xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 6cef7136f04e2b24a6db289e78720d6d8c60274e)
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 1.63 2008/11/11 10:46:58 sephe Exp $
69  */
70 
71 #include "opt_compat.h"
72 #include "opt_inet.h"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_tcpdebug.h"
76 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/callout.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mpipe.h>
84 #include <sys/mbuf.h>
85 #ifdef INET6
86 #include <sys/domain.h>
87 #endif
88 #include <sys/proc.h>
89 #include <sys/priv.h>
90 #include <sys/socket.h>
91 #include <sys/socketvar.h>
92 #include <sys/protosw.h>
93 #include <sys/random.h>
94 #include <sys/in_cksum.h>
95 #include <sys/ktr.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_timer2.h>
120 #include <netinet/tcp_var.h>
121 #include <netinet6/tcp6_var.h>
122 #include <netinet/tcpip.h>
123 #ifdef TCPDEBUG
124 #include <netinet/tcp_debug.h>
125 #endif
126 #include <netinet6/ip6protosw.h>
127 
128 #ifdef IPSEC
129 #include <netinet6/ipsec.h>
130 #include <netproto/key/key.h>
131 #ifdef INET6
132 #include <netinet6/ipsec6.h>
133 #endif
134 #endif
135 
136 #ifdef FAST_IPSEC
137 #include <netproto/ipsec/ipsec.h>
138 #ifdef INET6
139 #include <netproto/ipsec/ipsec6.h>
140 #endif
141 #define	IPSEC
142 #endif
143 
144 #include <sys/md5.h>
145 #include <machine/smp.h>
146 
147 #include <sys/msgport2.h>
148 #include <sys/mplock2.h>
149 #include <net/netmsg2.h>
150 
151 #if !defined(KTR_TCP)
152 #define KTR_TCP		KTR_ALL
153 #endif
154 KTR_INFO_MASTER(tcp);
155 /*
156 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
158 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
159 */
160 #define logtcp(name)	KTR_LOG(tcp_ ## name)
161 
162 struct inpcbinfo tcbinfo[MAXCPU];
163 struct tcpcbackqhead tcpcbackq[MAXCPU];
164 
165 int tcp_mssdflt = TCP_MSS;
166 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
167     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
168 
169 #ifdef INET6
170 int tcp_v6mssdflt = TCP6_MSS;
171 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
172     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
173 #endif
174 
175 /*
176  * Minimum MSS we accept and use. This prevents DoS attacks where
177  * we are forced to a ridiculous low MSS like 20 and send hundreds
178  * of packets instead of one. The effect scales with the available
179  * bandwidth and quickly saturates the CPU and network interface
180  * with packet generation and sending. Set to zero to disable MINMSS
181  * checking. This setting prevents us from sending too small packets.
182  */
183 int tcp_minmss = TCP_MINMSS;
184 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
185     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
186 
187 #if 0
188 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
189 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
190     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
191 #endif
192 
193 int tcp_do_rfc1323 = 1;
194 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
195     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
196 
197 static int tcp_tcbhashsize = 0;
198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
199      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
200 
201 static int do_tcpdrain = 1;
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
203      "Enable tcp_drain routine for extra help when low on mbufs");
204 
205 static int icmp_may_rst = 1;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
207     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
208 
209 static int tcp_isn_reseed_interval = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
211     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
212 
213 /*
214  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
215  * by default, but with generous values which should allow maximal
216  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
217  *
218  * The reason for doing this is that the limiter is the only mechanism we
219  * have which seems to do a really good job preventing receiver RX rings
220  * on network interfaces from getting blown out.  Even though GigE/10GigE
221  * is supposed to flow control it looks like either it doesn't actually
222  * do it or Open Source drivers do not properly enable it.
223  *
224  * People using the limiter to reduce bottlenecks on slower WAN connections
225  * should set the slop to 20 (2 packets).
226  */
227 static int tcp_inflight_enable = 1;
228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
229     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
230 
231 static int tcp_inflight_debug = 0;
232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
233     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
234 
235 static int tcp_inflight_min = 6144;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
237     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
238 
239 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
240 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
241     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
242 
243 static int tcp_inflight_stab = 50;
244 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
245     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
246 
247 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
248 static struct malloc_pipe tcptemp_mpipe;
249 
250 static void tcp_willblock(void);
251 static void tcp_notify (struct inpcb *, int);
252 
253 struct tcp_stats tcpstats_percpu[MAXCPU];
254 #ifdef SMP
255 static int
256 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
257 {
258 	int cpu, error = 0;
259 
260 	for (cpu = 0; cpu < ncpus; ++cpu) {
261 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
262 					sizeof(struct tcp_stats))))
263 			break;
264 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
265 				       sizeof(struct tcp_stats))))
266 			break;
267 	}
268 
269 	return (error);
270 }
271 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
272     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
273 #else
274 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
275     &tcpstat, tcp_stats, "TCP statistics");
276 #endif
277 
278 /*
279  * Target size of TCP PCB hash tables. Must be a power of two.
280  *
281  * Note that this can be overridden by the kernel environment
282  * variable net.inet.tcp.tcbhashsize
283  */
284 #ifndef TCBHASHSIZE
285 #define	TCBHASHSIZE	512
286 #endif
287 
288 /*
289  * This is the actual shape of what we allocate using the zone
290  * allocator.  Doing it this way allows us to protect both structures
291  * using the same generation count, and also eliminates the overhead
292  * of allocating tcpcbs separately.  By hiding the structure here,
293  * we avoid changing most of the rest of the code (although it needs
294  * to be changed, eventually, for greater efficiency).
295  */
296 #define	ALIGNMENT	32
297 #define	ALIGNM1		(ALIGNMENT - 1)
298 struct	inp_tp {
299 	union {
300 		struct	inpcb inp;
301 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
302 	} inp_tp_u;
303 	struct	tcpcb tcb;
304 	struct	tcp_callout inp_tp_rexmt;
305 	struct	tcp_callout inp_tp_persist;
306 	struct	tcp_callout inp_tp_keep;
307 	struct	tcp_callout inp_tp_2msl;
308 	struct	tcp_callout inp_tp_delack;
309 	struct	netmsg_tcp_timer inp_tp_timermsg;
310 };
311 #undef ALIGNMENT
312 #undef ALIGNM1
313 
314 /*
315  * Tcp initialization
316  */
317 void
318 tcp_init(void)
319 {
320 	struct inpcbporthead *porthashbase;
321 	u_long porthashmask;
322 	int hashsize = TCBHASHSIZE;
323 	int cpu;
324 
325 	/*
326 	 * note: tcptemp is used for keepalives, and it is ok for an
327 	 * allocation to fail so do not specify MPF_INT.
328 	 */
329 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
330 		    25, -1, 0, NULL);
331 
332 	tcp_delacktime = TCPTV_DELACK;
333 	tcp_keepinit = TCPTV_KEEP_INIT;
334 	tcp_keepidle = TCPTV_KEEP_IDLE;
335 	tcp_keepintvl = TCPTV_KEEPINTVL;
336 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
337 	tcp_msl = TCPTV_MSL;
338 	tcp_rexmit_min = TCPTV_MIN;
339 	tcp_rexmit_slop = TCPTV_CPU_VAR;
340 
341 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
342 	if (!powerof2(hashsize)) {
343 		kprintf("WARNING: TCB hash size not a power of 2\n");
344 		hashsize = 512; /* safe default */
345 	}
346 	tcp_tcbhashsize = hashsize;
347 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
348 
349 	for (cpu = 0; cpu < ncpus2; cpu++) {
350 		in_pcbinfo_init(&tcbinfo[cpu]);
351 		tcbinfo[cpu].cpu = cpu;
352 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
353 		    &tcbinfo[cpu].hashmask);
354 		tcbinfo[cpu].porthashbase = porthashbase;
355 		tcbinfo[cpu].porthashmask = porthashmask;
356 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
357 		    &tcbinfo[cpu].wildcardhashmask);
358 		tcbinfo[cpu].ipi_size = sizeof(struct inp_tp);
359 		TAILQ_INIT(&tcpcbackq[cpu]);
360 	}
361 
362 	tcp_reass_maxseg = nmbclusters / 16;
363 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
364 
365 #ifdef INET6
366 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
367 #else
368 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
369 #endif
370 	if (max_protohdr < TCP_MINPROTOHDR)
371 		max_protohdr = TCP_MINPROTOHDR;
372 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
373 		panic("tcp_init");
374 #undef TCP_MINPROTOHDR
375 
376 	/*
377 	 * Initialize TCP statistics counters for each CPU.
