xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 0b801f329d72f1dc48cfc2fe613e4817e3b87cbc)
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68  * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.60 2008/08/15 21:37:16 nth Exp $
69  */
70 
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
83 #include <sys/mbuf.h>
84 #ifdef INET6
85 #include <sys/domain.h>
86 #endif
87 #include <sys/proc.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/protosw.h>
91 #include <sys/random.h>
92 #include <sys/in_cksum.h>
93 #include <sys/ktr.h>
94 
95 #include <vm/vm_zone.h>
96 
97 #include <net/route.h>
98 #include <net/if.h>
99 #include <net/netisr.h>
100 
101 #define	_IP_VHL
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
112 #ifdef INET6
113 #include <netinet/icmp6.h>
114 #endif
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
122 #ifdef TCPDEBUG
123 #include <netinet/tcp_debug.h>
124 #endif
125 #include <netinet6/ip6protosw.h>
126 
127 #ifdef IPSEC
128 #include <netinet6/ipsec.h>
129 #ifdef INET6
130 #include <netinet6/ipsec6.h>
131 #endif
132 #endif
133 
134 #ifdef FAST_IPSEC
135 #include <netproto/ipsec/ipsec.h>
136 #ifdef INET6
137 #include <netproto/ipsec/ipsec6.h>
138 #endif
139 #define	IPSEC
140 #endif
141 
142 #include <sys/md5.h>
143 #include <sys/msgport2.h>
144 #include <machine/smp.h>
145 
146 #include <net/netmsg2.h>
147 
148 #if !defined(KTR_TCP)
149 #define KTR_TCP		KTR_ALL
150 #endif
151 KTR_INFO_MASTER(tcp);
152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
155 #define logtcp(name)	KTR_LOG(tcp_ ## name)
156 
157 struct inpcbinfo tcbinfo[MAXCPU];
158 struct tcpcbackqhead tcpcbackq[MAXCPU];
159 
160 int tcp_mssdflt = TCP_MSS;
161 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
162     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
163 
164 #ifdef INET6
165 int tcp_v6mssdflt = TCP6_MSS;
166 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
167     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
168 #endif
169 
170 #if 0
171 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
172 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
173     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
174 #endif
175 
176 int tcp_do_rfc1323 = 1;
177 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
178     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
179 
180 int tcp_do_rfc1644 = 0;
181 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
182     &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
183 
184 static int tcp_tcbhashsize = 0;
185 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
186      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
187 
188 static int do_tcpdrain = 1;
189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
190      "Enable tcp_drain routine for extra help when low on mbufs");
191 
192 /* XXX JH */
193 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
194     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
195 
196 static int icmp_may_rst = 1;
197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
198     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
199 
200 static int tcp_isn_reseed_interval = 0;
201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
202     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
203 
204 /*
205  * TCP bandwidth limiting sysctls.  Note that the default lower bound of
206  * 1024 exists only for debugging.  A good production default would be
207  * something like 6100.
208  */
209 static int tcp_inflight_enable = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
211     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
212 
213 static int tcp_inflight_debug = 0;
214 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
215     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
216 
217 static int tcp_inflight_min = 6144;
218 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
219     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
220 
221 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
222 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
223     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
224 
225 static int tcp_inflight_stab = 20;
226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
227     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
228 
229 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
230 static struct malloc_pipe tcptemp_mpipe;
231 
232 static void tcp_willblock(void);
233 static void tcp_cleartaocache (void);
234 static void tcp_notify (struct inpcb *, int);
235 
236 struct tcp_stats tcpstats_percpu[MAXCPU];
237 #ifdef SMP
238 static int
239 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
240 {
241 	int cpu, error = 0;
242 
243 	for (cpu = 0; cpu < ncpus; ++cpu) {
244 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
245 					sizeof(struct tcp_stats))))
246 			break;
247 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
248 				       sizeof(struct tcp_stats))))
249 			break;
250 	}
251 
252 	return (error);
253 }
254 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
255     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
256 #else
257 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
258     &tcpstat, tcp_stats, "TCP statistics");
259 #endif
260 
261 /*
262  * Target size of TCP PCB hash tables. Must be a power of two.
263  *
264  * Note that this can be overridden by the kernel environment
265  * variable net.inet.tcp.tcbhashsize
266  */
267 #ifndef TCBHASHSIZE
268 #define	TCBHASHSIZE	512
269 #endif
270 
271 /*
272  * This is the actual shape of what we allocate using the zone
273  * allocator.  Doing it this way allows us to protect both structures
274  * using the same generation count, and also eliminates the overhead
275  * of allocating tcpcbs separately.  By hiding the structure here,
276  * we avoid changing most of the rest of the code (although it needs
277  * to be changed, eventually, for greater efficiency).
278  */
279 #define	ALIGNMENT	32
280 #define	ALIGNM1		(ALIGNMENT - 1)
281 struct	inp_tp {
282 	union {
283 		struct	inpcb inp;
284 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
285 	} inp_tp_u;
286 	struct	tcpcb tcb;
287 	struct	callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
288 	struct	callout inp_tp_delack;
289 };
290 #undef ALIGNMENT
291 #undef ALIGNM1
292 
293 /*
294  * Tcp initialization
295  */
296 void
297 tcp_init(void)
298 {
299 	struct inpcbporthead *porthashbase;
300 	u_long porthashmask;
301 	struct vm_zone *ipi_zone;
302 	int hashsize = TCBHASHSIZE;
303 	int cpu;
304 
305 	/*
306 	 * note: tcptemp is used for keepalives, and it is ok for an
307 	 * allocation to fail so do not specify MPF_INT.
308 	 */
309 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
310 		    25, -1, 0, NULL);
311 
312 	tcp_ccgen = 1;
313 	tcp_cleartaocache();
314 
315 	tcp_delacktime = TCPTV_DELACK;
316 	tcp_keepinit = TCPTV_KEEP_INIT;
317 	tcp_keepidle = TCPTV_KEEP_IDLE;
318 	tcp_keepintvl = TCPTV_KEEPINTVL;
319 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
320 	tcp_msl = TCPTV_MSL;
321 	tcp_rexmit_min = TCPTV_MIN;
322 	tcp_rexmit_slop = TCPTV_CPU_VAR;
323 
324 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
325 	if (!powerof2(hashsize)) {
326 		kprintf("WARNING: TCB hash size not a power of 2\n");
327 		hashsize = 512; /* safe default */
328 	}
329 	tcp_tcbhashsize = hashsize;
330 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
331 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
332 			 ZONE_INTERRUPT, 0);
333 
334 	for (cpu = 0; cpu < ncpus2; cpu++) {
335 		in_pcbinfo_init(&tcbinfo[cpu]);
336 		tcbinfo[cpu].cpu = cpu;
337 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
338 		    &tcbinfo[cpu].hashmask);
339 		tcbinfo[cpu].porthashbase = porthashbase;
340 		tcbinfo[cpu].porthashmask = porthashmask;
341 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
342 		    &tcbinfo[cpu].wildcardhashmask);
343 		tcbinfo[cpu].ipi_zone = ipi_zone;
344 		TAILQ_INIT(&tcpcbackq[cpu]);
345 	}
346 
347 	tcp_reass_maxseg = nmbclusters / 16;
348 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
349 
350 #ifdef INET6
351 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
352 #else
353 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
354 #endif
355 	if (max_protohdr < TCP_MINPROTOHDR)
356 		max_protohdr = TCP_MINPROTOHDR;
357 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
358 		panic("tcp_init");
359 #undef TCP_MINPROTOHDR
360 
361 	/*
362 	 * Initialize TCP statistics counters for each CPU.
