1/*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 *	The Regents of the University of California.  All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 *    notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 *    notice, this list of conditions and the following disclaimer in the
14 *    documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 *    may be used to endorse or promote products derived from this software
17 *    without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
32 */
33
34#include <sys/cdefs.h>
35__FBSDID("$FreeBSD$");
36
37#include "opt_inet.h"
38#include "opt_inet6.h"
39#include "opt_ipsec.h"
40#include "opt_kern_tls.h"
41#include "opt_tcpdebug.h"
42
43#include <sys/param.h>
44#include <sys/systm.h>
45#include <sys/arb.h>
46#include <sys/domain.h>
47#ifdef TCP_HHOOK
48#include <sys/hhook.h>
49#endif
50#include <sys/kernel.h>
51#ifdef KERN_TLS
52#include <sys/ktls.h>
53#endif
54#include <sys/lock.h>
55#include <sys/mbuf.h>
56#include <sys/mutex.h>
57#include <sys/protosw.h>
58#include <sys/qmath.h>
59#include <sys/sdt.h>
60#include <sys/socket.h>
61#include <sys/socketvar.h>
62#include <sys/sysctl.h>
63#include <sys/stats.h>
64
65#include <net/if.h>
66#include <net/route.h>
67#include <net/route/nhop.h>
68#include <net/vnet.h>
69
70#include <netinet/in.h>
71#include <netinet/in_kdtrace.h>
72#include <netinet/in_systm.h>
73#include <netinet/ip.h>
74#include <netinet/in_pcb.h>
75#include <netinet/ip_var.h>
76#include <netinet/ip_options.h>
77#ifdef INET6
78#include <netinet6/in6_pcb.h>
79#include <netinet/ip6.h>
80#include <netinet6/ip6_var.h>
81#endif
82#include <netinet/tcp.h>
83#define	TCPOUTFLAGS
84#include <netinet/tcp_fsm.h>
85#include <netinet/tcp_log_buf.h>
86#include <netinet/tcp_seq.h>
87#include <netinet/tcp_timer.h>
88#include <netinet/tcp_var.h>
89#include <netinet/tcpip.h>
90#include <netinet/cc/cc.h>
91#include <netinet/tcp_fastopen.h>
92#ifdef TCPPCAP
93#include <netinet/tcp_pcap.h>
94#endif
95#ifdef TCPDEBUG
96#include <netinet/tcp_debug.h>
97#endif
98#ifdef TCP_OFFLOAD
99#include <netinet/tcp_offload.h>
100#endif
101
102#include <netipsec/ipsec_support.h>
103
104#include <netinet/udp.h>
105#include <netinet/udp_var.h>
106#include <machine/in_cksum.h>
107
108#include <security/mac/mac_framework.h>
109
110VNET_DEFINE(int, path_mtu_discovery) = 1;
111SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
112	&VNET_NAME(path_mtu_discovery), 1,
113	"Enable Path MTU Discovery");
114
115VNET_DEFINE(int, tcp_do_tso) = 1;
116SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
117	&VNET_NAME(tcp_do_tso), 0,
118	"Enable TCP Segmentation Offload");
119
120VNET_DEFINE(int, tcp_sendspace) = 1024*32;
121#define	V_tcp_sendspace	VNET(tcp_sendspace)
122SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
123	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
124
125VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
126SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
127	&VNET_NAME(tcp_do_autosndbuf), 0,
128	"Enable automatic send buffer sizing");
129
130VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
131SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
132	&VNET_NAME(tcp_autosndbuf_inc), 0,
133	"Incrementor step size of automatic send buffer");
134
135VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
136SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
137	&VNET_NAME(tcp_autosndbuf_max), 0,
138	"Max size of automatic send buffer");
139
140VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
141#define	V_tcp_sendbuf_auto_lowat	VNET(tcp_sendbuf_auto_lowat)
142SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
143	&VNET_NAME(tcp_sendbuf_auto_lowat), 0,
144	"Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
145
146/*
147 * Make sure that either retransmit or persist timer is set for SYN, FIN and
148 * non-ACK.
149 */
150#define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
151	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
152	    tcp_timer_active((tp), TT_REXMT) ||				\
153	    tcp_timer_active((tp), TT_PERSIST),				\
154	    ("neither rexmt nor persist timer is set"))
155
156static void inline	cc_after_idle(struct tcpcb *tp);
157
158#ifdef TCP_HHOOK
159/*
160 * Wrapper for the TCP established output helper hook.
161 */
162void
163hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
164    struct tcpopt *to, uint32_t len, int tso)
165{
166	struct tcp_hhook_data hhook_data;
167
168	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
169		hhook_data.tp = tp;
170		hhook_data.th = th;
171		hhook_data.to = to;
172		hhook_data.len = len;
173		hhook_data.tso = tso;
174
175		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
176		    tp->osd);
177	}
178}
179#endif
180
181/*
182 * CC wrapper hook functions
183 */
184static void inline
185cc_after_idle(struct tcpcb *tp)
186{
187	INP_WLOCK_ASSERT(tp->t_inpcb);
188
189	if (CC_ALGO(tp)->after_idle != NULL)
190		CC_ALGO(tp)->after_idle(tp->ccv);
191}
192
193/*
194 * Tcp output routine: figure out what should be sent and send it.
195 */
196int
197tcp_output(struct tcpcb *tp)
198{
199	struct socket *so = tp->t_inpcb->inp_socket;
200	int32_t len;
201	uint32_t recwin, sendwin;
202	int off, flags, error = 0;	/* Keep compiler happy */
203	u_int if_hw_tsomaxsegcount = 0;
204	u_int if_hw_tsomaxsegsize = 0;
205	struct mbuf *m;
206	struct ip *ip = NULL;
207#ifdef TCPDEBUG
208	struct ipovly *ipov = NULL;
209#endif
210	struct tcphdr *th;
211	u_char opt[TCP_MAXOLEN];
212	unsigned ipoptlen, optlen, hdrlen, ulen;
213#if defined(IPSEC) || defined(IPSEC_SUPPORT)
214	unsigned ipsec_optlen = 0;
215#endif
216	int idle, sendalot, curticks;
217	int sack_rxmit, sack_bytes_rxmt;
218	struct sackhole *p;
219	int tso, mtu;
220	struct tcpopt to;
221	struct udphdr *udp = NULL;
222	unsigned int wanted_cookie = 0;
223	unsigned int dont_sendalot = 0;
224#if 0
225	int maxburst = TCP_MAXBURST;
226#endif
227#ifdef INET6
228	struct ip6_hdr *ip6 = NULL;
229	int isipv6;
230
231	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
232#endif
233#ifdef KERN_TLS
234	const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
235#else
236	const bool hw_tls = false;
237#endif
238
239	NET_EPOCH_ASSERT();
240	INP_WLOCK_ASSERT(tp->t_inpcb);
241
242#ifdef TCP_OFFLOAD
243	if (tp->t_flags & TF_TOE)
244		return (tcp_offload_output(tp));
245#endif
246
247	/*
248	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
249	 * only allow the initial SYN or SYN|ACK and those sent
250	 * by the retransmit timer.
251	 */
252	if (IS_FASTOPEN(tp->t_flags) &&
253	    ((tp->t_state == TCPS_SYN_SENT) ||
254	     (tp->t_state == TCPS_SYN_RECEIVED)) &&
255	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
256	    (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
257		return (0);
258
259	/*
260	 * Determine length of data that should be transmitted,
261	 * and flags that will be used.
262	 * If there is some data or critical controls (SYN, RST)
263	 * to send, then transmit; otherwise, investigate further.
264	 */
265	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
266	if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
267	    (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
268		cc_after_idle(tp);
269	tp->t_flags &= ~TF_LASTIDLE;
270	if (idle) {
271		if (tp->t_flags & TF_MORETOCOME) {
272			tp->t_flags |= TF_LASTIDLE;
273			idle = 0;
274		}
275	}
276again:
277	/*
278	 * If we've recently taken a timeout, snd_max will be greater than
279	 * snd_nxt.  There may be SACK information that allows us to avoid
280	 * resending already delivered data.  Adjust snd_nxt accordingly.
281	 */
282	if ((tp->t_flags & TF_SACK_PERMIT) &&
283	    SEQ_LT(tp->snd_nxt, tp->snd_max))
284		tcp_sack_adjust(tp);
285	sendalot = 0;
286	tso = 0;
287	mtu = 0;
288	off = tp->snd_nxt - tp->snd_una;
289	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
290
291	flags = tcp_outflags[tp->t_state];
292	/*
293	 * Send any SACK-generated retransmissions.  If we're explicitly trying
294	 * to send out new data (when sendalot is 1), bypass this function.
295	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
296	 * we're replacing a (future) new transmission with a retransmission
297	 * now, and we previously incremented snd_cwnd in tcp_input().
