xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 211d4362597aee676ecea315377d5cb13da26bb5)
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  */
69 
70 #include "opt_compat.h"
71 #include "opt_inet.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/socketops.h>
91 #include <sys/socketvar.h>
92 #include <sys/protosw.h>
93 #include <sys/random.h>
94 #include <sys/in_cksum.h>
95 #include <sys/ktr.h>
96 
97 #include <net/route.h>
98 #include <net/if.h>
99 #include <net/netisr.h>
100 
101 #define	_IP_VHL
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
112 #ifdef INET6
113 #include <netinet/icmp6.h>
114 #endif
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_timer2.h>
120 #include <netinet/tcp_var.h>
121 #include <netinet6/tcp6_var.h>
122 #include <netinet/tcpip.h>
123 #ifdef TCPDEBUG
124 #include <netinet/tcp_debug.h>
125 #endif
126 #include <netinet6/ip6protosw.h>
127 
128 #ifdef IPSEC
129 #include <netinet6/ipsec.h>
130 #include <netproto/key/key.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 /*
155 KTR_INFO_MASTER(tcp);
156 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
158 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
159 #define logtcp(name)	KTR_LOG(tcp_ ## name)
160 */
161 
162 #define TCP_IW_MAXSEGS_DFLT	4
163 #define TCP_IW_CAPSEGS_DFLT	3
164 
165 struct inpcbinfo tcbinfo[MAXCPU];
166 struct tcpcbackqhead tcpcbackq[MAXCPU];
167 
168 static struct lwkt_token tcp_port_token =
169 		LWKT_TOKEN_INITIALIZER(tcp_port_token);
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 static int icmp_may_rst = 1;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
213     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
214 
215 static int tcp_isn_reseed_interval = 0;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
217     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
218 
219 /*
220  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
221  * by default, but with generous values which should allow maximal
222  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
223  *
224  * The reason for doing this is that the limiter is the only mechanism we
225  * have which seems to do a really good job preventing receiver RX rings
226  * on network interfaces from getting blown out.  Even though GigE/10GigE
227  * is supposed to flow control it looks like either it doesn't actually
228  * do it or Open Source drivers do not properly enable it.
229  *
230  * People using the limiter to reduce bottlenecks on slower WAN connections
231  * should set the slop to 20 (2 packets).
232  */
233 static int tcp_inflight_enable = 1;
234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
235     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
236 
237 static int tcp_inflight_debug = 0;
238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
239     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
240 
241 static int tcp_inflight_min = 6144;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
243     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
244 
245 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
247     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
248 
249 static int tcp_inflight_stab = 50;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
251     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
252 
253 static int tcp_do_rfc3390 = 1;
254 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
255     &tcp_do_rfc3390, 0,
256     "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
257 
258 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
259 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW,
260     &tcp_iw_maxsegs, 0, "TCP IW segments max");
261 
262 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
263 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW,
264     &tcp_iw_capsegs, 0, "TCP IW segments");
265 
266 int tcp_low_rtobase = 1;
267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW,
268     &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)");
269 
270 static int tcp_do_ncr = 1;
271 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW,
272     &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)");
273 
274 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
275 static struct malloc_pipe tcptemp_mpipe;
276 
277 static void tcp_willblock(void);
278 static void tcp_notify (struct inpcb *, int);
279 
280 struct tcp_stats tcpstats_percpu[MAXCPU];
281 
282 static int
283 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
284 {
285 	int cpu, error = 0;
286 
287 	for (cpu = 0; cpu < ncpus; ++cpu) {
288 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
289 					sizeof(struct tcp_stats))))
290 			break;
291 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
292 				       sizeof(struct tcp_stats))))
293 			break;
294 	}
295 
296 	return (error);
297 }
298 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
299     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
300 
301 /*
302  * Target size of TCP PCB hash tables. Must be a power of two.
303  *
304  * Note that this can be overridden by the kernel environment
305  * variable net.inet.tcp.tcbhashsize
306  */
307 #ifndef TCBHASHSIZE
308 #define	TCBHASHSIZE	512
309 #endif
310 
311 /*
312  * This is the actual shape of what we allocate using the zone
313  * allocator.  Doing it this way allows us to protect both structures
314  * using the same generation count, and also eliminates the overhead
315  * of allocating tcpcbs separately.  By hiding the structure here,
316  * we avoid changing most of the rest of the code (although it needs
317  * to be changed, eventually, for greater efficiency).
318  */
319 #define	ALIGNMENT	32
320 #define	ALIGNM1		(ALIGNMENT - 1)
321 struct	inp_tp {
322 	union {
323 		struct	inpcb inp;
324 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
325 	} inp_tp_u;
326 	struct	tcpcb tcb;
327 	struct	tcp_callout inp_tp_rexmt;
328 	struct	tcp_callout inp_tp_persist;
329 	struct	tcp_callout inp_tp_keep;
330 	struct	tcp_callout inp_tp_2msl;
331 	struct	tcp_callout inp_tp_delack;
332 	struct	netmsg_tcp_timer inp_tp_timermsg;
333 };
334 #undef ALIGNMENT
335 #undef ALIGNM1
336 
337 /*
338  * Tcp initialization
339  */
340 void
341 tcp_init(void)
342 {
343 	struct inpcbporthead *porthashbase;
344 	struct inpcbinfo *ticb;
345 	u_long porthashmask;
346 	int hashsize = TCBHASHSIZE;
347 	int cpu;
348 
349 	/*
350 	 * note: tcptemp is used for keepalives, and it is ok for an
351 	 * allocation to fail so do not specify MPF_INT.
352 	 */
353 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
354 		    25, -1, 0, NULL, NULL, NULL);
355 
356 	tcp_delacktime = TCPTV_DELACK;
357 	tcp_keepinit = TCPTV_KEEP_INIT;
358 	tcp_keepidle = TCPTV_KEEP_IDLE;
359 	tcp_keepintvl = TCPTV_KEEPINTVL;
360 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
361 	tcp_msl = TCPTV_MSL;
362 	tcp_rexmit_min = TCPTV_MIN;
363 	tcp_rexmit_slop = TCPTV_CPU_VAR;
364 
365 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
366 	if (!powerof2(hashsize)) {
367 		kprintf("WARNING: TCB hash size not a power of 2\n");
368 		hashsize = 512; /* safe default */
369 	}
370 	tcp_tcbhashsize = hashsize;
371 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
372 
373 	for (cpu = 0; cpu < ncpus2; cpu++) {
374 		ticb = &tcbinfo[cpu];
375 		in_pcbinfo_init(ticb);
376 		ticb->cpu = cpu;
377 		ticb->hashbase = hashinit(hashsize, M_PCB,
378 					  &ticb->hashmask);
379 		ticb->porthashbase = porthashbase;
380 		ticb->porthashmask = porthashmask;
381 		ticb->porttoken = &tcp_port_token;
382 #if 0
383 		ticb->porthashbase = hashinit(hashsize, M_PCB,
384 					      &ticb->porthashmask);
385 #endif
386 		ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
387 						  &ticb->wildcardhashmask);
388 		ticb->ipi_size = sizeof(struct inp_tp);
389 		TAILQ_INIT(&tcpcbackq[cpu]);
390 	}
391 
392 	tcp_reass_maxseg = nmbclusters / 16;
393 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
394 
395 #ifdef INET6
396 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
397 #else
398 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
399 #endif
400 	if (max_protohdr < TCP_MINPROTOHDR)
401 		max_protohdr = TCP_MINPROTOHDR;
402 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
403 		panic("tcp_init");
404 #undef TCP_MINPROTOHDR
405 
406 	/*
407 	 * Initialize TCP statistics counters for each CPU.
408 	 */
409 	for (cpu = 0; cpu < ncpus; ++cpu) {
410 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
411 	}
412 
413 	syncache_init();
414 	netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP);
415 }
416 
417 static void
418 tcp_willblock(void)
419 {
420 	struct tcpcb *tp;
421 	int cpu = mycpu->gd_cpuid;
422 
423 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
424 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
425 		tp->t_flags &= ~TF_ONOUTPUTQ;
426 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
427 		tcp_output(tp);
428 	}
429 }
430 
431 /*
432  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
433  * tcp_template used to store this data in mbufs, but we now recopy it out
434  * of the tcpcb each time to conserve mbufs.
