xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 895c1f850848cf50a3243e134345f0d4ff33ac59)
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/priv.h>
89 #include <sys/socket.h>
90 #include <sys/socketvar.h>
91 #include <sys/protosw.h>
92 #include <sys/random.h>
93 #include <sys/in_cksum.h>
94 #include <sys/ktr.h>
95 
96 #include <vm/vm_zone.h>
97 
98 #include <net/route.h>
99 #include <net/if.h>
100 #include <net/netisr.h>
101 
102 #define	_IP_VHL
103 #include <netinet/in.h>
104 #include <netinet/in_systm.h>
105 #include <netinet/ip.h>
106 #include <netinet/ip6.h>
107 #include <netinet/in_pcb.h>
108 #include <netinet6/in6_pcb.h>
109 #include <netinet/in_var.h>
110 #include <netinet/ip_var.h>
111 #include <netinet6/ip6_var.h>
112 #include <netinet/ip_icmp.h>
113 #ifdef INET6
114 #include <netinet/icmp6.h>
115 #endif
116 #include <netinet/tcp.h>
117 #include <netinet/tcp_fsm.h>
118 #include <netinet/tcp_seq.h>
119 #include <netinet/tcp_timer.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	callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
298 	struct	callout inp_tp_delack;
299 	struct	netmsg_tcp_timer inp_tp_timermsg;
300 };
301 #undef ALIGNMENT
302 #undef ALIGNM1
303 
304 /*
305  * Tcp initialization
306  */
307 void
308 tcp_init(void)
309 {
310 	struct inpcbporthead *porthashbase;
311 	u_long porthashmask;
312 	struct vm_zone *ipi_zone;
313 	int hashsize = TCBHASHSIZE;
314 	int cpu;
315 
316 	/*
317 	 * note: tcptemp is used for keepalives, and it is ok for an
318 	 * allocation to fail so do not specify MPF_INT.
319 	 */
320 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
321 		    25, -1, 0, NULL);
322 
323 	tcp_ccgen = 1;
324 	tcp_cleartaocache();
325 
326 	tcp_delacktime = TCPTV_DELACK;
327 	tcp_keepinit = TCPTV_KEEP_INIT;
328 	tcp_keepidle = TCPTV_KEEP_IDLE;
329 	tcp_keepintvl = TCPTV_KEEPINTVL;
330 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
331 	tcp_msl = TCPTV_MSL;
332 	tcp_rexmit_min = TCPTV_MIN;
333 	tcp_rexmit_slop = TCPTV_CPU_VAR;
334 
335 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
336 	if (!powerof2(hashsize)) {
337 		kprintf("WARNING: TCB hash size not a power of 2\n");
338 		hashsize = 512; /* safe default */
339 	}
340 	tcp_tcbhashsize = hashsize;
341 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
342 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
343 			 ZONE_INTERRUPT, 0);
344 
345 	for (cpu = 0; cpu < ncpus2; cpu++) {
346 		in_pcbinfo_init(&tcbinfo[cpu]);
347 		tcbinfo[cpu].cpu = cpu;
348 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
349 		    &tcbinfo[cpu].hashmask);
350 		tcbinfo[cpu].porthashbase = porthashbase;
351 		tcbinfo[cpu].porthashmask = porthashmask;
352 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
353 		    &tcbinfo[cpu].wildcardhashmask);
354 		tcbinfo[cpu].ipi_zone = ipi_zone;
355 		TAILQ_INIT(&tcpcbackq[cpu]);
356 	}
357 
358 	tcp_reass_maxseg = nmbclusters / 16;
359 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
360 
361 #ifdef INET6
362 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
363 #else
364 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
365 #endif
366 	if (max_protohdr < TCP_MINPROTOHDR)
367 		max_protohdr = TCP_MINPROTOHDR;
368 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
369 		panic("tcp_init");
370 #undef TCP_MINPROTOHDR
371 
372 	/*
373 	 * Initialize TCP statistics counters for each CPU.
374 	 */
375 #ifdef SMP
376 	for (cpu = 0; cpu < ncpus; ++cpu) {
377 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
378 	}
379 #else
380 	bzero(&tcpstat, sizeof(struct tcp_stats));
381 #endif
382 
383 	syncache_init();
384 	tcp_thread_init();
385 }
386 
387 void
388 tcpmsg_service_loop(void *dummy)
389 {
390 	struct netmsg *msg;
391 	int mplocked;
392 
393 	/*
394 	 * Thread was started with TDF_MPSAFE
395 	 */
396 	mplocked = 0;
397 
398 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
399 		do {
400 			logtcp(rxmsg);
401 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
402 						  mplocked);
403 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
404 
405 		logtcp(delayed);
406 		tcp_willblock(mplocked);
407 		logtcp(wait);
408 	}
409 }
410 
411 static void
412 tcp_willblock(int mplocked)
413 {
414 	struct tcpcb *tp;
415 	int cpu = mycpu->gd_cpuid;
416 	int unlock = 0;
417 
418 	if (!mplocked && !tcp_mpsafe_proto) {
419 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
420 			return;
421 
422 		get_mplock();
423 		mplocked = 1;
424 		unlock = 1;
425 	}
426 
427 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
428 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
429 		tp->t_flags &= ~TF_ONOUTPUTQ;
430 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
431 		tcp_output(tp);
432 	}
433 
434 	if (unlock)
435 		rel_mplock();
436 }
437 
438 
439 /*
440  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
441  * tcp_template used to store this data in mbufs, but we now recopy it out
442  * of the tcpcb each time to conserve mbufs.
443  */
444 void
445 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
446 {
447 	struct inpcb *inp = tp->t_inpcb;
448 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
449 
450 #ifdef INET6
451 	if (inp->inp_vflag & INP_IPV6) {
452 		struct ip6_hdr *ip6;
453 
454 		ip6 = (struct ip6_hdr *)ip_ptr;
455 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
456 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
457 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
458 			(IPV6_VERSION & IPV6_VERSION_MASK);
459 		ip6->ip6_nxt = IPPROTO_TCP;
460 		ip6->ip6_plen = sizeof(struct tcphdr);
461 		ip6->ip6_src = inp->in6p_laddr;
462 		ip6->ip6_dst = inp->in6p_faddr;
463 		tcp_hdr->th_sum = 0;
464 	} else
465 #endif
466 	{
467 		struct ip *ip = (struct ip *) ip_ptr;
468 
469 		ip->ip_vhl = IP_VHL_BORING;
470 		ip->ip_tos = 0;
471 		ip->ip_len = 0;
472 		ip->ip_id = 0;
473 		ip->ip_off = 0;
474 		ip->ip_ttl = 0;
475 		ip->ip_sum = 0;
476 		ip->ip_p = IPPROTO_TCP;
477 		ip->ip_src = inp->inp_laddr;
478 		ip->ip_dst = inp->inp_faddr;
479 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
480 				    ip->ip_dst.s_addr,
481 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
482 	}
483 
484 	tcp_hdr->th_sport = inp->inp_lport;
485 	tcp_hdr->th_dport = inp->inp_fport;
486 	tcp_hdr->th_seq = 0;
487 	tcp_hdr->th_ack = 0;
488 	tcp_hdr->th_x2 = 0;
489 	tcp_hdr->th_off = 5;
490 	tcp_hdr->th_flags = 0;
491 	tcp_hdr->th_win = 0;
492 	tcp_hdr->th_urp = 0;
493 }
494 
495 /*
496  * Create template to be used to send tcp packets on a connection.
