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 #ifdef INET6
215 struct ip6_hdr *ip6 = NULL;
216 const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
217 #endif
218 #ifdef KERN_TLS
219 const bool hw_tls = tp->t_nic_ktls_xmit != 0;
220 #else
221 const bool hw_tls = false;
222 #endif
223
224 NET_EPOCH_ASSERT();
225 INP_WLOCK_ASSERT(inp);
226
227 #ifdef TCP_OFFLOAD
228 if (tp->t_flags & TF_TOE)
229 return (tcp_offload_output(tp));
230 #endif
231
232 /*
233 * For TFO connections in SYN_SENT or SYN_RECEIVED,
234 * only allow the initial SYN or SYN|ACK and those sent
235 * by the retransmit timer.
236 */
237 if ((tp->t_flags & TF_FASTOPEN) &&
238 ((tp->t_state == TCPS_SYN_SENT) ||
239 (tp->t_state == TCPS_SYN_RECEIVED)) &&
240 SEQ_GT(tp->snd_max, tp->snd_una) && /* SYN or SYN|ACK sent */
241 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
242 return (0);
243
244 /*
245 * Determine length of data that should be transmitted,
246 * and flags that will be used.
247 * If there is some data or critical controls (SYN, RST)
248 * to send, then transmit; otherwise, investigate further.
249 */
250 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
251 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
252 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
253 cc_after_idle(tp);
254 tp->t_flags &= ~TF_LASTIDLE;
255 if (idle) {
256 if (tp->t_flags & TF_MORETOCOME) {
257 tp->t_flags |= TF_LASTIDLE;
258 idle = 0;
259 }
260 }
261 again:
262 sendwin = 0;
263 /*
264 * If we've recently taken a timeout, snd_max will be greater than
265 * snd_nxt. There may be SACK information that allows us to avoid
266 * resending already delivered data. Adjust snd_nxt accordingly.
267 */
268 if ((tp->t_flags & TF_SACK_PERMIT) &&
269 (tp->sackhint.nexthole != NULL) &&
270 !IN_FASTRECOVERY(tp->t_flags)) {
271 sendwin = tcp_sack_adjust(tp);
272 }
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 + sendwin);
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) &&
295 (IN_FASTRECOVERY(tp->t_flags) ||
296 (SEQ_LT(tp->snd_nxt, tp->snd_max) && (tp->t_dupacks >= tcprexmtthresh))) &&
297 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
298 int32_t cwin;
299
300 if (IN_FASTRECOVERY(tp->t_flags)) {
301 cwin = imax(sendwin - tcp_compute_pipe(tp), 0);
302 } else {
303 cwin = imax(sendwin - off, 0);
304 }
305 /* Do not retransmit SACK segments beyond snd_recover */
306 if (SEQ_GT(p->end, tp->snd_recover)) {
307 /*
308 * (At least) part of sack hole extends beyond
309 * snd_recover. Check to see if we can rexmit data
310 * for this hole.
311 */
312 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
313 /*
314 * Can't rexmit any more data for this hole.
315 * That data will be rexmitted in the next
316 * sack recovery episode, when snd_recover
317 * moves past p->rxmit.
318 */
319 p = NULL;
320 goto after_sack_rexmit;
321 } else {
322 /* Can rexmit part of the current hole */
323 len = SEQ_SUB(tp->snd_recover, p->rxmit);
324 if (cwin <= len) {
325 len = cwin;
326 } else {
327 sendalot = 1;
328 }
329 }
330 } else {
331 len = SEQ_SUB(p->end, p->rxmit);
332 if (cwin <= len) {
333 len = cwin;
334 } else {
335 sendalot = 1;
336 }
337 }
338 /* we could have transmitted from the scoreboard,
339 * but sendwin (expected flightsize) - pipe didn't
340 * allow any transmission.
341 * Bypass recalculating the possible transmission
342 * length further down by setting sack_rxmit.
343 * Wouldn't be here if there would have been
344 * nothing in the scoreboard to transmit.
345 */
346 sack_rxmit = 1;
347 if (len > 0) {
348 off = SEQ_SUB(p->rxmit, tp->snd_una);
349 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
350 __func__, off));
351 }
352 }
353 after_sack_rexmit:
354 /*
355 * Get standard flags, and add SYN or FIN if requested by 'hidden'
356 * state flags.
357 */
358 if (tp->t_flags & TF_NEEDFIN)
359 flags |= TH_FIN;
360 if (tp->t_flags & TF_NEEDSYN)
361 flags |= TH_SYN;
362
363 SOCK_SENDBUF_LOCK(so);
364 /*
365 * If in persist timeout with window of 0, send 1 byte.
366 * Otherwise, if window is small but nonzero
367 * and timer expired, we will send what we can
368 * and go to transmit state.
369 */
370 if (tp->t_flags & TF_FORCEDATA) {
371 if (sendwin == 0) {
372 /*
373 * If we still have some data to send, then
374 * clear the FIN bit. Usually this would
375 * happen below when it realizes that we
376 * aren't sending all the data. However,
377 * if we have exactly 1 byte of unsent data,
378 * then it won't clear the FIN bit below,
379 * and if we are in persist state, we wind
380 * up sending the packet without recording
381 * that we sent the FIN bit.
382 *
383 * We can't just blindly clear the FIN bit,
384 * because if we don't have any more data
385 * to send then the probe will be the FIN
386 * itself.
387 */
388 if (off < sbused(&so->so_snd))
389 flags &= ~TH_FIN;
390 sendwin = 1;
391 } else {
392 tcp_timer_activate(tp, TT_PERSIST, 0);
393 tp->t_rxtshift = 0;
394 }
395 }
396
397 /*
398 * If snd_nxt == snd_max and we have transmitted a FIN, the
399 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
400 * a negative length. This can also occur when TCP opens up
401 * its congestion window while receiving additional duplicate
402 * acks after fast-retransmit because TCP will reset snd_nxt
403 * to snd_max after the fast-retransmit.
404 *
405 * In the normal retransmit-FIN-only case, however, snd_nxt will
406 * be set to snd_una, the offset will be 0, and the length may
407 * wind up 0.
408 *
409 * If sack_rxmit is true we are retransmitting from the scoreboard
410 * in which case len is already set.
411 */
412 if (sack_rxmit == 0) {
413 if ((sack_bytes_rxmt == 0) || SEQ_LT(tp->snd_nxt, tp->snd_max)) {
414 len = imin(sbavail(&so->so_snd), sendwin) - off;
415 } else {
416 /*
417 * We are inside of a SACK recovery episode and are
418 * sending new data, having retransmitted all the
419 * data possible in the scoreboard.
420 */
421 len = imin(sbavail(&so->so_snd) - off,
422 sendwin - tcp_compute_pipe(tp));
423 }
424 }
425
426 /*
427 * Lop off SYN bit if it has already been sent. However, if this
428 * is SYN-SENT state and if segment contains data and if we don't
429 * know that foreign host supports TAO, suppress sending segment.
430 */
431 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
432 if (tp->t_state != TCPS_SYN_RECEIVED)
433 flags &= ~TH_SYN;
434 /*
435 * When sending additional segments following a TFO SYN|ACK,
436 * do not include the SYN bit.
437 */
438 if ((tp->t_flags & TF_FASTOPEN) &&
439 (tp->t_state == TCPS_SYN_RECEIVED))
440 flags &= ~TH_SYN;
441 off--, len++;
442 }
443
444 /*
445 * Be careful not to send data and/or FIN on SYN segments.
446 * This measure is needed to prevent interoperability problems
447 * with not fully conformant TCP implementations.
