xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 3cf8dfbcc5e851c3571a9caa420ccd50eb89a824)
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_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
83 #include <sys/mbuf.h>
84 #ifdef INET6
85 #include <sys/domain.h>
86 #endif
87 #include <sys/proc.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/protosw.h>
91 #include <sys/random.h>
92 #include <sys/in_cksum.h>
93 #include <sys/ktr.h>
94 
95 #include <vm/vm_zone.h>
96 
97 #include <net/route.h>
98 #include <net/if.h>
99 #include <net/netisr.h>
100 
101 #define	_IP_VHL
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
112 #ifdef INET6
113 #include <netinet/icmp6.h>
114 #endif
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_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 #ifdef INET6
131 #include <netinet6/ipsec6.h>
132 #endif
133 #endif
134 
135 #ifdef FAST_IPSEC
136 #include <netproto/ipsec/ipsec.h>
137 #ifdef INET6
138 #include <netproto/ipsec/ipsec6.h>
139 #endif
140 #define	IPSEC
141 #endif
142 
143 #include <sys/md5.h>
144 #include <sys/msgport2.h>
145 #include <machine/smp.h>
146 
147 #include <net/netmsg2.h>
148 
149 #if !defined(KTR_TCP)
150 #define KTR_TCP		KTR_ALL
151 #endif
152 KTR_INFO_MASTER(tcp);
153 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
154 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
155 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
156 #define logtcp(name)	KTR_LOG(tcp_ ## name)
157 
158 struct inpcbinfo tcbinfo[MAXCPU];
159 struct tcpcbackqhead tcpcbackq[MAXCPU];
160 
161 int tcp_mpsafe_proto = 0;
162 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
163 
164 static int tcp_mpsafe_thread = 0;
165 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
166 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
167 	   &tcp_mpsafe_thread, 0,
168 	   "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
169 
170 int tcp_mssdflt = TCP_MSS;
171 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
172     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
173 
174 #ifdef INET6
175 int tcp_v6mssdflt = TCP6_MSS;
176 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
177     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
178 #endif
179 
180 #if 0
181 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
182 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
183     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
184 #endif
185 
186 int tcp_do_rfc1323 = 1;
187 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
188     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
189 
190 int tcp_do_rfc1644 = 0;
191 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
192     &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
193 
194 static int tcp_tcbhashsize = 0;
195 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
196      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
197 
198 static int do_tcpdrain = 1;
199 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
200      "Enable tcp_drain routine for extra help when low on mbufs");
201 
202 /* XXX JH */
203 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
204     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
205 
206 static int icmp_may_rst = 1;
207 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
208     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
209 
210 static int tcp_isn_reseed_interval = 0;
211 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
212     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
213 
214 /*
215  * TCP bandwidth limiting sysctls.  Note that the default lower bound of
216  * 1024 exists only for debugging.  A good production default would be
217  * something like 6100.
218  */
219 static int tcp_inflight_enable = 0;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
221     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
222 
223 static int tcp_inflight_debug = 0;
224 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
225     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
226 
227 static int tcp_inflight_min = 6144;
228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
229     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
230 
231 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
233     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
234 
235 static int tcp_inflight_stab = 20;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
237     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
238 
239 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
240 static struct malloc_pipe tcptemp_mpipe;
241 
242 static void tcp_willblock(int);
243 static void tcp_cleartaocache (void);
244 static void tcp_notify (struct inpcb *, int);
245 
246 struct tcp_stats tcpstats_percpu[MAXCPU];
247 #ifdef SMP
248 static int
249 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
250 {
251 	int cpu, error = 0;
252 
253 	for (cpu = 0; cpu < ncpus; ++cpu) {
254 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
255 					sizeof(struct tcp_stats))))
256 			break;
257 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
258 				       sizeof(struct tcp_stats))))
259 			break;
260 	}
261 
262 	return (error);
263 }
264 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
265     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
266 #else
267 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
268     &tcpstat, tcp_stats, "TCP statistics");
269 #endif
270 
271 /*
272  * Target size of TCP PCB hash tables. Must be a power of two.
273  *
274  * Note that this can be overridden by the kernel environment
275  * variable net.inet.tcp.tcbhashsize
276  */
277 #ifndef TCBHASHSIZE
278 #define	TCBHASHSIZE	512
279 #endif
280 
281 /*
282  * This is the actual shape of what we allocate using the zone
283  * allocator.  Doing it this way allows us to protect both structures
284  * using the same generation count, and also eliminates the overhead
285  * of allocating tcpcbs separately.  By hiding the structure here,
286  * we avoid changing most of the rest of the code (although it needs
287  * to be changed, eventually, for greater efficiency).
288  */
289 #define	ALIGNMENT	32
290 #define	ALIGNM1		(ALIGNMENT - 1)
291 struct	inp_tp {
292 	union {
293 		struct	inpcb inp;
294 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
295 	} inp_tp_u;
296 	struct	tcpcb tcb;
297 	struct	tcp_callout inp_tp_rexmt;
298 	struct	tcp_callout inp_tp_persist;
299 	struct	tcp_callout inp_tp_keep;
300 	struct	tcp_callout inp_tp_2msl;
301 	struct	tcp_callout inp_tp_delack;
302 	struct	netmsg_tcp_timer inp_tp_timermsg;
303 };
304 #undef ALIGNMENT
305 #undef ALIGNM1
306 
307 /*
308  * Tcp initialization
309  */
310 void
311 tcp_init(void)
312 {
313 	struct inpcbporthead *porthashbase;
314 	u_long porthashmask;
315 	struct vm_zone *ipi_zone;
316 	int hashsize = TCBHASHSIZE;
317 	int cpu;
318 
319 	/*
320 	 * note: tcptemp is used for keepalives, and it is ok for an
321 	 * allocation to fail so do not specify MPF_INT.
322 	 */
323 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
324 		    25, -1, 0, NULL);
325 
326 	tcp_ccgen = 1;
327 	tcp_cleartaocache();
328 
329 	tcp_delacktime = TCPTV_DELACK;
330 	tcp_keepinit = TCPTV_KEEP_INIT;
331 	tcp_keepidle = TCPTV_KEEP_IDLE;
332 	tcp_keepintvl = TCPTV_KEEPINTVL;
333 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
334 	tcp_msl = TCPTV_MSL;
335 	tcp_rexmit_min = TCPTV_MIN;
336 	tcp_rexmit_slop = TCPTV_CPU_VAR;
337 
338 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
339 	if (!powerof2(hashsize)) {
340 		kprintf("WARNING: TCB hash size not a power of 2\n");
341 		hashsize = 512; /* safe default */
342 	}
343 	tcp_tcbhashsize = hashsize;
344 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
345 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
346 			 ZONE_INTERRUPT, 0);
347 
348 	for (cpu = 0; cpu < ncpus2; cpu++) {
349 		in_pcbinfo_init(&tcbinfo[cpu]);
350 		tcbinfo[cpu].cpu = cpu;
351 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
352 		    &tcbinfo[cpu].hashmask);
353 		tcbinfo[cpu].porthashbase = porthashbase;
354 		tcbinfo[cpu].porthashmask = porthashmask;
355 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
356 		    &tcbinfo[cpu].wildcardhashmask);
357 		tcbinfo[cpu].ipi_zone = ipi_zone;
358 		TAILQ_INIT(&tcpcbackq[cpu]);
359 	}
360 
361 	tcp_reass_maxseg = nmbclusters / 16;
362 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
363 
364 #ifdef INET6
365 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
366 #else
367 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
368 #endif
369 	if (max_protohdr < TCP_MINPROTOHDR)
370 		max_protohdr = TCP_MINPROTOHDR;
371 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
372 		panic("tcp_init");
373 #undef TCP_MINPROTOHDR
374 
375 	/*
376 	 * Initialize TCP statistics counters for each CPU.
