xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 90ea502b8c5d21f908cedff6680ee2bc9e74ce74)
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 <sys/msgport2.h>
146 #include <machine/smp.h>
147 
148 #include <net/netmsg2.h>
149 
150 #if !defined(KTR_TCP)
151 #define KTR_TCP		KTR_ALL
152 #endif
153 KTR_INFO_MASTER(tcp);
154 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
155 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
157 #define logtcp(name)	KTR_LOG(tcp_ ## name)
158 
159 struct inpcbinfo tcbinfo[MAXCPU];
160 struct tcpcbackqhead tcpcbackq[MAXCPU];
161 
162 int tcp_mpsafe_proto = 0;
163 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
164 
165 static int tcp_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE;
166 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
168 	   &tcp_mpsafe_thread, 0,
169 	   "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
170 
171 int tcp_mssdflt = TCP_MSS;
172 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
173     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
174 
175 #ifdef INET6
176 int tcp_v6mssdflt = TCP6_MSS;
177 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
178     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
179 #endif
180 
181 /*
182  * Minimum MSS we accept and use. This prevents DoS attacks where
183  * we are forced to a ridiculous low MSS like 20 and send hundreds
184  * of packets instead of one. The effect scales with the available
185  * bandwidth and quickly saturates the CPU and network interface
186  * with packet generation and sending. Set to zero to disable MINMSS
187  * checking. This setting prevents us from sending too small packets.
188  */
189 int tcp_minmss = TCP_MINMSS;
190 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
191     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
192 
193 #if 0
194 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
195 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
196     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
197 #endif
198 
199 int tcp_do_rfc1323 = 1;
200 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
201     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
202 
203 static int tcp_tcbhashsize = 0;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
205      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
206 
207 static int do_tcpdrain = 1;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
209      "Enable tcp_drain routine for extra help when low on mbufs");
210 
211 /* XXX JH */
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
213     &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
214 
215 static int icmp_may_rst = 1;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
217     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
218 
219 static int tcp_isn_reseed_interval = 0;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
221     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
222 
223 /*
224  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
225  * by default, but with generous values which should allow maximal
226  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
227  *
228  * The reason for doing this is that the limiter is the only mechanism we
229  * have which seems to do a really good job preventing receiver RX rings
230  * on network interfaces from getting blown out.  Even though GigE/10GigE
231  * is supposed to flow control it looks like either it doesn't actually
232  * do it or Open Source drivers do not properly enable it.
233  *
234  * People using the limiter to reduce bottlenecks on slower WAN connections
235  * should set the slop to 20 (2 packets).
236  */
237 static int tcp_inflight_enable = 1;
238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
239     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
240 
241 static int tcp_inflight_debug = 0;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
243     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
244 
245 static int tcp_inflight_min = 6144;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
247     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
248 
249 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
251     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
252 
253 static int tcp_inflight_stab = 50;
254 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
255     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
256 
257 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
258 static struct malloc_pipe tcptemp_mpipe;
259 
260 static void tcp_willblock(int);
261 static void tcp_notify (struct inpcb *, int);
262 
263 struct tcp_stats tcpstats_percpu[MAXCPU];
264 #ifdef SMP
265 static int
266 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
267 {
268 	int cpu, error = 0;
269 
270 	for (cpu = 0; cpu < ncpus; ++cpu) {
271 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
272 					sizeof(struct tcp_stats))))
273 			break;
274 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
275 				       sizeof(struct tcp_stats))))
276 			break;
277 	}
278 
279 	return (error);
280 }
281 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
282     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
283 #else
284 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
285     &tcpstat, tcp_stats, "TCP statistics");
286 #endif
287 
288 /*
289  * Target size of TCP PCB hash tables. Must be a power of two.
290  *
291  * Note that this can be overridden by the kernel environment
292  * variable net.inet.tcp.tcbhashsize
293  */
294 #ifndef TCBHASHSIZE
295 #define	TCBHASHSIZE	512
296 #endif
297 
298 /*
299  * This is the actual shape of what we allocate using the zone
300  * allocator.  Doing it this way allows us to protect both structures
301  * using the same generation count, and also eliminates the overhead
302  * of allocating tcpcbs separately.  By hiding the structure here,
303  * we avoid changing most of the rest of the code (although it needs
304  * to be changed, eventually, for greater efficiency).
305  */
306 #define	ALIGNMENT	32
307 #define	ALIGNM1		(ALIGNMENT - 1)
308 struct	inp_tp {
309 	union {
310 		struct	inpcb inp;
311 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
312 	} inp_tp_u;
313 	struct	tcpcb tcb;
314 	struct	tcp_callout inp_tp_rexmt;
315 	struct	tcp_callout inp_tp_persist;
316 	struct	tcp_callout inp_tp_keep;
317 	struct	tcp_callout inp_tp_2msl;
318 	struct	tcp_callout inp_tp_delack;
319 	struct	netmsg_tcp_timer inp_tp_timermsg;
320 };
321 #undef ALIGNMENT
322 #undef ALIGNM1
323 
324 /*
325  * Tcp initialization
326  */
327 void
328 tcp_init(void)
329 {
330 	struct inpcbporthead *porthashbase;
331 	u_long porthashmask;
332 	struct vm_zone *ipi_zone;
333 	int hashsize = TCBHASHSIZE;
334 	int cpu;
335 
336 	/*
337 	 * note: tcptemp is used for keepalives, and it is ok for an
338 	 * allocation to fail so do not specify MPF_INT.
339 	 */
340 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
341 		    25, -1, 0, NULL);
342 
343 	tcp_delacktime = TCPTV_DELACK;
344 	tcp_keepinit = TCPTV_KEEP_INIT;
345 	tcp_keepidle = TCPTV_KEEP_IDLE;
346 	tcp_keepintvl = TCPTV_KEEPINTVL;
347 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
348 	tcp_msl = TCPTV_MSL;
349 	tcp_rexmit_min = TCPTV_MIN;
350 	tcp_rexmit_slop = TCPTV_CPU_VAR;
351 
352 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
353 	if (!powerof2(hashsize)) {
354 		kprintf("WARNING: TCB hash size not a power of 2\n");
355 		hashsize = 512; /* safe default */
356 	}
357 	tcp_tcbhashsize = hashsize;
358 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
359 	ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
360 			 ZONE_INTERRUPT, 0);
361 
362 	for (cpu = 0; cpu < ncpus2; cpu++) {
363 		in_pcbinfo_init(&tcbinfo[cpu]);
364 		tcbinfo[cpu].cpu = cpu;
365 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
366 		    &tcbinfo[cpu].hashmask);
367 		tcbinfo[cpu].porthashbase = porthashbase;
368 		tcbinfo[cpu].porthashmask = porthashmask;
369 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
370 		    &tcbinfo[cpu].wildcardhashmask);
371 		tcbinfo[cpu].ipi_zone = ipi_zone;
372 		TAILQ_INIT(&tcpcbackq[cpu]);
373 	}
374 
375 	tcp_reass_maxseg = nmbclusters / 16;
376 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
377 
378 #ifdef INET6
379 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
380 #else
381 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
382 #endif
383 	if (max_protohdr < TCP_MINPROTOHDR)
384 		max_protohdr = TCP_MINPROTOHDR;
385 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
386 		panic("tcp_init");
387 #undef TCP_MINPROTOHDR
388 
389 	/*
390 	 * Initialize TCP statistics counters for each CPU.
