xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 6693db176654a0f25095ec64d0a74d58dcf0e47e)
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68  * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.63 2008/11/11 10:46:58 sephe Exp $
69  */
70 
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
83 #include <sys/mbuf.h>
84 #ifdef INET6
85 #include <sys/domain.h>
86 #endif
87 #include <sys/proc.h>
88 #include <sys/priv.h>
89 #include <sys/socket.h>
90 #include <sys/socketvar.h>
91 #include <sys/protosw.h>
92 #include <sys/random.h>
93 #include <sys/in_cksum.h>
94 #include <sys/ktr.h>
95 
96 #include <vm/vm_zone.h>
97 
98 #include <net/route.h>
99 #include <net/if.h>
100 #include <net/netisr.h>
101 
102 #define	_IP_VHL
103 #include <netinet/in.h>
104 #include <netinet/in_systm.h>
105 #include <netinet/ip.h>
106 #include <netinet/ip6.h>
107 #include <netinet/in_pcb.h>
108 #include <netinet6/in6_pcb.h>
109 #include <netinet/in_var.h>
110 #include <netinet/ip_var.h>
111 #include <netinet6/ip6_var.h>
112 #include <netinet/ip_icmp.h>
113 #ifdef INET6
114 #include <netinet/icmp6.h>
115 #endif
116 #include <netinet/tcp.h>
117 #include <netinet/tcp_fsm.h>
118 #include <netinet/tcp_seq.h>
119 #include <netinet/tcp_timer.h>
120 #include <netinet/tcp_timer2.h>
121 #include <netinet/tcp_var.h>
122 #include <netinet6/tcp6_var.h>
123 #include <netinet/tcpip.h>
124 #ifdef TCPDEBUG
125 #include <netinet/tcp_debug.h>
126 #endif
127 #include <netinet6/ip6protosw.h>
128 
129 #ifdef IPSEC
130 #include <netinet6/ipsec.h>
131 #ifdef INET6
132 #include <netinet6/ipsec6.h>
133 #endif
134 #endif
135 
136 #ifdef FAST_IPSEC
137 #include <netproto/ipsec/ipsec.h>
138 #ifdef INET6
139 #include <netproto/ipsec/ipsec6.h>
140 #endif
141 #define	IPSEC
142 #endif
143 
144 #include <sys/md5.h>
145 #include <machine/smp.h>
146 
147 #include <sys/msgport2.h>
148 #include <sys/mplock2.h>
149 #include <net/netmsg2.h>
150 
151 #if !defined(KTR_TCP)
152 #define KTR_TCP		KTR_ALL
153 #endif
154 KTR_INFO_MASTER(tcp);
155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
158 #define logtcp(name)	KTR_LOG(tcp_ ## name)
159 
160 struct inpcbinfo tcbinfo[MAXCPU];
161 struct tcpcbackqhead tcpcbackq[MAXCPU];
162 
163 int tcp_mpsafe_proto = 0;
164 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
165 
166 static int tcp_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE;
167 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
168 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
169 	   &tcp_mpsafe_thread, 0,
170 	   "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
171 
172 int tcp_mssdflt = TCP_MSS;
173 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
174     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
175 
176 #ifdef INET6
177 int tcp_v6mssdflt = TCP6_MSS;
178 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
179     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
180 #endif
181 
182 /*
183  * Minimum MSS we accept and use. This prevents DoS attacks where
184  * we are forced to a ridiculous low MSS like 20 and send hundreds
185  * of packets instead of one. The effect scales with the available
186  * bandwidth and quickly saturates the CPU and network interface
187  * with packet generation and sending. Set to zero to disable MINMSS
188  * checking. This setting prevents us from sending too small packets.
189  */
190 int tcp_minmss = TCP_MINMSS;
191 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
192     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
193 
194 #if 0
195 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
196 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
197     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
198 #endif
199 
200 int tcp_do_rfc1323 = 1;
201 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
202     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
203 
204 static int tcp_tcbhashsize = 0;
205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
206      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
207 
208 static int do_tcpdrain = 1;
209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
210      "Enable tcp_drain routine for extra help when low on mbufs");
211 
212 /* XXX JH */
213 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
214     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
215 
216 static int icmp_may_rst = 1;
217 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
218     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
219 
220 static int tcp_isn_reseed_interval = 0;
221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
222     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
223 
224 /*
225  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
226  * by default, but with generous values which should allow maximal
227  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
228  *
229  * The reason for doing this is that the limiter is the only mechanism we
230  * have which seems to do a really good job preventing receiver RX rings
231  * on network interfaces from getting blown out.  Even though GigE/10GigE
232  * is supposed to flow control it looks like either it doesn't actually
233  * do it or Open Source drivers do not properly enable it.
234  *
235  * People using the limiter to reduce bottlenecks on slower WAN connections
236  * should set the slop to 20 (2 packets).
237  */
238 static int tcp_inflight_enable = 1;
239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
240     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
241 
242 static int tcp_inflight_debug = 0;
243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
244     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
245 
246 static int tcp_inflight_min = 6144;
247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
248     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
249 
250 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
252     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
253 
254 static int tcp_inflight_stab = 50;
255 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
256     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
257 
258 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
259 static struct malloc_pipe tcptemp_mpipe;
260 
261 static void tcp_willblock(int);
262 static void tcp_notify (struct inpcb *, int);
263 
264 struct tcp_stats tcpstats_percpu[MAXCPU];
265 #ifdef SMP
266 static int
267 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
268 {
269 	int cpu, error = 0;
270 
271 	for (cpu = 0; cpu < ncpus; ++cpu) {
272 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
273 					sizeof(struct tcp_stats))))
274 			break;
275 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
276 				       sizeof(struct tcp_stats))))
277 			break;
278 	}
279 
280 	return (error);
281 }
282 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
283     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
284 #else
285 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
286     &tcpstat, tcp_stats, "TCP statistics");
287 #endif
288 
289 /*
290  * Target size of TCP PCB hash tables. Must be a power of two.
291  *
292  * Note that this can be overridden by the kernel environment
293  * variable net.inet.tcp.tcbhashsize
294  */
295 #ifndef TCBHASHSIZE
296 #define	TCBHASHSIZE	512
297 #endif
298 
299 /*
300  * This is the actual shape of what we allocate using the zone
301  * allocator.  Doing it this way allows us to protect both structures
302  * using the same generation count, and also eliminates the overhead
303  * of allocating tcpcbs separately.  By hiding the structure here,
304  * we avoid changing most of the rest of the code (although it needs
305  * to be changed, eventually, for greater efficiency).
306  */
307 #define	ALIGNMENT	32
308 #define	ALIGNM1		(ALIGNMENT - 1)
309 struct	inp_tp {
310 	union {
311 		struct	inpcb inp;
312 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
313 	} inp_tp_u;
314 	struct	tcpcb tcb;
315 	struct	tcp_callout inp_tp_rexmt;
316 	struct	tcp_callout inp_tp_persist;
317 	struct	tcp_callout inp_tp_keep;
318 	struct	tcp_callout inp_tp_2msl;
319 	struct	tcp_callout inp_tp_delack;
320 	struct	netmsg_tcp_timer inp_tp_timermsg;
321 };
322 #undef ALIGNMENT
323 #undef ALIGNM1
324 
325 /*
326  * Tcp initialization
327  */
328 void
329 tcp_init(void)
330 {
331 	struct inpcbporthead *porthashbase;
332 	u_long porthashmask;
333 	struct vm_zone *ipi_zone;
334 	int hashsize = TCBHASHSIZE;
335 	int cpu;
336 
337 	/*
338 	 * note: tcptemp is used for keepalives, and it is ok for an
339 	 * allocation to fail so do not specify MPF_INT.
