xref: /dflybsd-src/sys/netinet/tcp_subr.c (revision 14343ad3b815bafa1bcec3656de2d614fcc75bec)
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 		tcp_reass_qsize--;
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 *inpb;
1044 	struct tcpcb *tcpb;
1045 	struct tseg_qent *te;
1046 
1047 	LIST_FOREACH(inpb, head, inp_list) {
1048 		if (inpb->inp_flags & INP_PLACEMARKER)
1049 			continue;
1050 		if ((tcpb = intotcpcb(inpb))) {
1051 			while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1052 				LIST_REMOVE(te, tqe_q);
1053 				m_freem(te->tqe_m);
1054 				FREE(te, M_TSEGQ);
1055 				tcp_reass_qsize--;
1056 			}
1057 		}
1058 	}
1059 }
1060 
1061 #ifdef SMP
1062 struct netmsg_tcp_drain {
1063 	struct netmsg		nm_netmsg;
1064 	struct inpcbhead	*nm_head;
1065 };
1066 
1067 static void
1068 tcp_drain_handler(netmsg_t netmsg)
1069 {
1070 	struct netmsg_tcp_drain *nm = (void *)netmsg;
1071 
1072 	tcp_drain_oncpu(nm->nm_head);
1073 	lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1074 }
1075 #endif
1076 
1077 void
1078 tcp_drain(void)
1079 {
1080 #ifdef SMP
1081 	int cpu;
1082 #endif
1083 
1084 	if (!do_tcpdrain)
1085 		return;
1086 
1087 	/*
1088 	 * Walk the tcpbs, if existing, and flush the reassembly queue,
1089 	 * if there is one...
1090 	 * XXX: The "Net/3" implementation doesn't imply that the TCP
1091 	 *	reassembly queue should be flushed, but in a situation
1092 	 *	where we're really low on mbufs, this is potentially
1093 	 *	useful.
1094 	 */
1095 #ifdef SMP
1096 	for (cpu = 0; cpu < ncpus2; cpu++) {
1097 		struct netmsg_tcp_drain *msg;
1098 
1099 		if (cpu == mycpu->gd_cpuid) {
1100 			tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1101 		} else {
1102 			msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1103 				    M_LWKTMSG, M_NOWAIT);
1104 			if (msg == NULL)
1105 				continue;
1106 			netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1107 				    0, tcp_drain_handler);
1108 			msg->nm_head = &tcbinfo[cpu].pcblisthead;
1109 			lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1110 		}
1111 	}
1112 #else
1113 	tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1114 #endif
1115 }
1116 
1117 /*
1118  * Notify a tcp user of an asynchronous error;
1119  * store error as soft error, but wake up user
1120  * (for now, won't do anything until can select for soft error).
1121  *
1122  * Do not wake up user since there currently is no mechanism for
1123  * reporting soft errors (yet - a kqueue filter may be added).
1124  */
1125 static void
1126 tcp_notify(struct inpcb *inp, int error)
1127 {
1128 	struct tcpcb *tp = intotcpcb(inp);
1129 
1130 	/*
1131 	 * Ignore some errors if we are hooked up.
1132 	 * If connection hasn't completed, has retransmitted several times,
1133 	 * and receives a second error, give up now.  This is better
1134 	 * than waiting a long time to establish a connection that
1135 	 * can never complete.
1136 	 */
1137 	if (tp->t_state == TCPS_ESTABLISHED &&
1138 	     (error == EHOSTUNREACH || error == ENETUNREACH ||
1139 	      error == EHOSTDOWN)) {
1140 		return;
1141 	} else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1142 	    tp->t_softerror)
1143 		tcp_drop(tp, error);
1144 	else
1145 		tp->t_softerror = error;
1146 #if 0
1147 	wakeup(&so->so_timeo);
1148 	sorwakeup(so);
1149 	sowwakeup(so);
1150 #endif
1151 }
1152 
1153 static int
1154 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1155 {
1156 	int error, i, n;
1157 	struct inpcb *marker;
1158 	struct inpcb *inp;
1159 	inp_gen_t gencnt;
1160 	globaldata_t gd;
1161 	int origcpu, ccpu;
1162 
1163 	error = 0;
1164 	n = 0;
1165 
1166 	/*
1167 	 * The process of preparing the TCB list is too time-consuming and
1168 	 * resource-intensive to repeat twice on every request.
1169 	 */
1170 	if (req->oldptr == NULL) {
1171 		for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1172 			gd = globaldata_find(ccpu);
1173 			n += tcbinfo[gd->gd_cpuid].ipi_count;
1174 		}
1175 		req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1176 		return (0);
1177 	}
1178 
1179 	if (req->newptr != NULL)
1180 		return (EPERM);
1181 
1182 	marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1183 	marker->inp_flags |= INP_PLACEMARKER;
1184 
1185 	/*
1186 	 * OK, now we're committed to doing something.  Run the inpcb list
1187 	 * for each cpu in the system and construct the output.  Use a
1188 	 * list placemarker to deal with list changes occuring during
1189 	 * copyout blockages (but otherwise depend on being on the correct
1190 	 * cpu to avoid races).