497  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
498  * use for this function is in keepalives, which use tcp_respond.
499  */
500 struct tcptemp *
501 tcp_maketemplate(struct tcpcb *tp)
502 {
503 	struct tcptemp *tmp;
504 
505 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
506 		return (NULL);
507 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
508 	return (tmp);
509 }
510 
511 void
512 tcp_freetemplate(struct tcptemp *tmp)
513 {
514 	mpipe_free(&tcptemp_mpipe, tmp);
515 }
516 
517 /*
518  * Send a single message to the TCP at address specified by
519  * the given TCP/IP header.  If m == NULL, then we make a copy
520  * of the tcpiphdr at ti and send directly to the addressed host.
521  * This is used to force keep alive messages out using the TCP
522  * template for a connection.  If flags are given then we send
523  * a message back to the TCP which originated the * segment ti,
524  * and discard the mbuf containing it and any other attached mbufs.
525  *
526  * In any case the ack and sequence number of the transmitted
527  * segment are as specified by the parameters.
528  *
529  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
530  */
531 void
532 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
533 	    tcp_seq ack, tcp_seq seq, int flags)
534 {
535 	int tlen;
536 	int win = 0;
537 	struct route *ro = NULL;
538 	struct route sro;
539 	struct ip *ip = ipgen;
540 	struct tcphdr *nth;
541 	int ipflags = 0;
542 	struct route_in6 *ro6 = NULL;
543 	struct route_in6 sro6;
544 	struct ip6_hdr *ip6 = ipgen;
545 	boolean_t use_tmpro = TRUE;
546 #ifdef INET6
547 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
548 #else
549 	const boolean_t isipv6 = FALSE;
550 #endif
551 
552 	if (tp != NULL) {
553 		if (!(flags & TH_RST)) {
554 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
555 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
556 				win = (long)TCP_MAXWIN << tp->rcv_scale;
557 		}
558 		/*
559 		 * Don't use the route cache of a listen socket,
560 		 * it is not MPSAFE; use temporary route cache.
561 		 */
562 		if (tp->t_state != TCPS_LISTEN) {
563 			if (isipv6)
564 				ro6 = &tp->t_inpcb->in6p_route;
565 			else
566 				ro = &tp->t_inpcb->inp_route;
567 			use_tmpro = FALSE;
568 		}
569 	}
570 	if (use_tmpro) {
571 		if (isipv6) {
572 			ro6 = &sro6;
573 			bzero(ro6, sizeof *ro6);
574 		} else {
575 			ro = &sro;
576 			bzero(ro, sizeof *ro);
577 		}
578 	}
579 	if (m == NULL) {
580 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
581 		if (m == NULL)
582 			return;
583 		tlen = 0;
584 		m->m_data += max_linkhdr;
585 		if (isipv6) {
586 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
587 			ip6 = mtod(m, struct ip6_hdr *);
588 			nth = (struct tcphdr *)(ip6 + 1);
589 		} else {
590 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
591 			ip = mtod(m, struct ip *);
592 			nth = (struct tcphdr *)(ip + 1);
593 		}
594 		bcopy(th, nth, sizeof(struct tcphdr));
595 		flags = TH_ACK;
596 	} else {
597 		m_freem(m->m_next);
598 		m->m_next = NULL;
599 		m->m_data = (caddr_t)ipgen;
600 		/* m_len is set later */
601 		tlen = 0;
602 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
603 		if (isipv6) {
604 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
605 			nth = (struct tcphdr *)(ip6 + 1);
606 		} else {
607 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
608 			nth = (struct tcphdr *)(ip + 1);
609 		}
610 		if (th != nth) {
611 			/*
612 			 * this is usually a case when an extension header
613 			 * exists between the IPv6 header and the
614 			 * TCP header.
615 			 */
616 			nth->th_sport = th->th_sport;
617 			nth->th_dport = th->th_dport;
618 		}
619 		xchg(nth->th_dport, nth->th_sport, n_short);
620 #undef xchg
621 	}
622 	if (isipv6) {
623 		ip6->ip6_flow = 0;
624 		ip6->ip6_vfc = IPV6_VERSION;
625 		ip6->ip6_nxt = IPPROTO_TCP;
626 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
627 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
628 	} else {
629 		tlen += sizeof(struct tcpiphdr);
630 		ip->ip_len = tlen;
631 		ip->ip_ttl = ip_defttl;
632 	}
633 	m->m_len = tlen;
634 	m->m_pkthdr.len = tlen;
635 	m->m_pkthdr.rcvif = (struct ifnet *) NULL;
636 	nth->th_seq = htonl(seq);
637 	nth->th_ack = htonl(ack);
638 	nth->th_x2 = 0;
639 	nth->th_off = sizeof(struct tcphdr) >> 2;
640 	nth->th_flags = flags;
641 	if (tp != NULL)
642 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
643 	else
644 		nth->th_win = htons((u_short)win);
645 	nth->th_urp = 0;
646 	if (isipv6) {
647 		nth->th_sum = 0;
648 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
649 					sizeof(struct ip6_hdr),
650 					tlen - sizeof(struct ip6_hdr));
651 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
652 					       (ro6 && ro6->ro_rt) ?
653 						ro6->ro_rt->rt_ifp : NULL);
654 	} else {
655 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
656 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
657 		m->m_pkthdr.csum_flags = CSUM_TCP;
658 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
659 	}
660 #ifdef TCPDEBUG
661 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
662 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
663 #endif
664 	if (isipv6) {
665 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
666 			   tp ? tp->t_inpcb : NULL);
667 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
668 			RTFREE(ro6->ro_rt);
669 			ro6->ro_rt = NULL;
670 		}
671 	} else {
672 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
673 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
674 			RTFREE(ro->ro_rt);
675 			ro->ro_rt = NULL;
676 		}
677 	}
678 }
679 
680 /*
681  * Create a new TCP control block, making an
682  * empty reassembly queue and hooking it to the argument
683  * protocol control block.  The `inp' parameter must have
684  * come from the zone allocator set up in tcp_init().
685  */
686 struct tcpcb *
687 tcp_newtcpcb(struct inpcb *inp)
688 {
689 	struct inp_tp *it;
690 	struct tcpcb *tp;
691 #ifdef INET6
692 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
693 #else
694 	const boolean_t isipv6 = FALSE;
695 #endif
696 
697 	it = (struct inp_tp *)inp;
698 	tp = &it->tcb;
699 	bzero(tp, sizeof(struct tcpcb));
700 	LIST_INIT(&tp->t_segq);
701 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
702 
703 	/* Set up our timeouts. */
704 	callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
705 	callout_init(tp->tt_persist = &it->inp_tp_persist);
706 	callout_init(tp->tt_keep = &it->inp_tp_keep);
707 	callout_init(tp->tt_2msl = &it->inp_tp_2msl);
708 	callout_init(tp->tt_delack = &it->inp_tp_delack);
709 
710 	tp->tt_msg = &it->inp_tp_timermsg;
711 	if (isipv6) {
712 		/* Don't mess with IPv6; always create timer message */
713 		tcp_create_timermsg(tp);
714 	} else {
715 		/*
716 		 * Zero out timer message.  We don't create it here,
717 		 * since the current CPU may not be the owner of this
718 		 * inpcb.