435  */
436 void
437 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso)
438 {
439 	struct inpcb *inp = tp->t_inpcb;
440 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
441 
442 #ifdef INET6
443 	if (inp->inp_vflag & INP_IPV6) {
444 		struct ip6_hdr *ip6;
445 
446 		ip6 = (struct ip6_hdr *)ip_ptr;
447 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
448 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
449 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
450 			(IPV6_VERSION & IPV6_VERSION_MASK);
451 		ip6->ip6_nxt = IPPROTO_TCP;
452 		ip6->ip6_plen = sizeof(struct tcphdr);
453 		ip6->ip6_src = inp->in6p_laddr;
454 		ip6->ip6_dst = inp->in6p_faddr;
455 		tcp_hdr->th_sum = 0;
456 	} else
457 #endif
458 	{
459 		struct ip *ip = (struct ip *) ip_ptr;
460 		u_int plen;
461 
462 		ip->ip_vhl = IP_VHL_BORING;
463 		ip->ip_tos = 0;
464 		ip->ip_len = 0;
465 		ip->ip_id = 0;
466 		ip->ip_off = 0;
467 		ip->ip_ttl = 0;
468 		ip->ip_sum = 0;
469 		ip->ip_p = IPPROTO_TCP;
470 		ip->ip_src = inp->inp_laddr;
471 		ip->ip_dst = inp->inp_faddr;
472 
473 		if (tso)
474 			plen = htons(IPPROTO_TCP);
475 		else
476 			plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP);
477 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
478 		    ip->ip_dst.s_addr, plen);
479 	}
480 
481 	tcp_hdr->th_sport = inp->inp_lport;
482 	tcp_hdr->th_dport = inp->inp_fport;
483 	tcp_hdr->th_seq = 0;
484 	tcp_hdr->th_ack = 0;
485 	tcp_hdr->th_x2 = 0;
486 	tcp_hdr->th_off = 5;
487 	tcp_hdr->th_flags = 0;
488 	tcp_hdr->th_win = 0;
489 	tcp_hdr->th_urp = 0;
490 }
491 
492 /*
493  * Create template to be used to send tcp packets on a connection.
494  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
495  * use for this function is in keepalives, which use tcp_respond.
496  */
497 struct tcptemp *
498 tcp_maketemplate(struct tcpcb *tp)
499 {
500 	struct tcptemp *tmp;
501 
502 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
503 		return (NULL);
504 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE);
505 	return (tmp);
506 }
507 
508 void
509 tcp_freetemplate(struct tcptemp *tmp)
510 {
511 	mpipe_free(&tcptemp_mpipe, tmp);
512 }
513 
514 /*
515  * Send a single message to the TCP at address specified by
516  * the given TCP/IP header.  If m == NULL, then we make a copy
517  * of the tcpiphdr at ti and send directly to the addressed host.
518  * This is used to force keep alive messages out using the TCP
519  * template for a connection.  If flags are given then we send
520  * a message back to the TCP which originated the * segment ti,
521  * and discard the mbuf containing it and any other attached mbufs.
522  *
523  * In any case the ack and sequence number of the transmitted
524  * segment are as specified by the parameters.
525  *
526  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
527  */
528 void
529 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
530 	    tcp_seq ack, tcp_seq seq, int flags)
531 {
532 	int tlen;
533 	int win = 0;
534 	struct route *ro = NULL;
535 	struct route sro;
536 	struct ip *ip = ipgen;
537 	struct tcphdr *nth;
538 	int ipflags = 0;
539 	struct route_in6 *ro6 = NULL;
540 	struct route_in6 sro6;
541 	struct ip6_hdr *ip6 = ipgen;
542 	boolean_t use_tmpro = TRUE;
543 #ifdef INET6
544 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
545 #else
546 	const boolean_t isipv6 = FALSE;
547 #endif
548 
549 	if (tp != NULL) {
550 		if (!(flags & TH_RST)) {
551 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
552 			if (win < 0)
553 				win = 0;
554 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
555 				win = (long)TCP_MAXWIN << tp->rcv_scale;
556 		}
557 		/*
558 		 * Don't use the route cache of a listen socket,
559 		 * it is not MPSAFE; use temporary route cache.
560 		 */
561 		if (tp->t_state != TCPS_LISTEN) {
562 			if (isipv6)
563 				ro6 = &tp->t_inpcb->in6p_route;
564 			else
565 				ro = &tp->t_inpcb->inp_route;
566 			use_tmpro = FALSE;
567 		}
568 	}
569 	if (use_tmpro) {
570 		if (isipv6) {
571 			ro6 = &sro6;
572 			bzero(ro6, sizeof *ro6);
573 		} else {
574 			ro = &sro;
575 			bzero(ro, sizeof *ro);
576 		}
577 	}
578 	if (m == NULL) {
579 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
580 		if (m == NULL)
581 			return;
582 		tlen = 0;
583 		m->m_data += max_linkhdr;
584 		if (isipv6) {
585 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
586 			ip6 = mtod(m, struct ip6_hdr *);
587 			nth = (struct tcphdr *)(ip6 + 1);
588 		} else {
589 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
590 			ip = mtod(m, struct ip *);
591 			nth = (struct tcphdr *)(ip + 1);
592 		}
593 		bcopy(th, nth, sizeof(struct tcphdr));
594 		flags = TH_ACK;
595 	} else {
596 		m_freem(m->m_next);
597 		m->m_next = NULL;
598 		m->m_data = (caddr_t)ipgen;
599 		/* m_len is set later */
600 		tlen = 0;
601 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
602 		if (isipv6) {
603 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
604 			nth = (struct tcphdr *)(ip6 + 1);
605 		} else {
606 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
607 			nth = (struct tcphdr *)(ip + 1);
608 		}
609 		if (th != nth) {
610 			/*
611 			 * this is usually a case when an extension header
612 			 * exists between the IPv6 header and the
613 			 * TCP header.
614 			 */
615 			nth->th_sport = th->th_sport;
616 			nth->th_dport = th->th_dport;
617 		}
618 		xchg(nth->th_dport, nth->th_sport, n_short);
619 #undef xchg
620 	}
621 	if (isipv6) {
622 		ip6->ip6_flow = 0;
623 		ip6->ip6_vfc = IPV6_VERSION;
624 		ip6->ip6_nxt = IPPROTO_TCP;
625 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
626 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
627 	} else {
628 		tlen += sizeof(struct tcpiphdr);
629 		ip->ip_len = tlen;
630 		ip->ip_ttl = ip_defttl;
631 	}
632 	m->m_len = tlen;
633 	m->m_pkthdr.len = tlen;
634 	m->m_pkthdr.rcvif = NULL;
635 	nth->th_seq = htonl(seq);
636 	nth->th_ack = htonl(ack);
637 	nth->th_x2 = 0;
638 	nth->th_off = sizeof(struct tcphdr) >> 2;
639 	nth->th_flags = flags;
640 	if (tp != NULL)
641 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
642 	else
643 		nth->th_win = htons((u_short)win);
644 	nth->th_urp = 0;
645 	if (isipv6) {
646 		nth->th_sum = 0;
647 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
648 					sizeof(struct ip6_hdr),
649 					tlen - sizeof(struct ip6_hdr));
650 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
651 					       (ro6 && ro6->ro_rt) ?
652 						ro6->ro_rt->rt_ifp : NULL);
653 	} else {
654 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
655 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
656 		m->m_pkthdr.csum_flags = CSUM_TCP;
657 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
658 		m->m_pkthdr.csum_thlen = sizeof(struct tcphdr);
659 	}
660 #ifdef TCPDEBUG
661 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
662 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
663 #endif
664 	if (isipv6) {
665 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
666 			   tp ? tp->t_inpcb : NULL);
667 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
668 			RTFREE(ro6->ro_rt);
669 			ro6->ro_rt = NULL;
670 		}
671 	} else {
672 		ipflags |= IP_DEBUGROUTE;
673 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
674 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
675 			RTFREE(ro->ro_rt);
676 			ro->ro_rt = NULL;
677 		}
678 	}
679 }
680 
681 /*
682  * Create a new TCP control block, making an
683  * empty reassembly queue and hooking it to the argument
684  * protocol control block.  The `inp' parameter must have
685  * come from the zone allocator set up in tcp_init().
686  */
687 struct tcpcb *
688 tcp_newtcpcb(struct inpcb *inp)
689 {
690 	struct inp_tp *it;
691 	struct tcpcb *tp;
692 #ifdef INET6
693 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
694 #else
695 	const boolean_t isipv6 = FALSE;
696 #endif
697 
698 	it = (struct inp_tp *)inp;
699 	tp = &it->tcb;
700 	bzero(tp, sizeof(struct tcpcb));
701 	TAILQ_INIT(&tp->t_segq);
702 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
703 	tp->t_rxtthresh = tcprexmtthresh;
704 
705 	/* Set up our timeouts. */
706 	tp->tt_rexmt = &it->inp_tp_rexmt;
707 	tp->tt_persist = &it->inp_tp_persist;
708 	tp->tt_keep = &it->inp_tp_keep;
709 	tp->tt_2msl = &it->inp_tp_2msl;
710 	tp->tt_delack = &it->inp_tp_delack;
711 	tcp_inittimers(tp);
712 
713 	/*
714 	 * Zero out timer message.  We don't create it here,
715 	 * since the current CPU may not be the owner of this
716 	 * inpcb.