298	 */
299	/*
300	 * Still in sack recovery , reset rxmit flag to zero.
301	 */
302	sack_rxmit = 0;
303	sack_bytes_rxmt = 0;
304	len = 0;
305	p = NULL;
306	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
307	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
308		uint32_t cwin;
309
310		cwin =
311		    imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
312		/* Do not retransmit SACK segments beyond snd_recover */
313		if (SEQ_GT(p->end, tp->snd_recover)) {
314			/*
315			 * (At least) part of sack hole extends beyond
316			 * snd_recover. Check to see if we can rexmit data
317			 * for this hole.
318			 */
319			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
320				/*
321				 * Can't rexmit any more data for this hole.
322				 * That data will be rexmitted in the next
323				 * sack recovery episode, when snd_recover
324				 * moves past p->rxmit.
325				 */
326				p = NULL;
327				goto after_sack_rexmit;
328			} else
329				/* Can rexmit part of the current hole */
330				len = ((int32_t)ulmin(cwin,
331						   tp->snd_recover - p->rxmit));
332		} else
333			len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
334		off = p->rxmit - tp->snd_una;
335		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
336		    __func__, off));
337		if (len > 0) {
338			sack_rxmit = 1;
339			sendalot = 1;
340			TCPSTAT_INC(tcps_sack_rexmits);
341			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
342			    min(len, tcp_maxseg(tp)));
343		}
344	}
345after_sack_rexmit:
346	/*
347	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
348	 * state flags.
349	 */
350	if (tp->t_flags & TF_NEEDFIN)
351		flags |= TH_FIN;
352	if (tp->t_flags & TF_NEEDSYN)
353		flags |= TH_SYN;
354
355	SOCKBUF_LOCK(&so->so_snd);
356	/*
357	 * If in persist timeout with window of 0, send 1 byte.
358	 * Otherwise, if window is small but nonzero
359	 * and timer expired, we will send what we can
360	 * and go to transmit state.
361	 */
362	if (tp->t_flags & TF_FORCEDATA) {
363		if (sendwin == 0) {
364			/*
365			 * If we still have some data to send, then
366			 * clear the FIN bit.  Usually this would
367			 * happen below when it realizes that we
368			 * aren't sending all the data.  However,
369			 * if we have exactly 1 byte of unsent data,
370			 * then it won't clear the FIN bit below,
371			 * and if we are in persist state, we wind
372			 * up sending the packet without recording
373			 * that we sent the FIN bit.
374			 *
375			 * We can't just blindly clear the FIN bit,
376			 * because if we don't have any more data
377			 * to send then the probe will be the FIN
378			 * itself.
379			 */
380			if (off < sbused(&so->so_snd))
381				flags &= ~TH_FIN;
382			sendwin = 1;
383		} else {
384			tcp_timer_activate(tp, TT_PERSIST, 0);
385			tp->t_rxtshift = 0;
386		}
387	}
388
389	/*
390	 * If snd_nxt == snd_max and we have transmitted a FIN, the
391	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
392	 * a negative length.  This can also occur when TCP opens up
393	 * its congestion window while receiving additional duplicate
394	 * acks after fast-retransmit because TCP will reset snd_nxt
395	 * to snd_max after the fast-retransmit.
396	 *
397	 * In the normal retransmit-FIN-only case, however, snd_nxt will
398	 * be set to snd_una, the offset will be 0, and the length may
399	 * wind up 0.
400	 *
401	 * If sack_rxmit is true we are retransmitting from the scoreboard
402	 * in which case len is already set.
403	 */
404	if (sack_rxmit == 0) {
405		if (sack_bytes_rxmt == 0)
406			len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
407			    off);
408		else {
409			int32_t cwin;
410
411                        /*
412			 * We are inside of a SACK recovery episode and are
413			 * sending new data, having retransmitted all the
414			 * data possible in the scoreboard.
415			 */
416			len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
417			    off);
418			/*
419			 * Don't remove this (len > 0) check !
420			 * We explicitly check for len > 0 here (although it
421			 * isn't really necessary), to work around a gcc
422			 * optimization issue - to force gcc to compute
423			 * len above. Without this check, the computation
424			 * of len is bungled by the optimizer.
425			 */
426			if (len > 0) {
427				cwin = tp->snd_cwnd -
428					(tp->snd_nxt - tp->snd_recover) -
429					sack_bytes_rxmt;
430				if (cwin < 0)
431					cwin = 0;
432				len = imin(len, cwin);
433			}
434		}
435	}
436
437	/*
438	 * Lop off SYN bit if it has already been sent.  However, if this
439	 * is SYN-SENT state and if segment contains data and if we don't
440	 * know that foreign host supports TAO, suppress sending segment.
441	 */
442	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
443		if (tp->t_state != TCPS_SYN_RECEIVED)
444			flags &= ~TH_SYN;
445		/*
446		 * When sending additional segments following a TFO SYN|ACK,
447		 * do not include the SYN bit.
448		 */
449		if (IS_FASTOPEN(tp->t_flags) &&
450		    (tp->t_state == TCPS_SYN_RECEIVED))
451			flags &= ~TH_SYN;
452		off--, len++;
453	}
454
455	/*
456	 * Be careful not to send data and/or FIN on SYN segments.
457	 * This measure is needed to prevent interoperability problems
458	 * with not fully conformant TCP implementations.
459	 */
460	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
461		len = 0;
462		flags &= ~TH_FIN;
463	}
464
465	/*
466	 * On TFO sockets, ensure no data is sent in the following cases:
467	 *
468	 *  - When retransmitting SYN|ACK on a passively-created socket
469	 *
470	 *  - When retransmitting SYN on an actively created socket
471	 *
472	 *  - When sending a zero-length cookie (cookie request) on an
473	 *    actively created socket
474	 *
475	 *  - When the socket is in the CLOSED state (RST is being sent)
476	 */
477	if (IS_FASTOPEN(tp->t_flags) &&
478	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
479	     ((tp->t_state == TCPS_SYN_SENT) &&
480	      (tp->t_tfo_client_cookie_len == 0)) ||
481	     (flags & TH_RST)))
482		len = 0;
483	if (len <= 0) {
484		/*
485		 * If FIN has been sent but not acked,
486		 * but we haven't been called to retransmit,
487		 * len will be < 0.  Otherwise, window shrank
488		 * after we sent into it.  If window shrank to 0,
489		 * cancel pending retransmit, pull snd_nxt back
490		 * to (closed) window, and set the persist timer
491		 * if it isn't already going.  If the window didn't
492		 * close completely, just wait for an ACK.
493		 *
494		 * We also do a general check here to ensure that
495		 * we will set the persist timer when we have data
496		 * to send, but a 0-byte window. This makes sure
497		 * the persist timer is set even if the packet
498		 * hits one of the "goto send" lines below.
499		 */
500		len = 0;
501		if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
502			(off < (int) sbavail(&so->so_snd))) {
503			tcp_timer_activate(tp, TT_REXMT, 0);
504			tp->t_rxtshift = 0;
505			tp->snd_nxt = tp->snd_una;
506			if (!tcp_timer_active(tp, TT_PERSIST))
507				tcp_setpersist(tp);
508		}
509	}
510
511	/* len will be >= 0 after this point. */
512	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
513
514	tcp_sndbuf_autoscale(tp, so, sendwin);
515
516	/*
517	 * Decide if we can use TCP Segmentation Offloading (if supported by
518	 * hardware).
519	 *
520	 * TSO may only be used if we are in a pure bulk sending state.  The
521	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
522	 * IP options prevent using TSO.  With TSO the TCP header is the same
523	 * (except for the sequence number) for all generated packets.  This
524	 * makes it impossible to transmit any options which vary per generated
525	 * segment or packet.
526	 *
527	 * IPv4 handling has a clear separation of ip options and ip header
528	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
529	 * the right thing below to provide length of just ip options and thus
530	 * checking for ipoptlen is enough to decide if ip options are present.
531	 */
532#if defined(IPSEC) || defined(IPSEC_SUPPORT)
533	/*
534	 * Pre-calculate here as we save another lookup into the darknesses
535	 * of IPsec that way and can actually decide if TSO is ok.