719 		 */
720 		bzero(tp->tt_msg, sizeof(*tp->tt_msg));
721 	}
722 
723 	if (tcp_do_rfc1323)
724 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
725 	if (tcp_do_rfc1644)
726 		tp->t_flags |= TF_REQ_CC;
727 	tp->t_inpcb = inp;	/* XXX */
728 	tp->t_state = TCPS_CLOSED;
729 	/*
730 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
731 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
732 	 * reasonable initial retransmit time.
733 	 */
734 	tp->t_srtt = TCPTV_SRTTBASE;
735 	tp->t_rttvar =
736 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
737 	tp->t_rttmin = tcp_rexmit_min;
738 	tp->t_rxtcur = TCPTV_RTOBASE;
739 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
740 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
741 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
742 	tp->t_rcvtime = ticks;
743 	/*
744 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
745 	 * because the socket may be bound to an IPv6 wildcard address,
746 	 * which may match an IPv4-mapped IPv6 address.
747 	 */
748 	inp->inp_ip_ttl = ip_defttl;
749 	inp->inp_ppcb = tp;
750 	tcp_sack_tcpcb_init(tp);
751 	return (tp);		/* XXX */
752 }
753 
754 /*
755  * Drop a TCP connection, reporting the specified error.
756  * If connection is synchronized, then send a RST to peer.
757  */
758 struct tcpcb *
759 tcp_drop(struct tcpcb *tp, int error)
760 {
761 	struct socket *so = tp->t_inpcb->inp_socket;
762 
763 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
764 		tp->t_state = TCPS_CLOSED;
765 		tcp_output(tp);
766 		tcpstat.tcps_drops++;
767 	} else
768 		tcpstat.tcps_conndrops++;
769 	if (error == ETIMEDOUT && tp->t_softerror)
770 		error = tp->t_softerror;
771 	so->so_error = error;
772 	return (tcp_close(tp));
773 }
774 
775 #ifdef SMP
776 
777 struct netmsg_remwildcard {
778 	struct netmsg		nm_netmsg;
779 	struct inpcb		*nm_inp;
780 	struct inpcbinfo	*nm_pcbinfo;
781 #if defined(INET6)
782 	int			nm_isinet6;
783 #else
784 	int			nm_unused01;
785 #endif
786 };
787 
788 /*
789  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
790  * inp can be detached.  We do this by cycling through the cpus, ending up
791  * on the cpu controlling the inp last and then doing the disconnect.
792  */
793 static void
794 in_pcbremwildcardhash_handler(struct netmsg *msg0)
795 {
796 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
797 	int cpu;
798 
799 	cpu = msg->nm_pcbinfo->cpu;
800 
801 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
802 		/* note: detach removes any wildcard hash entry */
803 #ifdef INET6
804 		if (msg->nm_isinet6)
805 			in6_pcbdetach(msg->nm_inp);
806 		else
807 #endif
808 			in_pcbdetach(msg->nm_inp);
809 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
810 	} else {
811 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
812 		cpu = (cpu + 1) % ncpus2;
813 		msg->nm_pcbinfo = &tcbinfo[cpu];
814 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
815 	}
816 }
817 
818 #endif
819 
820 /*
821  * Close a TCP control block:
822  *	discard all space held by the tcp
823  *	discard internet protocol block
824  *	wake up any sleepers
825  */
826 struct tcpcb *
827 tcp_close(struct tcpcb *tp)
828 {
829 	struct tseg_qent *q;
830 	struct inpcb *inp = tp->t_inpcb;
831 	struct socket *so = inp->inp_socket;
832 	struct rtentry *rt;
833 	boolean_t dosavessthresh;
834 #ifdef SMP
835 	int cpu;
836 #endif
837 #ifdef INET6
838 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
839 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
840 #else
841 	const boolean_t isipv6 = FALSE;
842 #endif
843 
844 	/*
845 	 * The tp is not instantly destroyed in the wildcard case.  Setting
846 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
847 	 * messing with it, though it should be noted that this change may
848 	 * not take effect on other cpus until we have chained the wildcard
849 	 * hash removal.
850 	 *
851 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
852 	 * update and prevent other tcp protocol threads from accepting new
853 	 * connections on the listen socket we might be trying to close down.
854 	 */
855 	KKASSERT(tp->t_state != TCPS_TERMINATING);
856 	tp->t_state = TCPS_TERMINATING;
857 
858 	/*
859 	 * Make sure that all of our timers are stopped before we
860 	 * delete the PCB.
861 	 */
862 	callout_stop(tp->tt_rexmt);
863 	callout_stop(tp->tt_persist);
864 	callout_stop(tp->tt_keep);
865 	callout_stop(tp->tt_2msl);
866 	callout_stop(tp->tt_delack);
867 
868 	if (tp->t_flags & TF_ONOUTPUTQ) {
869 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
870 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
871 		tp->t_flags &= ~TF_ONOUTPUTQ;
872 	}
873 
874 	/*
875 	 * If we got enough samples through the srtt filter,
876 	 * save the rtt and rttvar in the routing entry.
877 	 * 'Enough' is arbitrarily defined as the 16 samples.
878 	 * 16 samples is enough for the srtt filter to converge
879 	 * to within 5% of the correct value; fewer samples and
880 	 * we could save a very bogus rtt.
881 	 *
882 	 * Don't update the default route's characteristics and don't
883 	 * update anything that the user "locked".
884 	 */
885 	if (tp->t_rttupdated >= 16) {
886 		u_long i = 0;
887 
888 		if (isipv6) {
889 			struct sockaddr_in6 *sin6;
890 
891 			if ((rt = inp->in6p_route.ro_rt) == NULL)
892 				goto no_valid_rt;
893 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
894 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
895 				goto no_valid_rt;
896 		} else
897 			if ((rt = inp->inp_route.ro_rt) == NULL ||
898 			    ((struct sockaddr_in *)rt_key(rt))->
899 			     sin_addr.s_addr == INADDR_ANY)
900 				goto no_valid_rt;
901 
902 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
903 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
904 			if (rt->rt_rmx.rmx_rtt && i)
905 				/*
906 				 * filter this update to half the old & half
907 				 * the new values, converting scale.
908 				 * See route.h and tcp_var.h for a
909 				 * description of the scaling constants.