1191 	 */
1192 	origcpu = mycpu->gd_cpuid;
1193 	for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1194 		globaldata_t rgd;
1195 		caddr_t inp_ppcb;
1196 		struct xtcpcb xt;
1197 		int cpu_id;
1198 
1199 		cpu_id = (origcpu + ccpu) % ncpus;
1200 		if ((smp_active_mask & (1 << cpu_id)) == 0)
1201 			continue;
1202 		rgd = globaldata_find(cpu_id);
1203 		lwkt_setcpu_self(rgd);
1204 
1205 		gencnt = tcbinfo[cpu_id].ipi_gencnt;
1206 		n = tcbinfo[cpu_id].ipi_count;
1207 
1208 		LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1209 		i = 0;
1210 		while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1211 			/*
1212 			 * process a snapshot of pcbs, ignoring placemarkers
1213 			 * and using our own to allow SYSCTL_OUT to block.
1214 			 */
1215 			LIST_REMOVE(marker, inp_list);
1216 			LIST_INSERT_AFTER(inp, marker, inp_list);
1217 
1218 			if (inp->inp_flags & INP_PLACEMARKER)
1219 				continue;
1220 			if (inp->inp_gencnt > gencnt)
1221 				continue;
1222 			if (prison_xinpcb(req->td, inp))
1223 				continue;
1224 
1225 			xt.xt_len = sizeof xt;
1226 			bcopy(inp, &xt.xt_inp, sizeof *inp);
1227 			inp_ppcb = inp->inp_ppcb;
1228 			if (inp_ppcb != NULL)
1229 				bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1230 			else
1231 				bzero(&xt.xt_tp, sizeof xt.xt_tp);
1232 			if (inp->inp_socket)
1233 				sotoxsocket(inp->inp_socket, &xt.xt_socket);
1234 			if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1235 				break;
1236 			++i;
1237 		}
1238 		LIST_REMOVE(marker, inp_list);
1239 		if (error == 0 && i < n) {
1240 			bzero(&xt, sizeof xt);
1241 			xt.xt_len = sizeof xt;
1242 			while (i < n) {
1243 				error = SYSCTL_OUT(req, &xt, sizeof xt);
1244 				if (error)
1245 					break;
1246 				++i;
1247 			}
1248 		}
1249 	}
1250 
1251 	/*
1252 	 * Make sure we are on the same cpu we were on originally, since
1253 	 * higher level callers expect this.  Also don't pollute caches with
1254 	 * migrated userland data by (eventually) returning to userland
1255 	 * on a different cpu.
1256 	 */
1257 	lwkt_setcpu_self(globaldata_find(origcpu));
1258 	kfree(marker, M_TEMP);
1259 	return (error);
1260 }
1261 
1262 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1263 	    tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1264 
1265 static int
1266 tcp_getcred(SYSCTL_HANDLER_ARGS)
1267 {
1268 	struct sockaddr_in addrs[2];
1269 	struct inpcb *inp;
1270 	int cpu;
1271 	int error;
1272 
1273 	error = priv_check(req->td, PRIV_ROOT);
1274 	if (error != 0)
1275 		return (error);
1276 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1277 	if (error != 0)
1278 		return (error);
1279 	crit_enter();
1280 	cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1281 	    addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1282 	inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1283 	    addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1284 	if (inp == NULL || inp->inp_socket == NULL) {
1285 		error = ENOENT;
1286 		goto out;
1287 	}
1288 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1289 out:
1290 	crit_exit();
1291 	return (error);
1292 }
1293 
1294 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1295     0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1296 
1297 #ifdef INET6
1298 static int
1299 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1300 {
1301 	struct sockaddr_in6 addrs[2];
1302 	struct inpcb *inp;
1303 	int error;
1304 	boolean_t mapped = FALSE;
1305 
1306 	error = priv_check(req->td, PRIV_ROOT);
1307 	if (error != 0)
1308 		return (error);
1309 	error = SYSCTL_IN(req, addrs, sizeof addrs);
1310 	if (error != 0)
1311 		return (error);
1312 	if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1313 		if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1314 			mapped = TRUE;
1315 		else
1316 			return (EINVAL);
1317 	}
1318 	crit_enter();
1319 	if (mapped) {
1320 		inp = in_pcblookup_hash(&tcbinfo[0],
1321 		    *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1322 		    addrs[1].sin6_port,
1323 		    *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1324 		    addrs[0].sin6_port,
1325 		    0, NULL);
1326 	} else {
1327 		inp = in6_pcblookup_hash(&tcbinfo[0],
1328 		    &addrs[1].sin6_addr, addrs[1].sin6_port,
1329 		    &addrs[0].sin6_addr, addrs[0].sin6_port,
1330 		    0, NULL);
1331 	}
1332 	if (inp == NULL || inp->inp_socket == NULL) {
1333 		error = ENOENT;
1334 		goto out;
1335 	}
1336 	error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1337 out:
1338 	crit_exit();
1339 	return (error);
1340 }
1341 
1342 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1343 	    0, 0,
1344 	    tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1345 #endif