391 	 */
392 #ifdef SMP
393 	for (cpu = 0; cpu < ncpus; ++cpu) {
394 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
395 	}
396 #else
397 	bzero(&tcpstat, sizeof(struct tcp_stats));
398 #endif
399 
400 	syncache_init();
401 	tcp_thread_init();
402 }
403 
404 void
405 tcpmsg_service_loop(void *dummy)
406 {
407 	struct netmsg *msg;
408 	int mplocked;
409 
410 	/*
411 	 * Thread was started with TDF_MPSAFE
412 	 */
413 	mplocked = 0;
414 
415 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
416 		do {
417 			logtcp(rxmsg);
418 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
419 						  mplocked);
420 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
421 
422 		logtcp(delayed);
423 		tcp_willblock(mplocked);
424 		logtcp(wait);
425 	}
426 }
427 
428 static void
429 tcp_willblock(int mplocked)
430 {
431 	struct tcpcb *tp;
432 	int cpu = mycpu->gd_cpuid;
433 	int unlock = 0;
434 
435 	if (!mplocked && !tcp_mpsafe_proto) {
436 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
437 			return;
438 
439 		get_mplock();
440 		mplocked = 1;
441 		unlock = 1;
442 	}
443 
444 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
445 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
446 		tp->t_flags &= ~TF_ONOUTPUTQ;
447 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
448 		tcp_output(tp);
449 	}
450 
451 	if (unlock)
452 		rel_mplock();
453 }
454 
455 
456 /*
457  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
458  * tcp_template used to store this data in mbufs, but we now recopy it out
459  * of the tcpcb each time to conserve mbufs.
460  */
461 void
462 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
463 {
464 	struct inpcb *inp = tp->t_inpcb;
465 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
466 
467 #ifdef INET6
468 	if (inp->inp_vflag & INP_IPV6) {
469 		struct ip6_hdr *ip6;
470 
471 		ip6 = (struct ip6_hdr *)ip_ptr;
472 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
473 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
474 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
475 			(IPV6_VERSION & IPV6_VERSION_MASK);
476 		ip6->ip6_nxt = IPPROTO_TCP;
477 		ip6->ip6_plen = sizeof(struct tcphdr);
478 		ip6->ip6_src = inp->in6p_laddr;
479 		ip6->ip6_dst = inp->in6p_faddr;
480 		tcp_hdr->th_sum = 0;
481 	} else
482 #endif
483 	{
484 		struct ip *ip = (struct ip *) ip_ptr;
485 
486 		ip->ip_vhl = IP_VHL_BORING;
487 		ip->ip_tos = 0;
488 		ip->ip_len = 0;
489 		ip->ip_id = 0;
490 		ip->ip_off = 0;
491 		ip->ip_ttl = 0;
492 		ip->ip_sum = 0;
493 		ip->ip_p = IPPROTO_TCP;
494 		ip->ip_src = inp->inp_laddr;
495 		ip->ip_dst = inp->inp_faddr;
496 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
497 				    ip->ip_dst.s_addr,
498 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
499 	}
500 
501 	tcp_hdr->th_sport = inp->inp_lport;
502 	tcp_hdr->th_dport = inp->inp_fport;
503 	tcp_hdr->th_seq = 0;
504 	tcp_hdr->th_ack = 0;
505 	tcp_hdr->th_x2 = 0;
506 	tcp_hdr->th_off = 5;
507 	tcp_hdr->th_flags = 0;
508 	tcp_hdr->th_win = 0;
509 	tcp_hdr->th_urp = 0;
510 }
511 
512 /*
513  * Create template to be used to send tcp packets on a connection.
514  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
515  * use for this function is in keepalives, which use tcp_respond.
516  */
517 struct tcptemp *
518 tcp_maketemplate(struct tcpcb *tp)
519 {
520 	struct tcptemp *tmp;
521 
522 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
523 		return (NULL);
524 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
525 	return (tmp);
526 }
527 
528 void
529 tcp_freetemplate(struct tcptemp *tmp)
530 {
531 	mpipe_free(&tcptemp_mpipe, tmp);
532 }
533 
534 /*
535  * Send a single message to the TCP at address specified by
536  * the given TCP/IP header.  If m == NULL, then we make a copy
537  * of the tcpiphdr at ti and send directly to the addressed host.
538  * This is used to force keep alive messages out using the TCP
539  * template for a connection.  If flags are given then we send
540  * a message back to the TCP which originated the * segment ti,
541  * and discard the mbuf containing it and any other attached mbufs.
542  *
543  * In any case the ack and sequence number of the transmitted
544  * segment are as specified by the parameters.
545  *
546  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
547  */
548 void
549 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
550 	    tcp_seq ack, tcp_seq seq, int flags)
551 {
552 	int tlen;
553 	int win = 0;
554 	struct route *ro = NULL;
555 	struct route sro;
556 	struct ip *ip = ipgen;
557 	struct tcphdr *nth;
558 	int ipflags = 0;
559 	struct route_in6 *ro6 = NULL;
560 	struct route_in6 sro6;
561 	struct ip6_hdr *ip6 = ipgen;
562 	boolean_t use_tmpro = TRUE;
563 #ifdef INET6
564 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
565 #else
566 	const boolean_t isipv6 = FALSE;
567 #endif
568 
569 	if (tp != NULL) {
570 		if (!(flags & TH_RST)) {
571 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
572 			if (win < 0)
573 				win = 0;
574 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
575 				win = (long)TCP_MAXWIN << tp->rcv_scale;
576 		}
577 		/*
578 		 * Don't use the route cache of a listen socket,
579 		 * it is not MPSAFE; use temporary route cache.
580 		 */
581 		if (tp->t_state != TCPS_LISTEN) {
582 			if (isipv6)
583 				ro6 = &tp->t_inpcb->in6p_route;
584 			else
585 				ro = &tp->t_inpcb->inp_route;
586 			use_tmpro = FALSE;
587 		}
588 	}
589 	if (use_tmpro) {
590 		if (isipv6) {
591 			ro6 = &sro6;
592 			bzero(ro6, sizeof *ro6);
593 		} else {
594 			ro = &sro;
595 			bzero(ro, sizeof *ro);
596 		}
597 	}
598 	if (m == NULL) {
599 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
600 		if (m == NULL)
601 			return;
602 		tlen = 0;
603 		m->m_data += max_linkhdr;
604 		if (isipv6) {
605 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
606 			ip6 = mtod(m, struct ip6_hdr *);
607 			nth = (struct tcphdr *)(ip6 + 1);
608 		} else {
609 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
610 			ip = mtod(m, struct ip *);
611 			nth = (struct tcphdr *)(ip + 1);
612 		}
613 		bcopy(th, nth, sizeof(struct tcphdr));
614 		flags = TH_ACK;
615 	} else {
616 		m_freem(m->m_next);
617 		m->m_next = NULL;
618 		m->m_data = (caddr_t)ipgen;
619 		/* m_len is set later */
620 		tlen = 0;
621 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
622 		if (isipv6) {
623 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
624 			nth = (struct tcphdr *)(ip6 + 1);
625 		} else {
626 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
627 			nth = (struct tcphdr *)(ip + 1);
628 		}
629 		if (th != nth) {
630 			/*
631 			 * this is usually a case when an extension header
632 			 * exists between the IPv6 header and the
633 			 * TCP header.
