xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 0bb7d8c82a64940013681cf515d16f3e62eb7e3c)
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_inet.h"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_tcpdebug.h"
76 
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/callout.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mpipe.h>
84 #include <sys/mbuf.h>
85 #ifdef INET6
86 #include <sys/domain.h>
87 #endif
88 #include <sys/proc.h>
89 #include <sys/priv.h>
90 #include <sys/socket.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 KTR_INFO_MASTER(tcp);
155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
158 #define logtcp(name)	KTR_LOG(tcp_ ## name)
159 
160 struct inpcbinfo tcbinfo[MAXCPU];
161 struct tcpcbackqhead tcpcbackq[MAXCPU];
162 
163 int tcp_mpsafe_proto = 0;
164 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
165 
166 static int tcp_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE;
167 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
168 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
169 	   &tcp_mpsafe_thread, 0,
170 	   "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
171 
172 int tcp_mssdflt = TCP_MSS;
173 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
174     &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
175 
176 #ifdef INET6
177 int tcp_v6mssdflt = TCP6_MSS;
178 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
179     &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
180 #endif
181 
182 /*
183  * Minimum MSS we accept and use. This prevents DoS attacks where
184  * we are forced to a ridiculous low MSS like 20 and send hundreds
185  * of packets instead of one. The effect scales with the available
186  * bandwidth and quickly saturates the CPU and network interface
187  * with packet generation and sending. Set to zero to disable MINMSS
188  * checking. This setting prevents us from sending too small packets.
189  */
190 int tcp_minmss = TCP_MINMSS;
191 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
192     &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
193 
194 #if 0
195 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
196 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
197     &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
198 #endif
199 
200 int tcp_do_rfc1323 = 1;
201 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
202     &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
203 
204 static int tcp_tcbhashsize = 0;
205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
206      &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
207 
208 static int do_tcpdrain = 1;
209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
210      "Enable tcp_drain routine for extra help when low on mbufs");
211 
212 static int icmp_may_rst = 1;
213 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
214     "Certain ICMP unreachable messages may abort connections in SYN_SENT");
215 
216 static int tcp_isn_reseed_interval = 0;
217 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
218     &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
219 
220 /*
221  * TCP bandwidth limiting sysctls.  The inflight limiter is now turned on
222  * by default, but with generous values which should allow maximal
223  * bandwidth.  In particular, the slop defaults to 50 (5 packets).
224  *
225  * The reason for doing this is that the limiter is the only mechanism we
226  * have which seems to do a really good job preventing receiver RX rings
227  * on network interfaces from getting blown out.  Even though GigE/10GigE
228  * is supposed to flow control it looks like either it doesn't actually
229  * do it or Open Source drivers do not properly enable it.
230  *
231  * People using the limiter to reduce bottlenecks on slower WAN connections
232  * should set the slop to 20 (2 packets).
233  */
234 static int tcp_inflight_enable = 1;
235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
236     &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
237 
238 static int tcp_inflight_debug = 0;
239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
240     &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
241 
242 static int tcp_inflight_min = 6144;
243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
244     &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
245 
246 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
248     &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
249 
250 static int tcp_inflight_stab = 50;
251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
252     &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
253 
254 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
255 static struct malloc_pipe tcptemp_mpipe;
256 
257 static void tcp_willblock(int);
258 static void tcp_notify (struct inpcb *, int);
259 
260 struct tcp_stats tcpstats_percpu[MAXCPU];
261 #ifdef SMP
262 static int
263 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
264 {
265 	int cpu, error = 0;
266 
267 	for (cpu = 0; cpu < ncpus; ++cpu) {
268 		if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
269 					sizeof(struct tcp_stats))))
270 			break;
271 		if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
272 				       sizeof(struct tcp_stats))))
273 			break;
274 	}
275 
276 	return (error);
277 }
278 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
279     0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
280 #else
281 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
282     &tcpstat, tcp_stats, "TCP statistics");
283 #endif
284 
285 /*
286  * Target size of TCP PCB hash tables. Must be a power of two.
287  *
288  * Note that this can be overridden by the kernel environment
289  * variable net.inet.tcp.tcbhashsize
290  */
291 #ifndef TCBHASHSIZE
292 #define	TCBHASHSIZE	512
293 #endif
294 
295 /*
296  * This is the actual shape of what we allocate using the zone
297  * allocator.  Doing it this way allows us to protect both structures
298  * using the same generation count, and also eliminates the overhead
299  * of allocating tcpcbs separately.  By hiding the structure here,
300  * we avoid changing most of the rest of the code (although it needs
301  * to be changed, eventually, for greater efficiency).
302  */
303 #define	ALIGNMENT	32
304 #define	ALIGNM1		(ALIGNMENT - 1)
305 struct	inp_tp {
306 	union {
307 		struct	inpcb inp;
308 		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
309 	} inp_tp_u;
310 	struct	tcpcb tcb;
311 	struct	tcp_callout inp_tp_rexmt;
312 	struct	tcp_callout inp_tp_persist;
313 	struct	tcp_callout inp_tp_keep;
314 	struct	tcp_callout inp_tp_2msl;
315 	struct	tcp_callout inp_tp_delack;
316 	struct	netmsg_tcp_timer inp_tp_timermsg;
317 };
318 #undef ALIGNMENT
319 #undef ALIGNM1
320 
321 /*
322  * Tcp initialization
323  */
324 void
325 tcp_init(void)
326 {
327 	struct inpcbporthead *porthashbase;
328 	u_long porthashmask;
329 	int hashsize = TCBHASHSIZE;
330 	int cpu;
331 
332 	/*
333 	 * note: tcptemp is used for keepalives, and it is ok for an
334 	 * allocation to fail so do not specify MPF_INT.
335 	 */
336 	mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
337 		    25, -1, 0, NULL);
338 
339 	tcp_delacktime = TCPTV_DELACK;
340 	tcp_keepinit = TCPTV_KEEP_INIT;
341 	tcp_keepidle = TCPTV_KEEP_IDLE;
342 	tcp_keepintvl = TCPTV_KEEPINTVL;
343 	tcp_maxpersistidle = TCPTV_KEEP_IDLE;
344 	tcp_msl = TCPTV_MSL;
345 	tcp_rexmit_min = TCPTV_MIN;
346 	tcp_rexmit_slop = TCPTV_CPU_VAR;
347 
348 	TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
349 	if (!powerof2(hashsize)) {
350 		kprintf("WARNING: TCB hash size not a power of 2\n");
351 		hashsize = 512; /* safe default */
352 	}
353 	tcp_tcbhashsize = hashsize;
354 	porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
355 
356 	for (cpu = 0; cpu < ncpus2; cpu++) {
357 		in_pcbinfo_init(&tcbinfo[cpu]);
358 		tcbinfo[cpu].cpu = cpu;
359 		tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
360 		    &tcbinfo[cpu].hashmask);
361 		tcbinfo[cpu].porthashbase = porthashbase;
362 		tcbinfo[cpu].porthashmask = porthashmask;
363 		tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
364 		    &tcbinfo[cpu].wildcardhashmask);
365 		tcbinfo[cpu].ipi_size = sizeof(struct inp_tp);
366 		TAILQ_INIT(&tcpcbackq[cpu]);
367 	}
368 
369 	tcp_reass_maxseg = nmbclusters / 16;
370 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
371 
372 #ifdef INET6
373 #define	TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
374 #else
375 #define	TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
376 #endif
377 	if (max_protohdr < TCP_MINPROTOHDR)
378 		max_protohdr = TCP_MINPROTOHDR;
379 	if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
380 		panic("tcp_init");
381 #undef TCP_MINPROTOHDR
382 
383 	/*
384 	 * Initialize TCP statistics counters for each CPU.
385 	 */
386 #ifdef SMP
387 	for (cpu = 0; cpu < ncpus; ++cpu) {
388 		bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
389 	}
390 #else
391 	bzero(&tcpstat, sizeof(struct tcp_stats));
392 #endif
393 
394 	syncache_init();
395 	tcp_thread_init();
396 }
397 
398 void
399 tcpmsg_service_loop(void *dummy)
400 {
401 	struct netmsg *msg;
402 	int mplocked;
403 
404 	/*
405 	 * Threads always start mpsafe.