340 	 */
341 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
342 		    25, -1, 0, NULL);
343 
344 	tcp_delacktime = TCPTV_DELACK;
345 	tcp_keepinit = TCPTV_KEEP_INIT;
346 	tcp_keepidle = TCPTV_KEEP_IDLE;
347 	tcp_keepintvl = TCPTV_KEEPINTVL;
348 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
349 	tcp_msl = TCPTV_MSL;
350 	tcp_rexmit_min = TCPTV_MIN;
351 	tcp_rexmit_slop = TCPTV_CPU_VAR;
352 
353 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
354 	if (!powerof2(hashsize)) {
355 		kprintf("WARNING: TCB hash size not a power of 2\n");
356 		hashsize = 512; /* safe default */
357 	}
358 	tcp_tcbhashsize = hashsize;
359 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
360 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
361 			 ZONE_INTERRUPT, 0);
362 
363 	for (cpu = 0; cpu < ncpus2; cpu++) {
364 		in_pcbinfo_init(&tcbinfo[cpu]);
365 		tcbinfo[cpu].cpu = cpu;
366 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
367 		    &tcbinfo[cpu].hashmask);
368 		tcbinfo[cpu].porthashbase = porthashbase;
369 		tcbinfo[cpu].porthashmask = porthashmask;
370 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
371 		    &tcbinfo[cpu].wildcardhashmask);
372 		tcbinfo[cpu].ipi_zone = ipi_zone;
373 		TAILQ_INIT(&tcpcbackq[cpu]);
374 	}
375 
376 	tcp_reass_maxseg = nmbclusters / 16;
377 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
378 
379 #ifdef INET6
380 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
381 #else
382 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
383 #endif
384 	if (max_protohdr < TCP_MINPROTOHDR)
385 		max_protohdr = TCP_MINPROTOHDR;
386 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
387 		panic("tcp_init");
388 #undef TCP_MINPROTOHDR
389 
390 	/*
391 	 * Initialize TCP statistics counters for each CPU.
392 	 */
393 #ifdef SMP
394 	for (cpu = 0; cpu < ncpus; ++cpu) {
395 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
396 	}
397 #else
398 	bzero(&tcpstat, sizeof(struct tcp_stats));
399 #endif
400 
401 	syncache_init();
402 	tcp_thread_init();
403 }
404 
405 void
406 tcpmsg_service_loop(void *dummy)
407 {
408 	struct netmsg *msg;
409 	int mplocked;
410 
411 	/*
412 	 * Thread was started with TDF_MPSAFE
413 	 */
414 	mplocked = 0;
415 
416 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
417 		do {
418 			logtcp(rxmsg);
419 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
420 						  mplocked);
421 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
422 
423 		logtcp(delayed);
424 		tcp_willblock(mplocked);
425 		logtcp(wait);
426 	}
427 }
428 
429 static void
430 tcp_willblock(int mplocked)
431 {
432 	struct tcpcb *tp;
433 	int cpu = mycpu->gd_cpuid;
434 	int unlock = 0;
435 
436 	if (!mplocked && !tcp_mpsafe_proto) {
437 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
438 			return;
439 
440 		get_mplock();
441 		mplocked = 1;
442 		unlock = 1;
443 	}
444 
445 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
446 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
447 		tp->t_flags &= ~TF_ONOUTPUTQ;
448 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
449 		tcp_output(tp);
450 	}
451 
452 	if (unlock)
453 		rel_mplock();
454 }
455 
456 
457 /*
458  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
459  * tcp_template used to store this data in mbufs, but we now recopy it out
460  * of the tcpcb each time to conserve mbufs.
461  */
462 void
463 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
464 {
465 	struct inpcb *inp = tp->t_inpcb;
466 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
467 
468 #ifdef INET6
469 	if (inp->inp_vflag & INP_IPV6) {
470 		struct ip6_hdr *ip6;
471 
472 		ip6 = (struct ip6_hdr *)ip_ptr;
473 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
474 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
475 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
476 			(IPV6_VERSION & IPV6_VERSION_MASK);
477 		ip6->ip6_nxt = IPPROTO_TCP;
478 		ip6->ip6_plen = sizeof(struct tcphdr);
479 		ip6->ip6_src = inp->in6p_laddr;
480 		ip6->ip6_dst = inp->in6p_faddr;
481 		tcp_hdr->th_sum = 0;
482 	} else
483 #endif
484 	{
485 		struct ip *ip = (struct ip *) ip_ptr;
486 
487 		ip->ip_vhl = IP_VHL_BORING;
488 		ip->ip_tos = 0;
489 		ip->ip_len = 0;
490 		ip->ip_id = 0;
491 		ip->ip_off = 0;
492 		ip->ip_ttl = 0;
493 		ip->ip_sum = 0;
494 		ip->ip_p = IPPROTO_TCP;
495 		ip->ip_src = inp->inp_laddr;
496 		ip->ip_dst = inp->inp_faddr;
497 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
498 				    ip->ip_dst.s_addr,
499 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
500 	}
501 
502 	tcp_hdr->th_sport = inp->inp_lport;
503 	tcp_hdr->th_dport = inp->inp_fport;
504 	tcp_hdr->th_seq = 0;
505 	tcp_hdr->th_ack = 0;
506 	tcp_hdr->th_x2 = 0;
507 	tcp_hdr->th_off = 5;
508 	tcp_hdr->th_flags = 0;
509 	tcp_hdr->th_win = 0;
510 	tcp_hdr->th_urp = 0;
511 }
512 
513 /*
514  * Create template to be used to send tcp packets on a connection.
515  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
516  * use for this function is in keepalives, which use tcp_respond.
517  */
518 struct tcptemp *
519 tcp_maketemplate(struct tcpcb *tp)
520 {
521 	struct tcptemp *tmp;
522 
523 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
524 		return (NULL);
525 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
526 	return (tmp);
527 }
528 
529 void
530 tcp_freetemplate(struct tcptemp *tmp)
531 {
532 	mpipe_free(&tcptemp_mpipe, tmp);
533 }
534 
535 /*
536  * Send a single message to the TCP at address specified by
537  * the given TCP/IP header.  If m == NULL, then we make a copy
538  * of the tcpiphdr at ti and send directly to the addressed host.
539  * This is used to force keep alive messages out using the TCP
540  * template for a connection.  If flags are given then we send
541  * a message back to the TCP which originated the * segment ti,
542  * and discard the mbuf containing it and any other attached mbufs.
543  *
544  * In any case the ack and sequence number of the transmitted
545  * segment are as specified by the parameters.
546  *
547  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
548  */
549 void
550 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
551 	    tcp_seq ack, tcp_seq seq, int flags)
552 {
553 	int tlen;
554 	int win = 0;
555 	struct route *ro = NULL;
556 	struct route sro;
557 	struct ip *ip = ipgen;
558 	struct tcphdr *nth;
559 	int ipflags = 0;
560 	struct route_in6 *ro6 = NULL;
561 	struct route_in6 sro6;
562 	struct ip6_hdr *ip6 = ipgen;
563 	boolean_t use_tmpro = TRUE;
564 #ifdef INET6
565 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
566 #else
567 	const boolean_t isipv6 = FALSE;
568 #endif
569 
570 	if (tp != NULL) {
571 		if (!(flags & TH_RST)) {
572 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
573 			if (win < 0)
574 				win = 0;
575 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
576 				win = (long)TCP_MAXWIN << tp->rcv_scale;
577 		}
578 		/*
579 		 * Don't use the route cache of a listen socket,
580 		 * it is not MPSAFE; use temporary route cache.