378 	 */
379 #ifdef SMP
380 	for (cpu = 0; cpu < ncpus; ++cpu) {
381 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
382 	}
383 #else
384 	bzero(&tcpstat, sizeof(struct tcp_stats));
385 #endif
386 
387 	syncache_init();
388 	netisr_register_rollup(tcp_willblock);
389 }
390 
391 static void
392 tcp_willblock(void)
393 {
394 	struct tcpcb *tp;
395 	int cpu = mycpu->gd_cpuid;
396 
397 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
398 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
399 		tp->t_flags &= ~TF_ONOUTPUTQ;
400 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
401 		tcp_output(tp);
402 	}
403 }
404 
405 /*
406  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
407  * tcp_template used to store this data in mbufs, but we now recopy it out
408  * of the tcpcb each time to conserve mbufs.
409  */
410 void
411 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
412 {
413 	struct inpcb *inp = tp->t_inpcb;
414 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
415 
416 #ifdef INET6
417 	if (inp->inp_vflag & INP_IPV6) {
418 		struct ip6_hdr *ip6;
419 
420 		ip6 = (struct ip6_hdr *)ip_ptr;
421 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
422 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
423 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
424 			(IPV6_VERSION & IPV6_VERSION_MASK);
425 		ip6->ip6_nxt = IPPROTO_TCP;
426 		ip6->ip6_plen = sizeof(struct tcphdr);
427 		ip6->ip6_src = inp->in6p_laddr;
428 		ip6->ip6_dst = inp->in6p_faddr;
429 		tcp_hdr->th_sum = 0;
430 	} else
431 #endif
432 	{
433 		struct ip *ip = (struct ip *) ip_ptr;
434 
435 		ip->ip_vhl = IP_VHL_BORING;
436 		ip->ip_tos = 0;
437 		ip->ip_len = 0;
438 		ip->ip_id = 0;
439 		ip->ip_off = 0;
440 		ip->ip_ttl = 0;
441 		ip->ip_sum = 0;
442 		ip->ip_p = IPPROTO_TCP;
443 		ip->ip_src = inp->inp_laddr;
444 		ip->ip_dst = inp->inp_faddr;
445 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
446 				    ip->ip_dst.s_addr,
447 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
448 	}
449 
450 	tcp_hdr->th_sport = inp->inp_lport;
451 	tcp_hdr->th_dport = inp->inp_fport;
452 	tcp_hdr->th_seq = 0;
453 	tcp_hdr->th_ack = 0;
454 	tcp_hdr->th_x2 = 0;
455 	tcp_hdr->th_off = 5;
456 	tcp_hdr->th_flags = 0;
457 	tcp_hdr->th_win = 0;
458 	tcp_hdr->th_urp = 0;
459 }
460 
461 /*
462  * Create template to be used to send tcp packets on a connection.
463  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
464  * use for this function is in keepalives, which use tcp_respond.
465  */
466 struct tcptemp *
467 tcp_maketemplate(struct tcpcb *tp)
468 {
469 	struct tcptemp *tmp;
470 
471 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
472 		return (NULL);
473 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
474 	return (tmp);
475 }
476 
477 void
478 tcp_freetemplate(struct tcptemp *tmp)
479 {
480 	mpipe_free(&tcptemp_mpipe, tmp);
481 }
482 
483 /*
484  * Send a single message to the TCP at address specified by
485  * the given TCP/IP header.  If m == NULL, then we make a copy
486  * of the tcpiphdr at ti and send directly to the addressed host.
487  * This is used to force keep alive messages out using the TCP
488  * template for a connection.  If flags are given then we send
489  * a message back to the TCP which originated the * segment ti,
490  * and discard the mbuf containing it and any other attached mbufs.
491  *
492  * In any case the ack and sequence number of the transmitted
493  * segment are as specified by the parameters.
494  *
495  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
496  */
497 void
498 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
499 	    tcp_seq ack, tcp_seq seq, int flags)
500 {
501 	int tlen;
502 	int win = 0;
503 	struct route *ro = NULL;
504 	struct route sro;
505 	struct ip *ip = ipgen;
506 	struct tcphdr *nth;
507 	int ipflags = 0;
508 	struct route_in6 *ro6 = NULL;
509 	struct route_in6 sro6;
510 	struct ip6_hdr *ip6 = ipgen;
511 	boolean_t use_tmpro = TRUE;
512 #ifdef INET6
513 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
514 #else
515 	const boolean_t isipv6 = FALSE;
516 #endif
517 
518 	if (tp != NULL) {
519 		if (!(flags & TH_RST)) {
520 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
521 			if (win < 0)
522 				win = 0;
523 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
524 				win = (long)TCP_MAXWIN << tp->rcv_scale;
525 		}
526 		/*
527 		 * Don't use the route cache of a listen socket,
528 		 * it is not MPSAFE; use temporary route cache.
529 		 */
530 		if (tp->t_state != TCPS_LISTEN) {
531 			if (isipv6)
532 				ro6 = &tp->t_inpcb->in6p_route;
533 			else
534 				ro = &tp->t_inpcb->inp_route;
535 			use_tmpro = FALSE;
536 		}
537 	}
538 	if (use_tmpro) {
539 		if (isipv6) {
540 			ro6 = &sro6;
541 			bzero(ro6, sizeof *ro6);
542 		} else {
543 			ro = &sro;
544 			bzero(ro, sizeof *ro);
545 		}
546 	}
547 	if (m == NULL) {
548 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
549 		if (m == NULL)
550 			return;
551 		tlen = 0;
552 		m->m_data += max_linkhdr;
553 		if (isipv6) {
554 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
555 			ip6 = mtod(m, struct ip6_hdr *);
556 			nth = (struct tcphdr *)(ip6 + 1);
557 		} else {
558 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
559 			ip = mtod(m, struct ip *);
560 			nth = (struct tcphdr *)(ip + 1);
561 		}
562 		bcopy(th, nth, sizeof(struct tcphdr));
563 		flags = TH_ACK;
564 	} else {
565 		m_freem(m->m_next);
566 		m->m_next = NULL;
567 		m->m_data = (caddr_t)ipgen;
568 		/* m_len is set later */
569 		tlen = 0;
570 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
571 		if (isipv6) {
572 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
573 			nth = (struct tcphdr *)(ip6 + 1);
574 		} else {
575 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
576 			nth = (struct tcphdr *)(ip + 1);
577 		}
578 		if (th != nth) {
579 			/*
580 			 * this is usually a case when an extension header
581 			 * exists between the IPv6 header and the
582 			 * TCP header.
583 			 */
584 			nth->th_sport = th->th_sport;
585 			nth->th_dport = th->th_dport;
586 		}
587 		xchg(nth->th_dport, nth->th_sport, n_short);
588 #undef xchg
589 	}
590 	if (isipv6) {
591 		ip6->ip6_flow = 0;
592 		ip6->ip6_vfc = IPV6_VERSION;
593 		ip6->ip6_nxt = IPPROTO_TCP;
594 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
595 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
596 	} else {
597 		tlen += sizeof(struct tcpiphdr);
598 		ip->ip_len = tlen;
599 		ip->ip_ttl = ip_defttl;
600 	}
601 	m->m_len = tlen;
602 	m->m_pkthdr.len = tlen;
603 	m->m_pkthdr.rcvif = NULL;
604 	nth->th_seq = htonl(seq);
605 	nth->th_ack = htonl(ack);
606 	nth->th_x2 = 0;
607 	nth->th_off = sizeof(struct tcphdr) >> 2;
608 	nth->th_flags = flags;
609 	if (tp != NULL)
610 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
611 	else
612 		nth->th_win = htons((u_short)win);
613 	nth->th_urp = 0;
614 	if (isipv6) {
615 		nth->th_sum = 0;
616 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
617 					sizeof(struct ip6_hdr),
618 					tlen - sizeof(struct ip6_hdr));
619 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
620 					       (ro6 && ro6->ro_rt) ?
