xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision bc49aa1be5400e3bdd801519c6936e8947d5d432)
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) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
36  *
37  * License terms: all terms for the DragonFly license above plus the following:
38  *
39  * 4. All advertising materials mentioning features or use of this software
40  *    must display the following acknowledgement:
41  *
42  *	This product includes software developed by Jeffrey M. Hsu
43  *	for the DragonFly Project.
44  *
45  *    This requirement may be waived with permission from Jeffrey Hsu.
46  *    This requirement will sunset and may be removed on July 8 2005,
47  *    after which the standard DragonFly license (as shown above) will
48  *    apply.
49  */
50 
51 /*
52  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
53  *	The Regents of the University of California.  All rights reserved.
54  *
55  * Redistribution and use in source and binary forms, with or without
56  * modification, are permitted provided that the following conditions
57  * are met:
58  * 1. Redistributions of source code must retain the above copyright
59  *    notice, this list of conditions and the following disclaimer.
60  * 2. Redistributions in binary form must reproduce the above copyright
61  *    notice, this list of conditions and the following disclaimer in the
62  *    documentation and/or other materials provided with the distribution.
63  * 3. All advertising materials mentioning features or use of this software
64  *    must display the following acknowledgement:
65  *	This product includes software developed by the University of
66  *	California, Berkeley and its contributors.
67  * 4. Neither the name of the University nor the names of its contributors
68  *    may be used to endorse or promote products derived from this software
69  *    without specific prior written permission.
70  *
71  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
72  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
73  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
74  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
75  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
76  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
77  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
78  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
79  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
80  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
81  * SUCH DAMAGE.
82  *
83  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
84  * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
85  * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.36 2004/07/08 22:07:35 hsu Exp $
86  */
87 
88 #include "opt_compat.h"
89 #include "opt_inet6.h"
90 #include "opt_ipsec.h"
91 #include "opt_tcpdebug.h"
92 
93 #include <sys/param.h>
94 #include <sys/systm.h>
95 #include <sys/callout.h>
96 #include <sys/kernel.h>
97 #include <sys/sysctl.h>
98 #include <sys/malloc.h>
99 #include <sys/mpipe.h>
100 #include <sys/mbuf.h>
101 #ifdef INET6
102 #include <sys/domain.h>
103 #endif
104 #include <sys/proc.h>
105 #include <sys/socket.h>
106 #include <sys/socketvar.h>
107 #include <sys/protosw.h>
108 #include <sys/random.h>
109 #include <sys/in_cksum.h>
110 
111 #include <vm/vm_zone.h>
112 
113 #include <net/route.h>
114 #include <net/if.h>
115 #include <net/netisr.h>
116 
117 #define	_IP_VHL
118 #include <netinet/in.h>
119 #include <netinet/in_systm.h>
120 #include <netinet/ip.h>
121 #include <netinet/ip6.h>
122 #include <netinet/in_pcb.h>
123 #include <netinet6/in6_pcb.h>
124 #include <netinet/in_var.h>
125 #include <netinet/ip_var.h>
126 #include <netinet6/ip6_var.h>
127 #include <netinet/tcp.h>
128 #include <netinet/tcp_fsm.h>
129 #include <netinet/tcp_seq.h>
130 #include <netinet/tcp_timer.h>
131 #include <netinet/tcp_var.h>
132 #include <netinet6/tcp6_var.h>
133 #include <netinet/tcpip.h>
134 #ifdef TCPDEBUG
135 #include <netinet/tcp_debug.h>
136 #endif
137 #include <netinet6/ip6protosw.h>
138 
139 #ifdef IPSEC
140 #include <netinet6/ipsec.h>
141 #ifdef INET6
142 #include <netinet6/ipsec6.h>
143 #endif
144 #endif
145 
146 #ifdef FAST_IPSEC
147 #include <netipsec/ipsec.h>
148 #ifdef INET6
149 #include <netipsec/ipsec6.h>
150 #endif
151 #define	IPSEC
152 #endif
153 
154 #include <sys/md5.h>
155 
156 #include <sys/msgport2.h>
157 
158 #include <machine/smp.h>
159 
160 int tcp_mssdflt = TCP_MSS;
161 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
162     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
163 
164 #ifdef INET6
165 int tcp_v6mssdflt = TCP6_MSS;
166 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
167     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
168 #endif
169 
170 #if 0
171 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
172 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
173     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
174 #endif
175 
176 int tcp_do_rfc1323 = 1;
177 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
178     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
179 
180 int tcp_do_rfc1644 = 0;
181 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
182     &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
183 
184 static int tcp_tcbhashsize = 0;
185 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
186      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
187 
188 static int do_tcpdrain = 1;
189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
190      "Enable tcp_drain routine for extra help when low on mbufs");
191 
192 /* XXX JH */
193 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
194     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
195 
196 static int icmp_may_rst = 1;
197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
198     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
199 
200 static int tcp_isn_reseed_interval = 0;
201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
202     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
203 
204 /*
205  * TCP bandwidth limiting sysctls.  Note that the default lower bound of
206  * 1024 exists only for debugging.  A good production default would be
207  * something like 6100.
208  */
209 static int tcp_inflight_enable = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
211     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
212 
213 static int tcp_inflight_debug = 0;
214 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
215     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
216 
217 static int tcp_inflight_min = 6144;
218 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
219     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
220 
221 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
222 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
223     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
224 
225 static int tcp_inflight_stab = 20;
226 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
227     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
228 
229 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
230 static struct malloc_pipe tcptemp_mpipe;
231 
232 static void tcp_cleartaocache (void);
233 static void tcp_notify (struct inpcb *, int);
234 
235 struct tcp_stats tcpstats_ary[MAXCPU];
236 #ifdef SMP
237 static int
238 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
239 {
240 	int cpu, error = 0;
241 
242 	for (cpu = 0; cpu < ncpus; ++cpu) {
243 		if ((error = SYSCTL_OUT(req, (void *)&tcpstats_ary[cpu],
244 					sizeof(struct tcp_stats))))
245 			break;
246 		if ((error = SYSCTL_IN(req, (void *)&tcpstats_ary[cpu],
247 				       sizeof(struct tcp_stats))))
248 			break;
249 	}
250 
251 	return (error);
252 }
253 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
254     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
255 #else
256 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
257     &tcpstat, tcp_stats, "TCP statistics");
258 #endif
259 
260 /*
261  * Target size of TCP PCB hash tables. Must be a power of two.
262  *
263  * Note that this can be overridden by the kernel environment
264  * variable net.inet.tcp.tcbhashsize
265  */
266 #ifndef TCBHASHSIZE
267 #define	TCBHASHSIZE	512
268 #endif
269 
270 /*
271  * This is the actual shape of what we allocate using the zone
272  * allocator.  Doing it this way allows us to protect both structures
273  * using the same generation count, and also eliminates the overhead
274  * of allocating tcpcbs separately.  By hiding the structure here,
275  * we avoid changing most of the rest of the code (although it needs
276  * to be changed, eventually, for greater efficiency).