363 	 */
364 #ifdef SMP
365 	for (cpu = 0; cpu < ncpus; ++cpu) {
366 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
367 	}
368 #else
369 	bzero(&tcpstat, sizeof(struct tcp_stats));
370 #endif
371 
372 	syncache_init();
373 	tcp_thread_init();
374 }
375 
376 void
377 tcpmsg_service_loop(void *dummy)
378 {
379 	struct netmsg *msg;
380 
381 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
382 		do {
383 			logtcp(rxmsg);
384 			msg->nm_dispatch(msg);
385 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
386 		logtcp(delayed);
387 		tcp_willblock();
388 		logtcp(wait);
389 	}
390 }
391 
392 static void
393 tcp_willblock(void)
394 {
395 	struct tcpcb *tp;
396 	int cpu = mycpu->gd_cpuid;
397 
398 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
399 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
400 		tp->t_flags &= ~TF_ONOUTPUTQ;
401 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
402 		tcp_output(tp);
403 	}
404 }
405 
406 
407 /*
408  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
409  * tcp_template used to store this data in mbufs, but we now recopy it out
410  * of the tcpcb each time to conserve mbufs.
411  */
412 void
413 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
414 {
415 	struct inpcb *inp = tp->t_inpcb;
416 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
417 
418 #ifdef INET6
419 	if (inp->inp_vflag & INP_IPV6) {
420 		struct ip6_hdr *ip6;
421 
422 		ip6 = (struct ip6_hdr *)ip_ptr;
423 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
424 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
425 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
426 			(IPV6_VERSION & IPV6_VERSION_MASK);
427 		ip6->ip6_nxt = IPPROTO_TCP;
428 		ip6->ip6_plen = sizeof(struct tcphdr);
429 		ip6->ip6_src = inp->in6p_laddr;
430 		ip6->ip6_dst = inp->in6p_faddr;
431 		tcp_hdr->th_sum = 0;
432 	} else
433 #endif
434 	{
435 		struct ip *ip = (struct ip *) ip_ptr;
436 
437 		ip->ip_vhl = IP_VHL_BORING;
438 		ip->ip_tos = 0;
439 		ip->ip_len = 0;
440 		ip->ip_id = 0;
441 		ip->ip_off = 0;
442 		ip->ip_ttl = 0;
443 		ip->ip_sum = 0;
444 		ip->ip_p = IPPROTO_TCP;
445 		ip->ip_src = inp->inp_laddr;
446 		ip->ip_dst = inp->inp_faddr;
447 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
448 				    ip->ip_dst.s_addr,
449 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
450 	}
451 
452 	tcp_hdr->th_sport = inp->inp_lport;
453 	tcp_hdr->th_dport = inp->inp_fport;
454 	tcp_hdr->th_seq = 0;
455 	tcp_hdr->th_ack = 0;
456 	tcp_hdr->th_x2 = 0;
457 	tcp_hdr->th_off = 5;
458 	tcp_hdr->th_flags = 0;
459 	tcp_hdr->th_win = 0;
460 	tcp_hdr->th_urp = 0;
461 }
462 
463 /*
464  * Create template to be used to send tcp packets on a connection.
465  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
466  * use for this function is in keepalives, which use tcp_respond.
467  */
468 struct tcptemp *
469 tcp_maketemplate(struct tcpcb *tp)
470 {
471 	struct tcptemp *tmp;
472 
473 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
474 		return (NULL);
475 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
476 	return (tmp);
477 }
478 
479 void
480 tcp_freetemplate(struct tcptemp *tmp)
481 {
482 	mpipe_free(&tcptemp_mpipe, tmp);
483 }
484 
485 /*
486  * Send a single message to the TCP at address specified by
487  * the given TCP/IP header.  If m == NULL, then we make a copy
488  * of the tcpiphdr at ti and send directly to the addressed host.
489  * This is used to force keep alive messages out using the TCP
490  * template for a connection.  If flags are given then we send
491  * a message back to the TCP which originated the * segment ti,
492  * and discard the mbuf containing it and any other attached mbufs.
493  *
494  * In any case the ack and sequence number of the transmitted
495  * segment are as specified by the parameters.
496  *
497  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
498  */
499 void
500 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
501 	    tcp_seq ack, tcp_seq seq, int flags)
502 {
503 	int tlen;
504 	int win = 0;
505 	struct route *ro = NULL;
506 	struct route sro;
507 	struct ip *ip = ipgen;
508 	struct tcphdr *nth;
509 	int ipflags = 0;
510 	struct route_in6 *ro6 = NULL;
511 	struct route_in6 sro6;
512 	struct ip6_hdr *ip6 = ipgen;
513 #ifdef INET6
514 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
515 #else
516 	const boolean_t isipv6 = FALSE;
517 #endif
518 
519 	if (tp != NULL) {
520 		if (!(flags & TH_RST)) {
521 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
522 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
523 				win = (long)TCP_MAXWIN << tp->rcv_scale;
524 		}
525 		if (isipv6)
526 			ro6 = &tp->t_inpcb->in6p_route;
527 		else
528 			ro = &tp->t_inpcb->inp_route;
529 	} else {
530 		if (isipv6) {
531 			ro6 = &sro6;
532 			bzero(ro6, sizeof *ro6);
533 		} else {
534 			ro = &sro;
535 			bzero(ro, sizeof *ro);
536 		}
537 	}
538 	if (m == NULL) {
539 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
540 		if (m == NULL)
541 			return;
542 		tlen = 0;
543 		m->m_data += max_linkhdr;
544 		if (isipv6) {
545 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
546 			ip6 = mtod(m, struct ip6_hdr *);
547 			nth = (struct tcphdr *)(ip6 + 1);
548 		} else {
549 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
550 			ip = mtod(m, struct ip *);
551 			nth = (struct tcphdr *)(ip + 1);
552 		}
553 		bcopy(th, nth, sizeof(struct tcphdr));
554 		flags = TH_ACK;
555 	} else {
556 		m_freem(m->m_next);
557 		m->m_next = NULL;
558 		m->m_data = (caddr_t)ipgen;
559 		/* m_len is set later */
560 		tlen = 0;
561 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
562 		if (isipv6) {
563 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
564 			nth = (struct tcphdr *)(ip6 + 1);
565 		} else {
566 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
567 			nth = (struct tcphdr *)(ip + 1);
568 		}
569 		if (th != nth) {
570 			/*
571 			 * this is usually a case when an extension header
572 			 * exists between the IPv6 header and the
573 			 * TCP header.