621 						ro6->ro_rt->rt_ifp : NULL);
622 	} else {
623 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
624 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
625 		m->m_pkthdr.csum_flags = CSUM_TCP;
626 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
627 	}
628 #ifdef TCPDEBUG
629 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
630 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
631 #endif
632 	if (isipv6) {
633 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
634 			   tp ? tp->t_inpcb : NULL);
635 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
636 			RTFREE(ro6->ro_rt);
637 			ro6->ro_rt = NULL;
638 		}
639 	} else {
640 		ipflags |= IP_DEBUGROUTE;
641 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
642 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
643 			RTFREE(ro->ro_rt);
644 			ro->ro_rt = NULL;
645 		}
646 	}
647 }
648 
649 /*
650  * Create a new TCP control block, making an
651  * empty reassembly queue and hooking it to the argument
652  * protocol control block.  The `inp' parameter must have
653  * come from the zone allocator set up in tcp_init().
654  */
655 struct tcpcb *
656 tcp_newtcpcb(struct inpcb *inp)
657 {
658 	struct inp_tp *it;
659 	struct tcpcb *tp;
660 #ifdef INET6
661 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
662 #else
663 	const boolean_t isipv6 = FALSE;
664 #endif
665 
666 	it = (struct inp_tp *)inp;
667 	tp = &it->tcb;
668 	bzero(tp, sizeof(struct tcpcb));
669 	LIST_INIT(&tp->t_segq);
670 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
671 
672 	/* Set up our timeouts. */
673 	tp->tt_rexmt = &it->inp_tp_rexmt;
674 	tp->tt_persist = &it->inp_tp_persist;
675 	tp->tt_keep = &it->inp_tp_keep;
676 	tp->tt_2msl = &it->inp_tp_2msl;
677 	tp->tt_delack = &it->inp_tp_delack;
678 	tcp_inittimers(tp);
679 
680 	/*
681 	 * Zero out timer message.  We don't create it here,
682 	 * since the current CPU may not be the owner of this
683 	 * inpcb.
684 	 */
685 	tp->tt_msg = &it->inp_tp_timermsg;
686 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
687 
688 	if (tcp_do_rfc1323)
689 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
690 	tp->t_inpcb = inp;	/* XXX */
691 	tp->t_state = TCPS_CLOSED;
692 	/*
693 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
694 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
695 	 * reasonable initial retransmit time.
696 	 */
697 	tp->t_srtt = TCPTV_SRTTBASE;
698 	tp->t_rttvar =
699 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
700 	tp->t_rttmin = tcp_rexmit_min;
701 	tp->t_rxtcur = TCPTV_RTOBASE;
702 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
703 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
704 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
705 	tp->t_rcvtime = ticks;
706 	/*
707 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
708 	 * because the socket may be bound to an IPv6 wildcard address,
709 	 * which may match an IPv4-mapped IPv6 address.
710 	 */
711 	inp->inp_ip_ttl = ip_defttl;
712 	inp->inp_ppcb = tp;
713 	tcp_sack_tcpcb_init(tp);
714 	return (tp);		/* XXX */
715 }
716 
717 /*
718  * Drop a TCP connection, reporting the specified error.
719  * If connection is synchronized, then send a RST to peer.
720  */
721 struct tcpcb *
722 tcp_drop(struct tcpcb *tp, int error)
723 {
724 	struct socket *so = tp->t_inpcb->inp_socket;
725 
726 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
727 		tp->t_state = TCPS_CLOSED;
728 		tcp_output(tp);
729 		tcpstat.tcps_drops++;
730 	} else
731 		tcpstat.tcps_conndrops++;
732 	if (error == ETIMEDOUT && tp->t_softerror)
733 		error = tp->t_softerror;
734 	so->so_error = error;
735 	return (tcp_close(tp));
736 }
737 
738 #ifdef SMP
739 
740 struct netmsg_remwildcard {
741 	struct netmsg		nm_netmsg;
742 	struct inpcb		*nm_inp;
743 	struct inpcbinfo	*nm_pcbinfo;
744 #if defined(INET6)
745 	int			nm_isinet6;
746 #else
747 	int			nm_unused01;
748 #endif
749 };
750 
751 /*
752  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
753  * inp can be detached.  We do this by cycling through the cpus, ending up
754  * on the cpu controlling the inp last and then doing the disconnect.
755  */
756 static void
757 in_pcbremwildcardhash_handler(struct netmsg *msg0)
758 {
759 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
760 	int cpu;
761 
762 	cpu = msg->nm_pcbinfo->cpu;
763 
764 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
765 		/* note: detach removes any wildcard hash entry */
766 #ifdef INET6
767 		if (msg->nm_isinet6)
768 			in6_pcbdetach(msg->nm_inp);
769 		else
770 #endif
771 			in_pcbdetach(msg->nm_inp);
772 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
773 	} else {
774 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
775 		cpu = (cpu + 1) % ncpus2;
776 		msg->nm_pcbinfo = &tcbinfo[cpu];
777 		lwkt_forwardmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
778 	}
779 }
780 
781 #endif
782 
783 /*
784  * Close a TCP control block:
785  *	discard all space held by the tcp
786  *	discard internet protocol block
787  *	wake up any sleepers
788  */
789 struct tcpcb *
790 tcp_close(struct tcpcb *tp)
791 {
792 	struct tseg_qent *q;
793 	struct inpcb *inp = tp->t_inpcb;
794 	struct socket *so = inp->inp_socket;
795 	struct rtentry *rt;
796 	boolean_t dosavessthresh;
797 #ifdef SMP
798 	int cpu;
799 #endif
800 #ifdef INET6
801 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
802 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
803 #else
804 	const boolean_t isipv6 = FALSE;
805 #endif
806 
807 	/*
808 	 * The tp is not instantly destroyed in the wildcard case.  Setting
809 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
810 	 * messing with it, though it should be noted that this change may
811 	 * not take effect on other cpus until we have chained the wildcard
812 	 * hash removal.
813 	 *
814 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
815 	 * update and prevent other tcp protocol threads from accepting new
816 	 * connections on the listen socket we might be trying to close down.
817 	 */
818 	KKASSERT(tp->t_state != TCPS_TERMINATING);
819 	tp->t_state = TCPS_TERMINATING;
820 
821 	/*
822 	 * Make sure that all of our timers are stopped before we
823 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
824 	 * timers are never used.  If timer message is never created
825 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
826 	 */
827 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
828 		tcp_callout_stop(tp, tp->tt_rexmt);
829 		tcp_callout_stop(tp, tp->tt_persist);
830 		tcp_callout_stop(tp, tp->tt_keep);
831 		tcp_callout_stop(tp, tp->tt_2msl);
832 		tcp_callout_stop(tp, tp->tt_delack);
833 	}
834 
835 	if (tp->t_flags & TF_ONOUTPUTQ) {
836 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
837 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
838 		tp->t_flags &= ~TF_ONOUTPUTQ;
839 	}
840 
841 	/*
842 	 * If we got enough samples through the srtt filter,
843 	 * save the rtt and rttvar in the routing entry.
844 	 * 'Enough' is arbitrarily defined as the 16 samples.
845 	 * 16 samples is enough for the srtt filter to converge
846 	 * to within 5% of the correct value; fewer samples and
847 	 * we could save a very bogus rtt.
848 	 *
849 	 * Don't update the default route's characteristics and don't
850 	 * update anything that the user "locked".
851 	 */
852 	if (tp->t_rttupdated >= 16) {
853 		u_long i = 0;
854 
855 		if (isipv6) {
856 			struct sockaddr_in6 *sin6;
857 
858 			if ((rt = inp->in6p_route.ro_rt) == NULL)
859 				goto no_valid_rt;
860 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
861 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
862 				goto no_valid_rt;
863 		} else
864 			if ((rt = inp->inp_route.ro_rt) == NULL ||
865 			    ((struct sockaddr_in *)rt_key(rt))->
866 			     sin_addr.s_addr == INADDR_ANY)
867 				goto no_valid_rt;
868 
869 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
870 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
871 			if (rt->rt_rmx.rmx_rtt && i)
872 				/*
873 				 * filter this update to half the old & half
874 				 * the new values, converting scale.
