xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 716da9586b1ae170aab06fc4c81d74c5f3370b58)
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68  * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.62 2008/10/30 10:50:18 sephe Exp $
69  */
70 
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
83 #include <sys/mbuf.h>
84 #ifdef INET6
85 #include <sys/domain.h>
86 #endif
87 #include <sys/proc.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/protosw.h>
91 #include <sys/random.h>
92 #include <sys/in_cksum.h>
93 #include <sys/ktr.h>
94 
95 #include <vm/vm_zone.h>
96 
97 #include <net/route.h>
98 #include <net/if.h>
99 #include <net/netisr.h>
100 
101 #define	_IP_VHL
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
112 #ifdef INET6
113 #include <netinet/icmp6.h>
114 #endif
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
122 #ifdef TCPDEBUG
123 #include <netinet/tcp_debug.h>
124 #endif
125 #include <netinet6/ip6protosw.h>
126 
127 #ifdef IPSEC
128 #include <netinet6/ipsec.h>
129 #ifdef INET6
130 #include <netinet6/ipsec6.h>
131 #endif
132 #endif
133 
134 #ifdef FAST_IPSEC
135 #include <netproto/ipsec/ipsec.h>
136 #ifdef INET6
137 #include <netproto/ipsec/ipsec6.h>
138 #endif
139 #define	IPSEC
140 #endif
141 
142 #include <sys/md5.h>
143 #include <sys/msgport2.h>
144 #include <machine/smp.h>
145 
146 #include <net/netmsg2.h>
147 
148 #if !defined(KTR_TCP)
149 #define KTR_TCP		KTR_ALL
150 #endif
151 KTR_INFO_MASTER(tcp);
152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
155 #define logtcp(name)	KTR_LOG(tcp_ ## name)
156 
157 struct inpcbinfo tcbinfo[MAXCPU];
158 struct tcpcbackqhead tcpcbackq[MAXCPU];
159 
160 int tcp_mpsafe_proto = 0;
161 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
162 
163 static int tcp_mpsafe_thread = 0;
164 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
166 	   &tcp_mpsafe_thread, 0,
167 	   "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
168 
169 int tcp_mssdflt = TCP_MSS;
170 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
171     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
172 
173 #ifdef INET6
174 int tcp_v6mssdflt = TCP6_MSS;
175 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
176     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
177 #endif
178 
179 #if 0
180 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
181 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
182     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
183 #endif
184 
185 int tcp_do_rfc1323 = 1;
186 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
187     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
188 
189 int tcp_do_rfc1644 = 0;
190 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
191     &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
192 
193 static int tcp_tcbhashsize = 0;
194 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
195      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
196 
197 static int do_tcpdrain = 1;
198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
199      "Enable tcp_drain routine for extra help when low on mbufs");
200 
201 /* XXX JH */
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
203     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
204 
205 static int icmp_may_rst = 1;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
207     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
208 
209 static int tcp_isn_reseed_interval = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
211     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
212 
213 /*
214  * TCP bandwidth limiting sysctls.  Note that the default lower bound of
215  * 1024 exists only for debugging.  A good production default would be
216  * something like 6100.
217  */
218 static int tcp_inflight_enable = 0;
219 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
220     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
221 
222 static int tcp_inflight_debug = 0;
223 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
224     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
225 
226 static int tcp_inflight_min = 6144;
227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
228     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
229 
230 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
232     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
233 
234 static int tcp_inflight_stab = 20;
235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
236     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
237 
238 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
239 static struct malloc_pipe tcptemp_mpipe;
240 
241 static void tcp_willblock(int);
242 static void tcp_cleartaocache (void);
243 static void tcp_notify (struct inpcb *, int);
244 
245 struct tcp_stats tcpstats_percpu[MAXCPU];
246 #ifdef SMP
247 static int
248 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
249 {
250 	int cpu, error = 0;
251 
252 	for (cpu = 0; cpu < ncpus; ++cpu) {
253 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
254 					sizeof(struct tcp_stats))))
255 			break;
256 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
257 				       sizeof(struct tcp_stats))))
258 			break;
259 	}
260 
261 	return (error);
262 }
263 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
264     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
265 #else
266 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
267     &tcpstat, tcp_stats, "TCP statistics");
268 #endif
269 
270 /*
271  * Target size of TCP PCB hash tables. Must be a power of two.
272  *
273  * Note that this can be overridden by the kernel environment
274  * variable net.inet.tcp.tcbhashsize
275  */
276 #ifndef TCBHASHSIZE
277 #define	TCBHASHSIZE	512
278 #endif
279 
280 /*
281  * This is the actual shape of what we allocate using the zone
282  * allocator.  Doing it this way allows us to protect both structures
283  * using the same generation count, and also eliminates the overhead
284  * of allocating tcpcbs separately.  By hiding the structure here,
285  * we avoid changing most of the rest of the code (although it needs
286  * to be changed, eventually, for greater efficiency).
287  */
288 #define	ALIGNMENT	32
289 #define	ALIGNM1		(ALIGNMENT - 1)
290 struct	inp_tp {
291 	union {
292 		struct	inpcb inp;
293 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
294 	} inp_tp_u;
295 	struct	tcpcb tcb;
296 	struct	callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
297 	struct	callout inp_tp_delack;
298 };
299 #undef ALIGNMENT
300 #undef ALIGNM1
301 
302 /*
303  * Tcp initialization
304  */
305 void
306 tcp_init(void)
307 {
308 	struct inpcbporthead *porthashbase;
309 	u_long porthashmask;
310 	struct vm_zone *ipi_zone;
311 	int hashsize = TCBHASHSIZE;
312 	int cpu;
313 
314 	/*
315 	 * note: tcptemp is used for keepalives, and it is ok for an
316 	 * allocation to fail so do not specify MPF_INT.
317 	 */
318 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
319 		    25, -1, 0, NULL);
320 
321 	tcp_ccgen = 1;
322 	tcp_cleartaocache();
323 
324 	tcp_delacktime = TCPTV_DELACK;
325 	tcp_keepinit = TCPTV_KEEP_INIT;
326 	tcp_keepidle = TCPTV_KEEP_IDLE;
327 	tcp_keepintvl = TCPTV_KEEPINTVL;
328 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
329 	tcp_msl = TCPTV_MSL;
330 	tcp_rexmit_min = TCPTV_MIN;
331 	tcp_rexmit_slop = TCPTV_CPU_VAR;
332 
333 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
334 	if (!powerof2(hashsize)) {
335 		kprintf("WARNING: TCB hash size not a power of 2\n");
336 		hashsize = 512; /* safe default */
337 	}
338 	tcp_tcbhashsize = hashsize;
339 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
340 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
341 			 ZONE_INTERRUPT, 0);
342 
343 	for (cpu = 0; cpu < ncpus2; cpu++) {
344 		in_pcbinfo_init(&tcbinfo[cpu]);
345 		tcbinfo[cpu].cpu = cpu;
346 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
347 		    &tcbinfo[cpu].hashmask);
348 		tcbinfo[cpu].porthashbase = porthashbase;
349 		tcbinfo[cpu].porthashmask = porthashmask;
350 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
351 		    &tcbinfo[cpu].wildcardhashmask);
352 		tcbinfo[cpu].ipi_zone = ipi_zone;
353 		TAILQ_INIT(&tcpcbackq[cpu]);
354 	}
355 
356 	tcp_reass_maxseg = nmbclusters / 16;
357 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
358 
359 #ifdef INET6
360 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
361 #else
362 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
363 #endif
364 	if (max_protohdr < TCP_MINPROTOHDR)
365 		max_protohdr = TCP_MINPROTOHDR;
366 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
367 		panic("tcp_init");
368 #undef TCP_MINPROTOHDR
369 
370 	/*
371 	 * Initialize TCP statistics counters for each CPU.