1346 
1347 struct netmsg_tcp_notify {
1348 	struct netmsg	nm_nmsg;
1349 	void		(*nm_notify)(struct inpcb *, int);
1350 	struct in_addr	nm_faddr;
1351 	int		nm_arg;
1352 };
1353 
1354 static void
1355 tcp_notifyall_oncpu(struct netmsg *netmsg)
1356 {
1357 	struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1358 	int nextcpu;
1359 
1360 	in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1361 			nmsg->nm_arg, nmsg->nm_notify);
1362 
1363 	nextcpu = mycpuid + 1;
1364 	if (nextcpu < ncpus2)
1365 		lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1366 	else
1367 		lwkt_replymsg(&netmsg->nm_lmsg, 0);
1368 }
1369 
1370 void
1371 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1372 {
1373 	struct ip *ip = vip;
1374 	struct tcphdr *th;
1375 	struct in_addr faddr;
1376 	struct inpcb *inp;
1377 	struct tcpcb *tp;
1378 	void (*notify)(struct inpcb *, int) = tcp_notify;
1379 	tcp_seq icmpseq;
1380 	int arg, cpu;
1381 
1382 	if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1383 		return;
1384 	}
1385 
1386 	faddr = ((struct sockaddr_in *)sa)->sin_addr;
1387 	if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1388 		return;
1389 
1390 	arg = inetctlerrmap[cmd];
1391 	if (cmd == PRC_QUENCH) {
1392 		notify = tcp_quench;
1393 	} else if (icmp_may_rst &&
1394 		   (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1395 		    cmd == PRC_UNREACH_PORT ||
1396 		    cmd == PRC_TIMXCEED_INTRANS) &&
1397 		   ip != NULL) {
1398 		notify = tcp_drop_syn_sent;
1399 	} else if (cmd == PRC_MSGSIZE) {
1400 		struct icmp *icmp = (struct icmp *)
1401 		    ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1402 
1403 		arg = ntohs(icmp->icmp_nextmtu);
1404 		notify = tcp_mtudisc;
1405 	} else if (PRC_IS_REDIRECT(cmd)) {
1406 		ip = NULL;
1407 		notify = in_rtchange;
1408 	} else if (cmd == PRC_HOSTDEAD) {
1409 		ip = NULL;
1410 	}
1411 
1412 	if (ip != NULL) {
1413 		crit_enter();
1414 		th = (struct tcphdr *)((caddr_t)ip +
1415 				       (IP_VHL_HL(ip->ip_vhl) << 2));
1416 		cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1417 				  ip->ip_src.s_addr, th->th_sport);
1418 		inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1419 					ip->ip_src, th->th_sport, 0, NULL);
1420 		if ((inp != NULL) && (inp->inp_socket != NULL)) {
1421 			icmpseq = htonl(th->th_seq);
1422 			tp = intotcpcb(inp);
1423 			if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1424 			    SEQ_LT(icmpseq, tp->snd_max))
1425 				(*notify)(inp, arg);
1426 		} else {
1427 			struct in_conninfo inc;
1428 
1429 			inc.inc_fport = th->th_dport;
1430 			inc.inc_lport = th->th_sport;
1431 			inc.inc_faddr = faddr;
1432 			inc.inc_laddr = ip->ip_src;
1433 #ifdef INET6
1434 			inc.inc_isipv6 = 0;
1435 #endif
1436 			syncache_unreach(&inc, th);
1437 		}
1438 		crit_exit();
1439 	} else {
1440 		struct netmsg_tcp_notify nmsg;
1441 
1442 		KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1443 		netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1444 			    0, tcp_notifyall_oncpu);
1445 		nmsg.nm_faddr = faddr;
1446 		nmsg.nm_arg = arg;
1447 		nmsg.nm_notify = notify;
1448 
1449 		lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1450 	}
1451 }
1452 
1453 #ifdef INET6
1454 void
1455 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1456 {
1457 	struct tcphdr th;
1458 	void (*notify) (struct inpcb *, int) = tcp_notify;
1459 	struct ip6_hdr *ip6;
1460 	struct mbuf *m;
1461 	struct ip6ctlparam *ip6cp = NULL;
1462 	const struct sockaddr_in6 *sa6_src = NULL;
1463 	int off;
1464 	struct tcp_portonly {
1465 		u_int16_t th_sport;
1466 		u_int16_t th_dport;
1467 	} *thp;
1468 	int arg;
1469 
1470 	if (sa->sa_family != AF_INET6 ||
1471 	    sa->sa_len != sizeof(struct sockaddr_in6))
1472 		return;
1473 
1474 	arg = 0;
1475 	if (cmd == PRC_QUENCH)
1476 		notify = tcp_quench;
1477 	else if (cmd == PRC_MSGSIZE) {
1478 		struct ip6ctlparam *ip6cp = d;
1479 		struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1480 
1481 		arg = ntohl(icmp6->icmp6_mtu);
1482 		notify = tcp_mtudisc;
1483 	} else if (!PRC_IS_REDIRECT(cmd) &&
1484 		 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1485 		return;
1486 	}
1487 
1488 	/* if the parameter is from icmp6, decode it. */
1489 	if (d != NULL) {
1490 		ip6cp = (struct ip6ctlparam *)d;
1491 		m = ip6cp->ip6c_m;
1492 		ip6 = ip6cp->ip6c_ip6;
1493 		off = ip6cp->ip6c_off;
1494 		sa6_src = ip6cp->ip6c_src;
1495 	} else {
1496 		m = NULL;
1497 		ip6 = NULL;
1498 		off = 0;	/* fool gcc */
1499 		sa6_src = &sa6_any;
1500 	}
1501 
1502 	if (ip6 != NULL) {
1503 		struct in_conninfo inc;
1504 		/*
1505 		 * XXX: We assume that when IPV6 is non NULL,
1506 		 * M and OFF are valid.