377 	 */
378 #ifdef SMP
379 	for (cpu = 0; cpu < ncpus; ++cpu) {
380 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
381 	}
382 #else
383 	bzero(&tcpstat, sizeof(struct tcp_stats));
384 #endif
385 
386 	syncache_init();
387 	tcp_thread_init();
388 }
389 
390 void
391 tcpmsg_service_loop(void *dummy)
392 {
393 	struct netmsg *msg;
394 	int mplocked;
395 
396 	/*
397 	 * Thread was started with TDF_MPSAFE
398 	 */
399 	mplocked = 0;
400 
401 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
402 		do {
403 			logtcp(rxmsg);
404 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
405 						  mplocked);
406 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
407 
408 		logtcp(delayed);
409 		tcp_willblock(mplocked);
410 		logtcp(wait);
411 	}
412 }
413 
414 static void
415 tcp_willblock(int mplocked)
416 {
417 	struct tcpcb *tp;
418 	int cpu = mycpu->gd_cpuid;
419 	int unlock = 0;
420 
421 	if (!mplocked && !tcp_mpsafe_proto) {
422 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
423 			return;
424 
425 		get_mplock();
426 		mplocked = 1;
427 		unlock = 1;
428 	}
429 
430 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
431 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
432 		tp->t_flags &= ~TF_ONOUTPUTQ;
433 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
434 		tcp_output(tp);
435 	}
436 
437 	if (unlock)
438 		rel_mplock();
439 }
440 
441 
442 /*
443  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
444  * tcp_template used to store this data in mbufs, but we now recopy it out
445  * of the tcpcb each time to conserve mbufs.
446  */
447 void
448 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
449 {
450 	struct inpcb *inp = tp->t_inpcb;
451 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
452 
453 #ifdef INET6
454 	if (inp->inp_vflag & INP_IPV6) {
455 		struct ip6_hdr *ip6;
456 
457 		ip6 = (struct ip6_hdr *)ip_ptr;
458 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
459 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
460 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
461 			(IPV6_VERSION & IPV6_VERSION_MASK);
462 		ip6->ip6_nxt = IPPROTO_TCP;
463 		ip6->ip6_plen = sizeof(struct tcphdr);
464 		ip6->ip6_src = inp->in6p_laddr;
465 		ip6->ip6_dst = inp->in6p_faddr;
466 		tcp_hdr->th_sum = 0;
467 	} else
468 #endif
469 	{
470 		struct ip *ip = (struct ip *) ip_ptr;
471 
472 		ip->ip_vhl = IP_VHL_BORING;
473 		ip->ip_tos = 0;
474 		ip->ip_len = 0;
475 		ip->ip_id = 0;
476 		ip->ip_off = 0;
477 		ip->ip_ttl = 0;
478 		ip->ip_sum = 0;
479 		ip->ip_p = IPPROTO_TCP;
480 		ip->ip_src = inp->inp_laddr;
481 		ip->ip_dst = inp->inp_faddr;
482 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
483 				    ip->ip_dst.s_addr,
484 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
485 	}
486 
487 	tcp_hdr->th_sport = inp->inp_lport;
488 	tcp_hdr->th_dport = inp->inp_fport;
489 	tcp_hdr->th_seq = 0;
490 	tcp_hdr->th_ack = 0;
491 	tcp_hdr->th_x2 = 0;
492 	tcp_hdr->th_off = 5;
493 	tcp_hdr->th_flags = 0;
494 	tcp_hdr->th_win = 0;
495 	tcp_hdr->th_urp = 0;
496 }
497 
498 /*
499  * Create template to be used to send tcp packets on a connection.
500  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
501  * use for this function is in keepalives, which use tcp_respond.
502  */
503 struct tcptemp *
504 tcp_maketemplate(struct tcpcb *tp)
505 {
506 	struct tcptemp *tmp;
507 
508 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
509 		return (NULL);
510 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
511 	return (tmp);
512 }
513 
514 void
515 tcp_freetemplate(struct tcptemp *tmp)
516 {
517 	mpipe_free(&tcptemp_mpipe, tmp);
518 }
519 
520 /*
521  * Send a single message to the TCP at address specified by
522  * the given TCP/IP header.  If m == NULL, then we make a copy
523  * of the tcpiphdr at ti and send directly to the addressed host.
524  * This is used to force keep alive messages out using the TCP
525  * template for a connection.  If flags are given then we send
526  * a message back to the TCP which originated the * segment ti,
527  * and discard the mbuf containing it and any other attached mbufs.
528  *
529  * In any case the ack and sequence number of the transmitted
530  * segment are as specified by the parameters.
531  *
532  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
533  */
534 void
535 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
536 	    tcp_seq ack, tcp_seq seq, int flags)
537 {
538 	int tlen;
539 	int win = 0;
540 	struct route *ro = NULL;
541 	struct route sro;
542 	struct ip *ip = ipgen;
543 	struct tcphdr *nth;
544 	int ipflags = 0;
545 	struct route_in6 *ro6 = NULL;
546 	struct route_in6 sro6;
547 	struct ip6_hdr *ip6 = ipgen;
548 	boolean_t use_tmpro = TRUE;
549 #ifdef INET6
550 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
551 #else
552 	const boolean_t isipv6 = FALSE;
553 #endif
554 
555 	if (tp != NULL) {
556 		if (!(flags & TH_RST)) {
557 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
558 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
559 				win = (long)TCP_MAXWIN << tp->rcv_scale;
560 		}
561 		/*
562 		 * Don't use the route cache of a listen socket,
563 		 * it is not MPSAFE; use temporary route cache.
564 		 */
565 		if (tp->t_state != TCPS_LISTEN) {
566 			if (isipv6)
567 				ro6 = &tp->t_inpcb->in6p_route;
568 			else
569 				ro = &tp->t_inpcb->inp_route;
570 			use_tmpro = FALSE;
571 		}
572 	}
573 	if (use_tmpro) {
574 		if (isipv6) {
575 			ro6 = &sro6;
576 			bzero(ro6, sizeof *ro6);
577 		} else {
578 			ro = &sro;
579 			bzero(ro, sizeof *ro);
580 		}
581 	}
582 	if (m == NULL) {
583 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
584 		if (m == NULL)
585 			return;
586 		tlen = 0;
587 		m->m_data += max_linkhdr;
588 		if (isipv6) {
589 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
590 			ip6 = mtod(m, struct ip6_hdr *);
591 			nth = (struct tcphdr *)(ip6 + 1);
592 		} else {
593 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
594 			ip = mtod(m, struct ip *);
595 			nth = (struct tcphdr *)(ip + 1);
596 		}
597 		bcopy(th, nth, sizeof(struct tcphdr));
598 		flags = TH_ACK;
599 	} else {
600 		m_freem(m->m_next);
601 		m->m_next = NULL;
602 		m->m_data = (caddr_t)ipgen;
603 		/* m_len is set later */
604 		tlen = 0;
605 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
606 		if (isipv6) {
607 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
608 			nth = (struct tcphdr *)(ip6 + 1);
609 		} else {
610 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
611 			nth = (struct tcphdr *)(ip + 1);
612 		}
613 		if (th != nth) {
614 			/*
615 			 * this is usually a case when an extension header
616 			 * exists between the IPv6 header and the
617 			 * TCP header.