406 	 */
407 	mplocked = 0;
408 
409 	while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
410 		do {
411 			logtcp(rxmsg);
412 			mplocked = netmsg_service(msg, tcp_mpsafe_thread,
413 						  mplocked);
414 		} while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
415 
416 		logtcp(delayed);
417 		tcp_willblock(mplocked);
418 		logtcp(wait);
419 	}
420 }
421 
422 static void
423 tcp_willblock(int mplocked)
424 {
425 	struct tcpcb *tp;
426 	int cpu = mycpu->gd_cpuid;
427 	int unlock = 0;
428 
429 	if (!mplocked && !tcp_mpsafe_proto) {
430 		if (TAILQ_EMPTY(&tcpcbackq[cpu]))
431 			return;
432 
433 		get_mplock();
434 		mplocked = 1;
435 		unlock = 1;
436 	}
437 
438 	while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
439 		KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
440 		tp->t_flags &= ~TF_ONOUTPUTQ;
441 		TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
442 		tcp_output(tp);
443 	}
444 
445 	if (unlock)
446 		rel_mplock();
447 }
448 
449 
450 /*
451  * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
452  * tcp_template used to store this data in mbufs, but we now recopy it out
453  * of the tcpcb each time to conserve mbufs.
454  */
455 void
456 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
457 {
458 	struct inpcb *inp = tp->t_inpcb;
459 	struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
460 
461 #ifdef INET6
462 	if (inp->inp_vflag & INP_IPV6) {
463 		struct ip6_hdr *ip6;
464 
465 		ip6 = (struct ip6_hdr *)ip_ptr;
466 		ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
467 			(inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
468 		ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
469 			(IPV6_VERSION & IPV6_VERSION_MASK);
470 		ip6->ip6_nxt = IPPROTO_TCP;
471 		ip6->ip6_plen = sizeof(struct tcphdr);
472 		ip6->ip6_src = inp->in6p_laddr;
473 		ip6->ip6_dst = inp->in6p_faddr;
474 		tcp_hdr->th_sum = 0;
475 	} else
476 #endif
477 	{
478 		struct ip *ip = (struct ip *) ip_ptr;
479 
480 		ip->ip_vhl = IP_VHL_BORING;
481 		ip->ip_tos = 0;
482 		ip->ip_len = 0;
483 		ip->ip_id = 0;
484 		ip->ip_off = 0;
485 		ip->ip_ttl = 0;
486 		ip->ip_sum = 0;
487 		ip->ip_p = IPPROTO_TCP;
488 		ip->ip_src = inp->inp_laddr;
489 		ip->ip_dst = inp->inp_faddr;
490 		tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
491 				    ip->ip_dst.s_addr,
492 				    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
493 	}
494 
495 	tcp_hdr->th_sport = inp->inp_lport;
496 	tcp_hdr->th_dport = inp->inp_fport;
497 	tcp_hdr->th_seq = 0;
498 	tcp_hdr->th_ack = 0;
499 	tcp_hdr->th_x2 = 0;
500 	tcp_hdr->th_off = 5;
501 	tcp_hdr->th_flags = 0;
502 	tcp_hdr->th_win = 0;
503 	tcp_hdr->th_urp = 0;
504 }
505 
506 /*
507  * Create template to be used to send tcp packets on a connection.
508  * Allocates an mbuf and fills in a skeletal tcp/ip header.  The only
509  * use for this function is in keepalives, which use tcp_respond.
510  */
511 struct tcptemp *
512 tcp_maketemplate(struct tcpcb *tp)
513 {
514 	struct tcptemp *tmp;
515 
516 	if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
517 		return (NULL);
518 	tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
519 	return (tmp);
520 }
521 
522 void
523 tcp_freetemplate(struct tcptemp *tmp)
524 {
525 	mpipe_free(&tcptemp_mpipe, tmp);
526 }
527 
528 /*
529  * Send a single message to the TCP at address specified by
530  * the given TCP/IP header.  If m == NULL, then we make a copy
531  * of the tcpiphdr at ti and send directly to the addressed host.
532  * This is used to force keep alive messages out using the TCP
533  * template for a connection.  If flags are given then we send
534  * a message back to the TCP which originated the * segment ti,
535  * and discard the mbuf containing it and any other attached mbufs.
536  *
537  * In any case the ack and sequence number of the transmitted
538  * segment are as specified by the parameters.
539  *
540  * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
541  */
542 void
543 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
544 	    tcp_seq ack, tcp_seq seq, int flags)
545 {
546 	int tlen;
547 	int win = 0;
548 	struct route *ro = NULL;
549 	struct route sro;
550 	struct ip *ip = ipgen;
551 	struct tcphdr *nth;
552 	int ipflags = 0;
553 	struct route_in6 *ro6 = NULL;
554 	struct route_in6 sro6;
555 	struct ip6_hdr *ip6 = ipgen;
556 	boolean_t use_tmpro = TRUE;
557 #ifdef INET6
558 	boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
559 #else
560 	const boolean_t isipv6 = FALSE;
561 #endif
562 
563 	if (tp != NULL) {
564 		if (!(flags & TH_RST)) {
565 			win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
566 			if (win < 0)
567 				win = 0;
568 			if (win > (long)TCP_MAXWIN << tp->rcv_scale)
569 				win = (long)TCP_MAXWIN << tp->rcv_scale;
570 		}
571 		/*
572 		 * Don't use the route cache of a listen socket,
573 		 * it is not MPSAFE; use temporary route cache.
574 		 */
575 		if (tp->t_state != TCPS_LISTEN) {
576 			if (isipv6)
577 				ro6 = &tp->t_inpcb->in6p_route;
578 			else
579 				ro = &tp->t_inpcb->inp_route;
580 			use_tmpro = FALSE;
581 		}
582 	}
583 	if (use_tmpro) {
584 		if (isipv6) {
585 			ro6 = &sro6;
586 			bzero(ro6, sizeof *ro6);
587 		} else {
588 			ro = &sro;
589 			bzero(ro, sizeof *ro);
590 		}
591 	}
592 	if (m == NULL) {
593 		m = m_gethdr(MB_DONTWAIT, MT_HEADER);
594 		if (m == NULL)
595 			return;
596 		tlen = 0;
597 		m->m_data += max_linkhdr;
598 		if (isipv6) {
599 			bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
600 			ip6 = mtod(m, struct ip6_hdr *);
601 			nth = (struct tcphdr *)(ip6 + 1);
602 		} else {
603 			bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
604 			ip = mtod(m, struct ip *);
605 			nth = (struct tcphdr *)(ip + 1);
606 		}
607 		bcopy(th, nth, sizeof(struct tcphdr));
608 		flags = TH_ACK;
609 	} else {
610 		m_freem(m->m_next);
611 		m->m_next = NULL;
612 		m->m_data = (caddr_t)ipgen;
613 		/* m_len is set later */
614 		tlen = 0;
615 #define	xchg(a, b, type) { type t; t = a; a = b; b = t; }
616 		if (isipv6) {
617 			xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
618 			nth = (struct tcphdr *)(ip6 + 1);
619 		} else {
620 			xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
621 			nth = (struct tcphdr *)(ip + 1);
622 		}
623 		if (th != nth) {
624 			/*
625 			 * this is usually a case when an extension header
626 			 * exists between the IPv6 header and the
627 			 * TCP header.
628 			 */
629 			nth->th_sport = th->th_sport;
630 			nth->th_dport = th->th_dport;
631 		}
632 		xchg(nth->th_dport, nth->th_sport, n_short);
633 #undef xchg
634 	}
635 	if (isipv6) {
636 		ip6->ip6_flow = 0;
637 		ip6->ip6_vfc = IPV6_VERSION;
638 		ip6->ip6_nxt = IPPROTO_TCP;
639 		ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
640 		tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
641 	} else {
642 		tlen += sizeof(struct tcpiphdr);
643 		ip->ip_len = tlen;
644 		ip->ip_ttl = ip_defttl;
645 	}
646 	m->m_len = tlen;
647 	m->m_pkthdr.len = tlen;
648 	m->m_pkthdr.rcvif = NULL;
649 	nth->th_seq = htonl(seq);
650 	nth->th_ack = htonl(ack);
651 	nth->th_x2 = 0;
652 	nth->th_off = sizeof(struct tcphdr) >> 2;
653 	nth->th_flags = flags;
654 	if (tp != NULL)
655 		nth->th_win = htons((u_short) (win >> tp->rcv_scale));
656 	else
657 		nth->th_win = htons((u_short)win);
658 	nth->th_urp = 0;
659 	if (isipv6) {
660 		nth->th_sum = 0;
661 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
662 					sizeof(struct ip6_hdr),
663 					tlen - sizeof(struct ip6_hdr));
664 		ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
665 					       (ro6 && ro6->ro_rt) ?