1507 		 */
1508 
1509 		/* check if we can safely examine src and dst ports */
1510 		if (m->m_pkthdr.len < off + sizeof *thp)
1511 			return;
1512 
1513 		bzero(&th, sizeof th);
1514 		m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1515 
1516 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1517 		    (struct sockaddr *)ip6cp->ip6c_src,
1518 		    th.th_sport, cmd, arg, notify);
1519 
1520 		inc.inc_fport = th.th_dport;
1521 		inc.inc_lport = th.th_sport;
1522 		inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1523 		inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1524 		inc.inc_isipv6 = 1;
1525 		syncache_unreach(&inc, &th);
1526 	} else
1527 		in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1528 		    (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1529 }
1530 #endif
1531 
1532 /*
1533  * Following is where TCP initial sequence number generation occurs.
1534  *
1535  * There are two places where we must use initial sequence numbers:
1536  * 1.  In SYN-ACK packets.
1537  * 2.  In SYN packets.
1538  *
1539  * All ISNs for SYN-ACK packets are generated by the syncache.  See
1540  * tcp_syncache.c for details.
1541  *
1542  * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1543  * depends on this property.  In addition, these ISNs should be
1544  * unguessable so as to prevent connection hijacking.  To satisfy
1545  * the requirements of this situation, the algorithm outlined in
1546  * RFC 1948 is used to generate sequence numbers.
1547  *
1548  * Implementation details:
1549  *
1550  * Time is based off the system timer, and is corrected so that it
1551  * increases by one megabyte per second.  This allows for proper
1552  * recycling on high speed LANs while still leaving over an hour
1553  * before rollover.
1554  *
1555  * net.inet.tcp.isn_reseed_interval controls the number of seconds
1556  * between seeding of isn_secret.  This is normally set to zero,
1557  * as reseeding should not be necessary.
1558  *
1559  */
1560 
1561 #define	ISN_BYTES_PER_SECOND 1048576
1562 
1563 u_char isn_secret[32];
1564 int isn_last_reseed;
1565 MD5_CTX isn_ctx;
1566 
1567 tcp_seq
1568 tcp_new_isn(struct tcpcb *tp)
1569 {
1570 	u_int32_t md5_buffer[4];
1571 	tcp_seq new_isn;
1572 
1573 	/* Seed if this is the first use, reseed if requested. */
1574 	if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1575 	     (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1576 		< (u_int)ticks))) {
1577 		read_random_unlimited(&isn_secret, sizeof isn_secret);
1578 		isn_last_reseed = ticks;
1579 	}
1580 
1581 	/* Compute the md5 hash and return the ISN. */
1582 	MD5Init(&isn_ctx);
1583 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1584 	MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1585 #ifdef INET6
1586 	if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1587 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1588 			  sizeof(struct in6_addr));
1589 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1590 			  sizeof(struct in6_addr));
1591 	} else
1592 #endif
1593 	{
1594 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1595 			  sizeof(struct in_addr));
1596 		MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1597 			  sizeof(struct in_addr));
1598 	}
1599 	MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1600 	MD5Final((u_char *) &md5_buffer, &isn_ctx);
1601 	new_isn = (tcp_seq) md5_buffer[0];
1602 	new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1603 	return (new_isn);
1604 }
1605 
1606 /*
1607  * When a source quench is received, close congestion window
1608  * to one segment.  We will gradually open it again as we proceed.
1609  */
1610 void
1611 tcp_quench(struct inpcb *inp, int error)
1612 {
1613 	struct tcpcb *tp = intotcpcb(inp);
1614 
1615 	if (tp != NULL) {
1616 		tp->snd_cwnd = tp->t_maxseg;
1617 		tp->snd_wacked = 0;
1618 	}
1619 }
1620 
1621 /*
1622  * When a specific ICMP unreachable message is received and the
1623  * connection state is SYN-SENT, drop the connection.  This behavior
1624  * is controlled by the icmp_may_rst sysctl.