618 			 */
619 			nth->th_sport = th->th_sport;
620 			nth->th_dport = th->th_dport;
621 		}
622 		xchg(nth->th_dport, nth->th_sport, n_short);
623 #undef xchg
624 	}
625 	if (isipv6) {
626 		ip6->ip6_flow = 0;
627 		ip6->ip6_vfc = IPV6_VERSION;
628 		ip6->ip6_nxt = IPPROTO_TCP;
629 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
630 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
631 	} else {
632 		tlen += sizeof(struct tcpiphdr);
633 		ip->ip_len = tlen;
634 		ip->ip_ttl = ip_defttl;
635 	}
636 	m->m_len = tlen;
637 	m->m_pkthdr.len = tlen;
638 	m->m_pkthdr.rcvif = (struct ifnet *) NULL;
639 	nth->th_seq = htonl(seq);
640 	nth->th_ack = htonl(ack);
641 	nth->th_x2 = 0;
642 	nth->th_off = sizeof(struct tcphdr) >> 2;
643 	nth->th_flags = flags;
644 	if (tp != NULL)
645 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
646 	else
647 		nth->th_win = htons((u_short)win);
648 	nth->th_urp = 0;
649 	if (isipv6) {
650 		nth->th_sum = 0;
651 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
652 					sizeof(struct ip6_hdr),
653 					tlen - sizeof(struct ip6_hdr));
654 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
655 					       (ro6 && ro6->ro_rt) ?
656 						ro6->ro_rt->rt_ifp : NULL);
657 	} else {
658 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
659 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
660 		m->m_pkthdr.csum_flags = CSUM_TCP;
661 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
662 	}
663 #ifdef TCPDEBUG
664 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
665 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
666 #endif
667 	if (isipv6) {
668 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
669 			   tp ? tp->t_inpcb : NULL);
670 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
671 			RTFREE(ro6->ro_rt);
672 			ro6->ro_rt = NULL;
673 		}
674 	} else {
675 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
676 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
677 			RTFREE(ro->ro_rt);
678 			ro->ro_rt = NULL;
679 		}
680 	}
681 }
682 
683 /*
684  * Create a new TCP control block, making an
685  * empty reassembly queue and hooking it to the argument
686  * protocol control block.  The `inp' parameter must have
687  * come from the zone allocator set up in tcp_init().
688  */
689 struct tcpcb *
690 tcp_newtcpcb(struct inpcb *inp)
691 {
692 	struct inp_tp *it;
693 	struct tcpcb *tp;
694 #ifdef INET6
695 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
696 #else
697 	const boolean_t isipv6 = FALSE;
698 #endif
699 
700 	it = (struct inp_tp *)inp;
701 	tp = &it->tcb;
702 	bzero(tp, sizeof(struct tcpcb));
703 	LIST_INIT(&tp->t_segq);
704 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
705 
706 	/* Set up our timeouts. */
707 	tp->tt_rexmt = &it->inp_tp_rexmt;
708 	tp->tt_persist = &it->inp_tp_persist;
709 	tp->tt_keep = &it->inp_tp_keep;
710 	tp->tt_2msl = &it->inp_tp_2msl;
711 	tp->tt_delack = &it->inp_tp_delack;
712 	tcp_inittimers(tp);
713 
714 	tp->tt_msg = &it->inp_tp_timermsg;
715 	if (isipv6) {
716 		/* Don't mess with IPv6; always create timer message */
717 		tcp_create_timermsg(tp);
718 	} else {
719 		/*
720 		 * Zero out timer message.  We don't create it here,
721 		 * since the current CPU may not be the owner of this
722 		 * inpcb.
723 		 */
724 		bzero(tp->tt_msg, sizeof(*tp->tt_msg));
725 	}
726 
727 	if (tcp_do_rfc1323)
728 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
729 	if (tcp_do_rfc1644)
730 		tp->t_flags |= TF_REQ_CC;
731 	tp->t_inpcb = inp;	/* XXX */
732 	tp->t_state = TCPS_CLOSED;
733 	/*
734 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
735 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
736 	 * reasonable initial retransmit time.
737 	 */
738 	tp->t_srtt = TCPTV_SRTTBASE;
739 	tp->t_rttvar =
740 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
741 	tp->t_rttmin = tcp_rexmit_min;
742 	tp->t_rxtcur = TCPTV_RTOBASE;
743 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
744 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
745 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
746 	tp->t_rcvtime = ticks;
747 	/*
748 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
749 	 * because the socket may be bound to an IPv6 wildcard address,
750 	 * which may match an IPv4-mapped IPv6 address.
751 	 */
752 	inp->inp_ip_ttl = ip_defttl;
753 	inp->inp_ppcb = tp;
754 	tcp_sack_tcpcb_init(tp);
755 	return (tp);		/* XXX */
756 }
757 
758 /*
759  * Drop a TCP connection, reporting the specified error.
760  * If connection is synchronized, then send a RST to peer.
761  */
762 struct tcpcb *
763 tcp_drop(struct tcpcb *tp, int error)
764 {
765 	struct socket *so = tp->t_inpcb->inp_socket;
766 
767 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
768 		tp->t_state = TCPS_CLOSED;
769 		tcp_output(tp);
770 		tcpstat.tcps_drops++;
771 	} else
772 		tcpstat.tcps_conndrops++;
773 	if (error == ETIMEDOUT && tp->t_softerror)
774 		error = tp->t_softerror;
775 	so->so_error = error;
776 	return (tcp_close(tp));
777 }
778 
779 #ifdef SMP
780 
781 struct netmsg_remwildcard {
782 	struct netmsg		nm_netmsg;
783 	struct inpcb		*nm_inp;
784 	struct inpcbinfo	*nm_pcbinfo;
785 #if defined(INET6)
786 	int			nm_isinet6;
787 #else
788 	int			nm_unused01;
789 #endif
790 };
791 
792 /*
793  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
794  * inp can be detached.  We do this by cycling through the cpus, ending up
795  * on the cpu controlling the inp last and then doing the disconnect.
796  */
797 static void
798 in_pcbremwildcardhash_handler(struct netmsg *msg0)
799 {
800 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
801 	int cpu;
802 
803 	cpu = msg->nm_pcbinfo->cpu;
804 
805 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
806 		/* note: detach removes any wildcard hash entry */
807 #ifdef INET6
808 		if (msg->nm_isinet6)
809 			in6_pcbdetach(msg->nm_inp);
810 		else
811 #endif
812 			in_pcbdetach(msg->nm_inp);
813 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
814 	} else {
815 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
816 		cpu = (cpu + 1) % ncpus2;
817 		msg->nm_pcbinfo = &tcbinfo[cpu];
818 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
819 	}
820 }
821 
822 #endif
823 
824 /*
825  * Close a TCP control block:
826  *	discard all space held by the tcp
827  *	discard internet protocol block
828  *	wake up any sleepers
829  */
830 struct tcpcb *
831 tcp_close(struct tcpcb *tp)
832 {
833 	struct tseg_qent *q;
834 	struct inpcb *inp = tp->t_inpcb;
835 	struct socket *so = inp->inp_socket;
836 	struct rtentry *rt;
837 	boolean_t dosavessthresh;
838 #ifdef SMP
839 	int cpu;
840 #endif
841 #ifdef INET6
842 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
843 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
844 #else
845 	const boolean_t isipv6 = FALSE;
846 #endif
847 
848 	/*
849 	 * The tp is not instantly destroyed in the wildcard case.  Setting
850 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
851 	 * messing with it, though it should be noted that this change may
852 	 * not take effect on other cpus until we have chained the wildcard
853 	 * hash removal.