277  */
278 #define	ALIGNMENT	32
279 #define	ALIGNM1		(ALIGNMENT - 1)
280 struct	inp_tp {
281 	union {
282 		struct	inpcb inp;
283 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
284 	} inp_tp_u;
285 	struct	tcpcb tcb;
286 	struct	callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
287 	struct	callout inp_tp_delack;
288 };
289 #undef ALIGNMENT
290 #undef ALIGNM1
291 
292 /*
293  * Tcp initialization
294  */
295 void
296 tcp_init()
297 {
298 	struct inpcbporthead *porthashbase;
299 	u_long porthashmask;
300 	struct vm_zone *ipi_zone;
301 	int hashsize = TCBHASHSIZE;
302 	int cpu;
303 
304 	/*
305 	 * note: tcptemp is used for keepalives, and it is ok for an
306 	 * allocation to fail so do not specify MPF_INT.
307 	 */
308 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
309 		    25, -1, 0, NULL);
310 
311 	tcp_ccgen = 1;
312 	tcp_cleartaocache();
313 
314 	tcp_delacktime = TCPTV_DELACK;
315 	tcp_keepinit = TCPTV_KEEP_INIT;
316 	tcp_keepidle = TCPTV_KEEP_IDLE;
317 	tcp_keepintvl = TCPTV_KEEPINTVL;
318 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
319 	tcp_msl = TCPTV_MSL;
320 	tcp_rexmit_min = TCPTV_MIN;
321 	tcp_rexmit_slop = TCPTV_CPU_VAR;
322 
323 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
324 	if (!powerof2(hashsize)) {
325 		printf("WARNING: TCB hash size not a power of 2\n");
326 		hashsize = 512; /* safe default */
327 	}
328 	tcp_tcbhashsize = hashsize;
329 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
330 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
331 			 ZONE_INTERRUPT, 0);
332 
333 	for (cpu = 0; cpu < ncpus2; cpu++) {
334 		in_pcbinfo_init(&tcbinfo[cpu]);
335 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
336 		    &tcbinfo[cpu].hashmask);
337 		tcbinfo[cpu].porthashbase = porthashbase;
338 		tcbinfo[cpu].porthashmask = porthashmask;
339 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
340 		    &tcbinfo[cpu].wildcardhashmask);
341 		tcbinfo[cpu].ipi_zone = ipi_zone;
342 	}
343 
344 	tcp_reass_maxseg = nmbclusters / 16;
345 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
346 
347 #ifdef INET6
348 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
349 #else
350 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
351 #endif
352 	if (max_protohdr < TCP_MINPROTOHDR)
353 		max_protohdr = TCP_MINPROTOHDR;
354 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
355 		panic("tcp_init");
356 #undef TCP_MINPROTOHDR
357 
358 	/*
359 	 * Initialize TCP statistics.
360 	 *
361 	 * It is layed out as an array which is has one element for UP,
362 	 * and SMP_MAXCPU elements for SMP.  This allows us to retain
363 	 * the access mechanism from userland for both UP and SMP.
364 	 */
365 #ifdef SMP
366 	for (cpu = 0; cpu < ncpus; ++cpu) {
367 		bzero(&tcpstats_ary[cpu], sizeof(struct tcp_stats));
368 	}
369 #else
370 	bzero(&tcpstat, sizeof(struct tcp_stats));
371 #endif
372 
373 	syncache_init();
374 	tcp_thread_init();
375 }
376 
377 /*
378  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
379  * tcp_template used to store this data in mbufs, but we now recopy it out
380  * of the tcpcb each time to conserve mbufs.
381  */
382 void
383 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
384 {
385 	struct inpcb *inp = tp->t_inpcb;
386 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
387 
388 #ifdef INET6
389 	if (inp->inp_vflag & INP_IPV6) {
390 		struct ip6_hdr *ip6;
391 
392 		ip6 = (struct ip6_hdr *)ip_ptr;
393 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
394 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
395 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
396 			(IPV6_VERSION & IPV6_VERSION_MASK);
397 		ip6->ip6_nxt = IPPROTO_TCP;
398 		ip6->ip6_plen = sizeof(struct tcphdr);
399 		ip6->ip6_src = inp->in6p_laddr;
400 		ip6->ip6_dst = inp->in6p_faddr;
401 		tcp_hdr->th_sum = 0;
402 	} else
403 #endif
404 	{
405 		struct ip *ip = (struct ip *) ip_ptr;
406 
407 		ip->ip_vhl = IP_VHL_BORING;
408 		ip->ip_tos = 0;
409 		ip->ip_len = 0;
410 		ip->ip_id = 0;
411 		ip->ip_off = 0;
412 		ip->ip_ttl = 0;
413 		ip->ip_sum = 0;
414 		ip->ip_p = IPPROTO_TCP;
415 		ip->ip_src = inp->inp_laddr;
416 		ip->ip_dst = inp->inp_faddr;
417 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
418 				    ip->ip_dst.s_addr,
419 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
420 	}
421 
422 	tcp_hdr->th_sport = inp->inp_lport;
423 	tcp_hdr->th_dport = inp->inp_fport;
424 	tcp_hdr->th_seq = 0;
425 	tcp_hdr->th_ack = 0;
426 	tcp_hdr->th_x2 = 0;
427 	tcp_hdr->th_off = 5;
428 	tcp_hdr->th_flags = 0;
429 	tcp_hdr->th_win = 0;
430 	tcp_hdr->th_urp = 0;
431 }
432 
433 /*
434  * Create template to be used to send tcp packets on a connection.
435  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
436  * use for this function is in keepalives, which use tcp_respond.
437  */
438 struct tcptemp *
439 tcp_maketemplate(struct tcpcb *tp)
440 {
441 	struct tcptemp *tmp;
442 
443 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
444 		return (NULL);
445 	tcp_fillheaders(tp, (void *)&tmp->tt_ipgen, (void *)&tmp->tt_t);
446 	return (tmp);
447 }
448 
449 void
450 tcp_freetemplate(struct tcptemp *tmp)
451 {
452 	mpipe_free(&tcptemp_mpipe, tmp);
453 }
454 
455 /*
456  * Send a single message to the TCP at address specified by
457  * the given TCP/IP header.  If m == NULL, then we make a copy
458  * of the tcpiphdr at ti and send directly to the addressed host.
459  * This is used to force keep alive messages out using the TCP
460  * template for a connection.  If flags are given then we send
461  * a message back to the TCP which originated the * segment ti,
462  * and discard the mbuf containing it and any other attached mbufs.
463  *
464  * In any case the ack and sequence number of the transmitted
465  * segment are as specified by the parameters.
