2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
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.
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
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 $
70 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
85 #include <sys/domain.h>
89 #include <sys/socket.h>
90 #include <sys/socketvar.h>
91 #include <sys/protosw.h>
92 #include <sys/random.h>
93 #include <sys/in_cksum.h>
96 #include <net/route.h>
98 #include <net/netisr.h>
101 #include <netinet/in.h>
102 #include <netinet/in_systm.h>
103 #include <netinet/ip.h>
104 #include <netinet/ip6.h>
105 #include <netinet/in_pcb.h>
106 #include <netinet6/in6_pcb.h>
107 #include <netinet/in_var.h>
108 #include <netinet/ip_var.h>
109 #include <netinet6/ip6_var.h>
110 #include <netinet/ip_icmp.h>
112 #include <netinet/icmp6.h>
114 #include <netinet/tcp.h>
115 #include <netinet/tcp_fsm.h>
116 #include <netinet/tcp_seq.h>
117 #include <netinet/tcp_timer.h>
118 #include <netinet/tcp_timer2.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
123 #include <netinet/tcp_debug.h>
125 #include <netinet6/ip6protosw.h>
128 #include <netinet6/ipsec.h>
129 #include <netproto/key/key.h>
131 #include <netinet6/ipsec6.h>
136 #include <netproto/ipsec/ipsec.h>
138 #include <netproto/ipsec/ipsec6.h>
144 #include <machine/smp.h>
146 #include <sys/msgport2.h>
147 #include <sys/mplock2.h>
148 #include <net/netmsg2.h>
150 #if !defined(KTR_TCP)
151 #define KTR_TCP KTR_ALL
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)
161 #define TCP_IW_MAXSEGS_DFLT 4
162 #define TCP_IW_CAPSEGS_DFLT 3
164 struct inpcbinfo tcbinfo[MAXCPU];
165 struct tcpcbackqhead tcpcbackq[MAXCPU];
167 static struct lwkt_token tcp_port_token =
168 LWKT_TOKEN_INITIALIZER(tcp_port_token);
170 int tcp_mssdflt = TCP_MSS;
171 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
172 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
175 int tcp_v6mssdflt = TCP6_MSS;
176 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
177 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
181 * Minimum MSS we accept and use. This prevents DoS attacks where
182 * we are forced to a ridiculous low MSS like 20 and send hundreds
183 * of packets instead of one. The effect scales with the available
184 * bandwidth and quickly saturates the CPU and network interface
185 * with packet generation and sending. Set to zero to disable MINMSS
186 * checking. This setting prevents us from sending too small packets.
188 int tcp_minmss = TCP_MINMSS;
189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
190 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
193 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
194 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
195 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
198 int tcp_do_rfc1323 = 1;
199 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
200 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
202 static int tcp_tcbhashsize = 0;
203 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
204 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
206 static int do_tcpdrain = 1;
207 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
208 "Enable tcp_drain routine for extra help when low on mbufs");
210 static int icmp_may_rst = 1;
211 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
212 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
214 static int tcp_isn_reseed_interval = 0;
215 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
216 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
219 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
220 * by default, but with generous values which should allow maximal
221 * bandwidth. In particular, the slop defaults to 50 (5 packets).
223 * The reason for doing this is that the limiter is the only mechanism we
224 * have which seems to do a really good job preventing receiver RX rings
225 * on network interfaces from getting blown out. Even though GigE/10GigE
226 * is supposed to flow control it looks like either it doesn't actually
227 * do it or Open Source drivers do not properly enable it.
229 * People using the limiter to reduce bottlenecks on slower WAN connections
230 * should set the slop to 20 (2 packets).
