xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 7a8ad07e4feda5d9f79caa4ef40f29e9a683e154)
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  */
69 
70 #include "opt_compat.h"
71 #include "opt_inet.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 <net/route.h>
97 #include <net/if.h>
98 #include <net/netisr.h>
99 
100 #define	_IP_VHL
101 #include <netinet/in.h>
102 #include <netinet/in_systm.h>
103 #include <netinet/ip.h>
104 #include <netinet/ip6.h>
105 #include <netinet/in_pcb.h>
106 #include <netinet6/in6_pcb.h>
107 #include <netinet/in_var.h>
108 #include <netinet/ip_var.h>
109 #include <netinet6/ip6_var.h>
110 #include <netinet/ip_icmp.h>
111 #ifdef INET6
112 #include <netinet/icmp6.h>
113 #endif
114 #include <netinet/tcp.h>
115 #include <netinet/tcp_fsm.h>
116 #include <netinet/tcp_seq.h>
117 #include <netinet/tcp_timer.h>
118 #include <netinet/tcp_timer2.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
122 #ifdef TCPDEBUG
123 #include <netinet/tcp_debug.h>
124 #endif
125 #include <netinet6/ip6protosw.h>
126 
127 #ifdef IPSEC
128 #include <netinet6/ipsec.h>
129 #include <netproto/key/key.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 <machine/smp.h>
145 
146 #include <sys/msgport2.h>
147 #include <sys/mplock2.h>
148 #include <net/netmsg2.h>
149 
150 #if !defined(KTR_TCP)
151 #define KTR_TCP		KTR_ALL
152 #endif
153 /*
154 KTR_INFO_MASTER(tcp);
155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
158 #define logtcp(name)	KTR_LOG(tcp_ ## name)
159 */
160 
161 struct inpcbinfo tcbinfo[MAXCPU];
162 struct tcpcbackqhead tcpcbackq[MAXCPU];
163 
164 static struct lwkt_token tcp_port_token =
165 		LWKT_TOKEN_INITIALIZER(tcp_port_token);
166 
167 int tcp_mssdflt = TCP_MSS;
168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
169     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
170 
171 #ifdef INET6
172 int tcp_v6mssdflt = TCP6_MSS;
173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
174     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
175 #endif
176 
177 /*
178  * Minimum MSS we accept and use. This prevents DoS attacks where
179  * we are forced to a ridiculous low MSS like 20 and send hundreds
180  * of packets instead of one. The effect scales with the available
181  * bandwidth and quickly saturates the CPU and network interface
182  * with packet generation and sending. Set to zero to disable MINMSS
183  * checking. This setting prevents us from sending too small packets.
184  */
185 int tcp_minmss = TCP_MINMSS;
186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
187     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
188 
189 #if 0
190 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
191 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
192     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
193 #endif
194 
195 int tcp_do_rfc1323 = 1;
196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
197     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
198 
199 static int tcp_tcbhashsize = 0;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
201      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
202 
203 static int do_tcpdrain = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
205      "Enable tcp_drain routine for extra help when low on mbufs");
206 
207 static int icmp_may_rst = 1;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
209     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
210 
211 static int tcp_isn_reseed_interval = 0;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
213     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
214 
215 /*
216  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
217  * by default, but with generous values which should allow maximal
218  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
219  *
220  * The reason for doing this is that the limiter is the only mechanism we
221  * have which seems to do a really good job preventing receiver RX rings
222  * on network interfaces from getting blown out.  Even though GigE/10GigE
223  * is supposed to flow control it looks like either it doesn't actually
224  * do it or Open Source drivers do not properly enable it.
225  *
226  * People using the limiter to reduce bottlenecks on slower WAN connections
227  * should set the slop to 20 (2 packets).
228  */
229 static int tcp_inflight_enable = 1;
230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
231     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
232 
233 static int tcp_inflight_debug = 0;
234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
235     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
236 
237 static int tcp_inflight_min = 6144;
238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
239     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
240 
241 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
243     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
244 
245 static int tcp_inflight_stab = 50;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
247     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
248 
249 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
250 static struct malloc_pipe tcptemp_mpipe;
251 
252 static void tcp_willblock(void);
253 static void tcp_notify (struct inpcb *, int);
254 
255 struct tcp_stats tcpstats_percpu[MAXCPU];
256 #ifdef SMP
257 static int
258 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
259 {
260 	int cpu, error = 0;
261 
262 	for (cpu = 0; cpu < ncpus; ++cpu) {
263 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
264 					sizeof(struct tcp_stats))))
265 			break;
266 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
267 				       sizeof(struct tcp_stats))))
268 			break;
269 	}
270 
271 	return (error);
272 }
273 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
274     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
275 #else
276 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
277     &tcpstat, tcp_stats, "TCP statistics");
278 #endif
279 
280 /*
281  * Target size of TCP PCB hash tables. Must be a power of two.
282  *
283  * Note that this can be overridden by the kernel environment
284  * variable net.inet.tcp.tcbhashsize
285  */
286 #ifndef TCBHASHSIZE
287 #define	TCBHASHSIZE	512
288 #endif
289 
290 /*
291  * This is the actual shape of what we allocate using the zone
292  * allocator.  Doing it this way allows us to protect both structures
293  * using the same generation count, and also eliminates the overhead
294  * of allocating tcpcbs separately.  By hiding the structure here,
295  * we avoid changing most of the rest of the code (although it needs
296  * to be changed, eventually, for greater efficiency).
297  */
298 #define	ALIGNMENT	32
299 #define	ALIGNM1		(ALIGNMENT - 1)
300 struct	inp_tp {
301 	union {
302 		struct	inpcb inp;
303 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
304 	} inp_tp_u;
305 	struct	tcpcb tcb;
306 	struct	tcp_callout inp_tp_rexmt;
307 	struct	tcp_callout inp_tp_persist;
308 	struct	tcp_callout inp_tp_keep;
309 	struct	tcp_callout inp_tp_2msl;
310 	struct	tcp_callout inp_tp_delack;
311 	struct	netmsg_tcp_timer inp_tp_timermsg;
312 };
313 #undef ALIGNMENT
314 #undef ALIGNM1
315 
316 /*
317  * Tcp initialization
318  */
319 void
320 tcp_init(void)
321 {
322 	struct inpcbporthead *porthashbase;
323 	struct inpcbinfo *ticb;
324 	u_long porthashmask;
325 	int hashsize = TCBHASHSIZE;
326 	int cpu;
327 
328 	/*
329 	 * note: tcptemp is used for keepalives, and it is ok for an
330 	 * allocation to fail so do not specify MPF_INT.
331 	 */
332 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
333 		    25, -1, 0, NULL, NULL, NULL);
334 
335 	tcp_delacktime = TCPTV_DELACK;
336 	tcp_keepinit = TCPTV_KEEP_INIT;
337 	tcp_keepidle = TCPTV_KEEP_IDLE;
338 	tcp_keepintvl = TCPTV_KEEPINTVL;
339 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
340 	tcp_msl = TCPTV_MSL;
341 	tcp_rexmit_min = TCPTV_MIN;
342 	tcp_rexmit_slop = TCPTV_CPU_VAR;
343 
344 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
345 	if (!powerof2(hashsize)) {
346 		kprintf("WARNING: TCB hash size not a power of 2\n");
347 		hashsize = 512; /* safe default */
348 	}
349 	tcp_tcbhashsize = hashsize;
350 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
351 
352 	for (cpu = 0; cpu < ncpus2; cpu++) {
353 		ticb = &tcbinfo[cpu];
354 		in_pcbinfo_init(ticb);
355 		ticb->cpu = cpu;
356 		ticb->hashbase = hashinit(hashsize, M_PCB,
357 					  &ticb->hashmask);
358 		ticb->porthashbase = porthashbase;
359 		ticb->porthashmask = porthashmask;
360 		ticb->porttoken = &tcp_port_token;
361 #if 0
362 		ticb->porthashbase = hashinit(hashsize, M_PCB,
363 					      &ticb->porthashmask);
364 #endif
365 		ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
366 						  &ticb->wildcardhashmask);
367 		ticb->ipi_size = sizeof(struct inp_tp);
368 		TAILQ_INIT(&tcpcbackq[cpu]);
369 	}
370 
371 	tcp_reass_maxseg = nmbclusters / 16;
372 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
373 
374 #ifdef INET6
375 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
376 #else
377 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
378 #endif
379 	if (max_protohdr < TCP_MINPROTOHDR)
380 		max_protohdr = TCP_MINPROTOHDR;
381 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
382 		panic("tcp_init");
383 #undef TCP_MINPROTOHDR
384 
385 	/*
386 	 * Initialize TCP statistics counters for each CPU.