1625  */
1626 void
1627 tcp_drop_syn_sent(struct inpcb *inp, int error)
1628 {
1629 	struct tcpcb *tp = intotcpcb(inp);
1630 
1631 	if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1632 		tcp_drop(tp, error);
1633 }
1634 
1635 /*
1636  * When a `need fragmentation' ICMP is received, update our idea of the MSS
1637  * based on the new value in the route.  Also nudge TCP to send something,
1638  * since we know the packet we just sent was dropped.
1639  * This duplicates some code in the tcp_mss() function in tcp_input.c.
1640  */
1641 void
1642 tcp_mtudisc(struct inpcb *inp, int mtu)
1643 {
1644 	struct tcpcb *tp = intotcpcb(inp);
1645 	struct rtentry *rt;
1646 	struct socket *so = inp->inp_socket;
1647 	int maxopd, mss;
1648 #ifdef INET6
1649 	boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1650 #else
1651 	const boolean_t isipv6 = FALSE;
1652 #endif
1653 
1654 	if (tp == NULL)
1655 		return;
1656 
1657 	/*
1658 	 * If no MTU is provided in the ICMP message, use the
1659 	 * next lower likely value, as specified in RFC 1191.
1660 	 */
1661 	if (mtu == 0) {
1662 		int oldmtu;
1663 
1664 		oldmtu = tp->t_maxopd +
1665 		    (isipv6 ?
1666 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1667 		     sizeof(struct tcpiphdr));
1668 		mtu = ip_next_mtu(oldmtu, 0);
1669 	}
1670 
1671 	if (isipv6)
1672 		rt = tcp_rtlookup6(&inp->inp_inc);
1673 	else
1674 		rt = tcp_rtlookup(&inp->inp_inc);
1675 	if (rt != NULL) {
1676 		if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1677 			mtu = rt->rt_rmx.rmx_mtu;
1678 
1679 		maxopd = mtu -
1680 		    (isipv6 ?
1681 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1682 		     sizeof(struct tcpiphdr));
1683 
1684 		/*
1685 		 * XXX - The following conditional probably violates the TCP
1686 		 * spec.  The problem is that, since we don't know the
1687 		 * other end's MSS, we are supposed to use a conservative
1688 		 * default.  But, if we do that, then MTU discovery will
1689 		 * never actually take place, because the conservative
1690 		 * default is much less than the MTUs typically seen
1691 		 * on the Internet today.  For the moment, we'll sweep
1692 		 * this under the carpet.
1693 		 *
1694 		 * The conservative default might not actually be a problem
1695 		 * if the only case this occurs is when sending an initial
1696 		 * SYN with options and data to a host we've never talked
1697 		 * to before.  Then, they will reply with an MSS value which
1698 		 * will get recorded and the new parameters should get
1699 		 * recomputed.  For Further Study.
1700 		 */
1701 		if (rt->rt_rmx.rmx_mssopt  && rt->rt_rmx.rmx_mssopt < maxopd)
1702 			maxopd = rt->rt_rmx.rmx_mssopt;
1703 	} else
1704 		maxopd = mtu -
1705 		    (isipv6 ?
1706 		     sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1707 		     sizeof(struct tcpiphdr));
1708 
1709 	if (tp->t_maxopd <= maxopd)
1710 		return;
1711 	tp->t_maxopd = maxopd;
1712 
1713 	mss = maxopd;
1714 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1715 			   (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1716 		mss -= TCPOLEN_TSTAMP_APPA;
1717 
1718 	/* round down to multiple of MCLBYTES */
1719 #if	(MCLBYTES & (MCLBYTES - 1)) == 0    /* test if MCLBYTES power of 2 */
1720 	if (mss > MCLBYTES)
1721 		mss &= ~(MCLBYTES - 1);
1722 #else
1723 	if (mss > MCLBYTES)
1724 		mss = (mss / MCLBYTES) * MCLBYTES;
1725 #endif
1726 
1727 	if (so->so_snd.ssb_hiwat < mss)
1728 		mss = so->so_snd.ssb_hiwat;
1729 
1730 	tp->t_maxseg = mss;
1731 	tp->t_rtttime = 0;
1732 	tp->snd_nxt = tp->snd_una;
1733 	tcp_output(tp);
1734 	tcpstat.tcps_mturesent++;
1735 }
1736 
1737 /*
1738  * Look-up the routing entry to the peer of this inpcb.  If no route
1739  * is found and it cannot be allocated the return NULL.  This routine
1740  * is called by TCP routines that access the rmx structure and by tcp_mss
1741  * to get the interface MTU.