875 				 * See route.h and tcp_var.h for a
876 				 * description of the scaling constants.
877 				 */
878 				rt->rt_rmx.rmx_rtt =
879 				    (rt->rt_rmx.rmx_rtt + i) / 2;
880 			else
881 				rt->rt_rmx.rmx_rtt = i;
882 			tcpstat.tcps_cachedrtt++;
883 		}
884 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
885 			i = tp->t_rttvar *
886 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
887 			if (rt->rt_rmx.rmx_rttvar && i)
888 				rt->rt_rmx.rmx_rttvar =
889 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
890 			else
891 				rt->rt_rmx.rmx_rttvar = i;
892 			tcpstat.tcps_cachedrttvar++;
893 		}
894 		/*
895 		 * The old comment here said:
896 		 * update the pipelimit (ssthresh) if it has been updated
897 		 * already or if a pipesize was specified & the threshhold
898 		 * got below half the pipesize.  I.e., wait for bad news
899 		 * before we start updating, then update on both good
900 		 * and bad news.
901 		 *
902 		 * But we want to save the ssthresh even if no pipesize is
903 		 * specified explicitly in the route, because such
904 		 * connections still have an implicit pipesize specified
905 		 * by the global tcp_sendspace.  In the absence of a reliable
906 		 * way to calculate the pipesize, it will have to do.
907 		 */
908 		i = tp->snd_ssthresh;
909 		if (rt->rt_rmx.rmx_sendpipe != 0)
910 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
911 		else
912 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
913 		if (dosavessthresh ||
914 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
915 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
916 			/*
917 			 * convert the limit from user data bytes to
918 			 * packets then to packet data bytes.
919 			 */
920 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
921 			if (i < 2)
922 				i = 2;
923 			i *= tp->t_maxseg +
924 			     (isipv6 ?
925 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
926 			      sizeof(struct tcpiphdr));
927 			if (rt->rt_rmx.rmx_ssthresh)
928 				rt->rt_rmx.rmx_ssthresh =
929 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
930 			else
931 				rt->rt_rmx.rmx_ssthresh = i;
932 			tcpstat.tcps_cachedssthresh++;
933 		}
934 	}
935 
936 no_valid_rt:
937 	/* free the reassembly queue, if any */
938 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
939 		LIST_REMOVE(q, tqe_q);
940 		m_freem(q->tqe_m);
941 		FREE(q, M_TSEGQ);
942 		atomic_add_int(&tcp_reass_qsize, -1);
943 	}
944 	/* throw away SACK blocks in scoreboard*/
945 	if (TCP_DO_SACK(tp))
946 		tcp_sack_cleanup(&tp->scb);
947 
948 	inp->inp_ppcb = NULL;
949 	soisdisconnected(so);
950 	/* note: pcb detached later on */
951 
952 	tcp_destroy_timermsg(tp);
953 	if (tp->t_flags & TF_SYNCACHE)
954 		syncache_destroy(tp);
955 
956 	/*
957 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
958 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
959 	 * for each protocol thread and must be removed in the context of
960 	 * that thread.  This is accomplished by chaining the message
961 	 * through the cpus.
962 	 *
963 	 * If the inp is not wildcarded we simply detach, which will remove
964 	 * the any hashes still present for this inp.
965 	 */
966 #ifdef SMP
967 	if (inp->inp_flags & INP_WILDCARD_MP) {
968 		struct netmsg_remwildcard *msg;
969 
970 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
971 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
972 			      M_LWKTMSG, M_INTWAIT);
973 		netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
974 			    0, in_pcbremwildcardhash_handler);
975 #ifdef INET6
976 		msg->nm_isinet6 = isafinet6;
977 #endif
978 		msg->nm_inp = inp;
979 		msg->nm_pcbinfo = &tcbinfo[cpu];
980 		lwkt_sendmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
981 	} else
982 #endif
983 	{
984 		/* note: detach removes any wildcard hash entry */
985 #ifdef INET6
986 		if (isafinet6)
987 			in6_pcbdetach(inp);
988 		else
989 #endif
990 			in_pcbdetach(inp);
991 	}
992 	tcpstat.tcps_closed++;
993 	return (NULL);
994 }
995 
996 static __inline void
997 tcp_drain_oncpu(struct inpcbhead *head)
998 {
999 	struct inpcb *marker;
1000 	struct inpcb *inpb;
1001 	struct tcpcb *tcpb;
1002 	struct tseg_qent *te;
1003 
1004 	/*
1005 	 * Allows us to block while running the list
1006 	 */
1007 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1008 	marker->inp_flags |= INP_PLACEMARKER;
1009 	LIST_INSERT_HEAD(head, marker, inp_list);
1010 
1011 	while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1012 		if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1013 		    (tcpb = intotcpcb(inpb)) != NULL &&
1014 		    (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1015 			LIST_REMOVE(te, tqe_q);
1016 			m_freem(te->tqe_m);
1017 			FREE(te, M_TSEGQ);
1018 			atomic_add_int(&tcp_reass_qsize, -1);
1019 			/* retry */
1020 		} else {
1021 			LIST_REMOVE(marker, inp_list);
1022 			LIST_INSERT_AFTER(inpb, marker, inp_list);
1023 		}
1024 	}
1025 	LIST_REMOVE(marker, inp_list);
1026 	kfree(marker, M_TEMP);
1027 }
1028 
1029 #ifdef SMP
1030 struct netmsg_tcp_drain {
1031 	struct netmsg		nm_netmsg;
1032 	struct inpcbhead	*nm_head;
1033 };
1034 
1035 static void
1036 tcp_drain_handler(netmsg_t netmsg)
1037 {
1038 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1039 
1040 	tcp_drain_oncpu(nm->nm_head);
1041 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1042 }
1043 #endif
1044 
1045 void
1046 tcp_drain(void)
1047 {
1048 #ifdef SMP
1049 	int cpu;
1050 #endif
1051 
1052 	if (!do_tcpdrain)
1053 		return;
1054 
1055 	/*
1056 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1057 	 * if there is one...
1058 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1059 	 *	reassembly queue should be flushed, but in a situation
1060 	 *	where we're really low on mbufs, this is potentially
1061 	 *	useful.
1062 	 */
1063 #ifdef SMP
1064 	for (cpu = 0; cpu < ncpus2; cpu++) {
1065 		struct netmsg_tcp_drain *msg;
1066 
1067 		if (cpu == mycpu->gd_cpuid) {
1068 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1069 		} else {
1070 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1071 				    M_LWKTMSG, M_NOWAIT);
1072 			if (msg == NULL)
1073 				continue;
1074 			netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1075 				    0, tcp_drain_handler);
1076 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1077 			lwkt_sendmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
1078 		}
1079 	}
1080 #else
1081 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1082 #endif
1083 }
1084 
1085 /*
1086  * Notify a tcp user of an asynchronous error;
1087  * store error as soft error, but wake up user
1088  * (for now, won't do anything until can select for soft error).
1089  *
1090  * Do not wake up user since there currently is no mechanism for
1091  * reporting soft errors (yet - a kqueue filter may be added).
1092  */
1093 static void
1094 tcp_notify(struct inpcb *inp, int error)
1095 {
1096 	struct tcpcb *tp = intotcpcb(inp);
1097 
1098 	/*
1099 	 * Ignore some errors if we are hooked up.
1100 	 * If connection hasn't completed, has retransmitted several times,
1101 	 * and receives a second error, give up now.  This is better
1102 	 * than waiting a long time to establish a connection that
1103 	 * can never complete.