232 static int tcp_inflight_enable = 1;
233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
234 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
236 static int tcp_inflight_debug = 0;
237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
238 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
240 static int tcp_inflight_min = 6144;
241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
242 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
244 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
246 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
248 static int tcp_inflight_stab = 50;
249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
250 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
252 static int tcp_do_rfc3390 = 1;
253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
255 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
257 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
258 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW,
259 &tcp_iw_maxsegs, 0, "TCP IW segments max");
261 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
262 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW,
263 &tcp_iw_capsegs, 0, "TCP IW segments");
265 int tcp_low_rtobase = 1;
266 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW,
267 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)");
269 static int tcp_do_ncr = 1;
270 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW,
271 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)");
273 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
274 static struct malloc_pipe tcptemp_mpipe;
276 static void tcp_willblock(void);
277 static void tcp_notify (struct inpcb *, int);
279 struct tcp_stats tcpstats_percpu[MAXCPU];
282 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
286 for (cpu = 0; cpu < ncpus; ++cpu) {
287 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
288 sizeof(struct tcp_stats))))
290 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
291 sizeof(struct tcp_stats))))
297 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
298 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
300 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
301 &tcpstat, tcp_stats, "TCP statistics");
305 * Target size of TCP PCB hash tables. Must be a power of two.
307 * Note that this can be overridden by the kernel environment
308 * variable net.inet.tcp.tcbhashsize
311 #define TCBHASHSIZE 512
315 * This is the actual shape of what we allocate using the zone
316 * allocator. Doing it this way allows us to protect both structures
317 * using the same generation count, and also eliminates the overhead
318 * of allocating tcpcbs separately. By hiding the structure here,
319 * we avoid changing most of the rest of the code (although it needs
320 * to be changed, eventually, for greater efficiency).
323 #define ALIGNM1 (ALIGNMENT - 1)
327 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
330 struct tcp_callout inp_tp_rexmt;
331 struct tcp_callout inp_tp_persist;
332 struct tcp_callout inp_tp_keep;
333 struct tcp_callout inp_tp_2msl;
334 struct tcp_callout inp_tp_delack;
335 struct netmsg_tcp_timer inp_tp_timermsg;
346 struct inpcbporthead *porthashbase;
347 struct inpcbinfo *ticb;
349 int hashsize = TCBHASHSIZE;
353 * note: tcptemp is used for keepalives, and it is ok for an
354 * allocation to fail so do not specify MPF_INT.
356 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
357 25, -1, 0, NULL, NULL, NULL);
359 tcp_delacktime = TCPTV_DELACK;
360 tcp_keepinit = TCPTV_KEEP_INIT;
361 tcp_keepidle = TCPTV_KEEP_IDLE;
362 tcp_keepintvl = TCPTV_KEEPINTVL;
363 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
365 tcp_rexmit_min = TCPTV_MIN;
366 tcp_rexmit_slop = TCPTV_CPU_VAR;
368 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
369 if (!powerof2(hashsize)) {
370 kprintf("WARNING: TCB hash size not a power of 2\n");
371 hashsize = 512; /* safe default */
373 tcp_tcbhashsize = hashsize;
374 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
376 for (cpu = 0; cpu < ncpus2; cpu++) {
377 ticb = &tcbinfo[cpu];
378 in_pcbinfo_init(ticb);
380 ticb->hashbase = hashinit(hashsize, M_PCB,
382 ticb->porthashbase = porthashbase;
383 ticb->porthashmask = porthashmask;
384 ticb->porttoken = &tcp_port_token;
386 ticb->porthashbase = hashinit(hashsize, M_PCB,
387 &ticb->porthashmask);
389 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
390 &ticb->wildcardhashmask);
391 ticb->ipi_size = sizeof(struct inp_tp);
392 TAILQ_INIT(&tcpcbackq[cpu]);
395 tcp_reass_maxseg = nmbclusters / 16;
396 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
399 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
401 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
403 if (max_protohdr < TCP_MINPROTOHDR)
404 max_protohdr = TCP_MINPROTOHDR;
405 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
407 #undef TCP_MINPROTOHDR
410 * Initialize TCP statistics counters for each CPU.
413 for (cpu = 0; cpu < ncpus; ++cpu) {
414 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
417 bzero(&tcpstat, sizeof(struct tcp_stats));
421 netisr_register_rollup(tcp_willblock);
428 int cpu = mycpu->gd_cpuid;
430 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
431 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
432 tp->t_flags &= ~TF_ONOUTPUTQ;
433 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
439 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
440 * tcp_template used to store this data in mbufs, but we now recopy it out
441 * of the tcpcb each time to conserve mbufs.