448 */
449 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
450 len = 0;
451 flags &= ~TH_FIN;
452 }
453
454 /*
455 * On TFO sockets, ensure no data is sent in the following cases:
456 *
457 * - When retransmitting SYN|ACK on a passively-created socket
458 *
459 * - When retransmitting SYN on an actively created socket
460 *
461 * - When sending a zero-length cookie (cookie request) on an
462 * actively created socket
463 *
464 * - When the socket is in the CLOSED state (RST is being sent)
465 */
466 if ((tp->t_flags & TF_FASTOPEN) &&
467 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
468 ((tp->t_state == TCPS_SYN_SENT) &&
469 (tp->t_tfo_client_cookie_len == 0)) ||
470 (flags & TH_RST)))
471 len = 0;
472
473 /* Without fast-open there should never be data sent on a SYN. */
474 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
475 len = 0;
476 }
477
478 if (len <= 0) {
479 /*
480 * If FIN has been sent but not acked,
481 * but we haven't been called to retransmit,
482 * len will be < 0. Otherwise, window shrank
483 * after we sent into it. If window shrank to 0,
484 * cancel pending retransmit, pull snd_nxt back
485 * to (closed) window, and set the persist timer
486 * if it isn't already going. If the window didn't
487 * close completely, just wait for an ACK.
488 *
489 * We also do a general check here to ensure that
490 * we will set the persist timer when we have data
491 * to send, but a 0-byte window. This makes sure
492 * the persist timer is set even if the packet
493 * hits one of the "goto send" lines below.
494 */
495 len = 0;
496 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
497 (off < (int) sbavail(&so->so_snd)) &&
498 !tcp_timer_active(tp, TT_PERSIST)) {
499 tcp_timer_activate(tp, TT_REXMT, 0);
500 tp->t_rxtshift = 0;
501 tp->snd_nxt = tp->snd_una;
502 if (!tcp_timer_active(tp, TT_PERSIST))
503 tcp_setpersist(tp);
504 }
505 }
506
507 /* len will be >= 0 after this point. */
508 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
509
510 tcp_sndbuf_autoscale(tp, so, sendwin);
511
512 /*
513 * Decide if we can use TCP Segmentation Offloading (if supported by
514 * hardware).
515 *
516 * TSO may only be used if we are in a pure bulk sending state. The
517 * presence of TCP-MD5, IP options (IPsec), and possibly SACK
518 * retransmits prevent using TSO. With TSO the TCP header is the same
519 * (except for the sequence number) for all generated packets. This
520 * makes it impossible to transmit any options which vary per generated
521 * segment or packet.
522 *
523 * IPv4 handling has a clear separation of ip options and ip header
524 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
525 * the right thing below to provide length of just ip options and thus
526 * checking for ipoptlen is enough to decide if ip options are present.
527 */
528 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
529 /*
530 * Pre-calculate here as we save another lookup into the darknesses
531 * of IPsec that way and can actually decide if TSO is ok.
532 */
533 #ifdef INET6
534 if (isipv6 && IPSEC_ENABLED(ipv6))
535 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
536 #ifdef INET
537 else
538 #endif
539 #endif /* INET6 */
540 #ifdef INET
541 if (IPSEC_ENABLED(ipv4))
542 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
543 #endif /* INET */
544 #endif /* IPSEC */
545 #ifdef INET6
546 if (isipv6)
547 ipoptlen = ip6_optlen(inp);
548 else
549 #endif
550 if (inp->inp_options)
551 ipoptlen = inp->inp_options->m_len -
552 offsetof(struct ipoption, ipopt_list);
553 else
554 ipoptlen = 0;
555 ipoptlen += ipsec_optlen;
556
557 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
558 (tp->t_port == 0) &&
559 ((tp->t_flags & TF_SIGNATURE) == 0) &&
560 ((sack_rxmit == 0) || V_tcp_sack_tso) &&
561 (ipoptlen == 0 || (ipoptlen == ipsec_optlen &&
562 (tp->t_flags2 & TF2_IPSEC_TSO) != 0)) &&
563 !(flags & TH_SYN))
564 tso = 1;
565
566 if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
567 tp->snd_una + sbused(&so->so_snd))) {
568 flags &= ~TH_FIN;
569 }
570
571 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
572 (long)TCP_MAXWIN << tp->rcv_scale);
573
574 /*
575 * Sender silly window avoidance. We transmit under the following
576 * conditions when len is non-zero:
577 *
578 * - We have a full segment (or more with TSO)
579 * - This is the last buffer in a write()/send() and we are
580 * either idle or running NODELAY
581 * - we've timed out (e.g. persist timer)
582 * - we have more then 1/2 the maximum send window's worth of
583 * data (receiver may be limited the window size)
584 * - we need to retransmit
585 */
586 if (len) {
587 if (len >= tp->t_maxseg)
588 goto send;
589 /*
590 * As the TCP header options are now
591 * considered when setting up the initial
592 * window, we would not send the last segment
593 * if we skip considering the option length here.
594 * Note: this may not work when tcp headers change
595 * very dynamically in the future.
596 */
597 if ((((tp->t_flags & TF_SIGNATURE) ?
598 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
599 ((tp->t_flags & TF_RCVD_TSTMP) ?
600 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
601 len) >= tp->t_maxseg)
602 goto send;
603 /*
604 * NOTE! on localhost connections an 'ack' from the remote
605 * end may occur synchronously with the output and cause
606 * us to flush a buffer queued with moretocome. XXX
607 *
608 * note: the len + off check is almost certainly unnecessary.
609 */
610 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
611 (idle || (tp->t_flags & TF_NODELAY)) &&
612 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
613 (tp->t_flags & TF_NOPUSH) == 0) {
614 goto send;
615 }
616 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
617 goto send;
618 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
619 goto send;
620 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
621 goto send;
622 if (sack_rxmit)
623 goto send;
624 }
625
626 /*
627 * Sending of standalone window updates.
628 *
629 * Window updates are important when we close our window due to a
630 * full socket buffer and are opening it again after the application
631 * reads data from it. Once the window has opened again and the
632 * remote end starts to send again the ACK clock takes over and
633 * provides the most current window information.
634 *
635 * We must avoid the silly window syndrome whereas every read
636 * from the receive buffer, no matter how small, causes a window
637 * update to be sent. We also should avoid sending a flurry of
638 * window updates when the socket buffer had queued a lot of data
639 * and the application is doing small reads.
640 *
641 * Prevent a flurry of pointless window updates by only sending
642 * an update when we can increase the advertized window by more
643 * than 1/4th of the socket buffer capacity. When the buffer is
644 * getting full or is very small be more aggressive and send an
645 * update whenever we can increase by two mss sized segments.
646 * In all other situations the ACK's to new incoming data will
647 * carry further window increases.
648 *
649 * Don't send an independent window update if a delayed
650 * ACK is pending (it will get piggy-backed on it) or the
651 * remote side already has done a half-close and won't send
652 * more data.
653 */
654 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
655 !(tp->t_flags & TF_DELACK) &&
656 !TCPS_HAVERCVDFIN(tp->t_state)) {
657 /*
658 * "adv" is the amount we could increase the window,
659 * taking into account that we are limited by
660 * TCP_MAXWIN << tp->rcv_scale.
661 */
662 int32_t adv;
663 int oldwin;
664
665 adv = recwin;
666 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
667 oldwin = (tp->rcv_adv - tp->rcv_nxt);
668 if (adv > oldwin)
669 adv -= oldwin;
670 else
671 adv = 0;
672 } else
673 oldwin = 0;
674
675 /*
676 * If the new window size ends up being the same as or less
677 * than the old size when it is scaled, then don't force
678 * a window update.
679 */
680 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
681 goto dontupdate;
682
683 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
684 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
685 recwin <= (so->so_rcv.sb_hiwat / 8) ||
686 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
687 adv >= TCP_MAXWIN << tp->rcv_scale))
688 goto send;
689 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
690 goto send;
691 }
692 dontupdate:
693
694 /*
695 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
696 * is also a catch-all for the retransmit timer timeout case.
697 */
698 if (tp->t_flags & TF_ACKNOW)
699 goto send;
700 if ((flags & TH_RST) ||
701 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
702 goto send;
703 if (SEQ_GT(tp->snd_up, tp->snd_una))
704 goto send;
705 /*
706 * If our state indicates that FIN should be sent
707 * and we have not yet done so, then we need to send.
708 */
709 if (flags & TH_FIN &&
710 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
711 goto send;
712 /*
713 * In SACK, it is possible for tcp_output to fail to send a segment
714 * after the retransmission timer has been turned off. Make sure
715 * that the retransmission timer is set.