854 	 *
855 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
856 	 * update and prevent other tcp protocol threads from accepting new
857 	 * connections on the listen socket we might be trying to close down.
858 	 */
859 	KKASSERT(tp->t_state != TCPS_TERMINATING);
860 	tp->t_state = TCPS_TERMINATING;
861 
862 	/*
863 	 * Make sure that all of our timers are stopped before we
864 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
865 	 * timers are never used.
866 	 */
867 	if (tp->tt_msg != NULL) {
868 		tcp_callout_stop(tp, tp->tt_rexmt);
869 		tcp_callout_stop(tp, tp->tt_persist);
870 		tcp_callout_stop(tp, tp->tt_keep);
871 		tcp_callout_stop(tp, tp->tt_2msl);
872 		tcp_callout_stop(tp, tp->tt_delack);
873 	}
874 
875 	if (tp->t_flags & TF_ONOUTPUTQ) {
876 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
877 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
878 		tp->t_flags &= ~TF_ONOUTPUTQ;
879 	}
880 
881 	/*
882 	 * If we got enough samples through the srtt filter,
883 	 * save the rtt and rttvar in the routing entry.
884 	 * 'Enough' is arbitrarily defined as the 16 samples.
885 	 * 16 samples is enough for the srtt filter to converge
886 	 * to within 5% of the correct value; fewer samples and
887 	 * we could save a very bogus rtt.
888 	 *
889 	 * Don't update the default route's characteristics and don't
890 	 * update anything that the user "locked".
891 	 */
892 	if (tp->t_rttupdated >= 16) {
893 		u_long i = 0;
894 
895 		if (isipv6) {
896 			struct sockaddr_in6 *sin6;
897 
898 			if ((rt = inp->in6p_route.ro_rt) == NULL)
899 				goto no_valid_rt;
900 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
901 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
902 				goto no_valid_rt;
903 		} else
904 			if ((rt = inp->inp_route.ro_rt) == NULL ||
905 			    ((struct sockaddr_in *)rt_key(rt))->
906 			     sin_addr.s_addr == INADDR_ANY)
907 				goto no_valid_rt;
908 
909 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
910 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
911 			if (rt->rt_rmx.rmx_rtt && i)
912 				/*
913 				 * filter this update to half the old & half
914 				 * the new values, converting scale.
915 				 * See route.h and tcp_var.h for a
916 				 * description of the scaling constants.
917 				 */
918 				rt->rt_rmx.rmx_rtt =
919 				    (rt->rt_rmx.rmx_rtt + i) / 2;
920 			else
921 				rt->rt_rmx.rmx_rtt = i;
922 			tcpstat.tcps_cachedrtt++;
923 		}
924 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
925 			i = tp->t_rttvar *
926 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
927 			if (rt->rt_rmx.rmx_rttvar && i)
928 				rt->rt_rmx.rmx_rttvar =
929 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
930 			else
931 				rt->rt_rmx.rmx_rttvar = i;
932 			tcpstat.tcps_cachedrttvar++;
933 		}
934 		/*
935 		 * The old comment here said:
936 		 * update the pipelimit (ssthresh) if it has been updated
937 		 * already or if a pipesize was specified & the threshhold
938 		 * got below half the pipesize.  I.e., wait for bad news
939 		 * before we start updating, then update on both good
940 		 * and bad news.
941 		 *
942 		 * But we want to save the ssthresh even if no pipesize is
943 		 * specified explicitly in the route, because such
944 		 * connections still have an implicit pipesize specified
945 		 * by the global tcp_sendspace.  In the absence of a reliable
946 		 * way to calculate the pipesize, it will have to do.
947 		 */
948 		i = tp->snd_ssthresh;
949 		if (rt->rt_rmx.rmx_sendpipe != 0)
950 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
951 		else
952 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
953 		if (dosavessthresh ||
954 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
955 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
956 			/*
957 			 * convert the limit from user data bytes to
958 			 * packets then to packet data bytes.
959 			 */
960 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
961 			if (i < 2)
962 				i = 2;
963 			i *= tp->t_maxseg +
964 			     (isipv6 ?
965 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
966 			      sizeof(struct tcpiphdr));
967 			if (rt->rt_rmx.rmx_ssthresh)
968 				rt->rt_rmx.rmx_ssthresh =
969 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
970 			else
971 				rt->rt_rmx.rmx_ssthresh = i;
972 			tcpstat.tcps_cachedssthresh++;
973 		}
974 	}
975 
976 no_valid_rt:
977 	/* free the reassembly queue, if any */
978 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
979 		LIST_REMOVE(q, tqe_q);
980 		m_freem(q->tqe_m);
981 		FREE(q, M_TSEGQ);
982 		tcp_reass_qsize--;
983 	}
984 	/* throw away SACK blocks in scoreboard*/
985 	if (TCP_DO_SACK(tp))
986 		tcp_sack_cleanup(&tp->scb);
987 
988 	inp->inp_ppcb = NULL;
989 	soisdisconnected(so);
990 
991 	tcp_destroy_timermsg(tp);
992 
993 	/*
994 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
995 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
996 	 * for each protocol thread and must be removed in the context of
997 	 * that thread.  This is accomplished by chaining the message
998 	 * through the cpus.
999 	 *
1000 	 * If the inp is not wildcarded we simply detach, which will remove
1001 	 * the any hashes still present for this inp.
1002 	 */
1003 #ifdef SMP
1004 	if (inp->inp_flags & INP_WILDCARD_MP) {
1005 		struct netmsg_remwildcard *msg;
1006 
1007 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1008 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
1009 			      M_LWKTMSG, M_INTWAIT);
1010 		netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1011 			    in_pcbremwildcardhash_handler);
1012 #ifdef INET6
1013 		msg->nm_isinet6 = isafinet6;
1014 #endif
1015 		msg->nm_inp = inp;
1016 		msg->nm_pcbinfo = &tcbinfo[cpu];
1017 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1018 	} else
1019 #endif
1020 	{
1021 		/* note: detach removes any wildcard hash entry */
1022 #ifdef INET6
1023 		if (isafinet6)
1024 			in6_pcbdetach(inp);
1025 		else
1026 #endif
1027 			in_pcbdetach(inp);
1028 	}
1029 	tcpstat.tcps_closed++;
1030 	return (NULL);
1031 }
1032 
1033 static __inline void
1034 tcp_drain_oncpu(struct inpcbhead *head)
1035 {
1036 	struct inpcb *inpb;
1037 	struct tcpcb *tcpb;
1038 	struct tseg_qent *te;
1039 
1040 	LIST_FOREACH(inpb, head, inp_list) {
1041 		if (inpb->inp_flags & INP_PLACEMARKER)
1042 			continue;
1043 		if ((tcpb = intotcpcb(inpb))) {
1044 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1045 				LIST_REMOVE(te, tqe_q);
1046 				m_freem(te->tqe_m);
1047 				FREE(te, M_TSEGQ);
1048 				tcp_reass_qsize--;
1049 			}
1050 		}
1051 	}
1052 }
1053 
1054 #ifdef SMP
1055 struct netmsg_tcp_drain {
1056 	struct netmsg		nm_netmsg;
1057 	struct inpcbhead	*nm_head;
1058 };
1059 
1060 static void
1061 tcp_drain_handler(netmsg_t netmsg)
1062 {
1063 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1064 
1065 	tcp_drain_oncpu(nm->nm_head);
1066 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1067 }
1068 #endif
1069 
1070 void
1071 tcp_drain(void)
1072 {
1073 #ifdef SMP
1074 	int cpu;
1075 #endif
1076 
1077 	if (!do_tcpdrain)
1078 		return;
1079 
1080 	/*
1081 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1082 	 * if there is one...
