xref: /freebsd/sys/netinet/tcp_subr.c (revision 35a04710d7286aa9538917fd7f8e417dbee95b82)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_compat.h"
36 #include "opt_inet.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_mac.h"
40 #include "opt_tcpdebug.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/callout.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/malloc.h>
48 #include <sys/mbuf.h>
49 #ifdef INET6
50 #include <sys/domain.h>
51 #endif
52 #include <sys/priv.h>
53 #include <sys/proc.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/protosw.h>
57 #include <sys/random.h>
58 
59 #include <vm/uma.h>
60 
61 #include <net/route.h>
62 #include <net/if.h>
63 
64 #include <netinet/in.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/ip.h>
67 #ifdef INET6
68 #include <netinet/ip6.h>
69 #endif
70 #include <netinet/in_pcb.h>
71 #ifdef INET6
72 #include <netinet6/in6_pcb.h>
73 #endif
74 #include <netinet/in_var.h>
75 #include <netinet/ip_var.h>
76 #ifdef INET6
77 #include <netinet6/ip6_var.h>
78 #include <netinet6/scope6_var.h>
79 #include <netinet6/nd6.h>
80 #endif
81 #include <netinet/ip_icmp.h>
82 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_seq.h>
85 #include <netinet/tcp_timer.h>
86 #include <netinet/tcp_var.h>
87 #include <netinet/tcp_syncache.h>
88 #ifdef INET6
89 #include <netinet6/tcp6_var.h>
90 #endif
91 #include <netinet/tcpip.h>
92 #ifdef TCPDEBUG
93 #include <netinet/tcp_debug.h>
94 #endif
95 #include <netinet6/ip6protosw.h>
96 
97 #ifdef IPSEC
98 #include <netipsec/ipsec.h>
99 #include <netipsec/xform.h>
100 #ifdef INET6
101 #include <netipsec/ipsec6.h>
102 #endif
103 #include <netipsec/key.h>
104 #endif /*IPSEC*/
105 
106 #include <machine/in_cksum.h>
107 #include <sys/md5.h>
108 
109 #include <security/mac/mac_framework.h>
110 
111 int	tcp_mssdflt = TCP_MSS;
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
113     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
114 
115 #ifdef INET6
116 int	tcp_v6mssdflt = TCP6_MSS;
117 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
118     CTLFLAG_RW, &tcp_v6mssdflt , 0,
119     "Default TCP Maximum Segment Size for IPv6");
120 #endif
121 
122 /*
123  * Minimum MSS we accept and use. This prevents DoS attacks where
124  * we are forced to a ridiculous low MSS like 20 and send hundreds
125  * of packets instead of one. The effect scales with the available
126  * bandwidth and quickly saturates the CPU and network interface
127  * with packet generation and sending. Set to zero to disable MINMSS
128  * checking. This setting prevents us from sending too small packets.
129  */
130 int	tcp_minmss = TCP_MINMSS;
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
132     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
133 
134 int	tcp_do_rfc1323 = 1;
135 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
136     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
137 
138 static int	tcp_log_debug = 0;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_debug, CTLFLAG_RW,
140     &tcp_log_debug, 0, "Log errors caused by incoming TCP segments");
141 
142 static int	tcp_tcbhashsize = 0;
143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN,
144     &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
145 
146 static int	do_tcpdrain = 1;
147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW,
148     &do_tcpdrain, 0,
149     "Enable tcp_drain routine for extra help when low on mbufs");
150 
151 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
152     &tcbinfo.ipi_count, 0, "Number of active PCBs");
153 
154 static int	icmp_may_rst = 1;
155 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW,
156     &icmp_may_rst, 0,
157     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
158 
159 static int	tcp_isn_reseed_interval = 0;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
161     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
162 
163 /*
164  * TCP bandwidth limiting sysctls.  Note that the default lower bound of
165  * 1024 exists only for debugging.  A good production default would be
166  * something like 6100.
167  */
168 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, inflight, CTLFLAG_RW, 0,
169     "TCP inflight data limiting");
170 
171 static int	tcp_inflight_enable = 1;
172 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, enable, CTLFLAG_RW,
173     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
174 
175 static int	tcp_inflight_debug = 0;
176 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, debug, CTLFLAG_RW,
177     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
178 
179 static int	tcp_inflight_rttthresh;
180 SYSCTL_PROC(_net_inet_tcp_inflight, OID_AUTO, rttthresh, CTLTYPE_INT|CTLFLAG_RW,
181     &tcp_inflight_rttthresh, 0, sysctl_msec_to_ticks, "I",
182     "RTT threshold below which inflight will deactivate itself");
183 
184 static int	tcp_inflight_min = 6144;
185 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW,
186     &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
187 
188 static int	tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
189 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW,
190     &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
191 
192 static int	tcp_inflight_stab = 20;
193 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW,
194     &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets");
195 
196 uma_zone_t sack_hole_zone;
197 
198 static struct inpcb *tcp_notify(struct inpcb *, int);
199 static void	tcp_isn_tick(void *);
200 
201 /*
202  * Target size of TCP PCB hash tables. Must be a power of two.
203  *
204  * Note that this can be overridden by the kernel environment
205  * variable net.inet.tcp.tcbhashsize
206  */
207 #ifndef TCBHASHSIZE
208 #define TCBHASHSIZE	512
209 #endif
210 
211 /*
212  * XXX
213  * Callouts should be moved into struct tcp directly.  They are currently
214  * separate because the tcpcb structure is exported to userland for sysctl
215  * parsing purposes, which do not know about callouts.
216  */
217 struct tcpcb_mem {
218 	struct	tcpcb		tcb;
219 	struct	tcp_timer	tt;
220 };
221 
222 static uma_zone_t tcpcb_zone;
223 MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers");
224 struct callout isn_callout;
225 static struct mtx isn_mtx;
226 
227 #define	ISN_LOCK_INIT()	mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF)
228 #define	ISN_LOCK()	mtx_lock(&isn_mtx)
229 #define	ISN_UNLOCK()	mtx_unlock(&isn_mtx)
230 
231 /*
232  * TCP initialization.
233  */
234 static void
235 tcp_zone_change(void *tag)
236 {
237 
238 	uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
239 	uma_zone_set_max(tcpcb_zone, maxsockets);
240 	tcp_tw_zone_change();
241 }
242 
243 static int
244 tcp_inpcb_init(void *mem, int size, int flags)
245 {
246 	struct inpcb *inp = mem;
247 
248 	INP_LOCK_INIT(inp, "inp", "tcpinp");
249 	return (0);
250 }
251 
252 void
253 tcp_init(void)
254 {
255 
256 	int hashsize = TCBHASHSIZE;
257 	tcp_delacktime = TCPTV_DELACK;
258 	tcp_keepinit = TCPTV_KEEP_INIT;
259 	tcp_keepidle = TCPTV_KEEP_IDLE;
260 	tcp_keepintvl = TCPTV_KEEPINTVL;
261 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
262 	tcp_msl = TCPTV_MSL;
263 	tcp_rexmit_min = TCPTV_MIN;
264 	if (tcp_rexmit_min < 1)
265 		tcp_rexmit_min = 1;
266 	tcp_rexmit_slop = TCPTV_CPU_VAR;
267 	tcp_inflight_rttthresh = TCPTV_INFLIGHT_RTTTHRESH;
268 	tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT;
269 
270 	INP_INFO_LOCK_INIT(&tcbinfo, "tcp");
271 	LIST_INIT(&tcb);
272 	tcbinfo.ipi_listhead = &tcb;
273 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
274 	if (!powerof2(hashsize)) {
275 		printf("WARNING: TCB hash size not a power of 2\n");
276 		hashsize = 512; /* safe default */
277 	}
278 	tcp_tcbhashsize = hashsize;
279 	tcbinfo.ipi_hashbase = hashinit(hashsize, M_PCB,
280 	    &tcbinfo.ipi_hashmask);
281 	tcbinfo.ipi_porthashbase = hashinit(hashsize, M_PCB,
282 	    &tcbinfo.ipi_porthashmask);
283 	tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb),
284 	    NULL, NULL, tcp_inpcb_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
285 	uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
286 #ifdef INET6
287 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
288 #else /* INET6 */
289 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
290 #endif /* INET6 */
291 	if (max_protohdr < TCP_MINPROTOHDR)
292 		max_protohdr = TCP_MINPROTOHDR;
293 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
294 		panic("tcp_init");
295 #undef TCP_MINPROTOHDR
296 	/*
297 	 * These have to be type stable for the benefit of the timers.
298 	 */
299 	tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem),
300 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
301 	uma_zone_set_max(tcpcb_zone, maxsockets);
302 	tcp_tw_init();
303 	syncache_init();
304 	tcp_hc_init();
305 	tcp_reass_init();
306 	ISN_LOCK_INIT();
307 	callout_init(&isn_callout, CALLOUT_MPSAFE);
308 	tcp_isn_tick(NULL);
309 	EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL,
310 		SHUTDOWN_PRI_DEFAULT);
311 	sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole),
312 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 	EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL,
314 		EVENTHANDLER_PRI_ANY);
315 }
316 
317 void
318 tcp_fini(void *xtp)
319 {
320 
321 	callout_stop(&isn_callout);
322 }
323 
324 /*
325  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
326  * tcp_template used to store this data in mbufs, but we now recopy it out
327  * of the tcpcb each time to conserve mbufs.