716 */
717 if ((tp->t_flags & TF_SACK_PERMIT) &&
718 SEQ_GT(tp->snd_max, tp->snd_una) &&
719 !tcp_timer_active(tp, TT_REXMT) &&
720 !tcp_timer_active(tp, TT_PERSIST)) {
721 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
722 goto just_return;
723 }
724 /*
725 * TCP window updates are not reliable, rather a polling protocol
726 * using ``persist'' packets is used to insure receipt of window
727 * updates. The three ``states'' for the output side are:
728 * idle not doing retransmits or persists
729 * persisting to move a small or zero window
730 * (re)transmitting and thereby not persisting
731 *
732 * tcp_timer_active(tp, TT_PERSIST)
733 * is true when we are in persist state.
734 * (tp->t_flags & TF_FORCEDATA)
735 * is set when we are called to send a persist packet.
736 * tcp_timer_active(tp, TT_REXMT)
737 * is set when we are retransmitting
738 * The output side is idle when both timers are zero.
739 *
740 * If send window is too small, there is data to transmit, and no
741 * retransmit or persist is pending, then go to persist state.
742 * If nothing happens soon, send when timer expires:
743 * if window is nonzero, transmit what we can,
744 * otherwise force out a byte.
745 */
746 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
747 !tcp_timer_active(tp, TT_PERSIST)) {
748 tp->t_rxtshift = 0;
749 tcp_setpersist(tp);
750 }
751
752 /*
753 * No reason to send a segment, just return.
754 */
755 just_return:
756 SOCK_SENDBUF_UNLOCK(so);
757 return (0);
758
759 send:
760 SOCK_SENDBUF_LOCK_ASSERT(so);
761 if (len > 0) {
762 if (len >= tp->t_maxseg)
763 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
764 else
765 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
766 }
767 /*
768 * Before ESTABLISHED, force sending of initial options
769 * unless TCP set not to do any options.
770 * NOTE: we assume that the IP/TCP header plus TCP options
771 * always fit in a single mbuf, leaving room for a maximum
772 * link header, i.e.
773 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
774 */
775 optlen = 0;
776 #ifdef INET6
777 if (isipv6)
778 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
779 else
780 #endif
781 hdrlen = sizeof (struct tcpiphdr);
782
783 if (flags & TH_SYN) {
784 tp->snd_nxt = tp->iss;
785 }
786
787 /*
788 * Compute options for segment.
789 * We only have to care about SYN and established connection
790 * segments. Options for SYN-ACK segments are handled in TCP
791 * syncache.
792 */
793 to.to_flags = 0;
794 if ((tp->t_flags & TF_NOOPT) == 0) {
795 /* Maximum segment size. */
796 if (flags & TH_SYN) {
797 to.to_mss = tcp_mssopt(&inp->inp_inc);
798 if (tp->t_port)
799 to.to_mss -= V_tcp_udp_tunneling_overhead;
800 to.to_flags |= TOF_MSS;
801
802 /*
803 * On SYN or SYN|ACK transmits on TFO connections,
804 * only include the TFO option if it is not a
805 * retransmit, as the presence of the TFO option may
806 * have caused the original SYN or SYN|ACK to have
807 * been dropped by a middlebox.
808 */
809 if ((tp->t_flags & TF_FASTOPEN) &&
810 (tp->t_rxtshift == 0)) {
811 if (tp->t_state == TCPS_SYN_RECEIVED) {
812 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
813 to.to_tfo_cookie =
814 (u_int8_t *)&tp->t_tfo_cookie.server;
815 to.to_flags |= TOF_FASTOPEN;
816 wanted_cookie = 1;
817 } else if (tp->t_state == TCPS_SYN_SENT) {
818 to.to_tfo_len =
819 tp->t_tfo_client_cookie_len;
820 to.to_tfo_cookie =
821 tp->t_tfo_cookie.client;
822 to.to_flags |= TOF_FASTOPEN;
823 wanted_cookie = 1;
824 /*
825 * If we wind up having more data to
826 * send with the SYN than can fit in
827 * one segment, don't send any more
828 * until the SYN|ACK comes back from
829 * the other end.
830 */
831 dont_sendalot = 1;
832 }
833 }
834 }
835 /* Window scaling. */
836 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
837 to.to_wscale = tp->request_r_scale;
838 to.to_flags |= TOF_SCALE;
839 }
840 /* Timestamps. */
841 if ((tp->t_flags & TF_RCVD_TSTMP) ||
842 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
843 curticks = tcp_ts_getticks();
844 to.to_tsval = curticks + tp->ts_offset;
845 to.to_tsecr = tp->ts_recent;
846 to.to_flags |= TOF_TS;
847 if (tp->t_rxtshift == 1)
848 tp->t_badrxtwin = curticks;
849 }
850
851 /* Set receive buffer autosizing timestamp. */
852 if (tp->rfbuf_ts == 0 &&
853 (so->so_rcv.sb_flags & SB_AUTOSIZE))
854 tp->rfbuf_ts = tcp_ts_getticks();
855
856 /* Selective ACK's. */
857 if (tp->t_flags & TF_SACK_PERMIT) {
858 if (flags & TH_SYN)
859 to.to_flags |= TOF_SACKPERM;
860 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
861 tp->rcv_numsacks > 0) {
862 to.to_flags |= TOF_SACK;
863 to.to_nsacks = tp->rcv_numsacks;
864 to.to_sacks = (u_char *)tp->sackblks;
865 }
866 }
867 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
868 /* TCP-MD5 (RFC2385). */
869 /*
870 * Check that TCP_MD5SIG is enabled in tcpcb to
871 * account the size needed to set this TCP option.
872 */
873 if (tp->t_flags & TF_SIGNATURE)
874 to.to_flags |= TOF_SIGNATURE;
875 #endif /* TCP_SIGNATURE */
876
877 /* Processing the options. */
878 hdrlen += optlen = tcp_addoptions(&to, opt);
879 /*
880 * If we wanted a TFO option to be added, but it was unable
881 * to fit, ensure no data is sent.
882 */
883 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
884 !(to.to_flags & TOF_FASTOPEN))
885 len = 0;
886 }
887 if (tp->t_port) {
888 if (V_tcp_udp_tunneling_port == 0) {
889 /* The port was removed?? */
890 SOCK_SENDBUF_UNLOCK(so);
891 return (EHOSTUNREACH);
892 }
893 hdrlen += sizeof(struct udphdr);
894 }
895 /*
896 * Adjust data length if insertion of options will
897 * bump the packet length beyond the t_maxseg length.
898 * Clear the FIN bit because we cut off the tail of
899 * the segment.
900 */
901 if (len + optlen + ipoptlen > tp->t_maxseg) {
902 flags &= ~TH_FIN;
903
904 if (tso) {
905 u_int if_hw_tsomax;
906 u_int moff;
907 int max_len;
908
909 /* extract TSO information */
910 if_hw_tsomax = tp->t_tsomax;
911 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
912 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
913
914 /*
915 * Limit a TSO burst to prevent it from
916 * overflowing or exceeding the maximum length
917 * allowed by the network interface:
918 */
919 KASSERT(ipoptlen == ipsec_optlen,
920 ("%s: TSO can't do IP options", __func__));
921
922 /*
923 * Check if we should limit by maximum payload
924 * length:
925 */
926 if (if_hw_tsomax != 0) {
927 /* compute maximum TSO length */
928 max_len = if_hw_tsomax - hdrlen -
929 ipsec_optlen - max_linkhdr;
930 if (max_len <= 0) {
931 len = 0;
932 } else if (len > max_len) {
933 sendalot = 1;
934 len = max_len;
935 }
936 }
937
938 /*
939 * Prevent the last segment from being
940 * fractional unless the send sockbuf can be
941 * emptied:
942 */
943 max_len = tp->t_maxseg - optlen - ipsec_optlen;
944 if (((uint32_t)off + (uint32_t)len) <
945 sbavail(&so->so_snd)) {
946 moff = len % max_len;
947 if (moff != 0) {
948 len -= moff;
949 sendalot = 1;
950 }
951 }
952
953 /*
954 * In case there are too many small fragments
955 * don't use TSO:
956 */
957 if (len <= max_len) {
958 len = max_len;
959 sendalot = 1;
960 tso = 0;
961 }
962
963 /*
964 * Send the FIN in a separate segment
965 * after the bulk sending is done.