574 			 */
575 			nth->th_sport = th->th_sport;
576 			nth->th_dport = th->th_dport;
577 		}
578 		xchg(nth->th_dport, nth->th_sport, n_short);
579 #undef xchg
580 	}
581 	if (isipv6) {
582 		ip6->ip6_flow = 0;
583 		ip6->ip6_vfc = IPV6_VERSION;
584 		ip6->ip6_nxt = IPPROTO_TCP;
585 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
586 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
587 	} else {
588 		tlen += sizeof(struct tcpiphdr);
589 		ip->ip_len = tlen;
590 		ip->ip_ttl = ip_defttl;
591 	}
592 	m->m_len = tlen;
593 	m->m_pkthdr.len = tlen;
594 	m->m_pkthdr.rcvif = (struct ifnet *) NULL;
595 	nth->th_seq = htonl(seq);
596 	nth->th_ack = htonl(ack);
597 	nth->th_x2 = 0;
598 	nth->th_off = sizeof(struct tcphdr) >> 2;
599 	nth->th_flags = flags;
600 	if (tp != NULL)
601 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
602 	else
603 		nth->th_win = htons((u_short)win);
604 	nth->th_urp = 0;
605 	if (isipv6) {
606 		nth->th_sum = 0;
607 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
608 					sizeof(struct ip6_hdr),
609 					tlen - sizeof(struct ip6_hdr));
610 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
611 					       (ro6 && ro6->ro_rt) ?
612 						ro6->ro_rt->rt_ifp : NULL);
613 	} else {
614 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
615 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
616 		m->m_pkthdr.csum_flags = CSUM_TCP;
617 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
618 	}
619 #ifdef TCPDEBUG
620 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
621 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
622 #endif
623 	if (isipv6) {
624 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
625 			   tp ? tp->t_inpcb : NULL);
626 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
627 			RTFREE(ro6->ro_rt);
628 			ro6->ro_rt = NULL;
629 		}
630 	} else {
631 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
632 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
633 			RTFREE(ro->ro_rt);
634 			ro->ro_rt = NULL;
635 		}
636 	}
637 }
638 
639 /*
640  * Create a new TCP control block, making an
641  * empty reassembly queue and hooking it to the argument
642  * protocol control block.  The `inp' parameter must have
643  * come from the zone allocator set up in tcp_init().
644  */
645 struct tcpcb *
646 tcp_newtcpcb(struct inpcb *inp)
647 {
648 	struct inp_tp *it;
649 	struct tcpcb *tp;
650 #ifdef INET6
651 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
652 #else
653 	const boolean_t isipv6 = FALSE;
654 #endif
655 
656 	it = (struct inp_tp *)inp;
657 	tp = &it->tcb;
658 	bzero(tp, sizeof(struct tcpcb));
659 	LIST_INIT(&tp->t_segq);
660 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
661 
662 	/* Set up our timeouts. */
663 	callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
664 	callout_init(tp->tt_persist = &it->inp_tp_persist);
665 	callout_init(tp->tt_keep = &it->inp_tp_keep);
666 	callout_init(tp->tt_2msl = &it->inp_tp_2msl);
667 	callout_init(tp->tt_delack = &it->inp_tp_delack);
668 
669 	if (tcp_do_rfc1323)
670 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
671 	if (tcp_do_rfc1644)
672 		tp->t_flags |= TF_REQ_CC;
673 	tp->t_inpcb = inp;	/* XXX */
674 	tp->t_state = TCPS_CLOSED;
675 	/*
676 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
677 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
678 	 * reasonable initial retransmit time.
679 	 */
680 	tp->t_srtt = TCPTV_SRTTBASE;
681 	tp->t_rttvar =
682 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
683 	tp->t_rttmin = tcp_rexmit_min;
684 	tp->t_rxtcur = TCPTV_RTOBASE;
685 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
686 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
687 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
688 	tp->t_rcvtime = ticks;
689 	/*
690 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
691 	 * because the socket may be bound to an IPv6 wildcard address,
692 	 * which may match an IPv4-mapped IPv6 address.
693 	 */
694 	inp->inp_ip_ttl = ip_defttl;
695 	inp->inp_ppcb = tp;
696 	tcp_sack_tcpcb_init(tp);
697 	return (tp);		/* XXX */
698 }
699 
700 /*
701  * Drop a TCP connection, reporting the specified error.
702  * If connection is synchronized, then send a RST to peer.
703  */
704 struct tcpcb *
705 tcp_drop(struct tcpcb *tp, int error)
706 {
707 	struct socket *so = tp->t_inpcb->inp_socket;
708 
709 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
710 		tp->t_state = TCPS_CLOSED;
711 		tcp_output(tp);
712 		tcpstat.tcps_drops++;
713 	} else
714 		tcpstat.tcps_conndrops++;
715 	if (error == ETIMEDOUT && tp->t_softerror)
716 		error = tp->t_softerror;
717 	so->so_error = error;
718 	return (tcp_close(tp));
719 }
720 
721 #ifdef SMP
722 
723 struct netmsg_remwildcard {
724 	struct netmsg		nm_netmsg;
725 	struct inpcb		*nm_inp;
726 	struct inpcbinfo	*nm_pcbinfo;
727 #if defined(INET6)
728 	int			nm_isinet6;
729 #else
730 	int			nm_unused01;
731 #endif
732 };
733 
734 /*
735  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
736  * inp can be detached.  We do this by cycling through the cpus, ending up
737  * on the cpu controlling the inp last and then doing the disconnect.
738  */
739 static void
740 in_pcbremwildcardhash_handler(struct netmsg *msg0)
741 {
742 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
743 	int cpu;
744 
745 	cpu = msg->nm_pcbinfo->cpu;
746 
747 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
748 		/* note: detach removes any wildcard hash entry */
749 #ifdef INET6
750 		if (msg->nm_isinet6)
751 			in6_pcbdetach(msg->nm_inp);
752 		else
753 #endif
754 			in_pcbdetach(msg->nm_inp);
755 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
756 	} else {
757 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
758 		cpu = (cpu + 1) % ncpus2;
759 		msg->nm_pcbinfo = &tcbinfo[cpu];
760 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
761 	}
762 }
763 
764 #endif
765 
766 /*
767  * Close a TCP control block:
768  *	discard all space held by the tcp
769  *	discard internet protocol block
770  *	wake up any sleepers
771  */
772 struct tcpcb *
773 tcp_close(struct tcpcb *tp)
774 {
775 	struct tseg_qent *q;
776 	struct inpcb *inp = tp->t_inpcb;
777 	struct socket *so = inp->inp_socket;
778 	struct rtentry *rt;
779 	boolean_t dosavessthresh;
780 #ifdef SMP
781 	int cpu;
782 #endif
783 #ifdef INET6
784 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
785 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
786 #else
787 	const boolean_t isipv6 = FALSE;
788 #endif
789 
790 	/*
791 	 * The tp is not instantly destroyed in the wildcard case.  Setting
792 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
793 	 * messing with it, though it should be noted that this change may
794 	 * not take effect on other cpus until we have chained the wildcard
795 	 * hash removal.
796 	 *
797 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
798 	 * update and prevent other tcp protocol threads from accepting new
799 	 * connections on the listen socket we might be trying to close down.
800 	 */
801 	KKASSERT(tp->t_state != TCPS_TERMINATING);
802 	tp->t_state = TCPS_TERMINATING;
803 
804 	/*
805 	 * Make sure that all of our timers are stopped before we
806 	 * delete the PCB.