387 	 */
388 #ifdef SMP
389 	for (cpu = 0; cpu < ncpus; ++cpu) {
390 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
391 	}
392 #else
393 	bzero(&tcpstat, sizeof(struct tcp_stats));
394 #endif
395 
396 	syncache_init();
397 	netisr_register_rollup(tcp_willblock);
398 }
399 
400 static void
401 tcp_willblock(void)
402 {
403 	struct tcpcb *tp;
404 	int cpu = mycpu->gd_cpuid;
405 
406 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
407 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
408 		tp->t_flags &= ~TF_ONOUTPUTQ;
409 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
410 		tcp_output(tp);
411 	}
412 }
413 
414 /*
415  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
416  * tcp_template used to store this data in mbufs, but we now recopy it out
417  * of the tcpcb each time to conserve mbufs.
418  */
419 void
420 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
421 {
422 	struct inpcb *inp = tp->t_inpcb;
423 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
424 
425 #ifdef INET6
426 	if (inp->inp_vflag & INP_IPV6) {
427 		struct ip6_hdr *ip6;
428 
429 		ip6 = (struct ip6_hdr *)ip_ptr;
430 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
431 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
432 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
433 			(IPV6_VERSION & IPV6_VERSION_MASK);
434 		ip6->ip6_nxt = IPPROTO_TCP;
435 		ip6->ip6_plen = sizeof(struct tcphdr);
436 		ip6->ip6_src = inp->in6p_laddr;
437 		ip6->ip6_dst = inp->in6p_faddr;
438 		tcp_hdr->th_sum = 0;
439 	} else
440 #endif
441 	{
442 		struct ip *ip = (struct ip *) ip_ptr;
443 
444 		ip->ip_vhl = IP_VHL_BORING;
445 		ip->ip_tos = 0;
446 		ip->ip_len = 0;
447 		ip->ip_id = 0;
448 		ip->ip_off = 0;
449 		ip->ip_ttl = 0;
450 		ip->ip_sum = 0;
451 		ip->ip_p = IPPROTO_TCP;
452 		ip->ip_src = inp->inp_laddr;
453 		ip->ip_dst = inp->inp_faddr;
454 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
455 				    ip->ip_dst.s_addr,
456 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
457 	}
458 
459 	tcp_hdr->th_sport = inp->inp_lport;
460 	tcp_hdr->th_dport = inp->inp_fport;
461 	tcp_hdr->th_seq = 0;
462 	tcp_hdr->th_ack = 0;
463 	tcp_hdr->th_x2 = 0;
464 	tcp_hdr->th_off = 5;
465 	tcp_hdr->th_flags = 0;
466 	tcp_hdr->th_win = 0;
467 	tcp_hdr->th_urp = 0;
468 }
469 
470 /*
471  * Create template to be used to send tcp packets on a connection.
472  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
473  * use for this function is in keepalives, which use tcp_respond.
474  */
475 struct tcptemp *
476 tcp_maketemplate(struct tcpcb *tp)
477 {
478 	struct tcptemp *tmp;
479 
480 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
481 		return (NULL);
482 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
483 	return (tmp);
484 }
485 
486 void
487 tcp_freetemplate(struct tcptemp *tmp)
488 {
489 	mpipe_free(&tcptemp_mpipe, tmp);
490 }
491 
492 /*
493  * Send a single message to the TCP at address specified by
494  * the given TCP/IP header.  If m == NULL, then we make a copy
495  * of the tcpiphdr at ti and send directly to the addressed host.
496  * This is used to force keep alive messages out using the TCP
497  * template for a connection.  If flags are given then we send
498  * a message back to the TCP which originated the * segment ti,
499  * and discard the mbuf containing it and any other attached mbufs.
500  *
501  * In any case the ack and sequence number of the transmitted
502  * segment are as specified by the parameters.
503  *
504  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
505  */
506 void
507 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
508 	    tcp_seq ack, tcp_seq seq, int flags)
509 {
510 	int tlen;
511 	int win = 0;
512 	struct route *ro = NULL;
513 	struct route sro;
514 	struct ip *ip = ipgen;
515 	struct tcphdr *nth;
516 	int ipflags = 0;
517 	struct route_in6 *ro6 = NULL;
518 	struct route_in6 sro6;
519 	struct ip6_hdr *ip6 = ipgen;
520 	boolean_t use_tmpro = TRUE;
521 #ifdef INET6
522 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
523 #else
524 	const boolean_t isipv6 = FALSE;
525 #endif
526 
527 	if (tp != NULL) {
528 		if (!(flags & TH_RST)) {
529 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
530 			if (win < 0)
531 				win = 0;
532 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
533 				win = (long)TCP_MAXWIN << tp->rcv_scale;
534 		}
535 		/*
536 		 * Don't use the route cache of a listen socket,
537 		 * it is not MPSAFE; use temporary route cache.
538 		 */
539 		if (tp->t_state != TCPS_LISTEN) {
540 			if (isipv6)
541 				ro6 = &tp->t_inpcb->in6p_route;
542 			else
543 				ro = &tp->t_inpcb->inp_route;
544 			use_tmpro = FALSE;
545 		}
546 	}
547 	if (use_tmpro) {
548 		if (isipv6) {
549 			ro6 = &sro6;
550 			bzero(ro6, sizeof *ro6);
551 		} else {
552 			ro = &sro;
553 			bzero(ro, sizeof *ro);
554 		}
555 	}
556 	if (m == NULL) {
557 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
558 		if (m == NULL)
559 			return;
560 		tlen = 0;
561 		m->m_data += max_linkhdr;
562 		if (isipv6) {
563 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
564 			ip6 = mtod(m, struct ip6_hdr *);
565 			nth = (struct tcphdr *)(ip6 + 1);
566 		} else {
567 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
568 			ip = mtod(m, struct ip *);
569 			nth = (struct tcphdr *)(ip + 1);
570 		}
571 		bcopy(th, nth, sizeof(struct tcphdr));
572 		flags = TH_ACK;
573 	} else {
574 		m_freem(m->m_next);
575 		m->m_next = NULL;
576 		m->m_data = (caddr_t)ipgen;
577 		/* m_len is set later */
578 		tlen = 0;
579 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
580 		if (isipv6) {
581 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
582 			nth = (struct tcphdr *)(ip6 + 1);
583 		} else {
584 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
585 			nth = (struct tcphdr *)(ip + 1);
586 		}
587 		if (th != nth) {
588 			/*
589 			 * this is usually a case when an extension header
590 			 * exists between the IPv6 header and the
591 			 * TCP header.
592 			 */
593 			nth->th_sport = th->th_sport;
594 			nth->th_dport = th->th_dport;
595 		}
596 		xchg(nth->th_dport, nth->th_sport, n_short);
597 #undef xchg
598 	}
599 	if (isipv6) {
600 		ip6->ip6_flow = 0;
601 		ip6->ip6_vfc = IPV6_VERSION;
602 		ip6->ip6_nxt = IPPROTO_TCP;
603 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
604 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
605 	} else {
606 		tlen += sizeof(struct tcpiphdr);
607 		ip->ip_len = tlen;
608 		ip->ip_ttl = ip_defttl;
609 	}
610 	m->m_len = tlen;
611 	m->m_pkthdr.len = tlen;
612 	m->m_pkthdr.rcvif = NULL;
613 	nth->th_seq = htonl(seq);
614 	nth->th_ack = htonl(ack);
615 	nth->th_x2 = 0;
616 	nth->th_off = sizeof(struct tcphdr) >> 2;
617 	nth->th_flags = flags;
618 	if (tp != NULL)
619 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
620 	else
621 		nth->th_win = htons((u_short)win);
622 	nth->th_urp = 0;
623 	if (isipv6) {
624 		nth->th_sum = 0;
625 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
626 					sizeof(struct ip6_hdr),
627 					tlen - sizeof(struct ip6_hdr));
628 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
629 					       (ro6 && ro6->ro_rt) ?