328  */
329 void
330 tcpip_fillheaders(struct inpcb *inp, void *ip_ptr, void *tcp_ptr)
331 {
332 	struct tcphdr *th = (struct tcphdr *)tcp_ptr;
333 
334 	INP_LOCK_ASSERT(inp);
335 
336 #ifdef INET6
337 	if ((inp->inp_vflag & INP_IPV6) != 0) {
338 		struct ip6_hdr *ip6;
339 
340 		ip6 = (struct ip6_hdr *)ip_ptr;
341 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
342 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
343 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
344 			(IPV6_VERSION & IPV6_VERSION_MASK);
345 		ip6->ip6_nxt = IPPROTO_TCP;
346 		ip6->ip6_plen = sizeof(struct tcphdr);
347 		ip6->ip6_src = inp->in6p_laddr;
348 		ip6->ip6_dst = inp->in6p_faddr;
349 	} else
350 #endif
351 	{
352 		struct ip *ip;
353 
354 		ip = (struct ip *)ip_ptr;
355 		ip->ip_v = IPVERSION;
356 		ip->ip_hl = 5;
357 		ip->ip_tos = inp->inp_ip_tos;
358 		ip->ip_len = 0;
359 		ip->ip_id = 0;
360 		ip->ip_off = 0;
361 		ip->ip_ttl = inp->inp_ip_ttl;
362 		ip->ip_sum = 0;
363 		ip->ip_p = IPPROTO_TCP;
364 		ip->ip_src = inp->inp_laddr;
365 		ip->ip_dst = inp->inp_faddr;
366 	}
367 	th->th_sport = inp->inp_lport;
368 	th->th_dport = inp->inp_fport;
369 	th->th_seq = 0;
370 	th->th_ack = 0;
371 	th->th_x2 = 0;
372 	th->th_off = 5;
373 	th->th_flags = 0;
374 	th->th_win = 0;
375 	th->th_urp = 0;
376 	th->th_sum = 0;		/* in_pseudo() is called later for ipv4 */
377 }
378 
379 /*
380  * Create template to be used to send tcp packets on a connection.
381  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
382  * use for this function is in keepalives, which use tcp_respond.
383  */
384 struct tcptemp *
385 tcpip_maketemplate(struct inpcb *inp)
386 {
387 	struct mbuf *m;
388 	struct tcptemp *n;
389 
390 	m = m_get(M_DONTWAIT, MT_DATA);
391 	if (m == NULL)
392 		return (0);
393 	m->m_len = sizeof(struct tcptemp);
394 	n = mtod(m, struct tcptemp *);
395 
396 	tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
397 	return (n);
398 }
399 
400 /*
401  * Send a single message to the TCP at address specified by
402  * the given TCP/IP header.  If m == NULL, then we make a copy
403  * of the tcpiphdr at ti and send directly to the addressed host.
404  * This is used to force keep alive messages out using the TCP
405  * template for a connection.  If flags are given then we send
406  * a message back to the TCP which originated the * segment ti,
407  * and discard the mbuf containing it and any other attached mbufs.
408  *
409  * In any case the ack and sequence number of the transmitted
410  * segment are as specified by the parameters.
411  *
412  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
413  */
414 void
415 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
416     tcp_seq ack, tcp_seq seq, int flags)
417 {
418 	int tlen;
419 	int win = 0;
420 	struct ip *ip;
421 	struct tcphdr *nth;
422 #ifdef INET6
423 	struct ip6_hdr *ip6;
424 	int isipv6;
425 #endif /* INET6 */
426 	int ipflags = 0;
427 	struct inpcb *inp;
428 
429 	KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL"));
430 
431 #ifdef INET6
432 	isipv6 = ((struct ip *)ipgen)->ip_v == 6;
433 	ip6 = ipgen;
434 #endif /* INET6 */
435 	ip = ipgen;
436 
437 	if (tp != NULL) {
438 		inp = tp->t_inpcb;
439 		KASSERT(inp != NULL, ("tcp control block w/o inpcb"));
440 		INP_LOCK_ASSERT(inp);
441 	} else
442 		inp = NULL;
443 
444 	if (tp != NULL) {
445 		if (!(flags & TH_RST)) {
446 			win = sbspace(&inp->inp_socket->so_rcv);
447 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
448 				win = (long)TCP_MAXWIN << tp->rcv_scale;
449 		}
450 	}
451 	if (m == NULL) {
452 		m = m_gethdr(M_DONTWAIT, MT_DATA);
453 		if (m == NULL)
454 			return;
455 		tlen = 0;
456 		m->m_data += max_linkhdr;
457 #ifdef INET6
458 		if (isipv6) {
459 			bcopy((caddr_t)ip6, mtod(m, caddr_t),
460 			      sizeof(struct ip6_hdr));
461 			ip6 = mtod(m, struct ip6_hdr *);
462 			nth = (struct tcphdr *)(ip6 + 1);
463 		} else
464 #endif /* INET6 */
465 	      {
466 		bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
467 		ip = mtod(m, struct ip *);
468 		nth = (struct tcphdr *)(ip + 1);
469 	      }
470 		bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
471 		flags = TH_ACK;
472 	} else {
473 		m_freem(m->m_next);
474 		m->m_next = NULL;
475 		m->m_data = (caddr_t)ipgen;
476 		/* m_len is set later */
477 		tlen = 0;
478 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
479 #ifdef INET6
480 		if (isipv6) {
481 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
482 			nth = (struct tcphdr *)(ip6 + 1);
483 		} else
484 #endif /* INET6 */
485 	      {
486 		xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
487 		nth = (struct tcphdr *)(ip + 1);
488 	      }
489 		if (th != nth) {
490 			/*
491 			 * this is usually a case when an extension header
492 			 * exists between the IPv6 header and the
493 			 * TCP header.
494 			 */
495 			nth->th_sport = th->th_sport;
496 			nth->th_dport = th->th_dport;
497 		}
498 		xchg(nth->th_dport, nth->th_sport, n_short);
499 #undef xchg
500 	}
501 #ifdef INET6
502 	if (isipv6) {
503 		ip6->ip6_flow = 0;
504 		ip6->ip6_vfc = IPV6_VERSION;
505 		ip6->ip6_nxt = IPPROTO_TCP;
506 		ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
507 						tlen));
508 		tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
509 	} else
510 #endif
511 	{
512 		tlen += sizeof (struct tcpiphdr);
513 		ip->ip_len = tlen;
514 		ip->ip_ttl = ip_defttl;
515 		if (path_mtu_discovery)
516 			ip->ip_off |= IP_DF;
517 	}
518 	m->m_len = tlen;
519 	m->m_pkthdr.len = tlen;
520 	m->m_pkthdr.rcvif = NULL;
521 #ifdef MAC
522 	if (inp != NULL) {
523 		/*
524 		 * Packet is associated with a socket, so allow the
525 		 * label of the response to reflect the socket label.
526 		 */
527 		INP_LOCK_ASSERT(inp);
528 		mac_inpcb_create_mbuf(inp, m);
529 	} else {
530 		/*
531 		 * Packet is not associated with a socket, so possibly
532 		 * update the label in place.
533 		 */
534 		mac_netinet_tcp_reply(m);
535 	}
536 #endif
537 	nth->th_seq = htonl(seq);
538 	nth->th_ack = htonl(ack);
539 	nth->th_x2 = 0;
540 	nth->th_off = sizeof (struct tcphdr) >> 2;
541 	nth->th_flags = flags;
542 	if (tp != NULL)
543 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
544 	else
545 		nth->th_win = htons((u_short)win);
546 	nth->th_urp = 0;
547 #ifdef INET6
548 	if (isipv6) {
549 		nth->th_sum = 0;
550 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
551 					sizeof(struct ip6_hdr),
552 					tlen - sizeof(struct ip6_hdr));
553 		ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb :
554 		    NULL, NULL);
555 	} else
556 #endif /* INET6 */
557 	{
558 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
559 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
560 		m->m_pkthdr.csum_flags = CSUM_TCP;
561 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
562 	}
563 #ifdef TCPDEBUG
564 	if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
565 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
566 #endif
567 #ifdef INET6
568 	if (isipv6)
569 		(void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp);
570 	else
571 #endif /* INET6 */
572 	(void) ip_output(m, NULL, NULL, ipflags, NULL, inp);
573 }
574 
575 /*
576  * Create a new TCP control block, making an
577  * empty reassembly queue and hooking it to the argument
578  * protocol control block.  The `inp' parameter must have
579  * come from the zone allocator set up in tcp_init().
580  */
581 struct tcpcb *
582 tcp_newtcpcb(struct inpcb *inp)
583 {
584 	struct tcpcb_mem *tm;
585 	struct tcpcb *tp;
586 #ifdef INET6
587 	int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
588 #endif /* INET6 */
589 
590 	tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO);
591 	if (tm == NULL)
592 		return (NULL);
593 	tp = &tm->tcb;
594 	tp->t_timers = &tm->tt;
595 	/*	LIST_INIT(&tp->t_segq); */	/* XXX covered by M_ZERO */
596 	tp->t_maxseg = tp->t_maxopd =
597 #ifdef INET6
598 		isipv6 ? tcp_v6mssdflt :
599 #endif /* INET6 */
600 		tcp_mssdflt;
601 
602 	/* Set up our timeouts. */
603 	callout_init(&tp->t_timers->tt_rexmt, CALLOUT_MPSAFE);
604 	callout_init(&tp->t_timers->tt_persist, CALLOUT_MPSAFE);
605 	callout_init(&tp->t_timers->tt_keep, CALLOUT_MPSAFE);
606 	callout_init(&tp->t_timers->tt_2msl, CALLOUT_MPSAFE);
607 	callout_init(&tp->t_timers->tt_delack, CALLOUT_MPSAFE);
608 
609 	if (tcp_do_rfc1323)
610 		tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
611 	if (tcp_do_sack)
612 		tp->t_flags |= TF_SACK_PERMIT;
613 	TAILQ_INIT(&tp->snd_holes);
614 	tp->t_inpcb = inp;	/* XXX */
615 	/*
616 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
617 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
618 	 * reasonable initial retransmit time.