910 				 */
911 				rt->rt_rmx.rmx_rtt =
912 				    (rt->rt_rmx.rmx_rtt + i) / 2;
913 			else
914 				rt->rt_rmx.rmx_rtt = i;
915 			tcpstat.tcps_cachedrtt++;
916 		}
917 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
918 			i = tp->t_rttvar *
919 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
920 			if (rt->rt_rmx.rmx_rttvar && i)
921 				rt->rt_rmx.rmx_rttvar =
922 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
923 			else
924 				rt->rt_rmx.rmx_rttvar = i;
925 			tcpstat.tcps_cachedrttvar++;
926 		}
927 		/*
928 		 * The old comment here said:
929 		 * update the pipelimit (ssthresh) if it has been updated
930 		 * already or if a pipesize was specified & the threshhold
931 		 * got below half the pipesize.  I.e., wait for bad news
932 		 * before we start updating, then update on both good
933 		 * and bad news.
934 		 *
935 		 * But we want to save the ssthresh even if no pipesize is
936 		 * specified explicitly in the route, because such
937 		 * connections still have an implicit pipesize specified
938 		 * by the global tcp_sendspace.  In the absence of a reliable
939 		 * way to calculate the pipesize, it will have to do.
940 		 */
941 		i = tp->snd_ssthresh;
942 		if (rt->rt_rmx.rmx_sendpipe != 0)
943 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
944 		else
945 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
946 		if (dosavessthresh ||
947 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
948 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
949 			/*
950 			 * convert the limit from user data bytes to
951 			 * packets then to packet data bytes.
952 			 */
953 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
954 			if (i < 2)
955 				i = 2;
956 			i *= tp->t_maxseg +
957 			     (isipv6 ?
958 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
959 			      sizeof(struct tcpiphdr));
960 			if (rt->rt_rmx.rmx_ssthresh)
961 				rt->rt_rmx.rmx_ssthresh =
962 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
963 			else
964 				rt->rt_rmx.rmx_ssthresh = i;
965 			tcpstat.tcps_cachedssthresh++;
966 		}
967 	}
968 
969 no_valid_rt:
970 	/* free the reassembly queue, if any */
971 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
972 		LIST_REMOVE(q, tqe_q);
973 		m_freem(q->tqe_m);
974 		FREE(q, M_TSEGQ);
975 		tcp_reass_qsize--;
976 	}
977 	/* throw away SACK blocks in scoreboard*/
978 	if (TCP_DO_SACK(tp))
979 		tcp_sack_cleanup(&tp->scb);
980 
981 	inp->inp_ppcb = NULL;
982 	soisdisconnected(so);
983 
984 	tcp_destroy_timermsg(tp);
985 
986 	/*
987 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
988 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
989 	 * for each protocol thread and must be removed in the context of
990 	 * that thread.  This is accomplished by chaining the message
991 	 * through the cpus.
992 	 *
993 	 * If the inp is not wildcarded we simply detach, which will remove
994 	 * the any hashes still present for this inp.
995 	 */
996 #ifdef SMP
997 	if (inp->inp_flags & INP_WILDCARD_MP) {
998 		struct netmsg_remwildcard *msg;
999 
1000 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1001 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
1002 			      M_LWKTMSG, M_INTWAIT);
1003 		netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1004 			    in_pcbremwildcardhash_handler);
1005 #ifdef INET6
1006 		msg->nm_isinet6 = isafinet6;
1007 #endif
1008 		msg->nm_inp = inp;
1009 		msg->nm_pcbinfo = &tcbinfo[cpu];
1010 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1011 	} else
1012 #endif
1013 	{
1014 		/* note: detach removes any wildcard hash entry */
1015 #ifdef INET6
1016 		if (isafinet6)
1017 			in6_pcbdetach(inp);
1018 		else
1019 #endif
1020 			in_pcbdetach(inp);
1021 	}
1022 	tcpstat.tcps_closed++;
1023 	return (NULL);
1024 }
1025 
1026 static __inline void
1027 tcp_drain_oncpu(struct inpcbhead *head)
1028 {
1029 	struct inpcb *inpb;
1030 	struct tcpcb *tcpb;
1031 	struct tseg_qent *te;
1032 
1033 	LIST_FOREACH(inpb, head, inp_list) {
1034 		if (inpb->inp_flags & INP_PLACEMARKER)
1035 			continue;
1036 		if ((tcpb = intotcpcb(inpb))) {
1037 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1038 				LIST_REMOVE(te, tqe_q);
1039 				m_freem(te->tqe_m);
1040 				FREE(te, M_TSEGQ);
1041 				tcp_reass_qsize--;
1042 			}
1043 		}
1044 	}
1045 }
1046 
1047 #ifdef SMP
1048 struct netmsg_tcp_drain {
1049 	struct netmsg		nm_netmsg;
1050 	struct inpcbhead	*nm_head;
1051 };
1052 
1053 static void
1054 tcp_drain_handler(netmsg_t netmsg)
1055 {
1056 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1057 
1058 	tcp_drain_oncpu(nm->nm_head);
1059 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1060 }
1061 #endif
1062 
1063 void
1064 tcp_drain(void)
1065 {
1066 #ifdef SMP
1067 	int cpu;
1068 #endif
1069 
1070 	if (!do_tcpdrain)
1071 		return;
1072 
1073 	/*
1074 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1075 	 * if there is one...
1076 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1077 	 *	reassembly queue should be flushed, but in a situation
1078 	 *	where we're really low on mbufs, this is potentially
1079 	 *	useful.
1080 	 */
1081 #ifdef SMP
1082 	for (cpu = 0; cpu < ncpus2; cpu++) {
1083 		struct netmsg_tcp_drain *msg;
1084 
1085 		if (cpu == mycpu->gd_cpuid) {
1086 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1087 		} else {
1088 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1089 				    M_LWKTMSG, M_NOWAIT);
1090 			if (msg == NULL)
1091 				continue;
1092 			netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1093 				    tcp_drain_handler);
1094 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1095 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1096 		}
1097 	}
1098 #else
1099 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1100 #endif
1101 }
1102 
1103 /*
1104  * Notify a tcp user of an asynchronous error;
1105  * store error as soft error, but wake up user
1106  * (for now, won't do anything until can select for soft error).
1107  *
1108  * Do not wake up user since there currently is no mechanism for
1109  * reporting soft errors (yet - a kqueue filter may be added).
1110  */
1111 static void
1112 tcp_notify(struct inpcb *inp, int error)
1113 {
1114 	struct tcpcb *tp = intotcpcb(inp);
1115 
1116 	/*
1117 	 * Ignore some errors if we are hooked up.
1118 	 * If connection hasn't completed, has retransmitted several times,
1119 	 * and receives a second error, give up now.  This is better
1120 	 * than waiting a long time to establish a connection that
1121 	 * can never complete.
1122 	 */
1123 	if (tp->t_state == TCPS_ESTABLISHED &&
1124 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1125 	      error == EHOSTDOWN)) {
1126 		return;
1127 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1128 	    tp->t_softerror)
1129 		tcp_drop(tp, error);
1130 	else
1131 		tp->t_softerror = error;
1132 #if 0
1133 	wakeup(&so->so_timeo);
1134 	sorwakeup(so);
1135 	sowwakeup(so);
1136 #endif
1137 }
1138 
1139 static int
1140 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1141 {
1142 	int error, i, n;
1143 	struct inpcb *marker;
1144 	struct inpcb *inp;
1145 	inp_gen_t gencnt;
1146 	globaldata_t gd;
1147 	int origcpu, ccpu;
1148 
1149 	error = 0;
1150 	n = 0;
1151 
1152 	/*
1153 	 * The process of preparing the TCB list is too time-consuming and
1154 	 * resource-intensive to repeat twice on every request.