536	 */
537#ifdef INET6
538	if (isipv6 && IPSEC_ENABLED(ipv6))
539		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
540#ifdef INET
541	else
542#endif
543#endif /* INET6 */
544#ifdef INET
545	if (IPSEC_ENABLED(ipv4))
546		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
547#endif /* INET */
548#endif /* IPSEC */
549#ifdef INET6
550	if (isipv6)
551		ipoptlen = ip6_optlen(tp->t_inpcb);
552	else
553#endif
554	if (tp->t_inpcb->inp_options)
555		ipoptlen = tp->t_inpcb->inp_options->m_len -
556				offsetof(struct ipoption, ipopt_list);
557	else
558		ipoptlen = 0;
559#if defined(IPSEC) || defined(IPSEC_SUPPORT)
560	ipoptlen += ipsec_optlen;
561#endif
562
563	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
564	    (tp->t_port == 0) &&
565	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
566	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
567	    ipoptlen == 0 && !(flags & TH_SYN))
568		tso = 1;
569
570	if (sack_rxmit) {
571		if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
572			flags &= ~TH_FIN;
573	} else {
574		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
575		    sbused(&so->so_snd)))
576			flags &= ~TH_FIN;
577	}
578
579	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
580	    (long)TCP_MAXWIN << tp->rcv_scale);
581
582	/*
583	 * Sender silly window avoidance.   We transmit under the following
584	 * conditions when len is non-zero:
585	 *
586	 *	- We have a full segment (or more with TSO)
587	 *	- This is the last buffer in a write()/send() and we are
588	 *	  either idle or running NODELAY
589	 *	- we've timed out (e.g. persist timer)
590	 *	- we have more then 1/2 the maximum send window's worth of
591	 *	  data (receiver may be limited the window size)
592	 *	- we need to retransmit
593	 */
594	if (len) {
595		if (len >= tp->t_maxseg)
596			goto send;
597		/*
598		 * As the TCP header options are now
599		 * considered when setting up the initial
600		 * window, we would not send the last segment
601		 * if we skip considering the option length here.
602		 * Note: this may not work when tcp headers change
603		 * very dynamically in the future.
604		 */
605		if ((((tp->t_flags & TF_SIGNATURE) ?
606			PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
607		    ((tp->t_flags & TF_RCVD_TSTMP) ?
608			PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
609		    len) >= tp->t_maxseg)
610			goto send;
611		/*
612		 * NOTE! on localhost connections an 'ack' from the remote
613		 * end may occur synchronously with the output and cause
614		 * us to flush a buffer queued with moretocome.  XXX
615		 *
616		 * note: the len + off check is almost certainly unnecessary.
617		 */
618		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
619		    (idle || (tp->t_flags & TF_NODELAY)) &&
620		    (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
621		    (tp->t_flags & TF_NOPUSH) == 0) {
622			goto send;
623		}
624		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
625			goto send;
626		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
627			goto send;
628		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
629			goto send;
630		if (sack_rxmit)
631			goto send;
632	}
633
634	/*
635	 * Sending of standalone window updates.
636	 *
637	 * Window updates are important when we close our window due to a
638	 * full socket buffer and are opening it again after the application
639	 * reads data from it.  Once the window has opened again and the
640	 * remote end starts to send again the ACK clock takes over and
641	 * provides the most current window information.
642	 *
643	 * We must avoid the silly window syndrome whereas every read
644	 * from the receive buffer, no matter how small, causes a window
645	 * update to be sent.  We also should avoid sending a flurry of
646	 * window updates when the socket buffer had queued a lot of data
647	 * and the application is doing small reads.
648	 *
649	 * Prevent a flurry of pointless window updates by only sending
650	 * an update when we can increase the advertized window by more
651	 * than 1/4th of the socket buffer capacity.  When the buffer is
652	 * getting full or is very small be more aggressive and send an
653	 * update whenever we can increase by two mss sized segments.
654	 * In all other situations the ACK's to new incoming data will
655	 * carry further window increases.
656	 *
657	 * Don't send an independent window update if a delayed
658	 * ACK is pending (it will get piggy-backed on it) or the
659	 * remote side already has done a half-close and won't send
660	 * more data.  Skip this if the connection is in T/TCP
661	 * half-open state.
662	 */
663	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
664	    !(tp->t_flags & TF_DELACK) &&
665	    !TCPS_HAVERCVDFIN(tp->t_state)) {
666		/*
667		 * "adv" is the amount we could increase the window,
668		 * taking into account that we are limited by
669		 * TCP_MAXWIN << tp->rcv_scale.
670		 */
671		int32_t adv;
672		int oldwin;
673
674		adv = recwin;
675		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
676			oldwin = (tp->rcv_adv - tp->rcv_nxt);
677			if (adv > oldwin)
678				adv -= oldwin;
679			else
680				adv = 0;
681		} else
682			oldwin = 0;
683
684		/*
685		 * If the new window size ends up being the same as or less
686		 * than the old size when it is scaled, then don't force
687		 * a window update.
688		 */
689		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
690			goto dontupdate;
691
692		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
693		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
694		     recwin <= (so->so_rcv.sb_hiwat / 8) ||
695		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
696		     adv >= TCP_MAXWIN << tp->rcv_scale))
697			goto send;
698		if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
699			goto send;
700	}
701dontupdate:
702
703	/*
704	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
705	 * is also a catch-all for the retransmit timer timeout case.
706	 */
707	if (tp->t_flags & TF_ACKNOW)
708		goto send;
709	if ((flags & TH_RST) ||
710	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
711		goto send;
712	if (SEQ_GT(tp->snd_up, tp->snd_una))
713		goto send;
714	/*
715	 * If our state indicates that FIN should be sent
716	 * and we have not yet done so, then we need to send.
717	 */
718	if (flags & TH_FIN &&
719	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
720		goto send;
721	/*
722	 * In SACK, it is possible for tcp_output to fail to send a segment
723	 * after the retransmission timer has been turned off.  Make sure
724	 * that the retransmission timer is set.
725	 */
726	if ((tp->t_flags & TF_SACK_PERMIT) &&
727	    SEQ_GT(tp->snd_max, tp->snd_una) &&
728	    !tcp_timer_active(tp, TT_REXMT) &&
729	    !tcp_timer_active(tp, TT_PERSIST)) {
730		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
731		goto just_return;
732	}
733	/*
734	 * TCP window updates are not reliable, rather a polling protocol
735	 * using ``persist'' packets is used to insure receipt of window
736	 * updates.  The three ``states'' for the output side are:
737	 *	idle			not doing retransmits or persists
738	 *	persisting		to move a small or zero window
739	 *	(re)transmitting	and thereby not persisting
740	 *
741	 * tcp_timer_active(tp, TT_PERSIST)
742	 *	is true when we are in persist state.
743	 * (tp->t_flags & TF_FORCEDATA)
744	 *	is set when we are called to send a persist packet.
745	 * tcp_timer_active(tp, TT_REXMT)
746	 *	is set when we are retransmitting
747	 * The output side is idle when both timers are zero.
748	 *
749	 * If send window is too small, there is data to transmit, and no
750	 * retransmit or persist is pending, then go to persist state.
751	 * If nothing happens soon, send when timer expires:
752	 * if window is nonzero, transmit what we can,
753	 * otherwise force out a byte.
754	 */
755	if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
756	    !tcp_timer_active(tp, TT_PERSIST)) {
757		tp->t_rxtshift = 0;
758		tcp_setpersist(tp);
759	}
760
761	/*
762	 * No reason to send a segment, just return.
763	 */
764just_return:
765	SOCKBUF_UNLOCK(&so->so_snd);
766	return (0);
767
768send:
769	SOCKBUF_LOCK_ASSERT(&so->so_snd);
770	if (len > 0) {
771		if (len >= tp->t_maxseg)
772			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
773		else
774			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
775	}
776	/*
777	 * Before ESTABLISHED, force sending of initial options
778	 * unless TCP set not to do any options.
779	 * NOTE: we assume that the IP/TCP header plus TCP options
780	 * always fit in a single mbuf, leaving room for a maximum
781	 * link header, i.e.
782	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
783	 */
784	optlen = 0;
785#ifdef INET6
786	if (isipv6)
787		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
788	else
789#endif
790		hdrlen = sizeof (struct tcpiphdr);
791
792	if (flags & TH_SYN) {
793		tp->snd_nxt = tp->iss;
794	}
795
796	/*
797	 * Compute options for segment.
798	 * We only have to care about SYN and established connection
799	 * segments.  Options for SYN-ACK segments are handled in TCP
800	 * syncache.
801	 */
802	to.to_flags = 0;
803	if ((tp->t_flags & TF_NOOPT) == 0) {
804		/* Maximum segment size. */
805		if (flags & TH_SYN) {
806			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
807			if (tp->t_port)
808				to.to_mss -= V_tcp_udp_tunneling_overhead;
809			to.to_flags |= TOF_MSS;
810
811			/*
812			 * On SYN or SYN|ACK transmits on TFO connections,
813			 * only include the TFO option if it is not a
814			 * retransmit, as the presence of the TFO option may
815			 * have caused the original SYN or SYN|ACK to have
816			 * been dropped by a middlebox.