634 			 */
635 			nth->th_sport = th->th_sport;
636 			nth->th_dport = th->th_dport;
637 		}
638 		xchg(nth->th_dport, nth->th_sport, n_short);
639 #undef xchg
640 	}
641 	if (isipv6) {
642 		ip6->ip6_flow = 0;
643 		ip6->ip6_vfc = IPV6_VERSION;
644 		ip6->ip6_nxt = IPPROTO_TCP;
645 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
646 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
647 	} else {
648 		tlen += sizeof(struct tcpiphdr);
649 		ip->ip_len = tlen;
650 		ip->ip_ttl = ip_defttl;
651 	}
652 	m->m_len = tlen;
653 	m->m_pkthdr.len = tlen;
654 	m->m_pkthdr.rcvif = NULL;
655 	nth->th_seq = htonl(seq);
656 	nth->th_ack = htonl(ack);
657 	nth->th_x2 = 0;
658 	nth->th_off = sizeof(struct tcphdr) >> 2;
659 	nth->th_flags = flags;
660 	if (tp != NULL)
661 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
662 	else
663 		nth->th_win = htons((u_short)win);
664 	nth->th_urp = 0;
665 	if (isipv6) {
666 		nth->th_sum = 0;
667 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
668 					sizeof(struct ip6_hdr),
669 					tlen - sizeof(struct ip6_hdr));
670 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
671 					       (ro6 && ro6->ro_rt) ?
672 						ro6->ro_rt->rt_ifp : NULL);
673 	} else {
674 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
675 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
676 		m->m_pkthdr.csum_flags = CSUM_TCP;
677 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
678 	}
679 #ifdef TCPDEBUG
680 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
681 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
682 #endif
683 	if (isipv6) {
684 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
685 			   tp ? tp->t_inpcb : NULL);
686 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
687 			RTFREE(ro6->ro_rt);
688 			ro6->ro_rt = NULL;
689 		}
690 	} else {
691 		ipflags |= IP_DEBUGROUTE;
692 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
693 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
694 			RTFREE(ro->ro_rt);
695 			ro->ro_rt = NULL;
696 		}
697 	}
698 }
699 
700 /*
701  * Create a new TCP control block, making an
702  * empty reassembly queue and hooking it to the argument
703  * protocol control block.  The `inp' parameter must have
704  * come from the zone allocator set up in tcp_init().
705  */
706 struct tcpcb *
707 tcp_newtcpcb(struct inpcb *inp)
708 {
709 	struct inp_tp *it;
710 	struct tcpcb *tp;
711 #ifdef INET6
712 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
713 #else
714 	const boolean_t isipv6 = FALSE;
715 #endif
716 
717 	it = (struct inp_tp *)inp;
718 	tp = &it->tcb;
719 	bzero(tp, sizeof(struct tcpcb));
720 	LIST_INIT(&tp->t_segq);
721 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
722 
723 	/* Set up our timeouts. */
724 	tp->tt_rexmt = &it->inp_tp_rexmt;
725 	tp->tt_persist = &it->inp_tp_persist;
726 	tp->tt_keep = &it->inp_tp_keep;
727 	tp->tt_2msl = &it->inp_tp_2msl;
728 	tp->tt_delack = &it->inp_tp_delack;
729 	tcp_inittimers(tp);
730 
731 	/*
732 	 * Zero out timer message.  We don't create it here,
733 	 * since the current CPU may not be the owner of this
734 	 * inpcb.
735 	 */
736 	tp->tt_msg = &it->inp_tp_timermsg;
737 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
738 
739 	if (tcp_do_rfc1323)
740 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
741 	tp->t_inpcb = inp;	/* XXX */
742 	tp->t_state = TCPS_CLOSED;
743 	/*
744 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
745 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
746 	 * reasonable initial retransmit time.
747 	 */
748 	tp->t_srtt = TCPTV_SRTTBASE;
749 	tp->t_rttvar =
750 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
751 	tp->t_rttmin = tcp_rexmit_min;
752 	tp->t_rxtcur = TCPTV_RTOBASE;
753 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
754 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
755 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
756 	tp->t_rcvtime = ticks;
757 	/*
758 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
759 	 * because the socket may be bound to an IPv6 wildcard address,
760 	 * which may match an IPv4-mapped IPv6 address.
761 	 */
762 	inp->inp_ip_ttl = ip_defttl;
763 	inp->inp_ppcb = tp;
764 	tcp_sack_tcpcb_init(tp);
765 	return (tp);		/* XXX */
766 }
767 
768 /*
769  * Drop a TCP connection, reporting the specified error.
770  * If connection is synchronized, then send a RST to peer.
771  */
772 struct tcpcb *
773 tcp_drop(struct tcpcb *tp, int error)
774 {
775 	struct socket *so = tp->t_inpcb->inp_socket;
776 
777 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
778 		tp->t_state = TCPS_CLOSED;
779 		tcp_output(tp);
780 		tcpstat.tcps_drops++;
781 	} else
782 		tcpstat.tcps_conndrops++;
783 	if (error == ETIMEDOUT && tp->t_softerror)
784 		error = tp->t_softerror;
785 	so->so_error = error;
786 	return (tcp_close(tp));
787 }
788 
789 #ifdef SMP
790 
791 struct netmsg_remwildcard {
792 	struct netmsg		nm_netmsg;
793 	struct inpcb		*nm_inp;
794 	struct inpcbinfo	*nm_pcbinfo;
795 #if defined(INET6)
796 	int			nm_isinet6;
797 #else
798 	int			nm_unused01;
799 #endif
800 };
801 
802 /*
803  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
804  * inp can be detached.  We do this by cycling through the cpus, ending up
805  * on the cpu controlling the inp last and then doing the disconnect.
806  */
807 static void
808 in_pcbremwildcardhash_handler(struct netmsg *msg0)
809 {
810 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
811 	int cpu;
812 
813 	cpu = msg->nm_pcbinfo->cpu;
814 
815 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
816 		/* note: detach removes any wildcard hash entry */
817 #ifdef INET6
818 		if (msg->nm_isinet6)
819 			in6_pcbdetach(msg->nm_inp);
820 		else
821 #endif
822 			in_pcbdetach(msg->nm_inp);
823 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
824 	} else {
825 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
826 		cpu = (cpu + 1) % ncpus2;
827 		msg->nm_pcbinfo = &tcbinfo[cpu];
828 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
829 	}
830 }
831 
832 #endif
833 
834 /*
835  * Close a TCP control block:
836  *	discard all space held by the tcp
837  *	discard internet protocol block
838  *	wake up any sleepers
839  */
840 struct tcpcb *
841 tcp_close(struct tcpcb *tp)
842 {
843 	struct tseg_qent *q;
844 	struct inpcb *inp = tp->t_inpcb;
845 	struct socket *so = inp->inp_socket;
846 	struct rtentry *rt;
847 	boolean_t dosavessthresh;
848 #ifdef SMP
849 	int cpu;
850 #endif
851 #ifdef INET6
852 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
853 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
854 #else
855 	const boolean_t isipv6 = FALSE;
856 #endif
857 
858 	/*
859 	 * The tp is not instantly destroyed in the wildcard case.  Setting
860 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
861 	 * messing with it, though it should be noted that this change may
862 	 * not take effect on other cpus until we have chained the wildcard
863 	 * hash removal.