630 						ro6->ro_rt->rt_ifp : NULL);
631 	} else {
632 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
633 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
634 		m->m_pkthdr.csum_flags = CSUM_TCP;
635 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
636 	}
637 #ifdef TCPDEBUG
638 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
639 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
640 #endif
641 	if (isipv6) {
642 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
643 			   tp ? tp->t_inpcb : NULL);
644 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
645 			RTFREE(ro6->ro_rt);
646 			ro6->ro_rt = NULL;
647 		}
648 	} else {
649 		ipflags |= IP_DEBUGROUTE;
650 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
651 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
652 			RTFREE(ro->ro_rt);
653 			ro->ro_rt = NULL;
654 		}
655 	}
656 }
657 
658 /*
659  * Create a new TCP control block, making an
660  * empty reassembly queue and hooking it to the argument
661  * protocol control block.  The `inp' parameter must have
662  * come from the zone allocator set up in tcp_init().
663  */
664 struct tcpcb *
665 tcp_newtcpcb(struct inpcb *inp)
666 {
667 	struct inp_tp *it;
668 	struct tcpcb *tp;
669 #ifdef INET6
670 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
671 #else
672 	const boolean_t isipv6 = FALSE;
673 #endif
674 
675 	it = (struct inp_tp *)inp;
676 	tp = &it->tcb;
677 	bzero(tp, sizeof(struct tcpcb));
678 	LIST_INIT(&tp->t_segq);
679 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
680 
681 	/* Set up our timeouts. */
682 	tp->tt_rexmt = &it->inp_tp_rexmt;
683 	tp->tt_persist = &it->inp_tp_persist;
684 	tp->tt_keep = &it->inp_tp_keep;
685 	tp->tt_2msl = &it->inp_tp_2msl;
686 	tp->tt_delack = &it->inp_tp_delack;
687 	tcp_inittimers(tp);
688 
689 	/*
690 	 * Zero out timer message.  We don't create it here,
691 	 * since the current CPU may not be the owner of this
692 	 * inpcb.
693 	 */
694 	tp->tt_msg = &it->inp_tp_timermsg;
695 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
696 
697 	if (tcp_do_rfc1323)
698 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
699 	tp->t_inpcb = inp;	/* XXX */
700 	tp->t_state = TCPS_CLOSED;
701 	/*
702 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
703 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
704 	 * reasonable initial retransmit time.
705 	 */
706 	tp->t_srtt = TCPTV_SRTTBASE;
707 	tp->t_rttvar =
708 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
709 	tp->t_rttmin = tcp_rexmit_min;
710 	tp->t_rxtcur = TCPTV_RTOBASE;
711 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
712 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
713 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
714 	tp->t_rcvtime = ticks;
715 	/*
716 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
717 	 * because the socket may be bound to an IPv6 wildcard address,
718 	 * which may match an IPv4-mapped IPv6 address.
719 	 */
720 	inp->inp_ip_ttl = ip_defttl;
721 	inp->inp_ppcb = tp;
722 	tcp_sack_tcpcb_init(tp);
723 	return (tp);		/* XXX */
724 }
725 
726 /*
727  * Drop a TCP connection, reporting the specified error.
728  * If connection is synchronized, then send a RST to peer.
729  */
730 struct tcpcb *
731 tcp_drop(struct tcpcb *tp, int error)
732 {
733 	struct socket *so = tp->t_inpcb->inp_socket;
734 
735 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
736 		tp->t_state = TCPS_CLOSED;
737 		tcp_output(tp);
738 		tcpstat.tcps_drops++;
739 	} else
740 		tcpstat.tcps_conndrops++;
741 	if (error == ETIMEDOUT && tp->t_softerror)
742 		error = tp->t_softerror;
743 	so->so_error = error;
744 	return (tcp_close(tp));
745 }
746 
747 #ifdef SMP
748 
749 struct netmsg_listen_detach {
750 	struct netmsg_base	base;
751 	struct tcpcb		*nm_tp;
752 };
753 
754 static void
755 tcp_listen_detach_handler(netmsg_t msg)
756 {
757 	struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
758 	struct tcpcb *tp = nmsg->nm_tp;
759 	int cpu = mycpuid, nextcpu;
760 
761 	if (tp->t_flags & TF_LISTEN)
762 		syncache_destroy(tp);
763 
764 	in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
765 
766 	nextcpu = cpu + 1;
767 	if (nextcpu < ncpus2)
768 		lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg);
769 	else
770 		lwkt_replymsg(&nmsg->base.lmsg, 0);
771 }
772 
773 #endif
774 
775 /*
776  * Close a TCP control block:
777  *	discard all space held by the tcp
778  *	discard internet protocol block
779  *	wake up any sleepers
780  */
781 struct tcpcb *
782 tcp_close(struct tcpcb *tp)
783 {
784 	struct tseg_qent *q;
785 	struct inpcb *inp = tp->t_inpcb;
786 	struct socket *so = inp->inp_socket;
787 	struct rtentry *rt;
788 	boolean_t dosavessthresh;
789 #ifdef INET6
790 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
791 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
792 #else
793 	const boolean_t isipv6 = FALSE;
794 #endif
795 
796 #ifdef SMP
797 	/*
798 	 * INP_WILDCARD_MP indicates that listen(2) has been called on
799 	 * this socket.  This implies:
800 	 * - A wildcard inp's hash is replicated for each protocol thread.
801 	 * - Syncache for this inp grows independently in each protocol
802 	 *   thread.
803 	 * - There is more than one cpu
804 	 *
805 	 * We have to chain a message to the rest of the protocol threads
806 	 * to cleanup the wildcard hash and the syncache.  The cleanup
807 	 * in the current protocol thread is defered till the end of this
808 	 * function.
809 	 *
810 	 * NOTE:
811 	 * After cleanup the inp's hash and syncache entries, this inp will
812 	 * no longer be available to the rest of the protocol threads, so we
813 	 * are safe to whack the inp in the following code.
814 	 */
815 	if (inp->inp_flags & INP_WILDCARD_MP) {
816 		struct netmsg_listen_detach nmsg;
817 
818 		KKASSERT(so->so_port == cpu_portfn(0));
819 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
820 		KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
821 
822 		netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
823 			    MSGF_PRIORITY, tcp_listen_detach_handler);
824 		nmsg.nm_tp = tp;
825 		lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0);
826 
827 		inp->inp_flags &= ~INP_WILDCARD_MP;
828 	}
829 #endif
830 
831 	KKASSERT(tp->t_state != TCPS_TERMINATING);
832 	tp->t_state = TCPS_TERMINATING;
833 
834 	/*
835 	 * Make sure that all of our timers are stopped before we
836 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
837 	 * timers are never used.  If timer message is never created
838 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
839 	 */
840 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
841 		tcp_callout_stop(tp, tp->tt_rexmt);
842 		tcp_callout_stop(tp, tp->tt_persist);
843 		tcp_callout_stop(tp, tp->tt_keep);
844 		tcp_callout_stop(tp, tp->tt_2msl);
845 		tcp_callout_stop(tp, tp->tt_delack);
846 	}
847 
848 	if (tp->t_flags & TF_ONOUTPUTQ) {
849 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
850 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
851 		tp->t_flags &= ~TF_ONOUTPUTQ;
852 	}
853 
854 	/*
855 	 * If we got enough samples through the srtt filter,
856 	 * save the rtt and rttvar in the routing entry.
857 	 * 'Enough' is arbitrarily defined as the 16 samples.
858 	 * 16 samples is enough for the srtt filter to converge
859 	 * to within 5% of the correct value; fewer samples and
860 	 * we could save a very bogus rtt.
861 	 *
862 	 * Don't update the default route's characteristics and don't
863 	 * update anything that the user "locked".