1083 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1084 	 *	reassembly queue should be flushed, but in a situation
1085 	 *	where we're really low on mbufs, this is potentially
1086 	 *	useful.
1087 	 */
1088 #ifdef SMP
1089 	for (cpu = 0; cpu < ncpus2; cpu++) {
1090 		struct netmsg_tcp_drain *msg;
1091 
1092 		if (cpu == mycpu->gd_cpuid) {
1093 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1094 		} else {
1095 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1096 				    M_LWKTMSG, M_NOWAIT);
1097 			if (msg == NULL)
1098 				continue;
1099 			netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1100 				    tcp_drain_handler);
1101 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1102 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1103 		}
1104 	}
1105 #else
1106 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1107 #endif
1108 }
1109 
1110 /*
1111  * Notify a tcp user of an asynchronous error;
1112  * store error as soft error, but wake up user
1113  * (for now, won't do anything until can select for soft error).
1114  *
1115  * Do not wake up user since there currently is no mechanism for
1116  * reporting soft errors (yet - a kqueue filter may be added).
1117  */
1118 static void
1119 tcp_notify(struct inpcb *inp, int error)
1120 {
1121 	struct tcpcb *tp = intotcpcb(inp);
1122 
1123 	/*
1124 	 * Ignore some errors if we are hooked up.
1125 	 * If connection hasn't completed, has retransmitted several times,
1126 	 * and receives a second error, give up now.  This is better
1127 	 * than waiting a long time to establish a connection that
1128 	 * can never complete.
1129 	 */
1130 	if (tp->t_state == TCPS_ESTABLISHED &&
1131 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1132 	      error == EHOSTDOWN)) {
1133 		return;
1134 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1135 	    tp->t_softerror)
1136 		tcp_drop(tp, error);
1137 	else
1138 		tp->t_softerror = error;
1139 #if 0
1140 	wakeup(&so->so_timeo);
1141 	sorwakeup(so);
1142 	sowwakeup(so);
1143 #endif
1144 }
1145 
1146 static int
1147 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1148 {
1149 	int error, i, n;
1150 	struct inpcb *marker;
1151 	struct inpcb *inp;
1152 	inp_gen_t gencnt;
1153 	globaldata_t gd;
1154 	int origcpu, ccpu;
1155 
1156 	error = 0;
1157 	n = 0;
1158 
1159 	/*
1160 	 * The process of preparing the TCB list is too time-consuming and
1161 	 * resource-intensive to repeat twice on every request.
1162 	 */
1163 	if (req->oldptr == NULL) {
1164 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1165 			gd = globaldata_find(ccpu);
1166 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1167 		}
1168 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1169 		return (0);
1170 	}
1171 
1172 	if (req->newptr != NULL)
1173 		return (EPERM);
1174 
1175 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1176 	marker->inp_flags |= INP_PLACEMARKER;
1177 
1178 	/*
1179 	 * OK, now we're committed to doing something.  Run the inpcb list
1180 	 * for each cpu in the system and construct the output.  Use a
1181 	 * list placemarker to deal with list changes occuring during
1182 	 * copyout blockages (but otherwise depend on being on the correct
1183 	 * cpu to avoid races).
1184 	 */
1185 	origcpu = mycpu->gd_cpuid;
1186 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1187 		globaldata_t rgd;
1188 		caddr_t inp_ppcb;
1189 		struct xtcpcb xt;
1190 		int cpu_id;
1191 
1192 		cpu_id = (origcpu + ccpu) % ncpus;
1193 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1194 			continue;
1195 		rgd = globaldata_find(cpu_id);
1196 		lwkt_setcpu_self(rgd);
1197 
1198 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1199 		n = tcbinfo[cpu_id].ipi_count;
1200 
1201 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1202 		i = 0;
1203 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1204 			/*
1205 			 * process a snapshot of pcbs, ignoring placemarkers
1206 			 * and using our own to allow SYSCTL_OUT to block.
1207 			 */
1208 			LIST_REMOVE(marker, inp_list);
1209 			LIST_INSERT_AFTER(inp, marker, inp_list);
1210 
1211 			if (inp->inp_flags & INP_PLACEMARKER)
1212 				continue;
1213 			if (inp->inp_gencnt > gencnt)
1214 				continue;
1215 			if (prison_xinpcb(req->td, inp))
1216 				continue;
1217 
1218 			xt.xt_len = sizeof xt;
1219 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1220 			inp_ppcb = inp->inp_ppcb;
1221 			if (inp_ppcb != NULL)
1222 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1223 			else
1224 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1225 			if (inp->inp_socket)
1226 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1227 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1228 				break;
1229 			++i;
1230 		}
1231 		LIST_REMOVE(marker, inp_list);
1232 		if (error == 0 && i < n) {
1233 			bzero(&xt, sizeof xt);
1234 			xt.xt_len = sizeof xt;
1235 			while (i < n) {
1236 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1237 				if (error)
1238 					break;
1239 				++i;
1240 			}
1241 		}
1242 	}
1243 
1244 	/*
1245 	 * Make sure we are on the same cpu we were on originally, since
1246 	 * higher level callers expect this.  Also don't pollute caches with
1247 	 * migrated userland data by (eventually) returning to userland
1248 	 * on a different cpu.
1249 	 */
1250 	lwkt_setcpu_self(globaldata_find(origcpu));
1251 	kfree(marker, M_TEMP);
1252 	return (error);
1253 }
1254 
1255 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1256 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1257 
1258 static int
1259 tcp_getcred(SYSCTL_HANDLER_ARGS)
1260 {
1261 	struct sockaddr_in addrs[2];
1262 	struct inpcb *inp;
1263 	int cpu;
1264 	int error;
1265 
1266 	error = suser(req->td);
1267 	if (error != 0)
1268 		return (error);
1269 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1270 	if (error != 0)
1271 		return (error);
1272 	crit_enter();
1273 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1274 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1275 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1276 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1277 	if (inp == NULL || inp->inp_socket == NULL) {
1278 		error = ENOENT;
1279 		goto out;
1280 	}
1281 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1282 out:
1283 	crit_exit();
1284 	return (error);
1285 }
1286 
1287 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1288     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1289 
1290 #ifdef INET6
1291 static int
1292 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1293 {
1294 	struct sockaddr_in6 addrs[2];
1295 	struct inpcb *inp;
1296 	int error;
1297 	boolean_t mapped = FALSE;
1298 
1299 	error = suser(req->td);
1300 	if (error != 0)
1301 		return (error);
1302 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1303 	if (error != 0)
1304 		return (error);
1305 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1306 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1307 			mapped = TRUE;
1308 		else
1309 			return (EINVAL);
1310 	}
1311 	crit_enter();
1312 	if (mapped) {
1313 		inp = in_pcblookup_hash(&tcbinfo[0],
1314 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1315 		    addrs[1].sin6_port,