372 	 */
373 #ifdef SMP
374 	for (cpu = 0; cpu < ncpus; ++cpu) {
375 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
376 	}
377 #else
378 	bzero(&tcpstat, sizeof(struct tcp_stats));
379 #endif
380 
381 	syncache_init();
382 	tcp_thread_init();
383 }
384 
385 void
386 tcpmsg_service_loop(void *dummy)
387 {
388 	struct netmsg *msg;
389 	int mplocked;
390 
391 	/*
392 	 * Thread was started with TDF_MPSAFE
393 	 */
394 	mplocked = 0;
395 
396 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
397 		do {
398 			logtcp(rxmsg);
399 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
400 						  mplocked);
401 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
402 
403 		logtcp(delayed);
404 		tcp_willblock(mplocked);
405 		logtcp(wait);
406 	}
407 }
408 
409 static void
410 tcp_willblock(int mplocked)
411 {
412 	struct tcpcb *tp;
413 	int cpu = mycpu->gd_cpuid;
414 	int unlock = 0;
415 
416 	if (!mplocked && !tcp_mpsafe_proto) {
417 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
418 			return;
419 
420 		get_mplock();
421 		mplocked = 1;
422 		unlock = 1;
423 	}
424 
425 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
426 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
427 		tp->t_flags &= ~TF_ONOUTPUTQ;
428 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
429 		tcp_output(tp);
430 	}
431 
432 	if (unlock)
433 		rel_mplock();
434 }
435 
436 
437 /*
438  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
439  * tcp_template used to store this data in mbufs, but we now recopy it out
440  * of the tcpcb each time to conserve mbufs.
441  */
442 void
443 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
444 {
445 	struct inpcb *inp = tp->t_inpcb;
446 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
447 
448 #ifdef INET6
449 	if (inp->inp_vflag & INP_IPV6) {
450 		struct ip6_hdr *ip6;
451 
452 		ip6 = (struct ip6_hdr *)ip_ptr;
453 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
454 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
455 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
456 			(IPV6_VERSION & IPV6_VERSION_MASK);
457 		ip6->ip6_nxt = IPPROTO_TCP;
458 		ip6->ip6_plen = sizeof(struct tcphdr);
459 		ip6->ip6_src = inp->in6p_laddr;
460 		ip6->ip6_dst = inp->in6p_faddr;
461 		tcp_hdr->th_sum = 0;
462 	} else
463 #endif
464 	{
465 		struct ip *ip = (struct ip *) ip_ptr;
466 
467 		ip->ip_vhl = IP_VHL_BORING;
468 		ip->ip_tos = 0;
469 		ip->ip_len = 0;
470 		ip->ip_id = 0;
471 		ip->ip_off = 0;
472 		ip->ip_ttl = 0;
473 		ip->ip_sum = 0;
474 		ip->ip_p = IPPROTO_TCP;
475 		ip->ip_src = inp->inp_laddr;
476 		ip->ip_dst = inp->inp_faddr;
477 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
478 				    ip->ip_dst.s_addr,
479 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
480 	}
481 
482 	tcp_hdr->th_sport = inp->inp_lport;
483 	tcp_hdr->th_dport = inp->inp_fport;
484 	tcp_hdr->th_seq = 0;
485 	tcp_hdr->th_ack = 0;
486 	tcp_hdr->th_x2 = 0;
487 	tcp_hdr->th_off = 5;
488 	tcp_hdr->th_flags = 0;
489 	tcp_hdr->th_win = 0;
490 	tcp_hdr->th_urp = 0;
491 }
492 
493 /*
494  * Create template to be used to send tcp packets on a connection.
495  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
496  * use for this function is in keepalives, which use tcp_respond.
497  */
498 struct tcptemp *
499 tcp_maketemplate(struct tcpcb *tp)
500 {
501 	struct tcptemp *tmp;
502 
503 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
504 		return (NULL);
505 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
506 	return (tmp);
507 }
508 
509 void
510 tcp_freetemplate(struct tcptemp *tmp)
511 {
512 	mpipe_free(&tcptemp_mpipe, tmp);
513 }
514 
515 /*
516  * Send a single message to the TCP at address specified by
517  * the given TCP/IP header.  If m == NULL, then we make a copy
518  * of the tcpiphdr at ti and send directly to the addressed host.
519  * This is used to force keep alive messages out using the TCP
520  * template for a connection.  If flags are given then we send
521  * a message back to the TCP which originated the * segment ti,
522  * and discard the mbuf containing it and any other attached mbufs.
523  *
524  * In any case the ack and sequence number of the transmitted
525  * segment are as specified by the parameters.
526  *
527  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
528  */
529 void
530 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
531 	    tcp_seq ack, tcp_seq seq, int flags)
532 {
533 	int tlen;
534 	int win = 0;
535 	struct route *ro = NULL;
536 	struct route sro;
537 	struct ip *ip = ipgen;
538 	struct tcphdr *nth;
539 	int ipflags = 0;
540 	struct route_in6 *ro6 = NULL;
541 	struct route_in6 sro6;
542 	struct ip6_hdr *ip6 = ipgen;
543 	boolean_t use_tmpro = TRUE;
544 #ifdef INET6
545 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
546 #else
547 	const boolean_t isipv6 = FALSE;
548 #endif
549 
550 	if (tp != NULL) {
551 		if (!(flags & TH_RST)) {
552 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
553 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
554 				win = (long)TCP_MAXWIN << tp->rcv_scale;
555 		}
556 		/*
557 		 * Don't use the route cache of a listen socket,
558 		 * it is not MPSAFE; use temporary route cache.
559 		 */
560 		if (tp->t_state != TCPS_LISTEN) {
561 			if (isipv6)
562 				ro6 = &tp->t_inpcb->in6p_route;
563 			else
564 				ro = &tp->t_inpcb->inp_route;
565 			use_tmpro = FALSE;
566 		}
567 	}
568 	if (use_tmpro) {
569 		if (isipv6) {
570 			ro6 = &sro6;
571 			bzero(ro6, sizeof *ro6);
572 		} else {
573 			ro = &sro;
574 			bzero(ro, sizeof *ro);
575 		}
576 	}
577 	if (m == NULL) {
578 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
579 		if (m == NULL)
580 			return;
581 		tlen = 0;
582 		m->m_data += max_linkhdr;
583 		if (isipv6) {
584 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
585 			ip6 = mtod(m, struct ip6_hdr *);
586 			nth = (struct tcphdr *)(ip6 + 1);
587 		} else {
588 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
589 			ip = mtod(m, struct ip *);
590 			nth = (struct tcphdr *)(ip + 1);
591 		}
592 		bcopy(th, nth, sizeof(struct tcphdr));
593 		flags = TH_ACK;
594 	} else {
595 		m_freem(m->m_next);
596 		m->m_next = NULL;
597 		m->m_data = (caddr_t)ipgen;
598 		/* m_len is set later */
599 		tlen = 0;
600 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
601 		if (isipv6) {
602 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
603 			nth = (struct tcphdr *)(ip6 + 1);
604 		} else {
605 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
606 			nth = (struct tcphdr *)(ip + 1);
607 		}
608 		if (th != nth) {
609 			/*
610 			 * this is usually a case when an extension header
611 			 * exists between the IPv6 header and the
612 			 * TCP header.