1155 	 */
1156 	if (req->oldptr == NULL) {
1157 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1158 			gd = globaldata_find(ccpu);
1159 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1160 		}
1161 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1162 		return (0);
1163 	}
1164 
1165 	if (req->newptr != NULL)
1166 		return (EPERM);
1167 
1168 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1169 	marker->inp_flags |= INP_PLACEMARKER;
1170 
1171 	/*
1172 	 * OK, now we're committed to doing something.  Run the inpcb list
1173 	 * for each cpu in the system and construct the output.  Use a
1174 	 * list placemarker to deal with list changes occuring during
1175 	 * copyout blockages (but otherwise depend on being on the correct
1176 	 * cpu to avoid races).
1177 	 */
1178 	origcpu = mycpu->gd_cpuid;
1179 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1180 		globaldata_t rgd;
1181 		caddr_t inp_ppcb;
1182 		struct xtcpcb xt;
1183 		int cpu_id;
1184 
1185 		cpu_id = (origcpu + ccpu) % ncpus;
1186 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1187 			continue;
1188 		rgd = globaldata_find(cpu_id);
1189 		lwkt_setcpu_self(rgd);
1190 
1191 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1192 		n = tcbinfo[cpu_id].ipi_count;
1193 
1194 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1195 		i = 0;
1196 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1197 			/*
1198 			 * process a snapshot of pcbs, ignoring placemarkers
1199 			 * and using our own to allow SYSCTL_OUT to block.
1200 			 */
1201 			LIST_REMOVE(marker, inp_list);
1202 			LIST_INSERT_AFTER(inp, marker, inp_list);
1203 
1204 			if (inp->inp_flags & INP_PLACEMARKER)
1205 				continue;
1206 			if (inp->inp_gencnt > gencnt)
1207 				continue;
1208 			if (prison_xinpcb(req->td, inp))
1209 				continue;
1210 
1211 			xt.xt_len = sizeof xt;
1212 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1213 			inp_ppcb = inp->inp_ppcb;
1214 			if (inp_ppcb != NULL)
1215 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1216 			else
1217 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1218 			if (inp->inp_socket)
1219 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1220 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1221 				break;
1222 			++i;
1223 		}
1224 		LIST_REMOVE(marker, inp_list);
1225 		if (error == 0 && i < n) {
1226 			bzero(&xt, sizeof xt);
1227 			xt.xt_len = sizeof xt;
1228 			while (i < n) {
1229 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1230 				if (error)
1231 					break;
1232 				++i;
1233 			}
1234 		}
1235 	}
1236 
1237 	/*
1238 	 * Make sure we are on the same cpu we were on originally, since
1239 	 * higher level callers expect this.  Also don't pollute caches with
1240 	 * migrated userland data by (eventually) returning to userland
1241 	 * on a different cpu.
1242 	 */
1243 	lwkt_setcpu_self(globaldata_find(origcpu));
1244 	kfree(marker, M_TEMP);
1245 	return (error);
1246 }
1247 
1248 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1249 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1250 
1251 static int
1252 tcp_getcred(SYSCTL_HANDLER_ARGS)
1253 {
1254 	struct sockaddr_in addrs[2];
1255 	struct inpcb *inp;
1256 	int cpu;
1257 	int error;
1258 
1259 	error = priv_check(req->td, PRIV_ROOT);
1260 	if (error != 0)
1261 		return (error);
1262 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1263 	if (error != 0)
1264 		return (error);
1265 	crit_enter();
1266 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1267 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1268 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1269 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1270 	if (inp == NULL || inp->inp_socket == NULL) {
1271 		error = ENOENT;
1272 		goto out;
1273 	}
1274 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1275 out:
1276 	crit_exit();
1277 	return (error);
1278 }
1279 
1280 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1281     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1282 
1283 #ifdef INET6
1284 static int
1285 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1286 {
1287 	struct sockaddr_in6 addrs[2];
1288 	struct inpcb *inp;
1289 	int error;
1290 	boolean_t mapped = FALSE;
1291 
1292 	error = priv_check(req->td, PRIV_ROOT);
1293 	if (error != 0)
1294 		return (error);
1295 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1296 	if (error != 0)
1297 		return (error);
1298 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1299 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1300 			mapped = TRUE;
1301 		else
1302 			return (EINVAL);
1303 	}
1304 	crit_enter();
1305 	if (mapped) {
1306 		inp = in_pcblookup_hash(&tcbinfo[0],
1307 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1308 		    addrs[1].sin6_port,
1309 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1310 		    addrs[0].sin6_port,
1311 		    0, NULL);
1312 	} else {
1313 		inp = in6_pcblookup_hash(&tcbinfo[0],
1314 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1315 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1316 		    0, NULL);
1317 	}
1318 	if (inp == NULL || inp->inp_socket == NULL) {
1319 		error = ENOENT;
1320 		goto out;
1321 	}
1322 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1323 out:
1324 	crit_exit();
1325 	return (error);
1326 }
1327 
1328 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1329 	    0, 0,
1330 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1331 #endif
1332 
1333 struct netmsg_tcp_notify {
1334 	struct netmsg	nm_nmsg;
1335 	void		(*nm_notify)(struct inpcb *, int);
1336 	struct in_addr	nm_faddr;
1337 	int		nm_arg;
1338 };
1339 
1340 static void
1341 tcp_notifyall_oncpu(struct netmsg *netmsg)
1342 {
1343 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1344 	int nextcpu;
1345 
1346 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1347 			nmsg->nm_arg, nmsg->nm_notify);
1348 
1349 	nextcpu = mycpuid + 1;
1350 	if (nextcpu < ncpus2)
1351 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1352 	else
1353 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1354 }
1355 
1356 void
1357 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1358 {
1359 	struct ip *ip = vip;
1360 	struct tcphdr *th;
1361 	struct in_addr faddr;
1362 	struct inpcb *inp;
1363 	struct tcpcb *tp;
1364 	void (*notify)(struct inpcb *, int) = tcp_notify;
1365 	tcp_seq icmpseq;
1366 	int arg, cpu;
1367 
1368 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1369 		return;
1370 	}
1371 
1372 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1373 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1374 		return;
1375 
1376 	arg = inetctlerrmap[cmd];
1377 	if (cmd == PRC_QUENCH) {
1378 		notify = tcp_quench;
1379 	} else if (icmp_may_rst &&
1380 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1381 		    cmd == PRC_UNREACH_PORT ||
