
1779 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1780 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1781 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1782 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1783 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1784 */
1785 int
tcp_addoptions(struct tcpopt * to,u_char * optp)1786 tcp_addoptions(struct tcpopt *to, u_char *optp)
1787 {
1788 u_int32_t mask, optlen = 0;
1789
1790 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1791 if ((to->to_flags & mask) != mask)
1792 continue;
1793 if (optlen == TCP_MAXOLEN)
1794 break;
1795 switch (to->to_flags & mask) {
1796 case TOF_MSS:
1797 while (optlen % 4) {
1798 optlen += TCPOLEN_NOP;
1799 *optp++ = TCPOPT_NOP;
1800 }
1801 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1802 continue;
1803 optlen += TCPOLEN_MAXSEG;
1804 *optp++ = TCPOPT_MAXSEG;
1805 *optp++ = TCPOLEN_MAXSEG;
1806 to->to_mss = htons(to->to_mss);
1807 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1808 optp += sizeof(to->to_mss);
1809 break;
1810 case TOF_SCALE:
1811 while (!optlen || optlen % 2 != 1) {
1812 optlen += TCPOLEN_NOP;
1813 *optp++ = TCPOPT_NOP;
1814 }
1815 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1816 continue;
1817 optlen += TCPOLEN_WINDOW;
1818 *optp++ = TCPOPT_WINDOW;
1819 *optp++ = TCPOLEN_WINDOW;
1820 *optp++ = to->to_wscale;
1821 break;
1822 case TOF_SACKPERM:
1823 while (optlen % 2) {
1824 optlen += TCPOLEN_NOP;
1825 *optp++ = TCPOPT_NOP;
1826 }
1827 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1828 continue;
1829 optlen += TCPOLEN_SACK_PERMITTED;
1830 *optp++ = TCPOPT_SACK_PERMITTED;
1831 *optp++ = TCPOLEN_SACK_PERMITTED;
1832 break;
1833 case TOF_TS:
1834 while (!optlen || optlen % 4 != 2) {
1835 optlen += TCPOLEN_NOP;
1836 *optp++ = TCPOPT_NOP;
1837 }
1838 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1839 continue;
1840 optlen += TCPOLEN_TIMESTAMP;
1841 *optp++ = TCPOPT_TIMESTAMP;
1842 *optp++ = TCPOLEN_TIMESTAMP;
1843 to->to_tsval = htonl(to->to_tsval);
1844 to->to_tsecr = htonl(to->to_tsecr);
1845 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1846 optp += sizeof(to->to_tsval);
1847 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1848 optp += sizeof(to->to_tsecr);
1849 break;
1850 case TOF_SIGNATURE:
1851 {
1852 int siglen = TCPOLEN_SIGNATURE - 2;
1853
1854 while (!optlen || optlen % 4 != 2) {
1855 optlen += TCPOLEN_NOP;
1856 *optp++ = TCPOPT_NOP;
1857 }
1858 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1859 to->to_flags &= ~TOF_SIGNATURE;
1860 continue;
1861 }
1862 optlen += TCPOLEN_SIGNATURE;
1863 *optp++ = TCPOPT_SIGNATURE;
1864 *optp++ = TCPOLEN_SIGNATURE;
1865 to->to_signature = optp;
1866 while (siglen--)
1867 *optp++ = 0;
1868 break;
1869 }
1870 case TOF_SACK:
1871 {
1872 int sackblks = 0;
1873 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1874 tcp_seq sack_seq;
1875
1876 while (!optlen || optlen % 4 != 2) {
1877 optlen += TCPOLEN_NOP;
1878 *optp++ = TCPOPT_NOP;
1879 }
1880 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1881 continue;
1882 optlen += TCPOLEN_SACKHDR;
1883 *optp++ = TCPOPT_SACK;
1884 sackblks = min(to->to_nsacks,
1885 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1886 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1887 while (sackblks--) {
1888 sack_seq = htonl(sack->start);
1889 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1890 optp += sizeof(sack_seq);
1891 sack_seq = htonl(sack->end);
1892 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1893 optp += sizeof(sack_seq);
1894 optlen += TCPOLEN_SACK;
1895 sack++;
1896 }
1897 TCPSTAT_INC(tcps_sack_send_blocks);
1898 break;
1899 }
1900 case TOF_FASTOPEN:
1901 {
1902 int total_len;
1903
1904 /* XXX is there any point to aligning this option? */
1905 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1906 if (TCP_MAXOLEN - optlen < total_len) {
1907 to->to_flags &= ~TOF_FASTOPEN;
1908 continue;
1909 }
1910 *optp++ = TCPOPT_FAST_OPEN;
1911 *optp++ = total_len;
1912 if (to->to_tfo_len > 0) {
1913 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1914 optp += to->to_tfo_len;
1915 }
1916 optlen += total_len;
1917 break;
1918 }
1919 default:
1920 panic("%s: unknown TCP option type", __func__);
1921 break;
1922 }
1923 }
1924
1925 /* Terminate and pad TCP options to a 4 byte boundary. */
1926 if (optlen % 4) {
1927 optlen += TCPOLEN_EOL;
1928 *optp++ = TCPOPT_EOL;
1929 }
1930 /*
1931 * According to RFC 793 (STD0007):
1932 * "The content of the header beyond the End-of-Option option
1933 * must be header padding (i.e., zero)."