817			 */
818			if (IS_FASTOPEN(tp->t_flags) &&
819			    (tp->t_rxtshift == 0)) {
820				if (tp->t_state == TCPS_SYN_RECEIVED) {
821					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
822					to.to_tfo_cookie =
823					    (u_int8_t *)&tp->t_tfo_cookie.server;
824					to.to_flags |= TOF_FASTOPEN;
825					wanted_cookie = 1;
826				} else if (tp->t_state == TCPS_SYN_SENT) {
827					to.to_tfo_len =
828					    tp->t_tfo_client_cookie_len;
829					to.to_tfo_cookie =
830					    tp->t_tfo_cookie.client;
831					to.to_flags |= TOF_FASTOPEN;
832					wanted_cookie = 1;
833					/*
834					 * If we wind up having more data to
835					 * send with the SYN than can fit in
836					 * one segment, don't send any more
837					 * until the SYN|ACK comes back from
838					 * the other end.
839					 */
840					dont_sendalot = 1;
841				}
842			}
843		}
844		/* Window scaling. */
845		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
846			to.to_wscale = tp->request_r_scale;
847			to.to_flags |= TOF_SCALE;
848		}
849		/* Timestamps. */
850		if ((tp->t_flags & TF_RCVD_TSTMP) ||
851		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
852			curticks = tcp_ts_getticks();
853			to.to_tsval = curticks + tp->ts_offset;
854			to.to_tsecr = tp->ts_recent;
855			to.to_flags |= TOF_TS;
856			if (tp->t_rxtshift == 1)
857				tp->t_badrxtwin = curticks;
858		}
859
860		/* Set receive buffer autosizing timestamp. */
861		if (tp->rfbuf_ts == 0 &&
862		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
863			tp->rfbuf_ts = tcp_ts_getticks();
864
865		/* Selective ACK's. */
866		if (tp->t_flags & TF_SACK_PERMIT) {
867			if (flags & TH_SYN)
868				to.to_flags |= TOF_SACKPERM;
869			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
870			    tp->rcv_numsacks > 0) {
871				to.to_flags |= TOF_SACK;
872				to.to_nsacks = tp->rcv_numsacks;
873				to.to_sacks = (u_char *)tp->sackblks;
874			}
875		}
876#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
877		/* TCP-MD5 (RFC2385). */
878		/*
879		 * Check that TCP_MD5SIG is enabled in tcpcb to
880		 * account the size needed to set this TCP option.
881		 */
882		if (tp->t_flags & TF_SIGNATURE)
883			to.to_flags |= TOF_SIGNATURE;
884#endif /* TCP_SIGNATURE */
885
886		/* Processing the options. */
887		hdrlen += optlen = tcp_addoptions(&to, opt);
888		/*
889		 * If we wanted a TFO option to be added, but it was unable
890		 * to fit, ensure no data is sent.
891		 */
892		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
893		    !(to.to_flags & TOF_FASTOPEN))
894			len = 0;
895	}
896	if (tp->t_port) {
897		if (V_tcp_udp_tunneling_port == 0) {
898			/* The port was removed?? */
899			SOCKBUF_UNLOCK(&so->so_snd);
900			return (EHOSTUNREACH);
901		}
902		hdrlen += sizeof(struct udphdr);
903	}
904	/*
905	 * Adjust data length if insertion of options will
906	 * bump the packet length beyond the t_maxseg length.
907	 * Clear the FIN bit because we cut off the tail of
908	 * the segment.
909	 */
910	if (len + optlen + ipoptlen > tp->t_maxseg) {
911		flags &= ~TH_FIN;
912
913		if (tso) {
914			u_int if_hw_tsomax;
915			u_int moff;
916			int max_len;
917
918			/* extract TSO information */
919			if_hw_tsomax = tp->t_tsomax;
920			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
921			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
922
923			/*
924			 * Limit a TSO burst to prevent it from
925			 * overflowing or exceeding the maximum length
926			 * allowed by the network interface:
927			 */
928			KASSERT(ipoptlen == 0,
929			    ("%s: TSO can't do IP options", __func__));
930
931			/*
932			 * Check if we should limit by maximum payload
933			 * length:
934			 */
935			if (if_hw_tsomax != 0) {
936				/* compute maximum TSO length */
937				max_len = (if_hw_tsomax - hdrlen -
938				    max_linkhdr);
939				if (max_len <= 0) {
940					len = 0;
941				} else if (len > max_len) {
942					sendalot = 1;
943					len = max_len;
944				}
945			}
946
947			/*
948			 * Prevent the last segment from being
949			 * fractional unless the send sockbuf can be
950			 * emptied:
951			 */
952			max_len = (tp->t_maxseg - optlen);
953			if (((uint32_t)off + (uint32_t)len) <
954			    sbavail(&so->so_snd)) {
955				moff = len % max_len;
956				if (moff != 0) {
957					len -= moff;
958					sendalot = 1;
959				}
960			}
961
962			/*
963			 * In case there are too many small fragments
964			 * don't use TSO:
965			 */
966			if (len <= max_len) {
967				len = max_len;
968				sendalot = 1;
969				tso = 0;
970			}
971
972			/*
973			 * Send the FIN in a separate segment
974			 * after the bulk sending is done.
975			 * We don't trust the TSO implementations
976			 * to clear the FIN flag on all but the
977			 * last segment.
978			 */
979			if (tp->t_flags & TF_NEEDFIN)
980				sendalot = 1;
981		} else {
982			if (optlen + ipoptlen >= tp->t_maxseg) {
983				/*
984				 * Since we don't have enough space to put
985				 * the IP header chain and the TCP header in
986				 * one packet as required by RFC 7112, don't
987				 * send it. Also ensure that at least one
988				 * byte of the payload can be put into the
989				 * TCP segment.
990				 */
991				SOCKBUF_UNLOCK(&so->so_snd);
992				error = EMSGSIZE;
993				sack_rxmit = 0;
994				goto out;
995			}
996			len = tp->t_maxseg - optlen - ipoptlen;
997			sendalot = 1;
998			if (dont_sendalot)
999				sendalot = 0;
1000		}
1001	} else
1002		tso = 0;
1003
1004	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1005	    ("%s: len > IP_MAXPACKET", __func__));
1006
1007/*#ifdef DIAGNOSTIC*/
1008#ifdef INET6
1009	if (max_linkhdr + hdrlen > MCLBYTES)
1010#else
1011	if (max_linkhdr + hdrlen > MHLEN)
1012#endif
1013		panic("tcphdr too big");
1014/*#endif*/
1015
1016	/*
1017	 * This KASSERT is here to catch edge cases at a well defined place.
1018	 * Before, those had triggered (random) panic conditions further down.
1019	 */
1020	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1021
1022	/*
1023	 * Grab a header mbuf, attaching a copy of data to
1024	 * be transmitted, and initialize the header from
1025	 * the template for sends on this connection.
1026	 */
1027	if (len) {
1028		struct mbuf *mb;
1029		struct sockbuf *msb;
1030		u_int moff;
1031
1032		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1033			TCPSTAT_INC(tcps_sndprobe);
1034#ifdef STATS
1035			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1036				stats_voi_update_abs_u32(tp->t_stats,
1037				VOI_TCP_RETXPB, len);
1038			else
1039				stats_voi_update_abs_u64(tp->t_stats,
1040				    VOI_TCP_TXPB, len);
1041#endif /* STATS */
1042		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1043			tp->t_sndrexmitpack++;
1044			TCPSTAT_INC(tcps_sndrexmitpack);
1045			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1046#ifdef STATS
1047			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1048			    len);
1049#endif /* STATS */
1050		} else {
1051			TCPSTAT_INC(tcps_sndpack);
1052			TCPSTAT_ADD(tcps_sndbyte, len);
1053#ifdef STATS
1054			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1055			    len);
1056#endif /* STATS */
1057		}
1058#ifdef INET6
1059		if (MHLEN < hdrlen + max_linkhdr)
1060			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1061		else
1062#endif
1063			m = m_gethdr(M_NOWAIT, MT_DATA);
1064
1065		if (m == NULL) {
1066			SOCKBUF_UNLOCK(&so->so_snd);
1067			error = ENOBUFS;
1068			sack_rxmit = 0;
1069			goto out;
1070		}
1071
1072		m->m_data += max_linkhdr;
1073		m->m_len = hdrlen;
1074
1075		/*
1076		 * Start the m_copy functions from the closest mbuf
1077		 * to the offset in the socket buffer chain.
1078		 */
1079		mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1080		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1081			m_copydata(mb, moff, len,
1082			    mtod(m, caddr_t) + hdrlen);
1083			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1084				sbsndptr_adv(&so->so_snd, mb, len);
1085			m->m_len += len;
1086		} else {
1087			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1088				msb = NULL;
1089			else
1090				msb = &so->so_snd;
1091			m->m_next = tcp_m_copym(mb, moff,
1092			    &len, if_hw_tsomaxsegcount,
1093			    if_hw_tsomaxsegsize, msb, hw_tls);
1094			if (len <= (tp->t_maxseg - optlen)) {
1095				/*
1096				 * Must have ran out of mbufs for the copy
1097				 * shorten it to no longer need tso. Lets
1098				 * not put on sendalot since we are low on
1099				 * mbufs.