966 * We don't trust the TSO implementations
967 * to clear the FIN flag on all but the
968 * last segment.
969 */
970 if (tp->t_flags & TF_NEEDFIN)
971 sendalot = 1;
972 } else {
973 if (optlen + ipoptlen >= tp->t_maxseg) {
974 /*
975 * Since we don't have enough space to put
976 * the IP header chain and the TCP header in
977 * one packet as required by RFC 7112, don't
978 * send it. Also ensure that at least one
979 * byte of the payload can be put into the
980 * TCP segment.
981 */
982 SOCK_SENDBUF_UNLOCK(so);
983 error = EMSGSIZE;
984 sack_rxmit = 0;
985 goto out;
986 }
987 len = tp->t_maxseg - optlen - ipoptlen;
988 sendalot = 1;
989 if (dont_sendalot)
990 sendalot = 0;
991 }
992 } else
993 tso = 0;
994
995 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
996 ("%s: len > IP_MAXPACKET", __func__));
997
998 /*#ifdef DIAGNOSTIC*/
999 #ifdef INET6
1000 if (max_linkhdr + hdrlen > MCLBYTES)
1001 #else
1002 if (max_linkhdr + hdrlen > MHLEN)
1003 #endif
1004 panic("tcphdr too big");
1005 /*#endif*/
1006
1007 /*
1008 * This KASSERT is here to catch edge cases at a well defined place.
1009 * Before, those had triggered (random) panic conditions further down.
1010 */
1011 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1012
1013 /*
1014 * Grab a header mbuf, attaching a copy of data to
1015 * be transmitted, and initialize the header from
1016 * the template for sends on this connection.
1017 */
1018 if (len) {
1019 struct mbuf *mb;
1020 struct sockbuf *msb;
1021 u_int moff;
1022
1023 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1024 TCPSTAT_INC(tcps_sndprobe);
1025 #ifdef STATS
1026 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1027 stats_voi_update_abs_u32(tp->t_stats,
1028 VOI_TCP_RETXPB, len);
1029 else
1030 stats_voi_update_abs_u64(tp->t_stats,
1031 VOI_TCP_TXPB, len);
1032 #endif /* STATS */
1033 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1034 tp->t_sndrexmitpack++;
1035 TCPSTAT_INC(tcps_sndrexmitpack);
1036 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1037 if (sack_rxmit) {
1038 TCPSTAT_INC(tcps_sack_rexmits);
1039 if (tso) {
1040 TCPSTAT_INC(tcps_sack_rexmits_tso);
1041 }
1042 TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1043 }
1044 #ifdef STATS
1045 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1046 len);
1047 #endif /* STATS */
1048 } else {
1049 TCPSTAT_INC(tcps_sndpack);
1050 TCPSTAT_ADD(tcps_sndbyte, len);
1051 #ifdef STATS
1052 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1053 len);
1054 #endif /* STATS */
1055 }
1056 #ifdef INET6
1057 if (MHLEN < hdrlen + max_linkhdr)
1058 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1059 else
1060 #endif
1061 m = m_gethdr(M_NOWAIT, MT_DATA);
1062
1063 if (m == NULL) {
1064 SOCK_SENDBUF_UNLOCK(so);
1065 error = ENOBUFS;
1066 sack_rxmit = 0;
1067 goto out;
1068 }
1069
1070 m->m_data += max_linkhdr;
1071 m->m_len = hdrlen;
1072
1073 /*
1074 * Start the m_copy functions from the closest mbuf
1075 * to the offset in the socket buffer chain.
1076 */
1077 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1078 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1079 m_copydata(mb, moff, len,
1080 mtod(m, caddr_t) + hdrlen);
1081 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1082 sbsndptr_adv(&so->so_snd, mb, len);
1083 m->m_len += len;
1084 } else {
1085 int32_t old_len;
1086
1087 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1088 msb = NULL;
1089 else
1090 msb = &so->so_snd;
1091 old_len = len;
1092 m->m_next = tcp_m_copym(mb, moff,
1093 &len, if_hw_tsomaxsegcount,
1094 if_hw_tsomaxsegsize, msb, hw_tls);
1095 if (old_len != len)
1096 flags &= ~TH_FIN;
1097 if (len <= (tp->t_maxseg - optlen)) {
1098 /*
1099 * Must have ran out of mbufs for the copy
1100 * shorten it to no longer need tso. Lets
1101 * not put on sendalot since we are low on
1102 * mbufs.
1103 */
1104 tso = 0;
1105 }
1106 if (m->m_next == NULL) {
1107 SOCK_SENDBUF_UNLOCK(so);
1108 (void) m_free(m);
1109 error = ENOBUFS;
1110 sack_rxmit = 0;
1111 goto out;
1112 }
1113 }
1114
1115 /*
1116 * If we're sending everything we've got, set PUSH.
1117 * (This will keep happy those implementations which only
1118 * give data to the user when a buffer fills or
1119 * a PUSH comes in.)
1120 */
1121 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1122 !(flags & TH_SYN))
1123 flags |= TH_PUSH;
1124 SOCK_SENDBUF_UNLOCK(so);
1125 } else {
1126 SOCK_SENDBUF_UNLOCK(so);
1127 if (tp->t_flags & TF_ACKNOW)
1128 TCPSTAT_INC(tcps_sndacks);
1129 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1130 TCPSTAT_INC(tcps_sndctrl);
1131 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1132 TCPSTAT_INC(tcps_sndurg);
1133 else
1134 TCPSTAT_INC(tcps_sndwinup);
1135
1136 m = m_gethdr(M_NOWAIT, MT_DATA);
1137 if (m == NULL) {
1138 error = ENOBUFS;
1139 sack_rxmit = 0;
1140 goto out;
1141 }
1142 #ifdef INET6
1143 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1144 MHLEN >= hdrlen) {
1145 M_ALIGN(m, hdrlen);
1146 } else
1147 #endif
1148 m->m_data += max_linkhdr;
1149 m->m_len = hdrlen;
1150 }
1151 SOCK_SENDBUF_UNLOCK_ASSERT(so);
1152 m->m_pkthdr.rcvif = (struct ifnet *)0;
1153 #ifdef MAC
1154 mac_inpcb_create_mbuf(inp, m);
1155 #endif
1156 #ifdef INET6
1157 if (isipv6) {
1158 ip6 = mtod(m, struct ip6_hdr *);
1159 if (tp->t_port) {
1160 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1161 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1162 udp->uh_dport = tp->t_port;
1163 ulen = hdrlen + len - sizeof(struct ip6_hdr);
1164 udp->uh_ulen = htons(ulen);
1165 th = (struct tcphdr *)(udp + 1);
1166 } else {
1167 th = (struct tcphdr *)(ip6 + 1);
1168 }
1169 tcpip_fillheaders(inp, tp->t_port, ip6, th);
1170 } else
1171 #endif /* INET6 */
1172 {
1173 ip = mtod(m, struct ip *);
1174 if (tp->t_port) {
1175 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1176 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1177 udp->uh_dport = tp->t_port;
1178 ulen = hdrlen + len - sizeof(struct ip);
1179 udp->uh_ulen = htons(ulen);
1180 th = (struct tcphdr *)(udp + 1);
1181 } else
1182 th = (struct tcphdr *)(ip + 1);
1183 tcpip_fillheaders(inp, tp->t_port, ip, th);
1184 }
1185
1186 /*
1187 * Fill in fields, remembering maximum advertised
1188 * window for use in delaying messages about window sizes.
1189 * If resending a FIN, be sure not to use a new sequence number.
1190 */
1191 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1192 tp->snd_nxt == tp->snd_max)
1193 tp->snd_nxt--;
1194 /*
1195 * If we are starting a connection, send ECN setup
1196 * SYN packet. If we are on a retransmit, we may
1197 * resend those bits a number of times as per
1198 * RFC 3168.