666 						ro6->ro_rt->rt_ifp : NULL);
667 	} else {
668 		nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
669 		    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
670 		m->m_pkthdr.csum_flags = CSUM_TCP;
671 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
672 	}
673 #ifdef TCPDEBUG
674 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
675 		tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
676 #endif
677 	if (isipv6) {
678 		ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
679 			   tp ? tp->t_inpcb : NULL);
680 		if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
681 			RTFREE(ro6->ro_rt);
682 			ro6->ro_rt = NULL;
683 		}
684 	} else {
685 		ipflags |= IP_DEBUGROUTE;
686 		ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
687 		if ((ro == &sro) && (ro->ro_rt != NULL)) {
688 			RTFREE(ro->ro_rt);
689 			ro->ro_rt = NULL;
690 		}
691 	}
692 }
693 
694 /*
695  * Create a new TCP control block, making an
696  * empty reassembly queue and hooking it to the argument
697  * protocol control block.  The `inp' parameter must have
698  * come from the zone allocator set up in tcp_init().
699  */
700 struct tcpcb *
701 tcp_newtcpcb(struct inpcb *inp)
702 {
703 	struct inp_tp *it;
704 	struct tcpcb *tp;
705 #ifdef INET6
706 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
707 #else
708 	const boolean_t isipv6 = FALSE;
709 #endif
710 
711 	it = (struct inp_tp *)inp;
712 	tp = &it->tcb;
713 	bzero(tp, sizeof(struct tcpcb));
714 	LIST_INIT(&tp->t_segq);
715 	tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
716 
717 	/* Set up our timeouts. */
718 	tp->tt_rexmt = &it->inp_tp_rexmt;
719 	tp->tt_persist = &it->inp_tp_persist;
720 	tp->tt_keep = &it->inp_tp_keep;
721 	tp->tt_2msl = &it->inp_tp_2msl;
722 	tp->tt_delack = &it->inp_tp_delack;
723 	tcp_inittimers(tp);
724 
725 	/*
726 	 * Zero out timer message.  We don't create it here,
727 	 * since the current CPU may not be the owner of this
728 	 * inpcb.
729 	 */
730 	tp->tt_msg = &it->inp_tp_timermsg;
731 	bzero(tp->tt_msg, sizeof(*tp->tt_msg));
732 
733 	if (tcp_do_rfc1323)
734 		tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
735 	tp->t_inpcb = inp;	/* XXX */
736 	tp->t_state = TCPS_CLOSED;
737 	/*
738 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
739 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
740 	 * reasonable initial retransmit time.
741 	 */
742 	tp->t_srtt = TCPTV_SRTTBASE;
743 	tp->t_rttvar =
744 	    ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
745 	tp->t_rttmin = tcp_rexmit_min;
746 	tp->t_rxtcur = TCPTV_RTOBASE;
747 	tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
748 	tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
749 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
750 	tp->t_rcvtime = ticks;
751 	/*
752 	 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
753 	 * because the socket may be bound to an IPv6 wildcard address,
754 	 * which may match an IPv4-mapped IPv6 address.
755 	 */
756 	inp->inp_ip_ttl = ip_defttl;
757 	inp->inp_ppcb = tp;
758 	tcp_sack_tcpcb_init(tp);
759 	return (tp);		/* XXX */
760 }
761 
762 /*
763  * Drop a TCP connection, reporting the specified error.
764  * If connection is synchronized, then send a RST to peer.
765  */
766 struct tcpcb *
767 tcp_drop(struct tcpcb *tp, int error)
768 {
769 	struct socket *so = tp->t_inpcb->inp_socket;
770 
771 	if (TCPS_HAVERCVDSYN(tp->t_state)) {
772 		tp->t_state = TCPS_CLOSED;
773 		tcp_output(tp);
774 		tcpstat.tcps_drops++;
775 	} else
776 		tcpstat.tcps_conndrops++;
777 	if (error == ETIMEDOUT && tp->t_softerror)
778 		error = tp->t_softerror;
779 	so->so_error = error;
780 	return (tcp_close(tp));
781 }
782 
783 #ifdef SMP
784 
785 struct netmsg_remwildcard {
786 	struct netmsg		nm_netmsg;
787 	struct inpcb		*nm_inp;
788 	struct inpcbinfo	*nm_pcbinfo;
789 #if defined(INET6)
790 	int			nm_isinet6;
791 #else
792 	int			nm_unused01;
793 #endif
794 };
795 
796 /*
797  * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
798  * inp can be detached.  We do this by cycling through the cpus, ending up
799  * on the cpu controlling the inp last and then doing the disconnect.
800  */
801 static void
802 in_pcbremwildcardhash_handler(struct netmsg *msg0)
803 {
804 	struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
805 	int cpu;
806 
807 	cpu = msg->nm_pcbinfo->cpu;
808 
809 	if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
810 		/* note: detach removes any wildcard hash entry */
811 #ifdef INET6
812 		if (msg->nm_isinet6)
813 			in6_pcbdetach(msg->nm_inp);
814 		else
815 #endif
816 			in_pcbdetach(msg->nm_inp);
817 		lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
818 	} else {
819 		in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
820 		cpu = (cpu + 1) % ncpus2;
821 		msg->nm_pcbinfo = &tcbinfo[cpu];
822 		lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
823 	}
824 }
825 
826 #endif
827 
828 /*
829  * Close a TCP control block:
830  *	discard all space held by the tcp
831  *	discard internet protocol block
832  *	wake up any sleepers
833  */
834 struct tcpcb *
835 tcp_close(struct tcpcb *tp)
836 {
837 	struct tseg_qent *q;
838 	struct inpcb *inp = tp->t_inpcb;
839 	struct socket *so = inp->inp_socket;
840 	struct rtentry *rt;
841 	boolean_t dosavessthresh;
842 #ifdef SMP
843 	int cpu;
844 #endif
845 #ifdef INET6
846 	boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
847 	boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
848 #else
849 	const boolean_t isipv6 = FALSE;
850 #endif
851 
852 	/*
853 	 * The tp is not instantly destroyed in the wildcard case.  Setting
854 	 * the state to TCPS_TERMINATING will prevent the TCP stack from
855 	 * messing with it, though it should be noted that this change may
856 	 * not take effect on other cpus until we have chained the wildcard
857 	 * hash removal.
858 	 *
859 	 * XXX we currently depend on the BGL to synchronize the tp->t_state
860 	 * update and prevent other tcp protocol threads from accepting new
861 	 * connections on the listen socket we might be trying to close down.
862 	 */
863 	KKASSERT(tp->t_state != TCPS_TERMINATING);
864 	tp->t_state = TCPS_TERMINATING;
865 
866 	/*
867 	 * Make sure that all of our timers are stopped before we
868 	 * delete the PCB.  For listen TCP socket (tp->tt_msg == NULL),
869 	 * timers are never used.  If timer message is never created
870 	 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
871 	 */
872 	if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
873 		tcp_callout_stop(tp, tp->tt_rexmt);
874 		tcp_callout_stop(tp, tp->tt_persist);
875 		tcp_callout_stop(tp, tp->tt_keep);
876 		tcp_callout_stop(tp, tp->tt_2msl);
877 		tcp_callout_stop(tp, tp->tt_delack);
878 	}
879 
880 	if (tp->t_flags & TF_ONOUTPUTQ) {
881 		KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
882 		TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
883 		tp->t_flags &= ~TF_ONOUTPUTQ;
884 	}
885 
886 	/*
887 	 * If we got enough samples through the srtt filter,
888 	 * save the rtt and rttvar in the routing entry.
889 	 * 'Enough' is arbitrarily defined as the 16 samples.
890 	 * 16 samples is enough for the srtt filter to converge
891 	 * to within 5% of the correct value; fewer samples and
892 	 * we could save a very bogus rtt.
893 	 *
894 	 * Don't update the default route's characteristics and don't
895 	 * update anything that the user "locked".