613 			 */
614 			nth->th_sport = th->th_sport;
615 			nth->th_dport = th->th_dport;
616 		}
617 		xchg(nth->th_dport, nth->th_sport, n_short);
618 #undef xchg
619 	}
620 	if (isipv6) {
621 		ip6->ip6_flow = 0;
622 		ip6->ip6_vfc = IPV6_VERSION;
623 		ip6->ip6_nxt = IPPROTO_TCP;
624 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
625 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
626 	} else {
627 		tlen += sizeof(struct tcpiphdr);
628 		ip->ip_len = tlen;
629 		ip->ip_ttl = ip_defttl;
630 	}
631 	m->m_len = tlen;
632 	m->m_pkthdr.len = tlen;
633 	m->m_pkthdr.rcvif = (struct ifnet *) NULL;
634 	nth->th_seq = htonl(seq);
635 	nth->th_ack = htonl(ack);
636 	nth->th_x2 = 0;
637 	nth->th_off = sizeof(struct tcphdr) >> 2;
638 	nth->th_flags = flags;
639 	if (tp != NULL)
640 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
641 	else
642 		nth->th_win = htons((u_short)win);
643 	nth->th_urp = 0;
644 	if (isipv6) {
645 		nth->th_sum = 0;
646 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
647 					sizeof(struct ip6_hdr),
648 					tlen - sizeof(struct ip6_hdr));
649 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
650 					       (ro6 && ro6->ro_rt) ?
651 						ro6->ro_rt->rt_ifp : NULL);
652 	} else {
653 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
654 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
655 		m->m_pkthdr.csum_flags = CSUM_TCP;
656 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
657 	}
658 #ifdef TCPDEBUG
659 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
660 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
661 #endif
662 	if (isipv6) {
663 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
664 			   tp ? tp->t_inpcb : NULL);
665 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
666 			RTFREE(ro6->ro_rt);
667 			ro6->ro_rt = NULL;
668 		}
669 	} else {
670 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
671 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
672 			RTFREE(ro->ro_rt);
673 			ro->ro_rt = NULL;
674 		}
675 	}
676 }
677 
678 /*
679  * Create a new TCP control block, making an
680  * empty reassembly queue and hooking it to the argument
681  * protocol control block.  The `inp' parameter must have
682  * come from the zone allocator set up in tcp_init().
683  */
684 struct tcpcb *
685 tcp_newtcpcb(struct inpcb *inp)
686 {
687 	struct inp_tp *it;
688 	struct tcpcb *tp;
689 #ifdef INET6
690 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
691 #else
692 	const boolean_t isipv6 = FALSE;
693 #endif
694 
695 	it = (struct inp_tp *)inp;
696 	tp = &it->tcb;
697 	bzero(tp, sizeof(struct tcpcb));
698 	LIST_INIT(&tp->t_segq);
699 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
700 
701 	/* Set up our timeouts. */
702 	callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
703 	callout_init(tp->tt_persist = &it->inp_tp_persist);
704 	callout_init(tp->tt_keep = &it->inp_tp_keep);
705 	callout_init(tp->tt_2msl = &it->inp_tp_2msl);
706 	callout_init(tp->tt_delack = &it->inp_tp_delack);
707 
708 	if (tcp_do_rfc1323)
709 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
710 	if (tcp_do_rfc1644)
711 		tp->t_flags |= TF_REQ_CC;
712 	tp->t_inpcb = inp;	/* XXX */
713 	tp->t_state = TCPS_CLOSED;
714 	/*
715 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
716 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
717 	 * reasonable initial retransmit time.
718 	 */
719 	tp->t_srtt = TCPTV_SRTTBASE;
720 	tp->t_rttvar =
721 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
722 	tp->t_rttmin = tcp_rexmit_min;
723 	tp->t_rxtcur = TCPTV_RTOBASE;
724 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
725 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
726 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
727 	tp->t_rcvtime = ticks;
728 	/*
729 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
730 	 * because the socket may be bound to an IPv6 wildcard address,
731 	 * which may match an IPv4-mapped IPv6 address.
732 	 */
733 	inp->inp_ip_ttl = ip_defttl;
734 	inp->inp_ppcb = tp;
735 	tcp_sack_tcpcb_init(tp);
736 	return (tp);		/* XXX */
737 }
738 
739 /*
740  * Drop a TCP connection, reporting the specified error.
741  * If connection is synchronized, then send a RST to peer.
742  */
743 struct tcpcb *
744 tcp_drop(struct tcpcb *tp, int error)
745 {
746 	struct socket *so = tp->t_inpcb->inp_socket;
747 
748 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
749 		tp->t_state = TCPS_CLOSED;
750 		tcp_output(tp);
751 		tcpstat.tcps_drops++;
752 	} else
753 		tcpstat.tcps_conndrops++;
754 	if (error == ETIMEDOUT && tp->t_softerror)
755 		error = tp->t_softerror;
756 	so->so_error = error;
757 	return (tcp_close(tp));
758 }
759 
760 #ifdef SMP
761 
762 struct netmsg_remwildcard {
763 	struct netmsg		nm_netmsg;
764 	struct inpcb		*nm_inp;
765 	struct inpcbinfo	*nm_pcbinfo;
766 #if defined(INET6)
767 	int			nm_isinet6;
768 #else
769 	int			nm_unused01;
770 #endif
771 };
772 
773 /*
774  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
775  * inp can be detached.  We do this by cycling through the cpus, ending up
776  * on the cpu controlling the inp last and then doing the disconnect.
777  */
778 static void
779 in_pcbremwildcardhash_handler(struct netmsg *msg0)
780 {
781 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
782 	int cpu;
783 
784 	cpu = msg->nm_pcbinfo->cpu;
785 
786 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
787 		/* note: detach removes any wildcard hash entry */
788 #ifdef INET6
789 		if (msg->nm_isinet6)
790 			in6_pcbdetach(msg->nm_inp);
791 		else
792 #endif
793 			in_pcbdetach(msg->nm_inp);
794 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
795 	} else {
796 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
797 		cpu = (cpu + 1) % ncpus2;
798 		msg->nm_pcbinfo = &tcbinfo[cpu];
799 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
800 	}
801 }
802 
803 #endif
804 
805 /*
806  * Close a TCP control block:
807  *	discard all space held by the tcp
808  *	discard internet protocol block
809  *	wake up any sleepers
810  */
811 struct tcpcb *
812 tcp_close(struct tcpcb *tp)
813 {
814 	struct tseg_qent *q;
815 	struct inpcb *inp = tp->t_inpcb;
816 	struct socket *so = inp->inp_socket;
817 	struct rtentry *rt;
818 	boolean_t dosavessthresh;
819 #ifdef SMP
820 	int cpu;
821 #endif
822 #ifdef INET6
823 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
824 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
825 #else
826 	const boolean_t isipv6 = FALSE;
827 #endif
828 
829 	/*
830 	 * The tp is not instantly destroyed in the wildcard case.  Setting
831 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
832 	 * messing with it, though it should be noted that this change may
833 	 * not take effect on other cpus until we have chained the wildcard
834 	 * hash removal.