619 	 */
620 	tp->t_srtt = TCPTV_SRTTBASE;
621 	tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
622 	tp->t_rttmin = tcp_rexmit_min;
623 	tp->t_rxtcur = TCPTV_RTOBASE;
624 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
625 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
626 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
627 	tp->t_rcvtime = ticks;
628 	tp->t_bw_rtttime = ticks;
629 	/*
630 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
631 	 * because the socket may be bound to an IPv6 wildcard address,
632 	 * which may match an IPv4-mapped IPv6 address.
633 	 */
634 	inp->inp_ip_ttl = ip_defttl;
635 	inp->inp_ppcb = tp;
636 	return (tp);		/* XXX */
637 }
638 
639 /*
640  * Drop a TCP connection, reporting
641  * the specified error.  If connection is synchronized,
642  * then send a RST to peer.
643  */
644 struct tcpcb *
645 tcp_drop(struct tcpcb *tp, int errno)
646 {
647 	struct socket *so = tp->t_inpcb->inp_socket;
648 
649 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
650 	INP_LOCK_ASSERT(tp->t_inpcb);
651 
652 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
653 		tp->t_state = TCPS_CLOSED;
654 		(void) tcp_output(tp);
655 		tcpstat.tcps_drops++;
656 	} else
657 		tcpstat.tcps_conndrops++;
658 	if (errno == ETIMEDOUT && tp->t_softerror)
659 		errno = tp->t_softerror;
660 	so->so_error = errno;
661 	return (tcp_close(tp));
662 }
663 
664 void
665 tcp_discardcb(struct tcpcb *tp)
666 {
667 	struct tseg_qent *q;
668 	struct inpcb *inp = tp->t_inpcb;
669 	struct socket *so = inp->inp_socket;
670 #ifdef INET6
671 	int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
672 #endif /* INET6 */
673 
674 	INP_LOCK_ASSERT(inp);
675 
676 	/*
677 	 * Make sure that all of our timers are stopped before we
678 	 * delete the PCB.
679 	 */
680 	callout_stop(&tp->t_timers->tt_rexmt);
681 	callout_stop(&tp->t_timers->tt_persist);
682 	callout_stop(&tp->t_timers->tt_keep);
683 	callout_stop(&tp->t_timers->tt_2msl);
684 	callout_stop(&tp->t_timers->tt_delack);
685 
686 	/*
687 	 * If we got enough samples through the srtt filter,
688 	 * save the rtt and rttvar in the routing entry.
689 	 * 'Enough' is arbitrarily defined as 4 rtt samples.
690 	 * 4 samples is enough for the srtt filter to converge
691 	 * to within enough % of the correct value; fewer samples
692 	 * and we could save a bogus rtt. The danger is not high
693 	 * as tcp quickly recovers from everything.
694 	 * XXX: Works very well but needs some more statistics!
695 	 */
696 	if (tp->t_rttupdated >= 4) {
697 		struct hc_metrics_lite metrics;
698 		u_long ssthresh;
699 
700 		bzero(&metrics, sizeof(metrics));
701 		/*
702 		 * Update the ssthresh always when the conditions below
703 		 * are satisfied. This gives us better new start value
704 		 * for the congestion avoidance for new connections.
705 		 * ssthresh is only set if packet loss occured on a session.
706 		 *
707 		 * XXXRW: 'so' may be NULL here, and/or socket buffer may be
708 		 * being torn down.  Ideally this code would not use 'so'.
709 		 */
710 		ssthresh = tp->snd_ssthresh;
711 		if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) {
712 			/*
713 			 * convert the limit from user data bytes to
714 			 * packets then to packet data bytes.
715 			 */
716 			ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg;
717 			if (ssthresh < 2)
718 				ssthresh = 2;
719 			ssthresh *= (u_long)(tp->t_maxseg +
720 #ifdef INET6
721 				      (isipv6 ? sizeof (struct ip6_hdr) +
722 					       sizeof (struct tcphdr) :
723 #endif
724 				       sizeof (struct tcpiphdr)
725 #ifdef INET6
726 				       )
727 #endif
728 				      );
729 		} else
730 			ssthresh = 0;
731 		metrics.rmx_ssthresh = ssthresh;
732 
733 		metrics.rmx_rtt = tp->t_srtt;
734 		metrics.rmx_rttvar = tp->t_rttvar;
735 		/* XXX: This wraps if the pipe is more than 4 Gbit per second */
736 		metrics.rmx_bandwidth = tp->snd_bandwidth;
737 		metrics.rmx_cwnd = tp->snd_cwnd;
738 		metrics.rmx_sendpipe = 0;
739 		metrics.rmx_recvpipe = 0;
740 
741 		tcp_hc_update(&inp->inp_inc, &metrics);
742 	}
743 
744 	/* free the reassembly queue, if any */
745 	while ((q = LIST_FIRST(&tp->t_segq)) != NULL) {
746 		LIST_REMOVE(q, tqe_q);
747 		m_freem(q->tqe_m);
748 		uma_zfree(tcp_reass_zone, q);
749 		tp->t_segqlen--;
750 		tcp_reass_qsize--;
751 	}
752 	tcp_free_sackholes(tp);
753 	inp->inp_ppcb = NULL;
754 	tp->t_inpcb = NULL;
755 	uma_zfree(tcpcb_zone, tp);
756 }
757 
758 /*
759  * Attempt to close a TCP control block, marking it as dropped, and freeing
760  * the socket if we hold the only reference.
761  */
762 struct tcpcb *
763 tcp_close(struct tcpcb *tp)
764 {
765 	struct inpcb *inp = tp->t_inpcb;
766 	struct socket *so;
767 
768 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
769 	INP_LOCK_ASSERT(inp);
770 
771 	in_pcbdrop(inp);
772 	tcpstat.tcps_closed++;
773 	KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL"));
774 	so = inp->inp_socket;
775 	soisdisconnected(so);
776 	if (inp->inp_vflag & INP_SOCKREF) {
777 		KASSERT(so->so_state & SS_PROTOREF,
778 		    ("tcp_close: !SS_PROTOREF"));
779 		inp->inp_vflag &= ~INP_SOCKREF;
780 		INP_UNLOCK(inp);
781 		ACCEPT_LOCK();
782 		SOCK_LOCK(so);
783 		so->so_state &= ~SS_PROTOREF;
784 		sofree(so);
785 		return (NULL);
786 	}
787 	return (tp);
788 }
789 
790 void
791 tcp_drain(void)
792 {
793 
794 	if (do_tcpdrain) {
795 		struct inpcb *inpb;
796 		struct tcpcb *tcpb;
797 		struct tseg_qent *te;
798 
799 	/*
800 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
801 	 * if there is one...
802 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
803 	 *      reassembly queue should be flushed, but in a situation
804 	 *	where we're really low on mbufs, this is potentially
805 	 *	usefull.
806 	 */
807 		INP_INFO_RLOCK(&tcbinfo);
808 		LIST_FOREACH(inpb, tcbinfo.ipi_listhead, inp_list) {
809 			if (inpb->inp_vflag & INP_TIMEWAIT)
810 				continue;
811 			INP_LOCK(inpb);
812 			if ((tcpb = intotcpcb(inpb)) != NULL) {
813 				while ((te = LIST_FIRST(&tcpb->t_segq))
814 			            != NULL) {
815 					LIST_REMOVE(te, tqe_q);
816 					m_freem(te->tqe_m);
817 					uma_zfree(tcp_reass_zone, te);
818 					tcpb->t_segqlen--;
819 					tcp_reass_qsize--;
820 				}
821 				tcp_clean_sackreport(tcpb);
822 			}
823 			INP_UNLOCK(inpb);
824 		}
825 		INP_INFO_RUNLOCK(&tcbinfo);
826 	}
827 }
828 
829 /*
830  * Notify a tcp user of an asynchronous error;
831  * store error as soft error, but wake up user
832  * (for now, won't do anything until can select for soft error).
833  *
834  * Do not wake up user since there currently is no mechanism for
835  * reporting soft errors (yet - a kqueue filter may be added).
836  */
837 static struct inpcb *
838 tcp_notify(struct inpcb *inp, int error)
839 {
840 	struct tcpcb *tp;
841 
842 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
843 	INP_LOCK_ASSERT(inp);
844 
845 	if ((inp->inp_vflag & INP_TIMEWAIT) ||
846 	    (inp->inp_vflag & INP_DROPPED))
847 		return (inp);
848 
849 	tp = intotcpcb(inp);
850 	KASSERT(tp != NULL, ("tcp_notify: tp == NULL"));
851 
852 	/*
853 	 * Ignore some errors if we are hooked up.
854 	 * If connection hasn't completed, has retransmitted several times,
855 	 * and receives a second error, give up now.  This is better
856 	 * than waiting a long time to establish a connection that
857 	 * can never complete.
858 	 */
859 	if (tp->t_state == TCPS_ESTABLISHED &&
860 	    (error == EHOSTUNREACH || error == ENETUNREACH ||
861 	     error == EHOSTDOWN)) {
862 		return (inp);
863 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
864 	    tp->t_softerror) {
865 		tp = tcp_drop(tp, error);
866 		if (tp != NULL)
867 			return (inp);
868 		else
869 			return (NULL);
870 	} else {
871 		tp->t_softerror = error;
872 		return (inp);
873 	}
874 #if 0
875 	wakeup( &so->so_timeo);
876 	sorwakeup(so);
877 	sowwakeup(so);
878 #endif
879 }
880 
881 static int
882 tcp_pcblist(SYSCTL_HANDLER_ARGS)
883 {
884 	int error, i, m, n, pcb_count;
885 	struct inpcb *inp, **inp_list;
886 	inp_gen_t gencnt;
887 	struct xinpgen xig;
888 
889 	/*
890 	 * The process of preparing the TCB list is too time-consuming and
891 	 * resource-intensive to repeat twice on every request.