717 	 */
718 	tp->tt_msg = &it->inp_tp_timermsg;
719 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
720 
721 	tp->t_keepinit = tcp_keepinit;
722 	tp->t_keepidle = tcp_keepidle;
723 	tp->t_keepintvl = tcp_keepintvl;
724 	tp->t_keepcnt = tcp_keepcnt;
725 	tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
726 
727 	if (tcp_do_ncr)
728 		tp->t_flags |= TF_NCR;
729 	if (tcp_do_rfc1323)
730 		tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP);
731 
732 	tp->t_inpcb = inp;	/* XXX */
733 	tp->t_state = TCPS_CLOSED;
734 	/*
735 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
736 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
737 	 * reasonable initial retransmit time.
738 	 */
739 	tp->t_srtt = TCPTV_SRTTBASE;
740 	tp->t_rttvar =
741 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
742 	tp->t_rttmin = tcp_rexmit_min;
743 	tp->t_rxtcur = TCPTV_RTOBASE;
744 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
745 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
746 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
747 	tp->snd_last = ticks;
748 	tp->t_rcvtime = ticks;
749 	/*
750 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
751 	 * because the socket may be bound to an IPv6 wildcard address,
752 	 * which may match an IPv4-mapped IPv6 address.
753 	 */
754 	inp->inp_ip_ttl = ip_defttl;
755 	inp->inp_ppcb = tp;
756 	tcp_sack_tcpcb_init(tp);
757 	return (tp);		/* XXX */
758 }
759 
760 /*
761  * Drop a TCP connection, reporting the specified error.
762  * If connection is synchronized, then send a RST to peer.
763  */
764 struct tcpcb *
765 tcp_drop(struct tcpcb *tp, int error)
766 {
767 	struct socket *so = tp->t_inpcb->inp_socket;
768 
769 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
770 		tp->t_state = TCPS_CLOSED;
771 		tcp_output(tp);
772 		tcpstat.tcps_drops++;
773 	} else
774 		tcpstat.tcps_conndrops++;
775 	if (error == ETIMEDOUT && tp->t_softerror)
776 		error = tp->t_softerror;
777 	so->so_error = error;
778 	return (tcp_close(tp));
779 }
780 
781 struct netmsg_listen_detach {
782 	struct netmsg_base	base;
783 	struct tcpcb		*nm_tp;
784 };
785 
786 static void
787 tcp_listen_detach_handler(netmsg_t msg)
788 {
789 	struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
790 	struct tcpcb *tp = nmsg->nm_tp;
791 	int cpu = mycpuid, nextcpu;
792 
793 	if (tp->t_flags & TF_LISTEN)
794 		syncache_destroy(tp);
795 
796 	in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
797 
798 	nextcpu = cpu + 1;
799 	if (nextcpu < ncpus2)
800 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nmsg->base.lmsg);
801 	else
802 		lwkt_replymsg(&nmsg->base.lmsg, 0);
803 }
804 
805 /*
806  * Close a TCP control block:
807  *	discard all space held by the tcp
808  *	discard internet protocol block
809  *	wake up any sleepers
810  */
811 struct tcpcb *
812 tcp_close(struct tcpcb *tp)
813 {
814 	struct tseg_qent *q;
815 	struct inpcb *inp = tp->t_inpcb;
816 	struct socket *so = inp->inp_socket;
817 	struct rtentry *rt;
818 	boolean_t dosavessthresh;
819 #ifdef INET6
820 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
821 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
822 #else
823 	const boolean_t isipv6 = FALSE;
824 #endif
825 
826 	/*
827 	 * INP_WILDCARD_MP indicates that listen(2) has been called on
828 	 * this socket.  This implies:
829 	 * - A wildcard inp's hash is replicated for each protocol thread.
830 	 * - Syncache for this inp grows independently in each protocol
831 	 *   thread.
832 	 * - There is more than one cpu
833 	 *
834 	 * We have to chain a message to the rest of the protocol threads
835 	 * to cleanup the wildcard hash and the syncache.  The cleanup
836 	 * in the current protocol thread is defered till the end of this
837 	 * function.
838 	 *
839 	 * NOTE:
840 	 * After cleanup the inp's hash and syncache entries, this inp will
841 	 * no longer be available to the rest of the protocol threads, so we
842 	 * are safe to whack the inp in the following code.
843 	 */
844 	if (inp->inp_flags & INP_WILDCARD_MP) {
845 		struct netmsg_listen_detach nmsg;
846 
847 		KKASSERT(so->so_port == netisr_portfn(0));
848 		KKASSERT(&curthread->td_msgport == netisr_portfn(0));
849 		KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
850 
851 		netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
852 			    MSGF_PRIORITY, tcp_listen_detach_handler);
853 		nmsg.nm_tp = tp;
854 		lwkt_domsg(netisr_portfn(1), &nmsg.base.lmsg, 0);
855 
856 		inp->inp_flags &= ~INP_WILDCARD_MP;
857 	}
858 
859 	KKASSERT(tp->t_state != TCPS_TERMINATING);
860 	tp->t_state = TCPS_TERMINATING;
861 
862 	/*
863 	 * Make sure that all of our timers are stopped before we
864 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
865 	 * timers are never used.  If timer message is never created
866 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
867 	 */
868 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
869 		tcp_callout_stop(tp, tp->tt_rexmt);
870 		tcp_callout_stop(tp, tp->tt_persist);
871 		tcp_callout_stop(tp, tp->tt_keep);
872 		tcp_callout_stop(tp, tp->tt_2msl);
873 		tcp_callout_stop(tp, tp->tt_delack);
874 	}
875 
876 	if (tp->t_flags & TF_ONOUTPUTQ) {
877 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
878 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
879 		tp->t_flags &= ~TF_ONOUTPUTQ;
880 	}
881 
882 	/*
883 	 * If we got enough samples through the srtt filter,
884 	 * save the rtt and rttvar in the routing entry.
885 	 * 'Enough' is arbitrarily defined as the 16 samples.
886 	 * 16 samples is enough for the srtt filter to converge
887 	 * to within 5% of the correct value; fewer samples and
888 	 * we could save a very bogus rtt.
889 	 *
890 	 * Don't update the default route's characteristics and don't
891 	 * update anything that the user "locked".
892 	 */
893 	if (tp->t_rttupdated >= 16) {
894 		u_long i = 0;
895 
896 		if (isipv6) {
897 			struct sockaddr_in6 *sin6;
898 
899 			if ((rt = inp->in6p_route.ro_rt) == NULL)
900 				goto no_valid_rt;
901 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
902 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
903 				goto no_valid_rt;
904 		} else
905 			if ((rt = inp->inp_route.ro_rt) == NULL ||
906 			    ((struct sockaddr_in *)rt_key(rt))->
907 			     sin_addr.s_addr == INADDR_ANY)
908 				goto no_valid_rt;
909 
910 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
911 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
912 			if (rt->rt_rmx.rmx_rtt && i)
913 				/*
914 				 * filter this update to half the old & half
915 				 * the new values, converting scale.
916 				 * See route.h and tcp_var.h for a
917 				 * description of the scaling constants.
918 				 */
919 				rt->rt_rmx.rmx_rtt =
920 				    (rt->rt_rmx.rmx_rtt + i) / 2;
921 			else
922 				rt->rt_rmx.rmx_rtt = i;
923 			tcpstat.tcps_cachedrtt++;
924 		}
925 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
926 			i = tp->t_rttvar *
927 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
928 			if (rt->rt_rmx.rmx_rttvar && i)
929 				rt->rt_rmx.rmx_rttvar =
930 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
931 			else
932 				rt->rt_rmx.rmx_rttvar = i;
933 			tcpstat.tcps_cachedrttvar++;
934 		}
935 		/*
936 		 * The old comment here said:
937 		 * update the pipelimit (ssthresh) if it has been updated
938 		 * already or if a pipesize was specified & the threshhold
939 		 * got below half the pipesize.  I.e., wait for bad news
940 		 * before we start updating, then update on both good
941 		 * and bad news.
942 		 *
943 		 * But we want to save the ssthresh even if no pipesize is
944 		 * specified explicitly in the route, because such
945 		 * connections still have an implicit pipesize specified
946 		 * by the global tcp_sendspace.  In the absence of a reliable
947 		 * way to calculate the pipesize, it will have to do.
948 		 */
949 		i = tp->snd_ssthresh;
950 		if (rt->rt_rmx.rmx_sendpipe != 0)
951 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
952 		else
953 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
954 		if (dosavessthresh ||
955 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
956 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
957 			/*
958 			 * convert the limit from user data bytes to
959 			 * packets then to packet data bytes.
960 			 */
961 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
962 			if (i < 2)
963 				i = 2;
964 			i *= tp->t_maxseg +
965 			     (isipv6 ?