466  *
467  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
468  */
469 void
470 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
471 	    tcp_seq ack, tcp_seq seq, int flags)
472 {
473 	int tlen;
474 	int win = 0;
475 	struct route *ro = NULL;
476 	struct route sro;
477 	struct ip *ip = ipgen;
478 	struct tcphdr *nth;
479 	int ipflags = 0;
480 	struct route_in6 *ro6 = NULL;
481 	struct route_in6 sro6;
482 	struct ip6_hdr *ip6 = ipgen;
483 #ifdef INET6
484 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
485 #else
486 	const boolean_t isipv6 = FALSE;
487 #endif
488 
489 	if (tp != NULL) {
490 		if (!(flags & TH_RST)) {
491 			win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
492 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
493 				win = (long)TCP_MAXWIN << tp->rcv_scale;
494 		}
495 		if (isipv6)
496 			ro6 = &tp->t_inpcb->in6p_route;
497 		else
498 			ro = &tp->t_inpcb->inp_route;
499 	} else {
500 		if (isipv6) {
501 			ro6 = &sro6;
502 			bzero(ro6, sizeof *ro6);
503 		} else {
504 			ro = &sro;
505 			bzero(ro, sizeof *ro);
506 		}
507 	}
508 	if (m == NULL) {
509 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
510 		if (m == NULL)
511 			return;
512 		tlen = 0;
513 		m->m_data += max_linkhdr;
514 		if (isipv6) {
515 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
516 			ip6 = mtod(m, struct ip6_hdr *);
517 			nth = (struct tcphdr *)(ip6 + 1);
518 		} else {
519 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
520 			ip = mtod(m, struct ip *);
521 			nth = (struct tcphdr *)(ip + 1);
522 		}
523 		bcopy(th, nth, sizeof(struct tcphdr));
524 		flags = TH_ACK;
525 	} else {
526 		m_freem(m->m_next);
527 		m->m_next = NULL;
528 		m->m_data = (caddr_t)ipgen;
529 		/* m_len is set later */
530 		tlen = 0;
531 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
532 		if (isipv6) {
533 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
534 			nth = (struct tcphdr *)(ip6 + 1);
535 		} else {
536 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
537 			nth = (struct tcphdr *)(ip + 1);
538 		}
539 		if (th != nth) {
540 			/*
541 			 * this is usually a case when an extension header
542 			 * exists between the IPv6 header and the
543 			 * TCP header.
544 			 */
545 			nth->th_sport = th->th_sport;
546 			nth->th_dport = th->th_dport;
547 		}
548 		xchg(nth->th_dport, nth->th_sport, n_short);
549 #undef xchg
550 	}
551 	if (isipv6) {
552 		ip6->ip6_flow = 0;
553 		ip6->ip6_vfc = IPV6_VERSION;
554 		ip6->ip6_nxt = IPPROTO_TCP;
555 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
556 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
557 	} else {
558 		tlen += sizeof(struct tcpiphdr);
559 		ip->ip_len = tlen;
560 		ip->ip_ttl = ip_defttl;
561 	}
562 	m->m_len = tlen;
563 	m->m_pkthdr.len = tlen;
564 	m->m_pkthdr.rcvif = (struct ifnet *) NULL;
565 	nth->th_seq = htonl(seq);
566 	nth->th_ack = htonl(ack);
567 	nth->th_x2 = 0;
568 	nth->th_off = sizeof(struct tcphdr) >> 2;
569 	nth->th_flags = flags;
570 	if (tp != NULL)
571 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
572 	else
573 		nth->th_win = htons((u_short)win);
574 	nth->th_urp = 0;
575 	if (isipv6) {
576 		nth->th_sum = 0;
577 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
578 					sizeof(struct ip6_hdr),
579 					tlen - sizeof(struct ip6_hdr));
580 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
581 					       (ro6 && ro6->ro_rt) ?
582 					           ro6->ro_rt->rt_ifp : NULL);
583 	} else {
584 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
585 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
586 		m->m_pkthdr.csum_flags = CSUM_TCP;
587 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
588 	}
589 #ifdef TCPDEBUG
590 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
591 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
592 #endif
593 	if (isipv6) {
594 		(void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
595 				 tp ? tp->t_inpcb : NULL);
596 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
597 			RTFREE(ro6->ro_rt);
598 			ro6->ro_rt = NULL;
599 		}
600 	} else {
601 		(void)ip_output(m, NULL, ro, ipflags, NULL,
602 				tp ? tp->t_inpcb : NULL);
603 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
604 			RTFREE(ro->ro_rt);
605 			ro->ro_rt = NULL;
606 		}
607 	}
608 }
609 
610 /*
611  * Create a new TCP control block, making an
612  * empty reassembly queue and hooking it to the argument
613  * protocol control block.  The `inp' parameter must have
614  * come from the zone allocator set up in tcp_init().
615  */
616 struct tcpcb *
617 tcp_newtcpcb(struct inpcb *inp)
618 {
619 	struct inp_tp *it;
620 	struct tcpcb *tp;
621 #ifdef INET6
622 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
623 #else
624 	const boolean_t isipv6 = FALSE;
625 #endif
626 
627 	it = (struct inp_tp *)inp;
628 	tp = &it->tcb;
629 	bzero(tp, sizeof(struct tcpcb));
630 	LIST_INIT(&tp->t_segq);
631 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
632 
633 	/* Set up our timeouts. */
634 	callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
635 	callout_init(tp->tt_persist = &it->inp_tp_persist);
636 	callout_init(tp->tt_keep = &it->inp_tp_keep);
637 	callout_init(tp->tt_2msl = &it->inp_tp_2msl);
638 	callout_init(tp->tt_delack = &it->inp_tp_delack);
639 
640 	if (tcp_do_rfc1323)
641 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
642 	if (tcp_do_rfc1644)
643 		tp->t_flags |= TF_REQ_CC;
644 	tp->t_inpcb = inp;	/* XXX */
645 	/*
646 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
647 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
648 	 * reasonable initial retransmit time.
649 	 */
650 	tp->t_srtt = TCPTV_SRTTBASE;
651 	tp->t_rttvar =
652 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
653 	tp->t_rttmin = tcp_rexmit_min;
654 	tp->t_rxtcur = TCPTV_RTOBASE;
655 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
656 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
657 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
658 	tp->t_rcvtime = ticks;
659 	tp->t_bw_rtttime = ticks;
660 	/*
661 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
662 	 * because the socket may be bound to an IPv6 wildcard address,
663 	 * which may match an IPv4-mapped IPv6 address.
664 	 */
665 	inp->inp_ip_ttl = ip_defttl;
666 	inp->inp_ppcb = (caddr_t)tp;
667 	return (tp);		/* XXX */
668 }
669 
670 /*
671  * Drop a TCP connection, reporting the specified error.
672  * If connection is synchronized, then send a RST to peer.
673  */
674 struct tcpcb *
675 tcp_drop(struct tcpcb *tp, int errno)
676 {
677 	struct socket *so = tp->t_inpcb->inp_socket;
678 
679 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
680 		tp->t_state = TCPS_CLOSED;
681 		(void) tcp_output(tp);
682 		tcpstat.tcps_drops++;
683 	} else
684 		tcpstat.tcps_conndrops++;
685 	if (errno == ETIMEDOUT && tp->t_softerror)
686 		errno = tp->t_softerror;
687 	so->so_error = errno;
688 	return (tcp_close(tp));
689 }
690 
691 #ifdef SMP
692 struct netmsg_remwildcard {
693 	struct lwkt_msg		nm_lmsg;
694 	struct inpcb		*nm_inp;
695 	struct inpcbinfo	*nm_pcbinfo;
696 };
697 
698 static int
699 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
700 {
701 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
702 
703 	in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
704 	lwkt_replymsg(&msg->nm_lmsg, 0);
705 	return (EASYNC);
706 }
707 #endif
708 
709 /*
710  * Close a TCP control block:
711  *	discard all space held by the tcp
712  *	discard internet protocol block
713  *	wake up any sleepers
714  */
715 struct tcpcb *
716 tcp_close(struct tcpcb *tp)
717 {
718 	struct tseg_qent *q;
719 	struct inpcb *inp = tp->t_inpcb;
720 	struct socket *so = inp->inp_socket;
721 	struct rtentry *rt;
722 	boolean_t dosavessthresh;
723 #ifdef SMP
724 	int cpu;
725 #endif
726 #ifdef INET6
727 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
728 #else
729 	const boolean_t isipv6 = FALSE;
730 #endif
731 
732 	/*
733 	 * Make sure that all of our timers are stopped before we
734 	 * delete the PCB.