1104 	 */
1105 	if (tp->t_state == TCPS_ESTABLISHED &&
1106 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1107 	      error == EHOSTDOWN)) {
1108 		return;
1109 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1110 	    tp->t_softerror)
1111 		tcp_drop(tp, error);
1112 	else
1113 		tp->t_softerror = error;
1114 #if 0
1115 	wakeup(&so->so_timeo);
1116 	sorwakeup(so);
1117 	sowwakeup(so);
1118 #endif
1119 }
1120 
1121 static int
1122 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1123 {
1124 	int error, i, n;
1125 	struct inpcb *marker;
1126 	struct inpcb *inp;
1127 	inp_gen_t gencnt;
1128 	globaldata_t gd;
1129 	int origcpu, ccpu;
1130 
1131 	error = 0;
1132 	n = 0;
1133 
1134 	/*
1135 	 * The process of preparing the TCB list is too time-consuming and
1136 	 * resource-intensive to repeat twice on every request.
1137 	 */
1138 	if (req->oldptr == NULL) {
1139 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1140 			gd = globaldata_find(ccpu);
1141 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1142 		}
1143 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1144 		return (0);
1145 	}
1146 
1147 	if (req->newptr != NULL)
1148 		return (EPERM);
1149 
1150 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1151 	marker->inp_flags |= INP_PLACEMARKER;
1152 
1153 	/*
1154 	 * OK, now we're committed to doing something.  Run the inpcb list
1155 	 * for each cpu in the system and construct the output.  Use a
1156 	 * list placemarker to deal with list changes occuring during
1157 	 * copyout blockages (but otherwise depend on being on the correct
1158 	 * cpu to avoid races).
1159 	 */
1160 	origcpu = mycpu->gd_cpuid;
1161 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1162 		globaldata_t rgd;
1163 		caddr_t inp_ppcb;
1164 		struct xtcpcb xt;
1165 		int cpu_id;
1166 
1167 		cpu_id = (origcpu + ccpu) % ncpus;
1168 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1169 			continue;
1170 		rgd = globaldata_find(cpu_id);
1171 		lwkt_setcpu_self(rgd);
1172 
1173 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1174 		n = tcbinfo[cpu_id].ipi_count;
1175 
1176 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1177 		i = 0;
1178 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1179 			/*
1180 			 * process a snapshot of pcbs, ignoring placemarkers
1181 			 * and using our own to allow SYSCTL_OUT to block.
1182 			 */
1183 			LIST_REMOVE(marker, inp_list);
1184 			LIST_INSERT_AFTER(inp, marker, inp_list);
1185 
1186 			if (inp->inp_flags & INP_PLACEMARKER)
1187 				continue;
1188 			if (inp->inp_gencnt > gencnt)
1189 				continue;
1190 			if (prison_xinpcb(req->td, inp))
1191 				continue;
1192 
1193 			xt.xt_len = sizeof xt;
1194 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1195 			inp_ppcb = inp->inp_ppcb;
1196 			if (inp_ppcb != NULL)
1197 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1198 			else
1199 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1200 			if (inp->inp_socket)
1201 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1202 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1203 				break;
1204 			++i;
1205 		}
1206 		LIST_REMOVE(marker, inp_list);
1207 		if (error == 0 && i < n) {
1208 			bzero(&xt, sizeof xt);
1209 			xt.xt_len = sizeof xt;
1210 			while (i < n) {
1211 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1212 				if (error)
1213 					break;
1214 				++i;
1215 			}
1216 		}
1217 	}
1218 
1219 	/*
1220 	 * Make sure we are on the same cpu we were on originally, since
1221 	 * higher level callers expect this.  Also don't pollute caches with
1222 	 * migrated userland data by (eventually) returning to userland
1223 	 * on a different cpu.
1224 	 */
1225 	lwkt_setcpu_self(globaldata_find(origcpu));
1226 	kfree(marker, M_TEMP);
1227 	return (error);
1228 }
1229 
1230 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1231 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1232 
1233 static int
1234 tcp_getcred(SYSCTL_HANDLER_ARGS)
1235 {
1236 	struct sockaddr_in addrs[2];
1237 	struct inpcb *inp;
1238 	int cpu;
1239 	int error;
1240 
1241 	error = priv_check(req->td, PRIV_ROOT);
1242 	if (error != 0)
1243 		return (error);
1244 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1245 	if (error != 0)
1246 		return (error);
1247 	crit_enter();
1248 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1249 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1250 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1251 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1252 	if (inp == NULL || inp->inp_socket == NULL) {
1253 		error = ENOENT;
1254 		goto out;
1255 	}
1256 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1257 out:
1258 	crit_exit();
1259 	return (error);
1260 }
1261 
1262 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1263     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1264 
1265 #ifdef INET6
1266 static int
1267 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1268 {
1269 	struct sockaddr_in6 addrs[2];
1270 	struct inpcb *inp;
1271 	int error;
1272 	boolean_t mapped = FALSE;
1273 
1274 	error = priv_check(req->td, PRIV_ROOT);
1275 	if (error != 0)
1276 		return (error);
1277 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1278 	if (error != 0)
1279 		return (error);
1280 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1281 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1282 			mapped = TRUE;
1283 		else
1284 			return (EINVAL);
1285 	}
1286 	crit_enter();
1287 	if (mapped) {
1288 		inp = in_pcblookup_hash(&tcbinfo[0],
1289 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1290 		    addrs[1].sin6_port,
1291 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1292 		    addrs[0].sin6_port,
1293 		    0, NULL);
1294 	} else {
1295 		inp = in6_pcblookup_hash(&tcbinfo[0],
1296 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1297 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1298 		    0, NULL);
1299 	}
1300 	if (inp == NULL || inp->inp_socket == NULL) {
1301 		error = ENOENT;
1302 		goto out;
1303 	}
1304 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1305 out:
1306 	crit_exit();
1307 	return (error);
1308 }
1309 
1310 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1311 	    0, 0,
1312 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1313 #endif
1314 
1315 struct netmsg_tcp_notify {
1316 	struct netmsg	nm_nmsg;
1317 	void		(*nm_notify)(struct inpcb *, int);
1318 	struct in_addr	nm_faddr;
1319 	int		nm_arg;
1320 };
1321 
1322 static void
1323 tcp_notifyall_oncpu(struct netmsg *netmsg)
1324 {
1325 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1326 	int nextcpu;
1327 
1328 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1329 			nmsg->nm_arg, nmsg->nm_notify);
1330 
1331 	nextcpu = mycpuid + 1;
1332 	if (nextcpu < ncpus2)
1333 		lwkt_forwardmsg(cpu_portfn(nextcpu), &netmsg->nm_lmsg);
1334 	else
1335 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1336 }
1337 
1338 void
1339 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1340 {
1341 	struct ip *ip = vip;
1342 	struct tcphdr *th;
1343 	struct in_addr faddr;
1344 	struct inpcb *inp;
1345 	struct tcpcb *tp;
1346 	void (*notify)(struct inpcb *, int) = tcp_notify;
1347 	tcp_seq icmpseq;
1348 	int arg, cpu;
1349 
1350 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1351 		return;
1352 	}
1353 
1354 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1355 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1356 		return;
1357 
1358 	arg = inetctlerrmap[cmd];
1359 	if (cmd == PRC_QUENCH) {
1360 		notify = tcp_quench;
1361 	} else if (icmp_may_rst &&
1362 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1363 		    cmd == PRC_UNREACH_PORT ||
1364 		    cmd == PRC_TIMXCEED_INTRANS) &&
1365 		   ip != NULL) {
1366 		notify = tcp_drop_syn_sent;
1367 	} else if (cmd == PRC_MSGSIZE) {
1368 		struct icmp *icmp = (struct icmp *)
1369 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1370 
1371 		arg = ntohs(icmp->icmp_nextmtu);
1372 		notify = tcp_mtudisc;
1373 	} else if (PRC_IS_REDIRECT(cmd)) {