966 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
967 			      sizeof(struct tcpiphdr));
968 			if (rt->rt_rmx.rmx_ssthresh)
969 				rt->rt_rmx.rmx_ssthresh =
970 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
971 			else
972 				rt->rt_rmx.rmx_ssthresh = i;
973 			tcpstat.tcps_cachedssthresh++;
974 		}
975 	}
976 
977 no_valid_rt:
978 	/* free the reassembly queue, if any */
979 	while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) {
980 		TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
981 		m_freem(q->tqe_m);
982 		kfree(q, M_TSEGQ);
983 		atomic_add_int(&tcp_reass_qsize, -1);
984 	}
985 	/* throw away SACK blocks in scoreboard*/
986 	if (TCP_DO_SACK(tp))
987 		tcp_sack_destroy(&tp->scb);
988 
989 	inp->inp_ppcb = NULL;
990 	soisdisconnected(so);
991 	/* note: pcb detached later on */
992 
993 	tcp_destroy_timermsg(tp);
994 
995 	if (tp->t_flags & TF_LISTEN)
996 		syncache_destroy(tp);
997 
998 	so_async_rcvd_drop(so);
999 
1000 	/*
1001 	 * NOTE:
1002 	 * pcbdetach removes any wildcard hash entry on the current CPU.
1003 	 */
1004 #ifdef INET6
1005 	if (isafinet6)
1006 		in6_pcbdetach(inp);
1007 	else
1008 #endif
1009 		in_pcbdetach(inp);
1010 
1011 	tcpstat.tcps_closed++;
1012 	return (NULL);
1013 }
1014 
1015 static __inline void
1016 tcp_drain_oncpu(struct inpcbhead *head)
1017 {
1018 	struct inpcb *marker;
1019 	struct inpcb *inpb;
1020 	struct tcpcb *tcpb;
1021 	struct tseg_qent *te;
1022 
1023 	/*
1024 	 * Allows us to block while running the list
1025 	 */
1026 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1027 	marker->inp_flags |= INP_PLACEMARKER;
1028 	LIST_INSERT_HEAD(head, marker, inp_list);
1029 
1030 	while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1031 		if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1032 		    (tcpb = intotcpcb(inpb)) != NULL &&
1033 		    (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) {
1034 			TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q);
1035 			if (te->tqe_th->th_flags & TH_FIN)
1036 				tcpb->t_flags &= ~TF_QUEDFIN;
1037 			m_freem(te->tqe_m);
1038 			kfree(te, M_TSEGQ);
1039 			atomic_add_int(&tcp_reass_qsize, -1);
1040 			/* retry */
1041 		} else {
1042 			LIST_REMOVE(marker, inp_list);
1043 			LIST_INSERT_AFTER(inpb, marker, inp_list);
1044 		}
1045 	}
1046 	LIST_REMOVE(marker, inp_list);
1047 	kfree(marker, M_TEMP);
1048 }
1049 
1050 struct netmsg_tcp_drain {
1051 	struct netmsg_base	base;
1052 	struct inpcbhead	*nm_head;
1053 };
1054 
1055 static void
1056 tcp_drain_handler(netmsg_t msg)
1057 {
1058 	struct netmsg_tcp_drain *nm = (void *)msg;
1059 
1060 	tcp_drain_oncpu(nm->nm_head);
1061 	lwkt_replymsg(&nm->base.lmsg, 0);
1062 }
1063 
1064 void
1065 tcp_drain(void)
1066 {
1067 	int cpu;
1068 
1069 	if (!do_tcpdrain)
1070 		return;
1071 
1072 	/*
1073 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1074 	 * if there is one...
1075 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1076 	 *	reassembly queue should be flushed, but in a situation
1077 	 *	where we're really low on mbufs, this is potentially
1078 	 *	useful.
1079 	 */
1080 	for (cpu = 0; cpu < ncpus2; cpu++) {
1081 		struct netmsg_tcp_drain *nm;
1082 
1083 		if (cpu == mycpu->gd_cpuid) {
1084 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1085 		} else {
1086 			nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1087 				     M_LWKTMSG, M_NOWAIT);
1088 			if (nm == NULL)
1089 				continue;
1090 			netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1091 				    0, tcp_drain_handler);
1092 			nm->nm_head = &tcbinfo[cpu].pcblisthead;
1093 			lwkt_sendmsg(netisr_portfn(cpu), &nm->base.lmsg);
1094 		}
1095 	}
1096 }
1097 
1098 /*
1099  * Notify a tcp user of an asynchronous error;
1100  * store error as soft error, but wake up user
1101  * (for now, won't do anything until can select for soft error).
1102  *
1103  * Do not wake up user since there currently is no mechanism for
1104  * reporting soft errors (yet - a kqueue filter may be added).
1105  */
1106 static void
1107 tcp_notify(struct inpcb *inp, int error)
1108 {
1109 	struct tcpcb *tp = intotcpcb(inp);
1110 
1111 	/*
1112 	 * Ignore some errors if we are hooked up.
1113 	 * If connection hasn't completed, has retransmitted several times,
1114 	 * and receives a second error, give up now.  This is better
1115 	 * than waiting a long time to establish a connection that
1116 	 * can never complete.
1117 	 */
1118 	if (tp->t_state == TCPS_ESTABLISHED &&
1119 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1120 	      error == EHOSTDOWN)) {
1121 		return;
1122 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1123 	    tp->t_softerror)
1124 		tcp_drop(tp, error);
1125 	else
1126 		tp->t_softerror = error;
1127 #if 0
1128 	wakeup(&so->so_timeo);
1129 	sorwakeup(so);
1130 	sowwakeup(so);
1131 #endif
1132 }
1133 
1134 static int
1135 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1136 {
1137 	int error, i, n;
1138 	struct inpcb *marker;
1139 	struct inpcb *inp;
1140 	globaldata_t gd;
1141 	int origcpu, ccpu;
1142 
1143 	error = 0;
1144 	n = 0;
1145 
1146 	/*
1147 	 * The process of preparing the TCB list is too time-consuming and
1148 	 * resource-intensive to repeat twice on every request.
1149 	 */
1150 	if (req->oldptr == NULL) {
1151 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1152 			gd = globaldata_find(ccpu);
1153 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1154 		}
1155 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1156 		return (0);
1157 	}
1158 
1159 	if (req->newptr != NULL)
1160 		return (EPERM);
1161 
1162 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1163 	marker->inp_flags |= INP_PLACEMARKER;
1164 
1165 	/*
1166 	 * OK, now we're committed to doing something.  Run the inpcb list
1167 	 * for each cpu in the system and construct the output.  Use a
1168 	 * list placemarker to deal with list changes occuring during
1169 	 * copyout blockages (but otherwise depend on being on the correct
1170 	 * cpu to avoid races).
1171 	 */
1172 	origcpu = mycpu->gd_cpuid;
1173 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1174 		globaldata_t rgd;
1175 		caddr_t inp_ppcb;
1176 		struct xtcpcb xt;
1177 		int cpu_id;
1178 
1179 		cpu_id = (origcpu + ccpu) % ncpus;
1180 		if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1181 			continue;
1182 		rgd = globaldata_find(cpu_id);
1183 		lwkt_setcpu_self(rgd);
1184 
1185 		n = tcbinfo[cpu_id].ipi_count;
1186 
1187 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1188 		i = 0;
1189 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1190 			/*
1191 			 * process a snapshot of pcbs, ignoring placemarkers
1192 			 * and using our own to allow SYSCTL_OUT to block.
1193 			 */
1194 			LIST_REMOVE(marker, inp_list);
1195 			LIST_INSERT_AFTER(inp, marker, inp_list);
1196 
1197 			if (inp->inp_flags & INP_PLACEMARKER)
1198 				continue;
1199 			if (prison_xinpcb(req->td, inp))
1200 				continue;
1201 
1202 			xt.xt_len = sizeof xt;
1203 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1204 			inp_ppcb = inp->inp_ppcb;
1205 			if (inp_ppcb != NULL)
1206 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1207 			else
1208 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1209 			if (inp->inp_socket)
1210 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1211 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1212 				break;
1213 			++i;
1214 		}
1215 		LIST_REMOVE(marker, inp_list);
1216 		if (error == 0 && i < n) {
1217 			bzero(&xt, sizeof xt);
1218 			xt.xt_len = sizeof xt;
1219 			while (i < n) {
1220 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1221 				if (error)
1222 					break;
1223 				++i;
1224 			}
1225 		}
1226 	}
1227 
1228 	/*
1229 	 * Make sure we are on the same cpu we were on originally, since
1230 	 * higher level callers expect this.  Also don't pollute caches with
1231 	 * migrated userland data by (eventually) returning to userland
1232 	 * on a different cpu.