807 	 */
808 	callout_stop(tp->tt_rexmt);
809 	callout_stop(tp->tt_persist);
810 	callout_stop(tp->tt_keep);
811 	callout_stop(tp->tt_2msl);
812 	callout_stop(tp->tt_delack);
813 
814 	if (tp->t_flags & TF_ONOUTPUTQ) {
815 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
816 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
817 		tp->t_flags &= ~TF_ONOUTPUTQ;
818 	}
819 
820 	/*
821 	 * If we got enough samples through the srtt filter,
822 	 * save the rtt and rttvar in the routing entry.
823 	 * 'Enough' is arbitrarily defined as the 16 samples.
824 	 * 16 samples is enough for the srtt filter to converge
825 	 * to within 5% of the correct value; fewer samples and
826 	 * we could save a very bogus rtt.
827 	 *
828 	 * Don't update the default route's characteristics and don't
829 	 * update anything that the user "locked".
830 	 */
831 	if (tp->t_rttupdated >= 16) {
832 		u_long i = 0;
833 
834 		if (isipv6) {
835 			struct sockaddr_in6 *sin6;
836 
837 			if ((rt = inp->in6p_route.ro_rt) == NULL)
838 				goto no_valid_rt;
839 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
840 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
841 				goto no_valid_rt;
842 		} else
843 			if ((rt = inp->inp_route.ro_rt) == NULL ||
844 			    ((struct sockaddr_in *)rt_key(rt))->
845 			     sin_addr.s_addr == INADDR_ANY)
846 				goto no_valid_rt;
847 
848 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
849 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
850 			if (rt->rt_rmx.rmx_rtt && i)
851 				/*
852 				 * filter this update to half the old & half
853 				 * the new values, converting scale.
854 				 * See route.h and tcp_var.h for a
855 				 * description of the scaling constants.
856 				 */
857 				rt->rt_rmx.rmx_rtt =
858 				    (rt->rt_rmx.rmx_rtt + i) / 2;
859 			else
860 				rt->rt_rmx.rmx_rtt = i;
861 			tcpstat.tcps_cachedrtt++;
862 		}
863 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
864 			i = tp->t_rttvar *
865 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
866 			if (rt->rt_rmx.rmx_rttvar && i)
867 				rt->rt_rmx.rmx_rttvar =
868 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
869 			else
870 				rt->rt_rmx.rmx_rttvar = i;
871 			tcpstat.tcps_cachedrttvar++;
872 		}
873 		/*
874 		 * The old comment here said:
875 		 * update the pipelimit (ssthresh) if it has been updated
876 		 * already or if a pipesize was specified & the threshhold
877 		 * got below half the pipesize.  I.e., wait for bad news
878 		 * before we start updating, then update on both good
879 		 * and bad news.
880 		 *
881 		 * But we want to save the ssthresh even if no pipesize is
882 		 * specified explicitly in the route, because such
883 		 * connections still have an implicit pipesize specified
884 		 * by the global tcp_sendspace.  In the absence of a reliable
885 		 * way to calculate the pipesize, it will have to do.
886 		 */
887 		i = tp->snd_ssthresh;
888 		if (rt->rt_rmx.rmx_sendpipe != 0)
889 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
890 		else
891 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
892 		if (dosavessthresh ||
893 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
894 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
895 			/*
896 			 * convert the limit from user data bytes to
897 			 * packets then to packet data bytes.
898 			 */
899 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
900 			if (i < 2)
901 				i = 2;
902 			i *= tp->t_maxseg +
903 			     (isipv6 ?
904 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
905 			      sizeof(struct tcpiphdr));
906 			if (rt->rt_rmx.rmx_ssthresh)
907 				rt->rt_rmx.rmx_ssthresh =
908 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
909 			else
910 				rt->rt_rmx.rmx_ssthresh = i;
911 			tcpstat.tcps_cachedssthresh++;
912 		}
913 	}
914 
915 no_valid_rt:
916 	/* free the reassembly queue, if any */
917 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
918 		LIST_REMOVE(q, tqe_q);
919 		m_freem(q->tqe_m);
920 		FREE(q, M_TSEGQ);
921 		tcp_reass_qsize--;
922 	}
923 	/* throw away SACK blocks in scoreboard*/
924 	if (TCP_DO_SACK(tp))
925 		tcp_sack_cleanup(&tp->scb);
926 
927 	inp->inp_ppcb = NULL;
928 	soisdisconnected(so);
929 	/*
930 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
931 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
932 	 * for each protocol thread and must be removed in the context of
933 	 * that thread.  This is accomplished by chaining the message
934 	 * through the cpus.
935 	 *
936 	 * If the inp is not wildcarded we simply detach, which will remove
937 	 * the any hashes still present for this inp.
938 	 */
939 #ifdef SMP
940 	if (inp->inp_flags & INP_WILDCARD_MP) {
941 		struct netmsg_remwildcard *msg;
942 
943 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
944 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
945 			      M_LWKTMSG, M_INTWAIT);
946 		netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
947 			    in_pcbremwildcardhash_handler);
948 #ifdef INET6
949 		msg->nm_isinet6 = isafinet6;
950 #endif
951 		msg->nm_inp = inp;
952 		msg->nm_pcbinfo = &tcbinfo[cpu];
953 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
954 	} else
955 #endif
956 	{
957 		/* note: detach removes any wildcard hash entry */
958 #ifdef INET6
959 		if (isafinet6)
960 			in6_pcbdetach(inp);
961 		else
962 #endif
963 			in_pcbdetach(inp);
964 	}
965 	tcpstat.tcps_closed++;
966 	return (NULL);
967 }
968 
969 static __inline void
970 tcp_drain_oncpu(struct inpcbhead *head)
971 {
972 	struct inpcb *inpb;
973 	struct tcpcb *tcpb;
974 	struct tseg_qent *te;
975 
976 	LIST_FOREACH(inpb, head, inp_list) {
977 		if (inpb->inp_flags & INP_PLACEMARKER)
978 			continue;
979 		if ((tcpb = intotcpcb(inpb))) {
980 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
981 				LIST_REMOVE(te, tqe_q);
982 				m_freem(te->tqe_m);
983 				FREE(te, M_TSEGQ);
984 				tcp_reass_qsize--;
985 			}
986 		}
987 	}
988 }
989 
990 #ifdef SMP
991 struct netmsg_tcp_drain {
992 	struct netmsg		nm_netmsg;
993 	struct inpcbhead	*nm_head;
994 };
995 
996 static void
997 tcp_drain_handler(netmsg_t netmsg)
998 {
999 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1000 
1001 	tcp_drain_oncpu(nm->nm_head);
1002 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1003 }
1004 #endif
1005 
1006 void
1007 tcp_drain(void)
1008 {
1009 #ifdef SMP
1010 	int cpu;
1011 #endif
1012 
1013 	if (!do_tcpdrain)
1014 		return;
1015 
1016 	/*
1017 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1018 	 * if there is one...
1019 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1020 	 *	reassembly queue should be flushed, but in a situation
1021 	 *	where we're really low on mbufs, this is potentially
1022 	 *	useful.