1934 * and later: "The padding is composed of zeros."
1935 */
1936 while (optlen % 4) {
1937 optlen += TCPOLEN_PAD;
1938 *optp++ = TCPOPT_PAD;
1939 }
1940
1941 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1942 return (optlen);
1943 }
1944
1945 /*
1946 * This is a copy of m_copym(), taking the TSO segment size/limit
1947 * constraints into account, and advancing the sndptr as it goes.
1948 */
1949 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)1950 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1951 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1952 {
1953 #ifdef KERN_TLS
1954 struct ktls_session *tls, *ntls;
1955 struct mbuf *start __diagused;
1956 #endif
1957 struct mbuf *n, **np;
1958 struct mbuf *top;
1959 int32_t off = off0;
1960 int32_t len = *plen;
1961 int32_t fragsize;
1962 int32_t len_cp = 0;
1963 int32_t *pkthdrlen;
1964 uint32_t mlen, frags;
1965 bool copyhdr;
1966
1967 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1968 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1969 if (off == 0 && m->m_flags & M_PKTHDR)
1970 copyhdr = true;
1971 else
1972 copyhdr = false;
1973 while (off > 0) {
1974 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1975 if (off < m->m_len)
1976 break;
1977 off -= m->m_len;
1978 if ((sb) && (m == sb->sb_sndptr)) {
1979 sb->sb_sndptroff += m->m_len;
1980 sb->sb_sndptr = m->m_next;
1981 }
1982 m = m->m_next;
1983 }
1984 np = ⊤
1985 top = NULL;
1986 pkthdrlen = NULL;
1987 #ifdef KERN_TLS
1988 if (hw_tls && (m->m_flags & M_EXTPG))
1989 tls = m->m_epg_tls;
1990 else
1991 tls = NULL;
1992 start = m;
1993 #endif
1994 while (len > 0) {
1995 if (m == NULL) {
1996 KASSERT(len == M_COPYALL,
1997 ("tcp_m_copym, length > size of mbuf chain"));
1998 *plen = len_cp;
1999 if (pkthdrlen != NULL)
2000 *pkthdrlen = len_cp;
2001 break;
2002 }
2003 #ifdef KERN_TLS
2004 if (hw_tls) {
2005 if (m->m_flags & M_EXTPG)
2006 ntls = m->m_epg_tls;
2007 else
2008 ntls = NULL;
2009
2010 /*
2011 * Avoid mixing TLS records with handshake
2012 * data or TLS records from different
2013 * sessions.
2014 */
2015 if (tls != ntls) {
2016 MPASS(m != start);
2017 *plen = len_cp;
2018 if (pkthdrlen != NULL)
2019 *pkthdrlen = len_cp;
2020 break;
2021 }
2022 }
2023 #endif
2024 mlen = min(len, m->m_len - off);
2025 if (seglimit) {
2026 /*
2027 * For M_EXTPG mbufs, add 3 segments
2028 * + 1 in case we are crossing page boundaries
2029 * + 2 in case the TLS hdr/trailer are used
2030 * It is cheaper to just add the segments
2031 * than it is to take the cache miss to look
2032 * at the mbuf ext_pgs state in detail.
2033 */
2034 if (m->m_flags & M_EXTPG) {
2035 fragsize = min(segsize, PAGE_SIZE);
2036 frags = 3;
2037 } else {
2038 fragsize = segsize;
2039 frags = 0;
2040 }
2041
2042 /* Break if we really can't fit anymore. */
2043 if ((frags + 1) >= seglimit) {
2044 *plen = len_cp;
2045 if (pkthdrlen != NULL)
2046 *pkthdrlen = len_cp;
2047 break;
2048 }
2049
2050 /*
2051 * Reduce size if you can't copy the whole
2052 * mbuf. If we can't copy the whole mbuf, also
2053 * adjust len so the loop will end after this
2054 * mbuf.