1233 	 */
1234 	lwkt_setcpu_self(globaldata_find(origcpu));
1235 	kfree(marker, M_TEMP);
1236 	return (error);
1237 }
1238 
1239 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1240 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1241 
1242 static int
1243 tcp_getcred(SYSCTL_HANDLER_ARGS)
1244 {
1245 	struct sockaddr_in addrs[2];
1246 	struct inpcb *inp;
1247 	int cpu;
1248 	int error;
1249 
1250 	error = priv_check(req->td, PRIV_ROOT);
1251 	if (error != 0)
1252 		return (error);
1253 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1254 	if (error != 0)
1255 		return (error);
1256 	crit_enter();
1257 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1258 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1259 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1260 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1261 	if (inp == NULL || inp->inp_socket == NULL) {
1262 		error = ENOENT;
1263 		goto out;
1264 	}
1265 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1266 out:
1267 	crit_exit();
1268 	return (error);
1269 }
1270 
1271 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1272     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1273 
1274 #ifdef INET6
1275 static int
1276 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1277 {
1278 	struct sockaddr_in6 addrs[2];
1279 	struct inpcb *inp;
1280 	int error;
1281 	boolean_t mapped = FALSE;
1282 
1283 	error = priv_check(req->td, PRIV_ROOT);
1284 	if (error != 0)
1285 		return (error);
1286 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1287 	if (error != 0)
1288 		return (error);
1289 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1290 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1291 			mapped = TRUE;
1292 		else
1293 			return (EINVAL);
1294 	}
1295 	crit_enter();
1296 	if (mapped) {
1297 		inp = in_pcblookup_hash(&tcbinfo[0],
1298 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1299 		    addrs[1].sin6_port,
1300 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1301 		    addrs[0].sin6_port,
1302 		    0, NULL);
1303 	} else {
1304 		inp = in6_pcblookup_hash(&tcbinfo[0],
1305 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1306 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1307 		    0, NULL);
1308 	}
1309 	if (inp == NULL || inp->inp_socket == NULL) {
1310 		error = ENOENT;
1311 		goto out;
1312 	}
1313 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1314 out:
1315 	crit_exit();
1316 	return (error);
1317 }
1318 
1319 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1320 	    0, 0,
1321 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1322 #endif
1323 
1324 struct netmsg_tcp_notify {
1325 	struct netmsg_base base;
1326 	void		(*nm_notify)(struct inpcb *, int);
1327 	struct in_addr	nm_faddr;
1328 	int		nm_arg;
1329 };
1330 
1331 static void
1332 tcp_notifyall_oncpu(netmsg_t msg)
1333 {
1334 	struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1335 	int nextcpu;
1336 
1337 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1338 			nm->nm_arg, nm->nm_notify);
1339 
1340 	nextcpu = mycpuid + 1;
1341 	if (nextcpu < ncpus2)
1342 		lwkt_forwardmsg(netisr_portfn(nextcpu), &nm->base.lmsg);
1343 	else
1344 		lwkt_replymsg(&nm->base.lmsg, 0);
1345 }
1346 
1347 void
1348 tcp_ctlinput(netmsg_t msg)
1349 {
1350 	int cmd = msg->ctlinput.nm_cmd;
1351 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1352 	struct ip *ip = msg->ctlinput.nm_extra;
1353 	struct tcphdr *th;
1354 	struct in_addr faddr;
1355 	struct inpcb *inp;
1356 	struct tcpcb *tp;
1357 	void (*notify)(struct inpcb *, int) = tcp_notify;
1358 	tcp_seq icmpseq;
1359 	int arg, cpu;
1360 
1361 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1362 		goto done;
1363 	}
1364 
1365 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1366 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1367 		goto done;
1368 
1369 	arg = inetctlerrmap[cmd];
1370 	if (cmd == PRC_QUENCH) {
1371 		notify = tcp_quench;
1372 	} else if (icmp_may_rst &&
1373 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1374 		    cmd == PRC_UNREACH_PORT ||
1375 		    cmd == PRC_TIMXCEED_INTRANS) &&
1376 		   ip != NULL) {
1377 		notify = tcp_drop_syn_sent;
1378 	} else if (cmd == PRC_MSGSIZE) {
1379 		struct icmp *icmp = (struct icmp *)
1380 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1381 
1382 		arg = ntohs(icmp->icmp_nextmtu);
1383 		notify = tcp_mtudisc;
1384 	} else if (PRC_IS_REDIRECT(cmd)) {
1385 		ip = NULL;
1386 		notify = in_rtchange;
1387 	} else if (cmd == PRC_HOSTDEAD) {
1388 		ip = NULL;
1389 	}
1390 
1391 	if (ip != NULL) {
1392 		crit_enter();
1393 		th = (struct tcphdr *)((caddr_t)ip +
1394 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1395 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1396 				  ip->ip_src.s_addr, th->th_sport);
1397 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1398 					ip->ip_src, th->th_sport, 0, NULL);
1399 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1400 			icmpseq = htonl(th->th_seq);
1401 			tp = intotcpcb(inp);
1402 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1403 			    SEQ_LT(icmpseq, tp->snd_max))
1404 				(*notify)(inp, arg);
1405 		} else {
1406 			struct in_conninfo inc;
1407 
1408 			inc.inc_fport = th->th_dport;
1409 			inc.inc_lport = th->th_sport;
1410 			inc.inc_faddr = faddr;
1411 			inc.inc_laddr = ip->ip_src;
1412 #ifdef INET6
1413 			inc.inc_isipv6 = 0;
1414 #endif
1415 			syncache_unreach(&inc, th);
1416 		}
1417 		crit_exit();
1418 	} else {
1419 		struct netmsg_tcp_notify *nm;
1420 
1421 		KKASSERT(&curthread->td_msgport == netisr_portfn(0));
1422 		nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1423 		netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1424 			    0, tcp_notifyall_oncpu);
1425 		nm->nm_faddr = faddr;
1426 		nm->nm_arg = arg;
1427 		nm->nm_notify = notify;
1428 
1429 		lwkt_sendmsg(netisr_portfn(0), &nm->base.lmsg);
1430 	}
1431 done:
1432 	lwkt_replymsg(&msg->lmsg, 0);
1433 }
1434 
1435 #ifdef INET6
1436 
1437 void
1438 tcp6_ctlinput(netmsg_t msg)
1439 {
1440 	int cmd = msg->ctlinput.nm_cmd;
1441 	struct sockaddr *sa = msg->ctlinput.nm_arg;
1442 	void *d = msg->ctlinput.nm_extra;
1443 	struct tcphdr th;
1444 	void (*notify) (struct inpcb *, int) = tcp_notify;
1445 	struct ip6_hdr *ip6;
1446 	struct mbuf *m;
1447 	struct ip6ctlparam *ip6cp = NULL;
1448 	const struct sockaddr_in6 *sa6_src = NULL;
1449 	int off;
1450 	struct tcp_portonly {
1451 		u_int16_t th_sport;
1452 		u_int16_t th_dport;
1453 	} *thp;
1454 	int arg;
1455 
1456 	if (sa->sa_family != AF_INET6 ||
1457 	    sa->sa_len != sizeof(struct sockaddr_in6)) {
1458 		goto out;
1459 	}
1460 
1461 	arg = 0;
1462 	if (cmd == PRC_QUENCH)
1463 		notify = tcp_quench;
1464 	else if (cmd == PRC_MSGSIZE) {
1465 		struct ip6ctlparam *ip6cp = d;
1466 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1467 
1468 		arg = ntohl(icmp6->icmp6_mtu);
1469 		notify = tcp_mtudisc;
1470 	} else if (!PRC_IS_REDIRECT(cmd) &&
1471 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1472 		goto out;
1473 	}
1474 
1475 	/* if the parameter is from icmp6, decode it. */
1476 	if (d != NULL) {
1477 		ip6cp = (struct ip6ctlparam *)d;
1478 		m = ip6cp->ip6c_m;
1479 		ip6 = ip6cp->ip6c_ip6;
1480 		off = ip6cp->ip6c_off;
1481 		sa6_src = ip6cp->ip6c_src;
1482 	} else {
1483 		m = NULL;
1484 		ip6 = NULL;
1485 		off = 0;	/* fool gcc */
1486 		sa6_src = &sa6_any;
1487 	}
1488 
1489 	if (ip6 != NULL) {
1490 		struct in_conninfo inc;
1491 		/*
1492 		 * XXX: We assume that when IPV6 is non NULL,
1493 		 * M and OFF are valid.