1374 		ip = NULL;
1375 		notify = in_rtchange;
1376 	} else if (cmd == PRC_HOSTDEAD) {
1377 		ip = NULL;
1378 	}
1379 
1380 	if (ip != NULL) {
1381 		crit_enter();
1382 		th = (struct tcphdr *)((caddr_t)ip +
1383 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1384 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1385 				  ip->ip_src.s_addr, th->th_sport);
1386 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1387 					ip->ip_src, th->th_sport, 0, NULL);
1388 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1389 			icmpseq = htonl(th->th_seq);
1390 			tp = intotcpcb(inp);
1391 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1392 			    SEQ_LT(icmpseq, tp->snd_max))
1393 				(*notify)(inp, arg);
1394 		} else {
1395 			struct in_conninfo inc;
1396 
1397 			inc.inc_fport = th->th_dport;
1398 			inc.inc_lport = th->th_sport;
1399 			inc.inc_faddr = faddr;
1400 			inc.inc_laddr = ip->ip_src;
1401 #ifdef INET6
1402 			inc.inc_isipv6 = 0;
1403 #endif
1404 			syncache_unreach(&inc, th);
1405 		}
1406 		crit_exit();
1407 	} else {
1408 		struct netmsg_tcp_notify nmsg;
1409 
1410 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1411 		netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1412 			    0, tcp_notifyall_oncpu);
1413 		nmsg.nm_faddr = faddr;
1414 		nmsg.nm_arg = arg;
1415 		nmsg.nm_notify = notify;
1416 
1417 		lwkt_domsg(cpu_portfn(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1418 	}
1419 }
1420 
1421 #ifdef INET6
1422 void
1423 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1424 {
1425 	struct tcphdr th;
1426 	void (*notify) (struct inpcb *, int) = tcp_notify;
1427 	struct ip6_hdr *ip6;
1428 	struct mbuf *m;
1429 	struct ip6ctlparam *ip6cp = NULL;
1430 	const struct sockaddr_in6 *sa6_src = NULL;
1431 	int off;
1432 	struct tcp_portonly {
1433 		u_int16_t th_sport;
1434 		u_int16_t th_dport;
1435 	} *thp;
1436 	int arg;
1437 
1438 	if (sa->sa_family != AF_INET6 ||
1439 	    sa->sa_len != sizeof(struct sockaddr_in6))
1440 		return;
1441 
1442 	arg = 0;
1443 	if (cmd == PRC_QUENCH)
1444 		notify = tcp_quench;
1445 	else if (cmd == PRC_MSGSIZE) {
1446 		struct ip6ctlparam *ip6cp = d;
1447 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1448 
1449 		arg = ntohl(icmp6->icmp6_mtu);
1450 		notify = tcp_mtudisc;
1451 	} else if (!PRC_IS_REDIRECT(cmd) &&
1452 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1453 		return;
1454 	}
1455 
1456 	/* if the parameter is from icmp6, decode it. */
1457 	if (d != NULL) {
1458 		ip6cp = (struct ip6ctlparam *)d;
1459 		m = ip6cp->ip6c_m;
1460 		ip6 = ip6cp->ip6c_ip6;
1461 		off = ip6cp->ip6c_off;
1462 		sa6_src = ip6cp->ip6c_src;
1463 	} else {
1464 		m = NULL;
1465 		ip6 = NULL;
1466 		off = 0;	/* fool gcc */
1467 		sa6_src = &sa6_any;
1468 	}
1469 
1470 	if (ip6 != NULL) {
1471 		struct in_conninfo inc;
1472 		/*
1473 		 * XXX: We assume that when IPV6 is non NULL,
1474 		 * M and OFF are valid.
1475 		 */
1476 
1477 		/* check if we can safely examine src and dst ports */
1478 		if (m->m_pkthdr.len < off + sizeof *thp)
1479 			return;
1480 
1481 		bzero(&th, sizeof th);
1482 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1483 
1484 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1485 		    (struct sockaddr *)ip6cp->ip6c_src,
1486 		    th.th_sport, cmd, arg, notify);
1487 
1488 		inc.inc_fport = th.th_dport;
1489 		inc.inc_lport = th.th_sport;
1490 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1491 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1492 		inc.inc_isipv6 = 1;
1493 		syncache_unreach(&inc, &th);
1494 	} else
1495 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1496 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1497 }
1498 #endif
1499 
1500 /*
1501  * Following is where TCP initial sequence number generation occurs.
1502  *
1503  * There are two places where we must use initial sequence numbers:
1504  * 1.  In SYN-ACK packets.
1505  * 2.  In SYN packets.
1506  *
1507  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1508  * tcp_syncache.c for details.
1509  *
1510  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1511  * depends on this property.  In addition, these ISNs should be
1512  * unguessable so as to prevent connection hijacking.  To satisfy
1513  * the requirements of this situation, the algorithm outlined in
1514  * RFC 1948 is used to generate sequence numbers.
1515  *
1516  * Implementation details:
1517  *
1518  * Time is based off the system timer, and is corrected so that it
1519  * increases by one megabyte per second.  This allows for proper
1520  * recycling on high speed LANs while still leaving over an hour
1521  * before rollover.
1522  *
1523  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1524  * between seeding of isn_secret.  This is normally set to zero,
1525  * as reseeding should not be necessary.
1526  *
1527  */
1528 
1529 #define	ISN_BYTES_PER_SECOND 1048576
1530 
1531 u_char isn_secret[32];
1532 int isn_last_reseed;
1533 MD5_CTX isn_ctx;
1534 
1535 tcp_seq
1536 tcp_new_isn(struct tcpcb *tp)
1537 {
1538 	u_int32_t md5_buffer[4];
1539 	tcp_seq new_isn;
1540 
1541 	/* Seed if this is the first use, reseed if requested. */
1542 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1543 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1544 		< (u_int)ticks))) {
1545 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1546 		isn_last_reseed = ticks;
1547 	}
1548 
1549 	/* Compute the md5 hash and return the ISN. */
1550 	MD5Init(&isn_ctx);
1551 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1552 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1553 #ifdef INET6
1554 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1555 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1556 			  sizeof(struct in6_addr));
1557 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1558 			  sizeof(struct in6_addr));
1559 	} else
1560 #endif
1561 	{
1562 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1563 			  sizeof(struct in_addr));
1564 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1565 			  sizeof(struct in_addr));
1566 	}
1567 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1568 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1569 	new_isn = (tcp_seq) md5_buffer[0];
1570 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1571 	return (new_isn);
1572 }
1573 
1574 /*
1575  * When a source quench is received, close congestion window
1576  * to one segment.  We will gradually open it again as we proceed.
1577  */
1578 void
1579 tcp_quench(struct inpcb *inp, int error)
1580 {
1581 	struct tcpcb *tp = intotcpcb(inp);
1582 
1583 	if (tp != NULL) {
1584 		tp->snd_cwnd = tp->t_maxseg;
1585 		tp->snd_wacked = 0;
1586 	}
1587 }
1588 
1589 /*
1590  * When a specific ICMP unreachable message is received and the
1591  * connection state is SYN-SENT, drop the connection.  This behavior
1592  * is controlled by the icmp_may_rst sysctl.
1593  */
1594 void
1595 tcp_drop_syn_sent(struct inpcb *inp, int error)
1596 {
1597 	struct tcpcb *tp = intotcpcb(inp);
1598 
1599 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1600 		tcp_drop(tp, error);
1601 }
1602 
1603 /*
1604  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1605  * based on the new value in the route.  Also nudge TCP to send something,
1606  * since we know the packet we just sent was dropped.
1607  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1608  */
1609 void
1610 tcp_mtudisc(struct inpcb *inp, int mtu)
1611 {
1612 	struct tcpcb *tp = intotcpcb(inp);
1613 	struct rtentry *rt;
1614 	struct socket *so = inp->inp_socket;
1615 	int maxopd, mss;
1616 #ifdef INET6
1617 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1618 #else
1619 	const boolean_t isipv6 = FALSE;
1620 #endif
1621 
1622 	if (tp == NULL)
1623 		return;
1624 
1625 	/*
1626 	 * If no MTU is provided in the ICMP message, use the
1627 	 * next lower likely value, as specified in RFC 1191.
1628 	 */
1629 	if (mtu == 0) {
1630 		int oldmtu;
1631 
1632 		oldmtu = tp->t_maxopd +
1633 		    (isipv6 ?
