xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision dce2f7a6000be387deb4457887c74bbf9991eba0)
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. Neither the name of the University nor the names of its contributors
47  *    may be used to endorse or promote products derived from this software
48  *    without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
60  * SUCH DAMAGE.
61  *
62  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
63  * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
64  */
65 
66 #include "opt_compat.h"
67 #include "opt_inet.h"
68 #include "opt_inet6.h"
69 #include "opt_ipsec.h"
70 #include "opt_tcpdebug.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/callout.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/malloc.h>
78 #include <sys/mpipe.h>
79 #include <sys/mbuf.h>
80 #ifdef INET6
81 #include <sys/domain.h>
82 #endif
83 #include <sys/proc.h>
84 #include <sys/priv.h>
85 #include <sys/socket.h>
86 #include <sys/socketops.h>
87 #include <sys/socketvar.h>
88 #include <sys/protosw.h>
89 #include <sys/random.h>
90 #include <sys/in_cksum.h>
91 #include <sys/ktr.h>
92 
93 #include <net/route.h>
94 #include <net/if.h>
95 #include <net/netisr2.h>
96 
97 #define	_IP_VHL
98 #include <netinet/in.h>
99 #include <netinet/in_systm.h>
100 #include <netinet/ip.h>
101 #include <netinet/ip6.h>
102 #include <netinet/in_pcb.h>
103 #include <netinet6/in6_pcb.h>
104 #include <netinet/in_var.h>
105 #include <netinet/ip_var.h>
106 #include <netinet6/ip6_var.h>
107 #include <netinet/ip_icmp.h>
108 #ifdef INET6
109 #include <netinet/icmp6.h>
110 #endif
111 #include <netinet/tcp.h>
112 #include <netinet/tcp_fsm.h>
113 #include <netinet/tcp_seq.h>
114 #include <netinet/tcp_timer.h>
115 #include <netinet/tcp_timer2.h>
116 #include <netinet/tcp_var.h>
117 #include <netinet6/tcp6_var.h>
118 #include <netinet/tcpip.h>
119 #ifdef TCPDEBUG
120 #include <netinet/tcp_debug.h>
121 #endif
122 #include <netinet6/ip6protosw.h>
123 
124 #ifdef IPSEC
125 #include <netinet6/ipsec.h>
126 #include <netproto/key/key.h>
127 #ifdef INET6
128 #include <netinet6/ipsec6.h>
129 #endif
130 #endif
131 
132 #ifdef FAST_IPSEC
133 #include <netproto/ipsec/ipsec.h>
134 #ifdef INET6
135 #include <netproto/ipsec/ipsec6.h>
136 #endif
137 #define	IPSEC
138 #endif
139 
140 #include <sys/md5.h>
141 #include <machine/smp.h>
142 
143 #include <sys/msgport2.h>
144 #include <sys/mplock2.h>
145 #include <net/netmsg2.h>
146 
147 #if !defined(KTR_TCP)
148 #define KTR_TCP		KTR_ALL
149 #endif
150 /*
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 
158 #define TCP_IW_MAXSEGS_DFLT	4
159 #define TCP_IW_CAPSEGS_DFLT	4
160 
161 struct inpcbinfo tcbinfo[MAXCPU];
162 struct tcpcbackqhead tcpcbackq[MAXCPU];
163 
164 int tcp_mssdflt = TCP_MSS;
165 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
166     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
167 
168 #ifdef INET6
169 int tcp_v6mssdflt = TCP6_MSS;
170 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
171     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
172 #endif
173 
174 /*
175  * Minimum MSS we accept and use. This prevents DoS attacks where
176  * we are forced to a ridiculous low MSS like 20 and send hundreds
177  * of packets instead of one. The effect scales with the available
178  * bandwidth and quickly saturates the CPU and network interface
179  * with packet generation and sending. Set to zero to disable MINMSS
180  * checking. This setting prevents us from sending too small packets.
181  */
182 int tcp_minmss = TCP_MINMSS;
183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
184     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
185 
186 #if 0
187 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
188 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
189     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
190 #endif
191 
192 int tcp_do_rfc1323 = 1;
193 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
194     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
195 
196 static int tcp_tcbhashsize = 0;
197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
198      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
199 
200 static int do_tcpdrain = 1;
201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
202      "Enable tcp_drain routine for extra help when low on mbufs");
203 
204 static int icmp_may_rst = 1;
205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
206     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
207 
208 static int tcp_isn_reseed_interval = 0;
209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
210     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
211 
212 /*
213  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
214  * by default, but with generous values which should allow maximal
215  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
216  *
217  * The reason for doing this is that the limiter is the only mechanism we
218  * have which seems to do a really good job preventing receiver RX rings
219  * on network interfaces from getting blown out.  Even though GigE/10GigE
220  * is supposed to flow control it looks like either it doesn't actually
221  * do it or Open Source drivers do not properly enable it.
222  *
223  * People using the limiter to reduce bottlenecks on slower WAN connections
224  * should set the slop to 20 (2 packets).
225  */
226 static int tcp_inflight_enable = 1;
227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
228     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
229 
230 static int tcp_inflight_debug = 0;
231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
232     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
233 
234 /*
235  * NOTE: tcp_inflight_start is essentially the starting receive window
236  *	 for a connection.  If set too low then fetches over tcp
237  *	 connections will take noticably longer to ramp-up over
238  *	 high-latency connections.  6144 is too low for a default,
239  *	 use something more reasonable.
240  */
241 static int tcp_inflight_start = 33792;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_start, CTLFLAG_RW,
243     &tcp_inflight_start, 0, "Start value for TCP inflight window");
244 
245 static int tcp_inflight_min = 6144;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
247     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
248 
249 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
251     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
252 
253 static int tcp_inflight_stab = 50;
254 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
255     &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)");
256 
257 static int tcp_inflight_adjrtt = 2;
258 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW,
259     &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)");
260 
261 static int tcp_do_rfc3390 = 1;
262 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
263     &tcp_do_rfc3390, 0,
264     "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
265 
266 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
267 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW,
268     &tcp_iw_maxsegs, 0, "TCP IW segments max");
269 
270 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
271 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW,
272     &tcp_iw_capsegs, 0, "TCP IW segments");
273 
274 int tcp_low_rtobase = 1;
275 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW,
276     &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)");
277 
278 static int tcp_do_ncr = 1;
279 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW,
280     &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)");
281 
282 int tcp_ncr_rxtthresh_max = 16;
283 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW,
284     &tcp_ncr_rxtthresh_max, 0,
285     "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
286 
287 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
288 static struct malloc_pipe tcptemp_mpipe;
289 
290 static void tcp_willblock(void);
291 static void tcp_notify (struct inpcb *, int);
292 
293 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign;
294 
295 static struct netmsg_base tcp_drain_netmsg[MAXCPU];
296 static void	tcp_drain_dispatch(netmsg_t nmsg);
297 
298 static int
299 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
300 {
301 	int cpu, error = 0;
302 
303 	for (cpu = 0; cpu < ncpus2; ++cpu) {
304 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
305 					sizeof(struct tcp_stats))))
306 			break;
307 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
308 				       sizeof(struct tcp_stats))))
309 			break;
310 	}
311 
312 	return (error);
313 }
314 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
315     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
316 
317 /*
318  * Target size of TCP PCB hash tables. Must be a power of two.
319  *
320  * Note that this can be overridden by the kernel environment
321  * variable net.inet.tcp.tcbhashsize
322  */
323 #ifndef TCBHASHSIZE
324 #define	TCBHASHSIZE	512
325 #endif
326 
327 /*
328  * This is the actual shape of what we allocate using the zone
329  * allocator.  Doing it this way allows us to protect both structures
330  * using the same generation count, and also eliminates the overhead
331  * of allocating tcpcbs separately.  By hiding the structure here,
332  * we avoid changing most of the rest of the code (although it needs
333  * to be changed, eventually, for greater efficiency).
334  */
335 #define	ALIGNMENT	32
336 #define	ALIGNM1		(ALIGNMENT - 1)
337 struct	inp_tp {
338 	union {
339 		struct	inpcb inp;
340 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
341 	} inp_tp_u;
342 	struct	tcpcb tcb;
343 	struct	tcp_callout inp_tp_rexmt;
344 	struct	tcp_callout inp_tp_persist;
345 	struct	tcp_callout inp_tp_keep;
346 	struct	tcp_callout inp_tp_2msl;
347 	struct	tcp_callout inp_tp_delack;
348 	struct	netmsg_tcp_timer inp_tp_timermsg;
349 	struct	netmsg_base inp_tp_sndmore;
350 };
351 #undef ALIGNMENT
352 #undef ALIGNM1
353 
354 /*
355  * Tcp initialization
356  */
357 void
358 tcp_init(void)
359 {
360 	struct inpcbportinfo *portinfo;
361 	struct inpcbinfo *ticb;
362 	int hashsize = TCBHASHSIZE;
363 	int cpu;
364 
365 	/*
366 	 * note: tcptemp is used for keepalives, and it is ok for an
367 	 * allocation to fail so do not specify MPF_INT.