864 	 */
865 	if (tp->t_rttupdated >= 16) {
866 		u_long i = 0;
867 
868 		if (isipv6) {
869 			struct sockaddr_in6 *sin6;
870 
871 			if ((rt = inp->in6p_route.ro_rt) == NULL)
872 				goto no_valid_rt;
873 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
874 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
875 				goto no_valid_rt;
876 		} else
877 			if ((rt = inp->inp_route.ro_rt) == NULL ||
878 			    ((struct sockaddr_in *)rt_key(rt))->
879 			     sin_addr.s_addr == INADDR_ANY)
880 				goto no_valid_rt;
881 
882 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
883 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
884 			if (rt->rt_rmx.rmx_rtt && i)
885 				/*
886 				 * filter this update to half the old & half
887 				 * the new values, converting scale.
888 				 * See route.h and tcp_var.h for a
889 				 * description of the scaling constants.
890 				 */
891 				rt->rt_rmx.rmx_rtt =
892 				    (rt->rt_rmx.rmx_rtt + i) / 2;
893 			else
894 				rt->rt_rmx.rmx_rtt = i;
895 			tcpstat.tcps_cachedrtt++;
896 		}
897 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
898 			i = tp->t_rttvar *
899 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
900 			if (rt->rt_rmx.rmx_rttvar && i)
901 				rt->rt_rmx.rmx_rttvar =
902 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
903 			else
904 				rt->rt_rmx.rmx_rttvar = i;
905 			tcpstat.tcps_cachedrttvar++;
906 		}
907 		/*
908 		 * The old comment here said:
909 		 * update the pipelimit (ssthresh) if it has been updated
910 		 * already or if a pipesize was specified & the threshhold
911 		 * got below half the pipesize.  I.e., wait for bad news
912 		 * before we start updating, then update on both good
913 		 * and bad news.
914 		 *
915 		 * But we want to save the ssthresh even if no pipesize is
916 		 * specified explicitly in the route, because such
917 		 * connections still have an implicit pipesize specified
918 		 * by the global tcp_sendspace.  In the absence of a reliable
919 		 * way to calculate the pipesize, it will have to do.
920 		 */
921 		i = tp->snd_ssthresh;
922 		if (rt->rt_rmx.rmx_sendpipe != 0)
923 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
924 		else
925 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
926 		if (dosavessthresh ||
927 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
928 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
929 			/*
930 			 * convert the limit from user data bytes to
931 			 * packets then to packet data bytes.
932 			 */
933 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
934 			if (i < 2)
935 				i = 2;
936 			i *= tp->t_maxseg +
937 			     (isipv6 ?
938 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
939 			      sizeof(struct tcpiphdr));
940 			if (rt->rt_rmx.rmx_ssthresh)
941 				rt->rt_rmx.rmx_ssthresh =
942 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
943 			else
944 				rt->rt_rmx.rmx_ssthresh = i;
945 			tcpstat.tcps_cachedssthresh++;
946 		}
947 	}
948 
949 no_valid_rt:
950 	/* free the reassembly queue, if any */
951 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
952 		LIST_REMOVE(q, tqe_q);
953 		m_freem(q->tqe_m);
954 		FREE(q, M_TSEGQ);
955 		atomic_add_int(&tcp_reass_qsize, -1);
956 	}
957 	/* throw away SACK blocks in scoreboard*/
958 	if (TCP_DO_SACK(tp))
959 		tcp_sack_cleanup(&tp->scb);
960 
961 	inp->inp_ppcb = NULL;
962 	soisdisconnected(so);
963 	/* note: pcb detached later on */
964 
965 	tcp_destroy_timermsg(tp);
966 
967 	if (tp->t_flags & TF_LISTEN)
968 		syncache_destroy(tp);
969 
970 	/*
971 	 * NOTE:
972 	 * pcbdetach removes any wildcard hash entry on the current CPU.
973 	 */
974 #ifdef INET6
975 	if (isafinet6)
976 		in6_pcbdetach(inp);
977 	else
978 #endif
979 		in_pcbdetach(inp);
980 
981 	tcpstat.tcps_closed++;
982 	return (NULL);
983 }
984 
985 static __inline void
986 tcp_drain_oncpu(struct inpcbhead *head)
987 {
988 	struct inpcb *marker;
989 	struct inpcb *inpb;
990 	struct tcpcb *tcpb;
991 	struct tseg_qent *te;
992 
993 	/*
994 	 * Allows us to block while running the list
995 	 */
996 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
997 	marker->inp_flags |= INP_PLACEMARKER;
998 	LIST_INSERT_HEAD(head, marker, inp_list);
999 
1000 	while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1001 		if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1002 		    (tcpb = intotcpcb(inpb)) != NULL &&
1003 		    (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1004 			LIST_REMOVE(te, tqe_q);
1005 			m_freem(te->tqe_m);
1006 			FREE(te, M_TSEGQ);
1007 			atomic_add_int(&tcp_reass_qsize, -1);
1008 			/* retry */
1009 		} else {
1010 			LIST_REMOVE(marker, inp_list);
1011 			LIST_INSERT_AFTER(inpb, marker, inp_list);
1012 		}
1013 	}
1014 	LIST_REMOVE(marker, inp_list);
1015 	kfree(marker, M_TEMP);
1016 }
1017 
1018 #ifdef SMP
1019 struct netmsg_tcp_drain {
1020 	struct netmsg_base	base;
1021 	struct inpcbhead	*nm_head;
1022 };
1023 
1024 static void
1025 tcp_drain_handler(netmsg_t msg)
1026 {
1027 	struct netmsg_tcp_drain *nm = (void *)msg;
1028 
1029 	tcp_drain_oncpu(nm->nm_head);
1030 	lwkt_replymsg(&nm->base.lmsg, 0);
1031 }
1032 #endif
1033 
1034 void
1035 tcp_drain(void)
1036 {
1037 #ifdef SMP
1038 	int cpu;
1039 #endif
1040 
1041 	if (!do_tcpdrain)
1042 		return;
1043 
1044 	/*
1045 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1046 	 * if there is one...
1047 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1048 	 *	reassembly queue should be flushed, but in a situation
1049 	 *	where we're really low on mbufs, this is potentially
1050 	 *	useful.
1051 	 */
1052 #ifdef SMP
1053 	for (cpu = 0; cpu < ncpus2; cpu++) {
1054 		struct netmsg_tcp_drain *nm;
1055 
1056 		if (cpu == mycpu->gd_cpuid) {
1057 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1058 		} else {
1059 			nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1060 				     M_LWKTMSG, M_NOWAIT);
1061 			if (nm == NULL)
1062 				continue;
1063 			netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1064 				    0, tcp_drain_handler);
1065 			nm->nm_head = &tcbinfo[cpu].pcblisthead;
1066 			lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg);
1067 		}
1068 	}
1069 #else
1070 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1071 #endif
1072 }
1073 
1074 /*
1075  * Notify a tcp user of an asynchronous error;
1076  * store error as soft error, but wake up user
1077  * (for now, won't do anything until can select for soft error).
1078  *
1079  * Do not wake up user since there currently is no mechanism for
1080  * reporting soft errors (yet - a kqueue filter may be added).
1081  */
1082 static void
1083 tcp_notify(struct inpcb *inp, int error)
1084 {
1085 	struct tcpcb *tp = intotcpcb(inp);
1086 
1087 	/*
1088 	 * Ignore some errors if we are hooked up.
1089 	 * If connection hasn't completed, has retransmitted several times,
1090 	 * and receives a second error, give up now.  This is better
1091 	 * than waiting a long time to establish a connection that
1092 	 * can never complete.
1093 	 */
1094 	if (tp->t_state == TCPS_ESTABLISHED &&
1095 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1096 	      error == EHOSTDOWN)) {
1097 		return;
1098 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1099 	    tp->t_softerror)
1100 		tcp_drop(tp, error);
1101 	else
1102 		tp->t_softerror = error;
1103 #if 0
1104 	wakeup(&so->so_timeo);
1105 	sorwakeup(so);
1106 	sowwakeup(so);
1107 #endif
1108 }
1109 
1110 static int
1111 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1112 {
1113 	int error, i, n;
1114 	struct inpcb *marker;
1115 	struct inpcb *inp;
1116 	globaldata_t gd;
1117 	int origcpu, ccpu;
1118 
1119 	error = 0;
1120 	n = 0;
1121 
1122 	/*
1123 	 * The process of preparing the TCB list is too time-consuming and
1124 	 * resource-intensive to repeat twice on every request.