1023 	 */
1024 #ifdef SMP
1025 	for (cpu = 0; cpu < ncpus2; cpu++) {
1026 		struct netmsg_tcp_drain *msg;
1027 
1028 		if (cpu == mycpu->gd_cpuid) {
1029 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1030 		} else {
1031 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1032 				    M_LWKTMSG, M_NOWAIT);
1033 			if (msg == NULL)
1034 				continue;
1035 			netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1036 				    tcp_drain_handler);
1037 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1038 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1039 		}
1040 	}
1041 #else
1042 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1043 #endif
1044 }
1045 
1046 /*
1047  * Notify a tcp user of an asynchronous error;
1048  * store error as soft error, but wake up user
1049  * (for now, won't do anything until can select for soft error).
1050  *
1051  * Do not wake up user since there currently is no mechanism for
1052  * reporting soft errors (yet - a kqueue filter may be added).
1053  */
1054 static void
1055 tcp_notify(struct inpcb *inp, int error)
1056 {
1057 	struct tcpcb *tp = intotcpcb(inp);
1058 
1059 	/*
1060 	 * Ignore some errors if we are hooked up.
1061 	 * If connection hasn't completed, has retransmitted several times,
1062 	 * and receives a second error, give up now.  This is better
1063 	 * than waiting a long time to establish a connection that
1064 	 * can never complete.
1065 	 */
1066 	if (tp->t_state == TCPS_ESTABLISHED &&
1067 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1068 	      error == EHOSTDOWN)) {
1069 		return;
1070 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1071 	    tp->t_softerror)
1072 		tcp_drop(tp, error);
1073 	else
1074 		tp->t_softerror = error;
1075 #if 0
1076 	wakeup(&so->so_timeo);
1077 	sorwakeup(so);
1078 	sowwakeup(so);
1079 #endif
1080 }
1081 
1082 static int
1083 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1084 {
1085 	int error, i, n;
1086 	struct inpcb *marker;
1087 	struct inpcb *inp;
1088 	inp_gen_t gencnt;
1089 	globaldata_t gd;
1090 	int origcpu, ccpu;
1091 
1092 	error = 0;
1093 	n = 0;
1094 
1095 	/*
1096 	 * The process of preparing the TCB list is too time-consuming and
1097 	 * resource-intensive to repeat twice on every request.
1098 	 */
1099 	if (req->oldptr == NULL) {
1100 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1101 			gd = globaldata_find(ccpu);
1102 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1103 		}
1104 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1105 		return (0);
1106 	}
1107 
1108 	if (req->newptr != NULL)
1109 		return (EPERM);
1110 
1111 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1112 	marker->inp_flags |= INP_PLACEMARKER;
1113 
1114 	/*
1115 	 * OK, now we're committed to doing something.  Run the inpcb list
1116 	 * for each cpu in the system and construct the output.  Use a
1117 	 * list placemarker to deal with list changes occuring during
1118 	 * copyout blockages (but otherwise depend on being on the correct
1119 	 * cpu to avoid races).
1120 	 */
1121 	origcpu = mycpu->gd_cpuid;
1122 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1123 		globaldata_t rgd;
1124 		caddr_t inp_ppcb;
1125 		struct xtcpcb xt;
1126 		int cpu_id;
1127 
1128 		cpu_id = (origcpu + ccpu) % ncpus;
1129 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1130 			continue;
1131 		rgd = globaldata_find(cpu_id);
1132 		lwkt_setcpu_self(rgd);
1133 
1134 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1135 		n = tcbinfo[cpu_id].ipi_count;
1136 
1137 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1138 		i = 0;
1139 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1140 			/*
1141 			 * process a snapshot of pcbs, ignoring placemarkers
1142 			 * and using our own to allow SYSCTL_OUT to block.
1143 			 */
1144 			LIST_REMOVE(marker, inp_list);
1145 			LIST_INSERT_AFTER(inp, marker, inp_list);
1146 
1147 			if (inp->inp_flags & INP_PLACEMARKER)
1148 				continue;
1149 			if (inp->inp_gencnt > gencnt)
1150 				continue;
1151 			if (prison_xinpcb(req->td, inp))
1152 				continue;
1153 
1154 			xt.xt_len = sizeof xt;
1155 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1156 			inp_ppcb = inp->inp_ppcb;
1157 			if (inp_ppcb != NULL)
1158 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1159 			else
1160 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1161 			if (inp->inp_socket)
1162 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1163 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1164 				break;
1165 			++i;
1166 		}
1167 		LIST_REMOVE(marker, inp_list);
1168 		if (error == 0 && i < n) {
1169 			bzero(&xt, sizeof xt);
1170 			xt.xt_len = sizeof xt;
1171 			while (i < n) {
1172 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1173 				if (error)
1174 					break;
1175 				++i;
1176 			}
1177 		}
1178 	}
1179 
1180 	/*
1181 	 * Make sure we are on the same cpu we were on originally, since
1182 	 * higher level callers expect this.  Also don't pollute caches with
1183 	 * migrated userland data by (eventually) returning to userland
1184 	 * on a different cpu.
1185 	 */
1186 	lwkt_setcpu_self(globaldata_find(origcpu));
1187 	kfree(marker, M_TEMP);
1188 	return (error);
1189 }
1190 
1191 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1192 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1193 
1194 static int
1195 tcp_getcred(SYSCTL_HANDLER_ARGS)
1196 {
1197 	struct sockaddr_in addrs[2];
1198 	struct inpcb *inp;
1199 	int cpu;
1200 	int error;
1201 
1202 	error = suser(req->td);
1203 	if (error != 0)
1204 		return (error);
1205 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1206 	if (error != 0)
1207 		return (error);
1208 	crit_enter();
1209 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1210 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1211 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1212 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1213 	if (inp == NULL || inp->inp_socket == NULL) {
1214 		error = ENOENT;
1215 		goto out;
1216 	}
1217 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1218 out:
1219 	crit_exit();
1220 	return (error);
1221 }
1222 
1223 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1224     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1225 
1226 #ifdef INET6
1227 static int
1228 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1229 {
1230 	struct sockaddr_in6 addrs[2];
1231 	struct inpcb *inp;
1232 	int error;
1233 	boolean_t mapped = FALSE;
1234 
1235 	error = suser(req->td);
1236 	if (error != 0)
1237 		return (error);
1238 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1239 	if (error != 0)
1240 		return (error);
1241 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1242 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1243 			mapped = TRUE;
1244 		else
1245 			return (EINVAL);
1246 	}
1247 	crit_enter();
1248 	if (mapped) {
1249 		inp = in_pcblookup_hash(&tcbinfo[0],
1250 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1251 		    addrs[1].sin6_port,
1252 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1253 		    addrs[0].sin6_port,
1254 		    0, NULL);
1255 	} else {
1256 		inp = in6_pcblookup_hash(&tcbinfo[0],
1257 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1258 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1259 		    0, NULL);
1260 	}
1261 	if (inp == NULL || inp->inp_socket == NULL) {
1262 		error = ENOENT;
1263 		goto out;
1264 	}
1265 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1266 out:
1267 	crit_exit();