368 	 */
369 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
370 		    25, -1, 0, NULL, NULL, NULL);
371 
372 	tcp_delacktime = TCPTV_DELACK;
373 	tcp_keepinit = TCPTV_KEEP_INIT;
374 	tcp_keepidle = TCPTV_KEEP_IDLE;
375 	tcp_keepintvl = TCPTV_KEEPINTVL;
376 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
377 	tcp_msl = TCPTV_MSL;
378 	tcp_rexmit_min = TCPTV_MIN;
379 	tcp_rexmit_slop = TCPTV_CPU_VAR;
380 
381 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
382 	if (!powerof2(hashsize)) {
383 		kprintf("WARNING: TCB hash size not a power of 2\n");
384 		hashsize = 512; /* safe default */
385 	}
386 	tcp_tcbhashsize = hashsize;
387 
388 	portinfo = kmalloc_cachealign(sizeof(*portinfo) * ncpus2, M_PCB,
389 	    M_WAITOK);
390 
391 	for (cpu = 0; cpu < ncpus2; cpu++) {
392 		ticb = &tcbinfo[cpu];
393 		in_pcbinfo_init(ticb, cpu, FALSE);
394 		ticb->hashbase = hashinit(hashsize, M_PCB,
395 					  &ticb->hashmask);
396 		in_pcbportinfo_init(&portinfo[cpu], hashsize, TRUE, cpu);
397 		ticb->portinfo = portinfo;
398 		ticb->portinfo_mask = ncpus2_mask;
399 		ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
400 						  &ticb->wildcardhashmask);
401 		ticb->localgrphashbase = hashinit(hashsize, M_PCB,
402 						  &ticb->localgrphashmask);
403 		ticb->ipi_size = sizeof(struct inp_tp);
404 		TAILQ_INIT(&tcpcbackq[cpu]);
405 	}
406 
407 	tcp_reass_maxseg = nmbclusters / 16;
408 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
409 
410 #ifdef INET6
411 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
412 #else
413 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
414 #endif
415 	if (max_protohdr < TCP_MINPROTOHDR)
416 		max_protohdr = TCP_MINPROTOHDR;
417 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
418 		panic("tcp_init");
419 #undef TCP_MINPROTOHDR
420 
421 	/*
422 	 * Initialize TCP statistics counters for each CPU.
423 	 */
424 	for (cpu = 0; cpu < ncpus2; ++cpu)
425 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
426 
427 	/*
428 	 * Initialize netmsgs for TCP drain
429 	 */
430 	for (cpu = 0; cpu < ncpus2; ++cpu) {
431 		netmsg_init(&tcp_drain_netmsg[cpu], NULL, &netisr_adone_rport,
432 		    MSGF_PRIORITY, tcp_drain_dispatch);
433 	}
434 
435 	syncache_init();
436 	netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP);
437 }
438 
439 static void
440 tcp_willblock(void)
441 {
442 	struct tcpcb *tp;
443 	int cpu = mycpu->gd_cpuid;
444 
445 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
446 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
447 		tp->t_flags &= ~TF_ONOUTPUTQ;
448 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
449 		tcp_output(tp);
450 	}
451 }
452 
453 /*
454  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
455  * tcp_template used to store this data in mbufs, but we now recopy it out
456  * of the tcpcb each time to conserve mbufs.
457  */
458 void
459 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso)
460 {
461 	struct inpcb *inp = tp->t_inpcb;
462 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
463 
464 #ifdef INET6
465 	if (INP_ISIPV6(inp)) {
466 		struct ip6_hdr *ip6;
467 
468 		ip6 = (struct ip6_hdr *)ip_ptr;
469 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
470 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
471 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
472 			(IPV6_VERSION & IPV6_VERSION_MASK);
473 		ip6->ip6_nxt = IPPROTO_TCP;
474 		ip6->ip6_plen = sizeof(struct tcphdr);
475 		ip6->ip6_src = inp->in6p_laddr;
476 		ip6->ip6_dst = inp->in6p_faddr;
477 		tcp_hdr->th_sum = 0;
478 	} else
479 #endif
480 	{
481 		struct ip *ip = (struct ip *) ip_ptr;
482 		u_int plen;
483 
484 		ip->ip_vhl = IP_VHL_BORING;
485 		ip->ip_tos = 0;
486 		ip->ip_len = 0;
487 		ip->ip_id = 0;
488 		ip->ip_off = 0;
489 		ip->ip_ttl = 0;
490 		ip->ip_sum = 0;
491 		ip->ip_p = IPPROTO_TCP;
492 		ip->ip_src = inp->inp_laddr;
493 		ip->ip_dst = inp->inp_faddr;
494 
495 		if (tso)
496 			plen = htons(IPPROTO_TCP);
497 		else
498 			plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP);
499 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
500 		    ip->ip_dst.s_addr, plen);
501 	}
502 
503 	tcp_hdr->th_sport = inp->inp_lport;
504 	tcp_hdr->th_dport = inp->inp_fport;
505 	tcp_hdr->th_seq = 0;
506 	tcp_hdr->th_ack = 0;
507 	tcp_hdr->th_x2 = 0;
508 	tcp_hdr->th_off = 5;
509 	tcp_hdr->th_flags = 0;
510 	tcp_hdr->th_win = 0;
511 	tcp_hdr->th_urp = 0;
512 }
513 
514 /*
515  * Create template to be used to send tcp packets on a connection.
516  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
517  * use for this function is in keepalives, which use tcp_respond.
518  */
519 struct tcptemp *
520 tcp_maketemplate(struct tcpcb *tp)
521 {
522 	struct tcptemp *tmp;
523 
524 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
525 		return (NULL);
526 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE);
527 	return (tmp);
528 }
529 
530 void
531 tcp_freetemplate(struct tcptemp *tmp)
532 {
533 	mpipe_free(&tcptemp_mpipe, tmp);
534 }
535 
536 /*
537  * Send a single message to the TCP at address specified by
538  * the given TCP/IP header.  If m == NULL, then we make a copy
539  * of the tcpiphdr at ti and send directly to the addressed host.
540  * This is used to force keep alive messages out using the TCP
541  * template for a connection.  If flags are given then we send
542  * a message back to the TCP which originated the * segment ti,
543  * and discard the mbuf containing it and any other attached mbufs.
544  *
545  * In any case the ack and sequence number of the transmitted
546  * segment are as specified by the parameters.
547  *
548  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
549  */
550 void
551 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
552 	    tcp_seq ack, tcp_seq seq, int flags)
553 {
554 	int tlen;
555 	long win = 0;
556 	struct route *ro = NULL;
557 	struct route sro;
558 	struct ip *ip = ipgen;
559 	struct tcphdr *nth;
560 	int ipflags = 0;
561 	struct route_in6 *ro6 = NULL;
562 	struct route_in6 sro6;
563 	struct ip6_hdr *ip6 = ipgen;
564 	boolean_t use_tmpro = TRUE;
565 #ifdef INET6
566 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
567 #else
568 	const boolean_t isipv6 = FALSE;
569 #endif
570 
571 	if (tp != NULL) {
572 		if (!(flags & TH_RST)) {
573 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
574 			if (win < 0)
575 				win = 0;
576 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
577 				win = (long)TCP_MAXWIN << tp->rcv_scale;
578 		}
579 		/*
580 		 * Don't use the route cache of a listen socket,
581 		 * it is not MPSAFE; use temporary route cache.
582 		 */
583 		if (tp->t_state != TCPS_LISTEN) {
584 			if (isipv6)
585 				ro6 = &tp->t_inpcb->in6p_route;
586 			else
587 				ro = &tp->t_inpcb->inp_route;
588 			use_tmpro = FALSE;
589 		}
590 	}
591 	if (use_tmpro) {
592 		if (isipv6) {
593 			ro6 = &sro6;
594 			bzero(ro6, sizeof *ro6);
595 		} else {
596 			ro = &sro;
597 			bzero(ro, sizeof *ro);
598 		}
599 	}
600 	if (m == NULL) {
601 		m = m_gethdr(M_NOWAIT, MT_HEADER);
602 		if (m == NULL)
603 			return;
604 		tlen = 0;
605 		m->m_data += max_linkhdr;
606 		if (isipv6) {
607 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
608 			ip6 = mtod(m, struct ip6_hdr *);
609 			nth = (struct tcphdr *)(ip6 + 1);
610 		} else {
611 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
612 			ip = mtod(m, struct ip *);
613 			nth = (struct tcphdr *)(ip + 1);
614 		}
615 		bcopy(th, nth, sizeof(struct tcphdr));
616 		flags = TH_ACK;
617 	} else {
618 		m_freem(m->m_next);
619 		m->m_next = NULL;
620 		m->m_data = (caddr_t)ipgen;
621 		/* m_len is set later */
622 		tlen = 0;
623 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
624 		if (isipv6) {
625 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
626 			nth = (struct tcphdr *)(ip6 + 1);
627 		} else {
628 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
629 			nth = (struct tcphdr *)(ip + 1);
630 		}
631 		if (th != nth) {
632 			/*
633 			 * this is usually a case when an extension header
634 			 * exists between the IPv6 header and the
635 			 * TCP header.
636 			 */
637 			nth->th_sport = th->th_sport;
638 			nth->th_dport = th->th_dport;
639 		}
640 		xchg(nth->th_dport, nth->th_sport, n_short);
641 #undef xchg
642 	}
643 	if (isipv6) {
644 		ip6->ip6_flow = 0;
645 		ip6->ip6_vfc = IPV6_VERSION;
646 		ip6->ip6_nxt = IPPROTO_TCP;
647 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
648 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
649 	} else {
650 		tlen += sizeof(struct tcpiphdr);
651 		ip->ip_len = tlen;
652 		ip->ip_ttl = ip_defttl;
653 	}
654 	m->m_len = tlen;
655 	m->m_pkthdr.len = tlen;
656 	m->m_pkthdr.rcvif = NULL;
657 	nth->th_seq = htonl(seq);
658 	nth->th_ack = htonl(ack);
659 	nth->th_x2 = 0;
660 	nth->th_off = sizeof(struct tcphdr) >> 2;
661 	nth->th_flags = flags;
662 	if (tp != NULL)
663 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
664 	else
665 		nth->th_win = htons((u_short)win);
666 	nth->th_urp = 0;
667 	if (isipv6) {
668 		nth->th_sum = 0;
669 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
670 					sizeof(struct ip6_hdr),
671 					tlen - sizeof(struct ip6_hdr));
672 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
673 					       (ro6 && ro6->ro_rt) ?
674 						ro6->ro_rt->rt_ifp : NULL);
675 	} else {
676 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
677 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
678 		m->m_pkthdr.csum_flags = CSUM_TCP;
679 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
680 		m->m_pkthdr.csum_thlen = sizeof(struct tcphdr);
681 	}
682 #ifdef TCPDEBUG
683 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
684 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
685 #endif
686 	if (isipv6) {
687 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
688 			   tp ? tp->t_inpcb : NULL);
689 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
690 			RTFREE(ro6->ro_rt);
691 			ro6->ro_rt = NULL;
692 		}
693 	} else {
694 		ipflags |= IP_DEBUGROUTE;
695 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
696 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
697 			RTFREE(ro->ro_rt);
698 			ro->ro_rt = NULL;
699 		}
700 	}
701 }
702 
703 /*
704  * Create a new TCP control block, making an
705  * empty reassembly queue and hooking it to the argument
706  * protocol control block.  The `inp' parameter must have
707  * come from the zone allocator set up in tcp_init().