444 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr, boolean_t tso)
446 struct inpcb *inp = tp->t_inpcb;
447 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
450 if (inp->inp_vflag & INP_IPV6) {
453 ip6 = (struct ip6_hdr *)ip_ptr;
454 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
455 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
456 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
457 (IPV6_VERSION & IPV6_VERSION_MASK);
458 ip6->ip6_nxt = IPPROTO_TCP;
459 ip6->ip6_plen = sizeof(struct tcphdr);
460 ip6->ip6_src = inp->in6p_laddr;
461 ip6->ip6_dst = inp->in6p_faddr;
466 struct ip *ip = (struct ip *) ip_ptr;
469 ip->ip_vhl = IP_VHL_BORING;
476 ip->ip_p = IPPROTO_TCP;
477 ip->ip_src = inp->inp_laddr;
478 ip->ip_dst = inp->inp_faddr;
481 plen = htons(IPPROTO_TCP);
483 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP);
484 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
485 ip->ip_dst.s_addr, plen);
488 tcp_hdr->th_sport = inp->inp_lport;
489 tcp_hdr->th_dport = inp->inp_fport;
494 tcp_hdr->th_flags = 0;
500 * Create template to be used to send tcp packets on a connection.
501 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
502 * use for this function is in keepalives, which use tcp_respond.
505 tcp_maketemplate(struct tcpcb *tp)
509 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
511 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE);
516 tcp_freetemplate(struct tcptemp *tmp)
518 mpipe_free(&tcptemp_mpipe, tmp);
522 * Send a single message to the TCP at address specified by
523 * the given TCP/IP header. If m == NULL, then we make a copy
524 * of the tcpiphdr at ti and send directly to the addressed host.
525 * This is used to force keep alive messages out using the TCP
526 * template for a connection. If flags are given then we send
527 * a message back to the TCP which originated the * segment ti,
528 * and discard the mbuf containing it and any other attached mbufs.
530 * In any case the ack and sequence number of the transmitted
531 * segment are as specified by the parameters.
533 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
536 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
537 tcp_seq ack, tcp_seq seq, int flags)
541 struct route *ro = NULL;
543 struct ip *ip = ipgen;
546 struct route_in6 *ro6 = NULL;
547 struct route_in6 sro6;
548 struct ip6_hdr *ip6 = ipgen;
549 boolean_t use_tmpro = TRUE;
551 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
553 const boolean_t isipv6 = FALSE;
557 if (!(flags & TH_RST)) {
558 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
561 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
562 win = (long)TCP_MAXWIN << tp->rcv_scale;
565 * Don't use the route cache of a listen socket,
566 * it is not MPSAFE; use temporary route cache.
568 if (tp->t_state != TCPS_LISTEN) {
570 ro6 = &tp->t_inpcb->in6p_route;
572 ro = &tp->t_inpcb->inp_route;
579 bzero(ro6, sizeof *ro6);
582 bzero(ro, sizeof *ro);
586 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
590 m->m_data += max_linkhdr;
592 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
593 ip6 = mtod(m, struct ip6_hdr *);
594 nth = (struct tcphdr *)(ip6 + 1);
596 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
597 ip = mtod(m, struct ip *);
598 nth = (struct tcphdr *)(ip + 1);
600 bcopy(th, nth, sizeof(struct tcphdr));
605 m->m_data = (caddr_t)ipgen;
606 /* m_len is set later */
608 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
610 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
611 nth = (struct tcphdr *)(ip6 + 1);
613 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
614 nth = (struct tcphdr *)(ip + 1);
618 * this is usually a case when an extension header
619 * exists between the IPv6 header and the
622 nth->th_sport = th->th_sport;
623 nth->th_dport = th->th_dport;
625 xchg(nth->th_dport, nth->th_sport, n_short);
630 ip6->ip6_vfc = IPV6_VERSION;
631 ip6->ip6_nxt = IPPROTO_TCP;
632 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
633 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
635 tlen += sizeof(struct tcpiphdr);
637 ip->ip_ttl = ip_defttl;
640 m->m_pkthdr.len = tlen;
641 m->m_pkthdr.rcvif = NULL;
642 nth->th_seq = htonl(seq);
643 nth->th_ack = htonl(ack);
645 nth->th_off = sizeof(struct tcphdr) >> 2;
646 nth->th_flags = flags;
648 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
650 nth->th_win = htons((u_short)win);
654 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
655 sizeof(struct ip6_hdr),
656 tlen - sizeof(struct ip6_hdr));
657 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
658 (ro6 && ro6->ro_rt) ?