1634 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1635 		     sizeof(struct tcpiphdr));
1636 		mtu = ip_next_mtu(oldmtu, 0);
1637 	}
1638 
1639 	if (isipv6)
1640 		rt = tcp_rtlookup6(&inp->inp_inc);
1641 	else
1642 		rt = tcp_rtlookup(&inp->inp_inc);
1643 	if (rt != NULL) {
1644 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1645 			mtu = rt->rt_rmx.rmx_mtu;
1646 
1647 		maxopd = mtu -
1648 		    (isipv6 ?
1649 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1650 		     sizeof(struct tcpiphdr));
1651 
1652 		/*
1653 		 * XXX - The following conditional probably violates the TCP
1654 		 * spec.  The problem is that, since we don't know the
1655 		 * other end's MSS, we are supposed to use a conservative
1656 		 * default.  But, if we do that, then MTU discovery will
1657 		 * never actually take place, because the conservative
1658 		 * default is much less than the MTUs typically seen
1659 		 * on the Internet today.  For the moment, we'll sweep
1660 		 * this under the carpet.
1661 		 *
1662 		 * The conservative default might not actually be a problem
1663 		 * if the only case this occurs is when sending an initial
1664 		 * SYN with options and data to a host we've never talked
1665 		 * to before.  Then, they will reply with an MSS value which
1666 		 * will get recorded and the new parameters should get
1667 		 * recomputed.  For Further Study.
1668 		 */
1669 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1670 			maxopd = rt->rt_rmx.rmx_mssopt;
1671 	} else
1672 		maxopd = mtu -
1673 		    (isipv6 ?
1674 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1675 		     sizeof(struct tcpiphdr));
1676 
1677 	if (tp->t_maxopd <= maxopd)
1678 		return;
1679 	tp->t_maxopd = maxopd;
1680 
1681 	mss = maxopd;
1682 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1683 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1684 		mss -= TCPOLEN_TSTAMP_APPA;
1685 
1686 	/* round down to multiple of MCLBYTES */
1687 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1688 	if (mss > MCLBYTES)
1689 		mss &= ~(MCLBYTES - 1);
1690 #else
1691 	if (mss > MCLBYTES)
1692 		mss = (mss / MCLBYTES) * MCLBYTES;
1693 #endif
1694 
1695 	if (so->so_snd.ssb_hiwat < mss)
1696 		mss = so->so_snd.ssb_hiwat;
1697 
1698 	tp->t_maxseg = mss;
1699 	tp->t_rtttime = 0;
1700 	tp->snd_nxt = tp->snd_una;
1701 	tcp_output(tp);
1702 	tcpstat.tcps_mturesent++;
1703 }
1704 
1705 /*
1706  * Look-up the routing entry to the peer of this inpcb.  If no route
1707  * is found and it cannot be allocated the return NULL.  This routine
1708  * is called by TCP routines that access the rmx structure and by tcp_mss
1709  * to get the interface MTU.
1710  */
1711 struct rtentry *
1712 tcp_rtlookup(struct in_conninfo *inc)
1713 {
1714 	struct route *ro = &inc->inc_route;
1715 
1716 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1717 		/* No route yet, so try to acquire one */
1718 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1719 			/*
1720 			 * unused portions of the structure MUST be zero'd
1721 			 * out because rtalloc() treats it as opaque data
1722 			 */
1723 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1724 			ro->ro_dst.sa_family = AF_INET;
1725 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1726 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1727 			    inc->inc_faddr;
1728 			rtalloc(ro);
1729 		}
1730 	}
1731 	return (ro->ro_rt);
1732 }
1733 
1734 #ifdef INET6
1735 struct rtentry *
1736 tcp_rtlookup6(struct in_conninfo *inc)
1737 {
1738 	struct route_in6 *ro6 = &inc->inc6_route;
1739 
1740 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1741 		/* No route yet, so try to acquire one */
1742 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1743 			/*
1744 			 * unused portions of the structure MUST be zero'd
1745 			 * out because rtalloc() treats it as opaque data
1746 			 */
1747 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1748 			ro6->ro_dst.sin6_family = AF_INET6;
1749 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1750 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1751 			rtalloc((struct route *)ro6);
1752 		}
1753 	}
1754 	return (ro6->ro_rt);
1755 }
1756 #endif
1757 
1758 #ifdef IPSEC
1759 /* compute ESP/AH header size for TCP, including outer IP header. */
1760 size_t
1761 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1762 {
1763 	struct inpcb *inp;
1764 	struct mbuf *m;
1765 	size_t hdrsiz;
1766 	struct ip *ip;
1767 	struct tcphdr *th;
1768 
1769 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1770 		return (0);
1771 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1772 	if (!m)
1773 		return (0);
1774 
1775 #ifdef INET6
1776 	if (inp->inp_vflag & INP_IPV6) {
1777 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1778 
1779 		th = (struct tcphdr *)(ip6 + 1);
1780 		m->m_pkthdr.len = m->m_len =
1781 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1782 		tcp_fillheaders(tp, ip6, th);
1783 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1784 	} else
1785 #endif
1786 	{
1787 		ip = mtod(m, struct ip *);
1788 		th = (struct tcphdr *)(ip + 1);
1789 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1790 		tcp_fillheaders(tp, ip, th);
1791 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1792 	}
1793 
1794 	m_free(m);
1795 	return (hdrsiz);
1796 }
1797 #endif
1798 
1799 /*
1800  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1801  *
1802  * This code attempts to calculate the bandwidth-delay product as a
1803  * means of determining the optimal window size to maximize bandwidth,
1804  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1805  * routers.  This code also does a fairly good job keeping RTTs in check
1806  * across slow links like modems.  We implement an algorithm which is very
1807  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1808  * transmitter side of a TCP connection and so only effects the transmit
1809  * side of the connection.
1810  *
1811  * BACKGROUND:  TCP makes no provision for the management of buffer space
1812  * at the end points or at the intermediate routers and switches.  A TCP
1813  * stream, whether using NewReno or not, will eventually buffer as
1814  * many packets as it is able and the only reason this typically works is
1815  * due to the fairly small default buffers made available for a connection
1816  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1817  * scaling it is now fairly easy for even a single TCP connection to blow-out
1818  * all available buffer space not only on the local interface, but on
1819  * intermediate routers and switches as well.  NewReno makes a misguided
1820  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1821  * then backing off, then steadily increasing the window again until another
1822  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1823  * is only made worse as network loads increase and the idea of intentionally
1824  * blowing out network buffers is, frankly, a terrible way to manage network
1825  * resources.
1826  *
1827  * It is far better to limit the transmit window prior to the failure
1828  * condition being achieved.  There are two general ways to do this:  First
1829  * you can 'scan' through different transmit window sizes and locate the
1830  * point where the RTT stops increasing, indicating that you have filled the
1831  * pipe, then scan backwards until you note that RTT stops decreasing, then
1832  * repeat ad-infinitum.  This method works in principle but has severe
1833  * implementation issues due to RTT variances, timer granularity, and
1834  * instability in the algorithm which can lead to many false positives and
1835  * create oscillations as well as interact badly with other TCP streams
1836  * implementing the same algorithm.
1837  *
1838  * The second method is to limit the window to the bandwidth delay product
1839  * of the link.  This is the method we implement.  RTT variances and our
1840  * own manipulation of the congestion window, bwnd, can potentially
1841  * destabilize the algorithm.  For this reason we have to stabilize the
1842  * elements used to calculate the window.  We do this by using the minimum
1843  * observed RTT, the long term average of the observed bandwidth, and
1844  * by adding two segments worth of slop.  It isn't perfect but it is able
1845  * to react to changing conditions and gives us a very stable basis on
1846  * which to extend the algorithm.
1847  */
1848 void
1849 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1850 {
1851 	u_long bw;
1852 	u_long bwnd;
1853 	int save_ticks;
1854 	int delta_ticks;
1855 
1856 	/*
1857 	 * If inflight_enable is disabled in the middle of a tcp connection,
1858 	 * make sure snd_bwnd is effectively disabled.
1859 	 */
1860 	if (!tcp_inflight_enable) {
1861 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1862 		tp->snd_bandwidth = 0;
1863 		return;
1864 	}
1865 
1866 	/*
1867 	 * Validate the delta time.  If a connection is new or has been idle
1868 	 * a long time we have to reset the bandwidth calculator.