1268 	return (error);
1269 }
1270 
1271 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1272 	    0, 0,
1273 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1274 #endif
1275 
1276 void
1277 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1278 {
1279 	struct ip *ip = vip;
1280 	struct tcphdr *th;
1281 	struct in_addr faddr;
1282 	struct inpcb *inp;
1283 	struct tcpcb *tp;
1284 	void (*notify)(struct inpcb *, int) = tcp_notify;
1285 	tcp_seq icmpseq;
1286 	int arg, cpu;
1287 
1288 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1289 		return;
1290 	}
1291 
1292 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1293 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1294 		return;
1295 
1296 	arg = inetctlerrmap[cmd];
1297 	if (cmd == PRC_QUENCH) {
1298 		notify = tcp_quench;
1299 	} else if (icmp_may_rst &&
1300 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1301 		    cmd == PRC_UNREACH_PORT ||
1302 		    cmd == PRC_TIMXCEED_INTRANS) &&
1303 		   ip != NULL) {
1304 		notify = tcp_drop_syn_sent;
1305 	} else if (cmd == PRC_MSGSIZE) {
1306 		struct icmp *icmp = (struct icmp *)
1307 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1308 
1309 		arg = ntohs(icmp->icmp_nextmtu);
1310 		notify = tcp_mtudisc;
1311 	} else if (PRC_IS_REDIRECT(cmd)) {
1312 		ip = NULL;
1313 		notify = in_rtchange;
1314 	} else if (cmd == PRC_HOSTDEAD) {
1315 		ip = NULL;
1316 	}
1317 
1318 	if (ip != NULL) {
1319 		crit_enter();
1320 		th = (struct tcphdr *)((caddr_t)ip +
1321 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1322 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1323 				  ip->ip_src.s_addr, th->th_sport);
1324 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1325 					ip->ip_src, th->th_sport, 0, NULL);
1326 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1327 			icmpseq = htonl(th->th_seq);
1328 			tp = intotcpcb(inp);
1329 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1330 			    SEQ_LT(icmpseq, tp->snd_max))
1331 				(*notify)(inp, arg);
1332 		} else {
1333 			struct in_conninfo inc;
1334 
1335 			inc.inc_fport = th->th_dport;
1336 			inc.inc_lport = th->th_sport;
1337 			inc.inc_faddr = faddr;
1338 			inc.inc_laddr = ip->ip_src;
1339 #ifdef INET6
1340 			inc.inc_isipv6 = 0;
1341 #endif
1342 			syncache_unreach(&inc, th);
1343 		}
1344 		crit_exit();
1345 	} else {
1346 		for (cpu = 0; cpu < ncpus2; cpu++) {
1347 			in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1348 					notify);
1349 		}
1350 	}
1351 }
1352 
1353 #ifdef INET6
1354 void
1355 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1356 {
1357 	struct tcphdr th;
1358 	void (*notify) (struct inpcb *, int) = tcp_notify;
1359 	struct ip6_hdr *ip6;
1360 	struct mbuf *m;
1361 	struct ip6ctlparam *ip6cp = NULL;
1362 	const struct sockaddr_in6 *sa6_src = NULL;
1363 	int off;
1364 	struct tcp_portonly {
1365 		u_int16_t th_sport;
1366 		u_int16_t th_dport;
1367 	} *thp;
1368 	int arg;
1369 
1370 	if (sa->sa_family != AF_INET6 ||
1371 	    sa->sa_len != sizeof(struct sockaddr_in6))
1372 		return;
1373 
1374 	arg = 0;
1375 	if (cmd == PRC_QUENCH)
1376 		notify = tcp_quench;
1377 	else if (cmd == PRC_MSGSIZE) {
1378 		struct ip6ctlparam *ip6cp = d;
1379 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1380 
1381 		arg = ntohl(icmp6->icmp6_mtu);
1382 		notify = tcp_mtudisc;
1383 	} else if (!PRC_IS_REDIRECT(cmd) &&
1384 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1385 		return;
1386 	}
1387 
1388 	/* if the parameter is from icmp6, decode it. */
1389 	if (d != NULL) {
1390 		ip6cp = (struct ip6ctlparam *)d;
1391 		m = ip6cp->ip6c_m;
1392 		ip6 = ip6cp->ip6c_ip6;
1393 		off = ip6cp->ip6c_off;
1394 		sa6_src = ip6cp->ip6c_src;
1395 	} else {
1396 		m = NULL;
1397 		ip6 = NULL;
1398 		off = 0;	/* fool gcc */
1399 		sa6_src = &sa6_any;
1400 	}
1401 
1402 	if (ip6 != NULL) {
1403 		struct in_conninfo inc;
1404 		/*
1405 		 * XXX: We assume that when IPV6 is non NULL,
1406 		 * M and OFF are valid.
1407 		 */
1408 
1409 		/* check if we can safely examine src and dst ports */
1410 		if (m->m_pkthdr.len < off + sizeof *thp)
1411 			return;
1412 
1413 		bzero(&th, sizeof th);
1414 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1415 
1416 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1417 		    (struct sockaddr *)ip6cp->ip6c_src,
1418 		    th.th_sport, cmd, arg, notify);
1419 
1420 		inc.inc_fport = th.th_dport;
1421 		inc.inc_lport = th.th_sport;
1422 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1423 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1424 		inc.inc_isipv6 = 1;
1425 		syncache_unreach(&inc, &th);
1426 	} else
1427 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1428 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1429 }
1430 #endif
1431 
1432 /*
1433  * Following is where TCP initial sequence number generation occurs.
1434  *
1435  * There are two places where we must use initial sequence numbers:
1436  * 1.  In SYN-ACK packets.
1437  * 2.  In SYN packets.
1438  *
1439  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1440  * tcp_syncache.c for details.
1441  *
1442  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1443  * depends on this property.  In addition, these ISNs should be
1444  * unguessable so as to prevent connection hijacking.  To satisfy
1445  * the requirements of this situation, the algorithm outlined in
1446  * RFC 1948 is used to generate sequence numbers.
1447  *
1448  * Implementation details:
1449  *
1450  * Time is based off the system timer, and is corrected so that it
1451  * increases by one megabyte per second.  This allows for proper
1452  * recycling on high speed LANs while still leaving over an hour
1453  * before rollover.
1454  *
1455  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1456  * between seeding of isn_secret.  This is normally set to zero,
1457  * as reseeding should not be necessary.
1458  *
1459  */
1460 
1461 #define	ISN_BYTES_PER_SECOND 1048576
1462 
1463 u_char isn_secret[32];
1464 int isn_last_reseed;
1465 MD5_CTX isn_ctx;
1466 
1467 tcp_seq
1468 tcp_new_isn(struct tcpcb *tp)
1469 {
1470 	u_int32_t md5_buffer[4];
1471 	tcp_seq new_isn;
1472 
1473 	/* Seed if this is the first use, reseed if requested. */
1474 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1475 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1476 		< (u_int)ticks))) {
1477 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1478 		isn_last_reseed = ticks;
1479 	}
1480 
1481 	/* Compute the md5 hash and return the ISN. */
1482 	MD5Init(&isn_ctx);
1483 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1484 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1485 #ifdef INET6
1486 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1487 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1488 			  sizeof(struct in6_addr));
1489 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1490 			  sizeof(struct in6_addr));
1491 	} else
1492 #endif
1493 	{
1494 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1495 			  sizeof(struct in_addr));
1496 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1497 			  sizeof(struct in_addr));
1498 	}
1499 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1500 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1501 	new_isn = (tcp_seq) md5_buffer[0];
1502 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1503 	return (new_isn);
1504 }
1505 
1506 /*
1507  * When a source quench is received, close congestion window
1508  * to one segment.  We will gradually open it again as we proceed.