892 	 */
893 	if (req->oldptr == NULL) {
894 		m = syncache_pcbcount();
895 		n = tcbinfo.ipi_count;
896 		req->oldidx = 2 * (sizeof xig)
897 			+ ((m + n) + n/8) * sizeof(struct xtcpcb);
898 		return (0);
899 	}
900 
901 	if (req->newptr != NULL)
902 		return (EPERM);
903 
904 	/*
905 	 * OK, now we're committed to doing something.
906 	 */
907 	INP_INFO_RLOCK(&tcbinfo);
908 	gencnt = tcbinfo.ipi_gencnt;
909 	n = tcbinfo.ipi_count;
910 	INP_INFO_RUNLOCK(&tcbinfo);
911 
912 	m = syncache_pcbcount();
913 
914 	error = sysctl_wire_old_buffer(req, 2 * (sizeof xig)
915 		+ (n + m) * sizeof(struct xtcpcb));
916 	if (error != 0)
917 		return (error);
918 
919 	xig.xig_len = sizeof xig;
920 	xig.xig_count = n + m;
921 	xig.xig_gen = gencnt;
922 	xig.xig_sogen = so_gencnt;
923 	error = SYSCTL_OUT(req, &xig, sizeof xig);
924 	if (error)
925 		return (error);
926 
927 	error = syncache_pcblist(req, m, &pcb_count);
928 	if (error)
929 		return (error);
930 
931 	inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
932 	if (inp_list == NULL)
933 		return (ENOMEM);
934 
935 	INP_INFO_RLOCK(&tcbinfo);
936 	for (inp = LIST_FIRST(tcbinfo.ipi_listhead), i = 0; inp != NULL && i
937 	    < n; inp = LIST_NEXT(inp, inp_list)) {
938 		INP_LOCK(inp);
939 		if (inp->inp_gencnt <= gencnt) {
940 			/*
941 			 * XXX: This use of cr_cansee(), introduced with
942 			 * TCP state changes, is not quite right, but for
943 			 * now, better than nothing.
944 			 */
945 			if (inp->inp_vflag & INP_TIMEWAIT) {
946 				if (intotw(inp) != NULL)
947 					error = cr_cansee(req->td->td_ucred,
948 					    intotw(inp)->tw_cred);
949 				else
950 					error = EINVAL;	/* Skip this inp. */
951 			} else
952 				error = cr_canseesocket(req->td->td_ucred,
953 				    inp->inp_socket);
954 			if (error == 0)
955 				inp_list[i++] = inp;
956 		}
957 		INP_UNLOCK(inp);
958 	}
959 	INP_INFO_RUNLOCK(&tcbinfo);
960 	n = i;
961 
962 	error = 0;
963 	for (i = 0; i < n; i++) {
964 		inp = inp_list[i];
965 		INP_LOCK(inp);
966 		if (inp->inp_gencnt <= gencnt) {
967 			struct xtcpcb xt;
968 			void *inp_ppcb;
969 
970 			bzero(&xt, sizeof(xt));
971 			xt.xt_len = sizeof xt;
972 			/* XXX should avoid extra copy */
973 			bcopy(inp, &xt.xt_inp, sizeof *inp);
974 			inp_ppcb = inp->inp_ppcb;
975 			if (inp_ppcb == NULL)
976 				bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
977 			else if (inp->inp_vflag & INP_TIMEWAIT) {
978 				bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
979 				xt.xt_tp.t_state = TCPS_TIME_WAIT;
980 			} else
981 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
982 			if (inp->inp_socket != NULL)
983 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
984 			else {
985 				bzero(&xt.xt_socket, sizeof xt.xt_socket);
986 				xt.xt_socket.xso_protocol = IPPROTO_TCP;
987 			}
988 			xt.xt_inp.inp_gencnt = inp->inp_gencnt;
989 			INP_UNLOCK(inp);
990 			error = SYSCTL_OUT(req, &xt, sizeof xt);
991 		} else
992 			INP_UNLOCK(inp);
993 
994 	}
995 	if (!error) {
996 		/*
997 		 * Give the user an updated idea of our state.
998 		 * If the generation differs from what we told
999 		 * her before, she knows that something happened
1000 		 * while we were processing this request, and it
1001 		 * might be necessary to retry.
1002 		 */
1003 		INP_INFO_RLOCK(&tcbinfo);
1004 		xig.xig_gen = tcbinfo.ipi_gencnt;
1005 		xig.xig_sogen = so_gencnt;
1006 		xig.xig_count = tcbinfo.ipi_count + pcb_count;
1007 		INP_INFO_RUNLOCK(&tcbinfo);
1008 		error = SYSCTL_OUT(req, &xig, sizeof xig);
1009 	}
1010 	free(inp_list, M_TEMP);
1011 	return (error);
1012 }
1013 
1014 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1015     tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1016 
1017 static int
1018 tcp_getcred(SYSCTL_HANDLER_ARGS)
1019 {
1020 	struct xucred xuc;
1021 	struct sockaddr_in addrs[2];
1022 	struct inpcb *inp;
1023 	int error;
1024 
1025 	error = priv_check(req->td, PRIV_NETINET_GETCRED);
1026 	if (error)
1027 		return (error);
1028 	error = SYSCTL_IN(req, addrs, sizeof(addrs));
1029 	if (error)
1030 		return (error);
1031 	INP_INFO_RLOCK(&tcbinfo);
1032 	inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
1033 	    addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1034 	if (inp == NULL) {
1035 		error = ENOENT;
1036 		goto outunlocked;
1037 	}
1038 	INP_LOCK(inp);
1039 	if (inp->inp_socket == NULL) {
1040 		error = ENOENT;
1041 		goto out;
1042 	}
1043 	error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1044 	if (error)
1045 		goto out;
1046 	cru2x(inp->inp_socket->so_cred, &xuc);
1047 out:
1048 	INP_UNLOCK(inp);
1049 outunlocked:
1050 	INP_INFO_RUNLOCK(&tcbinfo);
1051 	if (error == 0)
1052 		error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1053 	return (error);
1054 }
1055 
1056 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred,
1057     CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1058     tcp_getcred, "S,xucred", "Get the xucred of a TCP connection");
1059 
1060 #ifdef INET6
1061 static int
1062 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1063 {
1064 	struct xucred xuc;
1065 	struct sockaddr_in6 addrs[2];
1066 	struct inpcb *inp;
1067 	int error, mapped = 0;
1068 
1069 	error = priv_check(req->td, PRIV_NETINET_GETCRED);
1070 	if (error)
1071 		return (error);
1072 	error = SYSCTL_IN(req, addrs, sizeof(addrs));
1073 	if (error)
1074 		return (error);
1075 	if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 ||
1076 	    (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) {
1077 		return (error);
1078 	}
1079 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1080 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1081 			mapped = 1;
1082 		else
1083 			return (EINVAL);
1084 	}
1085 
1086 	INP_INFO_RLOCK(&tcbinfo);
1087 	if (mapped == 1)
1088 		inp = in_pcblookup_hash(&tcbinfo,
1089 			*(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1090 			addrs[1].sin6_port,
1091 			*(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1092 			addrs[0].sin6_port,
1093 			0, NULL);
1094 	else
1095 		inp = in6_pcblookup_hash(&tcbinfo,
1096 			&addrs[1].sin6_addr, addrs[1].sin6_port,
1097 			&addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1098 	if (inp == NULL) {
1099 		error = ENOENT;
1100 		goto outunlocked;
1101 	}
1102 	INP_LOCK(inp);
1103 	if (inp->inp_socket == NULL) {
1104 		error = ENOENT;
1105 		goto out;
1106 	}
1107 	error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1108 	if (error)
1109 		goto out;
1110 	cru2x(inp->inp_socket->so_cred, &xuc);
1111 out:
1112 	INP_UNLOCK(inp);
1113 outunlocked:
1114 	INP_INFO_RUNLOCK(&tcbinfo);
1115 	if (error == 0)
1116 		error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1117 	return (error);
1118 }
1119 
1120 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred,
1121     CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1122     tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection");
1123 #endif
1124 
1125 
1126 void
1127 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1128 {
1129 	struct ip *ip = vip;
1130 	struct tcphdr *th;
1131 	struct in_addr faddr;
1132 	struct inpcb *inp;
1133 	struct tcpcb *tp;
1134 	struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1135 	struct icmp *icp;
1136 	struct in_conninfo inc;
1137 	tcp_seq icmp_tcp_seq;
1138 	int mtu;
1139 
1140 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1141 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1142 		return;
1143 
1144 	if (cmd == PRC_MSGSIZE)
1145 		notify = tcp_mtudisc;
1146 	else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1147 		cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1148 		notify = tcp_drop_syn_sent;
1149 	/*
1150 	 * Redirects don't need to be handled up here.
1151 	 */
1152 	else if (PRC_IS_REDIRECT(cmd))
1153 		return;
1154 	/*
1155 	 * Source quench is depreciated.
1156 	 */
1157 	else if (cmd == PRC_QUENCH)
1158 		return;
1159 	/*
1160 	 * Hostdead is ugly because it goes linearly through all PCBs.
1161 	 * XXX: We never get this from ICMP, otherwise it makes an
1162 	 * excellent DoS attack on machines with many connections.