835 	 *
836 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
837 	 * update and prevent other tcp protocol threads from accepting new
838 	 * connections on the listen socket we might be trying to close down.
839 	 */
840 	KKASSERT(tp->t_state != TCPS_TERMINATING);
841 	tp->t_state = TCPS_TERMINATING;
842 
843 	/*
844 	 * Make sure that all of our timers are stopped before we
845 	 * delete the PCB.
846 	 */
847 	callout_stop(tp->tt_rexmt);
848 	callout_stop(tp->tt_persist);
849 	callout_stop(tp->tt_keep);
850 	callout_stop(tp->tt_2msl);
851 	callout_stop(tp->tt_delack);
852 
853 	if (tp->t_flags & TF_ONOUTPUTQ) {
854 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
855 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
856 		tp->t_flags &= ~TF_ONOUTPUTQ;
857 	}
858 
859 	/*
860 	 * If we got enough samples through the srtt filter,
861 	 * save the rtt and rttvar in the routing entry.
862 	 * 'Enough' is arbitrarily defined as the 16 samples.
863 	 * 16 samples is enough for the srtt filter to converge
864 	 * to within 5% of the correct value; fewer samples and
865 	 * we could save a very bogus rtt.
866 	 *
867 	 * Don't update the default route's characteristics and don't
868 	 * update anything that the user "locked".
869 	 */
870 	if (tp->t_rttupdated >= 16) {
871 		u_long i = 0;
872 
873 		if (isipv6) {
874 			struct sockaddr_in6 *sin6;
875 
876 			if ((rt = inp->in6p_route.ro_rt) == NULL)
877 				goto no_valid_rt;
878 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
879 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
880 				goto no_valid_rt;
881 		} else
882 			if ((rt = inp->inp_route.ro_rt) == NULL ||
883 			    ((struct sockaddr_in *)rt_key(rt))->
884 			     sin_addr.s_addr == INADDR_ANY)
885 				goto no_valid_rt;
886 
887 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
888 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
889 			if (rt->rt_rmx.rmx_rtt && i)
890 				/*
891 				 * filter this update to half the old & half
892 				 * the new values, converting scale.
893 				 * See route.h and tcp_var.h for a
894 				 * description of the scaling constants.
895 				 */
896 				rt->rt_rmx.rmx_rtt =
897 				    (rt->rt_rmx.rmx_rtt + i) / 2;
898 			else
899 				rt->rt_rmx.rmx_rtt = i;
900 			tcpstat.tcps_cachedrtt++;
901 		}
902 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
903 			i = tp->t_rttvar *
904 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
905 			if (rt->rt_rmx.rmx_rttvar && i)
906 				rt->rt_rmx.rmx_rttvar =
907 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
908 			else
909 				rt->rt_rmx.rmx_rttvar = i;
910 			tcpstat.tcps_cachedrttvar++;
911 		}
912 		/*
913 		 * The old comment here said:
914 		 * update the pipelimit (ssthresh) if it has been updated
915 		 * already or if a pipesize was specified & the threshhold
916 		 * got below half the pipesize.  I.e., wait for bad news
917 		 * before we start updating, then update on both good
918 		 * and bad news.
919 		 *
920 		 * But we want to save the ssthresh even if no pipesize is
921 		 * specified explicitly in the route, because such
922 		 * connections still have an implicit pipesize specified
923 		 * by the global tcp_sendspace.  In the absence of a reliable
924 		 * way to calculate the pipesize, it will have to do.
925 		 */
926 		i = tp->snd_ssthresh;
927 		if (rt->rt_rmx.rmx_sendpipe != 0)
928 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
929 		else
930 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
931 		if (dosavessthresh ||
932 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
933 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
934 			/*
935 			 * convert the limit from user data bytes to
936 			 * packets then to packet data bytes.
937 			 */
938 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
939 			if (i < 2)
940 				i = 2;
941 			i *= tp->t_maxseg +
942 			     (isipv6 ?
943 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
944 			      sizeof(struct tcpiphdr));
945 			if (rt->rt_rmx.rmx_ssthresh)
946 				rt->rt_rmx.rmx_ssthresh =
947 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
948 			else
949 				rt->rt_rmx.rmx_ssthresh = i;
950 			tcpstat.tcps_cachedssthresh++;
951 		}
952 	}
953 
954 no_valid_rt:
955 	/* free the reassembly queue, if any */
956 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
957 		LIST_REMOVE(q, tqe_q);
958 		m_freem(q->tqe_m);
959 		FREE(q, M_TSEGQ);
960 		tcp_reass_qsize--;
961 	}
962 	/* throw away SACK blocks in scoreboard*/
963 	if (TCP_DO_SACK(tp))
964 		tcp_sack_cleanup(&tp->scb);
965 
966 	inp->inp_ppcb = NULL;
967 	soisdisconnected(so);
968 	/*
969 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
970 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
971 	 * for each protocol thread and must be removed in the context of
972 	 * that thread.  This is accomplished by chaining the message
973 	 * through the cpus.
974 	 *
975 	 * If the inp is not wildcarded we simply detach, which will remove
976 	 * the any hashes still present for this inp.
977 	 */
978 #ifdef SMP
979 	if (inp->inp_flags & INP_WILDCARD_MP) {
980 		struct netmsg_remwildcard *msg;
981 
982 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
983 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
984 			      M_LWKTMSG, M_INTWAIT);
985 		netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
986 			    in_pcbremwildcardhash_handler);
987 #ifdef INET6
988 		msg->nm_isinet6 = isafinet6;
989 #endif
990 		msg->nm_inp = inp;
991 		msg->nm_pcbinfo = &tcbinfo[cpu];
992 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
993 	} else
994 #endif
995 	{
996 		/* note: detach removes any wildcard hash entry */
997 #ifdef INET6
998 		if (isafinet6)
999 			in6_pcbdetach(inp);
1000 		else
1001 #endif
1002 			in_pcbdetach(inp);
1003 	}
1004 	tcpstat.tcps_closed++;
1005 	return (NULL);
1006 }
1007 
1008 static __inline void
1009 tcp_drain_oncpu(struct inpcbhead *head)
1010 {
1011 	struct inpcb *inpb;
1012 	struct tcpcb *tcpb;
1013 	struct tseg_qent *te;
1014 
1015 	LIST_FOREACH(inpb, head, inp_list) {
1016 		if (inpb->inp_flags & INP_PLACEMARKER)
1017 			continue;
1018 		if ((tcpb = intotcpcb(inpb))) {
1019 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1020 				LIST_REMOVE(te, tqe_q);
1021 				m_freem(te->tqe_m);
1022 				FREE(te, M_TSEGQ);
1023 				tcp_reass_qsize--;
1024 			}
1025 		}
1026 	}
1027 }
1028 
1029 #ifdef SMP
1030 struct netmsg_tcp_drain {
1031 	struct netmsg		nm_netmsg;
1032 	struct inpcbhead	*nm_head;
1033 };
1034 
1035 static void
1036 tcp_drain_handler(netmsg_t netmsg)
1037 {
1038 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1039 
1040 	tcp_drain_oncpu(nm->nm_head);
1041 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1042 }
1043 #endif
1044 
1045 void
1046 tcp_drain(void)
1047 {
1048 #ifdef SMP
1049 	int cpu;
1050 #endif
1051 
1052 	if (!do_tcpdrain)
1053 		return;
1054 
1055 	/*
1056 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1057 	 * if there is one...