1509  */
1510 void
1511 tcp_quench(struct inpcb *inp, int error)
1512 {
1513 	struct tcpcb *tp = intotcpcb(inp);
1514 
1515 	if (tp != NULL) {
1516 		tp->snd_cwnd = tp->t_maxseg;
1517 		tp->snd_wacked = 0;
1518 	}
1519 }
1520 
1521 /*
1522  * When a specific ICMP unreachable message is received and the
1523  * connection state is SYN-SENT, drop the connection.  This behavior
1524  * is controlled by the icmp_may_rst sysctl.
1525  */
1526 void
1527 tcp_drop_syn_sent(struct inpcb *inp, int error)
1528 {
1529 	struct tcpcb *tp = intotcpcb(inp);
1530 
1531 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1532 		tcp_drop(tp, error);
1533 }
1534 
1535 /*
1536  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1537  * based on the new value in the route.  Also nudge TCP to send something,
1538  * since we know the packet we just sent was dropped.
1539  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1540  */
1541 void
1542 tcp_mtudisc(struct inpcb *inp, int mtu)
1543 {
1544 	struct tcpcb *tp = intotcpcb(inp);
1545 	struct rtentry *rt;
1546 	struct socket *so = inp->inp_socket;
1547 	int maxopd, mss;
1548 #ifdef INET6
1549 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1550 #else
1551 	const boolean_t isipv6 = FALSE;
1552 #endif
1553 
1554 	if (tp == NULL)
1555 		return;
1556 
1557 	/*
1558 	 * If no MTU is provided in the ICMP message, use the
1559 	 * next lower likely value, as specified in RFC 1191.
1560 	 */
1561 	if (mtu == 0) {
1562 		int oldmtu;
1563 
1564 		oldmtu = tp->t_maxopd +
1565 		    (isipv6 ?
1566 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1567 		     sizeof(struct tcpiphdr));
1568 		mtu = ip_next_mtu(oldmtu, 0);
1569 	}
1570 
1571 	if (isipv6)
1572 		rt = tcp_rtlookup6(&inp->inp_inc);
1573 	else
1574 		rt = tcp_rtlookup(&inp->inp_inc);
1575 	if (rt != NULL) {
1576 		struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1577 
1578 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1579 			mtu = rt->rt_rmx.rmx_mtu;
1580 
1581 		maxopd = mtu -
1582 		    (isipv6 ?
1583 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1584 		     sizeof(struct tcpiphdr));
1585 
1586 		/*
1587 		 * XXX - The following conditional probably violates the TCP
1588 		 * spec.  The problem is that, since we don't know the
1589 		 * other end's MSS, we are supposed to use a conservative
1590 		 * default.  But, if we do that, then MTU discovery will
1591 		 * never actually take place, because the conservative
1592 		 * default is much less than the MTUs typically seen
1593 		 * on the Internet today.  For the moment, we'll sweep
1594 		 * this under the carpet.
1595 		 *
1596 		 * The conservative default might not actually be a problem
1597 		 * if the only case this occurs is when sending an initial
1598 		 * SYN with options and data to a host we've never talked
1599 		 * to before.  Then, they will reply with an MSS value which
1600 		 * will get recorded and the new parameters should get
1601 		 * recomputed.  For Further Study.
1602 		 */
1603 		if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1604 			maxopd = taop->tao_mssopt;
1605 	} else
1606 		maxopd = mtu -
1607 		    (isipv6 ?
1608 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1609 		     sizeof(struct tcpiphdr));
1610 
1611 	if (tp->t_maxopd <= maxopd)
1612 		return;
1613 	tp->t_maxopd = maxopd;
1614 
1615 	mss = maxopd;
1616 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1617 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1618 		mss -= TCPOLEN_TSTAMP_APPA;
1619 
1620 	if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1621 			   (TF_REQ_CC | TF_RCVD_CC))
1622 		mss -= TCPOLEN_CC_APPA;
1623 
1624 	/* round down to multiple of MCLBYTES */
1625 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1626 	if (mss > MCLBYTES)
1627 		mss &= ~(MCLBYTES - 1);
1628 #else
1629 	if (mss > MCLBYTES)
1630 		mss = (mss / MCLBYTES) * MCLBYTES;
1631 #endif
1632 
1633 	if (so->so_snd.ssb_hiwat < mss)
1634 		mss = so->so_snd.ssb_hiwat;
1635 
1636 	tp->t_maxseg = mss;
1637 	tp->t_rtttime = 0;
1638 	tp->snd_nxt = tp->snd_una;
1639 	tcp_output(tp);
1640 	tcpstat.tcps_mturesent++;
1641 }
1642 
1643 /*
1644  * Look-up the routing entry to the peer of this inpcb.  If no route
1645  * is found and it cannot be allocated the return NULL.  This routine
1646  * is called by TCP routines that access the rmx structure and by tcp_mss
1647  * to get the interface MTU.
1648  */
1649 struct rtentry *
1650 tcp_rtlookup(struct in_conninfo *inc)
1651 {
1652 	struct route *ro = &inc->inc_route;
1653 
1654 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1655 		/* No route yet, so try to acquire one */
1656 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1657 			/*
1658 			 * unused portions of the structure MUST be zero'd
1659 			 * out because rtalloc() treats it as opaque data
1660 			 */
1661 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1662 			ro->ro_dst.sa_family = AF_INET;
1663 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1664 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1665 			    inc->inc_faddr;
1666 			rtalloc(ro);
1667 		}
1668 	}
1669 	return (ro->ro_rt);
1670 }
1671 
1672 #ifdef INET6
1673 struct rtentry *
1674 tcp_rtlookup6(struct in_conninfo *inc)
1675 {
1676 	struct route_in6 *ro6 = &inc->inc6_route;
1677 
1678 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1679 		/* No route yet, so try to acquire one */
1680 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1681 			/*
1682 			 * unused portions of the structure MUST be zero'd
1683 			 * out because rtalloc() treats it as opaque data
1684 			 */
1685 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1686 			ro6->ro_dst.sin6_family = AF_INET6;
1687 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1688 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1689 			rtalloc((struct route *)ro6);
1690 		}
1691 	}
1692 	return (ro6->ro_rt);
1693 }
1694 #endif
1695 
1696 #ifdef IPSEC
1697 /* compute ESP/AH header size for TCP, including outer IP header. */
1698 size_t
1699 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1700 {
1701 	struct inpcb *inp;
1702 	struct mbuf *m;
1703 	size_t hdrsiz;
1704 	struct ip *ip;
1705 	struct tcphdr *th;
1706 
1707 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1708 		return (0);
1709 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1710 	if (!m)
1711 		return (0);
1712 
1713 #ifdef INET6
1714 	if (inp->inp_vflag & INP_IPV6) {
1715 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1716 
1717 		th = (struct tcphdr *)(ip6 + 1);
1718 		m->m_pkthdr.len = m->m_len =
1719 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1720 		tcp_fillheaders(tp, ip6, th);
1721 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1722 	} else
1723 #endif
1724 	{
1725 		ip = mtod(m, struct ip *);
1726 		th = (struct tcphdr *)(ip + 1);
1727 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1728 		tcp_fillheaders(tp, ip, th);
1729 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1730 	}
1731 
1732 	m_free(m);
1733 	return (hdrsiz);
1734 }
1735 #endif
1736 
1737 /*
1738  * Return a pointer to the cached information about the remote host.