1742  */
1743 struct rtentry *
1744 tcp_rtlookup(struct in_conninfo *inc)
1745 {
1746 	struct route *ro = &inc->inc_route;
1747 
1748 	if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1749 		/* No route yet, so try to acquire one */
1750 		if (inc->inc_faddr.s_addr != INADDR_ANY) {
1751 			/*
1752 			 * unused portions of the structure MUST be zero'd
1753 			 * out because rtalloc() treats it as opaque data
1754 			 */
1755 			bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1756 			ro->ro_dst.sa_family = AF_INET;
1757 			ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1758 			((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1759 			    inc->inc_faddr;
1760 			rtalloc(ro);
1761 		}
1762 	}
1763 	return (ro->ro_rt);
1764 }
1765 
1766 #ifdef INET6
1767 struct rtentry *
1768 tcp_rtlookup6(struct in_conninfo *inc)
1769 {
1770 	struct route_in6 *ro6 = &inc->inc6_route;
1771 
1772 	if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1773 		/* No route yet, so try to acquire one */
1774 		if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1775 			/*
1776 			 * unused portions of the structure MUST be zero'd
1777 			 * out because rtalloc() treats it as opaque data
1778 			 */
1779 			bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1780 			ro6->ro_dst.sin6_family = AF_INET6;
1781 			ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1782 			ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1783 			rtalloc((struct route *)ro6);
1784 		}
1785 	}
1786 	return (ro6->ro_rt);
1787 }
1788 #endif
1789 
1790 #ifdef IPSEC
1791 /* compute ESP/AH header size for TCP, including outer IP header. */
1792 size_t
1793 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1794 {
1795 	struct inpcb *inp;
1796 	struct mbuf *m;
1797 	size_t hdrsiz;
1798 	struct ip *ip;
1799 	struct tcphdr *th;
1800 
1801 	if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1802 		return (0);
1803 	MGETHDR(m, MB_DONTWAIT, MT_DATA);
1804 	if (!m)
1805 		return (0);
1806 
1807 #ifdef INET6
1808 	if (inp->inp_vflag & INP_IPV6) {
1809 		struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1810 
1811 		th = (struct tcphdr *)(ip6 + 1);
1812 		m->m_pkthdr.len = m->m_len =
1813 		    sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1814 		tcp_fillheaders(tp, ip6, th);
1815 		hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1816 	} else
1817 #endif
1818 	{
1819 		ip = mtod(m, struct ip *);
1820 		th = (struct tcphdr *)(ip + 1);
1821 		m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1822 		tcp_fillheaders(tp, ip, th);
1823 		hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1824 	}
1825 
1826 	m_free(m);
1827 	return (hdrsiz);
1828 }
1829 #endif
1830 
1831 /*
1832  * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1833  *
1834  * This code attempts to calculate the bandwidth-delay product as a
1835  * means of determining the optimal window size to maximize bandwidth,
1836  * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1837  * routers.  This code also does a fairly good job keeping RTTs in check
1838  * across slow links like modems.  We implement an algorithm which is very
1839  * similar (but not meant to be) TCP/Vegas.  The code operates on the
1840  * transmitter side of a TCP connection and so only effects the transmit
1841  * side of the connection.
1842  *
1843  * BACKGROUND:  TCP makes no provision for the management of buffer space
1844  * at the end points or at the intermediate routers and switches.  A TCP
1845  * stream, whether using NewReno or not, will eventually buffer as
1846  * many packets as it is able and the only reason this typically works is
1847  * due to the fairly small default buffers made available for a connection
1848  * (typicaly 16K or 32K).  As machines use larger windows and/or window
1849  * scaling it is now fairly easy for even a single TCP connection to blow-out
1850  * all available buffer space not only on the local interface, but on
1851  * intermediate routers and switches as well.  NewReno makes a misguided
1852  * attempt to 'solve' this problem by waiting for an actual failure to occur,
1853  * then backing off, then steadily increasing the window again until another
1854  * failure occurs, ad-infinitum.  This results in terrible oscillation that
1855  * is only made worse as network loads increase and the idea of intentionally
1856  * blowing out network buffers is, frankly, a terrible way to manage network
1857  * resources.
1858  *
1859  * It is far better to limit the transmit window prior to the failure
1860  * condition being achieved.  There are two general ways to do this:  First
1861  * you can 'scan' through different transmit window sizes and locate the
1862  * point where the RTT stops increasing, indicating that you have filled the
1863  * pipe, then scan backwards until you note that RTT stops decreasing, then
1864  * repeat ad-infinitum.  This method works in principle but has severe
1865  * implementation issues due to RTT variances, timer granularity, and
1866  * instability in the algorithm which can lead to many false positives and
1867  * create oscillations as well as interact badly with other TCP streams
1868  * implementing the same algorithm.
1869  *
1870  * The second method is to limit the window to the bandwidth delay product
1871  * of the link.  This is the method we implement.  RTT variances and our
1872  * own manipulation of the congestion window, bwnd, can potentially
1873  * destabilize the algorithm.  For this reason we have to stabilize the
1874  * elements used to calculate the window.  We do this by using the minimum
1875  * observed RTT, the long term average of the observed bandwidth, and
1876  * by adding two segments worth of slop.  It isn't perfect but it is able
1877  * to react to changing conditions and gives us a very stable basis on
1878  * which to extend the algorithm.