864 	 *
865 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
866 	 * update and prevent other tcp protocol threads from accepting new
867 	 * connections on the listen socket we might be trying to close down.
868 	 */
869 	KKASSERT(tp->t_state != TCPS_TERMINATING);
870 	tp->t_state = TCPS_TERMINATING;
871 
872 	/*
873 	 * Make sure that all of our timers are stopped before we
874 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
875 	 * timers are never used.  If timer message is never created
876 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
877 	 */
878 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
879 		tcp_callout_stop(tp, tp->tt_rexmt);
880 		tcp_callout_stop(tp, tp->tt_persist);
881 		tcp_callout_stop(tp, tp->tt_keep);
882 		tcp_callout_stop(tp, tp->tt_2msl);
883 		tcp_callout_stop(tp, tp->tt_delack);
884 	}
885 
886 	if (tp->t_flags & TF_ONOUTPUTQ) {
887 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
888 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
889 		tp->t_flags &= ~TF_ONOUTPUTQ;
890 	}
891 
892 	/*
893 	 * If we got enough samples through the srtt filter,
894 	 * save the rtt and rttvar in the routing entry.
895 	 * 'Enough' is arbitrarily defined as the 16 samples.
896 	 * 16 samples is enough for the srtt filter to converge
897 	 * to within 5% of the correct value; fewer samples and
898 	 * we could save a very bogus rtt.
899 	 *
900 	 * Don't update the default route's characteristics and don't
901 	 * update anything that the user "locked".
902 	 */
903 	if (tp->t_rttupdated >= 16) {
904 		u_long i = 0;
905 
906 		if (isipv6) {
907 			struct sockaddr_in6 *sin6;
908 
909 			if ((rt = inp->in6p_route.ro_rt) == NULL)
910 				goto no_valid_rt;
911 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
912 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
913 				goto no_valid_rt;
914 		} else
915 			if ((rt = inp->inp_route.ro_rt) == NULL ||
916 			    ((struct sockaddr_in *)rt_key(rt))->
917 			     sin_addr.s_addr == INADDR_ANY)
918 				goto no_valid_rt;
919 
920 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
921 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
922 			if (rt->rt_rmx.rmx_rtt && i)
923 				/*
924 				 * filter this update to half the old & half
925 				 * the new values, converting scale.
926 				 * See route.h and tcp_var.h for a
927 				 * description of the scaling constants.
928 				 */
929 				rt->rt_rmx.rmx_rtt =
930 				    (rt->rt_rmx.rmx_rtt + i) / 2;
931 			else
932 				rt->rt_rmx.rmx_rtt = i;
933 			tcpstat.tcps_cachedrtt++;
934 		}
935 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
936 			i = tp->t_rttvar *
937 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
938 			if (rt->rt_rmx.rmx_rttvar && i)
939 				rt->rt_rmx.rmx_rttvar =
940 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
941 			else
942 				rt->rt_rmx.rmx_rttvar = i;
943 			tcpstat.tcps_cachedrttvar++;
944 		}
945 		/*
946 		 * The old comment here said:
947 		 * update the pipelimit (ssthresh) if it has been updated
948 		 * already or if a pipesize was specified & the threshhold
949 		 * got below half the pipesize.  I.e., wait for bad news
950 		 * before we start updating, then update on both good
951 		 * and bad news.
952 		 *
953 		 * But we want to save the ssthresh even if no pipesize is
954 		 * specified explicitly in the route, because such
955 		 * connections still have an implicit pipesize specified
956 		 * by the global tcp_sendspace.  In the absence of a reliable
957 		 * way to calculate the pipesize, it will have to do.
958 		 */
959 		i = tp->snd_ssthresh;
960 		if (rt->rt_rmx.rmx_sendpipe != 0)
961 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
962 		else
963 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
964 		if (dosavessthresh ||
965 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
966 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
967 			/*
968 			 * convert the limit from user data bytes to
969 			 * packets then to packet data bytes.
970 			 */
971 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
972 			if (i < 2)
973 				i = 2;
974 			i *= tp->t_maxseg +
975 			     (isipv6 ?
976 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
977 			      sizeof(struct tcpiphdr));
978 			if (rt->rt_rmx.rmx_ssthresh)
979 				rt->rt_rmx.rmx_ssthresh =
980 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
981 			else
982 				rt->rt_rmx.rmx_ssthresh = i;
983 			tcpstat.tcps_cachedssthresh++;
984 		}
985 	}
986 
987 no_valid_rt:
988 	/* free the reassembly queue, if any */
989 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
990 		LIST_REMOVE(q, tqe_q);
991 		m_freem(q->tqe_m);
992 		FREE(q, M_TSEGQ);
993 		tcp_reass_qsize--;
994 	}
995 	/* throw away SACK blocks in scoreboard*/
996 	if (TCP_DO_SACK(tp))
997 		tcp_sack_cleanup(&tp->scb);
998 
999 	inp->inp_ppcb = NULL;
1000 	soisdisconnected(so);
1001 
1002 	tcp_destroy_timermsg(tp);
1003 
1004 	/*
1005 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
1006 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
1007 	 * for each protocol thread and must be removed in the context of
1008 	 * that thread.  This is accomplished by chaining the message
1009 	 * through the cpus.
1010 	 *
1011 	 * If the inp is not wildcarded we simply detach, which will remove
1012 	 * the any hashes still present for this inp.
1013 	 */
1014 #ifdef SMP
1015 	if (inp->inp_flags & INP_WILDCARD_MP) {
1016 		struct netmsg_remwildcard *msg;
1017 
1018 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1019 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
1020 			      M_LWKTMSG, M_INTWAIT);
1021 		netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1022 			    0, in_pcbremwildcardhash_handler);
1023 #ifdef INET6
1024 		msg->nm_isinet6 = isafinet6;
1025 #endif
1026 		msg->nm_inp = inp;
1027 		msg->nm_pcbinfo = &tcbinfo[cpu];
1028 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1029 	} else
1030 #endif
1031 	{
1032 		/* note: detach removes any wildcard hash entry */
1033 #ifdef INET6
1034 		if (isafinet6)
1035 			in6_pcbdetach(inp);
1036 		else
1037 #endif
1038 			in_pcbdetach(inp);
1039 	}
1040 	tcpstat.tcps_closed++;
1041 	return (NULL);
1042 }
1043 
1044 static __inline void
1045 tcp_drain_oncpu(struct inpcbhead *head)
1046 {
1047 	struct inpcb *inpb;
1048 	struct tcpcb *tcpb;
1049 	struct tseg_qent *te;
1050 
1051 	LIST_FOREACH(inpb, head, inp_list) {
1052 		if (inpb->inp_flags & INP_PLACEMARKER)
1053 			continue;
1054 		if ((tcpb = intotcpcb(inpb))) {
1055 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1056 				LIST_REMOVE(te, tqe_q);
1057 				m_freem(te->tqe_m);
1058 				FREE(te, M_TSEGQ);
1059 				tcp_reass_qsize--;
1060 			}
1061 		}
1062 	}
1063 }
1064 
1065 #ifdef SMP
1066 struct netmsg_tcp_drain {
1067 	struct netmsg		nm_netmsg;
1068 	struct inpcbhead	*nm_head;
1069 };
1070 
1071 static void
1072 tcp_drain_handler(netmsg_t netmsg)
1073 {
1074 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1075 
1076 	tcp_drain_oncpu(nm->nm_head);
1077 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1078 }
1079 #endif
1080 
1081 void
1082 tcp_drain(void)
1083 {
1084 #ifdef SMP
1085 	int cpu;
1086 #endif
1087 
1088 	if (!do_tcpdrain)
1089 		return;
1090 
1091 	/*
1092 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1093 	 * if there is one...