896 	 */
897 	if (tp->t_rttupdated >= 16) {
898 		u_long i = 0;
899 
900 		if (isipv6) {
901 			struct sockaddr_in6 *sin6;
902 
903 			if ((rt = inp->in6p_route.ro_rt) == NULL)
904 				goto no_valid_rt;
905 			sin6 = (struct sockaddr_in6 *)rt_key(rt);
906 			if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
907 				goto no_valid_rt;
908 		} else
909 			if ((rt = inp->inp_route.ro_rt) == NULL ||
910 			    ((struct sockaddr_in *)rt_key(rt))->
911 			     sin_addr.s_addr == INADDR_ANY)
912 				goto no_valid_rt;
913 
914 		if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
915 			i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
916 			if (rt->rt_rmx.rmx_rtt && i)
917 				/*
918 				 * filter this update to half the old & half
919 				 * the new values, converting scale.
920 				 * See route.h and tcp_var.h for a
921 				 * description of the scaling constants.
922 				 */
923 				rt->rt_rmx.rmx_rtt =
924 				    (rt->rt_rmx.rmx_rtt + i) / 2;
925 			else
926 				rt->rt_rmx.rmx_rtt = i;
927 			tcpstat.tcps_cachedrtt++;
928 		}
929 		if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
930 			i = tp->t_rttvar *
931 			    (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
932 			if (rt->rt_rmx.rmx_rttvar && i)
933 				rt->rt_rmx.rmx_rttvar =
934 				    (rt->rt_rmx.rmx_rttvar + i) / 2;
935 			else
936 				rt->rt_rmx.rmx_rttvar = i;
937 			tcpstat.tcps_cachedrttvar++;
938 		}
939 		/*
940 		 * The old comment here said:
941 		 * update the pipelimit (ssthresh) if it has been updated
942 		 * already or if a pipesize was specified & the threshhold
943 		 * got below half the pipesize.  I.e., wait for bad news
944 		 * before we start updating, then update on both good
945 		 * and bad news.
946 		 *
947 		 * But we want to save the ssthresh even if no pipesize is
948 		 * specified explicitly in the route, because such
949 		 * connections still have an implicit pipesize specified
950 		 * by the global tcp_sendspace.  In the absence of a reliable
951 		 * way to calculate the pipesize, it will have to do.
952 		 */
953 		i = tp->snd_ssthresh;
954 		if (rt->rt_rmx.rmx_sendpipe != 0)
955 			dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
956 		else
957 			dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
958 		if (dosavessthresh ||
959 		    (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
960 		     (rt->rt_rmx.rmx_ssthresh != 0))) {
961 			/*
962 			 * convert the limit from user data bytes to
963 			 * packets then to packet data bytes.
964 			 */
965 			i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
966 			if (i < 2)
967 				i = 2;
968 			i *= tp->t_maxseg +
969 			     (isipv6 ?
970 			      sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
971 			      sizeof(struct tcpiphdr));
972 			if (rt->rt_rmx.rmx_ssthresh)
973 				rt->rt_rmx.rmx_ssthresh =
974 				    (rt->rt_rmx.rmx_ssthresh + i) / 2;
975 			else
976 				rt->rt_rmx.rmx_ssthresh = i;
977 			tcpstat.tcps_cachedssthresh++;
978 		}
979 	}
980 
981 no_valid_rt:
982 	/* free the reassembly queue, if any */
983 	while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
984 		LIST_REMOVE(q, tqe_q);
985 		m_freem(q->tqe_m);
986 		FREE(q, M_TSEGQ);
987 		atomic_add_int(&tcp_reass_qsize, -1);
988 	}
989 	/* throw away SACK blocks in scoreboard*/
990 	if (TCP_DO_SACK(tp))
991 		tcp_sack_cleanup(&tp->scb);
992 
993 	inp->inp_ppcb = NULL;
994 	soisdisconnected(so);
995 
996 	tcp_destroy_timermsg(tp);
997 	if (tp->t_flags & TF_SYNCACHE)
998 		syncache_destroy(tp);
999 
1000 	/*
1001 	 * Discard the inp.  In the SMP case a wildcard inp's hash (created
1002 	 * by a listen socket or an INADDR_ANY udp socket) is replicated
1003 	 * for each protocol thread and must be removed in the context of
1004 	 * that thread.  This is accomplished by chaining the message
1005 	 * through the cpus.
1006 	 *
1007 	 * If the inp is not wildcarded we simply detach, which will remove
1008 	 * the any hashes still present for this inp.
1009 	 */
1010 #ifdef SMP
1011 	if (inp->inp_flags & INP_WILDCARD_MP) {
1012 		struct netmsg_remwildcard *msg;
1013 
1014 		cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1015 		msg = kmalloc(sizeof(struct netmsg_remwildcard),
1016 			      M_LWKTMSG, M_INTWAIT);
1017 		netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1018 			    0, in_pcbremwildcardhash_handler);
1019 #ifdef INET6
1020 		msg->nm_isinet6 = isafinet6;
1021 #endif
1022 		msg->nm_inp = inp;
1023 		msg->nm_pcbinfo = &tcbinfo[cpu];
1024 		lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1025 	} else
1026 #endif
1027 	{
1028 		/* note: detach removes any wildcard hash entry */
1029 #ifdef INET6
1030 		if (isafinet6)
1031 			in6_pcbdetach(inp);
1032 		else
1033 #endif
1034 			in_pcbdetach(inp);
1035 	}
1036 	tcpstat.tcps_closed++;
1037 	return (NULL);
1038 }
1039 
1040 static __inline void
1041 tcp_drain_oncpu(struct inpcbhead *head)
1042 {
1043 	struct inpcb *marker;
1044 	struct inpcb *inpb;
1045 	struct tcpcb *tcpb;
1046 	struct tseg_qent *te;
1047 
1048 	/*
1049 	 * Allows us to block while running the list
1050 	 */
1051 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1052 	marker->inp_flags |= INP_PLACEMARKER;
1053 	LIST_INSERT_HEAD(head, marker, inp_list);
1054 
1055 	while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1056 		if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1057 		    (tcpb = intotcpcb(inpb)) != NULL &&
1058 		    (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1059 			LIST_REMOVE(te, tqe_q);
1060 			m_freem(te->tqe_m);
1061 			FREE(te, M_TSEGQ);
1062 			atomic_add_int(&tcp_reass_qsize, -1);
1063 			/* retry */
1064 		} else {
1065 			LIST_REMOVE(marker, inp_list);
1066 			LIST_INSERT_AFTER(inpb, marker, inp_list);
1067 		}
1068 	}
1069 	LIST_REMOVE(marker, inp_list);
1070 	kfree(marker, M_TEMP);
1071 }
1072 
1073 #ifdef SMP
1074 struct netmsg_tcp_drain {
1075 	struct netmsg		nm_netmsg;
1076 	struct inpcbhead	*nm_head;
1077 };
1078 
1079 static void
1080 tcp_drain_handler(netmsg_t netmsg)
1081 {
1082 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1083 
1084 	tcp_drain_oncpu(nm->nm_head);
1085 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1086 }
1087 #endif
1088 
1089 void
1090 tcp_drain(void)
1091 {
1092 #ifdef SMP
1093 	int cpu;
1094 #endif
1095 
1096 	if (!do_tcpdrain)
1097 		return;
1098 
1099 	/*
1100 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1101 	 * if there is one...
1102 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1103 	 *	reassembly queue should be flushed, but in a situation
1104 	 *	where we're really low on mbufs, this is potentially
1105 	 *	useful.
1106 	 */
1107 #ifdef SMP
1108 	for (cpu = 0; cpu < ncpus2; cpu++) {
1109 		struct netmsg_tcp_drain *msg;
1110 
1111 		if (cpu == mycpu->gd_cpuid) {
1112 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1113 		} else {
1114 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1115 				    M_LWKTMSG, M_NOWAIT);
1116 			if (msg == NULL)
1117 				continue;
1118 			netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1119 				    0, tcp_drain_handler);
1120 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1121 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1122 		}
1123 	}
1124 #else
1125 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1126 #endif
1127 }
1128 
1129 /*
1130  * Notify a tcp user of an asynchronous error;
1131  * store error as soft error, but wake up user
1132  * (for now, won't do anything until can select for soft error).