1125 	 */
1126 	if (req->oldptr == NULL) {
1127 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1128 			gd = globaldata_find(ccpu);
1129 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1130 		}
1131 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1132 		return (0);
1133 	}
1134 
1135 	if (req->newptr != NULL)
1136 		return (EPERM);
1137 
1138 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1139 	marker->inp_flags |= INP_PLACEMARKER;
1140 
1141 	/*
1142 	 * OK, now we're committed to doing something.  Run the inpcb list
1143 	 * for each cpu in the system and construct the output.  Use a
1144 	 * list placemarker to deal with list changes occuring during
1145 	 * copyout blockages (but otherwise depend on being on the correct
1146 	 * cpu to avoid races).
1147 	 */
1148 	origcpu = mycpu->gd_cpuid;
1149 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1150 		globaldata_t rgd;
1151 		caddr_t inp_ppcb;
1152 		struct xtcpcb xt;
1153 		int cpu_id;
1154 
1155 		cpu_id = (origcpu + ccpu) % ncpus;
1156 		if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1157 			continue;
1158 		rgd = globaldata_find(cpu_id);
1159 		lwkt_setcpu_self(rgd);
1160 
1161 		n = tcbinfo[cpu_id].ipi_count;
1162 
1163 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1164 		i = 0;
1165 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1166 			/*
1167 			 * process a snapshot of pcbs, ignoring placemarkers
1168 			 * and using our own to allow SYSCTL_OUT to block.
1169 			 */
1170 			LIST_REMOVE(marker, inp_list);
1171 			LIST_INSERT_AFTER(inp, marker, inp_list);
1172 
1173 			if (inp->inp_flags & INP_PLACEMARKER)
1174 				continue;
1175 			if (prison_xinpcb(req->td, inp))
1176 				continue;
1177 
1178 			xt.xt_len = sizeof xt;
1179 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1180 			inp_ppcb = inp->inp_ppcb;
1181 			if (inp_ppcb != NULL)
1182 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1183 			else
1184 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1185 			if (inp->inp_socket)
1186 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1187 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1188 				break;
1189 			++i;
1190 		}
1191 		LIST_REMOVE(marker, inp_list);
1192 		if (error == 0 && i < n) {
1193 			bzero(&xt, sizeof xt);
1194 			xt.xt_len = sizeof xt;
1195 			while (i < n) {
1196 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1197 				if (error)
1198 					break;
1199 				++i;
1200 			}
1201 		}
1202 	}
1203 
1204 	/*
1205 	 * Make sure we are on the same cpu we were on originally, since
1206 	 * higher level callers expect this.  Also don't pollute caches with
1207 	 * migrated userland data by (eventually) returning to userland
1208 	 * on a different cpu.
1209 	 */
1210 	lwkt_setcpu_self(globaldata_find(origcpu));
1211 	kfree(marker, M_TEMP);
1212 	return (error);
1213 }
1214 
1215 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1216 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1217 
1218 static int
1219 tcp_getcred(SYSCTL_HANDLER_ARGS)
1220 {
1221 	struct sockaddr_in addrs[2];
1222 	struct inpcb *inp;
1223 	int cpu;
1224 	int error;
1225 
1226 	error = priv_check(req->td, PRIV_ROOT);
1227 	if (error != 0)
1228 		return (error);
1229 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1230 	if (error != 0)
1231 		return (error);
1232 	crit_enter();
1233 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1234 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1235 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1236 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1237 	if (inp == NULL || inp->inp_socket == NULL) {
1238 		error = ENOENT;
1239 		goto out;
1240 	}
1241 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1242 out:
1243 	crit_exit();
1244 	return (error);
1245 }
1246 
1247 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1248     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1249 
1250 #ifdef INET6
1251 static int
1252 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1253 {
1254 	struct sockaddr_in6 addrs[2];
1255 	struct inpcb *inp;
1256 	int error;
1257 	boolean_t mapped = FALSE;
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 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1266 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1267 			mapped = TRUE;
1268 		else
1269 			return (EINVAL);
1270 	}
1271 	crit_enter();
1272 	if (mapped) {
1273 		inp = in_pcblookup_hash(&tcbinfo[0],
1274 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1275 		    addrs[1].sin6_port,
1276 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1277 		    addrs[0].sin6_port,
1278 		    0, NULL);
1279 	} else {
1280 		inp = in6_pcblookup_hash(&tcbinfo[0],
1281 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1282 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1283 		    0, NULL);
1284 	}
1285 	if (inp == NULL || inp->inp_socket == NULL) {
1286 		error = ENOENT;
1287 		goto out;
1288 	}
1289 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1290 out:
1291 	crit_exit();
1292 	return (error);
1293 }
1294 
1295 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1296 	    0, 0,
1297 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1298 #endif
1299 
1300 struct netmsg_tcp_notify {
1301 	struct netmsg_base base;
1302 	void		(*nm_notify)(struct inpcb *, int);
1303 	struct in_addr	nm_faddr;
1304 	int		nm_arg;
1305 };
1306 
1307 static void
1308 tcp_notifyall_oncpu(netmsg_t msg)
1309 {
1310 	struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1311 	int nextcpu;
1312 
1313 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1314 			nm->nm_arg, nm->nm_notify);
1315 
1316 	nextcpu = mycpuid + 1;
1317 	if (nextcpu < ncpus2)
1318 		lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg);
1319 	else
1320 		lwkt_replymsg(&nm->base.lmsg, 0);
1321 }
1322 
1323 void
1324 tcp_ctlinput(netmsg_t msg)
1325 {
1326 	int cmd = msg->ctlinput.nm_cmd;
1327 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1328 	struct ip *ip = msg->ctlinput.nm_extra;
1329 	struct tcphdr *th;
1330 	struct in_addr faddr;
1331 	struct inpcb *inp;
1332 	struct tcpcb *tp;
1333 	void (*notify)(struct inpcb *, int) = tcp_notify;
1334 	tcp_seq icmpseq;
1335 	int arg, cpu;
1336 
1337 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1338 		goto done;
1339 	}
1340 
1341 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1342 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1343 		goto done;
1344 
1345 	arg = inetctlerrmap[cmd];
1346 	if (cmd == PRC_QUENCH) {
1347 		notify = tcp_quench;
1348 	} else if (icmp_may_rst &&
1349 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1350 		    cmd == PRC_UNREACH_PORT ||
1351 		    cmd == PRC_TIMXCEED_INTRANS) &&
1352 		   ip != NULL) {
1353 		notify = tcp_drop_syn_sent;
1354 	} else if (cmd == PRC_MSGSIZE) {
1355 		struct icmp *icmp = (struct icmp *)
1356 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1357 
1358 		arg = ntohs(icmp->icmp_nextmtu);
1359 		notify = tcp_mtudisc;
1360 	} else if (PRC_IS_REDIRECT(cmd)) {
1361 		ip = NULL;
1362 		notify = in_rtchange;
1363 	} else if (cmd == PRC_HOSTDEAD) {
1364 		ip = NULL;
1365 	}
1366 
1367 	if (ip != NULL) {
1368 		crit_enter();
1369 		th = (struct tcphdr *)((caddr_t)ip +
1370 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1371 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1372 				  ip->ip_src.s_addr, th->th_sport);
1373 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1374 					ip->ip_src, th->th_sport, 0, NULL);
1375 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1376 			icmpseq = htonl(th->th_seq);
1377 			tp = intotcpcb(inp);
1378 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1379 			    SEQ_LT(icmpseq, tp->snd_max))