1094 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1095 	 *	reassembly queue should be flushed, but in a situation
1096 	 *	where we're really low on mbufs, this is potentially
1097 	 *	useful.
1098 	 */
1099 #ifdef SMP
1100 	for (cpu = 0; cpu < ncpus2; cpu++) {
1101 		struct netmsg_tcp_drain *msg;
1102 
1103 		if (cpu == mycpu->gd_cpuid) {
1104 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1105 		} else {
1106 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1107 				    M_LWKTMSG, M_NOWAIT);
1108 			if (msg == NULL)
1109 				continue;
1110 			netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1111 				    0, tcp_drain_handler);
1112 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1113 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1114 		}
1115 	}
1116 #else
1117 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1118 #endif
1119 }
1120 
1121 /*
1122  * Notify a tcp user of an asynchronous error;
1123  * store error as soft error, but wake up user
1124  * (for now, won't do anything until can select for soft error).
1125  *
1126  * Do not wake up user since there currently is no mechanism for
1127  * reporting soft errors (yet - a kqueue filter may be added).
1128  */
1129 static void
1130 tcp_notify(struct inpcb *inp, int error)
1131 {
1132 	struct tcpcb *tp = intotcpcb(inp);
1133 
1134 	/*
1135 	 * Ignore some errors if we are hooked up.
1136 	 * If connection hasn't completed, has retransmitted several times,
1137 	 * and receives a second error, give up now.  This is better
1138 	 * than waiting a long time to establish a connection that
1139 	 * can never complete.
1140 	 */
1141 	if (tp->t_state == TCPS_ESTABLISHED &&
1142 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1143 	      error == EHOSTDOWN)) {
1144 		return;
1145 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1146 	    tp->t_softerror)
1147 		tcp_drop(tp, error);
1148 	else
1149 		tp->t_softerror = error;
1150 #if 0
1151 	wakeup(&so->so_timeo);
1152 	sorwakeup(so);
1153 	sowwakeup(so);
1154 #endif
1155 }
1156 
1157 static int
1158 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1159 {
1160 	int error, i, n;
1161 	struct inpcb *marker;
1162 	struct inpcb *inp;
1163 	inp_gen_t gencnt;
1164 	globaldata_t gd;
1165 	int origcpu, ccpu;
1166 
1167 	error = 0;
1168 	n = 0;
1169 
1170 	/*
1171 	 * The process of preparing the TCB list is too time-consuming and
1172 	 * resource-intensive to repeat twice on every request.
1173 	 */
1174 	if (req->oldptr == NULL) {
1175 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1176 			gd = globaldata_find(ccpu);
1177 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1178 		}
1179 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1180 		return (0);
1181 	}
1182 
1183 	if (req->newptr != NULL)
1184 		return (EPERM);
1185 
1186 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1187 	marker->inp_flags |= INP_PLACEMARKER;
1188 
1189 	/*
1190 	 * OK, now we're committed to doing something.  Run the inpcb list
1191 	 * for each cpu in the system and construct the output.  Use a
1192 	 * list placemarker to deal with list changes occuring during
1193 	 * copyout blockages (but otherwise depend on being on the correct
1194 	 * cpu to avoid races).
1195 	 */
1196 	origcpu = mycpu->gd_cpuid;
1197 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1198 		globaldata_t rgd;
1199 		caddr_t inp_ppcb;
1200 		struct xtcpcb xt;
1201 		int cpu_id;
1202 
1203 		cpu_id = (origcpu + ccpu) % ncpus;
1204 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1205 			continue;
1206 		rgd = globaldata_find(cpu_id);
1207 		lwkt_setcpu_self(rgd);
1208 
1209 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1210 		n = tcbinfo[cpu_id].ipi_count;
1211 
1212 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1213 		i = 0;
1214 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1215 			/*
1216 			 * process a snapshot of pcbs, ignoring placemarkers
1217 			 * and using our own to allow SYSCTL_OUT to block.
1218 			 */
1219 			LIST_REMOVE(marker, inp_list);
1220 			LIST_INSERT_AFTER(inp, marker, inp_list);
1221 
1222 			if (inp->inp_flags & INP_PLACEMARKER)
1223 				continue;
1224 			if (inp->inp_gencnt > gencnt)
1225 				continue;
1226 			if (prison_xinpcb(req->td, inp))
1227 				continue;
1228 
1229 			xt.xt_len = sizeof xt;
1230 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1231 			inp_ppcb = inp->inp_ppcb;
1232 			if (inp_ppcb != NULL)
1233 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1234 			else
1235 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1236 			if (inp->inp_socket)
1237 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1238 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1239 				break;
1240 			++i;
1241 		}
1242 		LIST_REMOVE(marker, inp_list);
1243 		if (error == 0 && i < n) {
1244 			bzero(&xt, sizeof xt);
1245 			xt.xt_len = sizeof xt;
1246 			while (i < n) {
1247 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1248 				if (error)
1249 					break;
1250 				++i;
1251 			}
1252 		}
1253 	}
1254 
1255 	/*
1256 	 * Make sure we are on the same cpu we were on originally, since
1257 	 * higher level callers expect this.  Also don't pollute caches with
1258 	 * migrated userland data by (eventually) returning to userland
1259 	 * on a different cpu.