1316 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1317 		    addrs[0].sin6_port,
1318 		    0, NULL);
1319 	} else {
1320 		inp = in6_pcblookup_hash(&tcbinfo[0],
1321 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1322 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1323 		    0, NULL);
1324 	}
1325 	if (inp == NULL || inp->inp_socket == NULL) {
1326 		error = ENOENT;
1327 		goto out;
1328 	}
1329 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1330 out:
1331 	crit_exit();
1332 	return (error);
1333 }
1334 
1335 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1336 	    0, 0,
1337 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1338 #endif
1339 
1340 struct netmsg_tcp_notify {
1341 	struct netmsg	nm_nmsg;
1342 	void		(*nm_notify)(struct inpcb *, int);
1343 	struct in_addr	nm_faddr;
1344 	int		nm_arg;
1345 };
1346 
1347 static void
1348 tcp_notifyall_oncpu(struct netmsg *netmsg)
1349 {
1350 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1351 	int nextcpu;
1352 
1353 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1354 			nmsg->nm_arg, nmsg->nm_notify);
1355 
1356 	nextcpu = mycpuid + 1;
1357 	if (nextcpu < ncpus2)
1358 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1359 	else
1360 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1361 }
1362 
1363 void
1364 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1365 {
1366 	struct ip *ip = vip;
1367 	struct tcphdr *th;
1368 	struct in_addr faddr;
1369 	struct inpcb *inp;
1370 	struct tcpcb *tp;
1371 	void (*notify)(struct inpcb *, int) = tcp_notify;
1372 	tcp_seq icmpseq;
1373 	int arg, cpu;
1374 
1375 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1376 		return;
1377 	}
1378 
1379 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1380 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1381 		return;
1382 
1383 	arg = inetctlerrmap[cmd];
1384 	if (cmd == PRC_QUENCH) {
1385 		notify = tcp_quench;
1386 	} else if (icmp_may_rst &&
1387 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1388 		    cmd == PRC_UNREACH_PORT ||
1389 		    cmd == PRC_TIMXCEED_INTRANS) &&
1390 		   ip != NULL) {
1391 		notify = tcp_drop_syn_sent;
1392 	} else if (cmd == PRC_MSGSIZE) {
1393 		struct icmp *icmp = (struct icmp *)
1394 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1395 
1396 		arg = ntohs(icmp->icmp_nextmtu);
1397 		notify = tcp_mtudisc;
1398 	} else if (PRC_IS_REDIRECT(cmd)) {
1399 		ip = NULL;
1400 		notify = in_rtchange;
1401 	} else if (cmd == PRC_HOSTDEAD) {
1402 		ip = NULL;
1403 	}
1404 
1405 	if (ip != NULL) {
1406 		crit_enter();
1407 		th = (struct tcphdr *)((caddr_t)ip +
1408 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1409 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1410 				  ip->ip_src.s_addr, th->th_sport);
1411 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1412 					ip->ip_src, th->th_sport, 0, NULL);
1413 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1414 			icmpseq = htonl(th->th_seq);
1415 			tp = intotcpcb(inp);
1416 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1417 			    SEQ_LT(icmpseq, tp->snd_max))
1418 				(*notify)(inp, arg);
1419 		} else {
1420 			struct in_conninfo inc;
1421 
1422 			inc.inc_fport = th->th_dport;
1423 			inc.inc_lport = th->th_sport;
1424 			inc.inc_faddr = faddr;
1425 			inc.inc_laddr = ip->ip_src;
1426 #ifdef INET6
1427 			inc.inc_isipv6 = 0;
1428 #endif
1429 			syncache_unreach(&inc, th);
1430 		}
1431 		crit_exit();
1432 	} else {
1433 		struct netmsg_tcp_notify nmsg;
1434 
1435 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1436 		netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0,
1437 			    tcp_notifyall_oncpu);
1438 		nmsg.nm_faddr = faddr;
1439 		nmsg.nm_arg = arg;
1440 		nmsg.nm_notify = notify;
1441 
1442 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1443 	}
1444 }
1445 
1446 #ifdef INET6
1447 void
1448 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1449 {
1450 	struct tcphdr th;
1451 	void (*notify) (struct inpcb *, int) = tcp_notify;
1452 	struct ip6_hdr *ip6;
1453 	struct mbuf *m;
1454 	struct ip6ctlparam *ip6cp = NULL;
1455 	const struct sockaddr_in6 *sa6_src = NULL;
1456 	int off;
1457 	struct tcp_portonly {
1458 		u_int16_t th_sport;
1459 		u_int16_t th_dport;
1460 	} *thp;
1461 	int arg;
1462 
1463 	if (sa->sa_family != AF_INET6 ||
1464 	    sa->sa_len != sizeof(struct sockaddr_in6))
1465 		return;
1466 
1467 	arg = 0;
1468 	if (cmd == PRC_QUENCH)
1469 		notify = tcp_quench;
1470 	else if (cmd == PRC_MSGSIZE) {
1471 		struct ip6ctlparam *ip6cp = d;
1472 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1473 
1474 		arg = ntohl(icmp6->icmp6_mtu);
1475 		notify = tcp_mtudisc;
1476 	} else if (!PRC_IS_REDIRECT(cmd) &&
1477 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1478 		return;
1479 	}
1480 
1481 	/* if the parameter is from icmp6, decode it. */
1482 	if (d != NULL) {
1483 		ip6cp = (struct ip6ctlparam *)d;
1484 		m = ip6cp->ip6c_m;
1485 		ip6 = ip6cp->ip6c_ip6;
1486 		off = ip6cp->ip6c_off;
1487 		sa6_src = ip6cp->ip6c_src;
1488 	} else {
1489 		m = NULL;
1490 		ip6 = NULL;
1491 		off = 0;	/* fool gcc */
1492 		sa6_src = &sa6_any;
1493 	}
1494 
1495 	if (ip6 != NULL) {
1496 		struct in_conninfo inc;
1497 		/*
1498 		 * XXX: We assume that when IPV6 is non NULL,
1499 		 * M and OFF are valid.
1500 		 */
1501 
1502 		/* check if we can safely examine src and dst ports */
1503 		if (m->m_pkthdr.len < off + sizeof *thp)
1504 			return;
1505 
1506 		bzero(&th, sizeof th);
1507 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1508 
1509 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1510 		    (struct sockaddr *)ip6cp->ip6c_src,
1511 		    th.th_sport, cmd, arg, notify);
1512 
1513 		inc.inc_fport = th.th_dport;
1514 		inc.inc_lport = th.th_sport;
1515 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1516 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1517 		inc.inc_isipv6 = 1;
1518 		syncache_unreach(&inc, &th);
1519 	} else
1520 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1521 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1522 }
1523 #endif
1524 
1525 /*
1526  * Following is where TCP initial sequence number generation occurs.
1527  *
1528  * There are two places where we must use initial sequence numbers:
1529  * 1.  In SYN-ACK packets.
1530  * 2.  In SYN packets.
1531  *
1532  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1533  * tcp_syncache.c for details.
1534  *
1535  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1536  * depends on this property.  In addition, these ISNs should be
1537  * unguessable so as to prevent connection hijacking.  To satisfy
1538  * the requirements of this situation, the algorithm outlined in
1539  * RFC 1948 is used to generate sequence numbers.
1540  *
1541  * Implementation details:
1542  *
1543  * Time is based off the system timer, and is corrected so that it
1544  * increases by one megabyte per second.  This allows for proper
1545  * recycling on high speed LANs while still leaving over an hour
1546  * before rollover.
1547  *
1548  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1549  * between seeding of isn_secret.  This is normally set to zero,
1550  * as reseeding should not be necessary.