2055 */
2056 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2057 mlen = (seglimit - frags - 1) * fragsize;
2058 len = mlen;
2059 *plen = len_cp + len;
2060 if (pkthdrlen != NULL)
2061 *pkthdrlen = *plen;
2062 }
2063 frags += howmany(mlen, fragsize);
2064 if (frags == 0)
2065 frags++;
2066 seglimit -= frags;
2067 KASSERT(seglimit > 0,
2068 ("%s: seglimit went too low", __func__));
2069 }
2070 if (copyhdr)
2071 n = m_gethdr(M_NOWAIT, m->m_type);
2072 else
2073 n = m_get(M_NOWAIT, m->m_type);
2074 *np = n;
2075 if (n == NULL)
2076 goto nospace;
2077 if (copyhdr) {
2078 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2079 goto nospace;
2080 if (len == M_COPYALL)
2081 n->m_pkthdr.len -= off0;
2082 else
2083 n->m_pkthdr.len = len;
2084 pkthdrlen = &n->m_pkthdr.len;
2085 copyhdr = false;
2086 }
2087 n->m_len = mlen;
2088 len_cp += n->m_len;
2089 if (m->m_flags & (M_EXT | M_EXTPG)) {
2090 n->m_data = m->m_data + off;
2091 mb_dupcl(n, m);
2092 } else
2093 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2094 (u_int)n->m_len);
2095
2096 if (sb && (sb->sb_sndptr == m) &&
2097 ((n->m_len + off) >= m->m_len) && m->m_next) {
2098 sb->sb_sndptroff += m->m_len;
2099 sb->sb_sndptr = m->m_next;
2100 }
2101 off = 0;
2102 if (len != M_COPYALL) {
2103 len -= n->m_len;
2104 }
2105 m = m->m_next;
2106 np = &n->m_next;
2107 }
2108 return (top);
2109 nospace:
2110 m_freem(top);
2111 return (NULL);
2112 }
2113
2114 void
tcp_sndbuf_autoscale(struct tcpcb * tp,struct socket * so,uint32_t sendwin)2115 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2116 {
2117
2118 /*
2119 * Automatic sizing of send socket buffer. Often the send buffer
2120 * size is not optimally adjusted to the actual network conditions
2121 * at hand (delay bandwidth product). Setting the buffer size too
2122 * small limits throughput on links with high bandwidth and high
2123 * delay (eg. trans-continental/oceanic links). Setting the
2124 * buffer size too big consumes too much real kernel memory,
2125 * especially with many connections on busy servers.
2126 *
2127 * The criteria to step up the send buffer one notch are:
2128 * 1. receive window of remote host is larger than send buffer
2129 * (with a fudge factor of 5/4th);
2130 * 2. send buffer is filled to 7/8th with data (so we actually
2131 * have data to make use of it);
2132 * 3. send buffer fill has not hit maximal automatic size;
2133 * 4. our send window (slow start and cogestion controlled) is
2134 * larger than sent but unacknowledged data in send buffer.
2135 *
2136 * The remote host receive window scaling factor may limit the
2137 * growing of the send buffer before it reaches its allowed
2138 * maximum.
2139 *
2140 * It scales directly with slow start or congestion window
2141 * and does at most one step per received ACK. This fast
2142 * scaling has the drawback of growing the send buffer beyond
2143 * what is strictly necessary to make full use of a given
2144 * delay*bandwidth product. However testing has shown this not
2145 * to be much of an problem. At worst we are trading wasting
2146 * of available bandwidth (the non-use of it) for wasting some
2147 * socket buffer memory.
2148 *
2149 * TODO: Shrink send buffer during idle periods together
2150 * with congestion window. Requires another timer. Has to
2151 * wait for upcoming tcp timer rewrite.
2152 *
2153 * XXXGL: should there be used sbused() or sbavail()?
2154 */
2155 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2156 int lowat;
2157
2158 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2159 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2160 sbused(&so->so_snd) >=
2161 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2162 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2163 sendwin >= (sbused(&so->so_snd) -
2164 (tp->snd_nxt - tp->snd_una))) {
2165 if (!sbreserve_locked(so, SO_SND,
2166 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2167 V_tcp_autosndbuf_max), curthread))
2168 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2169 }
2170 }
2171 }
2172