1260 	 */
1261 	lwkt_setcpu_self(globaldata_find(origcpu));
1262 	kfree(marker, M_TEMP);
1263 	return (error);
1264 }
1265 
1266 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1267 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1268 
1269 static int
1270 tcp_getcred(SYSCTL_HANDLER_ARGS)
1271 {
1272 	struct sockaddr_in addrs[2];
1273 	struct inpcb *inp;
1274 	int cpu;
1275 	int error;
1276 
1277 	error = priv_check(req->td, PRIV_ROOT);
1278 	if (error != 0)
1279 		return (error);
1280 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1281 	if (error != 0)
1282 		return (error);
1283 	crit_enter();
1284 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1285 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1286 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1287 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1288 	if (inp == NULL || inp->inp_socket == NULL) {
1289 		error = ENOENT;
1290 		goto out;
1291 	}
1292 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1293 out:
1294 	crit_exit();
1295 	return (error);
1296 }
1297 
1298 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1299     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1300 
1301 #ifdef INET6
1302 static int
1303 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1304 {
1305 	struct sockaddr_in6 addrs[2];
1306 	struct inpcb *inp;
1307 	int error;
1308 	boolean_t mapped = FALSE;
1309 
1310 	error = priv_check(req->td, PRIV_ROOT);
1311 	if (error != 0)
1312 		return (error);
1313 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1314 	if (error != 0)
1315 		return (error);
1316 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1317 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1318 			mapped = TRUE;
1319 		else
1320 			return (EINVAL);
1321 	}
1322 	crit_enter();
1323 	if (mapped) {
1324 		inp = in_pcblookup_hash(&tcbinfo[0],
1325 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1326 		    addrs[1].sin6_port,
1327 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1328 		    addrs[0].sin6_port,
1329 		    0, NULL);
1330 	} else {
1331 		inp = in6_pcblookup_hash(&tcbinfo[0],
1332 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1333 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1334 		    0, NULL);
1335 	}
1336 	if (inp == NULL || inp->inp_socket == NULL) {
1337 		error = ENOENT;
1338 		goto out;
1339 	}
1340 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1341 out:
1342 	crit_exit();
1343 	return (error);
1344 }
1345 
1346 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1347 	    0, 0,
1348 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1349 #endif
1350 
1351 struct netmsg_tcp_notify {
1352 	struct netmsg	nm_nmsg;
1353 	void		(*nm_notify)(struct inpcb *, int);
1354 	struct in_addr	nm_faddr;
1355 	int		nm_arg;
1356 };
1357 
1358 static void
1359 tcp_notifyall_oncpu(struct netmsg *netmsg)
1360 {
1361 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1362 	int nextcpu;
1363 
1364 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1365 			nmsg->nm_arg, nmsg->nm_notify);
1366 
1367 	nextcpu = mycpuid + 1;
1368 	if (nextcpu < ncpus2)
1369 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1370 	else
1371 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1372 }
1373 
1374 void
1375 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1376 {
1377 	struct ip *ip = vip;
1378 	struct tcphdr *th;
1379 	struct in_addr faddr;
1380 	struct inpcb *inp;
1381 	struct tcpcb *tp;
1382 	void (*notify)(struct inpcb *, int) = tcp_notify;
1383 	tcp_seq icmpseq;
1384 	int arg, cpu;
1385 
1386 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1387 		return;
1388 	}
1389 
1390 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1391 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1392 		return;
1393 
1394 	arg = inetctlerrmap[cmd];
1395 	if (cmd == PRC_QUENCH) {
1396 		notify = tcp_quench;
1397 	} else if (icmp_may_rst &&
1398 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1399 		    cmd == PRC_UNREACH_PORT ||
1400 		    cmd == PRC_TIMXCEED_INTRANS) &&
1401 		   ip != NULL) {
1402 		notify = tcp_drop_syn_sent;
1403 	} else if (cmd == PRC_MSGSIZE) {
1404 		struct icmp *icmp = (struct icmp *)
1405 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1406 
1407 		arg = ntohs(icmp->icmp_nextmtu);
1408 		notify = tcp_mtudisc;
1409 	} else if (PRC_IS_REDIRECT(cmd)) {
1410 		ip = NULL;
1411 		notify = in_rtchange;
1412 	} else if (cmd == PRC_HOSTDEAD) {
1413 		ip = NULL;
1414 	}
1415 
1416 	if (ip != NULL) {
1417 		crit_enter();
1418 		th = (struct tcphdr *)((caddr_t)ip +
1419 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1420 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1421 				  ip->ip_src.s_addr, th->th_sport);
1422 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1423 					ip->ip_src, th->th_sport, 0, NULL);
1424 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1425 			icmpseq = htonl(th->th_seq);
1426 			tp = intotcpcb(inp);
1427 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1428 			    SEQ_LT(icmpseq, tp->snd_max))
1429 				(*notify)(inp, arg);
1430 		} else {
1431 			struct in_conninfo inc;
1432 
1433 			inc.inc_fport = th->th_dport;
1434 			inc.inc_lport = th->th_sport;
1435 			inc.inc_faddr = faddr;
1436 			inc.inc_laddr = ip->ip_src;
1437 #ifdef INET6
1438 			inc.inc_isipv6 = 0;
1439 #endif
1440 			syncache_unreach(&inc, th);
1441 		}
1442 		crit_exit();
1443 	} else {
1444 		struct netmsg_tcp_notify nmsg;
1445 
1446 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1447 		netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1448 			    0, tcp_notifyall_oncpu);
1449 		nmsg.nm_faddr = faddr;
1450 		nmsg.nm_arg = arg;
1451 		nmsg.nm_notify = notify;
1452 
1453 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1454 	}
1455 }
1456 
1457 #ifdef INET6
1458 void
1459 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1460 {
1461 	struct tcphdr th;
1462 	void (*notify) (struct inpcb *, int) = tcp_notify;
1463 	struct ip6_hdr *ip6;
1464 	struct mbuf *m;
1465 	struct ip6ctlparam *ip6cp = NULL;
1466 	const struct sockaddr_in6 *sa6_src = NULL;
1467 	int off;
1468 	struct tcp_portonly {
1469 		u_int16_t th_sport;
1470 		u_int16_t th_dport;
1471 	} *thp;
1472 	int arg;
1473 
1474 	if (sa->sa_family != AF_INET6 ||
1475 	    sa->sa_len != sizeof(struct sockaddr_in6))
1476 		return;
1477 
1478 	arg = 0;
1479 	if (cmd == PRC_QUENCH)
1480 		notify = tcp_quench;
1481 	else if (cmd == PRC_MSGSIZE) {
1482 		struct ip6ctlparam *ip6cp = d;
1483 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1484 
1485 		arg = ntohl(icmp6->icmp6_mtu);
1486 		notify = tcp_mtudisc;
1487 	} else if (!PRC_IS_REDIRECT(cmd) &&
1488 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1489 		return;
1490 	}
1491 
1492 	/* if the parameter is from icmp6, decode it. */
1493 	if (d != NULL) {
1494 		ip6cp = (struct ip6ctlparam *)d;
1495 		m = ip6cp->ip6c_m;
1496 		ip6 = ip6cp->ip6c_ip6;
1497 		off = ip6cp->ip6c_off;
1498 		sa6_src = ip6cp->ip6c_src;
1499 	} else {
1500 		m = NULL;
1501 		ip6 = NULL;
1502 		off = 0;	/* fool gcc */
1503 		sa6_src = &sa6_any;
1504 	}
1505 
1506 	if (ip6 != NULL) {
1507 		struct in_conninfo inc;
1508 		/*
1509 		 * XXX: We assume that when IPV6 is non NULL,
1510 		 * M and OFF are valid.
1511 		 */
1512 
1513 		/* check if we can safely examine src and dst ports */
1514 		if (m->m_pkthdr.len < off + sizeof *thp)
1515 			return;
1516 
1517 		bzero(&th, sizeof th);
1518 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1519 
1520 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1521 		    (struct sockaddr *)ip6cp->ip6c_src,
1522 		    th.th_sport, cmd, arg, notify);
1523 
1524 		inc.inc_fport = th.th_dport;
1525 		inc.inc_lport = th.th_sport;
1526 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1527 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1528 		inc.inc_isipv6 = 1;
1529 		syncache_unreach(&inc, &th);
1530 	} else
1531 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1532 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1533 }
1534 #endif
1535 
1536 /*
1537  * Following is where TCP initial sequence number generation occurs.
1538  *
1539  * There are two places where we must use initial sequence numbers:
1540  * 1.  In SYN-ACK packets.
1541  * 2.  In SYN packets.
1542  *
1543  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1544  * tcp_syncache.c for details.
1545  *
1546  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1547  * depends on this property.  In addition, these ISNs should be
1548  * unguessable so as to prevent connection hijacking.  To satisfy
1549  * the requirements of this situation, the algorithm outlined in
1550  * RFC 1948 is used to generate sequence numbers.