708  */
709 struct tcpcb *
710 tcp_newtcpcb(struct inpcb *inp)
711 {
712 	struct inp_tp *it;
713 	struct tcpcb *tp;
714 #ifdef INET6
715 	boolean_t isipv6 = INP_ISIPV6(inp);
716 #else
717 	const boolean_t isipv6 = FALSE;
718 #endif
719 
720 	it = (struct inp_tp *)inp;
721 	tp = &it->tcb;
722 	bzero(tp, sizeof(struct tcpcb));
723 	TAILQ_INIT(&tp->t_segq);
724 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
725 	tp->t_rxtthresh = tcprexmtthresh;
726 
727 	/* Set up our timeouts. */
728 	tp->tt_rexmt = &it->inp_tp_rexmt;
729 	tp->tt_persist = &it->inp_tp_persist;
730 	tp->tt_keep = &it->inp_tp_keep;
731 	tp->tt_2msl = &it->inp_tp_2msl;
732 	tp->tt_delack = &it->inp_tp_delack;
733 	tcp_inittimers(tp);
734 
735 	/*
736 	 * Zero out timer message.  We don't create it here,
737 	 * since the current CPU may not be the owner of this
738 	 * inpcb.
739 	 */
740 	tp->tt_msg = &it->inp_tp_timermsg;
741 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
742 
743 	tp->t_keepinit = tcp_keepinit;
744 	tp->t_keepidle = tcp_keepidle;
745 	tp->t_keepintvl = tcp_keepintvl;
746 	tp->t_keepcnt = tcp_keepcnt;
747 	tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
748 
749 	if (tcp_do_ncr)
750 		tp->t_flags |= TF_NCR;
751 	if (tcp_do_rfc1323)
752 		tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP);
753 
754 	tp->t_inpcb = inp;	/* XXX */
755 	tp->t_state = TCPS_CLOSED;
756 	/*
757 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
758 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
759 	 * reasonable initial retransmit time.
760 	 */
761 	tp->t_srtt = TCPTV_SRTTBASE;
762 	tp->t_rttvar =
763 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
764 	tp->t_rttmin = tcp_rexmit_min;
765 	tp->t_rxtcur = TCPTV_RTOBASE;
766 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
767 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
768 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
769 	tp->snd_last = ticks;
770 	tp->t_rcvtime = ticks;
771 	/*
772 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
773 	 * because the socket may be bound to an IPv6 wildcard address,
774 	 * which may match an IPv4-mapped IPv6 address.
775 	 */
776 	inp->inp_ip_ttl = ip_defttl;
777 	inp->inp_ppcb = tp;
778 	tcp_sack_tcpcb_init(tp);
779 
780 	tp->tt_sndmore = &it->inp_tp_sndmore;
781 	tcp_output_init(tp);
782 
783 	return (tp);		/* XXX */
784 }
785 
786 /*
787  * Drop a TCP connection, reporting the specified error.
788  * If connection is synchronized, then send a RST to peer.
789  */
790 struct tcpcb *
791 tcp_drop(struct tcpcb *tp, int error)
792 {
793 	struct socket *so = tp->t_inpcb->inp_socket;
794 
795 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
796 		tp->t_state = TCPS_CLOSED;
797 		tcp_output(tp);
798 		tcpstat.tcps_drops++;
799 	} else
800 		tcpstat.tcps_conndrops++;
801 	if (error == ETIMEDOUT && tp->t_softerror)
802 		error = tp->t_softerror;
803 	so->so_error = error;
804 	return (tcp_close(tp));
805 }
806 
807 struct netmsg_listen_detach {
808 	struct netmsg_base	base;
809 	struct tcpcb		*nm_tp;
810 	struct tcpcb		*nm_tp_inh;
811 };
812 
813 static void
814 tcp_listen_detach_handler(netmsg_t msg)
815 {
816 	struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
817 	struct tcpcb *tp = nmsg->nm_tp;
818 	int cpu = mycpuid, nextcpu;
819 
820 	if (tp->t_flags & TF_LISTEN)
821 		syncache_destroy(tp, nmsg->nm_tp_inh);
822 
823 	in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
824 
825 	nextcpu = cpu + 1;
826 	if (nextcpu < ncpus2)
827 		lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg);
828 	else
829 		lwkt_replymsg(&nmsg->base.lmsg, 0);
830 }
831 
832 /*
833  * Close a TCP control block:
834  *	discard all space held by the tcp
835  *	discard internet protocol block
836  *	wake up any sleepers
837  */
838 struct tcpcb *
839 tcp_close(struct tcpcb *tp)
840 {
841 	struct tseg_qent *q;
842 	struct inpcb *inp = tp->t_inpcb;
843 	struct inpcb *inp_inh = NULL;
844 	struct tcpcb *tp_inh = NULL;
845 	struct socket *so = inp->inp_socket;
846 	struct rtentry *rt;
847 	boolean_t dosavessthresh;
848 #ifdef INET6
849 	boolean_t isipv6 = INP_ISIPV6(inp);
850 #else
851 	const boolean_t isipv6 = FALSE;
852 #endif
853 
854 	if (tp->t_flags & TF_LISTEN) {
855 		/*
856 		 * Pending socket/syncache inheritance
857 		 *
858 		 * If this is a listen(2) socket, find another listen(2)
859 		 * socket in the same local group, which could inherit
860 		 * the syncache and sockets pending on the completion
861 		 * and incompletion queues.
862 		 *
863 		 * NOTE:
864 		 * Currently the inheritance could only happen on the
865 		 * listen(2) sockets w/ SO_REUSEPORT set.
866 		 */
867 		ASSERT_IN_NETISR(0);
868 		inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp);
869 		if (inp_inh != NULL)
870 			tp_inh = intotcpcb(inp_inh);
871 	}
872 
873 	/*
874 	 * INP_WILDCARD indicates that listen(2) has been called on
875 	 * this socket.  This implies:
876 	 * - A wildcard inp's hash is replicated for each protocol thread.
877 	 * - Syncache for this inp grows independently in each protocol
878 	 *   thread.
879 	 * - There is more than one cpu
880 	 *
881 	 * We have to chain a message to the rest of the protocol threads
882 	 * to cleanup the wildcard hash and the syncache.  The cleanup
883 	 * in the current protocol thread is defered till the end of this
884 	 * function (syncache_destroy and in_pcbdetach).
885 	 *
886 	 * NOTE:
887 	 * After cleanup the inp's hash and syncache entries, this inp will
888 	 * no longer be available to the rest of the protocol threads, so we
889 	 * are safe to whack the inp in the following code.
890 	 */
891 	if ((inp->inp_flags & INP_WILDCARD) && ncpus2 > 1) {
892 		struct netmsg_listen_detach nmsg;
893 
894 		KKASSERT(so->so_port == netisr_cpuport(0));
895 		ASSERT_IN_NETISR(0);
896 		KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
897 
898 		netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
899 			    MSGF_PRIORITY, tcp_listen_detach_handler);
900 		nmsg.nm_tp = tp;
901 		nmsg.nm_tp_inh = tp_inh;
902 		lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0);
903 	}
904 
905 	KKASSERT(tp->t_state != TCPS_TERMINATING);
906 	tp->t_state = TCPS_TERMINATING;
907 
908 	/*
909 	 * Make sure that all of our timers are stopped before we
910 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
911 	 * timers are never used.  If timer message is never created
912 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
913 	 */
914 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
915 		tcp_callout_stop(tp, tp->tt_rexmt);
916 		tcp_callout_stop(tp, tp->tt_persist);
917 		tcp_callout_stop(tp, tp->tt_keep);
918 		tcp_callout_stop(tp, tp->tt_2msl);
919 		tcp_callout_stop(tp, tp->tt_delack);
920 	}
921 
922 	if (tp->t_flags & TF_ONOUTPUTQ) {
923 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
924 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
925 		tp->t_flags &= ~TF_ONOUTPUTQ;
926 	}
927 
928 	/*
929 	 * If we got enough samples through the srtt filter,
930 	 * save the rtt and rttvar in the routing entry.
931 	 * 'Enough' is arbitrarily defined as the 16 samples.
932 	 * 16 samples is enough for the srtt filter to converge
933 	 * to within 5% of the correct value; fewer samples and
934 	 * we could save a very bogus rtt.
935 	 *
936 	 * Don't update the default route's characteristics and don't
937 	 * update anything that the user "locked".
938 	 */
939 	if (tp->t_rttupdated >= 16) {
940 		u_long i = 0;
941 
942 		if (isipv6) {
943 			struct sockaddr_in6 *sin6;
944 
945 			if ((rt = inp->in6p_route.ro_rt) == NULL)
946 				goto no_valid_rt;
947 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
948 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
949 				goto no_valid_rt;
950 		} else
951 			if ((rt = inp->inp_route.ro_rt) == NULL ||
952 			    ((struct sockaddr_in *)rt_key(rt))->
953 			     sin_addr.s_addr == INADDR_ANY)
954 				goto no_valid_rt;
955 
956 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
957 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
958 			if (rt->rt_rmx.rmx_rtt && i)
959 				/*
960 				 * filter this update to half the old & half
961 				 * the new values, converting scale.
962 				 * See route.h and tcp_var.h for a
963 				 * description of the scaling constants.
964 				 */
965 				rt->rt_rmx.rmx_rtt =
966 				    (rt->rt_rmx.rmx_rtt + i) / 2;
967 			else
968 				rt->rt_rmx.rmx_rtt = i;
969 			tcpstat.tcps_cachedrtt++;
970 		}
971 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
972 			i = tp->t_rttvar *
973 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
974 			if (rt->rt_rmx.rmx_rttvar && i)
975 				rt->rt_rmx.rmx_rttvar =
976 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
977 			else
978 				rt->rt_rmx.rmx_rttvar = i;
979 			tcpstat.tcps_cachedrttvar++;
980 		}
981 		/*
982 		 * The old comment here said:
983 		 * update the pipelimit (ssthresh) if it has been updated
984 		 * already or if a pipesize was specified & the threshhold
985 		 * got below half the pipesize.  I.e., wait for bad news
986 		 * before we start updating, then update on both good
987 		 * and bad news.