1133  *
1134  * Do not wake up user since there currently is no mechanism for
1135  * reporting soft errors (yet - a kqueue filter may be added).
1136  */
1137 static void
1138 tcp_notify(struct inpcb *inp, int error)
1139 {
1140 	struct tcpcb *tp = intotcpcb(inp);
1141 
1142 	/*
1143 	 * Ignore some errors if we are hooked up.
1144 	 * If connection hasn't completed, has retransmitted several times,
1145 	 * and receives a second error, give up now.  This is better
1146 	 * than waiting a long time to establish a connection that
1147 	 * can never complete.
1148 	 */
1149 	if (tp->t_state == TCPS_ESTABLISHED &&
1150 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1151 	      error == EHOSTDOWN)) {
1152 		return;
1153 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1154 	    tp->t_softerror)
1155 		tcp_drop(tp, error);
1156 	else
1157 		tp->t_softerror = error;
1158 #if 0
1159 	wakeup(&so->so_timeo);
1160 	sorwakeup(so);
1161 	sowwakeup(so);
1162 #endif
1163 }
1164 
1165 static int
1166 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1167 {
1168 	int error, i, n;
1169 	struct inpcb *marker;
1170 	struct inpcb *inp;
1171 	inp_gen_t gencnt;
1172 	globaldata_t gd;
1173 	int origcpu, ccpu;
1174 
1175 	error = 0;
1176 	n = 0;
1177 
1178 	/*
1179 	 * The process of preparing the TCB list is too time-consuming and
1180 	 * resource-intensive to repeat twice on every request.
1181 	 */
1182 	if (req->oldptr == NULL) {
1183 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1184 			gd = globaldata_find(ccpu);
1185 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1186 		}
1187 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1188 		return (0);
1189 	}
1190 
1191 	if (req->newptr != NULL)
1192 		return (EPERM);
1193 
1194 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1195 	marker->inp_flags |= INP_PLACEMARKER;
1196 
1197 	/*
1198 	 * OK, now we're committed to doing something.  Run the inpcb list
1199 	 * for each cpu in the system and construct the output.  Use a
1200 	 * list placemarker to deal with list changes occuring during
1201 	 * copyout blockages (but otherwise depend on being on the correct
1202 	 * cpu to avoid races).
1203 	 */
1204 	origcpu = mycpu->gd_cpuid;
1205 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1206 		globaldata_t rgd;
1207 		caddr_t inp_ppcb;
1208 		struct xtcpcb xt;
1209 		int cpu_id;
1210 
1211 		cpu_id = (origcpu + ccpu) % ncpus;
1212 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1213 			continue;
1214 		rgd = globaldata_find(cpu_id);
1215 		lwkt_setcpu_self(rgd);
1216 
1217 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1218 		n = tcbinfo[cpu_id].ipi_count;
1219 
1220 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1221 		i = 0;
1222 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1223 			/*
1224 			 * process a snapshot of pcbs, ignoring placemarkers
1225 			 * and using our own to allow SYSCTL_OUT to block.
1226 			 */
1227 			LIST_REMOVE(marker, inp_list);
1228 			LIST_INSERT_AFTER(inp, marker, inp_list);
1229 
1230 			if (inp->inp_flags & INP_PLACEMARKER)
1231 				continue;
1232 			if (inp->inp_gencnt > gencnt)
1233 				continue;
1234 			if (prison_xinpcb(req->td, inp))
1235 				continue;
1236 
1237 			xt.xt_len = sizeof xt;
1238 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1239 			inp_ppcb = inp->inp_ppcb;
1240 			if (inp_ppcb != NULL)
1241 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1242 			else
1243 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1244 			if (inp->inp_socket)
1245 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1246 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1247 				break;
1248 			++i;
1249 		}
1250 		LIST_REMOVE(marker, inp_list);
1251 		if (error == 0 && i < n) {
1252 			bzero(&xt, sizeof xt);
1253 			xt.xt_len = sizeof xt;
1254 			while (i < n) {
1255 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1256 				if (error)
1257 					break;
1258 				++i;
1259 			}
1260 		}
1261 	}
1262 
1263 	/*
1264 	 * Make sure we are on the same cpu we were on originally, since
1265 	 * higher level callers expect this.  Also don't pollute caches with
1266 	 * migrated userland data by (eventually) returning to userland
1267 	 * on a different cpu.
1268 	 */
1269 	lwkt_setcpu_self(globaldata_find(origcpu));
1270 	kfree(marker, M_TEMP);
1271 	return (error);
1272 }
1273 
1274 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1275 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1276 
1277 static int
1278 tcp_getcred(SYSCTL_HANDLER_ARGS)
1279 {
1280 	struct sockaddr_in addrs[2];
1281 	struct inpcb *inp;
1282 	int cpu;
1283 	int error;
1284 
1285 	error = priv_check(req->td, PRIV_ROOT);
1286 	if (error != 0)
1287 		return (error);
1288 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1289 	if (error != 0)
1290 		return (error);
1291 	crit_enter();
1292 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1293 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1294 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1295 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1296 	if (inp == NULL || inp->inp_socket == NULL) {
1297 		error = ENOENT;
1298 		goto out;
1299 	}
1300 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1301 out:
1302 	crit_exit();
1303 	return (error);
1304 }
1305 
1306 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1307     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1308 
1309 #ifdef INET6
1310 static int
1311 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1312 {
1313 	struct sockaddr_in6 addrs[2];
1314 	struct inpcb *inp;
1315 	int error;
1316 	boolean_t mapped = FALSE;
1317 
1318 	error = priv_check(req->td, PRIV_ROOT);
1319 	if (error != 0)
1320 		return (error);
1321 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1322 	if (error != 0)
1323 		return (error);
1324 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1325 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1326 			mapped = TRUE;
1327 		else
1328 			return (EINVAL);
1329 	}
1330 	crit_enter();
1331 	if (mapped) {
1332 		inp = in_pcblookup_hash(&tcbinfo[0],
1333 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1334 		    addrs[1].sin6_port,
1335 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1336 		    addrs[0].sin6_port,
1337 		    0, NULL);
1338 	} else {
1339 		inp = in6_pcblookup_hash(&tcbinfo[0],
1340 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1341 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1342 		    0, NULL);
1343 	}
1344 	if (inp == NULL || inp->inp_socket == NULL) {
1345 		error = ENOENT;
1346 		goto out;
1347 	}
1348 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1349 out:
1350 	crit_exit();
1351 	return (error);
1352 }
1353 
1354 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1355 	    0, 0,
1356 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1357 #endif
1358 
1359 struct netmsg_tcp_notify {
1360 	struct netmsg	nm_nmsg;
1361 	void		(*nm_notify)(struct inpcb *, int);
1362 	struct in_addr	nm_faddr;
1363 	int		nm_arg;
1364 };
1365 
1366 static void
1367 tcp_notifyall_oncpu(struct netmsg *netmsg)
1368 {
1369 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1370 	int nextcpu;
1371 
1372 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1373 			nmsg->nm_arg, nmsg->nm_notify);
1374 
1375 	nextcpu = mycpuid + 1;
1376 	if (nextcpu < ncpus2)
1377 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1378 	else
1379 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1380 }
1381 
1382 void
1383 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1384 {
1385 	struct ip *ip = vip;
1386 	struct tcphdr *th;
1387 	struct in_addr faddr;
1388 	struct inpcb *inp;
1389 	struct tcpcb *tp;
1390 	void (*notify)(struct inpcb *, int) = tcp_notify;
1391 	tcp_seq icmpseq;
1392 	int arg, cpu;
1393 
1394 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1395 		return;
1396 	}
1397 
1398 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1399 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1400 		return;
1401 
1402 	arg = inetctlerrmap[cmd];
1403 	if (cmd == PRC_QUENCH) {
1404 		notify = tcp_quench;
1405 	} else if (icmp_may_rst &&
1406 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1407 		    cmd == PRC_UNREACH_PORT ||