659 ro6->ro_rt->rt_ifp : NULL);
661 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
662 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
663 m->m_pkthdr.csum_flags = CSUM_TCP;
664 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
667 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
668 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
671 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
672 tp ? tp->t_inpcb : NULL);
673 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
678 ipflags |= IP_DEBUGROUTE;
679 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
680 if ((ro == &sro) && (ro->ro_rt != NULL)) {
688 * Create a new TCP control block, making an
689 * empty reassembly queue and hooking it to the argument
690 * protocol control block. The `inp' parameter must have
691 * come from the zone allocator set up in tcp_init().
694 tcp_newtcpcb(struct inpcb *inp)
699 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
701 const boolean_t isipv6 = FALSE;
704 it = (struct inp_tp *)inp;
706 bzero(tp, sizeof(struct tcpcb));
707 TAILQ_INIT(&tp->t_segq);
708 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
709 tp->t_rxtthresh = tcprexmtthresh;
711 /* Set up our timeouts. */
712 tp->tt_rexmt = &it->inp_tp_rexmt;
713 tp->tt_persist = &it->inp_tp_persist;
714 tp->tt_keep = &it->inp_tp_keep;
715 tp->tt_2msl = &it->inp_tp_2msl;
716 tp->tt_delack = &it->inp_tp_delack;
720 * Zero out timer message. We don't create it here,
721 * since the current CPU may not be the owner of this
724 tp->tt_msg = &it->inp_tp_timermsg;
725 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
727 tp->t_keepinit = tcp_keepinit;
728 tp->t_keepidle = tcp_keepidle;
729 tp->t_keepintvl = tcp_keepintvl;
730 tp->t_keepcnt = tcp_keepcnt;
731 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
734 tp->t_flags |= TF_NCR;
736 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP);
738 tp->t_inpcb = inp; /* XXX */
739 tp->t_state = TCPS_CLOSED;
741 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
742 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
743 * reasonable initial retransmit time.
745 tp->t_srtt = TCPTV_SRTTBASE;
747 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
748 tp->t_rttmin = tcp_rexmit_min;
749 tp->t_rxtcur = TCPTV_RTOBASE;
750 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
751 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
752 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
753 tp->snd_last = ticks;
754 tp->t_rcvtime = ticks;
756 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
757 * because the socket may be bound to an IPv6 wildcard address,
758 * which may match an IPv4-mapped IPv6 address.
760 inp->inp_ip_ttl = ip_defttl;
762 tcp_sack_tcpcb_init(tp);
763 return (tp); /* XXX */
767 * Drop a TCP connection, reporting the specified error.
768 * If connection is synchronized, then send a RST to peer.