1058 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1059 	 *	reassembly queue should be flushed, but in a situation
1060 	 *	where we're really low on mbufs, this is potentially
1061 	 *	useful.
1062 	 */
1063 #ifdef SMP
1064 	for (cpu = 0; cpu < ncpus2; cpu++) {
1065 		struct netmsg_tcp_drain *msg;
1066 
1067 		if (cpu == mycpu->gd_cpuid) {
1068 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1069 		} else {
1070 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1071 				    M_LWKTMSG, M_NOWAIT);
1072 			if (msg == NULL)
1073 				continue;
1074 			netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1075 				    tcp_drain_handler);
1076 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1077 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1078 		}
1079 	}
1080 #else
1081 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1082 #endif
1083 }
1084 
1085 /*
1086  * Notify a tcp user of an asynchronous error;
1087  * store error as soft error, but wake up user
1088  * (for now, won't do anything until can select for soft error).
1089  *
1090  * Do not wake up user since there currently is no mechanism for
1091  * reporting soft errors (yet - a kqueue filter may be added).
1092  */
1093 static void
1094 tcp_notify(struct inpcb *inp, int error)
1095 {
1096 	struct tcpcb *tp = intotcpcb(inp);
1097 
1098 	/*
1099 	 * Ignore some errors if we are hooked up.
1100 	 * If connection hasn't completed, has retransmitted several times,
1101 	 * and receives a second error, give up now.  This is better
1102 	 * than waiting a long time to establish a connection that
1103 	 * can never complete.
1104 	 */
1105 	if (tp->t_state == TCPS_ESTABLISHED &&
1106 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1107 	      error == EHOSTDOWN)) {
1108 		return;
1109 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1110 	    tp->t_softerror)
1111 		tcp_drop(tp, error);
1112 	else
1113 		tp->t_softerror = error;
1114 #if 0
1115 	wakeup(&so->so_timeo);
1116 	sorwakeup(so);
1117 	sowwakeup(so);
1118 #endif
1119 }
1120 
1121 static int
1122 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1123 {
1124 	int error, i, n;
1125 	struct inpcb *marker;
1126 	struct inpcb *inp;
1127 	inp_gen_t gencnt;
1128 	globaldata_t gd;
1129 	int origcpu, ccpu;
1130 
1131 	error = 0;
1132 	n = 0;
1133 
1134 	/*
1135 	 * The process of preparing the TCB list is too time-consuming and
1136 	 * resource-intensive to repeat twice on every request.
1137 	 */
1138 	if (req->oldptr == NULL) {
1139 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1140 			gd = globaldata_find(ccpu);
1141 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1142 		}
1143 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1144 		return (0);
1145 	}
1146 
1147 	if (req->newptr != NULL)
1148 		return (EPERM);
1149 
1150 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1151 	marker->inp_flags |= INP_PLACEMARKER;
1152 
1153 	/*
1154 	 * OK, now we're committed to doing something.  Run the inpcb list
1155 	 * for each cpu in the system and construct the output.  Use a
1156 	 * list placemarker to deal with list changes occuring during
1157 	 * copyout blockages (but otherwise depend on being on the correct
1158 	 * cpu to avoid races).
1159 	 */
1160 	origcpu = mycpu->gd_cpuid;
1161 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1162 		globaldata_t rgd;
1163 		caddr_t inp_ppcb;
1164 		struct xtcpcb xt;
1165 		int cpu_id;
1166 
1167 		cpu_id = (origcpu + ccpu) % ncpus;
1168 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1169 			continue;
1170 		rgd = globaldata_find(cpu_id);
1171 		lwkt_setcpu_self(rgd);
1172 
1173 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1174 		n = tcbinfo[cpu_id].ipi_count;
1175 
1176 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1177 		i = 0;
1178 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1179 			/*
1180 			 * process a snapshot of pcbs, ignoring placemarkers
1181 			 * and using our own to allow SYSCTL_OUT to block.
1182 			 */
1183 			LIST_REMOVE(marker, inp_list);
1184 			LIST_INSERT_AFTER(inp, marker, inp_list);
1185 
1186 			if (inp->inp_flags & INP_PLACEMARKER)
1187 				continue;
1188 			if (inp->inp_gencnt > gencnt)
1189 				continue;
1190 			if (prison_xinpcb(req->td, inp))
1191 				continue;
1192 
1193 			xt.xt_len = sizeof xt;
1194 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1195 			inp_ppcb = inp->inp_ppcb;
1196 			if (inp_ppcb != NULL)
1197 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1198 			else
1199 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1200 			if (inp->inp_socket)
1201 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1202 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1203 				break;
1204 			++i;
1205 		}
1206 		LIST_REMOVE(marker, inp_list);
1207 		if (error == 0 && i < n) {
1208 			bzero(&xt, sizeof xt);
1209 			xt.xt_len = sizeof xt;
1210 			while (i < n) {
1211 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1212 				if (error)
1213 					break;
1214 				++i;
1215 			}
1216 		}
1217 	}
1218 
1219 	/*
1220 	 * Make sure we are on the same cpu we were on originally, since
1221 	 * higher level callers expect this.  Also don't pollute caches with
1222 	 * migrated userland data by (eventually) returning to userland
1223 	 * on a different cpu.