1163 	 */
1164 	else if (cmd == PRC_HOSTDEAD)
1165 		ip = NULL;
1166 	else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0)
1167 		return;
1168 	if (ip != NULL) {
1169 		icp = (struct icmp *)((caddr_t)ip
1170 				      - offsetof(struct icmp, icmp_ip));
1171 		th = (struct tcphdr *)((caddr_t)ip
1172 				       + (ip->ip_hl << 2));
1173 		INP_INFO_WLOCK(&tcbinfo);
1174 		inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1175 		    ip->ip_src, th->th_sport, 0, NULL);
1176 		if (inp != NULL)  {
1177 			INP_LOCK(inp);
1178 			if (!(inp->inp_vflag & INP_TIMEWAIT) &&
1179 			    !(inp->inp_vflag & INP_DROPPED) &&
1180 			    !(inp->inp_socket == NULL)) {
1181 				icmp_tcp_seq = htonl(th->th_seq);
1182 				tp = intotcpcb(inp);
1183 				if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) &&
1184 				    SEQ_LT(icmp_tcp_seq, tp->snd_max)) {
1185 					if (cmd == PRC_MSGSIZE) {
1186 					    /*
1187 					     * MTU discovery:
1188 					     * If we got a needfrag set the MTU
1189 					     * in the route to the suggested new
1190 					     * value (if given) and then notify.
1191 					     */
1192 					    bzero(&inc, sizeof(inc));
1193 					    inc.inc_flags = 0;	/* IPv4 */
1194 					    inc.inc_faddr = faddr;
1195 
1196 					    mtu = ntohs(icp->icmp_nextmtu);
1197 					    /*
1198 					     * If no alternative MTU was
1199 					     * proposed, try the next smaller
1200 					     * one.  ip->ip_len has already
1201 					     * been swapped in icmp_input().
1202 					     */
1203 					    if (!mtu)
1204 						mtu = ip_next_mtu(ip->ip_len,
1205 						 1);
1206 					    if (mtu < max(296, (tcp_minmss)
1207 						 + sizeof(struct tcpiphdr)))
1208 						mtu = 0;
1209 					    if (!mtu)
1210 						mtu = tcp_mssdflt
1211 						 + sizeof(struct tcpiphdr);
1212 					    /*
1213 					     * Only cache the the MTU if it
1214 					     * is smaller than the interface
1215 					     * or route MTU.  tcp_mtudisc()
1216 					     * will do right thing by itself.
1217 					     */
1218 					    if (mtu <= tcp_maxmtu(&inc, NULL))
1219 						tcp_hc_updatemtu(&inc, mtu);
1220 					}
1221 
1222 					inp = (*notify)(inp, inetctlerrmap[cmd]);
1223 				}
1224 			}
1225 			if (inp != NULL)
1226 				INP_UNLOCK(inp);
1227 		} else {
1228 			inc.inc_fport = th->th_dport;
1229 			inc.inc_lport = th->th_sport;
1230 			inc.inc_faddr = faddr;
1231 			inc.inc_laddr = ip->ip_src;
1232 #ifdef INET6
1233 			inc.inc_isipv6 = 0;
1234 #endif
1235 			syncache_unreach(&inc, th);
1236 		}
1237 		INP_INFO_WUNLOCK(&tcbinfo);
1238 	} else
1239 		in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify);
1240 }
1241 
1242 #ifdef INET6
1243 void
1244 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1245 {
1246 	struct tcphdr th;
1247 	struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1248 	struct ip6_hdr *ip6;
1249 	struct mbuf *m;
1250 	struct ip6ctlparam *ip6cp = NULL;
1251 	const struct sockaddr_in6 *sa6_src = NULL;
1252 	int off;
1253 	struct tcp_portonly {
1254 		u_int16_t th_sport;
1255 		u_int16_t th_dport;
1256 	} *thp;
1257 
1258 	if (sa->sa_family != AF_INET6 ||
1259 	    sa->sa_len != sizeof(struct sockaddr_in6))
1260 		return;
1261 
1262 	if (cmd == PRC_MSGSIZE)
1263 		notify = tcp_mtudisc;
1264 	else if (!PRC_IS_REDIRECT(cmd) &&
1265 		 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1266 		return;
1267 	/* Source quench is depreciated. */
1268 	else if (cmd == PRC_QUENCH)
1269 		return;
1270 
1271 	/* if the parameter is from icmp6, decode it. */
1272 	if (d != NULL) {
1273 		ip6cp = (struct ip6ctlparam *)d;
1274 		m = ip6cp->ip6c_m;
1275 		ip6 = ip6cp->ip6c_ip6;
1276 		off = ip6cp->ip6c_off;
1277 		sa6_src = ip6cp->ip6c_src;
1278 	} else {
1279 		m = NULL;
1280 		ip6 = NULL;
1281 		off = 0;	/* fool gcc */
1282 		sa6_src = &sa6_any;
1283 	}
1284 
1285 	if (ip6 != NULL) {
1286 		struct in_conninfo inc;
1287 		/*
1288 		 * XXX: We assume that when IPV6 is non NULL,
1289 		 * M and OFF are valid.
1290 		 */
1291 
1292 		/* check if we can safely examine src and dst ports */
1293 		if (m->m_pkthdr.len < off + sizeof(*thp))
1294 			return;
1295 
1296 		bzero(&th, sizeof(th));
1297 		m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1298 
1299 		in6_pcbnotify(&tcbinfo, sa, th.th_dport,
1300 		    (struct sockaddr *)ip6cp->ip6c_src,
1301 		    th.th_sport, cmd, NULL, notify);
1302 
1303 		inc.inc_fport = th.th_dport;
1304 		inc.inc_lport = th.th_sport;
1305 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1306 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1307 		inc.inc_isipv6 = 1;
1308 		INP_INFO_WLOCK(&tcbinfo);
1309 		syncache_unreach(&inc, &th);
1310 		INP_INFO_WUNLOCK(&tcbinfo);
1311 	} else
1312 		in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src,
1313 			      0, cmd, NULL, notify);
1314 }
1315 #endif /* INET6 */
1316 
1317 
1318 /*
1319  * Following is where TCP initial sequence number generation occurs.
1320  *
1321  * There are two places where we must use initial sequence numbers:
1322  * 1.  In SYN-ACK packets.
1323  * 2.  In SYN packets.
1324  *
1325  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1326  * tcp_syncache.c for details.
1327  *
1328  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1329  * depends on this property.  In addition, these ISNs should be
1330  * unguessable so as to prevent connection hijacking.  To satisfy
1331  * the requirements of this situation, the algorithm outlined in
1332  * RFC 1948 is used, with only small modifications.
1333  *
1334  * Implementation details:
1335  *
1336  * Time is based off the system timer, and is corrected so that it
1337  * increases by one megabyte per second.  This allows for proper
1338  * recycling on high speed LANs while still leaving over an hour
1339  * before rollover.
1340  *
1341  * As reading the *exact* system time is too expensive to be done
1342  * whenever setting up a TCP connection, we increment the time
1343  * offset in two ways.  First, a small random positive increment
1344  * is added to isn_offset for each connection that is set up.
1345  * Second, the function tcp_isn_tick fires once per clock tick
1346  * and increments isn_offset as necessary so that sequence numbers
1347  * are incremented at approximately ISN_BYTES_PER_SECOND.  The
1348  * random positive increments serve only to ensure that the same
1349  * exact sequence number is never sent out twice (as could otherwise
1350  * happen when a port is recycled in less than the system tick
1351  * interval.)
1352  *
1353  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1354  * between seeding of isn_secret.  This is normally set to zero,
1355  * as reseeding should not be necessary.
1356  *
1357  * Locking of the global variables isn_secret, isn_last_reseed, isn_offset,
1358  * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock.  In
1359  * general, this means holding an exclusive (write) lock.
1360  */
1361 
1362 #define ISN_BYTES_PER_SECOND 1048576
1363 #define ISN_STATIC_INCREMENT 4096
1364 #define ISN_RANDOM_INCREMENT (4096 - 1)
1365 
1366 static u_char isn_secret[32];
1367 static int isn_last_reseed;
1368 static u_int32_t isn_offset, isn_offset_old;
1369 static MD5_CTX isn_ctx;
1370 
1371 tcp_seq
1372 tcp_new_isn(struct tcpcb *tp)
1373 {
1374 	u_int32_t md5_buffer[4];
1375 	tcp_seq new_isn;
1376 
1377 	INP_LOCK_ASSERT(tp->t_inpcb);
1378 
1379 	ISN_LOCK();
1380 	/* Seed if this is the first use, reseed if requested. */
1381 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1382 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1383 		< (u_int)ticks))) {
1384 		read_random(&isn_secret, sizeof(isn_secret));
1385 		isn_last_reseed = ticks;
1386 	}
1387 
1388 	/* Compute the md5 hash and return the ISN. */
1389 	MD5Init(&isn_ctx);
1390 	MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1391 	MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1392 #ifdef INET6
1393 	if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1394 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1395 			  sizeof(struct in6_addr));
1396 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1397 			  sizeof(struct in6_addr));
1398 	} else
1399 #endif
1400 	{
1401 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1402 			  sizeof(struct in_addr));
1403 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1404 			  sizeof(struct in_addr));
1405 	}
1406 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1407 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1408 	new_isn = (tcp_seq) md5_buffer[0];
1409 	isn_offset += ISN_STATIC_INCREMENT +
1410 		(arc4random() & ISN_RANDOM_INCREMENT);
1411 	new_isn += isn_offset;
1412 	ISN_UNLOCK();
1413 	return (new_isn);
1414 }
1415 
1416 /*
1417  * Increment the offset to the next ISN_BYTES_PER_SECOND / 100 boundary
1418  * to keep time flowing at a relatively constant rate.  If the random
1419  * increments have already pushed us past the projected offset, do nothing.