1382 		    cmd == PRC_TIMXCEED_INTRANS) &&
1383 		   ip != NULL) {
1384 		notify = tcp_drop_syn_sent;
1385 	} else if (cmd == PRC_MSGSIZE) {
1386 		struct icmp *icmp = (struct icmp *)
1387 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1388 
1389 		arg = ntohs(icmp->icmp_nextmtu);
1390 		notify = tcp_mtudisc;
1391 	} else if (PRC_IS_REDIRECT(cmd)) {
1392 		ip = NULL;
1393 		notify = in_rtchange;
1394 	} else if (cmd == PRC_HOSTDEAD) {
1395 		ip = NULL;
1396 	}
1397 
1398 	if (ip != NULL) {
1399 		crit_enter();
1400 		th = (struct tcphdr *)((caddr_t)ip +
1401 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1402 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1403 				  ip->ip_src.s_addr, th->th_sport);
1404 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1405 					ip->ip_src, th->th_sport, 0, NULL);
1406 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1407 			icmpseq = htonl(th->th_seq);
1408 			tp = intotcpcb(inp);
1409 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1410 			    SEQ_LT(icmpseq, tp->snd_max))
1411 				(*notify)(inp, arg);
1412 		} else {
1413 			struct in_conninfo inc;
1414 
1415 			inc.inc_fport = th->th_dport;
1416 			inc.inc_lport = th->th_sport;
1417 			inc.inc_faddr = faddr;
1418 			inc.inc_laddr = ip->ip_src;
1419 #ifdef INET6
1420 			inc.inc_isipv6 = 0;
1421 #endif
1422 			syncache_unreach(&inc, th);
1423 		}
1424 		crit_exit();
1425 	} else {
1426 		struct netmsg_tcp_notify nmsg;
1427 
1428 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1429 		netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0,
1430 			    tcp_notifyall_oncpu);
1431 		nmsg.nm_faddr = faddr;
1432 		nmsg.nm_arg = arg;
1433 		nmsg.nm_notify = notify;
1434 
1435 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1436 	}
1437 }
1438 
1439 #ifdef INET6
1440 void
1441 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1442 {
1443 	struct tcphdr th;
1444 	void (*notify) (struct inpcb *, int) = tcp_notify;
1445 	struct ip6_hdr *ip6;
1446 	struct mbuf *m;
1447 	struct ip6ctlparam *ip6cp = NULL;
1448 	const struct sockaddr_in6 *sa6_src = NULL;
1449 	int off;
1450 	struct tcp_portonly {
1451 		u_int16_t th_sport;
1452 		u_int16_t th_dport;
1453 	} *thp;
1454 	int arg;
1455 
1456 	if (sa->sa_family != AF_INET6 ||
1457 	    sa->sa_len != sizeof(struct sockaddr_in6))
1458 		return;
1459 
1460 	arg = 0;
1461 	if (cmd == PRC_QUENCH)
1462 		notify = tcp_quench;
1463 	else if (cmd == PRC_MSGSIZE) {
1464 		struct ip6ctlparam *ip6cp = d;
1465 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1466 
1467 		arg = ntohl(icmp6->icmp6_mtu);
1468 		notify = tcp_mtudisc;
1469 	} else if (!PRC_IS_REDIRECT(cmd) &&
1470 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1471 		return;
1472 	}
1473 
1474 	/* if the parameter is from icmp6, decode it. */
1475 	if (d != NULL) {
1476 		ip6cp = (struct ip6ctlparam *)d;
1477 		m = ip6cp->ip6c_m;
1478 		ip6 = ip6cp->ip6c_ip6;
1479 		off = ip6cp->ip6c_off;
1480 		sa6_src = ip6cp->ip6c_src;
1481 	} else {
1482 		m = NULL;
1483 		ip6 = NULL;
1484 		off = 0;	/* fool gcc */
1485 		sa6_src = &sa6_any;
1486 	}
1487 
1488 	if (ip6 != NULL) {
1489 		struct in_conninfo inc;
1490 		/*
1491 		 * XXX: We assume that when IPV6 is non NULL,
1492 		 * M and OFF are valid.
1493 		 */
1494 
1495 		/* check if we can safely examine src and dst ports */
1496 		if (m->m_pkthdr.len < off + sizeof *thp)
1497 			return;
1498 
1499 		bzero(&th, sizeof th);
1500 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1501 
1502 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1503 		    (struct sockaddr *)ip6cp->ip6c_src,
1504 		    th.th_sport, cmd, arg, notify);
1505 
1506 		inc.inc_fport = th.th_dport;
1507 		inc.inc_lport = th.th_sport;
1508 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1509 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1510 		inc.inc_isipv6 = 1;
1511 		syncache_unreach(&inc, &th);
1512 	} else
1513 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1514 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1515 }
1516 #endif
1517 
1518 /*
1519  * Following is where TCP initial sequence number generation occurs.
1520  *
1521  * There are two places where we must use initial sequence numbers:
1522  * 1.  In SYN-ACK packets.
1523  * 2.  In SYN packets.
1524  *
1525  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1526  * tcp_syncache.c for details.
1527  *
1528  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1529  * depends on this property.  In addition, these ISNs should be
1530  * unguessable so as to prevent connection hijacking.  To satisfy
1531  * the requirements of this situation, the algorithm outlined in
1532  * RFC 1948 is used to generate sequence numbers.
1533  *
1534  * Implementation details:
1535  *
1536  * Time is based off the system timer, and is corrected so that it
1537  * increases by one megabyte per second.  This allows for proper
1538  * recycling on high speed LANs while still leaving over an hour
1539  * before rollover.
1540  *
1541  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1542  * between seeding of isn_secret.  This is normally set to zero,
1543  * as reseeding should not be necessary.
1544  *
1545  */
1546 
1547 #define	ISN_BYTES_PER_SECOND 1048576
1548 
1549 u_char isn_secret[32];
1550 int isn_last_reseed;
1551 MD5_CTX isn_ctx;
1552 
1553 tcp_seq
1554 tcp_new_isn(struct tcpcb *tp)
1555 {
1556 	u_int32_t md5_buffer[4];
1557 	tcp_seq new_isn;
1558 
1559 	/* Seed if this is the first use, reseed if requested. */
1560 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1561 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1562 		< (u_int)ticks))) {
1563 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1564 		isn_last_reseed = ticks;
1565 	}
1566 
1567 	/* Compute the md5 hash and return the ISN. */
1568 	MD5Init(&isn_ctx);
1569 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1570 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1571 #ifdef INET6
1572 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1573 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1574 			  sizeof(struct in6_addr));
1575 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1576 			  sizeof(struct in6_addr));
1577 	} else
1578 #endif
1579 	{
1580 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1581 			  sizeof(struct in_addr));
1582 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1583 			  sizeof(struct in_addr));
1584 	}
1585 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1586 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1587 	new_isn = (tcp_seq) md5_buffer[0];
1588 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1589 	return (new_isn);
1590 }
1591 
1592 /*
1593  * When a source quench is received, close congestion window
1594  * to one segment.  We will gradually open it again as we proceed.