735 	 */
736 	callout_stop(tp->tt_rexmt);
737 	callout_stop(tp->tt_persist);
738 	callout_stop(tp->tt_keep);
739 	callout_stop(tp->tt_2msl);
740 	callout_stop(tp->tt_delack);
741 
742 	/*
743 	 * If we got enough samples through the srtt filter,
744 	 * save the rtt and rttvar in the routing entry.
745 	 * 'Enough' is arbitrarily defined as the 16 samples.
746 	 * 16 samples is enough for the srtt filter to converge
747 	 * to within 5% of the correct value; fewer samples and
748 	 * we could save a very bogus rtt.
749 	 *
750 	 * Don't update the default route's characteristics and don't
751 	 * update anything that the user "locked".
752 	 */
753 	if (tp->t_rttupdated >= 16) {
754 		u_long i = 0;
755 
756 		if (isipv6) {
757 			struct sockaddr_in6 *sin6;
758 
759 			if ((rt = inp->in6p_route.ro_rt) == NULL)
760 				goto no_valid_rt;
761 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
762 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
763 				goto no_valid_rt;
764 		} else
765 			if ((rt = inp->inp_route.ro_rt) == NULL ||
766 			    ((struct sockaddr_in *)rt_key(rt))->
767 			     sin_addr.s_addr == INADDR_ANY)
768 				goto no_valid_rt;
769 
770 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
771 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
772 			if (rt->rt_rmx.rmx_rtt && i)
773 				/*
774 				 * filter this update to half the old & half
775 				 * the new values, converting scale.
776 				 * See route.h and tcp_var.h for a
777 				 * description of the scaling constants.
778 				 */
779 				rt->rt_rmx.rmx_rtt =
780 				    (rt->rt_rmx.rmx_rtt + i) / 2;
781 			else
782 				rt->rt_rmx.rmx_rtt = i;
783 			tcpstat.tcps_cachedrtt++;
784 		}
785 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
786 			i = tp->t_rttvar *
787 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
788 			if (rt->rt_rmx.rmx_rttvar && i)
789 				rt->rt_rmx.rmx_rttvar =
790 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
791 			else
792 				rt->rt_rmx.rmx_rttvar = i;
793 			tcpstat.tcps_cachedrttvar++;
794 		}
795 		/*
796 		 * The old comment here said:
797 		 * update the pipelimit (ssthresh) if it has been updated
798 		 * already or if a pipesize was specified & the threshhold
799 		 * got below half the pipesize.  I.e., wait for bad news
800 		 * before we start updating, then update on both good
801 		 * and bad news.
802 		 *
803 		 * But we want to save the ssthresh even if no pipesize is
804 		 * specified explicitly in the route, because such
805 		 * connections still have an implicit pipesize specified
806 		 * by the global tcp_sendspace.  In the absence of a reliable
807 		 * way to calculate the pipesize, it will have to do.
808 		 */
809 		i = tp->snd_ssthresh;
810 		if (rt->rt_rmx.rmx_sendpipe != 0)
811 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
812 		else
813 			dosavessthresh = (i < so->so_snd.sb_hiwat/2);
814 		if (dosavessthresh ||
815 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
816 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
817 			/*
818 			 * convert the limit from user data bytes to
819 			 * packets then to packet data bytes.
820 			 */
821 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
822 			if (i < 2)
823 				i = 2;
824 			i *= tp->t_maxseg +
825 			     (isipv6 ?
826 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
827 			      sizeof(struct tcpiphdr));
828 			if (rt->rt_rmx.rmx_ssthresh)
829 				rt->rt_rmx.rmx_ssthresh =
830 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
831 			else
832 				rt->rt_rmx.rmx_ssthresh = i;
833 			tcpstat.tcps_cachedssthresh++;
834 		}
835 	}
836 
837 no_valid_rt:
838 	/* free the reassembly queue, if any */
839 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
840 		LIST_REMOVE(q, tqe_q);
841 		m_freem(q->tqe_m);
842 		FREE(q, M_TSEGQ);
843 		tcp_reass_qsize--;
844 	}
845 	inp->inp_ppcb = NULL;
846 	soisdisconnected(so);
847 
848 #ifdef SMP
849 	if (inp->inp_flags & INP_WILDCARD_MP) {
850 		for (cpu = 0; cpu < ncpus2; cpu ++) {
851 			struct netmsg_remwildcard *msg;
852 
853 			msg = malloc(sizeof(struct netmsg_remwildcard),
854 			    M_LWKTMSG, M_INTWAIT);
855 			lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
856 			    lwkt_cmd_func(in_pcbremwildcardhash_handler),
857 			    lwkt_cmd_op_none);
858 			msg->nm_inp = inp;
859 			msg->nm_pcbinfo = &tcbinfo[cpu];
860 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
861 		}
862 	}
863 #endif
864 
865 #ifdef INET6
866 	if (INP_CHECK_SOCKAF(so, AF_INET6))
867 		in6_pcbdetach(inp);
868 	else
869 #endif
870 		in_pcbdetach(inp);
871 	tcpstat.tcps_closed++;
872 	return (NULL);
873 }
874 
875 static __inline void
876 tcp_drain_oncpu(struct inpcbhead *head)
877 {
878 	struct inpcb *inpb;
879 	struct tcpcb *tcpb;
880 	struct tseg_qent *te;
881 
882 	LIST_FOREACH(inpb, head, inp_list) {
883 		if (inpb->inp_flags & INP_PLACEMARKER)
884 			continue;
885 		if ((tcpb = intotcpcb(inpb))) {
886 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
887 				LIST_REMOVE(te, tqe_q);
888 				m_freem(te->tqe_m);
889 				FREE(te, M_TSEGQ);
890 				tcp_reass_qsize--;
891 			}
892 		}
893 	}
894 }
895 
896 #ifdef SMP
897 struct netmsg_tcp_drain {
898 	struct lwkt_msg		nm_lmsg;
899 	struct inpcbhead	*nm_head;
900 };
901 
902 static int
903 tcp_drain_handler(lwkt_msg_t lmsg)
904 {
905 	struct netmsg_tcp_drain *nm = (void *)lmsg;
906 
907 	tcp_drain_oncpu(nm->nm_head);
908 	lwkt_replymsg(lmsg, 0);
909 	return(EASYNC);
910 }
911 #endif
912 
913 void
914 tcp_drain()
915 {
916 #ifdef SMP
917 	int cpu;
918 #endif
919 
920 	if (!do_tcpdrain)
921 		return;
922 
923 	/*
924 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
925 	 * if there is one...
926 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
927 	 *	reassembly queue should be flushed, but in a situation
928 	 *	where we're really low on mbufs, this is potentially
929 	 *	useful.