1551  *
1552  * Implementation details:
1553  *
1554  * Time is based off the system timer, and is corrected so that it
1555  * increases by one megabyte per second.  This allows for proper
1556  * recycling on high speed LANs while still leaving over an hour
1557  * before rollover.
1558  *
1559  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1560  * between seeding of isn_secret.  This is normally set to zero,
1561  * as reseeding should not be necessary.
1562  *
1563  */
1564 
1565 #define	ISN_BYTES_PER_SECOND 1048576
1566 
1567 u_char isn_secret[32];
1568 int isn_last_reseed;
1569 MD5_CTX isn_ctx;
1570 
1571 tcp_seq
1572 tcp_new_isn(struct tcpcb *tp)
1573 {
1574 	u_int32_t md5_buffer[4];
1575 	tcp_seq new_isn;
1576 
1577 	/* Seed if this is the first use, reseed if requested. */
1578 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1579 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1580 		< (u_int)ticks))) {
1581 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1582 		isn_last_reseed = ticks;
1583 	}
1584 
1585 	/* Compute the md5 hash and return the ISN. */
1586 	MD5Init(&isn_ctx);
1587 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1588 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1589 #ifdef INET6
1590 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1591 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1592 			  sizeof(struct in6_addr));
1593 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1594 			  sizeof(struct in6_addr));
1595 	} else
1596 #endif
1597 	{
1598 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1599 			  sizeof(struct in_addr));
1600 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1601 			  sizeof(struct in_addr));
1602 	}
1603 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1604 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1605 	new_isn = (tcp_seq) md5_buffer[0];
1606 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1607 	return (new_isn);
1608 }
1609 
1610 /*
1611  * When a source quench is received, close congestion window
1612  * to one segment.  We will gradually open it again as we proceed.
1613  */
1614 void
1615 tcp_quench(struct inpcb *inp, int error)
1616 {
1617 	struct tcpcb *tp = intotcpcb(inp);
1618 
1619 	if (tp != NULL) {
1620 		tp->snd_cwnd = tp->t_maxseg;
1621 		tp->snd_wacked = 0;
1622 	}
1623 }
1624 
1625 /*
1626  * When a specific ICMP unreachable message is received and the
1627  * connection state is SYN-SENT, drop the connection.  This behavior
1628  * is controlled by the icmp_may_rst sysctl.
1629  */
1630 void
1631 tcp_drop_syn_sent(struct inpcb *inp, int error)
1632 {
1633 	struct tcpcb *tp = intotcpcb(inp);
1634 
1635 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1636 		tcp_drop(tp, error);
1637 }
1638 
1639 /*
1640  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1641  * based on the new value in the route.  Also nudge TCP to send something,
1642  * since we know the packet we just sent was dropped.
1643  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1644  */
1645 void
1646 tcp_mtudisc(struct inpcb *inp, int mtu)
1647 {
1648 	struct tcpcb *tp = intotcpcb(inp);
1649 	struct rtentry *rt;
1650 	struct socket *so = inp->inp_socket;
1651 	int maxopd, mss;
1652 #ifdef INET6
1653 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1654 #else
1655 	const boolean_t isipv6 = FALSE;
1656 #endif
1657 
1658 	if (tp == NULL)
1659 		return;
1660 
1661 	/*
1662 	 * If no MTU is provided in the ICMP message, use the
1663 	 * next lower likely value, as specified in RFC 1191.
1664 	 */
1665 	if (mtu == 0) {
1666 		int oldmtu;
1667 
1668 		oldmtu = tp->t_maxopd +
1669 		    (isipv6 ?
1670 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1671 		     sizeof(struct tcpiphdr));
1672 		mtu = ip_next_mtu(oldmtu, 0);
1673 	}
1674 
1675 	if (isipv6)
1676 		rt = tcp_rtlookup6(&inp->inp_inc);
1677 	else
1678 		rt = tcp_rtlookup(&inp->inp_inc);
1679 	if (rt != NULL) {
1680 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1681 			mtu = rt->rt_rmx.rmx_mtu;
1682 
1683 		maxopd = mtu -
1684 		    (isipv6 ?
1685 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1686 		     sizeof(struct tcpiphdr));
1687 
1688 		/*
1689 		 * XXX - The following conditional probably violates the TCP
1690 		 * spec.  The problem is that, since we don't know the
1691 		 * other end's MSS, we are supposed to use a conservative
1692 		 * default.  But, if we do that, then MTU discovery will
1693 		 * never actually take place, because the conservative
1694 		 * default is much less than the MTUs typically seen
1695 		 * on the Internet today.  For the moment, we'll sweep
1696 		 * this under the carpet.
1697 		 *
1698 		 * The conservative default might not actually be a problem
1699 		 * if the only case this occurs is when sending an initial
1700 		 * SYN with options and data to a host we've never talked
1701 		 * to before.  Then, they will reply with an MSS value which
1702 		 * will get recorded and the new parameters should get
1703 		 * recomputed.  For Further Study.
1704 		 */
1705 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1706 			maxopd = rt->rt_rmx.rmx_mssopt;
1707 	} else
1708 		maxopd = mtu -
1709 		    (isipv6 ?
1710 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1711 		     sizeof(struct tcpiphdr));
1712 
1713 	if (tp->t_maxopd <= maxopd)
1714 		return;
1715 	tp->t_maxopd = maxopd;
1716 
1717 	mss = maxopd;
1718 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1719 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1720 		mss -= TCPOLEN_TSTAMP_APPA;
1721 
1722 	/* round down to multiple of MCLBYTES */
1723 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1724 	if (mss > MCLBYTES)
1725 		mss &= ~(MCLBYTES - 1);
1726 #else
1727 	if (mss > MCLBYTES)
1728 		mss = (mss / MCLBYTES) * MCLBYTES;
1729 #endif
1730 
1731 	if (so->so_snd.ssb_hiwat < mss)
1732 		mss = so->so_snd.ssb_hiwat;
1733 
1734 	tp->t_maxseg = mss;
1735 	tp->t_rtttime = 0;
1736 	tp->snd_nxt = tp->snd_una;
1737 	tcp_output(tp);
1738 	tcpstat.tcps_mturesent++;
1739 }
1740 
1741 /*
1742  * Look-up the routing entry to the peer of this inpcb.  If no route
1743  * is found and it cannot be allocated the return NULL.  This routine
1744  * is called by TCP routines that access the rmx structure and by tcp_mss
1745  * to get the interface MTU.