771 tcp_drop(struct tcpcb *tp, int error)
773 struct socket *so = tp->t_inpcb->inp_socket;
775 if (TCPS_HAVERCVDSYN(tp->t_state)) {
776 tp->t_state = TCPS_CLOSED;
778 tcpstat.tcps_drops++;
780 tcpstat.tcps_conndrops++;
781 if (error == ETIMEDOUT && tp->t_softerror)
782 error = tp->t_softerror;
783 so->so_error = error;
784 return (tcp_close(tp));
789 struct netmsg_listen_detach {
790 struct netmsg_base base;
795 tcp_listen_detach_handler(netmsg_t msg)
797 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
798 struct tcpcb *tp = nmsg->nm_tp;
799 int cpu = mycpuid, nextcpu;
801 if (tp->t_flags & TF_LISTEN)
802 syncache_destroy(tp);
804 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
807 if (nextcpu < ncpus2)
808 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg);
810 lwkt_replymsg(&nmsg->base.lmsg, 0);
816 * Close a TCP control block:
817 * discard all space held by the tcp
818 * discard internet protocol block
819 * wake up any sleepers
822 tcp_close(struct tcpcb *tp)
825 struct inpcb *inp = tp->t_inpcb;
826 struct socket *so = inp->inp_socket;
828 boolean_t dosavessthresh;
830 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
831 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
833 const boolean_t isipv6 = FALSE;
838 * INP_WILDCARD_MP indicates that listen(2) has been called on
839 * this socket. This implies:
840 * - A wildcard inp's hash is replicated for each protocol thread.
841 * - Syncache for this inp grows independently in each protocol
843 * - There is more than one cpu
845 * We have to chain a message to the rest of the protocol threads
846 * to cleanup the wildcard hash and the syncache. The cleanup
847 * in the current protocol thread is defered till the end of this
851 * After cleanup the inp's hash and syncache entries, this inp will
852 * no longer be available to the rest of the protocol threads, so we
853 * are safe to whack the inp in the following code.
855 if (inp->inp_flags & INP_WILDCARD_MP) {
856 struct netmsg_listen_detach nmsg;
858 KKASSERT(so->so_port == cpu_portfn(0));
859 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
860 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
862 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
863 MSGF_PRIORITY, tcp_listen_detach_handler);
865 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0);
867 inp->inp_flags &= ~INP_WILDCARD_MP;
871 KKASSERT(tp->t_state != TCPS_TERMINATING);
872 tp->t_state = TCPS_TERMINATING;
875 * Make sure that all of our timers are stopped before we
876 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
877 * timers are never used. If timer message is never created
878 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
880 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
881 tcp_callout_stop(tp, tp->tt_rexmt);
882 tcp_callout_stop(tp, tp->tt_persist);
883 tcp_callout_stop(tp, tp->tt_keep);
884 tcp_callout_stop(tp, tp->tt_2msl);
885 tcp_callout_stop(tp, tp->tt_delack);
888 if (tp->t_flags & TF_ONOUTPUTQ) {
889 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
890 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
891 tp->t_flags &= ~TF_ONOUTPUTQ;
895 * If we got enough samples through the srtt filter,
896 * save the rtt and rttvar in the routing entry.
897 * 'Enough' is arbitrarily defined as the 16 samples.
898 * 16 samples is enough for the srtt filter to converge
899 * to within 5% of the correct value; fewer samples and
900 * we could save a very bogus rtt.
902 * Don't update the default route's characteristics and don't
903 * update anything that the user "locked".
905 if (tp->t_rttupdated >= 16) {
909 struct sockaddr_in6 *sin6;
911 if ((rt = inp->in6p_route.ro_rt) == NULL)
913 sin6 = (struct sockaddr_in6 *)rt_key(rt);
914 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
917 if ((rt = inp->inp_route.ro_rt) == NULL ||
918 ((struct sockaddr_in *)rt_key(rt))->
919 sin_addr.s_addr == INADDR_ANY)
922 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
923 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
924 if (rt->rt_rmx.rmx_rtt && i)
926 * filter this update to half the old & half
927 * the new values, converting scale.
928 * See route.h and tcp_var.h for a
929 * description of the scaling constants.
932 (rt->rt_rmx.rmx_rtt + i) / 2;
934 rt->rt_rmx.rmx_rtt = i;
935 tcpstat.tcps_cachedrtt++;
937 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
939 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
940 if (rt->rt_rmx.rmx_rttvar && i)
941 rt->rt_rmx.rmx_rttvar =
942 (rt->rt_rmx.rmx_rttvar + i) / 2;
944 rt->rt_rmx.rmx_rttvar = i;
945 tcpstat.tcps_cachedrttvar++;
948 * The old comment here said:
949 * update the pipelimit (ssthresh) if it has been updated
950 * already or if a pipesize was specified & the threshhold
951 * got below half the pipesize. I.e., wait for bad news
952 * before we start updating, then update on both good
955 * But we want to save the ssthresh even if no pipesize is
956 * specified explicitly in the route, because such
957 * connections still have an implicit pipesize specified
958 * by the global tcp_sendspace. In the absence of a reliable
959 * way to calculate the pipesize, it will have to do.