581 		 */
582 		if (tp->t_state != TCPS_LISTEN) {
583 			if (isipv6)
584 				ro6 = &tp->t_inpcb->in6p_route;
585 			else
586 				ro = &tp->t_inpcb->inp_route;
587 			use_tmpro = FALSE;
588 		}
589 	}
590 	if (use_tmpro) {
591 		if (isipv6) {
592 			ro6 = &sro6;
593 			bzero(ro6, sizeof *ro6);
594 		} else {
595 			ro = &sro;
596 			bzero(ro, sizeof *ro);
597 		}
598 	}
599 	if (m == NULL) {
600 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
601 		if (m == NULL)
602 			return;
603 		tlen = 0;
604 		m->m_data += max_linkhdr;
605 		if (isipv6) {
606 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
607 			ip6 = mtod(m, struct ip6_hdr *);
608 			nth = (struct tcphdr *)(ip6 + 1);
609 		} else {
610 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
611 			ip = mtod(m, struct ip *);
612 			nth = (struct tcphdr *)(ip + 1);
613 		}
614 		bcopy(th, nth, sizeof(struct tcphdr));
615 		flags = TH_ACK;
616 	} else {
617 		m_freem(m->m_next);
618 		m->m_next = NULL;
619 		m->m_data = (caddr_t)ipgen;
620 		/* m_len is set later */
621 		tlen = 0;
622 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
623 		if (isipv6) {
624 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
625 			nth = (struct tcphdr *)(ip6 + 1);
626 		} else {
627 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
628 			nth = (struct tcphdr *)(ip + 1);
629 		}
630 		if (th != nth) {
631 			/*
632 			 * this is usually a case when an extension header
633 			 * exists between the IPv6 header and the
634 			 * TCP header.
635 			 */
636 			nth->th_sport = th->th_sport;
637 			nth->th_dport = th->th_dport;
638 		}
639 		xchg(nth->th_dport, nth->th_sport, n_short);
640 #undef xchg
641 	}
642 	if (isipv6) {
643 		ip6->ip6_flow = 0;
644 		ip6->ip6_vfc = IPV6_VERSION;
645 		ip6->ip6_nxt = IPPROTO_TCP;
646 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
647 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
648 	} else {
649 		tlen += sizeof(struct tcpiphdr);
650 		ip->ip_len = tlen;
651 		ip->ip_ttl = ip_defttl;
652 	}
653 	m->m_len = tlen;
654 	m->m_pkthdr.len = tlen;
655 	m->m_pkthdr.rcvif = NULL;
656 	nth->th_seq = htonl(seq);
657 	nth->th_ack = htonl(ack);
658 	nth->th_x2 = 0;
659 	nth->th_off = sizeof(struct tcphdr) >> 2;
660 	nth->th_flags = flags;
661 	if (tp != NULL)
662 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
663 	else
664 		nth->th_win = htons((u_short)win);
665 	nth->th_urp = 0;
666 	if (isipv6) {
667 		nth->th_sum = 0;
668 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
669 					sizeof(struct ip6_hdr),
670 					tlen - sizeof(struct ip6_hdr));
671 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
672 					       (ro6 && ro6->ro_rt) ?
673 						ro6->ro_rt->rt_ifp : NULL);
674 	} else {
675 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
676 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
677 		m->m_pkthdr.csum_flags = CSUM_TCP;
678 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
679 	}
680 #ifdef TCPDEBUG
681 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
682 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
683 #endif
684 	if (isipv6) {
685 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
686 			   tp ? tp->t_inpcb : NULL);
687 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
688 			RTFREE(ro6->ro_rt);
689 			ro6->ro_rt = NULL;
690 		}
691 	} else {
692 		ipflags |= IP_DEBUGROUTE;
693 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
694 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
695 			RTFREE(ro->ro_rt);
696 			ro->ro_rt = NULL;
697 		}
698 	}
699 }
700 
701 /*
702  * Create a new TCP control block, making an
703  * empty reassembly queue and hooking it to the argument
704  * protocol control block.  The `inp' parameter must have
705  * come from the zone allocator set up in tcp_init().
706  */
707 struct tcpcb *
708 tcp_newtcpcb(struct inpcb *inp)
709 {
710 	struct inp_tp *it;
711 	struct tcpcb *tp;
712 #ifdef INET6
713 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
714 #else
715 	const boolean_t isipv6 = FALSE;
716 #endif
717 
718 	it = (struct inp_tp *)inp;
719 	tp = &it->tcb;
720 	bzero(tp, sizeof(struct tcpcb));
721 	LIST_INIT(&tp->t_segq);
722 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
723 
724 	/* Set up our timeouts. */
725 	tp->tt_rexmt = &it->inp_tp_rexmt;
726 	tp->tt_persist = &it->inp_tp_persist;
727 	tp->tt_keep = &it->inp_tp_keep;
728 	tp->tt_2msl = &it->inp_tp_2msl;
729 	tp->tt_delack = &it->inp_tp_delack;
730 	tcp_inittimers(tp);
731 
732 	/*
733 	 * Zero out timer message.  We don't create it here,
734 	 * since the current CPU may not be the owner of this
735 	 * inpcb.
736 	 */
737 	tp->tt_msg = &it->inp_tp_timermsg;
738 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
739 
740 	if (tcp_do_rfc1323)
741 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
742 	tp->t_inpcb = inp;	/* XXX */
743 	tp->t_state = TCPS_CLOSED;
744 	/*
745 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
746 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
747 	 * reasonable initial retransmit time.
748 	 */
749 	tp->t_srtt = TCPTV_SRTTBASE;
750 	tp->t_rttvar =
751 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
752 	tp->t_rttmin = tcp_rexmit_min;
753 	tp->t_rxtcur = TCPTV_RTOBASE;
754 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
755 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
756 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
757 	tp->t_rcvtime = ticks;
758 	/*
759 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
760 	 * because the socket may be bound to an IPv6 wildcard address,
761 	 * which may match an IPv4-mapped IPv6 address.
762 	 */
763 	inp->inp_ip_ttl = ip_defttl;
764 	inp->inp_ppcb = tp;
765 	tcp_sack_tcpcb_init(tp);
766 	return (tp);		/* XXX */
767 }
768 
769 /*
770  * Drop a TCP connection, reporting the specified error.
771  * If connection is synchronized, then send a RST to peer.
772  */
773 struct tcpcb *
774 tcp_drop(struct tcpcb *tp, int error)
775 {
776 	struct socket *so = tp->t_inpcb->inp_socket;
777 
778 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
779 		tp->t_state = TCPS_CLOSED;
780 		tcp_output(tp);
781 		tcpstat.tcps_drops++;
782 	} else
783 		tcpstat.tcps_conndrops++;
784 	if (error == ETIMEDOUT && tp->t_softerror)
785 		error = tp->t_softerror;
786 	so->so_error = error;
787 	return (tcp_close(tp));
788 }
789 
790 #ifdef SMP
791 
792 struct netmsg_remwildcard {
793 	struct netmsg		nm_netmsg;
794 	struct inpcb		*nm_inp;
795 	struct inpcbinfo	*nm_pcbinfo;
796 #if defined(INET6)
797 	int			nm_isinet6;
798 #else
799 	int			nm_unused01;
800 #endif
801 };
802 
803 /*
804  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
805  * inp can be detached.  We do this by cycling through the cpus, ending up
806  * on the cpu controlling the inp last and then doing the disconnect.