1551  *
1552  */
1553 
1554 #define	ISN_BYTES_PER_SECOND 1048576
1555 
1556 u_char isn_secret[32];
1557 int isn_last_reseed;
1558 MD5_CTX isn_ctx;
1559 
1560 tcp_seq
1561 tcp_new_isn(struct tcpcb *tp)
1562 {
1563 	u_int32_t md5_buffer[4];
1564 	tcp_seq new_isn;
1565 
1566 	/* Seed if this is the first use, reseed if requested. */
1567 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1568 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1569 		< (u_int)ticks))) {
1570 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1571 		isn_last_reseed = ticks;
1572 	}
1573 
1574 	/* Compute the md5 hash and return the ISN. */
1575 	MD5Init(&isn_ctx);
1576 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1577 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1578 #ifdef INET6
1579 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1580 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1581 			  sizeof(struct in6_addr));
1582 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1583 			  sizeof(struct in6_addr));
1584 	} else
1585 #endif
1586 	{
1587 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1588 			  sizeof(struct in_addr));
1589 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1590 			  sizeof(struct in_addr));
1591 	}
1592 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1593 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1594 	new_isn = (tcp_seq) md5_buffer[0];
1595 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1596 	return (new_isn);
1597 }
1598 
1599 /*
1600  * When a source quench is received, close congestion window
1601  * to one segment.  We will gradually open it again as we proceed.
1602  */
1603 void
1604 tcp_quench(struct inpcb *inp, int error)
1605 {
1606 	struct tcpcb *tp = intotcpcb(inp);
1607 
1608 	if (tp != NULL) {
1609 		tp->snd_cwnd = tp->t_maxseg;
1610 		tp->snd_wacked = 0;
1611 	}
1612 }
1613 
1614 /*
1615  * When a specific ICMP unreachable message is received and the
1616  * connection state is SYN-SENT, drop the connection.  This behavior
1617  * is controlled by the icmp_may_rst sysctl.
1618  */
1619 void
1620 tcp_drop_syn_sent(struct inpcb *inp, int error)
1621 {
1622 	struct tcpcb *tp = intotcpcb(inp);
1623 
1624 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1625 		tcp_drop(tp, error);
1626 }
1627 
1628 /*
1629  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1630  * based on the new value in the route.  Also nudge TCP to send something,
1631  * since we know the packet we just sent was dropped.
1632  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1633  */
1634 void
1635 tcp_mtudisc(struct inpcb *inp, int mtu)
1636 {
1637 	struct tcpcb *tp = intotcpcb(inp);
1638 	struct rtentry *rt;
1639 	struct socket *so = inp->inp_socket;
1640 	int maxopd, mss;
1641 #ifdef INET6
1642 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1643 #else
1644 	const boolean_t isipv6 = FALSE;
1645 #endif
1646 
1647 	if (tp == NULL)
1648 		return;
1649 
1650 	/*
1651 	 * If no MTU is provided in the ICMP message, use the
1652 	 * next lower likely value, as specified in RFC 1191.
1653 	 */
1654 	if (mtu == 0) {
1655 		int oldmtu;
1656 
1657 		oldmtu = tp->t_maxopd +
1658 		    (isipv6 ?
1659 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1660 		     sizeof(struct tcpiphdr));
1661 		mtu = ip_next_mtu(oldmtu, 0);
1662 	}
1663 
1664 	if (isipv6)
1665 		rt = tcp_rtlookup6(&inp->inp_inc);
1666 	else
1667 		rt = tcp_rtlookup(&inp->inp_inc);
1668 	if (rt != NULL) {
1669 		struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1670 
1671 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1672 			mtu = rt->rt_rmx.rmx_mtu;
1673 
1674 		maxopd = mtu -
1675 		    (isipv6 ?
1676 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1677 		     sizeof(struct tcpiphdr));
1678 
1679 		/*
1680 		 * XXX - The following conditional probably violates the TCP
1681 		 * spec.  The problem is that, since we don't know the
1682 		 * other end's MSS, we are supposed to use a conservative
1683 		 * default.  But, if we do that, then MTU discovery will
1684 		 * never actually take place, because the conservative
1685 		 * default is much less than the MTUs typically seen
1686 		 * on the Internet today.  For the moment, we'll sweep
1687 		 * this under the carpet.
1688 		 *
1689 		 * The conservative default might not actually be a problem
1690 		 * if the only case this occurs is when sending an initial
1691 		 * SYN with options and data to a host we've never talked
1692 		 * to before.  Then, they will reply with an MSS value which
1693 		 * will get recorded and the new parameters should get
1694 		 * recomputed.  For Further Study.
1695 		 */
1696 		if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1697 			maxopd = taop->tao_mssopt;
1698 	} else
1699 		maxopd = mtu -
1700 		    (isipv6 ?
1701 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1702 		     sizeof(struct tcpiphdr));
1703 
1704 	if (tp->t_maxopd <= maxopd)
1705 		return;
1706 	tp->t_maxopd = maxopd;
1707 
1708 	mss = maxopd;
1709 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1710 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1711 		mss -= TCPOLEN_TSTAMP_APPA;
1712 
1713 	if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1714 			   (TF_REQ_CC | TF_RCVD_CC))
1715 		mss -= TCPOLEN_CC_APPA;
1716 
1717 	/* round down to multiple of MCLBYTES */
1718 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1719 	if (mss > MCLBYTES)
1720 		mss &= ~(MCLBYTES - 1);
1721 #else
1722 	if (mss > MCLBYTES)
1723 		mss = (mss / MCLBYTES) * MCLBYTES;
1724 #endif
1725 
1726 	if (so->so_snd.ssb_hiwat < mss)
1727 		mss = so->so_snd.ssb_hiwat;
1728 
1729 	tp->t_maxseg = mss;
1730 	tp->t_rtttime = 0;
1731 	tp->snd_nxt = tp->snd_una;
1732 	tcp_output(tp);
1733 	tcpstat.tcps_mturesent++;
1734 }
1735 
1736 /*
1737  * Look-up the routing entry to the peer of this inpcb.  If no route
1738  * is found and it cannot be allocated the return NULL.  This routine
1739  * is called by TCP routines that access the rmx structure and by tcp_mss
1740  * to get the interface MTU.
1741  */
1742 struct rtentry *
1743 tcp_rtlookup(struct in_conninfo *inc)
1744 {
1745 	struct route *ro = &inc->inc_route;
1746 
1747 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1748 		/* No route yet, so try to acquire one */
1749 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1750 			/*
1751 			 * unused portions of the structure MUST be zero'd
1752 			 * out because rtalloc() treats it as opaque data
1753 			 */
1754 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1755 			ro->ro_dst.sa_family = AF_INET;
1756 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1757 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1758 			    inc->inc_faddr;
1759 			rtalloc(ro);
1760 		}
1761 	}
1762 	return (ro->ro_rt);
1763 }
1764 
1765 #ifdef INET6
1766 struct rtentry *
1767 tcp_rtlookup6(struct in_conninfo *inc)
1768 {
1769 	struct route_in6 *ro6 = &inc->inc6_route;
1770 
1771 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1772 		/* No route yet, so try to acquire one */
1773 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1774 			/*
1775 			 * unused portions of the structure MUST be zero'd
1776 			 * out because rtalloc() treats it as opaque data
1777 			 */
1778 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1779 			ro6->ro_dst.sin6_family = AF_INET6;
1780 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1781 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1782 			rtalloc((struct route *)ro6);
1783 		}
1784 	}
1785 	return (ro6->ro_rt);
1786 }
1787 #endif
1788 
1789 #ifdef IPSEC
1790 /* compute ESP/AH header size for TCP, including outer IP header. */
1791 size_t
1792 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1793 {
1794 	struct inpcb *inp;
1795 	struct mbuf *m;
1796 	size_t hdrsiz;
1797 	struct ip *ip;
1798 	struct tcphdr *th;
1799 
1800 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1801 		return (0);
1802 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1803 	if (!m)
1804 		return (0);
1805 
1806 #ifdef INET6
1807 	if (inp->inp_vflag & INP_IPV6) {
1808 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1809 
1810 		th = (struct tcphdr *)(ip6 + 1);
1811 		m->m_pkthdr.len = m->m_len =
1812 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1813 		tcp_fillheaders(tp, ip6, th);
1814 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1815 	} else
1816 #endif
1817 	{
1818 		ip = mtod(m, struct ip *);
1819 		th = (struct tcphdr *)(ip + 1);
1820 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1821 		tcp_fillheaders(tp, ip, th);
1822 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1823 	}
1824 
1825 	m_free(m);
1826 	return (hdrsiz);
1827 }
1828 #endif
1829 
1830 /*
1831  * Return a pointer to the cached information about the remote host.