988 		 *
989 		 * But we want to save the ssthresh even if no pipesize is
990 		 * specified explicitly in the route, because such
991 		 * connections still have an implicit pipesize specified
992 		 * by the global tcp_sendspace.  In the absence of a reliable
993 		 * way to calculate the pipesize, it will have to do.
994 		 */
995 		i = tp->snd_ssthresh;
996 		if (rt->rt_rmx.rmx_sendpipe != 0)
997 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
998 		else
999 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
1000 		if (dosavessthresh ||
1001 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
1002 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
1003 			/*
1004 			 * convert the limit from user data bytes to
1005 			 * packets then to packet data bytes.
1006 			 */
1007 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
1008 			if (i < 2)
1009 				i = 2;
1010 			i *= tp->t_maxseg +
1011 			     (isipv6 ?
1012 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1013 			      sizeof(struct tcpiphdr));
1014 			if (rt->rt_rmx.rmx_ssthresh)
1015 				rt->rt_rmx.rmx_ssthresh =
1016 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
1017 			else
1018 				rt->rt_rmx.rmx_ssthresh = i;
1019 			tcpstat.tcps_cachedssthresh++;
1020 		}
1021 	}
1022 
1023 no_valid_rt:
1024 	/* free the reassembly queue, if any */
1025 	while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) {
1026 		TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1027 		m_freem(q->tqe_m);
1028 		kfree(q, M_TSEGQ);
1029 		atomic_add_int(&tcp_reass_qsize, -1);
1030 	}
1031 	/* throw away SACK blocks in scoreboard*/
1032 	if (TCP_DO_SACK(tp))
1033 		tcp_sack_destroy(&tp->scb);
1034 
1035 	inp->inp_ppcb = NULL;
1036 	soisdisconnected(so);
1037 	/* note: pcb detached later on */
1038 
1039 	tcp_destroy_timermsg(tp);
1040 	tcp_output_cancel(tp);
1041 
1042 	if (tp->t_flags & TF_LISTEN) {
1043 		syncache_destroy(tp, tp_inh);
1044 		if (inp_inh != NULL && inp_inh->inp_socket != NULL) {
1045 			/*
1046 			 * Pending sockets inheritance only needs
1047 			 * to be done once in the current thread,
1048 			 * i.e. netisr0.
1049 			 */
1050 			soinherit(so, inp_inh->inp_socket);
1051 		}
1052 	}
1053 
1054 	so_async_rcvd_drop(so);
1055 	/* Drop the reference for the asynchronized pru_rcvd */
1056 	sofree(so);
1057 
1058 	/*
1059 	 * NOTE:
1060 	 * pcbdetach removes any wildcard hash entry on the current CPU.
1061 	 */
1062 #ifdef INET6
1063 	if (isipv6)
1064 		in6_pcbdetach(inp);
1065 	else
1066 #endif
1067 		in_pcbdetach(inp);
1068 
1069 	tcpstat.tcps_closed++;
1070 	return (NULL);
1071 }
1072 
1073 static __inline void
1074 tcp_drain_oncpu(struct inpcbinfo *pcbinfo)
1075 {
1076 	struct inpcbhead *head = &pcbinfo->pcblisthead;
1077 	struct inpcb *inpb;
1078 
1079 	/*
1080 	 * Since we run in netisr, it is MP safe, even if
1081 	 * we block during the inpcb list iteration, i.e.
1082 	 * we don't need to use inpcb marker here.
1083 	 */
1084 	ASSERT_IN_NETISR(pcbinfo->cpu);
1085 
1086 	LIST_FOREACH(inpb, head, inp_list) {
1087 		struct tcpcb *tcpb;
1088 		struct tseg_qent *te;
1089 
1090 		if (inpb->inp_flags & INP_PLACEMARKER)
1091 			continue;
1092 
1093 		tcpb = intotcpcb(inpb);
1094 		KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL"));
1095 
1096 		if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) {
1097 			TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q);
1098 			if (te->tqe_th->th_flags & TH_FIN)
1099 				tcpb->t_flags &= ~TF_QUEDFIN;
1100 			m_freem(te->tqe_m);
1101 			kfree(te, M_TSEGQ);
1102 			atomic_add_int(&tcp_reass_qsize, -1);
1103 			/* retry */
1104 		}
1105 	}
1106 }
1107 
1108 static void
1109 tcp_drain_dispatch(netmsg_t nmsg)
1110 {
1111 	crit_enter();
1112 	lwkt_replymsg(&nmsg->lmsg, 0);  /* reply ASAP */
1113 	crit_exit();
1114 
1115 	tcp_drain_oncpu(&tcbinfo[mycpuid]);
1116 }
1117 
1118 static void
1119 tcp_drain_ipi(void *arg __unused)
1120 {
1121 	int cpu = mycpuid;
1122 	struct lwkt_msg *msg = &tcp_drain_netmsg[cpu].lmsg;
1123 
1124 	crit_enter();
1125 	if (msg->ms_flags & MSGF_DONE)
1126 		lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg);
1127 	crit_exit();
1128 }
1129 
1130 void
1131 tcp_drain(void)
1132 {
1133 	cpumask_t mask;
1134 
1135 	if (!do_tcpdrain)
1136 		return;
1137 
1138 	/*
1139 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1140 	 * if there is one...
1141 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1142 	 *	reassembly queue should be flushed, but in a situation
1143 	 *	where we're really low on mbufs, this is potentially
1144 	 *	useful.
1145 	 * YYY: We may consider run tcp_drain_oncpu directly here,
1146 	 *      however, that will require M_WAITOK memory allocation
1147 	 *      for the inpcb marker.
1148 	 */
1149 	CPUMASK_ASSBMASK(mask, ncpus2);
1150 	CPUMASK_ANDMASK(mask, smp_active_mask);
1151 	if (CPUMASK_TESTNZERO(mask))
1152 		lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL);
1153 }
1154 
1155 /*
1156  * Notify a tcp user of an asynchronous error;
1157  * store error as soft error, but wake up user
1158  * (for now, won't do anything until can select for soft error).
1159  *
1160  * Do not wake up user since there currently is no mechanism for
1161  * reporting soft errors (yet - a kqueue filter may be added).
1162  */
1163 static void
1164 tcp_notify(struct inpcb *inp, int error)
1165 {
1166 	struct tcpcb *tp = intotcpcb(inp);
1167 
1168 	/*
1169 	 * Ignore some errors if we are hooked up.
1170 	 * If connection hasn't completed, has retransmitted several times,
1171 	 * and receives a second error, give up now.  This is better
1172 	 * than waiting a long time to establish a connection that
1173 	 * can never complete.
1174 	 */
1175 	if (tp->t_state == TCPS_ESTABLISHED &&
1176 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1177 	      error == EHOSTDOWN)) {
1178 		return;
1179 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1180 	    tp->t_softerror)
1181 		tcp_drop(tp, error);
1182 	else
1183 		tp->t_softerror = error;
1184 #if 0
1185 	wakeup(&so->so_timeo);
1186 	sorwakeup(so);
1187 	sowwakeup(so);
1188 #endif
1189 }
1190 
1191 static int
1192 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1193 {
1194 	int error, i, n;
1195 	struct inpcb *marker;
1196 	struct inpcb *inp;
1197 	int origcpu, ccpu;
1198 
1199 	error = 0;
1200 	n = 0;
1201 
1202 	/*
1203 	 * The process of preparing the TCB list is too time-consuming and
1204 	 * resource-intensive to repeat twice on every request.
1205 	 */
1206 	if (req->oldptr == NULL) {
1207 		for (ccpu = 0; ccpu < ncpus2; ++ccpu)
1208 			n += tcbinfo[ccpu].ipi_count;
1209 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1210 		return (0);
1211 	}
1212 
1213 	if (req->newptr != NULL)
1214 		return (EPERM);
1215 
1216 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1217 	marker->inp_flags |= INP_PLACEMARKER;
1218 
1219 	/*
1220 	 * OK, now we're committed to doing something.  Run the inpcb list
1221 	 * for each cpu in the system and construct the output.  Use a
1222 	 * list placemarker to deal with list changes occuring during
1223 	 * copyout blockages (but otherwise depend on being on the correct
1224 	 * cpu to avoid races).
1225 	 */
1226 	origcpu = mycpu->gd_cpuid;
1227 	for (ccpu = 0; ccpu < ncpus2 && error == 0; ++ccpu) {
1228 		caddr_t inp_ppcb;
1229 		struct xtcpcb xt;
1230 
1231 		lwkt_migratecpu(ccpu);
1232 
1233 		n = tcbinfo[ccpu].ipi_count;
1234 
1235 		LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list);
1236 		i = 0;
1237 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1238 			/*
1239 			 * process a snapshot of pcbs, ignoring placemarkers
1240 			 * and using our own to allow SYSCTL_OUT to block.
1241 			 */
1242 			LIST_REMOVE(marker, inp_list);
1243 			LIST_INSERT_AFTER(inp, marker, inp_list);
1244 
1245 			if (inp->inp_flags & INP_PLACEMARKER)
1246 				continue;
1247 			if (prison_xinpcb(req->td, inp))
1248 				continue;
1249 
1250 			xt.xt_len = sizeof xt;
1251 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1252 			inp_ppcb = inp->inp_ppcb;
1253 			if (inp_ppcb != NULL)
1254 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1255 			else
1256 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1257 			if (inp->inp_socket)
1258 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1259 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1260 				break;
1261 			++i;
1262 		}
1263 		LIST_REMOVE(marker, inp_list);
1264 		if (error == 0 && i < n) {
1265 			bzero(&xt, sizeof xt);
1266 			xt.xt_len = sizeof xt;
1267 			while (i < n) {
1268 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1269 				if (error)
1270 					break;
1271 				++i;
1272 			}
1273 		}
1274 	}
1275 
1276 	/*
1277 	 * Make sure we are on the same cpu we were on originally, since
1278 	 * higher level callers expect this.  Also don't pollute caches with
1279 	 * migrated userland data by (eventually) returning to userland
1280 	 * on a different cpu.