930 	 */
931 #ifdef SMP
932 	for (cpu = 0; cpu < ncpus2; cpu++) {
933 		struct netmsg_tcp_drain *msg;
934 
935 		if (cpu == mycpu->gd_cpuid) {
936 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
937 		} else {
938 			msg = malloc(sizeof(struct netmsg_tcp_drain),
939 				    M_LWKTMSG, M_NOWAIT);
940 			if (msg == NULL)
941 				continue;
942 			lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
943 				lwkt_cmd_func(tcp_drain_handler),
944 				lwkt_cmd_op_none);
945 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
946 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
947 		}
948 	}
949 #else
950 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
951 #endif
952 }
953 
954 /*
955  * Notify a tcp user of an asynchronous error;
956  * store error as soft error, but wake up user
957  * (for now, won't do anything until can select for soft error).
958  *
959  * Do not wake up user since there currently is no mechanism for
960  * reporting soft errors (yet - a kqueue filter may be added).
961  */
962 static void
963 tcp_notify(struct inpcb *inp, int error)
964 {
965 	struct tcpcb *tp = intotcpcb(inp);
966 
967 	/*
968 	 * Ignore some errors if we are hooked up.
969 	 * If connection hasn't completed, has retransmitted several times,
970 	 * and receives a second error, give up now.  This is better
971 	 * than waiting a long time to establish a connection that
972 	 * can never complete.
973 	 */
974 	if (tp->t_state == TCPS_ESTABLISHED &&
975 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
976 	      error == EHOSTDOWN)) {
977 		return;
978 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
979 	    tp->t_softerror)
980 		tcp_drop(tp, error);
981 	else
982 		tp->t_softerror = error;
983 #if 0
984 	wakeup((caddr_t) &so->so_timeo);
985 	sorwakeup(so);
986 	sowwakeup(so);
987 #endif
988 }
989 
990 static int
991 tcp_pcblist(SYSCTL_HANDLER_ARGS)
992 {
993 	int error, i, n;
994 	struct inpcb *marker;
995 	struct inpcb *inp;
996 	inp_gen_t gencnt;
997 	struct xinpgen xig;
998 	globaldata_t gd;
999 	int origcpu, ccpu;
1000 
1001 	error = 0;
1002 	n = 0;
1003 
1004 	/*
1005 	 * The process of preparing the TCB list is too time-consuming and
1006 	 * resource-intensive to repeat twice on every request.
1007 	 */
1008 	if (req->oldptr == NULL) {
1009 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1010 			gd = globaldata_find(ccpu);
1011 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1012 		}
1013 		req->oldidx = 2 * ncpus * (sizeof xig) +
1014 		  (n + n/8) * sizeof(struct xtcpcb);
1015 		return (0);
1016 	}
1017 
1018 	if (req->newptr != NULL)
1019 		return (EPERM);
1020 
1021 	marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1022 	marker->inp_flags |= INP_PLACEMARKER;
1023 
1024 	/*
1025 	 * OK, now we're committed to doing something.  Run the inpcb list
1026 	 * for each cpu in the system and construct the output.  Use a
1027 	 * list placemarker to deal with list changes occuring during
1028 	 * copyout blockages (but otherwise depend on being on the correct
1029 	 * cpu to avoid races).
1030 	 */
1031 	origcpu = mycpu->gd_cpuid;
1032 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1033 		globaldata_t rgd;
1034 		caddr_t inp_ppcb;
1035 		struct xtcpcb xt;
1036 		int cpu_id;
1037 
1038 		cpu_id = (origcpu + ccpu) % ncpus;
1039 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1040 			continue;
1041 		rgd = globaldata_find(cpu_id);
1042 		lwkt_setcpu_self(rgd);
1043 
1044 		/* indicate change of CPU */
1045 		cpu_mb1();
1046 
1047 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1048 		n = tcbinfo[cpu_id].ipi_count;
1049 
1050 		xig.xig_len = sizeof xig;
1051 		xig.xig_count = n;
1052 		xig.xig_gen = gencnt;
1053 		xig.xig_sogen = so_gencnt;
1054 		xig.xig_cpu = cpu_id;
1055 		error = SYSCTL_OUT(req, &xig, sizeof xig);
1056 		if (error != 0)
1057 			break;
1058 
1059 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1060 		i = 0;
1061 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1062 			/*
1063 			 * process a snapshot of pcbs, ignoring placemarkers
1064 			 * and using our own to allow SYSCTL_OUT to block.
1065 			 */
1066 			LIST_REMOVE(marker, inp_list);
1067 			LIST_INSERT_AFTER(inp, marker, inp_list);
1068 
1069 			if (inp->inp_flags & INP_PLACEMARKER)
1070 				continue;
1071 			if (inp->inp_gencnt > gencnt)
1072 				continue;
1073 			if (prison_xinpcb(req->td, inp))
1074 				continue;
1075 
1076 			xt.xt_len = sizeof xt;
1077 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1078 			inp_ppcb = inp->inp_ppcb;
1079 			if (inp_ppcb != NULL)
1080 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1081 			else
1082 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1083 			if (inp->inp_socket)
1084 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1085 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1086 				break;
1087 			++i;
1088 		}
1089 		LIST_REMOVE(marker, inp_list);
1090 		if (error == 0 && i < n) {
1091 			bzero(&xt, sizeof(xt));
1092 			xt.xt_len = sizeof(xt);
1093 			while (i < n) {
1094 				error = SYSCTL_OUT(req, &xt, sizeof (xt));
1095 				if (error)
1096 					break;
1097 				++i;
1098 			}
1099 		}
1100 		if (error == 0) {
1101 			/*
1102 			 * Give the user an updated idea of our state.
1103 			 * If the generation differs from what we told
1104 			 * her before, she knows that something happened
1105 			 * while we were processing this request, and it
1106 			 * might be necessary to retry.
1107 			 */
1108 			xig.xig_gen = tcbinfo[cpu_id].ipi_gencnt;
1109 			xig.xig_sogen = so_gencnt;
1110 			xig.xig_count = tcbinfo[cpu_id].ipi_count;
1111 			error = SYSCTL_OUT(req, &xig, sizeof xig);
1112 		}
1113 	}
1114 
1115 	/*
1116 	 * Make sure we are on the same cpu we were on originally, since
1117 	 * higher level callers expect this.  Also don't pollute caches with
1118 	 * migrated userland data by (eventually) returning to userland
1119 	 * on a different cpu.