1380 				(*notify)(inp, arg);
1381 		} else {
1382 			struct in_conninfo inc;
1383 
1384 			inc.inc_fport = th->th_dport;
1385 			inc.inc_lport = th->th_sport;
1386 			inc.inc_faddr = faddr;
1387 			inc.inc_laddr = ip->ip_src;
1388 #ifdef INET6
1389 			inc.inc_isipv6 = 0;
1390 #endif
1391 			syncache_unreach(&inc, th);
1392 		}
1393 		crit_exit();
1394 	} else {
1395 		struct netmsg_tcp_notify *nm;
1396 
1397 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1398 		nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1399 		netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1400 			    0, tcp_notifyall_oncpu);
1401 		nm->nm_faddr = faddr;
1402 		nm->nm_arg = arg;
1403 		nm->nm_notify = notify;
1404 
1405 		lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg);
1406 	}
1407 done:
1408 	lwkt_replymsg(&msg->lmsg, 0);
1409 }
1410 
1411 #ifdef INET6
1412 
1413 void
1414 tcp6_ctlinput(netmsg_t msg)
1415 {
1416 	int cmd = msg->ctlinput.nm_cmd;
1417 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1418 	void *d = msg->ctlinput.nm_extra;
1419 	struct tcphdr th;
1420 	void (*notify) (struct inpcb *, int) = tcp_notify;
1421 	struct ip6_hdr *ip6;
1422 	struct mbuf *m;
1423 	struct ip6ctlparam *ip6cp = NULL;
1424 	const struct sockaddr_in6 *sa6_src = NULL;
1425 	int off;
1426 	struct tcp_portonly {
1427 		u_int16_t th_sport;
1428 		u_int16_t th_dport;
1429 	} *thp;
1430 	int arg;
1431 
1432 	if (sa->sa_family != AF_INET6 ||
1433 	    sa->sa_len != sizeof(struct sockaddr_in6)) {
1434 		goto out;
1435 	}
1436 
1437 	arg = 0;
1438 	if (cmd == PRC_QUENCH)
1439 		notify = tcp_quench;
1440 	else if (cmd == PRC_MSGSIZE) {
1441 		struct ip6ctlparam *ip6cp = d;
1442 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1443 
1444 		arg = ntohl(icmp6->icmp6_mtu);
1445 		notify = tcp_mtudisc;
1446 	} else if (!PRC_IS_REDIRECT(cmd) &&
1447 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1448 		goto out;
1449 	}
1450 
1451 	/* if the parameter is from icmp6, decode it. */
1452 	if (d != NULL) {
1453 		ip6cp = (struct ip6ctlparam *)d;
1454 		m = ip6cp->ip6c_m;
1455 		ip6 = ip6cp->ip6c_ip6;
1456 		off = ip6cp->ip6c_off;
1457 		sa6_src = ip6cp->ip6c_src;
1458 	} else {
1459 		m = NULL;
1460 		ip6 = NULL;
1461 		off = 0;	/* fool gcc */
1462 		sa6_src = &sa6_any;
1463 	}
1464 
1465 	if (ip6 != NULL) {
1466 		struct in_conninfo inc;
1467 		/*
1468 		 * XXX: We assume that when IPV6 is non NULL,
1469 		 * M and OFF are valid.
1470 		 */
1471 
1472 		/* check if we can safely examine src and dst ports */
1473 		if (m->m_pkthdr.len < off + sizeof *thp)
1474 			goto out;
1475 
1476 		bzero(&th, sizeof th);
1477 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1478 
1479 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1480 		    (struct sockaddr *)ip6cp->ip6c_src,
1481 		    th.th_sport, cmd, arg, notify);
1482 
1483 		inc.inc_fport = th.th_dport;
1484 		inc.inc_lport = th.th_sport;
1485 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1486 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1487 		inc.inc_isipv6 = 1;
1488 		syncache_unreach(&inc, &th);
1489 	} else {
1490 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1491 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1492 	}
1493 out:
1494 	lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1495 }
1496 
1497 #endif
1498 
1499 /*
1500  * Following is where TCP initial sequence number generation occurs.
1501  *
1502  * There are two places where we must use initial sequence numbers:
1503  * 1.  In SYN-ACK packets.
1504  * 2.  In SYN packets.
1505  *
1506  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1507  * tcp_syncache.c for details.
1508  *
1509  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1510  * depends on this property.  In addition, these ISNs should be
1511  * unguessable so as to prevent connection hijacking.  To satisfy
1512  * the requirements of this situation, the algorithm outlined in
1513  * RFC 1948 is used to generate sequence numbers.
1514  *
1515  * Implementation details:
1516  *
1517  * Time is based off the system timer, and is corrected so that it
1518  * increases by one megabyte per second.  This allows for proper
1519  * recycling on high speed LANs while still leaving over an hour
1520  * before rollover.
1521  *
1522  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1523  * between seeding of isn_secret.  This is normally set to zero,
1524  * as reseeding should not be necessary.
1525  *
1526  */
1527 
1528 #define	ISN_BYTES_PER_SECOND 1048576
1529 
1530 u_char isn_secret[32];
1531 int isn_last_reseed;
1532 MD5_CTX isn_ctx;
1533 
1534 tcp_seq
1535 tcp_new_isn(struct tcpcb *tp)
1536 {
1537 	u_int32_t md5_buffer[4];
1538 	tcp_seq new_isn;
1539 
1540 	/* Seed if this is the first use, reseed if requested. */
1541 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1542 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1543 		< (u_int)ticks))) {
1544 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1545 		isn_last_reseed = ticks;
1546 	}
1547 
1548 	/* Compute the md5 hash and return the ISN. */
1549 	MD5Init(&isn_ctx);
1550 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1551 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1552 #ifdef INET6
1553 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1554 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1555 			  sizeof(struct in6_addr));
1556 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1557 			  sizeof(struct in6_addr));
1558 	} else
1559 #endif
1560 	{
1561 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1562 			  sizeof(struct in_addr));
1563 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1564 			  sizeof(struct in_addr));
1565 	}
1566 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1567 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1568 	new_isn = (tcp_seq) md5_buffer[0];
1569 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1570 	return (new_isn);
1571 }
1572 
1573 /*
1574  * When a source quench is received, close congestion window
1575  * to one segment.  We will gradually open it again as we proceed.
1576  */
1577 void
1578 tcp_quench(struct inpcb *inp, int error)
1579 {
1580 	struct tcpcb *tp = intotcpcb(inp);
1581 
1582 	if (tp != NULL) {
1583 		tp->snd_cwnd = tp->t_maxseg;
1584 		tp->snd_wacked = 0;
1585 	}
1586 }
1587 
1588 /*
1589  * When a specific ICMP unreachable message is received and the
1590  * connection state is SYN-SENT, drop the connection.  This behavior
1591  * is controlled by the icmp_may_rst sysctl.
1592  */
1593 void
1594 tcp_drop_syn_sent(struct inpcb *inp, int error)
1595 {
1596 	struct tcpcb *tp = intotcpcb(inp);
1597 
1598 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1599 		tcp_drop(tp, error);
1600 }
1601 
1602 /*
1603  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1604  * based on the new value in the route.  Also nudge TCP to send something,
1605  * since we know the packet we just sent was dropped.
1606  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1607  */
1608 void
1609 tcp_mtudisc(struct inpcb *inp, int mtu)
1610 {
1611 	struct tcpcb *tp = intotcpcb(inp);
1612 	struct rtentry *rt;
1613 	struct socket *so = inp->inp_socket;
1614 	int maxopd, mss;
1615 #ifdef INET6
1616 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1617 #else
1618 	const boolean_t isipv6 = FALSE;
1619 #endif
1620 
1621 	if (tp == NULL)
1622 		return;
1623 
1624 	/*
1625 	 * If no MTU is provided in the ICMP message, use the
1626 	 * next lower likely value, as specified in RFC 1191.
1627 	 */
1628 	if (mtu == 0) {
1629 		int oldmtu;
1630 
1631 		oldmtu = tp->t_maxopd +
1632 		    (isipv6 ?