1595  */
1596 void
1597 tcp_quench(struct inpcb *inp, int error)
1598 {
1599 	struct tcpcb *tp = intotcpcb(inp);
1600 
1601 	if (tp != NULL) {
1602 		tp->snd_cwnd = tp->t_maxseg;
1603 		tp->snd_wacked = 0;
1604 	}
1605 }
1606 
1607 /*
1608  * When a specific ICMP unreachable message is received and the
1609  * connection state is SYN-SENT, drop the connection.  This behavior
1610  * is controlled by the icmp_may_rst sysctl.
1611  */
1612 void
1613 tcp_drop_syn_sent(struct inpcb *inp, int error)
1614 {
1615 	struct tcpcb *tp = intotcpcb(inp);
1616 
1617 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1618 		tcp_drop(tp, error);
1619 }
1620 
1621 /*
1622  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1623  * based on the new value in the route.  Also nudge TCP to send something,
1624  * since we know the packet we just sent was dropped.
1625  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1626  */
1627 void
1628 tcp_mtudisc(struct inpcb *inp, int mtu)
1629 {
1630 	struct tcpcb *tp = intotcpcb(inp);
1631 	struct rtentry *rt;
1632 	struct socket *so = inp->inp_socket;
1633 	int maxopd, mss;
1634 #ifdef INET6
1635 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1636 #else
1637 	const boolean_t isipv6 = FALSE;
1638 #endif
1639 
1640 	if (tp == NULL)
1641 		return;
1642 
1643 	/*
1644 	 * If no MTU is provided in the ICMP message, use the
1645 	 * next lower likely value, as specified in RFC 1191.
1646 	 */
1647 	if (mtu == 0) {
1648 		int oldmtu;
1649 
1650 		oldmtu = tp->t_maxopd +
1651 		    (isipv6 ?
1652 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1653 		     sizeof(struct tcpiphdr));
1654 		mtu = ip_next_mtu(oldmtu, 0);
1655 	}
1656 
1657 	if (isipv6)
1658 		rt = tcp_rtlookup6(&inp->inp_inc);
1659 	else
1660 		rt = tcp_rtlookup(&inp->inp_inc);
1661 	if (rt != NULL) {
1662 		struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1663 
1664 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1665 			mtu = rt->rt_rmx.rmx_mtu;
1666 
1667 		maxopd = mtu -
1668 		    (isipv6 ?
1669 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1670 		     sizeof(struct tcpiphdr));
1671 
1672 		/*
1673 		 * XXX - The following conditional probably violates the TCP
1674 		 * spec.  The problem is that, since we don't know the
1675 		 * other end's MSS, we are supposed to use a conservative
1676 		 * default.  But, if we do that, then MTU discovery will
1677 		 * never actually take place, because the conservative
1678 		 * default is much less than the MTUs typically seen
1679 		 * on the Internet today.  For the moment, we'll sweep
1680 		 * this under the carpet.
1681 		 *
1682 		 * The conservative default might not actually be a problem
1683 		 * if the only case this occurs is when sending an initial
1684 		 * SYN with options and data to a host we've never talked
1685 		 * to before.  Then, they will reply with an MSS value which
1686 		 * will get recorded and the new parameters should get
1687 		 * recomputed.  For Further Study.
1688 		 */
1689 		if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1690 			maxopd = taop->tao_mssopt;
1691 	} else
1692 		maxopd = mtu -
1693 		    (isipv6 ?
1694 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1695 		     sizeof(struct tcpiphdr));
1696 
1697 	if (tp->t_maxopd <= maxopd)
1698 		return;
1699 	tp->t_maxopd = maxopd;
1700 
1701 	mss = maxopd;
1702 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1703 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1704 		mss -= TCPOLEN_TSTAMP_APPA;
1705 
1706 	if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1707 			   (TF_REQ_CC | TF_RCVD_CC))
1708 		mss -= TCPOLEN_CC_APPA;
1709 
1710 	/* round down to multiple of MCLBYTES */
1711 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1712 	if (mss > MCLBYTES)
1713 		mss &= ~(MCLBYTES - 1);
1714 #else
1715 	if (mss > MCLBYTES)
1716 		mss = (mss / MCLBYTES) * MCLBYTES;
1717 #endif
1718 
1719 	if (so->so_snd.ssb_hiwat < mss)
1720 		mss = so->so_snd.ssb_hiwat;
1721 
1722 	tp->t_maxseg = mss;
1723 	tp->t_rtttime = 0;
1724 	tp->snd_nxt = tp->snd_una;
1725 	tcp_output(tp);
1726 	tcpstat.tcps_mturesent++;
1727 }
1728 
1729 /*
1730  * Look-up the routing entry to the peer of this inpcb.  If no route
1731  * is found and it cannot be allocated the return NULL.  This routine
1732  * is called by TCP routines that access the rmx structure and by tcp_mss
1733  * to get the interface MTU.
1734  */
1735 struct rtentry *
1736 tcp_rtlookup(struct in_conninfo *inc)
1737 {
1738 	struct route *ro = &inc->inc_route;
1739 
1740 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1741 		/* No route yet, so try to acquire one */
1742 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1743 			/*
1744 			 * unused portions of the structure MUST be zero'd
1745 			 * out because rtalloc() treats it as opaque data
1746 			 */
1747 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1748 			ro->ro_dst.sa_family = AF_INET;
1749 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1750 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1751 			    inc->inc_faddr;
1752 			rtalloc(ro);
1753 		}
1754 	}
1755 	return (ro->ro_rt);
1756 }
1757 
1758 #ifdef INET6
1759 struct rtentry *
1760 tcp_rtlookup6(struct in_conninfo *inc)
1761 {
1762 	struct route_in6 *ro6 = &inc->inc6_route;
1763 
1764 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1765 		/* No route yet, so try to acquire one */
1766 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1767 			/*
1768 			 * unused portions of the structure MUST be zero'd
1769 			 * out because rtalloc() treats it as opaque data
1770 			 */
1771 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1772 			ro6->ro_dst.sin6_family = AF_INET6;
1773 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1774 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1775 			rtalloc((struct route *)ro6);
1776 		}
1777 	}
1778 	return (ro6->ro_rt);
1779 }
1780 #endif
1781 
1782 #ifdef IPSEC
1783 /* compute ESP/AH header size for TCP, including outer IP header. */
1784 size_t
1785 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1786 {
1787 	struct inpcb *inp;
1788 	struct mbuf *m;
1789 	size_t hdrsiz;
1790 	struct ip *ip;
1791 	struct tcphdr *th;
1792 
1793 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1794 		return (0);
1795 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1796 	if (!m)
1797 		return (0);
1798 
1799 #ifdef INET6
1800 	if (inp->inp_vflag & INP_IPV6) {
1801 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1802 
1803 		th = (struct tcphdr *)(ip6 + 1);
1804 		m->m_pkthdr.len = m->m_len =
1805 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1806 		tcp_fillheaders(tp, ip6, th);
1807 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1808 	} else
1809 #endif
1810 	{
1811 		ip = mtod(m, struct ip *);
1812 		th = (struct tcphdr *)(ip + 1);
1813 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1814 		tcp_fillheaders(tp, ip, th);
1815 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1816 	}
1817 
1818 	m_free(m);
1819 	return (hdrsiz);
1820 }
1821 #endif
1822 
1823 /*
1824  * Return a pointer to the cached information about the remote host.