961 i = tp->snd_ssthresh;
962 if (rt->rt_rmx.rmx_sendpipe != 0)
963 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
965 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
966 if (dosavessthresh ||
967 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
968 (rt->rt_rmx.rmx_ssthresh != 0))) {
970 * convert the limit from user data bytes to
971 * packets then to packet data bytes.
973 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
978 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
979 sizeof(struct tcpiphdr));
980 if (rt->rt_rmx.rmx_ssthresh)
981 rt->rt_rmx.rmx_ssthresh =
982 (rt->rt_rmx.rmx_ssthresh + i) / 2;
984 rt->rt_rmx.rmx_ssthresh = i;
985 tcpstat.tcps_cachedssthresh++;
990 /* free the reassembly queue, if any */
991 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) {
992 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
995 atomic_add_int(&tcp_reass_qsize, -1);
997 /* throw away SACK blocks in scoreboard*/
999 tcp_sack_destroy(&tp->scb);
1001 inp->inp_ppcb = NULL;
1002 soisdisconnected(so);
1003 /* note: pcb detached later on */
1005 tcp_destroy_timermsg(tp);
1007 if (tp->t_flags & TF_LISTEN)
1008 syncache_destroy(tp);
1012 * pcbdetach removes any wildcard hash entry on the current CPU.
1021 tcpstat.tcps_closed++;
1025 static __inline void
1026 tcp_drain_oncpu(struct inpcbhead *head)
1028 struct inpcb *marker;
1031 struct tseg_qent *te;
1034 * Allows us to block while running the list
1036 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1037 marker->inp_flags |= INP_PLACEMARKER;
1038 LIST_INSERT_HEAD(head, marker, inp_list);
1040 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1041 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1042 (tcpb = intotcpcb(inpb)) != NULL &&
1043 (te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) {
1044 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q);
1045 if (te->tqe_th->th_flags & TH_FIN)
1046 tcpb->t_flags &= ~TF_QUEDFIN;
1049 atomic_add_int(&tcp_reass_qsize, -1);
1052 LIST_REMOVE(marker, inp_list);
1053 LIST_INSERT_AFTER(inpb, marker, inp_list);
1056 LIST_REMOVE(marker, inp_list);
1057 kfree(marker, M_TEMP);
1061 struct netmsg_tcp_drain {
1062 struct netmsg_base base;
1063 struct inpcbhead *nm_head;
1067 tcp_drain_handler(netmsg_t msg)
1069 struct netmsg_tcp_drain *nm = (void *)msg;
1071 tcp_drain_oncpu(nm->nm_head);
1072 lwkt_replymsg(&nm->base.lmsg, 0);
1087 * Walk the tcpbs, if existing, and flush the reassembly queue,
1088 * if there is one...
1089 * XXX: The "Net/3" implementation doesn't imply that the TCP
1090 * reassembly queue should be flushed, but in a situation
1091 * where we're really low on mbufs, this is potentially
1095 for (cpu = 0; cpu < ncpus2; cpu++) {
1096 struct netmsg_tcp_drain *nm;
1098 if (cpu == mycpu->gd_cpuid) {
1099 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1101 nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1102 M_LWKTMSG, M_NOWAIT);
1105 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1106 0, tcp_drain_handler);
1107 nm->nm_head = &tcbinfo[cpu].pcblisthead;
1108 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg);
1112 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1117 * Notify a tcp user of an asynchronous error;
1118 * store error as soft error, but wake up user
1119 * (for now, won't do anything until can select for soft error).
1121 * Do not wake up user since there currently is no mechanism for
1122 * reporting soft errors (yet - a kqueue filter may be added).