1633 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1634 		     sizeof(struct tcpiphdr));
1635 		mtu = ip_next_mtu(oldmtu, 0);
1636 	}
1637 
1638 	if (isipv6)
1639 		rt = tcp_rtlookup6(&inp->inp_inc);
1640 	else
1641 		rt = tcp_rtlookup(&inp->inp_inc);
1642 	if (rt != NULL) {
1643 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1644 			mtu = rt->rt_rmx.rmx_mtu;
1645 
1646 		maxopd = mtu -
1647 		    (isipv6 ?
1648 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1649 		     sizeof(struct tcpiphdr));
1650 
1651 		/*
1652 		 * XXX - The following conditional probably violates the TCP
1653 		 * spec.  The problem is that, since we don't know the
1654 		 * other end's MSS, we are supposed to use a conservative
1655 		 * default.  But, if we do that, then MTU discovery will
1656 		 * never actually take place, because the conservative
1657 		 * default is much less than the MTUs typically seen
1658 		 * on the Internet today.  For the moment, we'll sweep
1659 		 * this under the carpet.
1660 		 *
1661 		 * The conservative default might not actually be a problem
1662 		 * if the only case this occurs is when sending an initial
1663 		 * SYN with options and data to a host we've never talked
1664 		 * to before.  Then, they will reply with an MSS value which
1665 		 * will get recorded and the new parameters should get
1666 		 * recomputed.  For Further Study.
1667 		 */
1668 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1669 			maxopd = rt->rt_rmx.rmx_mssopt;
1670 	} else
1671 		maxopd = mtu -
1672 		    (isipv6 ?
1673 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1674 		     sizeof(struct tcpiphdr));
1675 
1676 	if (tp->t_maxopd <= maxopd)
1677 		return;
1678 	tp->t_maxopd = maxopd;
1679 
1680 	mss = maxopd;
1681 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1682 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1683 		mss -= TCPOLEN_TSTAMP_APPA;
1684 
1685 	/* round down to multiple of MCLBYTES */
1686 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1687 	if (mss > MCLBYTES)
1688 		mss &= ~(MCLBYTES - 1);
1689 #else
1690 	if (mss > MCLBYTES)
1691 		mss = (mss / MCLBYTES) * MCLBYTES;
1692 #endif
1693 
1694 	if (so->so_snd.ssb_hiwat < mss)
1695 		mss = so->so_snd.ssb_hiwat;
1696 
1697 	tp->t_maxseg = mss;
1698 	tp->t_rtttime = 0;
1699 	tp->snd_nxt = tp->snd_una;
1700 	tcp_output(tp);
1701 	tcpstat.tcps_mturesent++;
1702 }
1703 
1704 /*
1705  * Look-up the routing entry to the peer of this inpcb.  If no route
1706  * is found and it cannot be allocated the return NULL.  This routine
1707  * is called by TCP routines that access the rmx structure and by tcp_mss
1708  * to get the interface MTU.
1709  */
1710 struct rtentry *
1711 tcp_rtlookup(struct in_conninfo *inc)
1712 {
1713 	struct route *ro = &inc->inc_route;
1714 
1715 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1716 		/* No route yet, so try to acquire one */
1717 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1718 			/*
1719 			 * unused portions of the structure MUST be zero'd
1720 			 * out because rtalloc() treats it as opaque data
1721 			 */
1722 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1723 			ro->ro_dst.sa_family = AF_INET;
1724 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1725 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1726 			    inc->inc_faddr;
1727 			rtalloc(ro);
1728 		}
1729 	}
1730 	return (ro->ro_rt);
1731 }
1732 
1733 #ifdef INET6
1734 struct rtentry *
1735 tcp_rtlookup6(struct in_conninfo *inc)
1736 {
1737 	struct route_in6 *ro6 = &inc->inc6_route;
1738 
1739 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1740 		/* No route yet, so try to acquire one */
1741 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1742 			/*
1743 			 * unused portions of the structure MUST be zero'd
1744 			 * out because rtalloc() treats it as opaque data
1745 			 */
1746 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1747 			ro6->ro_dst.sin6_family = AF_INET6;
1748 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1749 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1750 			rtalloc((struct route *)ro6);
1751 		}
1752 	}
1753 	return (ro6->ro_rt);
1754 }
1755 #endif
1756 
1757 #ifdef IPSEC
1758 /* compute ESP/AH header size for TCP, including outer IP header. */
1759 size_t
1760 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1761 {
1762 	struct inpcb *inp;
1763 	struct mbuf *m;
1764 	size_t hdrsiz;
1765 	struct ip *ip;
1766 	struct tcphdr *th;
1767 
1768 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1769 		return (0);
1770 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1771 	if (!m)
1772 		return (0);
1773 
1774 #ifdef INET6
1775 	if (inp->inp_vflag & INP_IPV6) {
1776 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1777 
1778 		th = (struct tcphdr *)(ip6 + 1);
1779 		m->m_pkthdr.len = m->m_len =
1780 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1781 		tcp_fillheaders(tp, ip6, th);
1782 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1783 	} else
1784 #endif
1785 	{
1786 		ip = mtod(m, struct ip *);
1787 		th = (struct tcphdr *)(ip + 1);
1788 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1789 		tcp_fillheaders(tp, ip, th);
1790 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1791 	}
1792 
1793 	m_free(m);
1794 	return (hdrsiz);
1795 }
1796 #endif
1797 
1798 /*
1799  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1800  *
1801  * This code attempts to calculate the bandwidth-delay product as a
1802  * means of determining the optimal window size to maximize bandwidth,
1803  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1804  * routers.  This code also does a fairly good job keeping RTTs in check
1805  * across slow links like modems.  We implement an algorithm which is very
1806  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1807  * transmitter side of a TCP connection and so only effects the transmit
1808  * side of the connection.
1809  *
1810  * BACKGROUND:  TCP makes no provision for the management of buffer space
1811  * at the end points or at the intermediate routers and switches.  A TCP
1812  * stream, whether using NewReno or not, will eventually buffer as
1813  * many packets as it is able and the only reason this typically works is
1814  * due to the fairly small default buffers made available for a connection
1815  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1816  * scaling it is now fairly easy for even a single TCP connection to blow-out
1817  * all available buffer space not only on the local interface, but on
1818  * intermediate routers and switches as well.  NewReno makes a misguided
1819  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1820  * then backing off, then steadily increasing the window again until another
1821  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1822  * is only made worse as network loads increase and the idea of intentionally
1823  * blowing out network buffers is, frankly, a terrible way to manage network
1824  * resources.
1825  *
1826  * It is far better to limit the transmit window prior to the failure
1827  * condition being achieved.  There are two general ways to do this:  First
1828  * you can 'scan' through different transmit window sizes and locate the
1829  * point where the RTT stops increasing, indicating that you have filled the
1830  * pipe, then scan backwards until you note that RTT stops decreasing, then
1831  * repeat ad-infinitum.  This method works in principle but has severe
1832  * implementation issues due to RTT variances, timer granularity, and
1833  * instability in the algorithm which can lead to many false positives and
1834  * create oscillations as well as interact badly with other TCP streams
1835  * implementing the same algorithm.
1836  *
1837  * The second method is to limit the window to the bandwidth delay product
1838  * of the link.  This is the method we implement.  RTT variances and our
1839  * own manipulation of the congestion window, bwnd, can potentially
1840  * destabilize the algorithm.  For this reason we have to stabilize the
1841  * elements used to calculate the window.  We do this by using the minimum
1842  * observed RTT, the long term average of the observed bandwidth, and
1843  * by adding two segments worth of slop.  It isn't perfect but it is able
1844  * to react to changing conditions and gives us a very stable basis on
1845  * which to extend the algorithm.
1846  */
1847 void
1848 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1849 {
1850 	u_long bw;
1851 	u_long bwnd;
1852 	int save_ticks;
1853 	int delta_ticks;
1854 
1855 	/*
1856 	 * If inflight_enable is disabled in the middle of a tcp connection,
1857 	 * make sure snd_bwnd is effectively disabled.
1858 	 */
1859 	if (!tcp_inflight_enable) {
1860 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1861 		tp->snd_bandwidth = 0;
1862 		return;
1863 	}
1864 
1865 	/*
1866 	 * Validate the delta time.  If a connection is new or has been idle
1867 	 * a long time we have to reset the bandwidth calculator.