1408 		    cmd == PRC_TIMXCEED_INTRANS) &&
1409 		   ip != NULL) {
1410 		notify = tcp_drop_syn_sent;
1411 	} else if (cmd == PRC_MSGSIZE) {
1412 		struct icmp *icmp = (struct icmp *)
1413 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1414 
1415 		arg = ntohs(icmp->icmp_nextmtu);
1416 		notify = tcp_mtudisc;
1417 	} else if (PRC_IS_REDIRECT(cmd)) {
1418 		ip = NULL;
1419 		notify = in_rtchange;
1420 	} else if (cmd == PRC_HOSTDEAD) {
1421 		ip = NULL;
1422 	}
1423 
1424 	if (ip != NULL) {
1425 		crit_enter();
1426 		th = (struct tcphdr *)((caddr_t)ip +
1427 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1428 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1429 				  ip->ip_src.s_addr, th->th_sport);
1430 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1431 					ip->ip_src, th->th_sport, 0, NULL);
1432 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1433 			icmpseq = htonl(th->th_seq);
1434 			tp = intotcpcb(inp);
1435 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1436 			    SEQ_LT(icmpseq, tp->snd_max))
1437 				(*notify)(inp, arg);
1438 		} else {
1439 			struct in_conninfo inc;
1440 
1441 			inc.inc_fport = th->th_dport;
1442 			inc.inc_lport = th->th_sport;
1443 			inc.inc_faddr = faddr;
1444 			inc.inc_laddr = ip->ip_src;
1445 #ifdef INET6
1446 			inc.inc_isipv6 = 0;
1447 #endif
1448 			syncache_unreach(&inc, th);
1449 		}
1450 		crit_exit();
1451 	} else {
1452 		struct netmsg_tcp_notify nmsg;
1453 
1454 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1455 		netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1456 			    0, tcp_notifyall_oncpu);
1457 		nmsg.nm_faddr = faddr;
1458 		nmsg.nm_arg = arg;
1459 		nmsg.nm_notify = notify;
1460 
1461 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1462 	}
1463 }
1464 
1465 #ifdef INET6
1466 void
1467 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1468 {
1469 	struct tcphdr th;
1470 	void (*notify) (struct inpcb *, int) = tcp_notify;
1471 	struct ip6_hdr *ip6;
1472 	struct mbuf *m;
1473 	struct ip6ctlparam *ip6cp = NULL;
1474 	const struct sockaddr_in6 *sa6_src = NULL;
1475 	int off;
1476 	struct tcp_portonly {
1477 		u_int16_t th_sport;
1478 		u_int16_t th_dport;
1479 	} *thp;
1480 	int arg;
1481 
1482 	if (sa->sa_family != AF_INET6 ||
1483 	    sa->sa_len != sizeof(struct sockaddr_in6))
1484 		return;
1485 
1486 	arg = 0;
1487 	if (cmd == PRC_QUENCH)
1488 		notify = tcp_quench;
1489 	else if (cmd == PRC_MSGSIZE) {
1490 		struct ip6ctlparam *ip6cp = d;
1491 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1492 
1493 		arg = ntohl(icmp6->icmp6_mtu);
1494 		notify = tcp_mtudisc;
1495 	} else if (!PRC_IS_REDIRECT(cmd) &&
1496 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1497 		return;
1498 	}
1499 
1500 	/* if the parameter is from icmp6, decode it. */
1501 	if (d != NULL) {
1502 		ip6cp = (struct ip6ctlparam *)d;
1503 		m = ip6cp->ip6c_m;
1504 		ip6 = ip6cp->ip6c_ip6;
1505 		off = ip6cp->ip6c_off;
1506 		sa6_src = ip6cp->ip6c_src;
1507 	} else {
1508 		m = NULL;
1509 		ip6 = NULL;
1510 		off = 0;	/* fool gcc */
1511 		sa6_src = &sa6_any;
1512 	}
1513 
1514 	if (ip6 != NULL) {
1515 		struct in_conninfo inc;
1516 		/*
1517 		 * XXX: We assume that when IPV6 is non NULL,
1518 		 * M and OFF are valid.
1519 		 */
1520 
1521 		/* check if we can safely examine src and dst ports */
1522 		if (m->m_pkthdr.len < off + sizeof *thp)
1523 			return;
1524 
1525 		bzero(&th, sizeof th);
1526 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1527 
1528 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1529 		    (struct sockaddr *)ip6cp->ip6c_src,
1530 		    th.th_sport, cmd, arg, notify);
1531 
1532 		inc.inc_fport = th.th_dport;
1533 		inc.inc_lport = th.th_sport;
1534 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1535 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1536 		inc.inc_isipv6 = 1;
1537 		syncache_unreach(&inc, &th);
1538 	} else
1539 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1540 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1541 }
1542 #endif
1543 
1544 /*
1545  * Following is where TCP initial sequence number generation occurs.
1546  *
1547  * There are two places where we must use initial sequence numbers:
1548  * 1.  In SYN-ACK packets.
1549  * 2.  In SYN packets.
1550  *
1551  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1552  * tcp_syncache.c for details.
1553  *
1554  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1555  * depends on this property.  In addition, these ISNs should be
1556  * unguessable so as to prevent connection hijacking.  To satisfy
1557  * the requirements of this situation, the algorithm outlined in
1558  * RFC 1948 is used to generate sequence numbers.
1559  *
1560  * Implementation details:
1561  *
1562  * Time is based off the system timer, and is corrected so that it
1563  * increases by one megabyte per second.  This allows for proper
1564  * recycling on high speed LANs while still leaving over an hour
1565  * before rollover.
1566  *
1567  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1568  * between seeding of isn_secret.  This is normally set to zero,
1569  * as reseeding should not be necessary.
1570  *
1571  */
1572 
1573 #define	ISN_BYTES_PER_SECOND 1048576
1574 
1575 u_char isn_secret[32];
1576 int isn_last_reseed;
1577 MD5_CTX isn_ctx;
1578 
1579 tcp_seq
1580 tcp_new_isn(struct tcpcb *tp)
1581 {
1582 	u_int32_t md5_buffer[4];
1583 	tcp_seq new_isn;
1584 
1585 	/* Seed if this is the first use, reseed if requested. */
1586 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1587 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1588 		< (u_int)ticks))) {
1589 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1590 		isn_last_reseed = ticks;
1591 	}
1592 
1593 	/* Compute the md5 hash and return the ISN. */
1594 	MD5Init(&isn_ctx);
1595 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1596 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1597 #ifdef INET6
1598 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1599 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1600 			  sizeof(struct in6_addr));
1601 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1602 			  sizeof(struct in6_addr));
1603 	} else
1604 #endif
1605 	{
1606 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1607 			  sizeof(struct in_addr));
1608 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1609 			  sizeof(struct in_addr));
1610 	}
1611 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1612 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1613 	new_isn = (tcp_seq) md5_buffer[0];
1614 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1615 	return (new_isn);
1616 }
1617 
1618 /*
1619  * When a source quench is received, close congestion window
1620  * to one segment.  We will gradually open it again as we proceed.
1621  */
1622 void
1623 tcp_quench(struct inpcb *inp, int error)
1624 {
1625 	struct tcpcb *tp = intotcpcb(inp);
1626 
1627 	if (tp != NULL) {
1628 		tp->snd_cwnd = tp->t_maxseg;
1629 		tp->snd_wacked = 0;
1630 	}
1631 }
1632 
1633 /*
1634  * When a specific ICMP unreachable message is received and the
1635  * connection state is SYN-SENT, drop the connection.  This behavior
1636  * is controlled by the icmp_may_rst sysctl.
1637  */
1638 void
1639 tcp_drop_syn_sent(struct inpcb *inp, int error)
1640 {
1641 	struct tcpcb *tp = intotcpcb(inp);
1642 
1643 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1644 		tcp_drop(tp, error);
1645 }
1646 
1647 /*
1648  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1649  * based on the new value in the route.  Also nudge TCP to send something,
1650  * since we know the packet we just sent was dropped.
1651  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1652  */
1653 void
1654 tcp_mtudisc(struct inpcb *inp, int mtu)
1655 {
1656 	struct tcpcb *tp = intotcpcb(inp);
1657 	struct rtentry *rt;
1658 	struct socket *so = inp->inp_socket;
1659 	int maxopd, mss;
1660 #ifdef INET6
1661 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1662 #else
1663 	const boolean_t isipv6 = FALSE;
1664 #endif
1665 
1666 	if (tp == NULL)
1667 		return;
1668 
1669 	/*
1670 	 * If no MTU is provided in the ICMP message, use the
1671 	 * next lower likely value, as specified in RFC 1191.