1420  */
1421 static void
1422 tcp_isn_tick(void *xtp)
1423 {
1424 	u_int32_t projected_offset;
1425 
1426 	ISN_LOCK();
1427 	projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100;
1428 
1429 	if (SEQ_GT(projected_offset, isn_offset))
1430 		isn_offset = projected_offset;
1431 
1432 	isn_offset_old = isn_offset;
1433 	callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL);
1434 	ISN_UNLOCK();
1435 }
1436 
1437 /*
1438  * When a specific ICMP unreachable message is received and the
1439  * connection state is SYN-SENT, drop the connection.  This behavior
1440  * is controlled by the icmp_may_rst sysctl.
1441  */
1442 struct inpcb *
1443 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1444 {
1445 	struct tcpcb *tp;
1446 
1447 	INP_INFO_WLOCK_ASSERT(&tcbinfo);
1448 	INP_LOCK_ASSERT(inp);
1449 
1450 	if ((inp->inp_vflag & INP_TIMEWAIT) ||
1451 	    (inp->inp_vflag & INP_DROPPED))
1452 		return (inp);
1453 
1454 	tp = intotcpcb(inp);
1455 	if (tp->t_state != TCPS_SYN_SENT)
1456 		return (inp);
1457 
1458 	tp = tcp_drop(tp, errno);
1459 	if (tp != NULL)
1460 		return (inp);
1461 	else
1462 		return (NULL);
1463 }
1464 
1465 /*
1466  * When `need fragmentation' ICMP is received, update our idea of the MSS
1467  * based on the new value in the route.  Also nudge TCP to send something,
1468  * since we know the packet we just sent was dropped.
1469  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1470  */
1471 struct inpcb *
1472 tcp_mtudisc(struct inpcb *inp, int errno)
1473 {
1474 	struct tcpcb *tp;
1475 	struct socket *so = inp->inp_socket;
1476 	u_int maxmtu;
1477 	u_int romtu;
1478 	int mss;
1479 #ifdef INET6
1480 	int isipv6;
1481 #endif /* INET6 */
1482 
1483 	INP_LOCK_ASSERT(inp);
1484 	if ((inp->inp_vflag & INP_TIMEWAIT) ||
1485 	    (inp->inp_vflag & INP_DROPPED))
1486 		return (inp);
1487 
1488 	tp = intotcpcb(inp);
1489 	KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL"));
1490 
1491 #ifdef INET6
1492 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1493 #endif
1494 	maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */
1495 	romtu =
1496 #ifdef INET6
1497 	    isipv6 ? tcp_maxmtu6(&inp->inp_inc, NULL) :
1498 #endif /* INET6 */
1499 	    tcp_maxmtu(&inp->inp_inc, NULL);
1500 	if (!maxmtu)
1501 		maxmtu = romtu;
1502 	else
1503 		maxmtu = min(maxmtu, romtu);
1504 	if (!maxmtu) {
1505 		tp->t_maxopd = tp->t_maxseg =
1506 #ifdef INET6
1507 			isipv6 ? tcp_v6mssdflt :
1508 #endif /* INET6 */
1509 			tcp_mssdflt;
1510 		return (inp);
1511 	}
1512 	mss = maxmtu -
1513 #ifdef INET6
1514 		(isipv6 ? sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1515 #endif /* INET6 */
1516 		 sizeof(struct tcpiphdr)
1517 #ifdef INET6
1518 		 )
1519 #endif /* INET6 */
1520 		;
1521 
1522 	/*
1523 	 * XXX - The above conditional probably violates the TCP
1524 	 * spec.  The problem is that, since we don't know the
1525 	 * other end's MSS, we are supposed to use a conservative
1526 	 * default.  But, if we do that, then MTU discovery will
1527 	 * never actually take place, because the conservative
1528 	 * default is much less than the MTUs typically seen
1529 	 * on the Internet today.  For the moment, we'll sweep
1530 	 * this under the carpet.
1531 	 *
1532 	 * The conservative default might not actually be a problem
1533 	 * if the only case this occurs is when sending an initial
1534 	 * SYN with options and data to a host we've never talked
1535 	 * to before.  Then, they will reply with an MSS value which
1536 	 * will get recorded and the new parameters should get
1537 	 * recomputed.  For Further Study.
1538 	 */
1539 	if (tp->t_maxopd <= mss)
1540 		return (inp);
1541 	tp->t_maxopd = mss;
1542 
1543 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1544 	    (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1545 		mss -= TCPOLEN_TSTAMP_APPA;
1546 #if	(MCLBYTES & (MCLBYTES - 1)) == 0
1547 	if (mss > MCLBYTES)
1548 		mss &= ~(MCLBYTES-1);
1549 #else
1550 	if (mss > MCLBYTES)
1551 		mss = mss / MCLBYTES * MCLBYTES;
1552 #endif
1553 	if (so->so_snd.sb_hiwat < mss)
1554 		mss = so->so_snd.sb_hiwat;
1555 
1556 	tp->t_maxseg = mss;
1557 
1558 	tcpstat.tcps_mturesent++;
1559 	tp->t_rtttime = 0;
1560 	tp->snd_nxt = tp->snd_una;
1561 	tcp_free_sackholes(tp);
1562 	tp->snd_recover = tp->snd_max;
1563 	if (tp->t_flags & TF_SACK_PERMIT)
1564 		EXIT_FASTRECOVERY(tp);
1565 	tcp_output(tp);
1566 	return (inp);
1567 }
1568 
1569 /*
1570  * Look-up the routing entry to the peer of this inpcb.  If no route
1571  * is found and it cannot be allocated, then return NULL.  This routine
1572  * is called by TCP routines that access the rmx structure and by tcp_mss
1573  * to get the interface MTU.
1574  */
1575 u_long
1576 tcp_maxmtu(struct in_conninfo *inc, int *flags)
1577 {
1578 	struct route sro;
1579 	struct sockaddr_in *dst;
1580 	struct ifnet *ifp;
1581 	u_long maxmtu = 0;
1582 
1583 	KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer"));
1584 
1585 	bzero(&sro, sizeof(sro));
1586 	if (inc->inc_faddr.s_addr != INADDR_ANY) {
1587 	        dst = (struct sockaddr_in *)&sro.ro_dst;
1588 		dst->sin_family = AF_INET;
1589 		dst->sin_len = sizeof(*dst);
1590 		dst->sin_addr = inc->inc_faddr;
1591 		rtalloc_ign(&sro, RTF_CLONING);
1592 	}
1593 	if (sro.ro_rt != NULL) {
1594 		ifp = sro.ro_rt->rt_ifp;
1595 		if (sro.ro_rt->rt_rmx.rmx_mtu == 0)
1596 			maxmtu = ifp->if_mtu;
1597 		else
1598 			maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu);
1599 
1600 		/* Report additional interface capabilities. */
1601 		if (flags != NULL) {
1602 			if (ifp->if_capenable & IFCAP_TSO4 &&
1603 			    ifp->if_hwassist & CSUM_TSO)
1604 				*flags |= CSUM_TSO;
1605 		}
1606 		RTFREE(sro.ro_rt);
1607 	}
1608 	return (maxmtu);
1609 }
1610 
1611 #ifdef INET6
1612 u_long
1613 tcp_maxmtu6(struct in_conninfo *inc, int *flags)
1614 {
1615 	struct route_in6 sro6;
1616 	struct ifnet *ifp;
1617 	u_long maxmtu = 0;
1618 
1619 	KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer"));
1620 
1621 	bzero(&sro6, sizeof(sro6));
1622 	if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1623 		sro6.ro_dst.sin6_family = AF_INET6;
1624 		sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1625 		sro6.ro_dst.sin6_addr = inc->inc6_faddr;
1626 		rtalloc_ign((struct route *)&sro6, RTF_CLONING);
1627 	}
1628 	if (sro6.ro_rt != NULL) {
1629 		ifp = sro6.ro_rt->rt_ifp;
1630 		if (sro6.ro_rt->rt_rmx.rmx_mtu == 0)
1631 			maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp);
1632 		else
1633 			maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu,
1634 				     IN6_LINKMTU(sro6.ro_rt->rt_ifp));
1635 
1636 		/* Report additional interface capabilities. */
1637 		if (flags != NULL) {
1638 			if (ifp->if_capenable & IFCAP_TSO6 &&
1639 			    ifp->if_hwassist & CSUM_TSO)
1640 				*flags |= CSUM_TSO;
1641 		}
1642 		RTFREE(sro6.ro_rt);
1643 	}
1644 
1645 	return (maxmtu);
1646 }
1647 #endif /* INET6 */
1648 
1649 #ifdef IPSEC
1650 /* compute ESP/AH header size for TCP, including outer IP header. */
1651 size_t
1652 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1653 {
1654 	struct inpcb *inp;
1655 	struct mbuf *m;
1656 	size_t hdrsiz;
1657 	struct ip *ip;
1658 #ifdef INET6
1659 	struct ip6_hdr *ip6;
1660 #endif
1661 	struct tcphdr *th;
1662 
1663 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1664 		return (0);
1665 	MGETHDR(m, M_DONTWAIT, MT_DATA);
1666 	if (!m)
1667 		return (0);
1668 
1669 #ifdef INET6
1670 	if ((inp->inp_vflag & INP_IPV6) != 0) {
1671 		ip6 = mtod(m, struct ip6_hdr *);
1672 		th = (struct tcphdr *)(ip6 + 1);
1673 		m->m_pkthdr.len = m->m_len =
1674 			sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1675 		tcpip_fillheaders(inp, ip6, th);
1676 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1677 	} else
1678 #endif /* INET6 */
1679 	{
1680 		ip = mtod(m, struct ip *);
1681 		th = (struct tcphdr *)(ip + 1);
1682 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1683 		tcpip_fillheaders(inp, ip, th);
1684 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1685 	}
1686 
1687 	m_free(m);
1688 	return (hdrsiz);
1689 }
1690 #endif /* IPSEC */
1691 
1692 /*
1693  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1694  *
1695  * This code attempts to calculate the bandwidth-delay product as a
1696  * means of determining the optimal window size to maximize bandwidth,
1697  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1698  * routers.  This code also does a fairly good job keeping RTTs in check
1699  * across slow links like modems.  We implement an algorithm which is very
1700  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1701  * transmitter side of a TCP connection and so only effects the transmit
1702  * side of the connection.