1120 	 */
1121 	lwkt_setcpu_self(globaldata_find(origcpu));
1122 	free(marker, M_TEMP);
1123 	return (error);
1124 }
1125 
1126 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1127 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1128 
1129 static int
1130 tcp_getcred(SYSCTL_HANDLER_ARGS)
1131 {
1132 	struct sockaddr_in addrs[2];
1133 	struct inpcb *inp;
1134 	int cpu;
1135 	int error, s;
1136 
1137 	error = suser(req->td);
1138 	if (error != 0)
1139 		return (error);
1140 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1141 	if (error != 0)
1142 		return (error);
1143 	s = splnet();
1144 
1145 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1146 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1147 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1148 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1149 	if (inp == NULL || inp->inp_socket == NULL) {
1150 		error = ENOENT;
1151 		goto out;
1152 	}
1153 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1154 out:
1155 	splx(s);
1156 	return (error);
1157 }
1158 
1159 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1160     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1161 
1162 #ifdef INET6
1163 static int
1164 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1165 {
1166 	struct sockaddr_in6 addrs[2];
1167 	struct inpcb *inp;
1168 	int error, s;
1169 	boolean_t mapped = FALSE;
1170 
1171 	error = suser(req->td);
1172 	if (error != 0)
1173 		return (error);
1174 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1175 	if (error != 0)
1176 		return (error);
1177 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1178 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1179 			mapped = TRUE;
1180 		else
1181 			return (EINVAL);
1182 	}
1183 	s = splnet();
1184 	if (mapped) {
1185 		inp = in_pcblookup_hash(&tcbinfo[0],
1186 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1187 		    addrs[1].sin6_port,
1188 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1189 		    addrs[0].sin6_port,
1190 		    0, NULL);
1191 	} else {
1192 		inp = in6_pcblookup_hash(&tcbinfo[0],
1193 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1194 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1195 		    0, NULL);
1196 	}
1197 	if (inp == NULL || inp->inp_socket == NULL) {
1198 		error = ENOENT;
1199 		goto out;
1200 	}
1201 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1202 out:
1203 	splx(s);
1204 	return (error);
1205 }
1206 
1207 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1208 	    0, 0,
1209 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1210 #endif
1211 
1212 void
1213 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1214 {
1215 	struct ip *ip = vip;
1216 	struct tcphdr *th;
1217 	struct in_addr faddr;
1218 	struct inpcb *inp;
1219 	struct tcpcb *tp;
1220 	void (*notify)(struct inpcb *, int) = tcp_notify;
1221 	tcp_seq icmp_seq;
1222 	int cpu;
1223 	int s;
1224 
1225 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1226 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1227 		return;
1228 
1229 	if (cmd == PRC_QUENCH)
1230 		notify = tcp_quench;
1231 	else if (icmp_may_rst &&
1232 		 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT ||
1233 		  cmd == PRC_TIMXCEED_INTRANS) &&
1234 		 ip != NULL)
1235 		notify = tcp_drop_syn_sent;
1236 	else if (cmd == PRC_MSGSIZE)
1237 		notify = tcp_mtudisc;
1238 	else if (PRC_IS_REDIRECT(cmd)) {
1239 		ip = NULL;
1240 		notify = in_rtchange;
1241 	} else if (cmd == PRC_HOSTDEAD)
1242 		ip = NULL;
1243 	else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0)
1244 		return;
1245 	if (ip != NULL) {
1246 		s = splnet();
1247 		th = (struct tcphdr *)((caddr_t)ip +
1248 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1249 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1250 				  ip->ip_src.s_addr, th->th_sport);
1251 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1252 					ip->ip_src, th->th_sport, 0, NULL);
1253 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1254 			icmp_seq = htonl(th->th_seq);
1255 			tp = intotcpcb(inp);
1256 			if (SEQ_GEQ(icmp_seq, tp->snd_una) &&
1257 			    SEQ_LT(icmp_seq, tp->snd_max))
1258 				(*notify)(inp, inetctlerrmap[cmd]);
1259 		} else {
1260 			struct in_conninfo inc;
1261 
1262 			inc.inc_fport = th->th_dport;
1263 			inc.inc_lport = th->th_sport;
1264 			inc.inc_faddr = faddr;
1265 			inc.inc_laddr = ip->ip_src;
1266 #ifdef INET6
1267 			inc.inc_isipv6 = 0;
1268 #endif
1269 			syncache_unreach(&inc, th);
1270 		}
1271 		splx(s);
1272 	} else {
1273 		for (cpu = 0; cpu < ncpus2; cpu++) {
1274 			in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr,
1275 					inetctlerrmap[cmd], notify);
1276 		}
1277 	}
1278 }
1279 
1280 #ifdef INET6
1281 void
1282 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1283 {
1284 	struct tcphdr th;
1285 	void (*notify) (struct inpcb *, int) = tcp_notify;
1286 	struct ip6_hdr *ip6;
1287 	struct mbuf *m;
1288 	struct ip6ctlparam *ip6cp = NULL;
1289 	const struct sockaddr_in6 *sa6_src = NULL;
1290 	int off;
1291 	struct tcp_portonly {
1292 		u_int16_t th_sport;
1293 		u_int16_t th_dport;
1294 	} *thp;
1295 
1296 	if (sa->sa_family != AF_INET6 ||
1297 	    sa->sa_len != sizeof(struct sockaddr_in6))
1298 		return;
1299 
1300 	if (cmd == PRC_QUENCH)
1301 		notify = tcp_quench;
1302 	else if (cmd == PRC_MSGSIZE)
1303 		notify = tcp_mtudisc;
1304 	else if (!PRC_IS_REDIRECT(cmd) &&
1305 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1306 		return;
1307 
1308 	/* if the parameter is from icmp6, decode it. */
1309 	if (d != NULL) {
1310 		ip6cp = (struct ip6ctlparam *)d;
1311 		m = ip6cp->ip6c_m;
1312 		ip6 = ip6cp->ip6c_ip6;
1313 		off = ip6cp->ip6c_off;
1314 		sa6_src = ip6cp->ip6c_src;
1315 	} else {
1316 		m = NULL;
1317 		ip6 = NULL;
1318 		off = 0;	/* fool gcc */
1319 		sa6_src = &sa6_any;
1320 	}
1321 
1322 	if (ip6 != NULL) {
1323 		struct in_conninfo inc;
1324 		/*
1325 		 * XXX: We assume that when IPV6 is non NULL,
1326 		 * M and OFF are valid.
1327 		 */
1328 
1329 		/* check if we can safely examine src and dst ports */
1330 		if (m->m_pkthdr.len < off + sizeof *thp)
1331 			return;
1332 
1333 		bzero(&th, sizeof th);
1334 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1335 
1336 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1337 		    (struct sockaddr *)ip6cp->ip6c_src,
1338 		    th.th_sport, cmd, notify);
1339 
1340 		inc.inc_fport = th.th_dport;
1341 		inc.inc_lport = th.th_sport;
1342 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1343 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1344 		inc.inc_isipv6 = 1;
1345 		syncache_unreach(&inc, &th);
1346 	} else
1347 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1348 		    (const struct sockaddr *)sa6_src, 0, cmd, notify);
1349 }
1350 #endif
1351 
1352 /*
1353  * Following is where TCP initial sequence number generation occurs.
1354  *
1355  * There are two places where we must use initial sequence numbers:
1356  * 1.  In SYN-ACK packets.
1357  * 2.  In SYN packets.
1358  *
1359  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1360  * tcp_syncache.c for details.
1361  *
1362  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1363  * depends on this property.  In addition, these ISNs should be
1364  * unguessable so as to prevent connection hijacking.  To satisfy
1365  * the requirements of this situation, the algorithm outlined in
1366  * RFC 1948 is used to generate sequence numbers.
1367  *
1368  * Implementation details:
1369  *
1370  * Time is based off the system timer, and is corrected so that it
1371  * increases by one megabyte per second.  This allows for proper
1372  * recycling on high speed LANs while still leaving over an hour
1373  * before rollover.
1374  *
1375  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1376  * between seeding of isn_secret.  This is normally set to zero,
1377  * as reseeding should not be necessary.