807  */
808 static void
809 in_pcbremwildcardhash_handler(struct netmsg *msg0)
810 {
811 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
812 	int cpu;
813 
814 	cpu = msg->nm_pcbinfo->cpu;
815 
816 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
817 		/* note: detach removes any wildcard hash entry */
818 #ifdef INET6
819 		if (msg->nm_isinet6)
820 			in6_pcbdetach(msg->nm_inp);
821 		else
822 #endif
823 			in_pcbdetach(msg->nm_inp);
824 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
825 	} else {
826 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
827 		cpu = (cpu + 1) % ncpus2;
828 		msg->nm_pcbinfo = &tcbinfo[cpu];
829 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
830 	}
831 }
832 
833 #endif
834 
835 /*
836  * Close a TCP control block:
837  *	discard all space held by the tcp
838  *	discard internet protocol block
839  *	wake up any sleepers
840  */
841 struct tcpcb *
842 tcp_close(struct tcpcb *tp)
843 {
844 	struct tseg_qent *q;
845 	struct inpcb *inp = tp->t_inpcb;
846 	struct socket *so = inp->inp_socket;
847 	struct rtentry *rt;
848 	boolean_t dosavessthresh;
849 #ifdef SMP
850 	int cpu;
851 #endif
852 #ifdef INET6
853 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
854 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
855 #else
856 	const boolean_t isipv6 = FALSE;
857 #endif
858 
859 	/*
860 	 * The tp is not instantly destroyed in the wildcard case.  Setting
861 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
862 	 * messing with it, though it should be noted that this change may
863 	 * not take effect on other cpus until we have chained the wildcard
864 	 * hash removal.
865 	 *
866 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
867 	 * update and prevent other tcp protocol threads from accepting new
868 	 * connections on the listen socket we might be trying to close down.
869 	 */
870 	KKASSERT(tp->t_state != TCPS_TERMINATING);
871 	tp->t_state = TCPS_TERMINATING;
872 
873 	/*
874 	 * Make sure that all of our timers are stopped before we
875 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
876 	 * timers are never used.  If timer message is never created
877 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
878 	 */
879 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
880 		tcp_callout_stop(tp, tp->tt_rexmt);
881 		tcp_callout_stop(tp, tp->tt_persist);
882 		tcp_callout_stop(tp, tp->tt_keep);
883 		tcp_callout_stop(tp, tp->tt_2msl);
884 		tcp_callout_stop(tp, tp->tt_delack);
885 	}
886 
887 	if (tp->t_flags & TF_ONOUTPUTQ) {
888 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
889 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
890 		tp->t_flags &= ~TF_ONOUTPUTQ;
891 	}
892 
893 	/*
894 	 * If we got enough samples through the srtt filter,
895 	 * save the rtt and rttvar in the routing entry.
896 	 * 'Enough' is arbitrarily defined as the 16 samples.
897 	 * 16 samples is enough for the srtt filter to converge
898 	 * to within 5% of the correct value; fewer samples and
899 	 * we could save a very bogus rtt.
900 	 *
901 	 * Don't update the default route's characteristics and don't
902 	 * update anything that the user "locked".
903 	 */
904 	if (tp->t_rttupdated >= 16) {
905 		u_long i = 0;
906 
907 		if (isipv6) {
908 			struct sockaddr_in6 *sin6;
909 
910 			if ((rt = inp->in6p_route.ro_rt) == NULL)
911 				goto no_valid_rt;
912 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
913 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
914 				goto no_valid_rt;
915 		} else
916 			if ((rt = inp->inp_route.ro_rt) == NULL ||
917 			    ((struct sockaddr_in *)rt_key(rt))->
918 			     sin_addr.s_addr == INADDR_ANY)
919 				goto no_valid_rt;
920 
921 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
922 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
923 			if (rt->rt_rmx.rmx_rtt && i)
924 				/*
925 				 * filter this update to half the old & half
926 				 * the new values, converting scale.
927 				 * See route.h and tcp_var.h for a
928 				 * description of the scaling constants.
929 				 */
930 				rt->rt_rmx.rmx_rtt =
931 				    (rt->rt_rmx.rmx_rtt + i) / 2;
932 			else
933 				rt->rt_rmx.rmx_rtt = i;
934 			tcpstat.tcps_cachedrtt++;
935 		}
936 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
937 			i = tp->t_rttvar *
938 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
939 			if (rt->rt_rmx.rmx_rttvar && i)
940 				rt->rt_rmx.rmx_rttvar =
941 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
942 			else
943 				rt->rt_rmx.rmx_rttvar = i;
944 			tcpstat.tcps_cachedrttvar++;
945 		}
946 		/*
947 		 * The old comment here said:
948 		 * update the pipelimit (ssthresh) if it has been updated
949 		 * already or if a pipesize was specified & the threshhold
950 		 * got below half the pipesize.  I.e., wait for bad news
951 		 * before we start updating, then update on both good
952 		 * and bad news.
953 		 *
954 		 * But we want to save the ssthresh even if no pipesize is
955 		 * specified explicitly in the route, because such
956 		 * connections still have an implicit pipesize specified
957 		 * by the global tcp_sendspace.  In the absence of a reliable
958 		 * way to calculate the pipesize, it will have to do.
959 		 */
960 		i = tp->snd_ssthresh;
961 		if (rt->rt_rmx.rmx_sendpipe != 0)
962 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
963 		else
964 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
965 		if (dosavessthresh ||
966 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
967 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
968 			/*
969 			 * convert the limit from user data bytes to
970 			 * packets then to packet data bytes.
971 			 */
972 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
973 			if (i < 2)
974 				i = 2;
975 			i *= tp->t_maxseg +
976 			     (isipv6 ?
977 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
978 			      sizeof(struct tcpiphdr));
979 			if (rt->rt_rmx.rmx_ssthresh)
980 				rt->rt_rmx.rmx_ssthresh =
981 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
982 			else
983 				rt->rt_rmx.rmx_ssthresh = i;
984 			tcpstat.tcps_cachedssthresh++;
985 		}
986 	}
987 
988 no_valid_rt:
989 	/* free the reassembly queue, if any */
990 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
991 		LIST_REMOVE(q, tqe_q);
992 		m_freem(q->tqe_m);
993 		FREE(q, M_TSEGQ);
994 		tcp_reass_qsize--;
995 	}
996 	/* throw away SACK blocks in scoreboard*/
997 	if (TCP_DO_SACK(tp))
998 		tcp_sack_cleanup(&tp->scb);
999 
1000 	inp->inp_ppcb = NULL;
1001 	soisdisconnected(so);
1002 
1003 	tcp_destroy_timermsg(tp);
1004 
1005 	/*
1006 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
1007 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
1008 	 * for each protocol thread and must be removed in the context of
1009 	 * that thread.  This is accomplished by chaining the message
1010 	 * through the cpus.
1011 	 *
1012 	 * If the inp is not wildcarded we simply detach, which will remove
1013 	 * the any hashes still present for this inp.