1703  *
1704  * BACKGROUND:  TCP makes no provision for the management of buffer space
1705  * at the end points or at the intermediate routers and switches.  A TCP
1706  * stream, whether using NewReno or not, will eventually buffer as
1707  * many packets as it is able and the only reason this typically works is
1708  * due to the fairly small default buffers made available for a connection
1709  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1710  * scaling it is now fairly easy for even a single TCP connection to blow-out
1711  * all available buffer space not only on the local interface, but on
1712  * intermediate routers and switches as well.  NewReno makes a misguided
1713  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1714  * then backing off, then steadily increasing the window again until another
1715  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1716  * is only made worse as network loads increase and the idea of intentionally
1717  * blowing out network buffers is, frankly, a terrible way to manage network
1718  * resources.
1719  *
1720  * It is far better to limit the transmit window prior to the failure
1721  * condition being achieved.  There are two general ways to do this:  First
1722  * you can 'scan' through different transmit window sizes and locate the
1723  * point where the RTT stops increasing, indicating that you have filled the
1724  * pipe, then scan backwards until you note that RTT stops decreasing, then
1725  * repeat ad-infinitum.  This method works in principle but has severe
1726  * implementation issues due to RTT variances, timer granularity, and
1727  * instability in the algorithm which can lead to many false positives and
1728  * create oscillations as well as interact badly with other TCP streams
1729  * implementing the same algorithm.
1730  *
1731  * The second method is to limit the window to the bandwidth delay product
1732  * of the link.  This is the method we implement.  RTT variances and our
1733  * own manipulation of the congestion window, bwnd, can potentially
1734  * destabilize the algorithm.  For this reason we have to stabilize the
1735  * elements used to calculate the window.  We do this by using the minimum
1736  * observed RTT, the long term average of the observed bandwidth, and
1737  * by adding two segments worth of slop.  It isn't perfect but it is able
1738  * to react to changing conditions and gives us a very stable basis on
1739  * which to extend the algorithm.
1740  */
1741 void
1742 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1743 {
1744 	u_long bw;
1745 	u_long bwnd;
1746 	int save_ticks;
1747 
1748 	INP_LOCK_ASSERT(tp->t_inpcb);
1749 
1750 	/*
1751 	 * If inflight_enable is disabled in the middle of a tcp connection,
1752 	 * make sure snd_bwnd is effectively disabled.
1753 	 */
1754 	if (tcp_inflight_enable == 0 || tp->t_rttlow < tcp_inflight_rttthresh) {
1755 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1756 		tp->snd_bandwidth = 0;
1757 		return;
1758 	}
1759 
1760 	/*
1761 	 * Figure out the bandwidth.  Due to the tick granularity this
1762 	 * is a very rough number and it MUST be averaged over a fairly
1763 	 * long period of time.  XXX we need to take into account a link
1764 	 * that is not using all available bandwidth, but for now our
1765 	 * slop will ramp us up if this case occurs and the bandwidth later
1766 	 * increases.
1767 	 *
1768 	 * Note: if ticks rollover 'bw' may wind up negative.  We must
1769 	 * effectively reset t_bw_rtttime for this case.
1770 	 */
1771 	save_ticks = ticks;
1772 	if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1773 		return;
1774 
1775 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1776 	    (save_ticks - tp->t_bw_rtttime);
1777 	tp->t_bw_rtttime = save_ticks;
1778 	tp->t_bw_rtseq = ack_seq;
1779 	if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1780 		return;
1781 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1782 
1783 	tp->snd_bandwidth = bw;
1784 
1785 	/*
1786 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1787 	 * segments.  The additional slop puts us squarely in the sweet
1788 	 * spot and also handles the bandwidth run-up case and stabilization.
1789 	 * Without the slop we could be locking ourselves into a lower
1790 	 * bandwidth.
1791 	 *
1792 	 * Situations Handled:
1793 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1794 	 *	    high speed LANs, allowing larger TCP buffers to be
1795 	 *	    specified, and also does a good job preventing
1796 	 *	    over-queueing of packets over choke points like modems
1797 	 *	    (at least for the transmit side).
1798 	 *
1799 	 *	(2) Is able to handle changing network loads (bandwidth
1800 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1801 	 *	    increases).
1802 	 *
1803 	 *	(3) Theoretically should stabilize in the face of multiple
1804 	 *	    connections implementing the same algorithm (this may need
1805 	 *	    a little work).
1806 	 *
1807 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1808 	 *	    be adjusted with a sysctl but typically only needs to be
1809 	 *	    on very slow connections.  A value no smaller then 5
1810 	 *	    should be used, but only reduce this default if you have
1811 	 *	    no other choice.
1812 	 */
1813 #define USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1814 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10;
1815 #undef USERTT
1816 
1817 	if (tcp_inflight_debug > 0) {
1818 		static int ltime;
1819 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1820 			ltime = ticks;
1821 			printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1822 			    tp,
1823 			    bw,
1824 			    tp->t_rttbest,
1825 			    tp->t_srtt,
1826 			    bwnd
1827 			);
1828 		}
1829 	}
1830 	if ((long)bwnd < tcp_inflight_min)
1831 		bwnd = tcp_inflight_min;
1832 	if (bwnd > tcp_inflight_max)
1833 		bwnd = tcp_inflight_max;
1834 	if ((long)bwnd < tp->t_maxseg * 2)
1835 		bwnd = tp->t_maxseg * 2;
1836 	tp->snd_bwnd = bwnd;
1837 }
1838 
1839 #ifdef TCP_SIGNATURE
1840 /*
1841  * Callback function invoked by m_apply() to digest TCP segment data
1842  * contained within an mbuf chain.
1843  */
1844 static int
1845 tcp_signature_apply(void *fstate, void *data, u_int len)
1846 {
1847 
1848 	MD5Update(fstate, (u_char *)data, len);
1849 	return (0);
1850 }
1851 
1852 /*
1853  * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385)
1854  *
1855  * Parameters:
1856  * m		pointer to head of mbuf chain
1857  * off0		offset to TCP header within the mbuf chain
1858  * len		length of TCP segment data, excluding options
1859  * optlen	length of TCP segment options
1860  * buf		pointer to storage for computed MD5 digest
1861  * direction	direction of flow (IPSEC_DIR_INBOUND or OUTBOUND)
1862  *
1863  * We do this over ip, tcphdr, segment data, and the key in the SADB.
1864  * When called from tcp_input(), we can be sure that th_sum has been
1865  * zeroed out and verified already.
1866  *
1867  * This function is for IPv4 use only. Calling this function with an
1868  * IPv6 packet in the mbuf chain will yield undefined results.
1869  *
1870  * Return 0 if successful, otherwise return -1.
1871  *
1872  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1873  * search with the destination IP address, and a 'magic SPI' to be
1874  * determined by the application. This is hardcoded elsewhere to 1179
1875  * right now. Another branch of this code exists which uses the SPD to
1876  * specify per-application flows but it is unstable.
1877  */
1878 int
1879 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen,
1880     u_char *buf, u_int direction)
1881 {
1882 	union sockaddr_union dst;
1883 	struct ippseudo ippseudo;
1884 	MD5_CTX ctx;
1885 	int doff;
1886 	struct ip *ip;
1887 	struct ipovly *ipovly;
1888 	struct secasvar *sav;
1889 	struct tcphdr *th;
1890 	u_short savecsum;
1891 
1892 	KASSERT(m != NULL, ("NULL mbuf chain"));
1893 	KASSERT(buf != NULL, ("NULL signature pointer"));
1894 
1895 	/* Extract the destination from the IP header in the mbuf. */
1896 	ip = mtod(m, struct ip *);
1897 	bzero(&dst, sizeof(union sockaddr_union));
1898 	dst.sa.sa_len = sizeof(struct sockaddr_in);
1899 	dst.sa.sa_family = AF_INET;
1900 	dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ?
1901 	    ip->ip_src : ip->ip_dst;
1902 
1903 	/* Look up an SADB entry which matches the address of the peer. */
1904 	sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI));
1905 	if (sav == NULL) {
1906 		printf("%s: SADB lookup failed for %s\n", __func__,
1907 		    inet_ntoa(dst.sin.sin_addr));
1908 		return (EINVAL);
1909 	}
1910 
1911 	MD5Init(&ctx);
1912 	ipovly = (struct ipovly *)ip;
1913 	th = (struct tcphdr *)((u_char *)ip + off0);
1914 	doff = off0 + sizeof(struct tcphdr) + optlen;
1915 
1916 	/*
1917 	 * Step 1: Update MD5 hash with IP pseudo-header.