1869 	 */
1870 	save_ticks = ticks;
1871 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1872 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1873 		tp->t_bw_rtttime = ticks;
1874 		tp->t_bw_rtseq = ack_seq;
1875 		if (tp->snd_bandwidth == 0)
1876 			tp->snd_bandwidth = tcp_inflight_min;
1877 		return;
1878 	}
1879 	if (delta_ticks == 0)
1880 		return;
1881 
1882 	/*
1883 	 * Sanity check, plus ignore pure window update acks.
1884 	 */
1885 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1886 		return;
1887 
1888 	/*
1889 	 * Figure out the bandwidth.  Due to the tick granularity this
1890 	 * is a very rough number and it MUST be averaged over a fairly
1891 	 * long period of time.  XXX we need to take into account a link
1892 	 * that is not using all available bandwidth, but for now our
1893 	 * slop will ramp us up if this case occurs and the bandwidth later
1894 	 * increases.
1895 	 */
1896 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1897 	tp->t_bw_rtttime = save_ticks;
1898 	tp->t_bw_rtseq = ack_seq;
1899 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1900 
1901 	tp->snd_bandwidth = bw;
1902 
1903 	/*
1904 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1905 	 * segments.  The additional slop puts us squarely in the sweet
1906 	 * spot and also handles the bandwidth run-up case.  Without the
1907 	 * slop we could be locking ourselves into a lower bandwidth.
1908 	 *
1909 	 * Situations Handled:
1910 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1911 	 *	    high speed LANs, allowing larger TCP buffers to be
1912 	 *	    specified, and also does a good job preventing
1913 	 *	    over-queueing of packets over choke points like modems
1914 	 *	    (at least for the transmit side).
1915 	 *
1916 	 *	(2) Is able to handle changing network loads (bandwidth
1917 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1918 	 *	    increases).
1919 	 *
1920 	 *	(3) Theoretically should stabilize in the face of multiple
1921 	 *	    connections implementing the same algorithm (this may need
1922 	 *	    a little work).
1923 	 *
1924 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1925 	 *	    be adjusted with a sysctl but typically only needs to be on
1926 	 *	    very slow connections.  A value no smaller then 5 should
1927 	 *	    be used, but only reduce this default if you have no other
1928 	 *	    choice.
1929 	 */
1930 
1931 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1932 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1933 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1934 #undef USERTT
1935 
1936 	if (tcp_inflight_debug > 0) {
1937 		static int ltime;
1938 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1939 			ltime = ticks;
1940 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1941 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1942 		}
1943 	}
1944 	if ((long)bwnd < tcp_inflight_min)
1945 		bwnd = tcp_inflight_min;
1946 	if (bwnd > tcp_inflight_max)
1947 		bwnd = tcp_inflight_max;
1948 	if ((long)bwnd < tp->t_maxseg * 2)
1949 		bwnd = tp->t_maxseg * 2;
1950 	tp->snd_bwnd = bwnd;
1951 }
1952 
1953 #ifdef TCP_SIGNATURE
1954 /*
1955  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1956  *
1957  * We do this over ip, tcphdr, segment data, and the key in the SADB.
1958  * When called from tcp_input(), we can be sure that th_sum has been
1959  * zeroed out and verified already.
1960  *
1961  * Return 0 if successful, otherwise return -1.
1962  *
1963  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1964  * search with the destination IP address, and a 'magic SPI' to be
1965  * determined by the application. This is hardcoded elsewhere to 1179
1966  * right now. Another branch of this code exists which uses the SPD to
1967  * specify per-application flows but it is unstable.
1968  */
1969 int
1970 tcpsignature_compute(
1971 	struct mbuf *m,		/* mbuf chain */
1972 	int len,		/* length of TCP data */
1973 	int optlen,		/* length of TCP options */
1974 	u_char *buf,		/* storage for MD5 digest */
1975 	u_int direction)	/* direction of flow */
1976 {
1977 	struct ippseudo ippseudo;
1978 	MD5_CTX ctx;
1979 	int doff;
1980 	struct ip *ip;
1981 	struct ipovly *ipovly;
1982 	struct secasvar *sav;
1983 	struct tcphdr *th;
1984 #ifdef INET6
1985 	struct ip6_hdr *ip6;
1986 	struct in6_addr in6;
1987 	uint32_t plen;
1988 	uint16_t nhdr;
1989 #endif /* INET6 */
1990 	u_short savecsum;
1991 
1992 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
1993 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
1994 	/*
1995 	 * Extract the destination from the IP header in the mbuf.
1996 	 */
1997 	ip = mtod(m, struct ip *);
1998 #ifdef INET6
1999 	ip6 = NULL;     /* Make the compiler happy. */
2000 #endif /* INET6 */
2001 	/*
2002 	 * Look up an SADB entry which matches the address found in
2003 	 * the segment.
2004 	 */
2005 	switch (IP_VHL_V(ip->ip_vhl)) {
2006 	case IPVERSION:
2007 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2008 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2009 		break;
2010 #ifdef INET6
2011 	case (IPV6_VERSION >> 4):
2012 		ip6 = mtod(m, struct ip6_hdr *);
2013 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2014 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2015 		break;
2016 #endif /* INET6 */
2017 	default:
2018 		return (EINVAL);
2019 		/* NOTREACHED */
2020 		break;
2021 	}
2022 	if (sav == NULL) {
2023 		kprintf("%s: SADB lookup failed\n", __func__);
2024 		return (EINVAL);
2025 	}
2026 	MD5Init(&ctx);
2027 
2028 	/*
2029 	 * Step 1: Update MD5 hash with IP pseudo-header.
2030 	 *
2031 	 * XXX The ippseudo header MUST be digested in network byte order,
2032 	 * or else we'll fail the regression test. Assume all fields we've
2033 	 * been doing arithmetic on have been in host byte order.
2034 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2035 	 * tcp_output(), the underlying ip_len member has not yet been set.
2036 	 */
2037 	switch (IP_VHL_V(ip->ip_vhl)) {
2038 	case IPVERSION:
2039 		ipovly = (struct ipovly *)ip;
2040 		ippseudo.ippseudo_src = ipovly->ih_src;
2041 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2042 		ippseudo.ippseudo_pad = 0;
2043 		ippseudo.ippseudo_p = IPPROTO_TCP;
2044 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2045 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2046 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2047 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2048 		break;
2049 #ifdef INET6
2050 	/*
2051 	 * RFC 2385, 2.0  Proposal
2052 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2053 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2054 	 * extended next header value (to form 32 bits), and 32-bit segment
2055 	 * length.
2056 	 * Note: Upper-Layer Packet Length comes before Next Header.
2057 	 */
2058 	case (IPV6_VERSION >> 4):
2059 		in6 = ip6->ip6_src;
2060 		in6_clearscope(&in6);
2061 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2062 		in6 = ip6->ip6_dst;
2063 		in6_clearscope(&in6);
2064 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2065 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2066 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2067 		nhdr = 0;
2068 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2069 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2070 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2071 		nhdr = IPPROTO_TCP;
2072 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2073 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2074 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2075 		break;
2076 #endif /* INET6 */
2077 	default:
2078 		return (EINVAL);
2079 		/* NOTREACHED */
2080 		break;
2081 	}
2082 	/*
2083 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2084 	 * The TCP checksum must be set to zero.
2085 	 */
2086 	savecsum = th->th_sum;
2087 	th->th_sum = 0;
2088 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2089 	th->th_sum = savecsum;
2090 	/*
2091 	 * Step 3: Update MD5 hash with TCP segment data.
2092 	 *         Use m_apply() to avoid an early m_pullup().
2093 	 */
2094 	if (len > 0)
2095 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2096 	/*
2097 	 * Step 4: Update MD5 hash with shared secret.
2098 	 */
2099 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2100 	MD5Final(buf, &ctx);
2101 	key_sa_recordxfer(sav, m);
2102 	key_freesav(sav);
2103 	return (0);
2104 }
2105 
2106 int
2107 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2108 {
2109 
2110 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2111 	return (0);
2112 }
2113 #endif /* TCP_SIGNATURE */
2114