1739  * The cached information is stored in the protocol specific part of
1740  * the route metrics.
1741  */
1742 struct rmxp_tao *
1743 tcp_gettaocache(struct in_conninfo *inc)
1744 {
1745 	struct rtentry *rt;
1746 
1747 #ifdef INET6
1748 	if (inc->inc_isipv6)
1749 		rt = tcp_rtlookup6(inc);
1750 	else
1751 #endif
1752 		rt = tcp_rtlookup(inc);
1753 
1754 	/* Make sure this is a host route and is up. */
1755 	if (rt == NULL ||
1756 	    (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1757 		return (NULL);
1758 
1759 	return (rmx_taop(rt->rt_rmx));
1760 }
1761 
1762 /*
1763  * Clear all the TAO cache entries, called from tcp_init.
1764  *
1765  * XXX
1766  * This routine is just an empty one, because we assume that the routing
1767  * routing tables are initialized at the same time when TCP, so there is
1768  * nothing in the cache left over.
1769  */
1770 static void
1771 tcp_cleartaocache(void)
1772 {
1773 }
1774 
1775 /*
1776  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1777  *
1778  * This code attempts to calculate the bandwidth-delay product as a
1779  * means of determining the optimal window size to maximize bandwidth,
1780  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1781  * routers.  This code also does a fairly good job keeping RTTs in check
1782  * across slow links like modems.  We implement an algorithm which is very
1783  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1784  * transmitter side of a TCP connection and so only effects the transmit
1785  * side of the connection.
1786  *
1787  * BACKGROUND:  TCP makes no provision for the management of buffer space
1788  * at the end points or at the intermediate routers and switches.  A TCP
1789  * stream, whether using NewReno or not, will eventually buffer as
1790  * many packets as it is able and the only reason this typically works is
1791  * due to the fairly small default buffers made available for a connection
1792  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1793  * scaling it is now fairly easy for even a single TCP connection to blow-out
1794  * all available buffer space not only on the local interface, but on
1795  * intermediate routers and switches as well.  NewReno makes a misguided
1796  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1797  * then backing off, then steadily increasing the window again until another
1798  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1799  * is only made worse as network loads increase and the idea of intentionally
1800  * blowing out network buffers is, frankly, a terrible way to manage network
1801  * resources.
1802  *
1803  * It is far better to limit the transmit window prior to the failure
1804  * condition being achieved.  There are two general ways to do this:  First
1805  * you can 'scan' through different transmit window sizes and locate the
1806  * point where the RTT stops increasing, indicating that you have filled the
1807  * pipe, then scan backwards until you note that RTT stops decreasing, then
1808  * repeat ad-infinitum.  This method works in principle but has severe
1809  * implementation issues due to RTT variances, timer granularity, and
1810  * instability in the algorithm which can lead to many false positives and
1811  * create oscillations as well as interact badly with other TCP streams
1812  * implementing the same algorithm.
1813  *
1814  * The second method is to limit the window to the bandwidth delay product
1815  * of the link.  This is the method we implement.  RTT variances and our
1816  * own manipulation of the congestion window, bwnd, can potentially
1817  * destabilize the algorithm.  For this reason we have to stabilize the
1818  * elements used to calculate the window.  We do this by using the minimum
1819  * observed RTT, the long term average of the observed bandwidth, and
1820  * by adding two segments worth of slop.  It isn't perfect but it is able
1821  * to react to changing conditions and gives us a very stable basis on
1822  * which to extend the algorithm.
1823  */
1824 void
1825 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1826 {
1827 	u_long bw;
1828 	u_long bwnd;
1829 	int save_ticks;
1830 	int delta_ticks;
1831 
1832 	/*
1833 	 * If inflight_enable is disabled in the middle of a tcp connection,
1834 	 * make sure snd_bwnd is effectively disabled.
1835 	 */
1836 	if (!tcp_inflight_enable) {
1837 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1838 		tp->snd_bandwidth = 0;
1839 		return;
1840 	}
1841 
1842 	/*
1843 	 * Validate the delta time.  If a connection is new or has been idle
1844 	 * a long time we have to reset the bandwidth calculator.
1845 	 */
1846 	save_ticks = ticks;
1847 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1848 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1849 		tp->t_bw_rtttime = ticks;
1850 		tp->t_bw_rtseq = ack_seq;
1851 		if (tp->snd_bandwidth == 0)
1852 			tp->snd_bandwidth = tcp_inflight_min;
1853 		return;
1854 	}
1855 	if (delta_ticks == 0)
1856 		return;
1857 
1858 	/*
1859 	 * Sanity check, plus ignore pure window update acks.
1860 	 */
1861 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1862 		return;
1863 
1864 	/*
1865 	 * Figure out the bandwidth.  Due to the tick granularity this
1866 	 * is a very rough number and it MUST be averaged over a fairly
1867 	 * long period of time.  XXX we need to take into account a link
1868 	 * that is not using all available bandwidth, but for now our
1869 	 * slop will ramp us up if this case occurs and the bandwidth later
1870 	 * increases.
1871 	 */
1872 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1873 	tp->t_bw_rtttime = save_ticks;
1874 	tp->t_bw_rtseq = ack_seq;
1875 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1876 
1877 	tp->snd_bandwidth = bw;
1878 
1879 	/*
1880 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1881 	 * segments.  The additional slop puts us squarely in the sweet
1882 	 * spot and also handles the bandwidth run-up case.  Without the
1883 	 * slop we could be locking ourselves into a lower bandwidth.
1884 	 *
1885 	 * Situations Handled:
1886 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1887 	 *	    high speed LANs, allowing larger TCP buffers to be
1888 	 *	    specified, and also does a good job preventing
1889 	 *	    over-queueing of packets over choke points like modems
1890 	 *	    (at least for the transmit side).
1891 	 *
1892 	 *	(2) Is able to handle changing network loads (bandwidth
1893 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1894 	 *	    increases).
1895 	 *
1896 	 *	(3) Theoretically should stabilize in the face of multiple
1897 	 *	    connections implementing the same algorithm (this may need
1898 	 *	    a little work).
1899 	 *
1900 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1901 	 *	    be adjusted with a sysctl but typically only needs to be on
1902 	 *	    very slow connections.  A value no smaller then 5 should
1903 	 *	    be used, but only reduce this default if you have no other
1904 	 *	    choice.
1905 	 */
1906 
1907 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1908 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1909 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1910 #undef USERTT
1911 
1912 	if (tcp_inflight_debug > 0) {
1913 		static int ltime;
1914 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1915 			ltime = ticks;
1916 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1917 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1918 		}
1919 	}
1920 	if ((long)bwnd < tcp_inflight_min)
1921 		bwnd = tcp_inflight_min;
1922 	if (bwnd > tcp_inflight_max)
1923 		bwnd = tcp_inflight_max;
1924 	if ((long)bwnd < tp->t_maxseg * 2)
1925 		bwnd = tp->t_maxseg * 2;
1926 	tp->snd_bwnd = bwnd;
1927 }
1928