1879  */
1880 void
1881 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1882 {
1883 	u_long bw;
1884 	u_long bwnd;
1885 	int save_ticks;
1886 	int delta_ticks;
1887 
1888 	/*
1889 	 * If inflight_enable is disabled in the middle of a tcp connection,
1890 	 * make sure snd_bwnd is effectively disabled.
1891 	 */
1892 	if (!tcp_inflight_enable) {
1893 		tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1894 		tp->snd_bandwidth = 0;
1895 		return;
1896 	}
1897 
1898 	/*
1899 	 * Validate the delta time.  If a connection is new or has been idle
1900 	 * a long time we have to reset the bandwidth calculator.
1901 	 */
1902 	save_ticks = ticks;
1903 	delta_ticks = save_ticks - tp->t_bw_rtttime;
1904 	if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1905 		tp->t_bw_rtttime = ticks;
1906 		tp->t_bw_rtseq = ack_seq;
1907 		if (tp->snd_bandwidth == 0)
1908 			tp->snd_bandwidth = tcp_inflight_min;
1909 		return;
1910 	}
1911 	if (delta_ticks == 0)
1912 		return;
1913 
1914 	/*
1915 	 * Sanity check, plus ignore pure window update acks.
1916 	 */
1917 	if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1918 		return;
1919 
1920 	/*
1921 	 * Figure out the bandwidth.  Due to the tick granularity this
1922 	 * is a very rough number and it MUST be averaged over a fairly
1923 	 * long period of time.  XXX we need to take into account a link
1924 	 * that is not using all available bandwidth, but for now our
1925 	 * slop will ramp us up if this case occurs and the bandwidth later
1926 	 * increases.
1927 	 */
1928 	bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1929 	tp->t_bw_rtttime = save_ticks;
1930 	tp->t_bw_rtseq = ack_seq;
1931 	bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1932 
1933 	tp->snd_bandwidth = bw;
1934 
1935 	/*
1936 	 * Calculate the semi-static bandwidth delay product, plus two maximal
1937 	 * segments.  The additional slop puts us squarely in the sweet
1938 	 * spot and also handles the bandwidth run-up case.  Without the
1939 	 * slop we could be locking ourselves into a lower bandwidth.
1940 	 *
1941 	 * Situations Handled:
1942 	 *	(1) Prevents over-queueing of packets on LANs, especially on
1943 	 *	    high speed LANs, allowing larger TCP buffers to be
1944 	 *	    specified, and also does a good job preventing
1945 	 *	    over-queueing of packets over choke points like modems
1946 	 *	    (at least for the transmit side).
1947 	 *
1948 	 *	(2) Is able to handle changing network loads (bandwidth
1949 	 *	    drops so bwnd drops, bandwidth increases so bwnd
1950 	 *	    increases).
1951 	 *
1952 	 *	(3) Theoretically should stabilize in the face of multiple
1953 	 *	    connections implementing the same algorithm (this may need
1954 	 *	    a little work).
1955 	 *
1956 	 *	(4) Stability value (defaults to 20 = 2 maximal packets) can
1957 	 *	    be adjusted with a sysctl but typically only needs to be on
1958 	 *	    very slow connections.  A value no smaller then 5 should
1959 	 *	    be used, but only reduce this default if you have no other
1960 	 *	    choice.
1961 	 */
1962 
1963 #define	USERTT	((tp->t_srtt + tp->t_rttbest) / 2)
1964 	bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1965 	       tcp_inflight_stab * (int)tp->t_maxseg / 10;
1966 #undef USERTT
1967 
1968 	if (tcp_inflight_debug > 0) {
1969 		static int ltime;
1970 		if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1971 			ltime = ticks;
1972 			kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1973 				tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1974 		}
1975 	}
1976 	if ((long)bwnd < tcp_inflight_min)
1977 		bwnd = tcp_inflight_min;
1978 	if (bwnd > tcp_inflight_max)
1979 		bwnd = tcp_inflight_max;
1980 	if ((long)bwnd < tp->t_maxseg * 2)
1981 		bwnd = tp->t_maxseg * 2;
1982 	tp->snd_bwnd = bwnd;
1983 }
1984 
1985 #ifdef TCP_SIGNATURE
1986 /*
1987  * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1988  *
1989  * We do this over ip, tcphdr, segment data, and the key in the SADB.
1990  * When called from tcp_input(), we can be sure that th_sum has been
1991  * zeroed out and verified already.
1992  *
1993  * Return 0 if successful, otherwise return -1.
1994  *
1995  * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1996  * search with the destination IP address, and a 'magic SPI' to be
1997  * determined by the application. This is hardcoded elsewhere to 1179
1998  * right now. Another branch of this code exists which uses the SPD to
1999  * specify per-application flows but it is unstable.