1378  *
1379  */
1380 
1381 #define	ISN_BYTES_PER_SECOND 1048576
1382 
1383 u_char isn_secret[32];
1384 int isn_last_reseed;
1385 MD5_CTX isn_ctx;
1386 
1387 tcp_seq
1388 tcp_new_isn(struct tcpcb *tp)
1389 {
1390 	u_int32_t md5_buffer[4];
1391 	tcp_seq new_isn;
1392 
1393 	/* Seed if this is the first use, reseed if requested. */
1394 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1395 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1396 		< (u_int)ticks))) {
1397 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1398 		isn_last_reseed = ticks;
1399 	}
1400 
1401 	/* Compute the md5 hash and return the ISN. */
1402 	MD5Init(&isn_ctx);
1403 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1404 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1405 #ifdef INET6
1406 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1407 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1408 			  sizeof(struct in6_addr));
1409 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1410 			  sizeof(struct in6_addr));
1411 	} else
1412 #endif
1413 	{
1414 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1415 			  sizeof(struct in_addr));
1416 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1417 			  sizeof(struct in_addr));
1418 	}
1419 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1420 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1421 	new_isn = (tcp_seq) md5_buffer[0];
1422 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1423 	return (new_isn);
1424 }
1425 
1426 /*
1427  * When a source quench is received, close congestion window
1428  * to one segment.  We will gradually open it again as we proceed.
1429  */
1430 void
1431 tcp_quench(struct inpcb *inp, int errno)
1432 {
1433 	struct tcpcb *tp = intotcpcb(inp);
1434 
1435 	if (tp != NULL)
1436 		tp->snd_cwnd = tp->t_maxseg;
1437 }
1438 
1439 /*
1440  * When a specific ICMP unreachable message is received and the
1441  * connection state is SYN-SENT, drop the connection.  This behavior
1442  * is controlled by the icmp_may_rst sysctl.
1443  */
1444 void
1445 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1446 {
1447 	struct tcpcb *tp = intotcpcb(inp);
1448 
1449 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1450 		tcp_drop(tp, errno);
1451 }
1452 
1453 /*
1454  * When `need fragmentation' ICMP is received, update our idea of the MSS
1455  * based on the new value in the route.  Also nudge TCP to send something,
1456  * since we know the packet we just sent was dropped.
1457  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1458  */
1459 void
1460 tcp_mtudisc(struct inpcb *inp, int errno)
1461 {
1462 	struct tcpcb *tp = intotcpcb(inp);
1463 	struct rtentry *rt;
1464 	struct rmxp_tao *taop;
1465 	struct socket *so = inp->inp_socket;
1466 	int offered;
1467 	int mss;
1468 #ifdef INET6
1469 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1470 #else
1471 	const boolean_t isipv6 = FALSE;
1472 #endif
1473 
1474 	if (tp != NULL) {
1475 		if (isipv6)
1476 			rt = tcp_rtlookup6(&inp->inp_inc);
1477 		else
1478 			rt = tcp_rtlookup(&inp->inp_inc);
1479 		if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) {
1480 			tp->t_maxopd = tp->t_maxseg =
1481 			    isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
1482 			return;
1483 		}
1484 		taop = rmx_taop(rt->rt_rmx);
1485 		offered = taop->tao_mssopt;
1486 		mss = rt->rt_rmx.rmx_mtu -
1487 			(isipv6 ?
1488 			 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1489 			 sizeof(struct tcpiphdr));
1490 
1491 		if (offered != 0)
1492 			mss = min(mss, offered);
1493 		/*
1494 		 * XXX - The above conditional probably violates the TCP
1495 		 * spec.  The problem is that, since we don't know the
1496 		 * other end's MSS, we are supposed to use a conservative
1497 		 * default.  But, if we do that, then MTU discovery will
1498 		 * never actually take place, because the conservative
1499 		 * default is much less than the MTUs typically seen
1500 		 * on the Internet today.  For the moment, we'll sweep
1501 		 * this under the carpet.
1502 		 *
1503 		 * The conservative default might not actually be a problem
1504 		 * if the only case this occurs is when sending an initial
1505 		 * SYN with options and data to a host we've never talked
1506 		 * to before.  Then, they will reply with an MSS value which
1507 		 * will get recorded and the new parameters should get
1508 		 * recomputed.  For Further Study.
1509 		 */
1510 		if (tp->t_maxopd <= mss)
1511 			return;
1512 		tp->t_maxopd = mss;
1513 
1514 		if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1515 		    (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1516 			mss -= TCPOLEN_TSTAMP_APPA;
1517 		if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
1518 		    (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC)
1519 			mss -= TCPOLEN_CC_APPA;
1520 #if	(MCLBYTES & (MCLBYTES - 1)) == 0
1521 		if (mss > MCLBYTES)
1522 			mss &= ~(MCLBYTES - 1);
1523 #else
1524 		if (mss > MCLBYTES)
1525 			mss = mss / MCLBYTES * MCLBYTES;
1526 #endif
1527 		if (so->so_snd.sb_hiwat < mss)
1528 			mss = so->so_snd.sb_hiwat;
1529 
1530 		tp->t_maxseg = mss;
1531 
1532 		tcpstat.tcps_mturesent++;
1533 		tp->t_rtttime = 0;
1534 		tp->snd_nxt = tp->snd_una;
1535 		tcp_output(tp);
1536 	}
1537 }
1538 
1539 /*
1540  * Look-up the routing entry to the peer of this inpcb.  If no route
1541  * is found and it cannot be allocated the return NULL.  This routine
1542  * is called by TCP routines that access the rmx structure and by tcp_mss
1543  * to get the interface MTU.
1544  */
1545 struct rtentry *
1546 tcp_rtlookup(struct in_conninfo *inc)
1547 {
1548 	struct route *ro;
1549 	struct rtentry *rt;
1550 
1551 	ro = &inc->inc_route;
1552 	rt = ro->ro_rt;
1553 	if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1554 		/* No route yet, so try to acquire one */
1555 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1556 			/*
1557 			 * unused portions of the structure MUST be zero'd
1558 			 * out because rtalloc() treats it as opaque data
1559 			 */
1560 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1561 			ro->ro_dst.sa_family = AF_INET;
1562 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1563 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1564 			    inc->inc_faddr;
1565 			rtalloc(ro);
1566 			rt = ro->ro_rt;
1567 		}
1568 	}
1569 	return (rt);
1570 }
1571 
1572 #ifdef INET6
1573 struct rtentry *
1574 tcp_rtlookup6(struct in_conninfo *inc)
1575 {
1576 	struct route_in6 *ro6;
1577 	struct rtentry *rt;
1578 
1579 	ro6 = &inc->inc6_route;
1580 	rt = ro6->ro_rt;
1581 	if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1582 		/* No route yet, so try to acquire one */
1583 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1584 			/*
1585 			 * unused portions of the structure MUST be zero'd
1586 			 * out because rtalloc() treats it as opaque data
1587 			 */
1588 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1589 			ro6->ro_dst.sin6_family = AF_INET6;
1590 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1591 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1592 			rtalloc((struct route *)ro6);
1593 			rt = ro6->ro_rt;
1594 		}
1595 	}
1596 	return (rt);
1597 }
1598 #endif
1599 
1600 #ifdef IPSEC
1601 /* compute ESP/AH header size for TCP, including outer IP header. */
1602 size_t
1603 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1604 {
1605 	struct inpcb *inp;
1606 	struct mbuf *m;
1607 	size_t hdrsiz;
1608 	struct ip *ip;
1609 	struct tcphdr *th;
1610 
1611 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1612 		return (0);
1613 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1614 	if (!m)
1615 		return (0);
1616 
1617 #ifdef INET6
1618 	if (inp->inp_vflag & INP_IPV6) {
1619 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1620 
1621 		th = (struct tcphdr *)(ip6 + 1);
1622 		m->m_pkthdr.len = m->m_len =
1623 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1624 		tcp_fillheaders(tp, ip6, th);
1625 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1626 	} else
1627 #endif
1628 	{
1629 		ip = mtod(m, struct ip *);
1630 		th = (struct tcphdr *)(ip + 1);
1631 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1632 		tcp_fillheaders(tp, ip, th);
1633 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1634 	}
1635 
1636 	m_free(m);
1637 	return (hdrsiz);
1638 }
1639 #endif
1640 
1641 /*
1642  * Return a pointer to the cached information about the remote host.