1014 	 */
1015 #ifdef SMP
1016 	if (inp->inp_flags & INP_WILDCARD_MP) {
1017 		struct netmsg_remwildcard *msg;
1018 
1019 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1020 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
1021 			      M_LWKTMSG, M_INTWAIT);
1022 		netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1023 			    0, in_pcbremwildcardhash_handler);
1024 #ifdef INET6
1025 		msg->nm_isinet6 = isafinet6;
1026 #endif
1027 		msg->nm_inp = inp;
1028 		msg->nm_pcbinfo = &tcbinfo[cpu];
1029 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1030 	} else
1031 #endif
1032 	{
1033 		/* note: detach removes any wildcard hash entry */
1034 #ifdef INET6
1035 		if (isafinet6)
1036 			in6_pcbdetach(inp);
1037 		else
1038 #endif
1039 			in_pcbdetach(inp);
1040 	}
1041 	tcpstat.tcps_closed++;
1042 	return (NULL);
1043 }
1044 
1045 static __inline void
1046 tcp_drain_oncpu(struct inpcbhead *head)
1047 {
1048 	struct inpcb *inpb;
1049 	struct tcpcb *tcpb;
1050 	struct tseg_qent *te;
1051 
1052 	LIST_FOREACH(inpb, head, inp_list) {
1053 		if (inpb->inp_flags & INP_PLACEMARKER)
1054 			continue;
1055 		if ((tcpb = intotcpcb(inpb))) {
1056 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1057 				LIST_REMOVE(te, tqe_q);
1058 				m_freem(te->tqe_m);
1059 				FREE(te, M_TSEGQ);
1060 				tcp_reass_qsize--;
1061 			}
1062 		}
1063 	}
1064 }
1065 
1066 #ifdef SMP
1067 struct netmsg_tcp_drain {
1068 	struct netmsg		nm_netmsg;
1069 	struct inpcbhead	*nm_head;
1070 };
1071 
1072 static void
1073 tcp_drain_handler(netmsg_t netmsg)
1074 {
1075 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1076 
1077 	tcp_drain_oncpu(nm->nm_head);
1078 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1079 }
1080 #endif
1081 
1082 void
1083 tcp_drain(void)
1084 {
1085 #ifdef SMP
1086 	int cpu;
1087 #endif
1088 
1089 	if (!do_tcpdrain)
1090 		return;
1091 
1092 	/*
1093 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1094 	 * if there is one...
1095 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1096 	 *	reassembly queue should be flushed, but in a situation
1097 	 *	where we're really low on mbufs, this is potentially
1098 	 *	useful.
1099 	 */
1100 #ifdef SMP
1101 	for (cpu = 0; cpu < ncpus2; cpu++) {
1102 		struct netmsg_tcp_drain *msg;
1103 
1104 		if (cpu == mycpu->gd_cpuid) {
1105 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1106 		} else {
1107 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1108 				    M_LWKTMSG, M_NOWAIT);
1109 			if (msg == NULL)
1110 				continue;
1111 			netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1112 				    0, tcp_drain_handler);
1113 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1114 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1115 		}
1116 	}
1117 #else
1118 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1119 #endif
1120 }
1121 
1122 /*
1123  * Notify a tcp user of an asynchronous error;
1124  * store error as soft error, but wake up user
1125  * (for now, won't do anything until can select for soft error).
1126  *
1127  * Do not wake up user since there currently is no mechanism for
1128  * reporting soft errors (yet - a kqueue filter may be added).
1129  */
1130 static void
1131 tcp_notify(struct inpcb *inp, int error)
1132 {
1133 	struct tcpcb *tp = intotcpcb(inp);
1134 
1135 	/*
1136 	 * Ignore some errors if we are hooked up.
1137 	 * If connection hasn't completed, has retransmitted several times,
1138 	 * and receives a second error, give up now.  This is better
1139 	 * than waiting a long time to establish a connection that
1140 	 * can never complete.
1141 	 */
1142 	if (tp->t_state == TCPS_ESTABLISHED &&
1143 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1144 	      error == EHOSTDOWN)) {
1145 		return;
1146 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1147 	    tp->t_softerror)
1148 		tcp_drop(tp, error);
1149 	else
1150 		tp->t_softerror = error;
1151 #if 0
1152 	wakeup(&so->so_timeo);
1153 	sorwakeup(so);
1154 	sowwakeup(so);
1155 #endif
1156 }
1157 
1158 static int
1159 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1160 {
1161 	int error, i, n;
1162 	struct inpcb *marker;
1163 	struct inpcb *inp;
1164 	inp_gen_t gencnt;
1165 	globaldata_t gd;
1166 	int origcpu, ccpu;
1167 
1168 	error = 0;
1169 	n = 0;
1170 
1171 	/*
1172 	 * The process of preparing the TCB list is too time-consuming and
1173 	 * resource-intensive to repeat twice on every request.
1174 	 */
1175 	if (req->oldptr == NULL) {
1176 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1177 			gd = globaldata_find(ccpu);
1178 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1179 		}
1180 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1181 		return (0);
1182 	}
1183 
1184 	if (req->newptr != NULL)
1185 		return (EPERM);
1186 
1187 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1188 	marker->inp_flags |= INP_PLACEMARKER;
1189 
1190 	/*
1191 	 * OK, now we're committed to doing something.  Run the inpcb list
1192 	 * for each cpu in the system and construct the output.  Use a
1193 	 * list placemarker to deal with list changes occuring during
1194 	 * copyout blockages (but otherwise depend on being on the correct
1195 	 * cpu to avoid races).
1196 	 */
1197 	origcpu = mycpu->gd_cpuid;
1198 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1199 		globaldata_t rgd;
1200 		caddr_t inp_ppcb;
1201 		struct xtcpcb xt;
1202 		int cpu_id;
1203 
1204 		cpu_id = (origcpu + ccpu) % ncpus;
1205 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1206 			continue;
1207 		rgd = globaldata_find(cpu_id);
1208 		lwkt_setcpu_self(rgd);
1209 
1210 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1211 		n = tcbinfo[cpu_id].ipi_count;
1212 
1213 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1214 		i = 0;
1215 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1216 			/*
1217 			 * process a snapshot of pcbs, ignoring placemarkers
1218 			 * and using our own to allow SYSCTL_OUT to block.
1219 			 */
1220 			LIST_REMOVE(marker, inp_list);
1221 			LIST_INSERT_AFTER(inp, marker, inp_list);
1222 
1223 			if (inp->inp_flags & INP_PLACEMARKER)
1224 				continue;
1225 			if (inp->inp_gencnt > gencnt)
1226 				continue;
1227 			if (prison_xinpcb(req->td, inp))
1228 				continue;
1229 
1230 			xt.xt_len = sizeof xt;
1231 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1232 			inp_ppcb = inp->inp_ppcb;
1233 			if (inp_ppcb != NULL)
1234 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1235 			else
1236 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1237 			if (inp->inp_socket)
1238 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1239 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1240 				break;
1241 			++i;
1242 		}
1243 		LIST_REMOVE(marker, inp_list);
1244 		if (error == 0 && i < n) {
1245 			bzero(&xt, sizeof xt);
1246 			xt.xt_len = sizeof xt;
1247 			while (i < n) {
1248 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1249 				if (error)
1250 					break;
1251 				++i;
1252 			}
1253 		}
1254 	}
1255 
1256 	/*
1257 	 * Make sure we are on the same cpu we were on originally, since
1258 	 * higher level callers expect this.  Also don't pollute caches with
1259 	 * migrated userland data by (eventually) returning to userland
1260 	 * on a different cpu.