1494 		 */
1495 
1496 		/* check if we can safely examine src and dst ports */
1497 		if (m->m_pkthdr.len < off + sizeof *thp)
1498 			goto out;
1499 
1500 		bzero(&th, sizeof th);
1501 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1502 
1503 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1504 		    (struct sockaddr *)ip6cp->ip6c_src,
1505 		    th.th_sport, cmd, arg, notify);
1506 
1507 		inc.inc_fport = th.th_dport;
1508 		inc.inc_lport = th.th_sport;
1509 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1510 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1511 		inc.inc_isipv6 = 1;
1512 		syncache_unreach(&inc, &th);
1513 	} else {
1514 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1515 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1516 	}
1517 out:
1518 	lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1519 }
1520 
1521 #endif
1522 
1523 /*
1524  * Following is where TCP initial sequence number generation occurs.
1525  *
1526  * There are two places where we must use initial sequence numbers:
1527  * 1.  In SYN-ACK packets.
1528  * 2.  In SYN packets.
1529  *
1530  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1531  * tcp_syncache.c for details.
1532  *
1533  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1534  * depends on this property.  In addition, these ISNs should be
1535  * unguessable so as to prevent connection hijacking.  To satisfy
1536  * the requirements of this situation, the algorithm outlined in
1537  * RFC 1948 is used to generate sequence numbers.
1538  *
1539  * Implementation details:
1540  *
1541  * Time is based off the system timer, and is corrected so that it
1542  * increases by one megabyte per second.  This allows for proper
1543  * recycling on high speed LANs while still leaving over an hour
1544  * before rollover.
1545  *
1546  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1547  * between seeding of isn_secret.  This is normally set to zero,
1548  * as reseeding should not be necessary.
1549  *
1550  */
1551 
1552 #define	ISN_BYTES_PER_SECOND 1048576
1553 
1554 u_char isn_secret[32];
1555 int isn_last_reseed;
1556 MD5_CTX isn_ctx;
1557 
1558 tcp_seq
1559 tcp_new_isn(struct tcpcb *tp)
1560 {
1561 	u_int32_t md5_buffer[4];
1562 	tcp_seq new_isn;
1563 
1564 	/* Seed if this is the first use, reseed if requested. */
1565 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1566 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1567 		< (u_int)ticks))) {
1568 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1569 		isn_last_reseed = ticks;
1570 	}
1571 
1572 	/* Compute the md5 hash and return the ISN. */
1573 	MD5Init(&isn_ctx);
1574 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1575 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1576 #ifdef INET6
1577 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1578 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1579 			  sizeof(struct in6_addr));
1580 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1581 			  sizeof(struct in6_addr));
1582 	} else
1583 #endif
1584 	{
1585 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1586 			  sizeof(struct in_addr));
1587 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1588 			  sizeof(struct in_addr));
1589 	}
1590 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1591 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1592 	new_isn = (tcp_seq) md5_buffer[0];
1593 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1594 	return (new_isn);
1595 }
1596 
1597 /*
1598  * When a source quench is received, close congestion window
1599  * to one segment.  We will gradually open it again as we proceed.
1600  */
1601 void
1602 tcp_quench(struct inpcb *inp, int error)
1603 {
1604 	struct tcpcb *tp = intotcpcb(inp);
1605 
1606 	if (tp != NULL) {
1607 		tp->snd_cwnd = tp->t_maxseg;
1608 		tp->snd_wacked = 0;
1609 	}
1610 }
1611 
1612 /*
1613  * When a specific ICMP unreachable message is received and the
1614  * connection state is SYN-SENT, drop the connection.  This behavior
1615  * is controlled by the icmp_may_rst sysctl.
1616  */
1617 void
1618 tcp_drop_syn_sent(struct inpcb *inp, int error)
1619 {
1620 	struct tcpcb *tp = intotcpcb(inp);
1621 
1622 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1623 		tcp_drop(tp, error);
1624 }
1625 
1626 /*
1627  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1628  * based on the new value in the route.  Also nudge TCP to send something,
1629  * since we know the packet we just sent was dropped.
1630  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1631  */
1632 void
1633 tcp_mtudisc(struct inpcb *inp, int mtu)
1634 {
1635 	struct tcpcb *tp = intotcpcb(inp);
1636 	struct rtentry *rt;
1637 	struct socket *so = inp->inp_socket;
1638 	int maxopd, mss;
1639 #ifdef INET6
1640 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1641 #else
1642 	const boolean_t isipv6 = FALSE;
1643 #endif
1644 
1645 	if (tp == NULL)
1646 		return;
1647 
1648 	/*
1649 	 * If no MTU is provided in the ICMP message, use the
1650 	 * next lower likely value, as specified in RFC 1191.
1651 	 */
1652 	if (mtu == 0) {
1653 		int oldmtu;
1654 
1655 		oldmtu = tp->t_maxopd +
1656 		    (isipv6 ?
1657 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1658 		     sizeof(struct tcpiphdr));
1659 		mtu = ip_next_mtu(oldmtu, 0);
1660 	}
1661 
1662 	if (isipv6)
1663 		rt = tcp_rtlookup6(&inp->inp_inc);
1664 	else
1665 		rt = tcp_rtlookup(&inp->inp_inc);
1666 	if (rt != NULL) {
1667 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1668 			mtu = rt->rt_rmx.rmx_mtu;
1669 
1670 		maxopd = mtu -
1671 		    (isipv6 ?
1672 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1673 		     sizeof(struct tcpiphdr));
1674 
1675 		/*
1676 		 * XXX - The following conditional probably violates the TCP
1677 		 * spec.  The problem is that, since we don't know the
1678 		 * other end's MSS, we are supposed to use a conservative
1679 		 * default.  But, if we do that, then MTU discovery will
1680 		 * never actually take place, because the conservative
1681 		 * default is much less than the MTUs typically seen
1682 		 * on the Internet today.  For the moment, we'll sweep
1683 		 * this under the carpet.
1684 		 *
1685 		 * The conservative default might not actually be a problem
1686 		 * if the only case this occurs is when sending an initial
1687 		 * SYN with options and data to a host we've never talked
1688 		 * to before.  Then, they will reply with an MSS value which
1689 		 * will get recorded and the new parameters should get
1690 		 * recomputed.  For Further Study.
1691 		 */
1692 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1693 			maxopd = rt->rt_rmx.rmx_mssopt;
1694 	} else
1695 		maxopd = mtu -
1696 		    (isipv6 ?
1697 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1698 		     sizeof(struct tcpiphdr));
1699 
1700 	if (tp->t_maxopd <= maxopd)
1701 		return;
1702 	tp->t_maxopd = maxopd;
1703 
1704 	mss = maxopd;
1705 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1706 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1707 		mss -= TCPOLEN_TSTAMP_APPA;
1708 
1709 	/* round down to multiple of MCLBYTES */
1710 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1711 	if (mss > MCLBYTES)
1712 		mss &= ~(MCLBYTES - 1);
1713 #else
1714 	if (mss > MCLBYTES)
1715 		mss = (mss / MCLBYTES) * MCLBYTES;
1716 #endif
1717 
1718 	if (so->so_snd.ssb_hiwat < mss)
1719 		mss = so->so_snd.ssb_hiwat;
1720 
1721 	tp->t_maxseg = mss;
1722 	tp->t_rtttime = 0;
1723 	tp->snd_nxt = tp->snd_una;
1724 	tcp_output(tp);
1725 	tcpstat.tcps_mturesent++;
1726 }
1727 
1728 /*
1729  * Look-up the routing entry to the peer of this inpcb.  If no route
1730  * is found and it cannot be allocated the return NULL.  This routine
1731  * is called by TCP routines that access the rmx structure and by tcp_mss
1732  * to get the interface MTU.
1733  */
1734 struct rtentry *
1735 tcp_rtlookup(struct in_conninfo *inc)
1736 {
1737 	struct route *ro = &inc->inc_route;
1738 
1739 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1740 		/* No route yet, so try to acquire one */
1741 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1742 			/*
1743 			 * unused portions of the structure MUST be zero'd
1744 			 * out because rtalloc() treats it as opaque data
1745 			 */
1746 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1747 			ro->ro_dst.sa_family = AF_INET;
1748 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1749 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1750 			    inc->inc_faddr;
1751 			rtalloc(ro);
1752 		}
1753 	}
1754 	return (ro->ro_rt);
1755 }
1756 
1757 #ifdef INET6
1758 struct rtentry *
1759 tcp_rtlookup6(struct in_conninfo *inc)
1760 {
1761 	struct route_in6 *ro6 = &inc->inc6_route;
1762 
1763 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1764 		/* No route yet, so try to acquire one */
1765 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1766 			/*
1767 			 * unused portions of the structure MUST be zero'd
1768 			 * out because rtalloc() treats it as opaque data
1769 			 */
1770 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1771 			ro6->ro_dst.sin6_family = AF_INET6;
1772 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1773 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1774 			rtalloc((struct route *)ro6);
1775 		}
1776 	}
1777 	return (ro6->ro_rt);
1778 }
1779 #endif
1780 
1781 #ifdef IPSEC
1782 /* compute ESP/AH header size for TCP, including outer IP header. */
1783 size_t
1784 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1785 {
1786 	struct inpcb *inp;
1787 	struct mbuf *m;
1788 	size_t hdrsiz;
1789 	struct ip *ip;
1790 	struct tcphdr *th;
1791 
1792 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1793 		return (0);
1794 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1795 	if (!m)
1796 		return (0);
1797 
1798 #ifdef INET6
1799 	if (inp->inp_vflag & INP_IPV6) {
1800 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1801 
1802 		th = (struct tcphdr *)(ip6 + 1);
1803 		m->m_pkthdr.len = m->m_len =
1804 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1805 		tcp_fillheaders(tp, ip6, th, FALSE);
1806 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1807 	} else
1808 #endif
1809 	{
1810 		ip = mtod(m, struct ip *);
1811 		th = (struct tcphdr *)(ip + 1);
1812 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1813 		tcp_fillheaders(tp, ip, th, FALSE);
1814 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1815 	}
1816 
1817 	m_free(m);
1818 	return (hdrsiz);
1819 }
1820 #endif
1821 
1822 /*
1823  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1824  *
1825  * This code attempts to calculate the bandwidth-delay product as a
1826  * means of determining the optimal window size to maximize bandwidth,
1827  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1828  * routers.  This code also does a fairly good job keeping RTTs in check
1829  * across slow links like modems.  We implement an algorithm which is very
1830  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1831  * transmitter side of a TCP connection and so only effects the transmit
1832  * side of the connection.