1100				 */
1101				tso = 0;
1102			}
1103			if (m->m_next == NULL) {
1104				SOCKBUF_UNLOCK(&so->so_snd);
1105				(void) m_free(m);
1106				error = ENOBUFS;
1107				sack_rxmit = 0;
1108				goto out;
1109			}
1110		}
1111
1112		/*
1113		 * If we're sending everything we've got, set PUSH.
1114		 * (This will keep happy those implementations which only
1115		 * give data to the user when a buffer fills or
1116		 * a PUSH comes in.)
1117		 */
1118		if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1119		    !(flags & TH_SYN))
1120			flags |= TH_PUSH;
1121		SOCKBUF_UNLOCK(&so->so_snd);
1122	} else {
1123		SOCKBUF_UNLOCK(&so->so_snd);
1124		if (tp->t_flags & TF_ACKNOW)
1125			TCPSTAT_INC(tcps_sndacks);
1126		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1127			TCPSTAT_INC(tcps_sndctrl);
1128		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1129			TCPSTAT_INC(tcps_sndurg);
1130		else
1131			TCPSTAT_INC(tcps_sndwinup);
1132
1133		m = m_gethdr(M_NOWAIT, MT_DATA);
1134		if (m == NULL) {
1135			error = ENOBUFS;
1136			sack_rxmit = 0;
1137			goto out;
1138		}
1139#ifdef INET6
1140		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1141		    MHLEN >= hdrlen) {
1142			M_ALIGN(m, hdrlen);
1143		} else
1144#endif
1145		m->m_data += max_linkhdr;
1146		m->m_len = hdrlen;
1147	}
1148	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1149	m->m_pkthdr.rcvif = (struct ifnet *)0;
1150#ifdef MAC
1151	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1152#endif
1153#ifdef INET6
1154	if (isipv6) {
1155		ip6 = mtod(m, struct ip6_hdr *);
1156		if (tp->t_port) {
1157			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1158			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1159			udp->uh_dport = tp->t_port;
1160			ulen = hdrlen + len - sizeof(struct ip6_hdr);
1161			udp->uh_ulen = htons(ulen);
1162			th = (struct tcphdr *)(udp + 1);
1163		} else {
1164			th = (struct tcphdr *)(ip6 + 1);
1165		}
1166		tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th);
1167	} else
1168#endif /* INET6 */
1169	{
1170		ip = mtod(m, struct ip *);
1171#ifdef TCPDEBUG
1172		ipov = (struct ipovly *)ip;
1173#endif
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(tp->t_inpcb, 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 == 1) {
1201		if (tp->t_rxtshift >= 1) {
1202			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1203				flags |= TH_ECE|TH_CWR;
1204		} else
1205			flags |= TH_ECE|TH_CWR;
1206	}
1207	/* Handle parallel SYN for ECN */
1208	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1209	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1210			flags |= TH_ECE;
1211			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1212	}
1213
1214	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1215	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
1216		/*
1217		 * If the peer has ECN, mark data packets with
1218		 * ECN capable transmission (ECT).
1219		 * Ignore pure ack packets, retransmissions and window probes.
1220		 */
1221		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1222		    (sack_rxmit == 0) &&
1223		    !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1224		    SEQ_LT(tp->snd_una, tp->snd_max))) {
1225#ifdef INET6
1226			if (isipv6)
1227				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1228			else
1229#endif
1230				ip->ip_tos |= IPTOS_ECN_ECT0;
1231			TCPSTAT_INC(tcps_ecn_ect0);
1232			/*
1233			 * Reply with proper ECN notifications.
1234			 * Only set CWR on new data segments.
1235			 */
1236			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1237				flags |= TH_CWR;
1238				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1239			}
1240		}
1241		if (tp->t_flags2 & TF2_ECN_SND_ECE)
1242			flags |= TH_ECE;
1243	}
1244
1245	/*
1246	 * If we are doing retransmissions, then snd_nxt will
1247	 * not reflect the first unsent octet.  For ACK only
1248	 * packets, we do not want the sequence number of the
1249	 * retransmitted packet, we want the sequence number
1250	 * of the next unsent octet.  So, if there is no data
1251	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1252	 * when filling in ti_seq.  But if we are in persist
1253	 * state, snd_max might reflect one byte beyond the
1254	 * right edge of the window, so use snd_nxt in that
1255	 * case, since we know we aren't doing a retransmission.
1256	 * (retransmit and persist are mutually exclusive...)
1257	 */
1258	if (sack_rxmit == 0) {
1259		if (len || (flags & (TH_SYN|TH_FIN)) ||
1260		    tcp_timer_active(tp, TT_PERSIST))
1261			th->th_seq = htonl(tp->snd_nxt);
1262		else
1263			th->th_seq = htonl(tp->snd_max);
1264	} else {
1265		th->th_seq = htonl(p->rxmit);
1266		p->rxmit += len;
1267		tp->sackhint.sack_bytes_rexmit += len;
1268	}
1269	if (IN_RECOVERY(tp->t_flags)) {
1270		/*
1271		 * Account all bytes transmitted while
1272		 * IN_RECOVERY, simplifying PRR and
1273		 * Lost Retransmit Detection
1274		 */
1275		tp->sackhint.prr_out += len;
1276	}
1277	th->th_ack = htonl(tp->rcv_nxt);
1278	if (optlen) {
1279		bcopy(opt, th + 1, optlen);
1280		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1281	}
1282	th->th_flags = flags;
1283	/*
1284	 * Calculate receive window.  Don't shrink window,
1285	 * but avoid silly window syndrome.
1286	 * If a RST segment is sent, advertise a window of zero.
1287	 */
1288	if (flags & TH_RST) {
1289		recwin = 0;
1290	} else {
1291		if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1292		    recwin < tp->t_maxseg)
1293			recwin = 0;
1294		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1295		    recwin < (tp->rcv_adv - tp->rcv_nxt))
1296			recwin = (tp->rcv_adv - tp->rcv_nxt);
1297	}
1298	/*
1299	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1300	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1301	 * case is handled in syncache.
1302	 */
1303	if (flags & TH_SYN)
1304		th->th_win = htons((u_short)
1305				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1306	else {
1307		/* Avoid shrinking window with window scaling. */
1308		recwin = roundup2(recwin, 1 << tp->rcv_scale);
1309		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1310	}
1311
1312	/*
1313	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1314	 * a 0 window.  This may cause the remote transmitter to stall.  This
1315	 * flag tells soreceive() to disable delayed acknowledgements when
1316	 * draining the buffer.  This can occur if the receiver is attempting
1317	 * to read more data than can be buffered prior to transmitting on
1318	 * the connection.
1319	 */
1320	if (th->th_win == 0) {
1321		tp->t_sndzerowin++;
1322		tp->t_flags |= TF_RXWIN0SENT;
1323	} else
1324		tp->t_flags &= ~TF_RXWIN0SENT;
1325	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1326		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1327		th->th_flags |= TH_URG;
1328	} else
1329		/*
1330		 * If no urgent pointer to send, then we pull
1331		 * the urgent pointer to the left edge of the send window
1332		 * so that it doesn't drift into the send window on sequence
1333		 * number wraparound.
1334		 */
1335		tp->snd_up = tp->snd_una;		/* drag it along */
1336
1337	/*
1338	 * Put TCP length in extended header, and then
1339	 * checksum extended header and data.
1340	 */
1341	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1342
1343#if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1344	if (to.to_flags & TOF_SIGNATURE) {
1345		/*
1346		 * Calculate MD5 signature and put it into the place
1347		 * determined before.
1348		 * NOTE: since TCP options buffer doesn't point into
1349		 * mbuf's data, calculate offset and use it.
1350		 */
1351		if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1352		    (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1353			/*
1354			 * Do not send segment if the calculation of MD5
1355			 * digest has failed.
1356			 */
1357			m_freem(m);
1358			goto out;
1359		}
1360	}
1361#endif
1362#ifdef INET6
1363	if (isipv6) {
1364		/*
1365		 * There is no need to fill in ip6_plen right now.
1366		 * It will be filled later by ip6_output.
1367		 */
1368		if (tp->t_port) {
1369			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1370			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1371			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1372			th->th_sum = htons(0);
1373			UDPSTAT_INC(udps_opackets);
1374		} else {
1375			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1376			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1377			th->th_sum = in6_cksum_pseudo(ip6,
1378			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1379			    0);
1380		}
1381	}
1382#endif
1383#if defined(INET6) && defined(INET)
1384	else
1385#endif
1386#ifdef INET
1387	{
1388		if (tp->t_port) {
1389			m->m_pkthdr.csum_flags = CSUM_UDP;
1390			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1391			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1392			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1393			th->th_sum = htons(0);
1394			UDPSTAT_INC(udps_opackets);
1395		} else {
1396			m->m_pkthdr.csum_flags = CSUM_TCP;
1397			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1398			th->th_sum = in_pseudo(ip->ip_src.s_addr,
1399			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1400			    IPPROTO_TCP + len + optlen));
1401		}
1402
1403		/* IP version must be set here for ipv4/ipv6 checking later */
1404		KASSERT(ip->ip_v == IPVERSION,
1405		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1406	}
1407#endif
1408
1409	/*
1410	 * Enable TSO and specify the size of the segments.