1261 	 */
1262 	lwkt_setcpu_self(globaldata_find(origcpu));
1263 	kfree(marker, M_TEMP);
1264 	return (error);
1265 }
1266 
1267 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1268 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1269 
1270 static int
1271 tcp_getcred(SYSCTL_HANDLER_ARGS)
1272 {
1273 	struct sockaddr_in addrs[2];
1274 	struct inpcb *inp;
1275 	int cpu;
1276 	int error;
1277 
1278 	error = priv_check(req->td, PRIV_ROOT);
1279 	if (error != 0)
1280 		return (error);
1281 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1282 	if (error != 0)
1283 		return (error);
1284 	crit_enter();
1285 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1286 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1287 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1288 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1289 	if (inp == NULL || inp->inp_socket == NULL) {
1290 		error = ENOENT;
1291 		goto out;
1292 	}
1293 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1294 out:
1295 	crit_exit();
1296 	return (error);
1297 }
1298 
1299 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1300     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1301 
1302 #ifdef INET6
1303 static int
1304 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1305 {
1306 	struct sockaddr_in6 addrs[2];
1307 	struct inpcb *inp;
1308 	int error;
1309 	boolean_t mapped = FALSE;
1310 
1311 	error = priv_check(req->td, PRIV_ROOT);
1312 	if (error != 0)
1313 		return (error);
1314 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1315 	if (error != 0)
1316 		return (error);
1317 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1318 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1319 			mapped = TRUE;
1320 		else
1321 			return (EINVAL);
1322 	}
1323 	crit_enter();
1324 	if (mapped) {
1325 		inp = in_pcblookup_hash(&tcbinfo[0],
1326 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1327 		    addrs[1].sin6_port,
1328 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1329 		    addrs[0].sin6_port,
1330 		    0, NULL);
1331 	} else {
1332 		inp = in6_pcblookup_hash(&tcbinfo[0],
1333 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1334 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1335 		    0, NULL);
1336 	}
1337 	if (inp == NULL || inp->inp_socket == NULL) {
1338 		error = ENOENT;
1339 		goto out;
1340 	}
1341 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1342 out:
1343 	crit_exit();
1344 	return (error);
1345 }
1346 
1347 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1348 	    0, 0,
1349 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1350 #endif
1351 
1352 struct netmsg_tcp_notify {
1353 	struct netmsg	nm_nmsg;
1354 	void		(*nm_notify)(struct inpcb *, int);
1355 	struct in_addr	nm_faddr;
1356 	int		nm_arg;
1357 };
1358 
1359 static void
1360 tcp_notifyall_oncpu(struct netmsg *netmsg)
1361 {
1362 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1363 	int nextcpu;
1364 
1365 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1366 			nmsg->nm_arg, nmsg->nm_notify);
1367 
1368 	nextcpu = mycpuid + 1;
1369 	if (nextcpu < ncpus2)
1370 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1371 	else
1372 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1373 }
1374 
1375 void
1376 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1377 {
1378 	struct ip *ip = vip;
1379 	struct tcphdr *th;
1380 	struct in_addr faddr;
1381 	struct inpcb *inp;
1382 	struct tcpcb *tp;
1383 	void (*notify)(struct inpcb *, int) = tcp_notify;
1384 	tcp_seq icmpseq;
1385 	int arg, cpu;
1386 
1387 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1388 		return;
1389 	}
1390 
1391 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1392 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1393 		return;
1394 
1395 	arg = inetctlerrmap[cmd];
1396 	if (cmd == PRC_QUENCH) {
1397 		notify = tcp_quench;
1398 	} else if (icmp_may_rst &&
1399 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1400 		    cmd == PRC_UNREACH_PORT ||
1401 		    cmd == PRC_TIMXCEED_INTRANS) &&
1402 		   ip != NULL) {
1403 		notify = tcp_drop_syn_sent;
1404 	} else if (cmd == PRC_MSGSIZE) {
1405 		struct icmp *icmp = (struct icmp *)
1406 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1407 
1408 		arg = ntohs(icmp->icmp_nextmtu);
1409 		notify = tcp_mtudisc;
1410 	} else if (PRC_IS_REDIRECT(cmd)) {
1411 		ip = NULL;
1412 		notify = in_rtchange;
1413 	} else if (cmd == PRC_HOSTDEAD) {
1414 		ip = NULL;
1415 	}
1416 
1417 	if (ip != NULL) {
1418 		crit_enter();
1419 		th = (struct tcphdr *)((caddr_t)ip +
1420 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1421 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1422 				  ip->ip_src.s_addr, th->th_sport);
1423 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1424 					ip->ip_src, th->th_sport, 0, NULL);
1425 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1426 			icmpseq = htonl(th->th_seq);
1427 			tp = intotcpcb(inp);
1428 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1429 			    SEQ_LT(icmpseq, tp->snd_max))
1430 				(*notify)(inp, arg);
1431 		} else {
1432 			struct in_conninfo inc;
1433 
1434 			inc.inc_fport = th->th_dport;
1435 			inc.inc_lport = th->th_sport;
1436 			inc.inc_faddr = faddr;
1437 			inc.inc_laddr = ip->ip_src;
1438 #ifdef INET6
1439 			inc.inc_isipv6 = 0;
1440 #endif
1441 			syncache_unreach(&inc, th);
1442 		}
1443 		crit_exit();
1444 	} else {
1445 		struct netmsg_tcp_notify nmsg;
1446 
1447 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1448 		netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1449 			    0, tcp_notifyall_oncpu);
1450 		nmsg.nm_faddr = faddr;
1451 		nmsg.nm_arg = arg;
1452 		nmsg.nm_notify = notify;
1453 
1454 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1455 	}
1456 }
1457 
1458 #ifdef INET6
1459 void
1460 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1461 {
1462 	struct tcphdr th;
1463 	void (*notify) (struct inpcb *, int) = tcp_notify;
1464 	struct ip6_hdr *ip6;
1465 	struct mbuf *m;
1466 	struct ip6ctlparam *ip6cp = NULL;
1467 	const struct sockaddr_in6 *sa6_src = NULL;
1468 	int off;
1469 	struct tcp_portonly {
1470 		u_int16_t th_sport;
1471 		u_int16_t th_dport;
1472 	} *thp;
1473 	int arg;
1474 
1475 	if (sa->sa_family != AF_INET6 ||
1476 	    sa->sa_len != sizeof(struct sockaddr_in6))
1477 		return;
1478 
1479 	arg = 0;
1480 	if (cmd == PRC_QUENCH)
1481 		notify = tcp_quench;
1482 	else if (cmd == PRC_MSGSIZE) {
1483 		struct ip6ctlparam *ip6cp = d;
1484 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1485 
1486 		arg = ntohl(icmp6->icmp6_mtu);
1487 		notify = tcp_mtudisc;
1488 	} else if (!PRC_IS_REDIRECT(cmd) &&
1489 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1490 		return;
1491 	}
1492 
1493 	/* if the parameter is from icmp6, decode it. */
1494 	if (d != NULL) {
1495 		ip6cp = (struct ip6ctlparam *)d;
1496 		m = ip6cp->ip6c_m;
1497 		ip6 = ip6cp->ip6c_ip6;
1498 		off = ip6cp->ip6c_off;
1499 		sa6_src = ip6cp->ip6c_src;
1500 	} else {
1501 		m = NULL;
1502 		ip6 = NULL;
1503 		off = 0;	/* fool gcc */
1504 		sa6_src = &sa6_any;
1505 	}
1506 
1507 	if (ip6 != NULL) {
1508 		struct in_conninfo inc;
1509 		/*
1510 		 * XXX: We assume that when IPV6 is non NULL,
1511 		 * M and OFF are valid.
1512 		 */
1513 
1514 		/* check if we can safely examine src and dst ports */
1515 		if (m->m_pkthdr.len < off + sizeof *thp)
1516 			return;
1517 
1518 		bzero(&th, sizeof th);
1519 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1520 
1521 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1522 		    (struct sockaddr *)ip6cp->ip6c_src,
1523 		    th.th_sport, cmd, arg, notify);
1524 
1525 		inc.inc_fport = th.th_dport;
1526 		inc.inc_lport = th.th_sport;
1527 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1528 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1529 		inc.inc_isipv6 = 1;
1530 		syncache_unreach(&inc, &th);
1531 	} else
1532 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1533 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1534 }
1535 #endif
1536 
1537 /*
1538  * Following is where TCP initial sequence number generation occurs.