1833  *
1834  * BACKGROUND:  TCP makes no provision for the management of buffer space
1835  * at the end points or at the intermediate routers and switches.  A TCP
1836  * stream, whether using NewReno or not, will eventually buffer as
1837  * many packets as it is able and the only reason this typically works is
1838  * due to the fairly small default buffers made available for a connection
1839  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1840  * scaling it is now fairly easy for even a single TCP connection to blow-out
1841  * all available buffer space not only on the local interface, but on
1842  * intermediate routers and switches as well.  NewReno makes a misguided
1843  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1844  * then backing off, then steadily increasing the window again until another
1845  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1846  * is only made worse as network loads increase and the idea of intentionally
1847  * blowing out network buffers is, frankly, a terrible way to manage network
1848  * resources.
1849  *
1850  * It is far better to limit the transmit window prior to the failure
1851  * condition being achieved.  There are two general ways to do this:  First
1852  * you can 'scan' through different transmit window sizes and locate the
1853  * point where the RTT stops increasing, indicating that you have filled the
1854  * pipe, then scan backwards until you note that RTT stops decreasing, then
1855  * repeat ad-infinitum.  This method works in principle but has severe
1856  * implementation issues due to RTT variances, timer granularity, and
1857  * instability in the algorithm which can lead to many false positives and
1858  * create oscillations as well as interact badly with other TCP streams
1859  * implementing the same algorithm.
1860  *
1861  * The second method is to limit the window to the bandwidth delay product
1862  * of the link.  This is the method we implement.  RTT variances and our
1863  * own manipulation of the congestion window, bwnd, can potentially
1864  * destabilize the algorithm.  For this reason we have to stabilize the
1865  * elements used to calculate the window.  We do this by using the minimum
1866  * observed RTT, the long term average of the observed bandwidth, and
1867  * by adding two segments worth of slop.  It isn't perfect but it is able
1868  * to react to changing conditions and gives us a very stable basis on
1869  * which to extend the algorithm.
1870  */
1871 void
1872 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1873 {
1874 	u_long bw;
1875 	u_long bwnd;
1876 	int save_ticks;
1877 	int delta_ticks;
1878 
1879 	/*
1880 	 * If inflight_enable is disabled in the middle of a tcp connection,
1881 	 * make sure snd_bwnd is effectively disabled.
1882 	 */
1883 	if (!tcp_inflight_enable) {
1884 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1885 		tp->snd_bandwidth = 0;
1886 		return;
1887 	}
1888 
1889 	/*
1890 	 * Validate the delta time.  If a connection is new or has been idle
1891 	 * a long time we have to reset the bandwidth calculator.
1892 	 */
1893 	save_ticks = ticks;
1894 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1895 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1896 		tp->t_bw_rtttime = ticks;
1897 		tp->t_bw_rtseq = ack_seq;
1898 		if (tp->snd_bandwidth == 0)
1899 			tp->snd_bandwidth = tcp_inflight_min;
1900 		return;
1901 	}
1902 	if (delta_ticks == 0)
1903 		return;
1904 
1905 	/*
1906 	 * Sanity check, plus ignore pure window update acks.
1907 	 */
1908 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1909 		return;
1910 
1911 	/*
1912 	 * Figure out the bandwidth.  Due to the tick granularity this
1913 	 * is a very rough number and it MUST be averaged over a fairly
1914 	 * long period of time.  XXX we need to take into account a link
1915 	 * that is not using all available bandwidth, but for now our
1916 	 * slop will ramp us up if this case occurs and the bandwidth later
1917 	 * increases.
1918 	 */
1919 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1920 	tp->t_bw_rtttime = save_ticks;
1921 	tp->t_bw_rtseq = ack_seq;
1922 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1923 
1924 	tp->snd_bandwidth = bw;
1925 
1926 	/*
1927 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1928 	 * segments.  The additional slop puts us squarely in the sweet
1929 	 * spot and also handles the bandwidth run-up case.  Without the
1930 	 * slop we could be locking ourselves into a lower bandwidth.
1931 	 *
1932 	 * Situations Handled:
1933 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1934 	 *	    high speed LANs, allowing larger TCP buffers to be
1935 	 *	    specified, and also does a good job preventing
1936 	 *	    over-queueing of packets over choke points like modems
1937 	 *	    (at least for the transmit side).
1938 	 *
1939 	 *	(2) Is able to handle changing network loads (bandwidth
1940 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1941 	 *	    increases).
1942 	 *
1943 	 *	(3) Theoretically should stabilize in the face of multiple
1944 	 *	    connections implementing the same algorithm (this may need
1945 	 *	    a little work).
1946 	 *
1947 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1948 	 *	    be adjusted with a sysctl but typically only needs to be on
1949 	 *	    very slow connections.  A value no smaller then 5 should
1950 	 *	    be used, but only reduce this default if you have no other
1951 	 *	    choice.
1952 	 */
1953 
1954 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1955 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1956 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1957 #undef USERTT
1958 
1959 	if (tcp_inflight_debug > 0) {
1960 		static int ltime;
1961 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1962 			ltime = ticks;
1963 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1964 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1965 		}
1966 	}
1967 	if ((long)bwnd < tcp_inflight_min)
1968 		bwnd = tcp_inflight_min;
1969 	if (bwnd > tcp_inflight_max)
1970 		bwnd = tcp_inflight_max;
1971 	if ((long)bwnd < tp->t_maxseg * 2)
1972 		bwnd = tp->t_maxseg * 2;
1973 	tp->snd_bwnd = bwnd;
1974 }
1975 
1976 static void
1977 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs)
1978 {
1979 	struct rtentry *rt;
1980 	struct inpcb *inp = tp->t_inpcb;
1981 #ifdef INET6
1982 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
1983 #else
1984 	const boolean_t isipv6 = FALSE;
1985 #endif
1986 
1987 	/* XXX */
1988 	if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT)
1989 		tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
1990 	if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT)
1991 		tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
1992 
1993 	if (isipv6)
1994 		rt = tcp_rtlookup6(&inp->inp_inc);
1995 	else
1996 		rt = tcp_rtlookup(&inp->inp_inc);
1997 	if (rt == NULL ||
1998 	    rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT ||
1999 	    rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) {
2000 		*maxsegs = tcp_iw_maxsegs;
2001 		*capsegs = tcp_iw_capsegs;
2002 		return;
2003 	}
2004 	*maxsegs = rt->rt_rmx.rmx_iwmaxsegs;
2005 	*capsegs = rt->rt_rmx.rmx_iwcapsegs;
2006 }
2007 
2008 u_long
2009 tcp_initial_window(struct tcpcb *tp)
2010 {
2011 	if (tcp_do_rfc3390) {
2012 		/*
2013 		 * RFC3390:
2014 		 * "If the SYN or SYN/ACK is lost, the initial window
2015 		 *  used by a sender after a correctly transmitted SYN
2016 		 *  MUST be one segment consisting of MSS bytes."
2017 		 *
2018 		 * However, we do something a little bit more aggressive
2019 		 * then RFC3390 here:
2020 		 * - Only if time spent in the SYN or SYN|ACK retransmition
2021 		 *   >= 3 seconds, the IW is reduced.  We do this mainly
2022 		 *   because when RFC3390 is published, the initial RTO is
2023 		 *   still 3 seconds (the threshold we test here), while
2024 		 *   after RFC6298, the initial RTO is 1 second.  This
2025 		 *   behaviour probably still falls within the spirit of
2026 		 *   RFC3390.