1224 	 */
1225 	lwkt_setcpu_self(globaldata_find(origcpu));
1226 	kfree(marker, M_TEMP);
1227 	return (error);
1228 }
1229 
1230 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1231 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1232 
1233 static int
1234 tcp_getcred(SYSCTL_HANDLER_ARGS)
1235 {
1236 	struct sockaddr_in addrs[2];
1237 	struct inpcb *inp;
1238 	int cpu;
1239 	int error;
1240 
1241 	error = suser(req->td);
1242 	if (error != 0)
1243 		return (error);
1244 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1245 	if (error != 0)
1246 		return (error);
1247 	crit_enter();
1248 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1249 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1250 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1251 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1252 	if (inp == NULL || inp->inp_socket == NULL) {
1253 		error = ENOENT;
1254 		goto out;
1255 	}
1256 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1257 out:
1258 	crit_exit();
1259 	return (error);
1260 }
1261 
1262 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1263     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1264 
1265 #ifdef INET6
1266 static int
1267 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1268 {
1269 	struct sockaddr_in6 addrs[2];
1270 	struct inpcb *inp;
1271 	int error;
1272 	boolean_t mapped = FALSE;
1273 
1274 	error = suser(req->td);
1275 	if (error != 0)
1276 		return (error);
1277 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1278 	if (error != 0)
1279 		return (error);
1280 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1281 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1282 			mapped = TRUE;
1283 		else
1284 			return (EINVAL);
1285 	}
1286 	crit_enter();
1287 	if (mapped) {
1288 		inp = in_pcblookup_hash(&tcbinfo[0],
1289 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1290 		    addrs[1].sin6_port,
1291 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1292 		    addrs[0].sin6_port,
1293 		    0, NULL);
1294 	} else {
1295 		inp = in6_pcblookup_hash(&tcbinfo[0],
1296 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1297 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1298 		    0, NULL);
1299 	}
1300 	if (inp == NULL || inp->inp_socket == NULL) {
1301 		error = ENOENT;
1302 		goto out;
1303 	}
1304 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1305 out:
1306 	crit_exit();
1307 	return (error);
1308 }
1309 
1310 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1311 	    0, 0,
1312 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1313 #endif
1314 
1315 void
1316 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1317 {
1318 	struct ip *ip = vip;
1319 	struct tcphdr *th;
1320 	struct in_addr faddr;
1321 	struct inpcb *inp;
1322 	struct tcpcb *tp;
1323 	void (*notify)(struct inpcb *, int) = tcp_notify;
1324 	tcp_seq icmpseq;
1325 	int arg, cpu;
1326 
1327 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1328 		return;
1329 	}
1330 
1331 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1332 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1333 		return;
1334 
1335 	arg = inetctlerrmap[cmd];
1336 	if (cmd == PRC_QUENCH) {
1337 		notify = tcp_quench;
1338 	} else if (icmp_may_rst &&
1339 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1340 		    cmd == PRC_UNREACH_PORT ||
1341 		    cmd == PRC_TIMXCEED_INTRANS) &&
1342 		   ip != NULL) {
1343 		notify = tcp_drop_syn_sent;
1344 	} else if (cmd == PRC_MSGSIZE) {
1345 		struct icmp *icmp = (struct icmp *)
1346 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1347 
1348 		arg = ntohs(icmp->icmp_nextmtu);
1349 		notify = tcp_mtudisc;
1350 	} else if (PRC_IS_REDIRECT(cmd)) {
1351 		ip = NULL;
1352 		notify = in_rtchange;
1353 	} else if (cmd == PRC_HOSTDEAD) {
1354 		ip = NULL;
1355 	}
1356 
1357 	if (ip != NULL) {
1358 		crit_enter();
1359 		th = (struct tcphdr *)((caddr_t)ip +
1360 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1361 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1362 				  ip->ip_src.s_addr, th->th_sport);
1363 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1364 					ip->ip_src, th->th_sport, 0, NULL);
1365 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1366 			icmpseq = htonl(th->th_seq);
1367 			tp = intotcpcb(inp);
1368 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1369 			    SEQ_LT(icmpseq, tp->snd_max))
1370 				(*notify)(inp, arg);
1371 		} else {
1372 			struct in_conninfo inc;
1373 
1374 			inc.inc_fport = th->th_dport;
1375 			inc.inc_lport = th->th_sport;
1376 			inc.inc_faddr = faddr;
1377 			inc.inc_laddr = ip->ip_src;
1378 #ifdef INET6
1379 			inc.inc_isipv6 = 0;
1380 #endif
1381 			syncache_unreach(&inc, th);
1382 		}
1383 		crit_exit();
1384 	} else {
1385 		for (cpu = 0; cpu < ncpus2; cpu++) {
1386 			in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1387 					notify);
1388 		}
1389 	}
1390 }
1391 
1392 #ifdef INET6
1393 void
1394 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1395 {
1396 	struct tcphdr th;
1397 	void (*notify) (struct inpcb *, int) = tcp_notify;
1398 	struct ip6_hdr *ip6;
1399 	struct mbuf *m;
1400 	struct ip6ctlparam *ip6cp = NULL;
1401 	const struct sockaddr_in6 *sa6_src = NULL;
1402 	int off;
1403 	struct tcp_portonly {
1404 		u_int16_t th_sport;
1405 		u_int16_t th_dport;
1406 	} *thp;
1407 	int arg;
1408 
1409 	if (sa->sa_family != AF_INET6 ||
1410 	    sa->sa_len != sizeof(struct sockaddr_in6))
1411 		return;
1412 
1413 	arg = 0;
1414 	if (cmd == PRC_QUENCH)
1415 		notify = tcp_quench;
1416 	else if (cmd == PRC_MSGSIZE) {
1417 		struct ip6ctlparam *ip6cp = d;
1418 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1419 
1420 		arg = ntohl(icmp6->icmp6_mtu);
1421 		notify = tcp_mtudisc;
1422 	} else if (!PRC_IS_REDIRECT(cmd) &&
1423 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1424 		return;
1425 	}
1426 
1427 	/* if the parameter is from icmp6, decode it. */
1428 	if (d != NULL) {
1429 		ip6cp = (struct ip6ctlparam *)d;
1430 		m = ip6cp->ip6c_m;
1431 		ip6 = ip6cp->ip6c_ip6;
1432 		off = ip6cp->ip6c_off;
1433 		sa6_src = ip6cp->ip6c_src;
1434 	} else {
1435 		m = NULL;
1436 		ip6 = NULL;
1437 		off = 0;	/* fool gcc */
1438 		sa6_src = &sa6_any;
1439 	}
1440 
1441 	if (ip6 != NULL) {
1442 		struct in_conninfo inc;
1443 		/*
1444 		 * XXX: We assume that when IPV6 is non NULL,
1445 		 * M and OFF are valid.
1446 		 */
1447 
1448 		/* check if we can safely examine src and dst ports */
1449 		if (m->m_pkthdr.len < off + sizeof *thp)
1450 			return;
1451 
1452 		bzero(&th, sizeof th);
1453 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1454 
1455 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1456 		    (struct sockaddr *)ip6cp->ip6c_src,
1457 		    th.th_sport, cmd, arg, notify);
1458 
1459 		inc.inc_fport = th.th_dport;
1460 		inc.inc_lport = th.th_sport;
1461 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1462 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1463 		inc.inc_isipv6 = 1;
1464 		syncache_unreach(&inc, &th);
1465 	} else
1466 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1467 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1468 }
1469 #endif
1470 
1471 /*
1472  * Following is where TCP initial sequence number generation occurs.
1473  *
1474  * There are two places where we must use initial sequence numbers:
1475  * 1.  In SYN-ACK packets.
1476  * 2.  In SYN packets.
1477  *
1478  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1479  * tcp_syncache.c for details.
1480  *
1481  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1482  * depends on this property.  In addition, these ISNs should be
1483  * unguessable so as to prevent connection hijacking.  To satisfy
1484  * the requirements of this situation, the algorithm outlined in
1485  * RFC 1948 is used to generate sequence numbers.
1486  *
1487  * Implementation details:
1488  *
1489  * Time is based off the system timer, and is corrected so that it
1490  * increases by one megabyte per second.  This allows for proper
1491  * recycling on high speed LANs while still leaving over an hour
1492  * before rollover.
1493  *
1494  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1495  * between seeding of isn_secret.  This is normally set to zero,
1496  * as reseeding should not be necessary.
1497  *
1498  */
1499 
1500 #define	ISN_BYTES_PER_SECOND 1048576
1501 
1502 u_char isn_secret[32];
1503 int isn_last_reseed;
1504 MD5_CTX isn_ctx;
1505 
1506 tcp_seq
1507 tcp_new_isn(struct tcpcb *tp)
1508 {
1509 	u_int32_t md5_buffer[4];
1510 	tcp_seq new_isn;
1511 
1512 	/* Seed if this is the first use, reseed if requested. */
1513 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1514 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1515 		< (u_int)ticks))) {
1516 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1517 		isn_last_reseed = ticks;
1518 	}
1519 
1520 	/* Compute the md5 hash and return the ISN. */
1521 	MD5Init(&isn_ctx);
1522 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1523 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1524 #ifdef INET6
1525 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1526 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1527 			  sizeof(struct in6_addr));
1528 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1529 			  sizeof(struct in6_addr));
1530 	} else
1531 #endif
1532 	{
1533 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1534 			  sizeof(struct in_addr));
1535 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1536 			  sizeof(struct in_addr));
1537 	}
1538 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1539 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1540 	new_isn = (tcp_seq) md5_buffer[0];
1541 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1542 	return (new_isn);
1543 }
1544 
1545 /*
1546  * When a source quench is received, close congestion window
1547  * to one segment.  We will gradually open it again as we proceed.