1199 */
1200 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1201 flags |= tcp_ecn_output_syn_sent(tp);
1202 }
1203 /* Also handle parallel SYN for ECN */
1204 if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1205 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1206 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1207 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1208 (tp->t_flags2 & TF2_ECN_SND_ECE))
1209 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1210 #ifdef INET6
1211 if (isipv6) {
1212 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1213 ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1214 }
1215 else
1216 #endif
1217 {
1218 ip->ip_tos &= ~IPTOS_ECN_MASK;
1219 ip->ip_tos |= ect;
1220 }
1221 }
1222
1223 /*
1224 * If we are doing retransmissions, then snd_nxt will
1225 * not reflect the first unsent octet. For ACK only
1226 * packets, we do not want the sequence number of the
1227 * retransmitted packet, we want the sequence number
1228 * of the next unsent octet. So, if there is no data
1229 * (and no SYN or FIN), use snd_max instead of snd_nxt
1230 * when filling in ti_seq. But if we are in persist
1231 * state, snd_max might reflect one byte beyond the
1232 * right edge of the window, so use snd_nxt in that
1233 * case, since we know we aren't doing a retransmission.
1234 * (retransmit and persist are mutually exclusive...)
1235 */
1236 if (sack_rxmit == 0) {
1237 if (len || (flags & (TH_SYN|TH_FIN)) ||
1238 tcp_timer_active(tp, TT_PERSIST))
1239 th->th_seq = htonl(tp->snd_nxt);
1240 else
1241 th->th_seq = htonl(tp->snd_max);
1242 } else {
1243 th->th_seq = htonl(p->rxmit);
1244 p->rxmit += len;
1245 /*
1246 * Lost Retransmission Detection
1247 * trigger resending of a (then
1248 * still existing) hole, when
1249 * fack acks recoverypoint.
1250 */
1251 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1252 p->rxmit = tp->snd_recover;
1253 tp->sackhint.sack_bytes_rexmit += len;
1254 }
1255 if (IN_RECOVERY(tp->t_flags)) {
1256 /*
1257 * Account all bytes transmitted while
1258 * IN_RECOVERY, simplifying PRR and
1259 * Lost Retransmit Detection
1260 */
1261 tp->sackhint.prr_out += len;
1262 }
1263 th->th_ack = htonl(tp->rcv_nxt);
1264 if (optlen) {
1265 bcopy(opt, th + 1, optlen);
1266 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1267 }
1268 /*
1269 * Calculate receive window. Don't shrink window,
1270 * but avoid silly window syndrome.
1271 * If a RST segment is sent, advertise a window of zero.
1272 */
1273 if (flags & TH_RST) {
1274 recwin = 0;
1275 } else {
1276 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1277 recwin < tp->t_maxseg)
1278 recwin = 0;
1279 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1280 recwin < (tp->rcv_adv - tp->rcv_nxt))
1281 recwin = (tp->rcv_adv - tp->rcv_nxt);
1282 }
1283 /*
1284 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1285 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1286 * case is handled in syncache.
1287 */
1288 if (flags & TH_SYN)
1289 th->th_win = htons((u_short)
1290 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1291 else {
1292 /* Avoid shrinking window with window scaling. */
1293 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1294 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1295 }
1296
1297 /*
1298 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1299 * a 0 window. This may cause the remote transmitter to stall. This
1300 * flag tells soreceive() to disable delayed acknowledgements when
1301 * draining the buffer. This can occur if the receiver is attempting
1302 * to read more data than can be buffered prior to transmitting on
1303 * the connection.
1304 */
1305 if (th->th_win == 0) {
1306 tp->t_sndzerowin++;
1307 tp->t_flags |= TF_RXWIN0SENT;
1308 } else
1309 tp->t_flags &= ~TF_RXWIN0SENT;
1310 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1311 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1312 flags |= TH_URG;
1313 } else {
1314 /*
1315 * If no urgent pointer to send, then we pull
1316 * the urgent pointer to the left edge of the send window
1317 * so that it doesn't drift into the send window on sequence
1318 * number wraparound.
1319 */
1320 tp->snd_up = tp->snd_una; /* drag it along */
1321 }
1322 tcp_set_flags(th, flags);
1323
1324 /*
1325 * Put TCP length in extended header, and then
1326 * checksum extended header and data.
1327 */
1328 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1329
1330 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1331 if (to.to_flags & TOF_SIGNATURE) {
1332 /*
1333 * Calculate MD5 signature and put it into the place
1334 * determined before.
1335 * NOTE: since TCP options buffer doesn't point into
1336 * mbuf's data, calculate offset and use it.
1337 */
1338 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1339 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1340 /*
1341 * Do not send segment if the calculation of MD5
1342 * digest has failed.
1343 */
1344 m_freem(m);
1345 goto out;
1346 }
1347 }
1348 #endif
1349 #ifdef INET6
1350 if (isipv6) {
1351 /*
1352 * There is no need to fill in ip6_plen right now.
1353 * It will be filled later by ip6_output.
1354 */
1355 if (tp->t_port) {
1356 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1357 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1358 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1359 th->th_sum = htons(0);
1360 UDPSTAT_INC(udps_opackets);
1361 } else {
1362 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1363 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1364 th->th_sum = in6_cksum_pseudo(ip6,
1365 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1366 0);
1367 }
1368 }
1369 #endif
1370 #if defined(INET6) && defined(INET)
1371 else
1372 #endif
1373 #ifdef INET
1374 {
1375 if (tp->t_port) {
1376 m->m_pkthdr.csum_flags = CSUM_UDP;
1377 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1378 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1379 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1380 th->th_sum = htons(0);
1381 UDPSTAT_INC(udps_opackets);
1382 } else {
1383 m->m_pkthdr.csum_flags = CSUM_TCP;
1384 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1385 th->th_sum = in_pseudo(ip->ip_src.s_addr,
1386 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1387 IPPROTO_TCP + len + optlen));
1388 }
1389
1390 /* IP version must be set here for ipv4/ipv6 checking later */
1391 KASSERT(ip->ip_v == IPVERSION,
1392 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1393 }
1394 #endif
1395
1396 /*
1397 * Enable TSO and specify the size of the segments.
1398 * The TCP pseudo header checksum is always provided.
1399 */
1400 if (tso) {
1401 KASSERT(len > tp->t_maxseg - optlen - ipsec_optlen,
1402 ("%s: len <= tso_segsz", __func__));
1403 m->m_pkthdr.csum_flags |= CSUM_TSO;
1404 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen - ipsec_optlen;
1405 }
1406
1407 KASSERT(len + hdrlen == m_length(m, NULL),
1408 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1409 __func__, len, hdrlen, m_length(m, NULL)));
1410
1411 #ifdef TCP_HHOOK
1412 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1413 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1414 #endif
1415
1416 TCP_PROBE3(debug__output, tp, th, m);
1417
1418 /* We're getting ready to send; log now. */
1419 /* XXXMT: We are not honoring verbose logging. */
1420
1421 if (tcp_bblogging_on(tp))
1422 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1423 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1424 NULL);
1425 else
1426 lgb = NULL;
1427
1428 /*
1429 * Fill in IP length and desired time to live and
1430 * send to IP level. There should be a better way
1431 * to handle ttl and tos; we could keep them in
1432 * the template, but need a way to checksum without them.
1433 */
1434 /*
1435 * m->m_pkthdr.len should have been set before checksum calculation,
1436 * because in6_cksum() need it.
1437 */
1438 #ifdef INET6
1439 if (isipv6) {
1440 /*
1441 * we separately set hoplimit for every segment, since the
1442 * user might want to change the value via setsockopt.
1443 * Also, desired default hop limit might be changed via
1444 * Neighbor Discovery.
1445 */
1446 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1447
1448 /*
1449 * Set the packet size here for the benefit of DTrace probes.
1450 * ip6_output() will set it properly; it's supposed to include
1451 * the option header lengths as well.