1281 	 */
1282 	lwkt_migratecpu(origcpu);
1283 	kfree(marker, M_TEMP);
1284 	return (error);
1285 }
1286 
1287 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1288 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1289 
1290 static int
1291 tcp_getcred(SYSCTL_HANDLER_ARGS)
1292 {
1293 	struct sockaddr_in addrs[2];
1294 	struct ucred cred0, *cred = NULL;
1295 	struct inpcb *inp;
1296 	int cpu, origcpu, error;
1297 
1298 	error = priv_check(req->td, PRIV_ROOT);
1299 	if (error != 0)
1300 		return (error);
1301 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1302 	if (error != 0)
1303 		return (error);
1304 
1305 	origcpu = mycpuid;
1306 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1307 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1308 
1309 	lwkt_migratecpu(cpu);
1310 
1311 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1312 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1313 	if (inp == NULL || inp->inp_socket == NULL) {
1314 		error = ENOENT;
1315 	} else if (inp->inp_socket->so_cred != NULL) {
1316 		cred0 = *(inp->inp_socket->so_cred);
1317 		cred = &cred0;
1318 	}
1319 
1320 	lwkt_migratecpu(origcpu);
1321 
1322 	if (error)
1323 		return (error);
1324 
1325 	return SYSCTL_OUT(req, cred, sizeof(struct ucred));
1326 }
1327 
1328 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1329     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1330 
1331 #ifdef INET6
1332 static int
1333 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1334 {
1335 	struct sockaddr_in6 addrs[2];
1336 	struct inpcb *inp;
1337 	int error;
1338 
1339 	error = priv_check(req->td, PRIV_ROOT);
1340 	if (error != 0)
1341 		return (error);
1342 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1343 	if (error != 0)
1344 		return (error);
1345 	crit_enter();
1346 	inp = in6_pcblookup_hash(&tcbinfo[0],
1347 	    &addrs[1].sin6_addr, addrs[1].sin6_port,
1348 	    &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1349 	if (inp == NULL || inp->inp_socket == NULL) {
1350 		error = ENOENT;
1351 		goto out;
1352 	}
1353 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1354 out:
1355 	crit_exit();
1356 	return (error);
1357 }
1358 
1359 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1360 	    0, 0,
1361 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1362 #endif
1363 
1364 struct netmsg_tcp_notify {
1365 	struct netmsg_base base;
1366 	inp_notify_t	nm_notify;
1367 	struct in_addr	nm_faddr;
1368 	int		nm_arg;
1369 };
1370 
1371 static void
1372 tcp_notifyall_oncpu(netmsg_t msg)
1373 {
1374 	struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1375 	int nextcpu;
1376 
1377 	in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr,
1378 			nm->nm_arg, nm->nm_notify);
1379 
1380 	nextcpu = mycpuid + 1;
1381 	if (nextcpu < ncpus2)
1382 		lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg);
1383 	else
1384 		lwkt_replymsg(&nm->base.lmsg, 0);
1385 }
1386 
1387 inp_notify_t
1388 tcp_get_inpnotify(int cmd, const struct sockaddr *sa,
1389     int *arg, struct ip **ip0, int *cpuid)
1390 {
1391 	struct ip *ip = *ip0;
1392 	struct in_addr faddr;
1393 	inp_notify_t notify = tcp_notify;
1394 
1395 	faddr = ((const struct sockaddr_in *)sa)->sin_addr;
1396 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1397 		return NULL;
1398 
1399 	*arg = inetctlerrmap[cmd];
1400 	if (cmd == PRC_QUENCH) {
1401 		notify = tcp_quench;
1402 	} else if (icmp_may_rst &&
1403 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1404 		    cmd == PRC_UNREACH_PORT ||
1405 		    cmd == PRC_TIMXCEED_INTRANS) &&
1406 		   ip != NULL) {
1407 		notify = tcp_drop_syn_sent;
1408 	} else if (cmd == PRC_MSGSIZE) {
1409 		const struct icmp *icmp = (const struct icmp *)
1410 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1411 
1412 		*arg = ntohs(icmp->icmp_nextmtu);
1413 		notify = tcp_mtudisc;
1414 	} else if (PRC_IS_REDIRECT(cmd)) {
1415 		ip = NULL;
1416 		notify = in_rtchange;
1417 	} else if (cmd == PRC_HOSTDEAD) {
1418 		ip = NULL;
1419 	} else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1420 		return NULL;
1421 	}
1422 
1423 	if (cpuid != NULL) {
1424 		if (ip == NULL) {
1425 			/* Go through all CPUs */
1426 			*cpuid = ncpus;
1427 		} else {
1428 			const struct tcphdr *th;
1429 
1430 			th = (const struct tcphdr *)
1431 			    ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2));
1432 			*cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport,
1433 			    ip->ip_src.s_addr, th->th_sport);
1434 		}
1435 	}
1436 
1437 	*ip0 = ip;
1438 	return notify;
1439 }
1440 
1441 void
1442 tcp_ctlinput(netmsg_t msg)
1443 {
1444 	int cmd = msg->ctlinput.nm_cmd;
1445 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1446 	struct ip *ip = msg->ctlinput.nm_extra;
1447 	struct in_addr faddr;
1448 	inp_notify_t notify;
1449 	int arg, cpuid;
1450 
1451 	notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid);
1452 	if (notify == NULL)
1453 		goto done;
1454 
1455 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1456 	if (ip != NULL) {
1457 		const struct tcphdr *th;
1458 		struct inpcb *inp;
1459 
1460 		if (cpuid != mycpuid)
1461 			goto done;
1462 
1463 		th = (const struct tcphdr *)
1464 		    ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2));
1465 		inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport,
1466 					ip->ip_src, th->th_sport, 0, NULL);
1467 		if (inp != NULL && inp->inp_socket != NULL) {
1468 			tcp_seq icmpseq = htonl(th->th_seq);
1469 			struct tcpcb *tp = intotcpcb(inp);
1470 
1471 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1472 			    SEQ_LT(icmpseq, tp->snd_max))
1473 				notify(inp, arg);
1474 		} else {
1475 			struct in_conninfo inc;
1476 
1477 			inc.inc_fport = th->th_dport;
1478 			inc.inc_lport = th->th_sport;
1479 			inc.inc_faddr = faddr;
1480 			inc.inc_laddr = ip->ip_src;
1481 #ifdef INET6
1482 			inc.inc_isipv6 = 0;
1483 #endif
1484 			syncache_unreach(&inc, th);
1485 		}
1486 	} else if (msg->ctlinput.nm_direct) {
1487 		if (cpuid != ncpus && cpuid != mycpuid)
1488 			goto done;
1489 		if (mycpuid >= ncpus2)
1490 			goto done;
1491 
1492 		in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify);
1493 	} else {
1494 		struct netmsg_tcp_notify *nm;
1495 
1496 		ASSERT_IN_NETISR(0);
1497 		nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1498 		netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1499 			    0, tcp_notifyall_oncpu);
1500 		nm->nm_faddr = faddr;
1501 		nm->nm_arg = arg;
1502 		nm->nm_notify = notify;
1503 
1504 		lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg);
1505 	}
1506 done:
1507 	lwkt_replymsg(&msg->lmsg, 0);
1508 }
1509 
1510 #ifdef INET6
1511 
1512 void
1513 tcp6_ctlinput(netmsg_t msg)
1514 {
1515 	int cmd = msg->ctlinput.nm_cmd;
1516 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1517 	void *d = msg->ctlinput.nm_extra;
1518 	struct tcphdr th;
1519 	inp_notify_t notify = tcp_notify;
1520 	struct ip6_hdr *ip6;
1521 	struct mbuf *m;
1522 	struct ip6ctlparam *ip6cp = NULL;
1523 	const struct sockaddr_in6 *sa6_src = NULL;
1524 	int off;
1525 	struct tcp_portonly {
1526 		u_int16_t th_sport;
1527 		u_int16_t th_dport;
1528 	} *thp;
1529 	int arg;
1530 
1531 	if (sa->sa_family != AF_INET6 ||
1532 	    sa->sa_len != sizeof(struct sockaddr_in6)) {
1533 		goto out;
1534 	}
1535 
1536 	arg = 0;
1537 	if (cmd == PRC_QUENCH)
1538 		notify = tcp_quench;
1539 	else if (cmd == PRC_MSGSIZE) {
1540 		struct ip6ctlparam *ip6cp = d;
1541 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1542 
1543 		arg = ntohl(icmp6->icmp6_mtu);
1544 		notify = tcp_mtudisc;
1545 	} else if (!PRC_IS_REDIRECT(cmd) &&
1546 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1547 		goto out;
1548 	}
1549 
1550 	/* if the parameter is from icmp6, decode it. */
1551 	if (d != NULL) {
1552 		ip6cp = (struct ip6ctlparam *)d;
1553 		m = ip6cp->ip6c_m;
1554 		ip6 = ip6cp->ip6c_ip6;
1555 		off = ip6cp->ip6c_off;
1556 		sa6_src = ip6cp->ip6c_src;
1557 	} else {
1558 		m = NULL;
1559 		ip6 = NULL;
1560 		off = 0;	/* fool gcc */
1561 		sa6_src = &sa6_any;
1562 	}
1563 
1564 	if (ip6 != NULL) {
1565 		struct in_conninfo inc;
1566 		/*
1567 		 * XXX: We assume that when IPV6 is non NULL,
1568 		 * M and OFF are valid.
1569 		 */
1570 
1571 		/* check if we can safely examine src and dst ports */
1572 		if (m->m_pkthdr.len < off + sizeof *thp)
1573 			goto out;
1574 
1575 		bzero(&th, sizeof th);
1576 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1577 
1578 		in6_pcbnotify(&tcbinfo[0], sa, th.th_dport,
1579 		    (struct sockaddr *)ip6cp->ip6c_src,
1580 		    th.th_sport, cmd, arg, notify);
1581 
1582 		inc.inc_fport = th.th_dport;
1583 		inc.inc_lport = th.th_sport;
1584 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1585 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1586 		inc.inc_isipv6 = 1;
1587 		syncache_unreach(&inc, &th);
1588 	} else {
1589 		in6_pcbnotify(&tcbinfo[0], sa, 0,
1590 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1591 	}
1592 out:
1593 	lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1594 }
1595 
1596 #endif
1597 
1598 /*
1599  * Following is where TCP initial sequence number generation occurs.