1411	 * The TCP pseudo header checksum is always provided.
1412	 */
1413	if (tso) {
1414		KASSERT(len > tp->t_maxseg - optlen,
1415		    ("%s: len <= tso_segsz", __func__));
1416		m->m_pkthdr.csum_flags |= CSUM_TSO;
1417		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1418	}
1419
1420	KASSERT(len + hdrlen == m_length(m, NULL),
1421	    ("%s: mbuf chain shorter than expected: %d + %u != %u",
1422	    __func__, len, hdrlen, m_length(m, NULL)));
1423
1424#ifdef TCP_HHOOK
1425	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1426	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1427#endif
1428
1429#ifdef TCPDEBUG
1430	/*
1431	 * Trace.
1432	 */
1433	if (so->so_options & SO_DEBUG) {
1434		u_short save = 0;
1435#ifdef INET6
1436		if (!isipv6)
1437#endif
1438		{
1439			save = ipov->ih_len;
1440			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1441		}
1442		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1443#ifdef INET6
1444		if (!isipv6)
1445#endif
1446		ipov->ih_len = save;
1447	}
1448#endif /* TCPDEBUG */
1449	TCP_PROBE3(debug__output, tp, th, m);
1450
1451	/* We're getting ready to send; log now. */
1452	TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1453	    len, NULL, false);
1454
1455	/*
1456	 * Fill in IP length and desired time to live and
1457	 * send to IP level.  There should be a better way
1458	 * to handle ttl and tos; we could keep them in
1459	 * the template, but need a way to checksum without them.
1460	 */
1461	/*
1462	 * m->m_pkthdr.len should have been set before checksum calculation,
1463	 * because in6_cksum() need it.
1464	 */
1465#ifdef INET6
1466	if (isipv6) {
1467		/*
1468		 * we separately set hoplimit for every segment, since the
1469		 * user might want to change the value via setsockopt.
1470		 * Also, desired default hop limit might be changed via
1471		 * Neighbor Discovery.
1472		 */
1473		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1474
1475		/*
1476		 * Set the packet size here for the benefit of DTrace probes.
1477		 * ip6_output() will set it properly; it's supposed to include
1478		 * the option header lengths as well.
1479		 */
1480		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1481
1482		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1483			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1484		else
1485			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1486
1487		if (tp->t_state == TCPS_SYN_SENT)
1488			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1489
1490		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1491
1492#ifdef TCPPCAP
1493		/* Save packet, if requested. */
1494		tcp_pcap_add(th, m, &(tp->t_outpkts));
1495#endif
1496
1497		/* TODO: IPv6 IP6TOS_ECT bit on */
1498		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1499		    &tp->t_inpcb->inp_route6,
1500		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1501		    NULL, NULL, tp->t_inpcb);
1502
1503		if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1504			mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1505	}
1506#endif /* INET6 */
1507#if defined(INET) && defined(INET6)
1508	else
1509#endif
1510#ifdef INET
1511    {
1512	ip->ip_len = htons(m->m_pkthdr.len);
1513#ifdef INET6
1514	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1515		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1516#endif /* INET6 */
1517	/*
1518	 * If we do path MTU discovery, then we set DF on every packet.
1519	 * This might not be the best thing to do according to RFC3390
1520	 * Section 2. However the tcp hostcache migitates the problem
1521	 * so it affects only the first tcp connection with a host.
1522	 *
1523	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1524	 */
1525	if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1526		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1527		if (tp->t_port == 0 || len < V_tcp_minmss) {
1528			ip->ip_off |= htons(IP_DF);
1529		}
1530	} else {
1531		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1532	}
1533
1534	if (tp->t_state == TCPS_SYN_SENT)
1535		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1536
1537	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1538
1539#ifdef TCPPCAP
1540	/* Save packet, if requested. */
1541	tcp_pcap_add(th, m, &(tp->t_outpkts));
1542#endif
1543
1544	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1545	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1546	    tp->t_inpcb);
1547
1548	if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1549		mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1550    }
1551#endif /* INET */
1552
1553out:
1554	if (error == 0)
1555		tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0);
1556	/*
1557	 * In transmit state, time the transmission and arrange for
1558	 * the retransmit.  In persist state, just set snd_max.
1559	 */
1560	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1561	    !tcp_timer_active(tp, TT_PERSIST)) {
1562		tcp_seq startseq = tp->snd_nxt;
1563
1564		/*
1565		 * Advance snd_nxt over sequence space of this segment.
1566		 */
1567		if (flags & (TH_SYN|TH_FIN)) {
1568			if (flags & TH_SYN)
1569				tp->snd_nxt++;
1570			if (flags & TH_FIN) {
1571				tp->snd_nxt++;
1572				tp->t_flags |= TF_SENTFIN;
1573			}
1574		}
1575		if (sack_rxmit)
1576			goto timer;
1577		tp->snd_nxt += len;
1578		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1579			tp->snd_max = tp->snd_nxt;
1580			/*
1581			 * Time this transmission if not a retransmission and
1582			 * not currently timing anything.
1583			 */
1584			tp->t_sndtime = ticks;
1585			if (tp->t_rtttime == 0) {
1586				tp->t_rtttime = ticks;
1587				tp->t_rtseq = startseq;
1588				TCPSTAT_INC(tcps_segstimed);
1589			}
1590#ifdef STATS
1591			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1592				tp->t_flags |= TF_GPUTINPROG;
1593				tp->gput_seq = startseq;
1594				tp->gput_ack = startseq +
1595				    ulmin(sbavail(&so->so_snd) - off, sendwin);
1596				tp->gput_ts = tcp_ts_getticks();
1597			}
1598#endif /* STATS */
1599		}
1600
1601		/*
1602		 * Set retransmit timer if not currently set,
1603		 * and not doing a pure ack or a keep-alive probe.
1604		 * Initial value for retransmit timer is smoothed
1605		 * round-trip time + 2 * round-trip time variance.
1606		 * Initialize shift counter which is used for backoff
1607		 * of retransmit time.
1608		 */
1609timer:
1610		if (!tcp_timer_active(tp, TT_REXMT) &&
1611		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1612		     (tp->snd_nxt != tp->snd_una))) {
1613			if (tcp_timer_active(tp, TT_PERSIST)) {
1614				tcp_timer_activate(tp, TT_PERSIST, 0);
1615				tp->t_rxtshift = 0;
1616			}
1617			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1618		} else if (len == 0 && sbavail(&so->so_snd) &&
1619		    !tcp_timer_active(tp, TT_REXMT) &&
1620		    !tcp_timer_active(tp, TT_PERSIST)) {
1621			/*
1622			 * Avoid a situation where we do not set persist timer
1623			 * after a zero window condition. For example:
1624			 * 1) A -> B: packet with enough data to fill the window
1625			 * 2) B -> A: ACK for #1 + new data (0 window
1626			 *    advertisement)
1627			 * 3) A -> B: ACK for #2, 0 len packet
1628			 *
1629			 * In this case, A will not activate the persist timer,
1630			 * because it chose to send a packet. Unless tcp_output
1631			 * is called for some other reason (delayed ack timer,
1632			 * another input packet from B, socket syscall), A will
1633			 * not send zero window probes.
1634			 *
1635			 * So, if you send a 0-length packet, but there is data
1636			 * in the socket buffer, and neither the rexmt or
1637			 * persist timer is already set, then activate the
1638			 * persist timer.
1639			 */
1640			tp->t_rxtshift = 0;
1641			tcp_setpersist(tp);
1642		}
1643	} else {
1644		/*
1645		 * Persist case, update snd_max but since we are in
1646		 * persist mode (no window) we do not update snd_nxt.
1647		 */
1648		int xlen = len;
1649		if (flags & TH_SYN)
1650			++xlen;
1651		if (flags & TH_FIN) {
1652			++xlen;
1653			tp->t_flags |= TF_SENTFIN;
1654		}
1655		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1656			tp->snd_max = tp->snd_nxt + xlen;
1657	}
1658	if ((error == 0) &&
1659	    (TCPS_HAVEESTABLISHED(tp->t_state) &&
1660	     (tp->t_flags & TF_SACK_PERMIT) &&
1661	     tp->rcv_numsacks > 0)) {
1662		    /* Clean up any DSACK's sent */
1663		    tcp_clean_dsack_blocks(tp);
1664	}
1665	if (error) {
1666		/* Record the error. */
1667		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1668		    error, 0, NULL, false);
1669
1670		/*
1671		 * We know that the packet was lost, so back out the
1672		 * sequence number advance, if any.