1452 */
1453 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1454
1455 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1456 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1457 else
1458 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1459
1460 if (tp->t_state == TCPS_SYN_SENT)
1461 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1462
1463 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1464
1465 #ifdef TCPPCAP
1466 /* Save packet, if requested. */
1467 tcp_pcap_add(th, m, &(tp->t_outpkts));
1468 #endif
1469
1470 /* TODO: IPv6 IP6TOS_ECT bit on */
1471 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1472 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1473 NULL, NULL, inp);
1474
1475 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1476 mtu = inp->inp_route6.ro_nh->nh_mtu;
1477 }
1478 #endif /* INET6 */
1479 #if defined(INET) && defined(INET6)
1480 else
1481 #endif
1482 #ifdef INET
1483 {
1484 ip->ip_len = htons(m->m_pkthdr.len);
1485 #ifdef INET6
1486 if (inp->inp_vflag & INP_IPV6PROTO)
1487 ip->ip_ttl = in6_selecthlim(inp, NULL);
1488 #endif /* INET6 */
1489 /*
1490 * If we do path MTU discovery, then we set DF on every packet.
1491 * This might not be the best thing to do according to RFC3390
1492 * Section 2. However the tcp hostcache migitates the problem
1493 * so it affects only the first tcp connection with a host.
1494 *
1495 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1496 */
1497 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1498 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1499 if (tp->t_port == 0 || len < V_tcp_minmss) {
1500 ip->ip_off |= htons(IP_DF);
1501 }
1502 } else {
1503 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1504 }
1505
1506 if (tp->t_state == TCPS_SYN_SENT)
1507 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1508
1509 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1510
1511 #ifdef TCPPCAP
1512 /* Save packet, if requested. */
1513 tcp_pcap_add(th, m, &(tp->t_outpkts));
1514 #endif
1515
1516 error = ip_output(m, inp->inp_options, &inp->inp_route,
1517 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1518
1519 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1520 mtu = inp->inp_route.ro_nh->nh_mtu;
1521 }
1522 #endif /* INET */
1523
1524 if (lgb != NULL) {
1525 lgb->tlb_errno = error;
1526 lgb = NULL;
1527 }
1528 out:
1529 if (error == 0)
1530 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1531 /*
1532 * In transmit state, time the transmission and arrange for
1533 * the retransmit. In persist state, just set snd_max. In a closed
1534 * state just return.
1535 */
1536 if (flags & TH_RST) {
1537 TCPSTAT_INC(tcps_sndtotal);
1538 return (0);
1539 } else if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1540 !tcp_timer_active(tp, TT_PERSIST)) {
1541 tcp_seq startseq = tp->snd_nxt;
1542
1543 /*
1544 * Advance snd_nxt over sequence space of this segment.
1545 */
1546 if (flags & (TH_SYN|TH_FIN)) {
1547 if (flags & TH_SYN)
1548 tp->snd_nxt++;
1549 if (flags & TH_FIN) {
1550 tp->snd_nxt++;
1551 tp->t_flags |= TF_SENTFIN;
1552 }
1553 }
1554 if (sack_rxmit)
1555 goto timer;
1556 tp->snd_nxt += len;
1557 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1558 /*
1559 * Update "made progress" indication if we just
1560 * added new data to an empty socket buffer.
1561 */
1562 if (tp->snd_una == tp->snd_max)
1563 tp->t_acktime = ticks;
1564 tp->snd_max = tp->snd_nxt;
1565 /*
1566 * Time this transmission if not a retransmission and
1567 * not currently timing anything.
1568 */
1569 tp->t_sndtime = ticks;
1570 if (tp->t_rtttime == 0) {
1571 tp->t_rtttime = ticks;
1572 tp->t_rtseq = startseq;
1573 TCPSTAT_INC(tcps_segstimed);
1574 }
1575 #ifdef STATS
1576 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1577 tp->t_flags |= TF_GPUTINPROG;
1578 tp->gput_seq = startseq;
1579 tp->gput_ack = startseq +
1580 ulmin(sbavail(&so->so_snd) - off, sendwin);
1581 tp->gput_ts = tcp_ts_getticks();
1582 }
1583 #endif /* STATS */
1584 }
1585
1586 /*
1587 * Set retransmit timer if not currently set,
1588 * and not doing a pure ack or a keep-alive probe.
1589 * Initial value for retransmit timer is smoothed
1590 * round-trip time + 2 * round-trip time variance.
1591 * Initialize shift counter which is used for backoff
1592 * of retransmit time.
1593 */
1594 timer:
1595 if (!tcp_timer_active(tp, TT_REXMT) &&
1596 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1597 (tp->snd_nxt != tp->snd_una))) {
1598 if (tcp_timer_active(tp, TT_PERSIST)) {
1599 tcp_timer_activate(tp, TT_PERSIST, 0);
1600 tp->t_rxtshift = 0;
1601 }
1602 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1603 } else if (len == 0 && sbavail(&so->so_snd) &&
1604 !tcp_timer_active(tp, TT_REXMT) &&
1605 !tcp_timer_active(tp, TT_PERSIST)) {
1606 /*
1607 * Avoid a situation where we do not set persist timer
1608 * after a zero window condition. For example:
1609 * 1) A -> B: packet with enough data to fill the window
1610 * 2) B -> A: ACK for #1 + new data (0 window
1611 * advertisement)
1612 * 3) A -> B: ACK for #2, 0 len packet
1613 *
1614 * In this case, A will not activate the persist timer,
1615 * because it chose to send a packet. Unless tcp_output
1616 * is called for some other reason (delayed ack timer,
1617 * another input packet from B, socket syscall), A will
1618 * not send zero window probes.
1619 *
1620 * So, if you send a 0-length packet, but there is data
1621 * in the socket buffer, and neither the rexmt or
1622 * persist timer is already set, then activate the
1623 * persist timer.
1624 */
1625 tp->t_rxtshift = 0;
1626 tcp_setpersist(tp);
1627 }
1628 } else {
1629 /*
1630 * Persist case, update snd_max but since we are in
1631 * persist mode (no window) we do not update snd_nxt.
1632 */
1633 int xlen = len;
1634 if (flags & TH_SYN)
1635 ++xlen;
1636 if (flags & TH_FIN) {
1637 ++xlen;
1638 tp->t_flags |= TF_SENTFIN;
1639 }
1640 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1641 tp->snd_max = tp->snd_nxt + xlen;
1642 }
1643 if ((error == 0) &&
1644 (tp->rcv_numsacks > 0) &&
1645 TCPS_HAVEESTABLISHED(tp->t_state) &&
1646 (tp->t_flags & TF_SACK_PERMIT)) {
1647 /* Clean up any DSACK's sent */
1648 tcp_clean_dsack_blocks(tp);
1649 }
1650 if ((error == 0) &&
1651 sack_rxmit &&
1652 SEQ_LT(tp->snd_nxt, SEQ_MIN(p->rxmit, p->end))) {
1653 /*
1654 * When transmitting from SACK scoreboard
1655 * after an RTO, pull snd_nxt along.
1656 */
1657 tp->snd_nxt = SEQ_MIN(p->rxmit, p->end);
1658 }
1659 if (error) {
1660 /*
1661 * We know that the packet was lost, so back out the
1662 * sequence number advance, if any.
1663 *
1664 * If the error is EPERM the packet got blocked by the
1665 * local firewall. Normally we should terminate the
1666 * connection but the blocking may have been spurious
1667 * due to a firewall reconfiguration cycle. So we treat
1668 * it like a packet loss and let the retransmit timer and
1669 * timeouts do their work over time.
1670 * XXX: It is a POLA question whether calling tcp_drop right
1671 * away would be the really correct behavior instead.
1672 */
1673 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1674 !tcp_timer_active(tp, TT_PERSIST)) &&
1675 ((flags & TH_SYN) == 0) &&
1676 (error != EPERM)) {
1677 if (sack_rxmit) {
1678 p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1679 tp->sackhint.sack_bytes_rexmit -= len;
1680 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1681 ("sackhint bytes rtx >= 0"));
1682 KASSERT((flags & TH_FIN) == 0,
1683 ("error while FIN with SACK rxmit"));
1684 } else {
1685 tp->snd_nxt -= len;
1686 if (flags & TH_FIN)
1687 tp->snd_nxt--;
1688 }
1689 if (IN_RECOVERY(tp->t_flags))
1690 tp->sackhint.prr_out -= len;
1691 }
1692 SOCK_SENDBUF_UNLOCK_ASSERT(so); /* Check gotos. */
1693 switch (error) {
1694 case EACCES:
1695 case EPERM:
1696 tp->t_softerror = error;
1697 return (error);
1698 case ENOBUFS:
1699 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1700 tp->snd_cwnd = tcp_maxseg(tp);
1701 return (0);
1702 case EMSGSIZE:
1703 /*
1704 * For some reason the interface we used initially
1705 * to send segments changed to another or lowered
1706 * its MTU.