1672 	 */
1673 	if (mtu == 0) {
1674 		int oldmtu;
1675 
1676 		oldmtu = tp->t_maxopd +
1677 		    (isipv6 ?
1678 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1679 		     sizeof(struct tcpiphdr));
1680 		mtu = ip_next_mtu(oldmtu, 0);
1681 	}
1682 
1683 	if (isipv6)
1684 		rt = tcp_rtlookup6(&inp->inp_inc);
1685 	else
1686 		rt = tcp_rtlookup(&inp->inp_inc);
1687 	if (rt != NULL) {
1688 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1689 			mtu = rt->rt_rmx.rmx_mtu;
1690 
1691 		maxopd = mtu -
1692 		    (isipv6 ?
1693 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1694 		     sizeof(struct tcpiphdr));
1695 
1696 		/*
1697 		 * XXX - The following conditional probably violates the TCP
1698 		 * spec.  The problem is that, since we don't know the
1699 		 * other end's MSS, we are supposed to use a conservative
1700 		 * default.  But, if we do that, then MTU discovery will
1701 		 * never actually take place, because the conservative
1702 		 * default is much less than the MTUs typically seen
1703 		 * on the Internet today.  For the moment, we'll sweep
1704 		 * this under the carpet.
1705 		 *
1706 		 * The conservative default might not actually be a problem
1707 		 * if the only case this occurs is when sending an initial
1708 		 * SYN with options and data to a host we've never talked
1709 		 * to before.  Then, they will reply with an MSS value which
1710 		 * will get recorded and the new parameters should get
1711 		 * recomputed.  For Further Study.
1712 		 */
1713 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1714 			maxopd = rt->rt_rmx.rmx_mssopt;
1715 	} else
1716 		maxopd = mtu -
1717 		    (isipv6 ?
1718 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1719 		     sizeof(struct tcpiphdr));
1720 
1721 	if (tp->t_maxopd <= maxopd)
1722 		return;
1723 	tp->t_maxopd = maxopd;
1724 
1725 	mss = maxopd;
1726 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1727 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1728 		mss -= TCPOLEN_TSTAMP_APPA;
1729 
1730 	/* round down to multiple of MCLBYTES */
1731 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1732 	if (mss > MCLBYTES)
1733 		mss &= ~(MCLBYTES - 1);
1734 #else
1735 	if (mss > MCLBYTES)
1736 		mss = (mss / MCLBYTES) * MCLBYTES;
1737 #endif
1738 
1739 	if (so->so_snd.ssb_hiwat < mss)
1740 		mss = so->so_snd.ssb_hiwat;
1741 
1742 	tp->t_maxseg = mss;
1743 	tp->t_rtttime = 0;
1744 	tp->snd_nxt = tp->snd_una;
1745 	tcp_output(tp);
1746 	tcpstat.tcps_mturesent++;
1747 }
1748 
1749 /*
1750  * Look-up the routing entry to the peer of this inpcb.  If no route
1751  * is found and it cannot be allocated the return NULL.  This routine
1752  * is called by TCP routines that access the rmx structure and by tcp_mss
1753  * to get the interface MTU.
1754  */
1755 struct rtentry *
1756 tcp_rtlookup(struct in_conninfo *inc)
1757 {
1758 	struct route *ro = &inc->inc_route;
1759 
1760 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1761 		/* No route yet, so try to acquire one */
1762 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1763 			/*
1764 			 * unused portions of the structure MUST be zero'd
1765 			 * out because rtalloc() treats it as opaque data
1766 			 */
1767 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1768 			ro->ro_dst.sa_family = AF_INET;
1769 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1770 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1771 			    inc->inc_faddr;
1772 			rtalloc(ro);
1773 		}
1774 	}
1775 	return (ro->ro_rt);
1776 }
1777 
1778 #ifdef INET6
1779 struct rtentry *
1780 tcp_rtlookup6(struct in_conninfo *inc)
1781 {
1782 	struct route_in6 *ro6 = &inc->inc6_route;
1783 
1784 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1785 		/* No route yet, so try to acquire one */
1786 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1787 			/*
1788 			 * unused portions of the structure MUST be zero'd
1789 			 * out because rtalloc() treats it as opaque data
1790 			 */
1791 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1792 			ro6->ro_dst.sin6_family = AF_INET6;
1793 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1794 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1795 			rtalloc((struct route *)ro6);
1796 		}
1797 	}
1798 	return (ro6->ro_rt);
1799 }
1800 #endif
1801 
1802 #ifdef IPSEC
1803 /* compute ESP/AH header size for TCP, including outer IP header. */
1804 size_t
1805 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1806 {
1807 	struct inpcb *inp;
1808 	struct mbuf *m;
1809 	size_t hdrsiz;
1810 	struct ip *ip;
1811 	struct tcphdr *th;
1812 
1813 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1814 		return (0);
1815 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1816 	if (!m)
1817 		return (0);
1818 
1819 #ifdef INET6
1820 	if (inp->inp_vflag & INP_IPV6) {
1821 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1822 
1823 		th = (struct tcphdr *)(ip6 + 1);
1824 		m->m_pkthdr.len = m->m_len =
1825 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1826 		tcp_fillheaders(tp, ip6, th);
1827 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1828 	} else
1829 #endif
1830 	{
1831 		ip = mtod(m, struct ip *);
1832 		th = (struct tcphdr *)(ip + 1);
1833 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1834 		tcp_fillheaders(tp, ip, th);
1835 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1836 	}
1837 
1838 	m_free(m);
1839 	return (hdrsiz);
1840 }
1841 #endif
1842 
1843 /*
1844  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1845  *
1846  * This code attempts to calculate the bandwidth-delay product as a
1847  * means of determining the optimal window size to maximize bandwidth,
1848  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1849  * routers.  This code also does a fairly good job keeping RTTs in check
1850  * across slow links like modems.  We implement an algorithm which is very
1851  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1852  * transmitter side of a TCP connection and so only effects the transmit
1853  * side of the connection.
1854  *
1855  * BACKGROUND:  TCP makes no provision for the management of buffer space
1856  * at the end points or at the intermediate routers and switches.  A TCP
1857  * stream, whether using NewReno or not, will eventually buffer as
1858  * many packets as it is able and the only reason this typically works is
1859  * due to the fairly small default buffers made available for a connection
1860  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1861  * scaling it is now fairly easy for even a single TCP connection to blow-out
1862  * all available buffer space not only on the local interface, but on
1863  * intermediate routers and switches as well.  NewReno makes a misguided
1864  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1865  * then backing off, then steadily increasing the window again until another
1866  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1867  * is only made worse as network loads increase and the idea of intentionally
1868  * blowing out network buffers is, frankly, a terrible way to manage network
1869  * resources.
1870  *
1871  * It is far better to limit the transmit window prior to the failure
1872  * condition being achieved.  There are two general ways to do this:  First
1873  * you can 'scan' through different transmit window sizes and locate the
1874  * point where the RTT stops increasing, indicating that you have filled the
1875  * pipe, then scan backwards until you note that RTT stops decreasing, then
1876  * repeat ad-infinitum.  This method works in principle but has severe
1877  * implementation issues due to RTT variances, timer granularity, and
1878  * instability in the algorithm which can lead to many false positives and
1879  * create oscillations as well as interact badly with other TCP streams
1880  * implementing the same algorithm.
1881  *
1882  * The second method is to limit the window to the bandwidth delay product
1883  * of the link.  This is the method we implement.  RTT variances and our
1884  * own manipulation of the congestion window, bwnd, can potentially
1885  * destabilize the algorithm.  For this reason we have to stabilize the
1886  * elements used to calculate the window.  We do this by using the minimum
1887  * observed RTT, the long term average of the observed bandwidth, and
1888  * by adding two segments worth of slop.  It isn't perfect but it is able
1889  * to react to changing conditions and gives us a very stable basis on
1890  * which to extend the algorithm.
1891  */
1892 void
1893 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1894 {
1895 	u_long bw;
1896 	u_long bwnd;
1897 	int save_ticks;
1898 	int delta_ticks;
1899 
1900 	/*
1901 	 * If inflight_enable is disabled in the middle of a tcp connection,
1902 	 * make sure snd_bwnd is effectively disabled.