1832  * The cached information is stored in the protocol specific part of
1833  * the route metrics.
1834  */
1835 struct rmxp_tao *
1836 tcp_gettaocache(struct in_conninfo *inc)
1837 {
1838 	struct rtentry *rt;
1839 
1840 #ifdef INET6
1841 	if (inc->inc_isipv6)
1842 		rt = tcp_rtlookup6(inc);
1843 	else
1844 #endif
1845 		rt = tcp_rtlookup(inc);
1846 
1847 	/* Make sure this is a host route and is up. */
1848 	if (rt == NULL ||
1849 	    (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1850 		return (NULL);
1851 
1852 	return (rmx_taop(rt->rt_rmx));
1853 }
1854 
1855 /*
1856  * Clear all the TAO cache entries, called from tcp_init.
1857  *
1858  * XXX
1859  * This routine is just an empty one, because we assume that the routing
1860  * routing tables are initialized at the same time when TCP, so there is
1861  * nothing in the cache left over.
1862  */
1863 static void
1864 tcp_cleartaocache(void)
1865 {
1866 }
1867 
1868 /*
1869  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1870  *
1871  * This code attempts to calculate the bandwidth-delay product as a
1872  * means of determining the optimal window size to maximize bandwidth,
1873  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1874  * routers.  This code also does a fairly good job keeping RTTs in check
1875  * across slow links like modems.  We implement an algorithm which is very
1876  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1877  * transmitter side of a TCP connection and so only effects the transmit
1878  * side of the connection.
1879  *
1880  * BACKGROUND:  TCP makes no provision for the management of buffer space
1881  * at the end points or at the intermediate routers and switches.  A TCP
1882  * stream, whether using NewReno or not, will eventually buffer as
1883  * many packets as it is able and the only reason this typically works is
1884  * due to the fairly small default buffers made available for a connection
1885  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1886  * scaling it is now fairly easy for even a single TCP connection to blow-out
1887  * all available buffer space not only on the local interface, but on
1888  * intermediate routers and switches as well.  NewReno makes a misguided
1889  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1890  * then backing off, then steadily increasing the window again until another
1891  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1892  * is only made worse as network loads increase and the idea of intentionally
1893  * blowing out network buffers is, frankly, a terrible way to manage network
1894  * resources.
1895  *
1896  * It is far better to limit the transmit window prior to the failure
1897  * condition being achieved.  There are two general ways to do this:  First
1898  * you can 'scan' through different transmit window sizes and locate the
1899  * point where the RTT stops increasing, indicating that you have filled the
1900  * pipe, then scan backwards until you note that RTT stops decreasing, then
1901  * repeat ad-infinitum.  This method works in principle but has severe
1902  * implementation issues due to RTT variances, timer granularity, and
1903  * instability in the algorithm which can lead to many false positives and
1904  * create oscillations as well as interact badly with other TCP streams
1905  * implementing the same algorithm.
1906  *
1907  * The second method is to limit the window to the bandwidth delay product
1908  * of the link.  This is the method we implement.  RTT variances and our
1909  * own manipulation of the congestion window, bwnd, can potentially
1910  * destabilize the algorithm.  For this reason we have to stabilize the
1911  * elements used to calculate the window.  We do this by using the minimum
1912  * observed RTT, the long term average of the observed bandwidth, and
1913  * by adding two segments worth of slop.  It isn't perfect but it is able
1914  * to react to changing conditions and gives us a very stable basis on
1915  * which to extend the algorithm.
1916  */
1917 void
1918 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1919 {
1920 	u_long bw;
1921 	u_long bwnd;
1922 	int save_ticks;
1923 	int delta_ticks;
1924 
1925 	/*
1926 	 * If inflight_enable is disabled in the middle of a tcp connection,
1927 	 * make sure snd_bwnd is effectively disabled.
1928 	 */
1929 	if (!tcp_inflight_enable) {
1930 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1931 		tp->snd_bandwidth = 0;
1932 		return;
1933 	}
1934 
1935 	/*
1936 	 * Validate the delta time.  If a connection is new or has been idle
1937 	 * a long time we have to reset the bandwidth calculator.
1938 	 */
1939 	save_ticks = ticks;
1940 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1941 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1942 		tp->t_bw_rtttime = ticks;
1943 		tp->t_bw_rtseq = ack_seq;
1944 		if (tp->snd_bandwidth == 0)
1945 			tp->snd_bandwidth = tcp_inflight_min;
1946 		return;
1947 	}
1948 	if (delta_ticks == 0)
1949 		return;
1950 
1951 	/*
1952 	 * Sanity check, plus ignore pure window update acks.
1953 	 */
1954 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1955 		return;
1956 
1957 	/*
1958 	 * Figure out the bandwidth.  Due to the tick granularity this
1959 	 * is a very rough number and it MUST be averaged over a fairly
1960 	 * long period of time.  XXX we need to take into account a link
1961 	 * that is not using all available bandwidth, but for now our
1962 	 * slop will ramp us up if this case occurs and the bandwidth later
1963 	 * increases.
1964 	 */
1965 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1966 	tp->t_bw_rtttime = save_ticks;
1967 	tp->t_bw_rtseq = ack_seq;
1968 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1969 
1970 	tp->snd_bandwidth = bw;
1971 
1972 	/*
1973 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1974 	 * segments.  The additional slop puts us squarely in the sweet
1975 	 * spot and also handles the bandwidth run-up case.  Without the
1976 	 * slop we could be locking ourselves into a lower bandwidth.
1977 	 *
1978 	 * Situations Handled:
1979 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1980 	 *	    high speed LANs, allowing larger TCP buffers to be
1981 	 *	    specified, and also does a good job preventing
1982 	 *	    over-queueing of packets over choke points like modems
1983 	 *	    (at least for the transmit side).
1984 	 *
1985 	 *	(2) Is able to handle changing network loads (bandwidth
1986 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1987 	 *	    increases).
1988 	 *
1989 	 *	(3) Theoretically should stabilize in the face of multiple
1990 	 *	    connections implementing the same algorithm (this may need
1991 	 *	    a little work).
1992 	 *
1993 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1994 	 *	    be adjusted with a sysctl but typically only needs to be on
1995 	 *	    very slow connections.  A value no smaller then 5 should
1996 	 *	    be used, but only reduce this default if you have no other
1997 	 *	    choice.
1998 	 */
1999 
2000 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
2001 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
2002 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
2003 #undef USERTT
2004 
2005 	if (tcp_inflight_debug > 0) {
2006 		static int ltime;
2007 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2008 			ltime = ticks;
2009 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2010 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
2011 		}
2012 	}
2013 	if ((long)bwnd < tcp_inflight_min)
2014 		bwnd = tcp_inflight_min;
2015 	if (bwnd > tcp_inflight_max)
2016 		bwnd = tcp_inflight_max;
2017 	if ((long)bwnd < tp->t_maxseg * 2)
2018 		bwnd = tp->t_maxseg * 2;
2019 	tp->snd_bwnd = bwnd;
2020 }
2021