1707 * If TSO was active we either got an interface
1708 * without TSO capabilits or TSO was turned off.
1709 * If we obtained mtu from ip_output() then update
1710 * it and try again.
1711 */
1712 if (tso)
1713 tp->t_flags &= ~TF_TSO;
1714 if (mtu != 0) {
1715 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1716 goto again;
1717 }
1718 return (error);
1719 case EHOSTDOWN:
1720 case EHOSTUNREACH:
1721 case ENETDOWN:
1722 case ENETUNREACH:
1723 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1724 tp->t_softerror = error;
1725 return (0);
1726 }
1727 /* FALLTHROUGH */
1728 default:
1729 return (error);
1730 }
1731 }
1732 TCPSTAT_INC(tcps_sndtotal);
1733
1734 /*
1735 * Data sent (as far as we can tell).
1736 * If this advertises a larger window than any other segment,
1737 * then remember the size of the advertised window.
1738 * Any pending ACK has now been sent.
1739 */
1740 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1741 tp->rcv_adv = tp->rcv_nxt + recwin;
1742 tp->last_ack_sent = tp->rcv_nxt;
1743 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1744 if (tcp_timer_active(tp, TT_DELACK))
1745 tcp_timer_activate(tp, TT_DELACK, 0);
1746 if (sendalot)
1747 goto again;
1748 return (0);
1749 }
1750
1751 void
tcp_setpersist(struct tcpcb * tp)1752 tcp_setpersist(struct tcpcb *tp)
1753 {
1754 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1755 int tt;
1756 int maxunacktime;
1757
1758 tp->t_flags &= ~TF_PREVVALID;
1759 if (tcp_timer_active(tp, TT_REXMT))
1760 panic("tcp_setpersist: retransmit pending");
1761 /*
1762 * If the state is already closed, don't bother.
1763 */
1764 if (tp->t_state == TCPS_CLOSED)
1765 return;
1766
1767 /*
1768 * Start/restart persistence timer.
1769 */
1770 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1771 tcp_persmin, tcp_persmax);
1772 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1773 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1774 if (maxunacktime < 1)
1775 maxunacktime = 1;
1776 if (maxunacktime < tt)
1777 tt = maxunacktime;
1778 }
1779 tcp_timer_activate(tp, TT_PERSIST, tt);
1780 if (tp->t_rxtshift < V_tcp_retries)
1781 tp->t_rxtshift++;
1782 }
1783
1784 /*
1785 * Insert TCP options according to the supplied parameters to the place
1786 * optp in a consistent way. Can handle unaligned destinations.
1787 *
1788 * The order of the option processing is crucial for optimal packing and
1789 * alignment for the scarce option space.
1790 *
1791 * The optimal order for a SYN/SYN-ACK segment is:
1792 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1793 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1794 *
1795 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1796 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1797 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1798 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1799 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1800 */
1801 int
tcp_addoptions(struct tcpopt * to,u_char * optp)1802 tcp_addoptions(struct tcpopt *to, u_char *optp)
1803 {
1804 u_int32_t mask, optlen = 0;
1805
1806 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1807 if ((to->to_flags & mask) != mask)
1808 continue;
1809 if (optlen == TCP_MAXOLEN)
1810 break;
1811 switch (to->to_flags & mask) {
1812 case TOF_MSS:
1813 while (optlen % 4) {
1814 optlen += TCPOLEN_NOP;
1815 *optp++ = TCPOPT_NOP;
1816 }
1817 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1818 continue;
1819 optlen += TCPOLEN_MAXSEG;
1820 *optp++ = TCPOPT_MAXSEG;
1821 *optp++ = TCPOLEN_MAXSEG;
1822 to->to_mss = htons(to->to_mss);
1823 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1824 optp += sizeof(to->to_mss);
1825 break;
1826 case TOF_SCALE:
1827 while (!optlen || optlen % 2 != 1) {
1828 optlen += TCPOLEN_NOP;
1829 *optp++ = TCPOPT_NOP;
1830 }
1831 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1832 continue;
1833 optlen += TCPOLEN_WINDOW;
1834 *optp++ = TCPOPT_WINDOW;
1835 *optp++ = TCPOLEN_WINDOW;
1836 *optp++ = to->to_wscale;
1837 break;
1838 case TOF_SACKPERM:
1839 while (optlen % 2) {
1840 optlen += TCPOLEN_NOP;
1841 *optp++ = TCPOPT_NOP;
1842 }
1843 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1844 continue;
1845 optlen += TCPOLEN_SACK_PERMITTED;
1846 *optp++ = TCPOPT_SACK_PERMITTED;
1847 *optp++ = TCPOLEN_SACK_PERMITTED;
1848 break;
1849 case TOF_TS:
1850 while (!optlen || optlen % 4 != 2) {
1851 optlen += TCPOLEN_NOP;
1852 *optp++ = TCPOPT_NOP;
1853 }
1854 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1855 continue;
1856 optlen += TCPOLEN_TIMESTAMP;
1857 *optp++ = TCPOPT_TIMESTAMP;
1858 *optp++ = TCPOLEN_TIMESTAMP;
1859 to->to_tsval = htonl(to->to_tsval);
1860 to->to_tsecr = htonl(to->to_tsecr);
1861 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1862 optp += sizeof(to->to_tsval);
1863 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1864 optp += sizeof(to->to_tsecr);
1865 break;
1866 case TOF_SIGNATURE:
1867 {
1868 int siglen = TCPOLEN_SIGNATURE - 2;
1869
1870 while (!optlen || optlen % 4 != 2) {
1871 optlen += TCPOLEN_NOP;
1872 *optp++ = TCPOPT_NOP;
1873 }
1874 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1875 to->to_flags &= ~TOF_SIGNATURE;
1876 continue;
1877 }
1878 optlen += TCPOLEN_SIGNATURE;
1879 *optp++ = TCPOPT_SIGNATURE;
1880 *optp++ = TCPOLEN_SIGNATURE;
1881 to->to_signature = optp;
1882 while (siglen--)
1883 *optp++ = 0;
1884 break;
1885 }
1886 case TOF_SACK:
1887 {
1888 int sackblks = 0;
1889 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1890 tcp_seq sack_seq;
1891
1892 while (!optlen || optlen % 4 != 2) {
1893 optlen += TCPOLEN_NOP;
1894 *optp++ = TCPOPT_NOP;
1895 }
1896 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1897 continue;
1898 optlen += TCPOLEN_SACKHDR;
1899 *optp++ = TCPOPT_SACK;
1900 sackblks = min(to->to_nsacks,
1901 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1902 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1903 while (sackblks--) {
1904 sack_seq = htonl(sack->start);
1905 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1906 optp += sizeof(sack_seq);
1907 sack_seq = htonl(sack->end);
1908 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1909 optp += sizeof(sack_seq);
1910 optlen += TCPOLEN_SACK;
1911 sack++;
1912 }
1913 TCPSTAT_INC(tcps_sack_send_blocks);
1914 break;
1915 }
1916 case TOF_FASTOPEN:
1917 {
1918 int total_len;
1919
1920 /* XXX is there any point to aligning this option? */
1921 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1922 if (TCP_MAXOLEN - optlen < total_len) {
1923 to->to_flags &= ~TOF_FASTOPEN;
1924 continue;
1925 }
1926 *optp++ = TCPOPT_FAST_OPEN;
1927 *optp++ = total_len;
1928 if (to->to_tfo_len > 0) {
1929 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1930 optp += to->to_tfo_len;
1931 }
1932 optlen += total_len;
1933 break;
1934 }
1935 default:
1936 panic("%s: unknown TCP option type", __func__);
1937 break;
1938 }
1939 }
1940
1941 /* Terminate and pad TCP options to a 4 byte boundary. */
1942 if (optlen % 4) {
1943 optlen += TCPOLEN_EOL;
1944 *optp++ = TCPOPT_EOL;
1945 }
1946 /*
1947 * According to RFC 793 (STD0007):
1948 * "The content of the header beyond the End-of-Option option
1949 * must be header padding (i.e., zero)."