1903 	 */
1904 	if (!tcp_inflight_enable) {
1905 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1906 		tp->snd_bandwidth = 0;
1907 		return;
1908 	}
1909 
1910 	/*
1911 	 * Validate the delta time.  If a connection is new or has been idle
1912 	 * a long time we have to reset the bandwidth calculator.
1913 	 */
1914 	save_ticks = ticks;
1915 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1916 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1917 		tp->t_bw_rtttime = ticks;
1918 		tp->t_bw_rtseq = ack_seq;
1919 		if (tp->snd_bandwidth == 0)
1920 			tp->snd_bandwidth = tcp_inflight_min;
1921 		return;
1922 	}
1923 	if (delta_ticks == 0)
1924 		return;
1925 
1926 	/*
1927 	 * Sanity check, plus ignore pure window update acks.
1928 	 */
1929 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1930 		return;
1931 
1932 	/*
1933 	 * Figure out the bandwidth.  Due to the tick granularity this
1934 	 * is a very rough number and it MUST be averaged over a fairly
1935 	 * long period of time.  XXX we need to take into account a link
1936 	 * that is not using all available bandwidth, but for now our
1937 	 * slop will ramp us up if this case occurs and the bandwidth later
1938 	 * increases.
1939 	 */
1940 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1941 	tp->t_bw_rtttime = save_ticks;
1942 	tp->t_bw_rtseq = ack_seq;
1943 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1944 
1945 	tp->snd_bandwidth = bw;
1946 
1947 	/*
1948 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1949 	 * segments.  The additional slop puts us squarely in the sweet
1950 	 * spot and also handles the bandwidth run-up case.  Without the
1951 	 * slop we could be locking ourselves into a lower bandwidth.
1952 	 *
1953 	 * Situations Handled:
1954 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1955 	 *	    high speed LANs, allowing larger TCP buffers to be
1956 	 *	    specified, and also does a good job preventing
1957 	 *	    over-queueing of packets over choke points like modems
1958 	 *	    (at least for the transmit side).
1959 	 *
1960 	 *	(2) Is able to handle changing network loads (bandwidth
1961 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1962 	 *	    increases).
1963 	 *
1964 	 *	(3) Theoretically should stabilize in the face of multiple
1965 	 *	    connections implementing the same algorithm (this may need
1966 	 *	    a little work).
1967 	 *
1968 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1969 	 *	    be adjusted with a sysctl but typically only needs to be on
1970 	 *	    very slow connections.  A value no smaller then 5 should
1971 	 *	    be used, but only reduce this default if you have no other
1972 	 *	    choice.
1973 	 */
1974 
1975 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1976 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1977 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1978 #undef USERTT
1979 
1980 	if (tcp_inflight_debug > 0) {
1981 		static int ltime;
1982 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1983 			ltime = ticks;
1984 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1985 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1986 		}
1987 	}
1988 	if ((long)bwnd < tcp_inflight_min)
1989 		bwnd = tcp_inflight_min;
1990 	if (bwnd > tcp_inflight_max)
1991 		bwnd = tcp_inflight_max;
1992 	if ((long)bwnd < tp->t_maxseg * 2)
1993 		bwnd = tp->t_maxseg * 2;
1994 	tp->snd_bwnd = bwnd;
1995 }
1996 
1997 #ifdef TCP_SIGNATURE
1998 /*
1999  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2000  *
2001  * We do this over ip, tcphdr, segment data, and the key in the SADB.
2002  * When called from tcp_input(), we can be sure that th_sum has been
2003  * zeroed out and verified already.
2004  *
2005  * Return 0 if successful, otherwise return -1.
2006  *
2007  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2008  * search with the destination IP address, and a 'magic SPI' to be
2009  * determined by the application. This is hardcoded elsewhere to 1179
2010  * right now. Another branch of this code exists which uses the SPD to
2011  * specify per-application flows but it is unstable.
2012  */
2013 int
2014 tcpsignature_compute(
2015 	struct mbuf *m,		/* mbuf chain */
2016 	int len,		/* length of TCP data */
2017 	int optlen,		/* length of TCP options */
2018 	u_char *buf,		/* storage for MD5 digest */
2019 	u_int direction)	/* direction of flow */
2020 {
2021 	struct ippseudo ippseudo;
2022 	MD5_CTX ctx;
2023 	int doff;
2024 	struct ip *ip;
2025 	struct ipovly *ipovly;
2026 	struct secasvar *sav;
2027 	struct tcphdr *th;
2028 #ifdef INET6
2029 	struct ip6_hdr *ip6;
2030 	struct in6_addr in6;
2031 	uint32_t plen;
2032 	uint16_t nhdr;
2033 #endif /* INET6 */
2034 	u_short savecsum;
2035 
2036 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2037 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2038 	/*
2039 	 * Extract the destination from the IP header in the mbuf.
2040 	 */
2041 	ip = mtod(m, struct ip *);
2042 #ifdef INET6
2043 	ip6 = NULL;     /* Make the compiler happy. */
2044 #endif /* INET6 */
2045 	/*
2046 	 * Look up an SADB entry which matches the address found in
2047 	 * the segment.
2048 	 */
2049 	switch (IP_VHL_V(ip->ip_vhl)) {
2050 	case IPVERSION:
2051 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2052 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2053 		break;
2054 #ifdef INET6
2055 	case (IPV6_VERSION >> 4):
2056 		ip6 = mtod(m, struct ip6_hdr *);
2057 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2058 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2059 		break;
2060 #endif /* INET6 */
2061 	default:
2062 		return (EINVAL);
2063 		/* NOTREACHED */
2064 		break;
2065 	}
2066 	if (sav == NULL) {
2067 		kprintf("%s: SADB lookup failed\n", __func__);
2068 		return (EINVAL);
2069 	}
2070 	MD5Init(&ctx);
2071 
2072 	/*
2073 	 * Step 1: Update MD5 hash with IP pseudo-header.
2074 	 *
2075 	 * XXX The ippseudo header MUST be digested in network byte order,
2076 	 * or else we'll fail the regression test. Assume all fields we've
2077 	 * been doing arithmetic on have been in host byte order.
2078 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2079 	 * tcp_output(), the underlying ip_len member has not yet been set.
2080 	 */
2081 	switch (IP_VHL_V(ip->ip_vhl)) {
2082 	case IPVERSION:
2083 		ipovly = (struct ipovly *)ip;
2084 		ippseudo.ippseudo_src = ipovly->ih_src;
2085 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2086 		ippseudo.ippseudo_pad = 0;
2087 		ippseudo.ippseudo_p = IPPROTO_TCP;
2088 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2089 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2090 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2091 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2092 		break;
2093 #ifdef INET6
2094 	/*
2095 	 * RFC 2385, 2.0  Proposal
2096 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2097 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2098 	 * extended next header value (to form 32 bits), and 32-bit segment
2099 	 * length.
2100 	 * Note: Upper-Layer Packet Length comes before Next Header.
2101 	 */
2102 	case (IPV6_VERSION >> 4):
2103 		in6 = ip6->ip6_src;
2104 		in6_clearscope(&in6);
2105 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2106 		in6 = ip6->ip6_dst;
2107 		in6_clearscope(&in6);
2108 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2109 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2110 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2111 		nhdr = 0;
2112 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2113 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2114 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2115 		nhdr = IPPROTO_TCP;
2116 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2117 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2118 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2119 		break;
2120 #endif /* INET6 */
2121 	default:
2122 		return (EINVAL);
2123 		/* NOTREACHED */
2124 		break;
2125 	}
2126 	/*
2127 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2128 	 * The TCP checksum must be set to zero.
2129 	 */
2130 	savecsum = th->th_sum;
2131 	th->th_sum = 0;
2132 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2133 	th->th_sum = savecsum;
2134 	/*
2135 	 * Step 3: Update MD5 hash with TCP segment data.
2136 	 *         Use m_apply() to avoid an early m_pullup().
2137 	 */
2138 	if (len > 0)
2139 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2140 	/*
2141 	 * Step 4: Update MD5 hash with shared secret.
2142 	 */
2143 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2144 	MD5Final(buf, &ctx);
2145 	key_sa_recordxfer(sav, m);
2146 	key_freesav(sav);
2147 	return (0);
2148 }
2149 
2150 int
2151 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2152 {
2153 
2154 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2155 	return (0);
2156 }
2157 #endif /* TCP_SIGNATURE */
2158