1539  *
1540  * There are two places where we must use initial sequence numbers:
1541  * 1.  In SYN-ACK packets.
1542  * 2.  In SYN packets.
1543  *
1544  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1545  * tcp_syncache.c for details.
1546  *
1547  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1548  * depends on this property.  In addition, these ISNs should be
1549  * unguessable so as to prevent connection hijacking.  To satisfy
1550  * the requirements of this situation, the algorithm outlined in
1551  * RFC 1948 is used to generate sequence numbers.
1552  *
1553  * Implementation details:
1554  *
1555  * Time is based off the system timer, and is corrected so that it
1556  * increases by one megabyte per second.  This allows for proper
1557  * recycling on high speed LANs while still leaving over an hour
1558  * before rollover.
1559  *
1560  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1561  * between seeding of isn_secret.  This is normally set to zero,
1562  * as reseeding should not be necessary.
1563  *
1564  */
1565 
1566 #define	ISN_BYTES_PER_SECOND 1048576
1567 
1568 u_char isn_secret[32];
1569 int isn_last_reseed;
1570 MD5_CTX isn_ctx;
1571 
1572 tcp_seq
1573 tcp_new_isn(struct tcpcb *tp)
1574 {
1575 	u_int32_t md5_buffer[4];
1576 	tcp_seq new_isn;
1577 
1578 	/* Seed if this is the first use, reseed if requested. */
1579 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1580 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1581 		< (u_int)ticks))) {
1582 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1583 		isn_last_reseed = ticks;
1584 	}
1585 
1586 	/* Compute the md5 hash and return the ISN. */
1587 	MD5Init(&isn_ctx);
1588 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1589 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1590 #ifdef INET6
1591 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1592 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1593 			  sizeof(struct in6_addr));
1594 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1595 			  sizeof(struct in6_addr));
1596 	} else
1597 #endif
1598 	{
1599 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1600 			  sizeof(struct in_addr));
1601 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1602 			  sizeof(struct in_addr));
1603 	}
1604 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1605 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1606 	new_isn = (tcp_seq) md5_buffer[0];
1607 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1608 	return (new_isn);
1609 }
1610 
1611 /*
1612  * When a source quench is received, close congestion window
1613  * to one segment.  We will gradually open it again as we proceed.
1614  */
1615 void
1616 tcp_quench(struct inpcb *inp, int error)
1617 {
1618 	struct tcpcb *tp = intotcpcb(inp);
1619 
1620 	if (tp != NULL) {
1621 		tp->snd_cwnd = tp->t_maxseg;
1622 		tp->snd_wacked = 0;
1623 	}
1624 }
1625 
1626 /*
1627  * When a specific ICMP unreachable message is received and the
1628  * connection state is SYN-SENT, drop the connection.  This behavior
1629  * is controlled by the icmp_may_rst sysctl.
1630  */
1631 void
1632 tcp_drop_syn_sent(struct inpcb *inp, int error)
1633 {
1634 	struct tcpcb *tp = intotcpcb(inp);
1635 
1636 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1637 		tcp_drop(tp, error);
1638 }
1639 
1640 /*
1641  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1642  * based on the new value in the route.  Also nudge TCP to send something,
1643  * since we know the packet we just sent was dropped.
1644  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1645  */
1646 void
1647 tcp_mtudisc(struct inpcb *inp, int mtu)
1648 {
1649 	struct tcpcb *tp = intotcpcb(inp);
1650 	struct rtentry *rt;
1651 	struct socket *so = inp->inp_socket;
1652 	int maxopd, mss;
1653 #ifdef INET6
1654 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1655 #else
1656 	const boolean_t isipv6 = FALSE;
1657 #endif
1658 
1659 	if (tp == NULL)
1660 		return;
1661 
1662 	/*
1663 	 * If no MTU is provided in the ICMP message, use the
1664 	 * next lower likely value, as specified in RFC 1191.
1665 	 */
1666 	if (mtu == 0) {
1667 		int oldmtu;
1668 
1669 		oldmtu = tp->t_maxopd +
1670 		    (isipv6 ?
1671 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1672 		     sizeof(struct tcpiphdr));
1673 		mtu = ip_next_mtu(oldmtu, 0);
1674 	}
1675 
1676 	if (isipv6)
1677 		rt = tcp_rtlookup6(&inp->inp_inc);
1678 	else
1679 		rt = tcp_rtlookup(&inp->inp_inc);
1680 	if (rt != NULL) {
1681 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1682 			mtu = rt->rt_rmx.rmx_mtu;
1683 
1684 		maxopd = mtu -
1685 		    (isipv6 ?
1686 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1687 		     sizeof(struct tcpiphdr));
1688 
1689 		/*
1690 		 * XXX - The following conditional probably violates the TCP
1691 		 * spec.  The problem is that, since we don't know the
1692 		 * other end's MSS, we are supposed to use a conservative
1693 		 * default.  But, if we do that, then MTU discovery will
1694 		 * never actually take place, because the conservative
1695 		 * default is much less than the MTUs typically seen
1696 		 * on the Internet today.  For the moment, we'll sweep
1697 		 * this under the carpet.
1698 		 *
1699 		 * The conservative default might not actually be a problem
1700 		 * if the only case this occurs is when sending an initial
1701 		 * SYN with options and data to a host we've never talked
1702 		 * to before.  Then, they will reply with an MSS value which
1703 		 * will get recorded and the new parameters should get
1704 		 * recomputed.  For Further Study.
1705 		 */
1706 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1707 			maxopd = rt->rt_rmx.rmx_mssopt;
1708 	} else
1709 		maxopd = mtu -
1710 		    (isipv6 ?
1711 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1712 		     sizeof(struct tcpiphdr));
1713 
1714 	if (tp->t_maxopd <= maxopd)
1715 		return;
1716 	tp->t_maxopd = maxopd;
1717 
1718 	mss = maxopd;
1719 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1720 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1721 		mss -= TCPOLEN_TSTAMP_APPA;
1722 
1723 	/* round down to multiple of MCLBYTES */
1724 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1725 	if (mss > MCLBYTES)
1726 		mss &= ~(MCLBYTES - 1);
1727 #else
1728 	if (mss > MCLBYTES)
1729 		mss = (mss / MCLBYTES) * MCLBYTES;
1730 #endif
1731 
1732 	if (so->so_snd.ssb_hiwat < mss)
1733 		mss = so->so_snd.ssb_hiwat;
1734 
1735 	tp->t_maxseg = mss;
1736 	tp->t_rtttime = 0;
1737 	tp->snd_nxt = tp->snd_una;
1738 	tcp_output(tp);
1739 	tcpstat.tcps_mturesent++;
1740 }
1741 
1742 /*
1743  * Look-up the routing entry to the peer of this inpcb.  If no route
1744  * is found and it cannot be allocated the return NULL.  This routine
1745  * is called by TCP routines that access the rmx structure and by tcp_mss
1746  * to get the interface MTU.