1600  *
1601  * There are two places where we must use initial sequence numbers:
1602  * 1.  In SYN-ACK packets.
1603  * 2.  In SYN packets.
1604  *
1605  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1606  * tcp_syncache.c for details.
1607  *
1608  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1609  * depends on this property.  In addition, these ISNs should be
1610  * unguessable so as to prevent connection hijacking.  To satisfy
1611  * the requirements of this situation, the algorithm outlined in
1612  * RFC 1948 is used to generate sequence numbers.
1613  *
1614  * Implementation details:
1615  *
1616  * Time is based off the system timer, and is corrected so that it
1617  * increases by one megabyte per second.  This allows for proper
1618  * recycling on high speed LANs while still leaving over an hour
1619  * before rollover.
1620  *
1621  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1622  * between seeding of isn_secret.  This is normally set to zero,
1623  * as reseeding should not be necessary.
1624  *
1625  */
1626 
1627 #define	ISN_BYTES_PER_SECOND 1048576
1628 
1629 u_char isn_secret[32];
1630 int isn_last_reseed;
1631 MD5_CTX isn_ctx;
1632 
1633 tcp_seq
1634 tcp_new_isn(struct tcpcb *tp)
1635 {
1636 	u_int32_t md5_buffer[4];
1637 	tcp_seq new_isn;
1638 
1639 	/* Seed if this is the first use, reseed if requested. */
1640 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1641 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1642 		< (u_int)ticks))) {
1643 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1644 		isn_last_reseed = ticks;
1645 	}
1646 
1647 	/* Compute the md5 hash and return the ISN. */
1648 	MD5Init(&isn_ctx);
1649 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1650 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1651 #ifdef INET6
1652 	if (INP_ISIPV6(tp->t_inpcb)) {
1653 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1654 			  sizeof(struct in6_addr));
1655 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1656 			  sizeof(struct in6_addr));
1657 	} else
1658 #endif
1659 	{
1660 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1661 			  sizeof(struct in_addr));
1662 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1663 			  sizeof(struct in_addr));
1664 	}
1665 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1666 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1667 	new_isn = (tcp_seq) md5_buffer[0];
1668 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1669 	return (new_isn);
1670 }
1671 
1672 /*
1673  * When a source quench is received, close congestion window
1674  * to one segment.  We will gradually open it again as we proceed.
1675  */
1676 void
1677 tcp_quench(struct inpcb *inp, int error)
1678 {
1679 	struct tcpcb *tp = intotcpcb(inp);
1680 
1681 	KASSERT(tp != NULL, ("tcp_quench: tp is NULL"));
1682 	tp->snd_cwnd = tp->t_maxseg;
1683 	tp->snd_wacked = 0;
1684 }
1685 
1686 /*
1687  * When a specific ICMP unreachable message is received and the
1688  * connection state is SYN-SENT, drop the connection.  This behavior
1689  * is controlled by the icmp_may_rst sysctl.
1690  */
1691 void
1692 tcp_drop_syn_sent(struct inpcb *inp, int error)
1693 {
1694 	struct tcpcb *tp = intotcpcb(inp);
1695 
1696 	KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL"));
1697 	if (tp->t_state == TCPS_SYN_SENT)
1698 		tcp_drop(tp, error);
1699 }
1700 
1701 /*
1702  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1703  * based on the new value in the route.  Also nudge TCP to send something,
1704  * since we know the packet we just sent was dropped.
1705  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1706  */
1707 void
1708 tcp_mtudisc(struct inpcb *inp, int mtu)
1709 {
1710 	struct tcpcb *tp = intotcpcb(inp);
1711 	struct rtentry *rt;
1712 	struct socket *so = inp->inp_socket;
1713 	int maxopd, mss;
1714 #ifdef INET6
1715 	boolean_t isipv6 = INP_ISIPV6(inp);
1716 #else
1717 	const boolean_t isipv6 = FALSE;
1718 #endif
1719 
1720 	KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL"));
1721 
1722 	/*
1723 	 * If no MTU is provided in the ICMP message, use the
1724 	 * next lower likely value, as specified in RFC 1191.
1725 	 */
1726 	if (mtu == 0) {
1727 		int oldmtu;
1728 
1729 		oldmtu = tp->t_maxopd +
1730 		    (isipv6 ?
1731 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1732 		     sizeof(struct tcpiphdr));
1733 		mtu = ip_next_mtu(oldmtu, 0);
1734 	}
1735 
1736 	if (isipv6)
1737 		rt = tcp_rtlookup6(&inp->inp_inc);
1738 	else
1739 		rt = tcp_rtlookup(&inp->inp_inc);
1740 	if (rt != NULL) {
1741 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1742 			mtu = rt->rt_rmx.rmx_mtu;
1743 
1744 		maxopd = mtu -
1745 		    (isipv6 ?
1746 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1747 		     sizeof(struct tcpiphdr));
1748 
1749 		/*
1750 		 * XXX - The following conditional probably violates the TCP
1751 		 * spec.  The problem is that, since we don't know the
1752 		 * other end's MSS, we are supposed to use a conservative
1753 		 * default.  But, if we do that, then MTU discovery will
1754 		 * never actually take place, because the conservative
1755 		 * default is much less than the MTUs typically seen
1756 		 * on the Internet today.  For the moment, we'll sweep
1757 		 * this under the carpet.
1758 		 *
1759 		 * The conservative default might not actually be a problem
1760 		 * if the only case this occurs is when sending an initial
1761 		 * SYN with options and data to a host we've never talked
1762 		 * to before.  Then, they will reply with an MSS value which
1763 		 * will get recorded and the new parameters should get
1764 		 * recomputed.  For Further Study.
1765 		 */
1766 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1767 			maxopd = rt->rt_rmx.rmx_mssopt;
1768 	} else
1769 		maxopd = mtu -
1770 		    (isipv6 ?
1771 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1772 		     sizeof(struct tcpiphdr));
1773 
1774 	if (tp->t_maxopd <= maxopd)
1775 		return;
1776 	tp->t_maxopd = maxopd;
1777 
1778 	mss = maxopd;
1779 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1780 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1781 		mss -= TCPOLEN_TSTAMP_APPA;
1782 
1783 	/* round down to multiple of MCLBYTES */
1784 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1785 	if (mss > MCLBYTES)
1786 		mss &= ~(MCLBYTES - 1);
1787 #else
1788 	if (mss > MCLBYTES)
1789 		mss = (mss / MCLBYTES) * MCLBYTES;
1790 #endif
1791 
1792 	if (so->so_snd.ssb_hiwat < mss)
1793 		mss = so->so_snd.ssb_hiwat;
1794 
1795 	tp->t_maxseg = mss;
1796 	tp->t_rtttime = 0;
1797 	tp->snd_nxt = tp->snd_una;
1798 	tcp_output(tp);
1799 	tcpstat.tcps_mturesent++;
1800 }
1801 
1802 /*
1803  * Look-up the routing entry to the peer of this inpcb.  If no route
1804  * is found and it cannot be allocated the return NULL.  This routine
1805  * is called by TCP routines that access the rmx structure and by tcp_mss
1806  * to get the interface MTU.
1807  */
1808 struct rtentry *
1809 tcp_rtlookup(struct in_conninfo *inc)
1810 {
1811 	struct route *ro = &inc->inc_route;
1812 
1813 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1814 		/* No route yet, so try to acquire one */
1815 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1816 			/*
1817 			 * unused portions of the structure MUST be zero'd
1818 			 * out because rtalloc() treats it as opaque data
1819 			 */
1820 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1821 			ro->ro_dst.sa_family = AF_INET;
1822 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1823 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1824 			    inc->inc_faddr;
1825 			rtalloc(ro);
1826 		}
1827 	}
1828 	return (ro->ro_rt);
1829 }
1830 
1831 #ifdef INET6
1832 struct rtentry *
1833 tcp_rtlookup6(struct in_conninfo *inc)
1834 {
1835 	struct route_in6 *ro6 = &inc->inc6_route;
1836 
1837 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1838 		/* No route yet, so try to acquire one */
1839 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1840 			/*
1841 			 * unused portions of the structure MUST be zero'd
1842 			 * out because rtalloc() treats it as opaque data
1843 			 */
1844 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1845 			ro6->ro_dst.sin6_family = AF_INET6;
1846 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1847 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1848 			rtalloc((struct route *)ro6);
1849 		}
1850 	}
1851 	return (ro6->ro_rt);
1852 }
1853 #endif
1854 
1855 #ifdef IPSEC
1856 /* compute ESP/AH header size for TCP, including outer IP header. */
1857 size_t
1858 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1859 {
1860 	struct inpcb *inp;
1861 	struct mbuf *m;
1862 	size_t hdrsiz;
1863 	struct ip *ip;
1864 	struct tcphdr *th;
1865 
1866 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1867 		return (0);
1868 	MGETHDR(m, M_NOWAIT, MT_DATA);
1869 	if (!m)
1870 		return (0);
1871 
1872 #ifdef INET6
1873 	if (INP_ISIPV6(inp)) {
1874 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1875 
1876 		th = (struct tcphdr *)(ip6 + 1);
1877 		m->m_pkthdr.len = m->m_len =
1878 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1879 		tcp_fillheaders(tp, ip6, th, FALSE);
1880 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1881 	} else
1882 #endif
1883 	{
1884 		ip = mtod(m, struct ip *);
1885 		th = (struct tcphdr *)(ip + 1);
1886 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1887 		tcp_fillheaders(tp, ip, th, FALSE);
1888 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1889 	}
1890 
1891 	m_free(m);
1892 	return (hdrsiz);
1893 }
1894 #endif
1895 
1896 /*
1897  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1898  *
1899  * This code attempts to calculate the bandwidth-delay product as a
1900  * means of determining the optimal window size to maximize bandwidth,
1901  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1902  * routers.  This code also does a fairly good job keeping RTTs in check
1903  * across slow links like modems.  We implement an algorithm which is very
1904  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1905  * transmitter side of a TCP connection and so only effects the transmit
1906  * side of the connection.