1548  */
1549 void
1550 tcp_quench(struct inpcb *inp, int error)
1551 {
1552 	struct tcpcb *tp = intotcpcb(inp);
1553 
1554 	if (tp != NULL) {
1555 		tp->snd_cwnd = tp->t_maxseg;
1556 		tp->snd_wacked = 0;
1557 	}
1558 }
1559 
1560 /*
1561  * When a specific ICMP unreachable message is received and the
1562  * connection state is SYN-SENT, drop the connection.  This behavior
1563  * is controlled by the icmp_may_rst sysctl.
1564  */
1565 void
1566 tcp_drop_syn_sent(struct inpcb *inp, int error)
1567 {
1568 	struct tcpcb *tp = intotcpcb(inp);
1569 
1570 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1571 		tcp_drop(tp, error);
1572 }
1573 
1574 /*
1575  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1576  * based on the new value in the route.  Also nudge TCP to send something,
1577  * since we know the packet we just sent was dropped.
1578  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1579  */
1580 void
1581 tcp_mtudisc(struct inpcb *inp, int mtu)
1582 {
1583 	struct tcpcb *tp = intotcpcb(inp);
1584 	struct rtentry *rt;
1585 	struct socket *so = inp->inp_socket;
1586 	int maxopd, mss;
1587 #ifdef INET6
1588 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1589 #else
1590 	const boolean_t isipv6 = FALSE;
1591 #endif
1592 
1593 	if (tp == NULL)
1594 		return;
1595 
1596 	/*
1597 	 * If no MTU is provided in the ICMP message, use the
1598 	 * next lower likely value, as specified in RFC 1191.
1599 	 */
1600 	if (mtu == 0) {
1601 		int oldmtu;
1602 
1603 		oldmtu = tp->t_maxopd +
1604 		    (isipv6 ?
1605 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1606 		     sizeof(struct tcpiphdr));
1607 		mtu = ip_next_mtu(oldmtu, 0);
1608 	}
1609 
1610 	if (isipv6)
1611 		rt = tcp_rtlookup6(&inp->inp_inc);
1612 	else
1613 		rt = tcp_rtlookup(&inp->inp_inc);
1614 	if (rt != NULL) {
1615 		struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1616 
1617 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1618 			mtu = rt->rt_rmx.rmx_mtu;
1619 
1620 		maxopd = mtu -
1621 		    (isipv6 ?
1622 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1623 		     sizeof(struct tcpiphdr));
1624 
1625 		/*
1626 		 * XXX - The following conditional probably violates the TCP
1627 		 * spec.  The problem is that, since we don't know the
1628 		 * other end's MSS, we are supposed to use a conservative
1629 		 * default.  But, if we do that, then MTU discovery will
1630 		 * never actually take place, because the conservative
1631 		 * default is much less than the MTUs typically seen
1632 		 * on the Internet today.  For the moment, we'll sweep
1633 		 * this under the carpet.
1634 		 *
1635 		 * The conservative default might not actually be a problem
1636 		 * if the only case this occurs is when sending an initial
1637 		 * SYN with options and data to a host we've never talked
1638 		 * to before.  Then, they will reply with an MSS value which
1639 		 * will get recorded and the new parameters should get
1640 		 * recomputed.  For Further Study.
1641 		 */
1642 		if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1643 			maxopd = taop->tao_mssopt;
1644 	} else
1645 		maxopd = mtu -
1646 		    (isipv6 ?
1647 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1648 		     sizeof(struct tcpiphdr));
1649 
1650 	if (tp->t_maxopd <= maxopd)
1651 		return;
1652 	tp->t_maxopd = maxopd;
1653 
1654 	mss = maxopd;
1655 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1656 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1657 		mss -= TCPOLEN_TSTAMP_APPA;
1658 
1659 	if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1660 			   (TF_REQ_CC | TF_RCVD_CC))
1661 		mss -= TCPOLEN_CC_APPA;
1662 
1663 	/* round down to multiple of MCLBYTES */
1664 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1665 	if (mss > MCLBYTES)
1666 		mss &= ~(MCLBYTES - 1);
1667 #else
1668 	if (mss > MCLBYTES)
1669 		mss = (mss / MCLBYTES) * MCLBYTES;
1670 #endif
1671 
1672 	if (so->so_snd.ssb_hiwat < mss)
1673 		mss = so->so_snd.ssb_hiwat;
1674 
1675 	tp->t_maxseg = mss;
1676 	tp->t_rtttime = 0;
1677 	tp->snd_nxt = tp->snd_una;
1678 	tcp_output(tp);
1679 	tcpstat.tcps_mturesent++;
1680 }
1681 
1682 /*
1683  * Look-up the routing entry to the peer of this inpcb.  If no route
1684  * is found and it cannot be allocated the return NULL.  This routine
1685  * is called by TCP routines that access the rmx structure and by tcp_mss
1686  * to get the interface MTU.
1687  */
1688 struct rtentry *
1689 tcp_rtlookup(struct in_conninfo *inc)
1690 {
1691 	struct route *ro = &inc->inc_route;
1692 
1693 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1694 		/* No route yet, so try to acquire one */
1695 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1696 			/*
1697 			 * unused portions of the structure MUST be zero'd
1698 			 * out because rtalloc() treats it as opaque data
1699 			 */
1700 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1701 			ro->ro_dst.sa_family = AF_INET;
1702 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1703 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1704 			    inc->inc_faddr;
1705 			rtalloc(ro);
1706 		}
1707 	}
1708 	return (ro->ro_rt);
1709 }
1710 
1711 #ifdef INET6
1712 struct rtentry *
1713 tcp_rtlookup6(struct in_conninfo *inc)
1714 {
1715 	struct route_in6 *ro6 = &inc->inc6_route;
1716 
1717 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1718 		/* No route yet, so try to acquire one */
1719 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1720 			/*
1721 			 * unused portions of the structure MUST be zero'd
1722 			 * out because rtalloc() treats it as opaque data
1723 			 */
1724 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1725 			ro6->ro_dst.sin6_family = AF_INET6;
1726 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1727 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1728 			rtalloc((struct route *)ro6);
1729 		}
1730 	}
1731 	return (ro6->ro_rt);
1732 }
1733 #endif
1734 
1735 #ifdef IPSEC
1736 /* compute ESP/AH header size for TCP, including outer IP header. */
1737 size_t
1738 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1739 {
1740 	struct inpcb *inp;
1741 	struct mbuf *m;
1742 	size_t hdrsiz;
1743 	struct ip *ip;
1744 	struct tcphdr *th;
1745 
1746 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1747 		return (0);
1748 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1749 	if (!m)
1750 		return (0);
1751 
1752 #ifdef INET6
1753 	if (inp->inp_vflag & INP_IPV6) {
1754 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1755 
1756 		th = (struct tcphdr *)(ip6 + 1);
1757 		m->m_pkthdr.len = m->m_len =
1758 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1759 		tcp_fillheaders(tp, ip6, th);
1760 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1761 	} else
1762 #endif
1763 	{
1764 		ip = mtod(m, struct ip *);
1765 		th = (struct tcphdr *)(ip + 1);
1766 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1767 		tcp_fillheaders(tp, ip, th);
1768 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1769 	}
1770 
1771 	m_free(m);
1772 	return (hdrsiz);
1773 }
1774 #endif
1775 
1776 /*
1777  * Return a pointer to the cached information about the remote host.