1747  */
1748 struct rtentry *
1749 tcp_rtlookup(struct in_conninfo *inc)
1750 {
1751 	struct route *ro = &inc->inc_route;
1752 
1753 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1754 		/* No route yet, so try to acquire one */
1755 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1756 			/*
1757 			 * unused portions of the structure MUST be zero'd
1758 			 * out because rtalloc() treats it as opaque data
1759 			 */
1760 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1761 			ro->ro_dst.sa_family = AF_INET;
1762 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1763 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1764 			    inc->inc_faddr;
1765 			rtalloc(ro);
1766 		}
1767 	}
1768 	return (ro->ro_rt);
1769 }
1770 
1771 #ifdef INET6
1772 struct rtentry *
1773 tcp_rtlookup6(struct in_conninfo *inc)
1774 {
1775 	struct route_in6 *ro6 = &inc->inc6_route;
1776 
1777 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1778 		/* No route yet, so try to acquire one */
1779 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1780 			/*
1781 			 * unused portions of the structure MUST be zero'd
1782 			 * out because rtalloc() treats it as opaque data
1783 			 */
1784 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1785 			ro6->ro_dst.sin6_family = AF_INET6;
1786 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1787 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1788 			rtalloc((struct route *)ro6);
1789 		}
1790 	}
1791 	return (ro6->ro_rt);
1792 }
1793 #endif
1794 
1795 #ifdef IPSEC
1796 /* compute ESP/AH header size for TCP, including outer IP header. */
1797 size_t
1798 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1799 {
1800 	struct inpcb *inp;
1801 	struct mbuf *m;
1802 	size_t hdrsiz;
1803 	struct ip *ip;
1804 	struct tcphdr *th;
1805 
1806 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1807 		return (0);
1808 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1809 	if (!m)
1810 		return (0);
1811 
1812 #ifdef INET6
1813 	if (inp->inp_vflag & INP_IPV6) {
1814 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1815 
1816 		th = (struct tcphdr *)(ip6 + 1);
1817 		m->m_pkthdr.len = m->m_len =
1818 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1819 		tcp_fillheaders(tp, ip6, th);
1820 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1821 	} else
1822 #endif
1823 	{
1824 		ip = mtod(m, struct ip *);
1825 		th = (struct tcphdr *)(ip + 1);
1826 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1827 		tcp_fillheaders(tp, ip, th);
1828 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1829 	}
1830 
1831 	m_free(m);
1832 	return (hdrsiz);
1833 }
1834 #endif
1835 
1836 /*
1837  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1838  *
1839  * This code attempts to calculate the bandwidth-delay product as a
1840  * means of determining the optimal window size to maximize bandwidth,
1841  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1842  * routers.  This code also does a fairly good job keeping RTTs in check
1843  * across slow links like modems.  We implement an algorithm which is very
1844  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1845  * transmitter side of a TCP connection and so only effects the transmit
1846  * side of the connection.
1847  *
1848  * BACKGROUND:  TCP makes no provision for the management of buffer space
1849  * at the end points or at the intermediate routers and switches.  A TCP
1850  * stream, whether using NewReno or not, will eventually buffer as
1851  * many packets as it is able and the only reason this typically works is
1852  * due to the fairly small default buffers made available for a connection
1853  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1854  * scaling it is now fairly easy for even a single TCP connection to blow-out
1855  * all available buffer space not only on the local interface, but on
1856  * intermediate routers and switches as well.  NewReno makes a misguided
1857  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1858  * then backing off, then steadily increasing the window again until another
1859  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1860  * is only made worse as network loads increase and the idea of intentionally
1861  * blowing out network buffers is, frankly, a terrible way to manage network
1862  * resources.
1863  *
1864  * It is far better to limit the transmit window prior to the failure
1865  * condition being achieved.  There are two general ways to do this:  First
1866  * you can 'scan' through different transmit window sizes and locate the
1867  * point where the RTT stops increasing, indicating that you have filled the
1868  * pipe, then scan backwards until you note that RTT stops decreasing, then
1869  * repeat ad-infinitum.  This method works in principle but has severe
1870  * implementation issues due to RTT variances, timer granularity, and
1871  * instability in the algorithm which can lead to many false positives and
1872  * create oscillations as well as interact badly with other TCP streams
1873  * implementing the same algorithm.
1874  *
1875  * The second method is to limit the window to the bandwidth delay product
1876  * of the link.  This is the method we implement.  RTT variances and our
1877  * own manipulation of the congestion window, bwnd, can potentially
1878  * destabilize the algorithm.  For this reason we have to stabilize the
1879  * elements used to calculate the window.  We do this by using the minimum
1880  * observed RTT, the long term average of the observed bandwidth, and
1881  * by adding two segments worth of slop.  It isn't perfect but it is able
1882  * to react to changing conditions and gives us a very stable basis on
1883  * which to extend the algorithm.
1884  */
1885 void
1886 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1887 {
1888 	u_long bw;
1889 	u_long bwnd;
1890 	int save_ticks;
1891 	int delta_ticks;
1892 
1893 	/*
1894 	 * If inflight_enable is disabled in the middle of a tcp connection,
1895 	 * make sure snd_bwnd is effectively disabled.
1896 	 */
1897 	if (!tcp_inflight_enable) {
1898 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1899 		tp->snd_bandwidth = 0;
1900 		return;
1901 	}
1902 
1903 	/*
1904 	 * Validate the delta time.  If a connection is new or has been idle
1905 	 * a long time we have to reset the bandwidth calculator.
1906 	 */
1907 	save_ticks = ticks;
1908 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1909 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1910 		tp->t_bw_rtttime = ticks;
1911 		tp->t_bw_rtseq = ack_seq;
1912 		if (tp->snd_bandwidth == 0)
1913 			tp->snd_bandwidth = tcp_inflight_min;
1914 		return;
1915 	}
1916 	if (delta_ticks == 0)
1917 		return;
1918 
1919 	/*
1920 	 * Sanity check, plus ignore pure window update acks.
1921 	 */
1922 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1923 		return;
1924 
1925 	/*
1926 	 * Figure out the bandwidth.  Due to the tick granularity this
1927 	 * is a very rough number and it MUST be averaged over a fairly
1928 	 * long period of time.  XXX we need to take into account a link
1929 	 * that is not using all available bandwidth, but for now our
1930 	 * slop will ramp us up if this case occurs and the bandwidth later
1931 	 * increases.
1932 	 */
1933 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1934 	tp->t_bw_rtttime = save_ticks;
1935 	tp->t_bw_rtseq = ack_seq;
1936 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1937 
1938 	tp->snd_bandwidth = bw;
1939 
1940 	/*
1941 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1942 	 * segments.  The additional slop puts us squarely in the sweet
1943 	 * spot and also handles the bandwidth run-up case.  Without the
1944 	 * slop we could be locking ourselves into a lower bandwidth.
1945 	 *
1946 	 * Situations Handled:
1947 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1948 	 *	    high speed LANs, allowing larger TCP buffers to be
1949 	 *	    specified, and also does a good job preventing
1950 	 *	    over-queueing of packets over choke points like modems
1951 	 *	    (at least for the transmit side).
1952 	 *
1953 	 *	(2) Is able to handle changing network loads (bandwidth
1954 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1955 	 *	    increases).
1956 	 *
1957 	 *	(3) Theoretically should stabilize in the face of multiple
1958 	 *	    connections implementing the same algorithm (this may need
1959 	 *	    a little work).
1960 	 *
1961 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1962 	 *	    be adjusted with a sysctl but typically only needs to be on
1963 	 *	    very slow connections.  A value no smaller then 5 should
1964 	 *	    be used, but only reduce this default if you have no other
1965 	 *	    choice.
1966 	 */
1967 
1968 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1969 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1970 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1971 #undef USERTT
1972 
1973 	if (tcp_inflight_debug > 0) {
1974 		static int ltime;
1975 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1976 			ltime = ticks;
1977 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1978 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1979 		}
1980 	}
1981 	if ((long)bwnd < tcp_inflight_min)
1982 		bwnd = tcp_inflight_min;
1983 	if (bwnd > tcp_inflight_max)
1984 		bwnd = tcp_inflight_max;
1985 	if ((long)bwnd < tp->t_maxseg * 2)
1986 		bwnd = tp->t_maxseg * 2;
1987 	tp->snd_bwnd = bwnd;
1988 }
1989