1746  */
1747 struct rtentry *
1748 tcp_rtlookup(struct in_conninfo *inc)
1749 {
1750 	struct route *ro = &inc->inc_route;
1751 
1752 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1753 		/* No route yet, so try to acquire one */
1754 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1755 			/*
1756 			 * unused portions of the structure MUST be zero'd
1757 			 * out because rtalloc() treats it as opaque data
1758 			 */
1759 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1760 			ro->ro_dst.sa_family = AF_INET;
1761 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1762 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1763 			    inc->inc_faddr;
1764 			rtalloc(ro);
1765 		}
1766 	}
1767 	return (ro->ro_rt);
1768 }
1769 
1770 #ifdef INET6
1771 struct rtentry *
1772 tcp_rtlookup6(struct in_conninfo *inc)
1773 {
1774 	struct route_in6 *ro6 = &inc->inc6_route;
1775 
1776 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1777 		/* No route yet, so try to acquire one */
1778 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1779 			/*
1780 			 * unused portions of the structure MUST be zero'd
1781 			 * out because rtalloc() treats it as opaque data
1782 			 */
1783 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1784 			ro6->ro_dst.sin6_family = AF_INET6;
1785 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1786 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1787 			rtalloc((struct route *)ro6);
1788 		}
1789 	}
1790 	return (ro6->ro_rt);
1791 }
1792 #endif
1793 
1794 #ifdef IPSEC
1795 /* compute ESP/AH header size for TCP, including outer IP header. */
1796 size_t
1797 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1798 {
1799 	struct inpcb *inp;
1800 	struct mbuf *m;
1801 	size_t hdrsiz;
1802 	struct ip *ip;
1803 	struct tcphdr *th;
1804 
1805 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1806 		return (0);
1807 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1808 	if (!m)
1809 		return (0);
1810 
1811 #ifdef INET6
1812 	if (inp->inp_vflag & INP_IPV6) {
1813 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1814 
1815 		th = (struct tcphdr *)(ip6 + 1);
1816 		m->m_pkthdr.len = m->m_len =
1817 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1818 		tcp_fillheaders(tp, ip6, th);
1819 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1820 	} else
1821 #endif
1822 	{
1823 		ip = mtod(m, struct ip *);
1824 		th = (struct tcphdr *)(ip + 1);
1825 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1826 		tcp_fillheaders(tp, ip, th);
1827 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1828 	}
1829 
1830 	m_free(m);
1831 	return (hdrsiz);
1832 }
1833 #endif
1834 
1835 /*
1836  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1837  *
1838  * This code attempts to calculate the bandwidth-delay product as a
1839  * means of determining the optimal window size to maximize bandwidth,
1840  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1841  * routers.  This code also does a fairly good job keeping RTTs in check
1842  * across slow links like modems.  We implement an algorithm which is very
1843  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1844  * transmitter side of a TCP connection and so only effects the transmit
1845  * side of the connection.
1846  *
1847  * BACKGROUND:  TCP makes no provision for the management of buffer space
1848  * at the end points or at the intermediate routers and switches.  A TCP
1849  * stream, whether using NewReno or not, will eventually buffer as
1850  * many packets as it is able and the only reason this typically works is
1851  * due to the fairly small default buffers made available for a connection
1852  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1853  * scaling it is now fairly easy for even a single TCP connection to blow-out
1854  * all available buffer space not only on the local interface, but on
1855  * intermediate routers and switches as well.  NewReno makes a misguided
1856  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1857  * then backing off, then steadily increasing the window again until another
1858  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1859  * is only made worse as network loads increase and the idea of intentionally
1860  * blowing out network buffers is, frankly, a terrible way to manage network
1861  * resources.
1862  *
1863  * It is far better to limit the transmit window prior to the failure
1864  * condition being achieved.  There are two general ways to do this:  First
1865  * you can 'scan' through different transmit window sizes and locate the
1866  * point where the RTT stops increasing, indicating that you have filled the
1867  * pipe, then scan backwards until you note that RTT stops decreasing, then
1868  * repeat ad-infinitum.  This method works in principle but has severe
1869  * implementation issues due to RTT variances, timer granularity, and
1870  * instability in the algorithm which can lead to many false positives and
1871  * create oscillations as well as interact badly with other TCP streams
1872  * implementing the same algorithm.
1873  *
1874  * The second method is to limit the window to the bandwidth delay product
1875  * of the link.  This is the method we implement.  RTT variances and our
1876  * own manipulation of the congestion window, bwnd, can potentially
1877  * destabilize the algorithm.  For this reason we have to stabilize the
1878  * elements used to calculate the window.  We do this by using the minimum
1879  * observed RTT, the long term average of the observed bandwidth, and
1880  * by adding two segments worth of slop.  It isn't perfect but it is able
1881  * to react to changing conditions and gives us a very stable basis on
1882  * which to extend the algorithm.
1883  */
1884 void
1885 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1886 {
1887 	u_long bw;
1888 	u_long bwnd;
1889 	int save_ticks;
1890 	int delta_ticks;
1891 
1892 	/*
1893 	 * If inflight_enable is disabled in the middle of a tcp connection,
1894 	 * make sure snd_bwnd is effectively disabled.
1895 	 */
1896 	if (!tcp_inflight_enable) {
1897 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1898 		tp->snd_bandwidth = 0;
1899 		return;
1900 	}
1901 
1902 	/*
1903 	 * Validate the delta time.  If a connection is new or has been idle
1904 	 * a long time we have to reset the bandwidth calculator.
1905 	 */
1906 	save_ticks = ticks;
1907 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1908 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1909 		tp->t_bw_rtttime = ticks;
1910 		tp->t_bw_rtseq = ack_seq;
1911 		if (tp->snd_bandwidth == 0)
1912 			tp->snd_bandwidth = tcp_inflight_min;
1913 		return;
1914 	}
1915 	if (delta_ticks == 0)
1916 		return;
1917 
1918 	/*
1919 	 * Sanity check, plus ignore pure window update acks.
1920 	 */
1921 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1922 		return;
1923 
1924 	/*
1925 	 * Figure out the bandwidth.  Due to the tick granularity this
1926 	 * is a very rough number and it MUST be averaged over a fairly
1927 	 * long period of time.  XXX we need to take into account a link
1928 	 * that is not using all available bandwidth, but for now our
1929 	 * slop will ramp us up if this case occurs and the bandwidth later
1930 	 * increases.
1931 	 */
1932 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1933 	tp->t_bw_rtttime = save_ticks;
1934 	tp->t_bw_rtseq = ack_seq;
1935 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1936 
1937 	tp->snd_bandwidth = bw;
1938 
1939 	/*
1940 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1941 	 * segments.  The additional slop puts us squarely in the sweet
1942 	 * spot and also handles the bandwidth run-up case.  Without the
1943 	 * slop we could be locking ourselves into a lower bandwidth.
1944 	 *
1945 	 * Situations Handled:
1946 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1947 	 *	    high speed LANs, allowing larger TCP buffers to be
1948 	 *	    specified, and also does a good job preventing
1949 	 *	    over-queueing of packets over choke points like modems
1950 	 *	    (at least for the transmit side).
1951 	 *
1952 	 *	(2) Is able to handle changing network loads (bandwidth
1953 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1954 	 *	    increases).
1955 	 *
1956 	 *	(3) Theoretically should stabilize in the face of multiple
1957 	 *	    connections implementing the same algorithm (this may need
1958 	 *	    a little work).
1959 	 *
1960 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1961 	 *	    be adjusted with a sysctl but typically only needs to be on
1962 	 *	    very slow connections.  A value no smaller then 5 should
1963 	 *	    be used, but only reduce this default if you have no other
1964 	 *	    choice.
1965 	 */
1966 
1967 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1968 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1969 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1970 #undef USERTT
1971 
1972 	if (tcp_inflight_debug > 0) {
1973 		static int ltime;
1974 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1975 			ltime = ticks;
1976 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1977 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1978 		}
1979 	}
1980 	if ((long)bwnd < tcp_inflight_min)
1981 		bwnd = tcp_inflight_min;
1982 	if (bwnd > tcp_inflight_max)
1983 		bwnd = tcp_inflight_max;
1984 	if ((long)bwnd < tp->t_maxseg * 2)
1985 		bwnd = tp->t_maxseg * 2;
1986 	tp->snd_bwnd = bwnd;
1987 }
1988