1907  *
1908  * BACKGROUND:  TCP makes no provision for the management of buffer space
1909  * at the end points or at the intermediate routers and switches.  A TCP
1910  * stream, whether using NewReno or not, will eventually buffer as
1911  * many packets as it is able and the only reason this typically works is
1912  * due to the fairly small default buffers made available for a connection
1913  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1914  * scaling it is now fairly easy for even a single TCP connection to blow-out
1915  * all available buffer space not only on the local interface, but on
1916  * intermediate routers and switches as well.  NewReno makes a misguided
1917  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1918  * then backing off, then steadily increasing the window again until another
1919  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1920  * is only made worse as network loads increase and the idea of intentionally
1921  * blowing out network buffers is, frankly, a terrible way to manage network
1922  * resources.
1923  *
1924  * It is far better to limit the transmit window prior to the failure
1925  * condition being achieved.  There are two general ways to do this:  First
1926  * you can 'scan' through different transmit window sizes and locate the
1927  * point where the RTT stops increasing, indicating that you have filled the
1928  * pipe, then scan backwards until you note that RTT stops decreasing, then
1929  * repeat ad-infinitum.  This method works in principle but has severe
1930  * implementation issues due to RTT variances, timer granularity, and
1931  * instability in the algorithm which can lead to many false positives and
1932  * create oscillations as well as interact badly with other TCP streams
1933  * implementing the same algorithm.
1934  *
1935  * The second method is to limit the window to the bandwidth delay product
1936  * of the link.  This is the method we implement.  RTT variances and our
1937  * own manipulation of the congestion window, bwnd, can potentially
1938  * destabilize the algorithm.  For this reason we have to stabilize the
1939  * elements used to calculate the window.  We do this by using the minimum
1940  * observed RTT, the long term average of the observed bandwidth, and
1941  * by adding two segments worth of slop.  It isn't perfect but it is able
1942  * to react to changing conditions and gives us a very stable basis on
1943  * which to extend the algorithm.
1944  */
1945 void
1946 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1947 {
1948 	u_long bw;
1949 	u_long ibw;
1950 	u_long bwnd;
1951 	int save_ticks;
1952 	int delta_ticks;
1953 
1954 	/*
1955 	 * If inflight_enable is disabled in the middle of a tcp connection,
1956 	 * make sure snd_bwnd is effectively disabled.
1957 	 */
1958 	if (!tcp_inflight_enable) {
1959 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1960 		tp->snd_bandwidth = 0;
1961 		return;
1962 	}
1963 
1964 	/*
1965 	 * Validate the delta time.  If a connection is new or has been idle
1966 	 * a long time we have to reset the bandwidth calculator.
1967 	 */
1968 	save_ticks = ticks;
1969 	cpu_ccfence();
1970 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1971 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1972 		tp->t_bw_rtttime = save_ticks;
1973 		tp->t_bw_rtseq = ack_seq;
1974 		if (tp->snd_bandwidth == 0)
1975 			tp->snd_bandwidth = tcp_inflight_start;
1976 		return;
1977 	}
1978 
1979 	/*
1980 	 * A delta of at least 1 tick is required.  Waiting 2 ticks will
1981 	 * result in better (bw) accuracy.  More than that and the ramp-up
1982 	 * will be too slow.
1983 	 */
1984 	if (delta_ticks == 0 || delta_ticks == 1)
1985 		return;
1986 
1987 	/*
1988 	 * Sanity check, plus ignore pure window update acks.
1989 	 */
1990 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1991 		return;
1992 
1993 	/*
1994 	 * Figure out the bandwidth.  Due to the tick granularity this
1995 	 * is a very rough number and it MUST be averaged over a fairly
1996 	 * long period of time.  XXX we need to take into account a link
1997 	 * that is not using all available bandwidth, but for now our
1998 	 * slop will ramp us up if this case occurs and the bandwidth later
1999 	 * increases.
2000 	 */
2001 	ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
2002 	tp->t_bw_rtttime = save_ticks;
2003 	tp->t_bw_rtseq = ack_seq;
2004 	bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4;
2005 
2006 	tp->snd_bandwidth = bw;
2007 
2008 	/*
2009 	 * Calculate the semi-static bandwidth delay product, plus two maximal
2010 	 * segments.  The additional slop puts us squarely in the sweet
2011 	 * spot and also handles the bandwidth run-up case.  Without the
2012 	 * slop we could be locking ourselves into a lower bandwidth.
2013 	 *
2014 	 * At very high speeds the bw calculation can become overly sensitive
2015 	 * and error prone when delta_ticks is low (e.g. usually 1).  To deal
2016 	 * with the problem the stab must be scaled to the bw.  A stab of 50
2017 	 * (the default) increases the bw for the purposes of the bwnd
2018 	 * calculation by 5%.
2019 	 *
2020 	 * Situations Handled:
2021 	 *	(1) Prevents over-queueing of packets on LANs, especially on
2022 	 *	    high speed LANs, allowing larger TCP buffers to be
2023 	 *	    specified, and also does a good job preventing
2024 	 *	    over-queueing of packets over choke points like modems
2025 	 *	    (at least for the transmit side).
2026 	 *
2027 	 *	(2) Is able to handle changing network loads (bandwidth
2028 	 *	    drops so bwnd drops, bandwidth increases so bwnd
2029 	 *	    increases).
2030 	 *
2031 	 *	(3) Theoretically should stabilize in the face of multiple
2032 	 *	    connections implementing the same algorithm (this may need
2033 	 *	    a little work).
2034 	 *
2035 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
2036 	 *	    be adjusted with a sysctl but typically only needs to be on
2037 	 *	    very slow connections.  A value no smaller then 5 should
2038 	 *	    be used, but only reduce this default if you have no other
2039 	 *	    choice.
2040 	 */
2041 
2042 #define	USERTT	((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2043 	bw += bw * tcp_inflight_stab / 1000;
2044 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
2045 	       (int)tp->t_maxseg * 2;
2046 #undef USERTT
2047 
2048 	if (tcp_inflight_debug > 0) {
2049 		static int ltime;
2050 		if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) {
2051 			ltime = save_ticks;
2052 			kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2053 				"bwnd %ld delta %d snd_win %ld\n",
2054 				tp, ibw, bw, tp->t_rttvar, tp->t_srtt,
2055 				bwnd, delta_ticks, tp->snd_wnd);
2056 		}
2057 	}
2058 	if ((long)bwnd < tcp_inflight_min)
2059 		bwnd = tcp_inflight_min;
2060 	if (bwnd > tcp_inflight_max)
2061 		bwnd = tcp_inflight_max;
2062 	if ((long)bwnd < tp->t_maxseg * 2)
2063 		bwnd = tp->t_maxseg * 2;
2064 	tp->snd_bwnd = bwnd;
2065 }
2066 
2067 static void
2068 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs)
2069 {
2070 	struct rtentry *rt;
2071 	struct inpcb *inp = tp->t_inpcb;
2072 #ifdef INET6
2073 	boolean_t isipv6 = INP_ISIPV6(inp);
2074 #else
2075 	const boolean_t isipv6 = FALSE;
2076 #endif
2077 
2078 	/* XXX */
2079 	if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT)
2080 		tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
2081 	if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT)
2082 		tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
2083 
2084 	if (isipv6)
2085 		rt = tcp_rtlookup6(&inp->inp_inc);
2086 	else
2087 		rt = tcp_rtlookup(&inp->inp_inc);
2088 	if (rt == NULL ||
2089 	    rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT ||
2090 	    rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) {
2091 		*maxsegs = tcp_iw_maxsegs;
2092 		*capsegs = tcp_iw_capsegs;
2093 		return;
2094 	}
2095 	*maxsegs = rt->rt_rmx.rmx_iwmaxsegs;
2096 	*capsegs = rt->rt_rmx.rmx_iwcapsegs;
2097 }
2098 
2099 u_long
2100 tcp_initial_window(struct tcpcb *tp)
2101 {
2102 	if (tcp_do_rfc3390) {
2103 		/*
2104 		 * RFC3390:
2105 		 * "If the SYN or SYN/ACK is lost, the initial window
2106 		 *  used by a sender after a correctly transmitted SYN
2107 		 *  MUST be one segment consisting of MSS bytes."
2108 		 *
2109 		 * However, we do something a little bit more aggressive
2110 		 * then RFC3390 here:
2111 		 * - Only if time spent in the SYN or SYN|ACK retransmition
2112 		 *   >= 3 seconds, the IW is reduced.  We do this mainly
2113 		 *   because when RFC3390 is published, the initial RTO is
2114 		 *   still 3 seconds (the threshold we test here), while
2115 		 *   after RFC6298, the initial RTO is 1 second.  This
2116 		 *   behaviour probably still falls within the spirit of
2117 		 *   RFC3390.
2118 		 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2119 		 *   Mainly to avoid sender and receiver deadlock until
2120 		 *   delayed ACK timer expires.  And even RFC2581 does not
2121 		 *   try to reduce IW upon SYN or SYN|ACK retransmition
2122 		 *   timeout.
2123 		 *
2124 		 * See also:
2125 		 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2126 		 */
2127 		if (tp->t_rxtsyn >= TCPTV_RTOBASE3) {
2128 			return (2 * tp->t_maxseg);
2129 		} else {
2130 			u_long maxsegs, capsegs;
2131 
2132 			tcp_rmx_iwsegs(tp, &maxsegs, &capsegs);
2133 			return min(maxsegs * tp->t_maxseg,
2134 				   max(2 * tp->t_maxseg, capsegs * 1460));
2135 		}
2136 	} else {
2137 		/*
2138 		 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2139 		 *
2140 		 * Mainly to avoid sender and receiver deadlock
2141 		 * until delayed ACK timer expires.
2142 		 */
2143 		return (2 * tp->t_maxseg);
2144 	}
2145 }
2146 
2147 #ifdef TCP_SIGNATURE
2148 /*
2149  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2150  *
2151  * We do this over ip, tcphdr, segment data, and the key in the SADB.