1868 	 */
1869 	save_ticks = ticks;
1870 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1871 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1872 		tp->t_bw_rtttime = ticks;
1873 		tp->t_bw_rtseq = ack_seq;
1874 		if (tp->snd_bandwidth == 0)
1875 			tp->snd_bandwidth = tcp_inflight_min;
1876 		return;
1877 	}
1878 	if (delta_ticks == 0)
1879 		return;
1880 
1881 	/*
1882 	 * Sanity check, plus ignore pure window update acks.
1883 	 */
1884 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1885 		return;
1886 
1887 	/*
1888 	 * Figure out the bandwidth.  Due to the tick granularity this
1889 	 * is a very rough number and it MUST be averaged over a fairly
1890 	 * long period of time.  XXX we need to take into account a link
1891 	 * that is not using all available bandwidth, but for now our
1892 	 * slop will ramp us up if this case occurs and the bandwidth later
1893 	 * increases.
1894 	 */
1895 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1896 	tp->t_bw_rtttime = save_ticks;
1897 	tp->t_bw_rtseq = ack_seq;
1898 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1899 
1900 	tp->snd_bandwidth = bw;
1901 
1902 	/*
1903 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1904 	 * segments.  The additional slop puts us squarely in the sweet
1905 	 * spot and also handles the bandwidth run-up case.  Without the
1906 	 * slop we could be locking ourselves into a lower bandwidth.
1907 	 *
1908 	 * Situations Handled:
1909 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1910 	 *	    high speed LANs, allowing larger TCP buffers to be
1911 	 *	    specified, and also does a good job preventing
1912 	 *	    over-queueing of packets over choke points like modems
1913 	 *	    (at least for the transmit side).
1914 	 *
1915 	 *	(2) Is able to handle changing network loads (bandwidth
1916 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1917 	 *	    increases).
1918 	 *
1919 	 *	(3) Theoretically should stabilize in the face of multiple
1920 	 *	    connections implementing the same algorithm (this may need
1921 	 *	    a little work).
1922 	 *
1923 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1924 	 *	    be adjusted with a sysctl but typically only needs to be on
1925 	 *	    very slow connections.  A value no smaller then 5 should
1926 	 *	    be used, but only reduce this default if you have no other
1927 	 *	    choice.
1928 	 */
1929 
1930 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1931 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1932 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1933 #undef USERTT
1934 
1935 	if (tcp_inflight_debug > 0) {
1936 		static int ltime;
1937 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1938 			ltime = ticks;
1939 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1940 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1941 		}
1942 	}
1943 	if ((long)bwnd < tcp_inflight_min)
1944 		bwnd = tcp_inflight_min;
1945 	if (bwnd > tcp_inflight_max)
1946 		bwnd = tcp_inflight_max;
1947 	if ((long)bwnd < tp->t_maxseg * 2)
1948 		bwnd = tp->t_maxseg * 2;
1949 	tp->snd_bwnd = bwnd;
1950 }
1951 
1952 #ifdef TCP_SIGNATURE
1953 /*
1954  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1955  *
1956  * We do this over ip, tcphdr, segment data, and the key in the SADB.
1957  * When called from tcp_input(), we can be sure that th_sum has been
1958  * zeroed out and verified already.
1959  *
1960  * Return 0 if successful, otherwise return -1.
1961  *
1962  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1963  * search with the destination IP address, and a 'magic SPI' to be
1964  * determined by the application. This is hardcoded elsewhere to 1179
1965  * right now. Another branch of this code exists which uses the SPD to
1966  * specify per-application flows but it is unstable.
1967  */
1968 int
1969 tcpsignature_compute(
1970 	struct mbuf *m,		/* mbuf chain */
1971 	int len,		/* length of TCP data */
1972 	int optlen,		/* length of TCP options */
1973 	u_char *buf,		/* storage for MD5 digest */
1974 	u_int direction)	/* direction of flow */
1975 {
1976 	struct ippseudo ippseudo;
1977 	MD5_CTX ctx;
1978 	int doff;
1979 	struct ip *ip;
1980 	struct ipovly *ipovly;
1981 	struct secasvar *sav;
1982 	struct tcphdr *th;
1983 #ifdef INET6
1984 	struct ip6_hdr *ip6;
1985 	struct in6_addr in6;
1986 	uint32_t plen;
1987 	uint16_t nhdr;
1988 #endif /* INET6 */
1989 	u_short savecsum;
1990 
1991 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
1992 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
1993 	/*
1994 	 * Extract the destination from the IP header in the mbuf.
1995 	 */
1996 	ip = mtod(m, struct ip *);
1997 #ifdef INET6
1998 	ip6 = NULL;     /* Make the compiler happy. */
1999 #endif /* INET6 */
2000 	/*
2001 	 * Look up an SADB entry which matches the address found in
2002 	 * the segment.
2003 	 */
2004 	switch (IP_VHL_V(ip->ip_vhl)) {
2005 	case IPVERSION:
2006 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2007 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2008 		break;
2009 #ifdef INET6
2010 	case (IPV6_VERSION >> 4):
2011 		ip6 = mtod(m, struct ip6_hdr *);
2012 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2013 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2014 		break;
2015 #endif /* INET6 */
2016 	default:
2017 		return (EINVAL);
2018 		/* NOTREACHED */
2019 		break;
2020 	}
2021 	if (sav == NULL) {
2022 		kprintf("%s: SADB lookup failed\n", __func__);
2023 		return (EINVAL);
2024 	}
2025 	MD5Init(&ctx);
2026 
2027 	/*
2028 	 * Step 1: Update MD5 hash with IP pseudo-header.
2029 	 *
2030 	 * XXX The ippseudo header MUST be digested in network byte order,
2031 	 * or else we'll fail the regression test. Assume all fields we've
2032 	 * been doing arithmetic on have been in host byte order.
2033 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2034 	 * tcp_output(), the underlying ip_len member has not yet been set.
2035 	 */
2036 	switch (IP_VHL_V(ip->ip_vhl)) {
2037 	case IPVERSION:
2038 		ipovly = (struct ipovly *)ip;
2039 		ippseudo.ippseudo_src = ipovly->ih_src;
2040 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2041 		ippseudo.ippseudo_pad = 0;
2042 		ippseudo.ippseudo_p = IPPROTO_TCP;
2043 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2044 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2045 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2046 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2047 		break;
2048 #ifdef INET6
2049 	/*
2050 	 * RFC 2385, 2.0  Proposal
2051 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2052 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2053 	 * extended next header value (to form 32 bits), and 32-bit segment
2054 	 * length.
2055 	 * Note: Upper-Layer Packet Length comes before Next Header.
2056 	 */
2057 	case (IPV6_VERSION >> 4):
2058 		in6 = ip6->ip6_src;
2059 		in6_clearscope(&in6);
2060 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2061 		in6 = ip6->ip6_dst;
2062 		in6_clearscope(&in6);
2063 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2064 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2065 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2066 		nhdr = 0;
2067 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2068 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2069 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2070 		nhdr = IPPROTO_TCP;
2071 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2072 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2073 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2074 		break;
2075 #endif /* INET6 */
2076 	default:
2077 		return (EINVAL);
2078 		/* NOTREACHED */
2079 		break;
2080 	}
2081 	/*
2082 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2083 	 * The TCP checksum must be set to zero.
2084 	 */
2085 	savecsum = th->th_sum;
2086 	th->th_sum = 0;
2087 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2088 	th->th_sum = savecsum;
2089 	/*
2090 	 * Step 3: Update MD5 hash with TCP segment data.
2091 	 *         Use m_apply() to avoid an early m_pullup().
2092 	 */
2093 	if (len > 0)
2094 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2095 	/*
2096 	 * Step 4: Update MD5 hash with shared secret.
2097 	 */
2098 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2099 	MD5Final(buf, &ctx);
2100 	key_sa_recordxfer(sav, m);
2101 	key_freesav(sav);
2102 	return (0);
2103 }
2104 
2105 int
2106 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2107 {
2108 
2109 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2110 	return (0);
2111 }
2112 #endif /* TCP_SIGNATURE */
2113