1673		 *
1674		 * If the error is EPERM the packet got blocked by the
1675		 * local firewall.  Normally we should terminate the
1676		 * connection but the blocking may have been spurious
1677		 * due to a firewall reconfiguration cycle.  So we treat
1678		 * it like a packet loss and let the retransmit timer and
1679		 * timeouts do their work over time.
1680		 * XXX: It is a POLA question whether calling tcp_drop right
1681		 * away would be the really correct behavior instead.
1682		 */
1683		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1684		    !tcp_timer_active(tp, TT_PERSIST)) &&
1685		    ((flags & TH_SYN) == 0) &&
1686		    (error != EPERM)) {
1687			if (sack_rxmit) {
1688				p->rxmit -= len;
1689				tp->sackhint.sack_bytes_rexmit -= len;
1690				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1691				    ("sackhint bytes rtx >= 0"));
1692			} else
1693				tp->snd_nxt -= len;
1694		}
1695		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1696		switch (error) {
1697		case EACCES:
1698		case EPERM:
1699			tp->t_softerror = error;
1700			return (error);
1701		case ENOBUFS:
1702			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1703			tp->snd_cwnd = tp->t_maxseg;
1704			return (0);
1705		case EMSGSIZE:
1706			/*
1707			 * For some reason the interface we used initially
1708			 * to send segments changed to another or lowered
1709			 * its MTU.
1710			 * If TSO was active we either got an interface
1711			 * without TSO capabilits or TSO was turned off.
1712			 * If we obtained mtu from ip_output() then update
1713			 * it and try again.
1714			 */
1715			if (tso)
1716				tp->t_flags &= ~TF_TSO;
1717			if (mtu != 0) {
1718				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1719				goto again;
1720			}
1721			return (error);
1722		case EHOSTDOWN:
1723		case EHOSTUNREACH:
1724		case ENETDOWN:
1725		case ENETUNREACH:
1726			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1727				tp->t_softerror = error;
1728				return (0);
1729			}
1730			/* FALLTHROUGH */
1731		default:
1732			return (error);
1733		}
1734	}
1735	TCPSTAT_INC(tcps_sndtotal);
1736
1737	/*
1738	 * Data sent (as far as we can tell).
1739	 * If this advertises a larger window than any other segment,
1740	 * then remember the size of the advertised window.
1741	 * Any pending ACK has now been sent.
1742	 */
1743	if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1744		tp->rcv_adv = tp->rcv_nxt + recwin;
1745	tp->last_ack_sent = tp->rcv_nxt;
1746	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1747	if (tcp_timer_active(tp, TT_DELACK))
1748		tcp_timer_activate(tp, TT_DELACK, 0);
1749#if 0
1750	/*
1751	 * This completely breaks TCP if newreno is turned on.  What happens
1752	 * is that if delayed-acks are turned on on the receiver, this code
1753	 * on the transmitter effectively destroys the TCP window, forcing
1754	 * it to four packets (1.5Kx4 = 6K window).
1755	 */
1756	if (sendalot && --maxburst)
1757		goto again;
1758#endif
1759	if (sendalot)
1760		goto again;
1761	return (0);
1762}
1763
1764void
1765tcp_setpersist(struct tcpcb *tp)
1766{
1767	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1768	int tt;
1769
1770	tp->t_flags &= ~TF_PREVVALID;
1771	if (tcp_timer_active(tp, TT_REXMT))
1772		panic("tcp_setpersist: retransmit pending");
1773	/*
1774	 * Start/restart persistence timer.
1775	 */
1776	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1777		      tcp_persmin, tcp_persmax);
1778	tcp_timer_activate(tp, TT_PERSIST, tt);
1779	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1780		tp->t_rxtshift++;
1781}
1782
1783/*
1784 * Insert TCP options according to the supplied parameters to the place
1785 * optp in a consistent way.  Can handle unaligned destinations.
1786 *
1787 * The order of the option processing is crucial for optimal packing and
1788 * alignment for the scarce option space.
1789 *
1790 * The optimal order for a SYN/SYN-ACK segment is:
1791 *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1792 *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1793 *
1794 * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1795 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1796 * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1797 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1798 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1799 */
1800int
1801tcp_addoptions(struct tcpopt *to, u_char *optp)
1802{
1803	u_int32_t mask, optlen = 0;
1804
1805	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1806		if ((to->to_flags & mask) != mask)
1807			continue;
1808		if (optlen == TCP_MAXOLEN)
1809			break;
1810		switch (to->to_flags & mask) {
1811		case TOF_MSS:
1812			while (optlen % 4) {
1813				optlen += TCPOLEN_NOP;
1814				*optp++ = TCPOPT_NOP;
1815			}
1816			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1817				continue;
1818			optlen += TCPOLEN_MAXSEG;
1819			*optp++ = TCPOPT_MAXSEG;
1820			*optp++ = TCPOLEN_MAXSEG;
1821			to->to_mss = htons(to->to_mss);
1822			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1823			optp += sizeof(to->to_mss);
1824			break;
1825		case TOF_SCALE:
1826			while (!optlen || optlen % 2 != 1) {
1827				optlen += TCPOLEN_NOP;
1828				*optp++ = TCPOPT_NOP;
1829			}
1830			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1831				continue;
1832			optlen += TCPOLEN_WINDOW;
1833			*optp++ = TCPOPT_WINDOW;
1834			*optp++ = TCPOLEN_WINDOW;
1835			*optp++ = to->to_wscale;
1836			break;
1837		case TOF_SACKPERM:
1838			while (optlen % 2) {
1839				optlen += TCPOLEN_NOP;
1840				*optp++ = TCPOPT_NOP;
1841			}
1842			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1843				continue;
1844			optlen += TCPOLEN_SACK_PERMITTED;
1845			*optp++ = TCPOPT_SACK_PERMITTED;
1846			*optp++ = TCPOLEN_SACK_PERMITTED;
1847			break;
1848		case TOF_TS:
1849			while (!optlen || optlen % 4 != 2) {
1850				optlen += TCPOLEN_NOP;
1851				*optp++ = TCPOPT_NOP;
1852			}
1853			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1854				continue;
1855			optlen += TCPOLEN_TIMESTAMP;
1856			*optp++ = TCPOPT_TIMESTAMP;
1857			*optp++ = TCPOLEN_TIMESTAMP;
1858			to->to_tsval = htonl(to->to_tsval);
1859			to->to_tsecr = htonl(to->to_tsecr);
1860			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1861			optp += sizeof(to->to_tsval);
1862			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1863			optp += sizeof(to->to_tsecr);
1864			break;
1865		case TOF_SIGNATURE:
1866			{
1867			int siglen = TCPOLEN_SIGNATURE - 2;
1868
1869			while (!optlen || optlen % 4 != 2) {
1870				optlen += TCPOLEN_NOP;
1871				*optp++ = TCPOPT_NOP;
1872			}
1873			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1874				to->to_flags &= ~TOF_SIGNATURE;
1875				continue;
1876			}
1877			optlen += TCPOLEN_SIGNATURE;
1878			*optp++ = TCPOPT_SIGNATURE;
1879			*optp++ = TCPOLEN_SIGNATURE;
1880			to->to_signature = optp;
1881			while (siglen--)
1882				 *optp++ = 0;
1883			break;
1884			}
1885		case TOF_SACK:
1886			{
1887			int sackblks = 0;
1888			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1889			tcp_seq sack_seq;
1890
1891			while (!optlen || optlen % 4 != 2) {
1892				optlen += TCPOLEN_NOP;
1893				*optp++ = TCPOPT_NOP;
1894			}
1895			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1896				continue;
1897			optlen += TCPOLEN_SACKHDR;
1898			*optp++ = TCPOPT_SACK;
1899			sackblks = min(to->to_nsacks,
1900					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1901			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1902			while (sackblks--) {
1903				sack_seq = htonl(sack->start);
1904				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1905				optp += sizeof(sack_seq);
1906				sack_seq = htonl(sack->end);
1907				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1908				optp += sizeof(sack_seq);
1909				optlen += TCPOLEN_SACK;
1910				sack++;
1911			}
1912			TCPSTAT_INC(tcps_sack_send_blocks);
1913			break;
1914			}
1915		case TOF_FASTOPEN:
1916			{
1917			int total_len;
1918
1919			/* XXX is there any point to aligning this option? */
1920			total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1921			if (TCP_MAXOLEN - optlen < total_len) {
1922				to->to_flags &= ~TOF_FASTOPEN;
1923				continue;
1924			}
1925			*optp++ = TCPOPT_FAST_OPEN;
1926			*optp++ = total_len;
1927			if (to->to_tfo_len > 0) {
1928				bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1929				optp += to->to_tfo_len;
1930			}
1931			optlen += total_len;
1932			break;
1933			}
1934		default:
1935			panic("%s: unknown TCP option type", __func__);
1936			break;
1937		}
1938	}
1939
1940	/* Terminate and pad TCP options to a 4 byte boundary. */
1941	if (optlen % 4) {
1942		optlen += TCPOLEN_EOL;
1943		*optp++ = TCPOPT_EOL;
1944	}
1945	/*
1946	 * According to RFC 793 (STD0007):
1947	 *   "The content of the header beyond the End-of-Option option
1948	 *    must be header padding (i.e., zero)."