2152  * When called from tcp_input(), we can be sure that th_sum has been
2153  * zeroed out and verified already.
2154  *
2155  * Return 0 if successful, otherwise return -1.
2156  *
2157  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2158  * search with the destination IP address, and a 'magic SPI' to be
2159  * determined by the application. This is hardcoded elsewhere to 1179
2160  * right now. Another branch of this code exists which uses the SPD to
2161  * specify per-application flows but it is unstable.
2162  */
2163 int
2164 tcpsignature_compute(
2165 	struct mbuf *m,		/* mbuf chain */
2166 	int len,		/* length of TCP data */
2167 	int optlen,		/* length of TCP options */
2168 	u_char *buf,		/* storage for MD5 digest */
2169 	u_int direction)	/* direction of flow */
2170 {
2171 	struct ippseudo ippseudo;
2172 	MD5_CTX ctx;
2173 	int doff;
2174 	struct ip *ip;
2175 	struct ipovly *ipovly;
2176 	struct secasvar *sav;
2177 	struct tcphdr *th;
2178 #ifdef INET6
2179 	struct ip6_hdr *ip6;
2180 	struct in6_addr in6;
2181 	uint32_t plen;
2182 	uint16_t nhdr;
2183 #endif /* INET6 */
2184 	u_short savecsum;
2185 
2186 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2187 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2188 	/*
2189 	 * Extract the destination from the IP header in the mbuf.
2190 	 */
2191 	ip = mtod(m, struct ip *);
2192 #ifdef INET6
2193 	ip6 = NULL;     /* Make the compiler happy. */
2194 #endif /* INET6 */
2195 	/*
2196 	 * Look up an SADB entry which matches the address found in
2197 	 * the segment.
2198 	 */
2199 	switch (IP_VHL_V(ip->ip_vhl)) {
2200 	case IPVERSION:
2201 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2202 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2203 		break;
2204 #ifdef INET6
2205 	case (IPV6_VERSION >> 4):
2206 		ip6 = mtod(m, struct ip6_hdr *);
2207 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2208 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2209 		break;
2210 #endif /* INET6 */
2211 	default:
2212 		return (EINVAL);
2213 		/* NOTREACHED */
2214 		break;
2215 	}
2216 	if (sav == NULL) {
2217 		kprintf("%s: SADB lookup failed\n", __func__);
2218 		return (EINVAL);
2219 	}
2220 	MD5Init(&ctx);
2221 
2222 	/*
2223 	 * Step 1: Update MD5 hash with IP pseudo-header.
2224 	 *
2225 	 * XXX The ippseudo header MUST be digested in network byte order,
2226 	 * or else we'll fail the regression test. Assume all fields we've
2227 	 * been doing arithmetic on have been in host byte order.
2228 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2229 	 * tcp_output(), the underlying ip_len member has not yet been set.
2230 	 */
2231 	switch (IP_VHL_V(ip->ip_vhl)) {
2232 	case IPVERSION:
2233 		ipovly = (struct ipovly *)ip;
2234 		ippseudo.ippseudo_src = ipovly->ih_src;
2235 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2236 		ippseudo.ippseudo_pad = 0;
2237 		ippseudo.ippseudo_p = IPPROTO_TCP;
2238 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2239 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2240 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2241 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2242 		break;
2243 #ifdef INET6
2244 	/*
2245 	 * RFC 2385, 2.0  Proposal
2246 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2247 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2248 	 * extended next header value (to form 32 bits), and 32-bit segment
2249 	 * length.
2250 	 * Note: Upper-Layer Packet Length comes before Next Header.
2251 	 */
2252 	case (IPV6_VERSION >> 4):
2253 		in6 = ip6->ip6_src;
2254 		in6_clearscope(&in6);
2255 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2256 		in6 = ip6->ip6_dst;
2257 		in6_clearscope(&in6);
2258 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2259 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2260 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2261 		nhdr = 0;
2262 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2263 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2264 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2265 		nhdr = IPPROTO_TCP;
2266 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2267 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2268 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2269 		break;
2270 #endif /* INET6 */
2271 	default:
2272 		return (EINVAL);
2273 		/* NOTREACHED */
2274 		break;
2275 	}
2276 	/*
2277 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2278 	 * The TCP checksum must be set to zero.
2279 	 */
2280 	savecsum = th->th_sum;
2281 	th->th_sum = 0;
2282 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2283 	th->th_sum = savecsum;
2284 	/*
2285 	 * Step 3: Update MD5 hash with TCP segment data.
2286 	 *         Use m_apply() to avoid an early m_pullup().
2287 	 */
2288 	if (len > 0)
2289 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2290 	/*
2291 	 * Step 4: Update MD5 hash with shared secret.
2292 	 */
2293 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2294 	MD5Final(buf, &ctx);
2295 	key_sa_recordxfer(sav, m);
2296 	key_freesav(sav);
2297 	return (0);
2298 }
2299 
2300 int
2301 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2302 {
2303 
2304 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2305 	return (0);
2306 }
2307 #endif /* TCP_SIGNATURE */
2308 
2309 static void
2310 tcp_drop_sysctl_dispatch(netmsg_t nmsg)
2311 {
2312 	struct lwkt_msg *lmsg = &nmsg->lmsg;
2313 	/* addrs[0] is a foreign socket, addrs[1] is a local one. */
2314 	struct sockaddr_storage *addrs = lmsg->u.ms_resultp;
2315 	int error;
2316 	struct sockaddr_in *fin, *lin;
2317 #ifdef INET6
2318 	struct sockaddr_in6 *fin6, *lin6;
2319 	struct in6_addr f6, l6;
2320 #endif
2321 	struct inpcb *inp;
2322 
2323 	switch (addrs[0].ss_family) {
2324 #ifdef INET6
2325 	case AF_INET6:
2326 		fin6 = (struct sockaddr_in6 *)&addrs[0];
2327 		lin6 = (struct sockaddr_in6 *)&addrs[1];
2328 		error = in6_embedscope(&f6, fin6, NULL, NULL);
2329 		if (error)
2330 			goto done;
2331 		error = in6_embedscope(&l6, lin6, NULL, NULL);
2332 		if (error)
2333 			goto done;
2334 		inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6,
2335 		    fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL);
2336 		break;
2337 #endif
2338 #ifdef INET
2339 	case AF_INET:
2340 		fin = (struct sockaddr_in *)&addrs[0];
2341 		lin = (struct sockaddr_in *)&addrs[1];
2342 		inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr,
2343 		    fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL);
2344 		break;
2345 #endif
2346 	default:
2347 		/*
2348 		 * Must not reach here, since the address family was
2349 		 * checked in sysctl handler.
2350 		 */
2351 		panic("unknown address family %d", addrs[0].ss_family);
2352 	}
2353 	if (inp != NULL) {
2354 		struct tcpcb *tp = intotcpcb(inp);
2355 
2356 		KASSERT((inp->inp_flags & INP_WILDCARD) == 0,
2357 		    ("in wildcard hash"));
2358 		KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2359 		KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket"));
2360 		tcp_drop(tp, ECONNABORTED);
2361 		error = 0;
2362 	} else {
2363 		error = ESRCH;
2364 	}
2365 #ifdef INET6
2366 done:
2367 #endif
2368 	lwkt_replymsg(lmsg, error);
2369 }
2370 
2371 static int
2372 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS)
2373 {
2374 	/* addrs[0] is a foreign socket, addrs[1] is a local one. */
2375 	struct sockaddr_storage addrs[2];
2376 	struct sockaddr_in *fin, *lin;
2377 #ifdef INET6
2378 	struct sockaddr_in6 *fin6, *lin6;
2379 #endif
2380 	struct netmsg_base nmsg;
2381 	struct lwkt_msg *lmsg = &nmsg.lmsg;
2382 	struct lwkt_port *port = NULL;
2383 	int error;
2384 
2385 	fin = lin = NULL;
2386 #ifdef INET6
2387 	fin6 = lin6 = NULL;
2388 #endif
2389 	error = 0;
2390 
2391 	if (req->oldptr != NULL || req->oldlen != 0)
2392 		return (EINVAL);
2393 	if (req->newptr == NULL)
2394 		return (EPERM);
2395 	if (req->newlen < sizeof(addrs))
2396 		return (ENOMEM);
2397 	error = SYSCTL_IN(req, &addrs, sizeof(addrs));
2398 	if (error)
2399 		return (error);
2400 
2401 	switch (addrs[0].ss_family) {
2402 #ifdef INET6
2403 	case AF_INET6:
2404 		fin6 = (struct sockaddr_in6 *)&addrs[0];
2405 		lin6 = (struct sockaddr_in6 *)&addrs[1];
2406 		if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
2407 		    lin6->sin6_len != sizeof(struct sockaddr_in6))
2408 			return (EINVAL);
2409 		if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) ||
2410 		    IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2411 			return (EADDRNOTAVAIL);
2412 #if 0
2413 		error = sa6_embedscope(fin6, V_ip6_use_defzone);
2414 		if (error)
2415 			return (error);
2416 		error = sa6_embedscope(lin6, V_ip6_use_defzone);
2417 		if (error)
2418 			return (error);
2419 #endif
2420 		port = tcp6_addrport();
2421 		break;
2422 #endif
2423 #ifdef INET
2424 	case AF_INET:
2425 		fin = (struct sockaddr_in *)&addrs[0];
2426 		lin = (struct sockaddr_in *)&addrs[1];
2427 		if (fin->sin_len != sizeof(struct sockaddr_in) ||
2428 		    lin->sin_len != sizeof(struct sockaddr_in))
2429 			return (EINVAL);
2430 		port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port,
2431 		    lin->sin_addr.s_addr, lin->sin_port);
2432 		break;
2433 #endif
2434 	default:
2435 		return (EINVAL);
2436 	}
2437 
2438 	netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0,
2439 	    tcp_drop_sysctl_dispatch);
2440 	lmsg->u.ms_resultp = addrs;
2441 	return lwkt_domsg(port, lmsg, 0);
2442 }
2443 
2444 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop,
2445     CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL,
2446     0, sysctl_tcp_drop, "", "Drop TCP connection");
2447