1778  * The cached information is stored in the protocol specific part of
1779  * the route metrics.
1780  */
1781 struct rmxp_tao *
1782 tcp_gettaocache(struct in_conninfo *inc)
1783 {
1784 	struct rtentry *rt;
1785 
1786 #ifdef INET6
1787 	if (inc->inc_isipv6)
1788 		rt = tcp_rtlookup6(inc);
1789 	else
1790 #endif
1791 		rt = tcp_rtlookup(inc);
1792 
1793 	/* Make sure this is a host route and is up. */
1794 	if (rt == NULL ||
1795 	    (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1796 		return (NULL);
1797 
1798 	return (rmx_taop(rt->rt_rmx));
1799 }
1800 
1801 /*
1802  * Clear all the TAO cache entries, called from tcp_init.
1803  *
1804  * XXX
1805  * This routine is just an empty one, because we assume that the routing
1806  * routing tables are initialized at the same time when TCP, so there is
1807  * nothing in the cache left over.
1808  */
1809 static void
1810 tcp_cleartaocache(void)
1811 {
1812 }
1813 
1814 /*
1815  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1816  *
1817  * This code attempts to calculate the bandwidth-delay product as a
1818  * means of determining the optimal window size to maximize bandwidth,
1819  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1820  * routers.  This code also does a fairly good job keeping RTTs in check
1821  * across slow links like modems.  We implement an algorithm which is very
1822  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1823  * transmitter side of a TCP connection and so only effects the transmit
1824  * side of the connection.
1825  *
1826  * BACKGROUND:  TCP makes no provision for the management of buffer space
1827  * at the end points or at the intermediate routers and switches.  A TCP
1828  * stream, whether using NewReno or not, will eventually buffer as
1829  * many packets as it is able and the only reason this typically works is
1830  * due to the fairly small default buffers made available for a connection
1831  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1832  * scaling it is now fairly easy for even a single TCP connection to blow-out
1833  * all available buffer space not only on the local interface, but on
1834  * intermediate routers and switches as well.  NewReno makes a misguided
1835  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1836  * then backing off, then steadily increasing the window again until another
1837  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1838  * is only made worse as network loads increase and the idea of intentionally
1839  * blowing out network buffers is, frankly, a terrible way to manage network
1840  * resources.
1841  *
1842  * It is far better to limit the transmit window prior to the failure
1843  * condition being achieved.  There are two general ways to do this:  First
1844  * you can 'scan' through different transmit window sizes and locate the
1845  * point where the RTT stops increasing, indicating that you have filled the
1846  * pipe, then scan backwards until you note that RTT stops decreasing, then
1847  * repeat ad-infinitum.  This method works in principle but has severe
1848  * implementation issues due to RTT variances, timer granularity, and
1849  * instability in the algorithm which can lead to many false positives and
1850  * create oscillations as well as interact badly with other TCP streams
1851  * implementing the same algorithm.
1852  *
1853  * The second method is to limit the window to the bandwidth delay product
1854  * of the link.  This is the method we implement.  RTT variances and our
1855  * own manipulation of the congestion window, bwnd, can potentially
1856  * destabilize the algorithm.  For this reason we have to stabilize the
1857  * elements used to calculate the window.  We do this by using the minimum
1858  * observed RTT, the long term average of the observed bandwidth, and
1859  * by adding two segments worth of slop.  It isn't perfect but it is able
1860  * to react to changing conditions and gives us a very stable basis on
1861  * which to extend the algorithm.
1862  */
1863 void
1864 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1865 {
1866 	u_long bw;
1867 	u_long bwnd;
1868 	int save_ticks;
1869 	int delta_ticks;
1870 
1871 	/*
1872 	 * If inflight_enable is disabled in the middle of a tcp connection,
1873 	 * make sure snd_bwnd is effectively disabled.
1874 	 */
1875 	if (!tcp_inflight_enable) {
1876 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1877 		tp->snd_bandwidth = 0;
1878 		return;
1879 	}
1880 
1881 	/*
1882 	 * Validate the delta time.  If a connection is new or has been idle
1883 	 * a long time we have to reset the bandwidth calculator.
1884 	 */
1885 	save_ticks = ticks;
1886 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1887 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1888 		tp->t_bw_rtttime = ticks;
1889 		tp->t_bw_rtseq = ack_seq;
1890 		if (tp->snd_bandwidth == 0)
1891 			tp->snd_bandwidth = tcp_inflight_min;
1892 		return;
1893 	}
1894 	if (delta_ticks == 0)
1895 		return;
1896 
1897 	/*
1898 	 * Sanity check, plus ignore pure window update acks.
1899 	 */
1900 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1901 		return;
1902 
1903 	/*
1904 	 * Figure out the bandwidth.  Due to the tick granularity this
1905 	 * is a very rough number and it MUST be averaged over a fairly
1906 	 * long period of time.  XXX we need to take into account a link
1907 	 * that is not using all available bandwidth, but for now our
1908 	 * slop will ramp us up if this case occurs and the bandwidth later
1909 	 * increases.
1910 	 */
1911 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1912 	tp->t_bw_rtttime = save_ticks;
1913 	tp->t_bw_rtseq = ack_seq;
1914 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1915 
1916 	tp->snd_bandwidth = bw;
1917 
1918 	/*
1919 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1920 	 * segments.  The additional slop puts us squarely in the sweet
1921 	 * spot and also handles the bandwidth run-up case.  Without the
1922 	 * slop we could be locking ourselves into a lower bandwidth.
1923 	 *
1924 	 * Situations Handled:
1925 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1926 	 *	    high speed LANs, allowing larger TCP buffers to be
1927 	 *	    specified, and also does a good job preventing
1928 	 *	    over-queueing of packets over choke points like modems
1929 	 *	    (at least for the transmit side).
1930 	 *
1931 	 *	(2) Is able to handle changing network loads (bandwidth
1932 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1933 	 *	    increases).
1934 	 *
1935 	 *	(3) Theoretically should stabilize in the face of multiple
1936 	 *	    connections implementing the same algorithm (this may need
1937 	 *	    a little work).
1938 	 *
1939 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1940 	 *	    be adjusted with a sysctl but typically only needs to be on
1941 	 *	    very slow connections.  A value no smaller then 5 should
1942 	 *	    be used, but only reduce this default if you have no other
1943 	 *	    choice.
1944 	 */
1945 
1946 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1947 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1948 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1949 #undef USERTT
1950 
1951 	if (tcp_inflight_debug > 0) {
1952 		static int ltime;
1953 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1954 			ltime = ticks;
1955 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1956 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1957 		}
1958 	}
1959 	if ((long)bwnd < tcp_inflight_min)
1960 		bwnd = tcp_inflight_min;
1961 	if (bwnd > tcp_inflight_max)
1962 		bwnd = tcp_inflight_max;
1963 	if ((long)bwnd < tp->t_maxseg * 2)
1964 		bwnd = tp->t_maxseg * 2;
1965 	tp->snd_bwnd = bwnd;
1966 }
1967