1825  * The cached information is stored in the protocol specific part of
1826  * the route metrics.
1827  */
1828 struct rmxp_tao *
1829 tcp_gettaocache(struct in_conninfo *inc)
1830 {
1831 	struct rtentry *rt;
1832 
1833 #ifdef INET6
1834 	if (inc->inc_isipv6)
1835 		rt = tcp_rtlookup6(inc);
1836 	else
1837 #endif
1838 		rt = tcp_rtlookup(inc);
1839 
1840 	/* Make sure this is a host route and is up. */
1841 	if (rt == NULL ||
1842 	    (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1843 		return (NULL);
1844 
1845 	return (rmx_taop(rt->rt_rmx));
1846 }
1847 
1848 /*
1849  * Clear all the TAO cache entries, called from tcp_init.
1850  *
1851  * XXX
1852  * This routine is just an empty one, because we assume that the routing
1853  * routing tables are initialized at the same time when TCP, so there is
1854  * nothing in the cache left over.
1855  */
1856 static void
1857 tcp_cleartaocache(void)
1858 {
1859 }
1860 
1861 /*
1862  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1863  *
1864  * This code attempts to calculate the bandwidth-delay product as a
1865  * means of determining the optimal window size to maximize bandwidth,
1866  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1867  * routers.  This code also does a fairly good job keeping RTTs in check
1868  * across slow links like modems.  We implement an algorithm which is very
1869  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1870  * transmitter side of a TCP connection and so only effects the transmit
1871  * side of the connection.
1872  *
1873  * BACKGROUND:  TCP makes no provision for the management of buffer space
1874  * at the end points or at the intermediate routers and switches.  A TCP
1875  * stream, whether using NewReno or not, will eventually buffer as
1876  * many packets as it is able and the only reason this typically works is
1877  * due to the fairly small default buffers made available for a connection
1878  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1879  * scaling it is now fairly easy for even a single TCP connection to blow-out
1880  * all available buffer space not only on the local interface, but on
1881  * intermediate routers and switches as well.  NewReno makes a misguided
1882  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1883  * then backing off, then steadily increasing the window again until another
1884  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1885  * is only made worse as network loads increase and the idea of intentionally
1886  * blowing out network buffers is, frankly, a terrible way to manage network
1887  * resources.
1888  *
1889  * It is far better to limit the transmit window prior to the failure
1890  * condition being achieved.  There are two general ways to do this:  First
1891  * you can 'scan' through different transmit window sizes and locate the
1892  * point where the RTT stops increasing, indicating that you have filled the
1893  * pipe, then scan backwards until you note that RTT stops decreasing, then
1894  * repeat ad-infinitum.  This method works in principle but has severe
1895  * implementation issues due to RTT variances, timer granularity, and
1896  * instability in the algorithm which can lead to many false positives and
1897  * create oscillations as well as interact badly with other TCP streams
1898  * implementing the same algorithm.
1899  *
1900  * The second method is to limit the window to the bandwidth delay product
1901  * of the link.  This is the method we implement.  RTT variances and our
1902  * own manipulation of the congestion window, bwnd, can potentially
1903  * destabilize the algorithm.  For this reason we have to stabilize the
1904  * elements used to calculate the window.  We do this by using the minimum
1905  * observed RTT, the long term average of the observed bandwidth, and
1906  * by adding two segments worth of slop.  It isn't perfect but it is able
1907  * to react to changing conditions and gives us a very stable basis on
1908  * which to extend the algorithm.
1909  */
1910 void
1911 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1912 {
1913 	u_long bw;
1914 	u_long bwnd;
1915 	int save_ticks;
1916 	int delta_ticks;
1917 
1918 	/*
1919 	 * If inflight_enable is disabled in the middle of a tcp connection,
1920 	 * make sure snd_bwnd is effectively disabled.
1921 	 */
1922 	if (!tcp_inflight_enable) {
1923 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1924 		tp->snd_bandwidth = 0;
1925 		return;
1926 	}
1927 
1928 	/*
1929 	 * Validate the delta time.  If a connection is new or has been idle
1930 	 * a long time we have to reset the bandwidth calculator.
1931 	 */
1932 	save_ticks = ticks;
1933 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1934 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1935 		tp->t_bw_rtttime = ticks;
1936 		tp->t_bw_rtseq = ack_seq;
1937 		if (tp->snd_bandwidth == 0)
1938 			tp->snd_bandwidth = tcp_inflight_min;
1939 		return;
1940 	}
1941 	if (delta_ticks == 0)
1942 		return;
1943 
1944 	/*
1945 	 * Sanity check, plus ignore pure window update acks.
1946 	 */
1947 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1948 		return;
1949 
1950 	/*
1951 	 * Figure out the bandwidth.  Due to the tick granularity this
1952 	 * is a very rough number and it MUST be averaged over a fairly
1953 	 * long period of time.  XXX we need to take into account a link
1954 	 * that is not using all available bandwidth, but for now our
1955 	 * slop will ramp us up if this case occurs and the bandwidth later
1956 	 * increases.
1957 	 */
1958 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1959 	tp->t_bw_rtttime = save_ticks;
1960 	tp->t_bw_rtseq = ack_seq;
1961 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1962 
1963 	tp->snd_bandwidth = bw;
1964 
1965 	/*
1966 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1967 	 * segments.  The additional slop puts us squarely in the sweet
1968 	 * spot and also handles the bandwidth run-up case.  Without the
1969 	 * slop we could be locking ourselves into a lower bandwidth.
1970 	 *
1971 	 * Situations Handled:
1972 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1973 	 *	    high speed LANs, allowing larger TCP buffers to be
1974 	 *	    specified, and also does a good job preventing
1975 	 *	    over-queueing of packets over choke points like modems
1976 	 *	    (at least for the transmit side).
1977 	 *
1978 	 *	(2) Is able to handle changing network loads (bandwidth
1979 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1980 	 *	    increases).
1981 	 *
1982 	 *	(3) Theoretically should stabilize in the face of multiple
1983 	 *	    connections implementing the same algorithm (this may need
1984 	 *	    a little work).
1985 	 *
1986 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1987 	 *	    be adjusted with a sysctl but typically only needs to be on
1988 	 *	    very slow connections.  A value no smaller then 5 should
1989 	 *	    be used, but only reduce this default if you have no other
1990 	 *	    choice.
1991 	 */
1992 
1993 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1994 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1995 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1996 #undef USERTT
1997 
1998 	if (tcp_inflight_debug > 0) {
1999 		static int ltime;
2000 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2001 			ltime = ticks;
2002 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2003 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
2004 		}
2005 	}
2006 	if ((long)bwnd < tcp_inflight_min)
2007 		bwnd = tcp_inflight_min;
2008 	if (bwnd > tcp_inflight_max)
2009 		bwnd = tcp_inflight_max;
2010 	if ((long)bwnd < tp->t_maxseg * 2)
2011 		bwnd = tp->t_maxseg * 2;
2012 	tp->snd_bwnd = bwnd;
2013 }
2014