2000  */
2001 int
2002 tcpsignature_compute(
2003 	struct mbuf *m,		/* mbuf chain */
2004 	int len,		/* length of TCP data */
2005 	int optlen,		/* length of TCP options */
2006 	u_char *buf,		/* storage for MD5 digest */
2007 	u_int direction)	/* direction of flow */
2008 {
2009 	struct ippseudo ippseudo;
2010 	MD5_CTX ctx;
2011 	int doff;
2012 	struct ip *ip;
2013 	struct ipovly *ipovly;
2014 	struct secasvar *sav;
2015 	struct tcphdr *th;
2016 #ifdef INET6
2017 	struct ip6_hdr *ip6;
2018 	struct in6_addr in6;
2019 	uint32_t plen;
2020 	uint16_t nhdr;
2021 #endif /* INET6 */
2022 	u_short savecsum;
2023 
2024 	KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2025 	KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2026 	/*
2027 	 * Extract the destination from the IP header in the mbuf.
2028 	 */
2029 	ip = mtod(m, struct ip *);
2030 #ifdef INET6
2031 	ip6 = NULL;     /* Make the compiler happy. */
2032 #endif /* INET6 */
2033 	/*
2034 	 * Look up an SADB entry which matches the address found in
2035 	 * the segment.
2036 	 */
2037 	switch (IP_VHL_V(ip->ip_vhl)) {
2038 	case IPVERSION:
2039 		sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2040 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2041 		break;
2042 #ifdef INET6
2043 	case (IPV6_VERSION >> 4):
2044 		ip6 = mtod(m, struct ip6_hdr *);
2045 		sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2046 				IPPROTO_TCP, htonl(TCP_SIG_SPI));
2047 		break;
2048 #endif /* INET6 */
2049 	default:
2050 		return (EINVAL);
2051 		/* NOTREACHED */
2052 		break;
2053 	}
2054 	if (sav == NULL) {
2055 		kprintf("%s: SADB lookup failed\n", __func__);
2056 		return (EINVAL);
2057 	}
2058 	MD5Init(&ctx);
2059 
2060 	/*
2061 	 * Step 1: Update MD5 hash with IP pseudo-header.
2062 	 *
2063 	 * XXX The ippseudo header MUST be digested in network byte order,
2064 	 * or else we'll fail the regression test. Assume all fields we've
2065 	 * been doing arithmetic on have been in host byte order.
2066 	 * XXX One cannot depend on ipovly->ih_len here. When called from
2067 	 * tcp_output(), the underlying ip_len member has not yet been set.
2068 	 */
2069 	switch (IP_VHL_V(ip->ip_vhl)) {
2070 	case IPVERSION:
2071 		ipovly = (struct ipovly *)ip;
2072 		ippseudo.ippseudo_src = ipovly->ih_src;
2073 		ippseudo.ippseudo_dst = ipovly->ih_dst;
2074 		ippseudo.ippseudo_pad = 0;
2075 		ippseudo.ippseudo_p = IPPROTO_TCP;
2076 		ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2077 		MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2078 		th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2079 		doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2080 		break;
2081 #ifdef INET6
2082 	/*
2083 	 * RFC 2385, 2.0  Proposal
2084 	 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2085 	 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2086 	 * extended next header value (to form 32 bits), and 32-bit segment
2087 	 * length.
2088 	 * Note: Upper-Layer Packet Length comes before Next Header.
2089 	 */
2090 	case (IPV6_VERSION >> 4):
2091 		in6 = ip6->ip6_src;
2092 		in6_clearscope(&in6);
2093 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2094 		in6 = ip6->ip6_dst;
2095 		in6_clearscope(&in6);
2096 		MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2097 		plen = htonl(len + sizeof(struct tcphdr) + optlen);
2098 		MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2099 		nhdr = 0;
2100 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2101 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2102 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2103 		nhdr = IPPROTO_TCP;
2104 		MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2105 		th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2106 		doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2107 		break;
2108 #endif /* INET6 */
2109 	default:
2110 		return (EINVAL);
2111 		/* NOTREACHED */
2112 		break;
2113 	}
2114 	/*
2115 	 * Step 2: Update MD5 hash with TCP header, excluding options.
2116 	 * The TCP checksum must be set to zero.
2117 	 */
2118 	savecsum = th->th_sum;
2119 	th->th_sum = 0;
2120 	MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2121 	th->th_sum = savecsum;
2122 	/*
2123 	 * Step 3: Update MD5 hash with TCP segment data.
2124 	 *         Use m_apply() to avoid an early m_pullup().
2125 	 */
2126 	if (len > 0)
2127 		m_apply(m, doff, len, tcpsignature_apply, &ctx);
2128 	/*
2129 	 * Step 4: Update MD5 hash with shared secret.
2130 	 */
2131 	MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2132 	MD5Final(buf, &ctx);
2133 	key_sa_recordxfer(sav, m);
2134 	key_freesav(sav);
2135 	return (0);
2136 }
2137 
2138 int
2139 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2140 {
2141 
2142 	MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2143 	return (0);
2144 }
2145 #endif /* TCP_SIGNATURE */
2146