1643  * The cached information is stored in the protocol specific part of
1644  * the route metrics.
1645  */
1646 struct rmxp_tao *
1647 tcp_gettaocache(struct in_conninfo *inc)
1648 {
1649 	struct rtentry *rt;
1650 
1651 #ifdef INET6
1652 	if (inc->inc_isipv6)
1653 		rt = tcp_rtlookup6(inc);
1654 	else
1655 #endif
1656 		rt = tcp_rtlookup(inc);
1657 
1658 	/* Make sure this is a host route and is up. */
1659 	if (rt == NULL ||
1660 	    (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1661 		return (NULL);
1662 
1663 	return (rmx_taop(rt->rt_rmx));
1664 }
1665 
1666 /*
1667  * Clear all the TAO cache entries, called from tcp_init.
1668  *
1669  * XXX
1670  * This routine is just an empty one, because we assume that the routing
1671  * routing tables are initialized at the same time when TCP, so there is
1672  * nothing in the cache left over.
1673  */
1674 static void
1675 tcp_cleartaocache()
1676 {
1677 }
1678 
1679 /*
1680  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1681  *
1682  * This code attempts to calculate the bandwidth-delay product as a
1683  * means of determining the optimal window size to maximize bandwidth,
1684  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1685  * routers.  This code also does a fairly good job keeping RTTs in check
1686  * across slow links like modems.  We implement an algorithm which is very
1687  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1688  * transmitter side of a TCP connection and so only effects the transmit
1689  * side of the connection.
1690  *
1691  * BACKGROUND:  TCP makes no provision for the management of buffer space
1692  * at the end points or at the intermediate routers and switches.  A TCP
1693  * stream, whether using NewReno or not, will eventually buffer as
1694  * many packets as it is able and the only reason this typically works is
1695  * due to the fairly small default buffers made available for a connection
1696  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1697  * scaling it is now fairly easy for even a single TCP connection to blow-out
1698  * all available buffer space not only on the local interface, but on
1699  * intermediate routers and switches as well.  NewReno makes a misguided
1700  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1701  * then backing off, then steadily increasing the window again until another
1702  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1703  * is only made worse as network loads increase and the idea of intentionally
1704  * blowing out network buffers is, frankly, a terrible way to manage network
1705  * resources.
1706  *
1707  * It is far better to limit the transmit window prior to the failure
1708  * condition being achieved.  There are two general ways to do this:  First
1709  * you can 'scan' through different transmit window sizes and locate the
1710  * point where the RTT stops increasing, indicating that you have filled the
1711  * pipe, then scan backwards until you note that RTT stops decreasing, then
1712  * repeat ad-infinitum.  This method works in principle but has severe
1713  * implementation issues due to RTT variances, timer granularity, and
1714  * instability in the algorithm which can lead to many false positives and
1715  * create oscillations as well as interact badly with other TCP streams
1716  * implementing the same algorithm.
1717  *
1718  * The second method is to limit the window to the bandwidth delay product
1719  * of the link.  This is the method we implement.  RTT variances and our
1720  * own manipulation of the congestion window, bwnd, can potentially
1721  * destabilize the algorithm.  For this reason we have to stabilize the
1722  * elements used to calculate the window.  We do this by using the minimum
1723  * observed RTT, the long term average of the observed bandwidth, and
1724  * by adding two segments worth of slop.  It isn't perfect but it is able
1725  * to react to changing conditions and gives us a very stable basis on
1726  * which to extend the algorithm.
1727  */
1728 void
1729 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1730 {
1731 	u_long bw;
1732 	u_long bwnd;
1733 	int save_ticks;
1734 
1735 	/*
1736 	 * If inflight_enable is disabled in the middle of a tcp connection,
1737 	 * make sure snd_bwnd is effectively disabled.
1738 	 */
1739 	if (!tcp_inflight_enable) {
1740 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1741 		tp->snd_bandwidth = 0;
1742 		return;
1743 	}
1744 
1745 	/*
1746 	 * Figure out the bandwidth.  Due to the tick granularity this
1747 	 * is a very rough number and it MUST be averaged over a fairly
1748 	 * long period of time.  XXX we need to take into account a link
1749 	 * that is not using all available bandwidth, but for now our
1750 	 * slop will ramp us up if this case occurs and the bandwidth later
1751 	 * increases.
1752 	 *
1753 	 * Note: if ticks rollover 'bw' may wind up negative.  We must
1754 	 * effectively reset t_bw_rtttime for this case.
1755 	 */
1756 	save_ticks = ticks;
1757 	if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1758 		return;
1759 
1760 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1761 	    (save_ticks - tp->t_bw_rtttime);
1762 	tp->t_bw_rtttime = save_ticks;
1763 	tp->t_bw_rtseq = ack_seq;
1764 	if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1765 		return;
1766 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1767 
1768 	tp->snd_bandwidth = bw;
1769 
1770 	/*
1771 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1772 	 * segments.  The additional slop puts us squarely in the sweet
1773 	 * spot and also handles the bandwidth run-up case.  Without the
1774 	 * slop we could be locking ourselves into a lower bandwidth.
1775 	 *
1776 	 * Situations Handled:
1777 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1778 	 *	    high speed LANs, allowing larger TCP buffers to be
1779 	 *	    specified, and also does a good job preventing
1780 	 *	    over-queueing of packets over choke points like modems
1781 	 *	    (at least for the transmit side).
1782 	 *
1783 	 *	(2) Is able to handle changing network loads (bandwidth
1784 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1785 	 *	    increases).
1786 	 *
1787 	 *	(3) Theoretically should stabilize in the face of multiple
1788 	 *	    connections implementing the same algorithm (this may need
1789 	 *	    a little work).
1790 	 *
1791 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1792 	 *	    be adjusted with a sysctl but typically only needs to be on
1793 	 *	    very slow connections.  A value no smaller then 5 should
1794 	 *	    be used, but only reduce this default if you have no other
1795 	 *	    choice.
1796 	 */
1797 
1798 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1799 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1800 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1801 #undef USERTT
1802 
1803 	if (tcp_inflight_debug > 0) {
1804 		static int ltime;
1805 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1806 			ltime = ticks;
1807 			printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1808 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1809 		}
1810 	}
1811 	if ((long)bwnd < tcp_inflight_min)
1812 		bwnd = tcp_inflight_min;
1813 	if (bwnd > tcp_inflight_max)
1814 		bwnd = tcp_inflight_max;
1815 	if ((long)bwnd < tp->t_maxseg * 2)
1816 		bwnd = tp->t_maxseg * 2;
1817 	tp->snd_bwnd = bwnd;
1818 }
1819