2027 		 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2028 		 *   Mainly to avoid sender and receiver deadlock until
2029 		 *   delayed ACK timer expires.  And even RFC2581 does not
2030 		 *   try to reduce IW upon SYN or SYN|ACK retransmition
2031 		 *   timeout.
2032 		 *
2033 		 * See also:
2034 		 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2035 		 */
2036 		if (tp->t_rxtsyn >= TCPTV_RTOBASE3) {
2037 			return (2 * tp->t_maxseg);
2038 		} else {
2039 			u_long maxsegs, capsegs;
2040 
2041 			tcp_rmx_iwsegs(tp, &maxsegs, &capsegs);
2042 			return min(maxsegs * tp->t_maxseg,
2043 				   max(2 * tp->t_maxseg, capsegs * 1460));
2044 		}
2045 	} else {
2046 		/*
2047 		 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2048 		 *
2049 		 * Mainly to avoid sender and receiver deadlock
2050 		 * until delayed ACK timer expires.
2051 		 */
2052 		return (2 * tp->t_maxseg);
2053 	}
2054 }
2055 
2056 #ifdef TCP_SIGNATURE
2057 /*
2058  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2059  *
2060  * We do this over ip, tcphdr, segment data, and the key in the SADB.
2061  * When called from tcp_input(), we can be sure that th_sum has been
2062  * zeroed out and verified already.
2063  *
2064  * Return 0 if successful, otherwise return -1.
2065  *
2066  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2067  * search with the destination IP address, and a 'magic SPI' to be
2068  * determined by the application. This is hardcoded elsewhere to 1179
2069  * right now. Another branch of this code exists which uses the SPD to
2070  * specify per-application flows but it is unstable.
2071  */
2072 int
2073 tcpsignature_compute(
2074 	struct mbuf *m,		/* mbuf chain */
2075 	int len,		/* length of TCP data */
2076 	int optlen,		/* length of TCP options */
2077 	u_char *buf,		/* storage for MD5 digest */
2078 	u_int direction)	/* direction of flow */
2079 {
2080 	struct ippseudo ippseudo;
2081 	MD5_CTX ctx;
2082 	int doff;
2083 	struct ip *ip;
2084 	struct ipovly *ipovly;
2085 	struct secasvar *sav;
2086 	struct tcphdr *th;
2087 #ifdef INET6
2088 	struct ip6_hdr *ip6;
2089 	struct in6_addr in6;
2090 	uint32_t plen;
2091 	uint16_t nhdr;
2092 #endif /* INET6 */
2093 	u_short savecsum;
2094 
2095 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2096 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2097 	/*
2098 	 * Extract the destination from the IP header in the mbuf.
2099 	 */
2100 	ip = mtod(m, struct ip *);
2101 #ifdef INET6
2102 	ip6 = NULL;     /* Make the compiler happy. */
2103 #endif /* INET6 */
2104 	/*
2105 	 * Look up an SADB entry which matches the address found in
2106 	 * the segment.
2107 	 */
2108 	switch (IP_VHL_V(ip->ip_vhl)) {
2109 	case IPVERSION:
2110 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2111 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2112 		break;
2113 #ifdef INET6
2114 	case (IPV6_VERSION >> 4):
2115 		ip6 = mtod(m, struct ip6_hdr *);
2116 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2117 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2118 		break;
2119 #endif /* INET6 */
2120 	default:
2121 		return (EINVAL);
2122 		/* NOTREACHED */
2123 		break;
2124 	}
2125 	if (sav == NULL) {
2126 		kprintf("%s: SADB lookup failed\n", __func__);
2127 		return (EINVAL);
2128 	}
2129 	MD5Init(&ctx);
2130 
2131 	/*
2132 	 * Step 1: Update MD5 hash with IP pseudo-header.
2133 	 *
2134 	 * XXX The ippseudo header MUST be digested in network byte order,
2135 	 * or else we'll fail the regression test. Assume all fields we've
2136 	 * been doing arithmetic on have been in host byte order.
2137 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2138 	 * tcp_output(), the underlying ip_len member has not yet been set.
2139 	 */
2140 	switch (IP_VHL_V(ip->ip_vhl)) {
2141 	case IPVERSION:
2142 		ipovly = (struct ipovly *)ip;
2143 		ippseudo.ippseudo_src = ipovly->ih_src;
2144 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2145 		ippseudo.ippseudo_pad = 0;
2146 		ippseudo.ippseudo_p = IPPROTO_TCP;
2147 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2148 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2149 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2150 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2151 		break;
2152 #ifdef INET6
2153 	/*
2154 	 * RFC 2385, 2.0  Proposal
2155 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2156 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2157 	 * extended next header value (to form 32 bits), and 32-bit segment
2158 	 * length.
2159 	 * Note: Upper-Layer Packet Length comes before Next Header.
2160 	 */
2161 	case (IPV6_VERSION >> 4):
2162 		in6 = ip6->ip6_src;
2163 		in6_clearscope(&in6);
2164 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2165 		in6 = ip6->ip6_dst;
2166 		in6_clearscope(&in6);
2167 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2168 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2169 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2170 		nhdr = 0;
2171 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2172 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2173 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2174 		nhdr = IPPROTO_TCP;
2175 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2176 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2177 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2178 		break;
2179 #endif /* INET6 */
2180 	default:
2181 		return (EINVAL);
2182 		/* NOTREACHED */
2183 		break;
2184 	}
2185 	/*
2186 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2187 	 * The TCP checksum must be set to zero.
2188 	 */
2189 	savecsum = th->th_sum;
2190 	th->th_sum = 0;
2191 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2192 	th->th_sum = savecsum;
2193 	/*
2194 	 * Step 3: Update MD5 hash with TCP segment data.
2195 	 *         Use m_apply() to avoid an early m_pullup().
2196 	 */
2197 	if (len > 0)
2198 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2199 	/*
2200 	 * Step 4: Update MD5 hash with shared secret.
2201 	 */
2202 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2203 	MD5Final(buf, &ctx);
2204 	key_sa_recordxfer(sav, m);
2205 	key_freesav(sav);
2206 	return (0);
2207 }
2208 
2209 int
2210 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2211 {
2212 
2213 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2214 	return (0);
2215 }
2216 #endif /* TCP_SIGNATURE */
2217 
2218 boolean_t
2219 tcp_tso_pullup(struct mbuf **mp, int hoff, struct ip **ip0, int *iphlen0,
2220     struct tcphdr **th0, int *thoff0)
2221 {
2222 	struct mbuf *m = *mp;
2223 	struct ip *ip;
2224 	int len, iphlen;
2225 	struct tcphdr *th;
2226 	int thoff;
2227 
2228 	len = hoff + sizeof(struct ip);
2229 
2230 	/* The fixed IP header must reside completely in the first mbuf. */
2231 	if (m->m_len < len) {
2232 		m = m_pullup(m, len);
2233 		if (m == NULL)
2234 			goto fail;
2235 	}
2236 
2237 	ip = mtodoff(m, struct ip *, hoff);
2238 	iphlen = IP_VHL_HL(ip->ip_vhl) << 2;
2239 
2240 	/* The full IP header must reside completely in the one mbuf. */
2241 	if (m->m_len < hoff + iphlen) {
2242 		m = m_pullup(m, hoff + iphlen);
2243 		if (m == NULL)
2244 			goto fail;
2245 		ip = mtodoff(m, struct ip *, hoff);
2246 	}
2247 
2248 	KASSERT(ip->ip_p == IPPROTO_TCP, ("not tcp %d", ip->ip_p));
2249 
2250 	if (m->m_len < hoff + iphlen + sizeof(struct tcphdr)) {
2251 		m = m_pullup(m, hoff + iphlen + sizeof(struct tcphdr));
2252 		if (m == NULL)
2253 			goto fail;
2254 		ip = mtodoff(m, struct ip *, hoff);
2255 	}
2256 
2257 	th = (struct tcphdr *)((caddr_t)ip + iphlen);
2258 	thoff = th->th_off << 2;
2259 
2260 	if (m->m_len < hoff + iphlen + thoff) {
2261 		m = m_pullup(m, hoff + iphlen + thoff);
2262 		if (m == NULL)
2263 			goto fail;
2264 		ip = mtodoff(m, struct ip *, hoff);
2265 		th = (struct tcphdr *)((caddr_t)ip + iphlen);
2266 	}
2267 
2268 	*mp = m;
2269 	*ip0 = ip;
2270 	*iphlen0 = iphlen;
2271 	*th0 = th;
2272 	*thoff0 = thoff;
2273 	return TRUE;
2274 
2275 fail:
2276 	if (m != NULL)
2277 		m_freem(m);
2278 	*mp = NULL;
2279 	return FALSE;
2280 }
2281