1918 	 *
1919 	 * XXX The ippseudo header MUST be digested in network byte order,
1920 	 * or else we'll fail the regression test. Assume all fields we've
1921 	 * been doing arithmetic on have been in host byte order.
1922 	 * XXX One cannot depend on ipovly->ih_len here. When called from
1923 	 * tcp_output(), the underlying ip_len member has not yet been set.
1924 	 */
1925 	ippseudo.ippseudo_src = ipovly->ih_src;
1926 	ippseudo.ippseudo_dst = ipovly->ih_dst;
1927 	ippseudo.ippseudo_pad = 0;
1928 	ippseudo.ippseudo_p = IPPROTO_TCP;
1929 	ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
1930 	MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
1931 
1932 	/*
1933 	 * Step 2: Update MD5 hash with TCP header, excluding options.
1934 	 * The TCP checksum must be set to zero.
1935 	 */
1936 	savecsum = th->th_sum;
1937 	th->th_sum = 0;
1938 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
1939 	th->th_sum = savecsum;
1940 
1941 	/*
1942 	 * Step 3: Update MD5 hash with TCP segment data.
1943 	 *         Use m_apply() to avoid an early m_pullup().
1944 	 */
1945 	if (len > 0)
1946 		m_apply(m, doff, len, tcp_signature_apply, &ctx);
1947 
1948 	/*
1949 	 * Step 4: Update MD5 hash with shared secret.
1950 	 */
1951 	MD5Update(&ctx, sav->key_auth->key_data, _KEYLEN(sav->key_auth));
1952 	MD5Final(buf, &ctx);
1953 
1954 	key_sa_recordxfer(sav, m);
1955 	KEY_FREESAV(&sav);
1956 	return (0);
1957 }
1958 #endif /* TCP_SIGNATURE */
1959 
1960 static int
1961 sysctl_drop(SYSCTL_HANDLER_ARGS)
1962 {
1963 	/* addrs[0] is a foreign socket, addrs[1] is a local one. */
1964 	struct sockaddr_storage addrs[2];
1965 	struct inpcb *inp;
1966 	struct tcpcb *tp;
1967 	struct tcptw *tw;
1968 	struct sockaddr_in *fin, *lin;
1969 #ifdef INET6
1970 	struct sockaddr_in6 *fin6, *lin6;
1971 	struct in6_addr f6, l6;
1972 #endif
1973 	int error;
1974 
1975 	inp = NULL;
1976 	fin = lin = NULL;
1977 #ifdef INET6
1978 	fin6 = lin6 = NULL;
1979 #endif
1980 	error = 0;
1981 
1982 	if (req->oldptr != NULL || req->oldlen != 0)
1983 		return (EINVAL);
1984 	if (req->newptr == NULL)
1985 		return (EPERM);
1986 	if (req->newlen < sizeof(addrs))
1987 		return (ENOMEM);
1988 	error = SYSCTL_IN(req, &addrs, sizeof(addrs));
1989 	if (error)
1990 		return (error);
1991 
1992 	switch (addrs[0].ss_family) {
1993 #ifdef INET6
1994 	case AF_INET6:
1995 		fin6 = (struct sockaddr_in6 *)&addrs[0];
1996 		lin6 = (struct sockaddr_in6 *)&addrs[1];
1997 		if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
1998 		    lin6->sin6_len != sizeof(struct sockaddr_in6))
1999 			return (EINVAL);
2000 		if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) {
2001 			if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2002 				return (EINVAL);
2003 			in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]);
2004 			in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]);
2005 			fin = (struct sockaddr_in *)&addrs[0];
2006 			lin = (struct sockaddr_in *)&addrs[1];
2007 			break;
2008 		}
2009 		error = sa6_embedscope(fin6, ip6_use_defzone);
2010 		if (error)
2011 			return (error);
2012 		error = sa6_embedscope(lin6, ip6_use_defzone);
2013 		if (error)
2014 			return (error);
2015 		break;
2016 #endif
2017 	case AF_INET:
2018 		fin = (struct sockaddr_in *)&addrs[0];
2019 		lin = (struct sockaddr_in *)&addrs[1];
2020 		if (fin->sin_len != sizeof(struct sockaddr_in) ||
2021 		    lin->sin_len != sizeof(struct sockaddr_in))
2022 			return (EINVAL);
2023 		break;
2024 	default:
2025 		return (EINVAL);
2026 	}
2027 	INP_INFO_WLOCK(&tcbinfo);
2028 	switch (addrs[0].ss_family) {
2029 #ifdef INET6
2030 	case AF_INET6:
2031 		inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port,
2032 		    &l6, lin6->sin6_port, 0, NULL);
2033 		break;
2034 #endif
2035 	case AF_INET:
2036 		inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port,
2037 		    lin->sin_addr, lin->sin_port, 0, NULL);
2038 		break;
2039 	}
2040 	if (inp != NULL) {
2041 		INP_LOCK(inp);
2042 		if (inp->inp_vflag & INP_TIMEWAIT) {
2043 			/*
2044 			 * XXXRW: There currently exists a state where an
2045 			 * inpcb is present, but its timewait state has been
2046 			 * discarded.  For now, don't allow dropping of this
2047 			 * type of inpcb.
2048 			 */
2049 			tw = intotw(inp);
2050 			if (tw != NULL)
2051 				tcp_twclose(tw, 0);
2052 			else
2053 				INP_UNLOCK(inp);
2054 		} else if (!(inp->inp_vflag & INP_DROPPED) &&
2055 			   !(inp->inp_socket->so_options & SO_ACCEPTCONN)) {
2056 			tp = intotcpcb(inp);
2057 			tp = tcp_drop(tp, ECONNABORTED);
2058 			if (tp != NULL)
2059 				INP_UNLOCK(inp);
2060 		} else
2061 			INP_UNLOCK(inp);
2062 	} else
2063 		error = ESRCH;
2064 	INP_INFO_WUNLOCK(&tcbinfo);
2065 	return (error);
2066 }
2067 
2068 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop,
2069     CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL,
2070     0, sysctl_drop, "", "Drop TCP connection");
2071 
2072 /*
2073  * Generate a standardized TCP log line for use throughout the
2074  * tcp subsystem.  Memory allocation is done with M_NOWAIT to
2075  * allow use in the interrupt context.
2076  *
2077  * NB: The caller MUST free(s, M_TCPLOG) the returned string.
2078  * NB: The function may return NULL if memory allocation failed.
2079  *
2080  * Due to header inclusion and ordering limitations the struct ip
2081  * and ip6_hdr pointers have to be passed as void pointers.
2082  */
2083 char *
2084 tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr,
2085     const void *ip6hdr)
2086 {
2087 	char *s, *sp;
2088 	size_t size;
2089 	struct ip *ip;
2090 #ifdef INET6
2091 	const struct ip6_hdr *ip6;
2092 
2093 	ip6 = (const struct ip6_hdr *)ip6hdr;
2094 #endif /* INET6 */
2095 	ip = (struct ip *)ip4hdr;
2096 
2097 	/*
2098 	 * The log line looks like this:
2099 	 * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2<SYN>"
2100 	 */
2101 	size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") +
2102 	    sizeof(PRINT_TH_FLAGS) + 1 +
2103 #ifdef INET6
2104 	    2 * INET6_ADDRSTRLEN;
2105 #else
2106 	    2 * INET_ADDRSTRLEN;
2107 #endif /* INET6 */
2108 
2109 	/* Is logging enabled? */
2110 	if (tcp_log_debug == 0 && tcp_log_in_vain == 0)
2111 		return (NULL);
2112 
2113 	s = malloc(size, M_TCPLOG, M_ZERO|M_NOWAIT);
2114 	if (s == NULL)
2115 		return (NULL);
2116 
2117 	strcat(s, "TCP: [");
2118 	sp = s + strlen(s);
2119 
2120 	if (inc && inc->inc_isipv6 == 0) {
2121 		inet_ntoa_r(inc->inc_faddr, sp);
2122 		sp = s + strlen(s);
2123 		sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2124 		sp = s + strlen(s);
2125 		inet_ntoa_r(inc->inc_laddr, sp);
2126 		sp = s + strlen(s);
2127 		sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2128 #ifdef INET6
2129 	} else if (inc) {
2130 		ip6_sprintf(sp, &inc->inc6_faddr);
2131 		sp = s + strlen(s);
2132 		sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2133 		sp = s + strlen(s);
2134 		ip6_sprintf(sp, &inc->inc6_laddr);
2135 		sp = s + strlen(s);
2136 		sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2137 	} else if (ip6 && th) {
2138 		ip6_sprintf(sp, &ip6->ip6_src);
2139 		sp = s + strlen(s);
2140 		sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2141 		sp = s + strlen(s);
2142 		ip6_sprintf(sp, &ip6->ip6_dst);
2143 		sp = s + strlen(s);
2144 		sprintf(sp, "]:%i", ntohs(th->th_dport));
2145 #endif /* INET6 */
2146 	} else if (ip && th) {
2147 		inet_ntoa_r(ip->ip_src, sp);
2148 		sp = s + strlen(s);
2149 		sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2150 		sp = s + strlen(s);
2151 		inet_ntoa_r(ip->ip_dst, sp);
2152 		sp = s + strlen(s);
2153 		sprintf(sp, "]:%i", ntohs(th->th_dport));
2154 	} else {
2155 		free(s, M_TCPLOG);
2156 		return (NULL);
2157 	}
2158 	sp = s + strlen(s);
2159 	if (th)
2160 		sprintf(sp, " tcpflags 0x%b", th->th_flags, PRINT_TH_FLAGS);
2161 	if (*(s + size - 1) != '\0')
2162 		panic("%s: string too long", __func__);
2163 	return (s);
2164 }
2165