1949	 *   and later: "The padding is composed of zeros."
1950	 */
1951	while (optlen % 4) {
1952		optlen += TCPOLEN_PAD;
1953		*optp++ = TCPOPT_PAD;
1954	}
1955
1956	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1957	return (optlen);
1958}
1959
1960/*
1961 * This is a copy of m_copym(), taking the TSO segment size/limit
1962 * constraints into account, and advancing the sndptr as it goes.
1963 */
1964struct mbuf *
1965tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1966    int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1967{
1968#ifdef KERN_TLS
1969	struct ktls_session *tls, *ntls;
1970	struct mbuf *start;
1971#endif
1972	struct mbuf *n, **np;
1973	struct mbuf *top;
1974	int32_t off = off0;
1975	int32_t len = *plen;
1976	int32_t fragsize;
1977	int32_t len_cp = 0;
1978	int32_t *pkthdrlen;
1979	uint32_t mlen, frags;
1980	bool copyhdr;
1981
1982	KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1983	KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1984	if (off == 0 && m->m_flags & M_PKTHDR)
1985		copyhdr = true;
1986	else
1987		copyhdr = false;
1988	while (off > 0) {
1989		KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1990		if (off < m->m_len)
1991			break;
1992		off -= m->m_len;
1993		if ((sb) && (m == sb->sb_sndptr)) {
1994			sb->sb_sndptroff += m->m_len;
1995			sb->sb_sndptr = m->m_next;
1996		}
1997		m = m->m_next;
1998	}
1999	np = &top;
2000	top = NULL;
2001	pkthdrlen = NULL;
2002#ifdef KERN_TLS
2003	if (hw_tls && (m->m_flags & M_EXTPG))
2004		tls = m->m_epg_tls;
2005	else
2006		tls = NULL;
2007	start = m;
2008#endif
2009	while (len > 0) {
2010		if (m == NULL) {
2011			KASSERT(len == M_COPYALL,
2012			    ("tcp_m_copym, length > size of mbuf chain"));
2013			*plen = len_cp;
2014			if (pkthdrlen != NULL)
2015				*pkthdrlen = len_cp;
2016			break;
2017		}
2018#ifdef KERN_TLS
2019		if (hw_tls) {
2020			if (m->m_flags & M_EXTPG)
2021				ntls = m->m_epg_tls;
2022			else
2023				ntls = NULL;
2024
2025			/*
2026			 * Avoid mixing TLS records with handshake
2027			 * data or TLS records from different
2028			 * sessions.
2029			 */
2030			if (tls != ntls) {
2031				MPASS(m != start);
2032				*plen = len_cp;
2033				if (pkthdrlen != NULL)
2034					*pkthdrlen = len_cp;
2035				break;
2036			}
2037		}
2038#endif
2039		mlen = min(len, m->m_len - off);
2040		if (seglimit) {
2041			/*
2042			 * For M_EXTPG mbufs, add 3 segments
2043			 * + 1 in case we are crossing page boundaries
2044			 * + 2 in case the TLS hdr/trailer are used
2045			 * It is cheaper to just add the segments
2046			 * than it is to take the cache miss to look
2047			 * at the mbuf ext_pgs state in detail.
2048			 */
2049			if (m->m_flags & M_EXTPG) {
2050				fragsize = min(segsize, PAGE_SIZE);
2051				frags = 3;
2052			} else {
2053				fragsize = segsize;
2054				frags = 0;
2055			}
2056
2057			/* Break if we really can't fit anymore. */
2058			if ((frags + 1) >= seglimit) {
2059				*plen =	len_cp;
2060				if (pkthdrlen != NULL)
2061					*pkthdrlen = len_cp;
2062				break;
2063			}
2064
2065			/*
2066			 * Reduce size if you can't copy the whole
2067			 * mbuf. If we can't copy the whole mbuf, also
2068			 * adjust len so the loop will end after this
2069			 * mbuf.
2070			 */
2071			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2072				mlen = (seglimit - frags - 1) * fragsize;
2073				len = mlen;
2074				*plen = len_cp + len;
2075				if (pkthdrlen != NULL)
2076					*pkthdrlen = *plen;
2077			}
2078			frags += howmany(mlen, fragsize);
2079			if (frags == 0)
2080				frags++;
2081			seglimit -= frags;
2082			KASSERT(seglimit > 0,
2083			    ("%s: seglimit went too low", __func__));
2084		}
2085		if (copyhdr)
2086			n = m_gethdr(M_NOWAIT, m->m_type);
2087		else
2088			n = m_get(M_NOWAIT, m->m_type);
2089		*np = n;
2090		if (n == NULL)
2091			goto nospace;
2092		if (copyhdr) {
2093			if (!m_dup_pkthdr(n, m, M_NOWAIT))
2094				goto nospace;
2095			if (len == M_COPYALL)
2096				n->m_pkthdr.len -= off0;
2097			else
2098				n->m_pkthdr.len = len;
2099			pkthdrlen = &n->m_pkthdr.len;
2100			copyhdr = false;
2101		}
2102		n->m_len = mlen;
2103		len_cp += n->m_len;
2104		if (m->m_flags & (M_EXT|M_EXTPG)) {
2105			n->m_data = m->m_data + off;
2106			mb_dupcl(n, m);
2107		} else
2108			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2109			    (u_int)n->m_len);
2110
2111		if (sb && (sb->sb_sndptr == m) &&
2112		    ((n->m_len + off) >= m->m_len) && m->m_next) {
2113			sb->sb_sndptroff += m->m_len;
2114			sb->sb_sndptr = m->m_next;
2115		}
2116		off = 0;
2117		if (len != M_COPYALL) {
2118			len -= n->m_len;
2119		}
2120		m = m->m_next;
2121		np = &n->m_next;
2122	}
2123	return (top);
2124nospace:
2125	m_freem(top);
2126	return (NULL);
2127}
2128
2129void
2130tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2131{
2132
2133	/*
2134	 * Automatic sizing of send socket buffer.  Often the send buffer
2135	 * size is not optimally adjusted to the actual network conditions
2136	 * at hand (delay bandwidth product).  Setting the buffer size too
2137	 * small limits throughput on links with high bandwidth and high
2138	 * delay (eg. trans-continental/oceanic links).  Setting the
2139	 * buffer size too big consumes too much real kernel memory,
2140	 * especially with many connections on busy servers.
2141	 *
2142	 * The criteria to step up the send buffer one notch are:
2143	 *  1. receive window of remote host is larger than send buffer
2144	 *     (with a fudge factor of 5/4th);
2145	 *  2. send buffer is filled to 7/8th with data (so we actually
2146	 *     have data to make use of it);
2147	 *  3. send buffer fill has not hit maximal automatic size;
2148	 *  4. our send window (slow start and cogestion controlled) is
2149	 *     larger than sent but unacknowledged data in send buffer.
2150	 *
2151	 * The remote host receive window scaling factor may limit the
2152	 * growing of the send buffer before it reaches its allowed
2153	 * maximum.
2154	 *
2155	 * It scales directly with slow start or congestion window
2156	 * and does at most one step per received ACK.  This fast
2157	 * scaling has the drawback of growing the send buffer beyond
2158	 * what is strictly necessary to make full use of a given
2159	 * delay*bandwidth product.  However testing has shown this not
2160	 * to be much of an problem.  At worst we are trading wasting
2161	 * of available bandwidth (the non-use of it) for wasting some
2162	 * socket buffer memory.
2163	 *
2164	 * TODO: Shrink send buffer during idle periods together
2165	 * with congestion window.  Requires another timer.  Has to
2166	 * wait for upcoming tcp timer rewrite.
2167	 *
2168	 * XXXGL: should there be used sbused() or sbavail()?
2169	 */
2170	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2171		int lowat;
2172
2173		lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2174		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2175		    sbused(&so->so_snd) >=
2176		    (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2177		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2178		    sendwin >= (sbused(&so->so_snd) -
2179		    (tp->snd_nxt - tp->snd_una))) {
2180			if (!sbreserve_locked(&so->so_snd,
2181			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2182			     V_tcp_autosndbuf_max), so, curthread))
2183				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2184		}
2185	}
2186}
2187