1950 * and later: "The padding is composed of zeros."
1951 */
1952 while (optlen % 4) {
1953 optlen += TCPOLEN_PAD;
1954 *optp++ = TCPOPT_PAD;
1955 }
1956
1957 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1958 return (optlen);
1959 }
1960
1961 /*
1962 * This is a copy of m_copym(), taking the TSO segment size/limit
1963 * constraints into account, and advancing the sndptr as it goes.
1964 */
1965 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)1966 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1967 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1968 {
1969 #ifdef KERN_TLS
1970 struct ktls_session *tls, *ntls;
1971 struct mbuf *start __diagused;
1972 #endif
1973 struct mbuf *n, **np;
1974 struct mbuf *top;
1975 int32_t off = off0;
1976 int32_t len = *plen;
1977 int32_t fragsize;
1978 int32_t len_cp = 0;
1979 int32_t *pkthdrlen;
1980 uint32_t mlen, frags;
1981 bool copyhdr;
1982
1983 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1984 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1985 if (off == 0 && m->m_flags & M_PKTHDR)
1986 copyhdr = true;
1987 else
1988 copyhdr = false;
1989 while (off > 0) {
1990 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1991 if (off < m->m_len)
1992 break;
1993 off -= m->m_len;
1994 if ((sb) && (m == sb->sb_sndptr)) {
1995 sb->sb_sndptroff += m->m_len;
1996 sb->sb_sndptr = m->m_next;
1997 }
1998 m = m->m_next;
1999 }
2000 np = ⊤
2001 top = NULL;
2002 pkthdrlen = NULL;
2003 #ifdef KERN_TLS
2004 if (hw_tls && (m->m_flags & M_EXTPG))
2005 tls = m->m_epg_tls;
2006 else
2007 tls = NULL;
2008 start = m;
2009 #endif
2010 while (len > 0) {
2011 if (m == NULL) {
2012 KASSERT(len == M_COPYALL,
2013 ("tcp_m_copym, length > size of mbuf chain"));
2014 *plen = len_cp;
2015 if (pkthdrlen != NULL)
2016 *pkthdrlen = len_cp;
2017 break;
2018 }
2019 #ifdef KERN_TLS
2020 if (hw_tls) {
2021 if (m->m_flags & M_EXTPG)
2022 ntls = m->m_epg_tls;
2023 else
2024 ntls = NULL;
2025
2026 /*
2027 * Avoid mixing TLS records with handshake
2028 * data or TLS records from different
2029 * sessions.
2030 */
2031 if (tls != ntls) {
2032 MPASS(m != start);
2033 *plen = len_cp;
2034 if (pkthdrlen != NULL)
2035 *pkthdrlen = len_cp;
2036 break;
2037 }
2038 }
2039 #endif
2040 mlen = min(len, m->m_len - off);
2041 if (seglimit) {
2042 /*
2043 * For M_EXTPG mbufs, add 3 segments
2044 * + 1 in case we are crossing page boundaries
2045 * + 2 in case the TLS hdr/trailer are used
2046 * It is cheaper to just add the segments
2047 * than it is to take the cache miss to look
2048 * at the mbuf ext_pgs state in detail.
2049 */
2050 if (m->m_flags & M_EXTPG) {
2051 fragsize = min(segsize, PAGE_SIZE);
2052 frags = 3;
2053 } else {
2054 fragsize = segsize;
2055 frags = 0;
2056 }
2057
2058 /* Break if we really can't fit anymore. */
2059 if ((frags + 1) >= seglimit) {
2060 *plen = len_cp;
2061 if (pkthdrlen != NULL)
2062 *pkthdrlen = len_cp;
2063 break;
2064 }
2065
2066 /*
2067 * Reduce size if you can't copy the whole
2068 * mbuf. If we can't copy the whole mbuf, also
2069 * adjust len so the loop will end after this
2070 * mbuf.
2071 */
2072 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2073 mlen = (seglimit - frags - 1) * fragsize;
2074 len = mlen;
2075 *plen = len_cp + len;
2076 if (pkthdrlen != NULL)
2077 *pkthdrlen = *plen;
2078 }
2079 frags += howmany(mlen, fragsize);
2080 if (frags == 0)
2081 frags++;
2082 seglimit -= frags;
2083 KASSERT(seglimit > 0,
2084 ("%s: seglimit went too low", __func__));
2085 }
2086 if (copyhdr)
2087 n = m_gethdr(M_NOWAIT, m->m_type);
2088 else
2089 n = m_get(M_NOWAIT, m->m_type);
2090 *np = n;
2091 if (n == NULL)
2092 goto nospace;
2093 if (copyhdr) {
2094 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2095 goto nospace;
2096 if (len == M_COPYALL)
2097 n->m_pkthdr.len -= off0;
2098 else
2099 n->m_pkthdr.len = len;
2100 pkthdrlen = &n->m_pkthdr.len;
2101 copyhdr = false;
2102 }
2103 n->m_len = mlen;
2104 len_cp += n->m_len;
2105 if (m->m_flags & (M_EXT | M_EXTPG)) {
2106 n->m_data = m->m_data + off;
2107 mb_dupcl(n, m);
2108 } else
2109 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2110 (u_int)n->m_len);
2111
2112 if (sb && (sb->sb_sndptr == m) &&
2113 ((n->m_len + off) >= m->m_len) && m->m_next) {
2114 sb->sb_sndptroff += m->m_len;
2115 sb->sb_sndptr = m->m_next;
2116 }
2117 off = 0;
2118 if (len != M_COPYALL) {
2119 len -= n->m_len;
2120 }
2121 m = m->m_next;
2122 np = &n->m_next;
2123 }
2124 return (top);
2125 nospace:
2126 m_freem(top);
2127 return (NULL);
2128 }
2129
2130 void
tcp_sndbuf_autoscale(struct tcpcb * tp,struct socket * so,uint32_t sendwin)2131 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2132 {
2133
2134 /*
2135 * Automatic sizing of send socket buffer. Often the send buffer
2136 * size is not optimally adjusted to the actual network conditions
2137 * at hand (delay bandwidth product). Setting the buffer size too
2138 * small limits throughput on links with high bandwidth and high
2139 * delay (eg. trans-continental/oceanic links). Setting the
2140 * buffer size too big consumes too much real kernel memory,
2141 * especially with many connections on busy servers.
2142 *
2143 * The criteria to step up the send buffer one notch are:
2144 * 1. receive window of remote host is larger than send buffer
2145 * (with a fudge factor of 5/4th);
2146 * 2. send buffer is filled to 7/8th with data (so we actually
2147 * have data to make use of it);
2148 * 3. send buffer fill has not hit maximal automatic size;
2149 * 4. our send window (slow start and cogestion controlled) is
2150 * larger than sent but unacknowledged data in send buffer.
2151 *
2152 * The remote host receive window scaling factor may limit the
2153 * growing of the send buffer before it reaches its allowed
2154 * maximum.
2155 *
2156 * It scales directly with slow start or congestion window
2157 * and does at most one step per received ACK. This fast
2158 * scaling has the drawback of growing the send buffer beyond
2159 * what is strictly necessary to make full use of a given
2160 * delay*bandwidth product. However testing has shown this not
2161 * to be much of an problem. At worst we are trading wasting
2162 * of available bandwidth (the non-use of it) for wasting some
2163 * socket buffer memory.
2164 *
2165 * TODO: Shrink send buffer during idle periods together
2166 * with congestion window. Requires another timer. Has to
2167 * wait for upcoming tcp timer rewrite.
2168 *
2169 * XXXGL: should there be used sbused() or sbavail()?
2170 */
2171 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2172 int lowat;
2173
2174 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2175 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2176 sbused(&so->so_snd) >=
2177 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2178 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2179 sendwin >= (sbused(&so->so_snd) -
2180 (tp->snd_nxt - tp->snd_una))) {
2181 if (!sbreserve_locked(so, SO_SND,
2182 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2183 V_tcp_autosndbuf_max), curthread))
2184 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2185 }
2186 }
2187 }
2188