1125 tcp_notify(struct inpcb *inp, int error)
1127 struct tcpcb *tp = intotcpcb(inp);
1130 * Ignore some errors if we are hooked up.
1131 * If connection hasn't completed, has retransmitted several times,
1132 * and receives a second error, give up now. This is better
1133 * than waiting a long time to establish a connection that
1134 * can never complete.
1136 if (tp->t_state == TCPS_ESTABLISHED &&
1137 (error == EHOSTUNREACH || error == ENETUNREACH ||
1138 error == EHOSTDOWN)) {
1140 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1142 tcp_drop(tp, error);
1144 tp->t_softerror = error;
1146 wakeup(&so->so_timeo);
1153 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1156 struct inpcb *marker;
1165 * The process of preparing the TCB list is too time-consuming and
1166 * resource-intensive to repeat twice on every request.
1168 if (req->oldptr == NULL) {
1169 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1170 gd = globaldata_find(ccpu);
1171 n += tcbinfo[gd->gd_cpuid].ipi_count;
1173 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1177 if (req->newptr != NULL)
1180 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1181 marker->inp_flags |= INP_PLACEMARKER;
1184 * OK, now we're committed to doing something. Run the inpcb list
1185 * for each cpu in the system and construct the output. Use a
1186 * list placemarker to deal with list changes occuring during
1187 * copyout blockages (but otherwise depend on being on the correct
1188 * cpu to avoid races).
1190 origcpu = mycpu->gd_cpuid;
1191 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1197 cpu_id = (origcpu + ccpu) % ncpus;
1198 if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1200 rgd = globaldata_find(cpu_id);
1201 lwkt_setcpu_self(rgd);
1203 n = tcbinfo[cpu_id].ipi_count;
1205 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1207 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1209 * process a snapshot of pcbs, ignoring placemarkers
1210 * and using our own to allow SYSCTL_OUT to block.
1212 LIST_REMOVE(marker, inp_list);
1213 LIST_INSERT_AFTER(inp, marker, inp_list);
1215 if (inp->inp_flags & INP_PLACEMARKER)
1217 if (prison_xinpcb(req->td, inp))
1220 xt.xt_len = sizeof xt;
1221 bcopy(inp, &xt.xt_inp, sizeof *inp);
1222 inp_ppcb = inp->inp_ppcb;
1223 if (inp_ppcb != NULL)
1224 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1226 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1227 if (inp->inp_socket)
1228 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1229 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1233 LIST_REMOVE(marker, inp_list);
1234 if (error == 0 && i < n) {
1235 bzero(&xt, sizeof xt);
1236 xt.xt_len = sizeof xt;
1238 error = SYSCTL_OUT(req, &xt, sizeof xt);
1247 * Make sure we are on the same cpu we were on originally, since
1248 * higher level callers expect this. Also don't pollute caches with
1249 * migrated userland data by (eventually) returning to userland
1250 * on a different cpu.
1252 lwkt_setcpu_self(globaldata_find(origcpu));
1253 kfree(marker, M_TEMP);
1257 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1258 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1261 tcp_getcred(SYSCTL_HANDLER_ARGS)
1263 struct sockaddr_in addrs[2];
1268 error = priv_check(req->td, PRIV_ROOT);
1271 error = SYSCTL_IN(req, addrs, sizeof addrs);
1275 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1276 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1277 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1278 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1279 if (inp == NULL || inp->inp_socket == NULL) {
1283 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1289 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1290 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1294 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1296 struct sockaddr_in6 addrs[2];
1299 boolean_t mapped = FALSE;
1301 error = priv_check(req->td, PRIV_ROOT);
1304 error = SYSCTL_IN(req, addrs, sizeof addrs);
1307 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1308 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1315 inp = in_pcblookup_hash(&tcbinfo[0],
1316 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1318 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1322 inp = in6_pcblookup_hash(&tcbinfo[0],
1323 &addrs[1].sin6_addr, addrs[1].sin6_port,
1324 &addrs[0].sin6_addr, addrs[0].sin6_port,
1327 if (inp == NULL || inp->inp_socket == NULL) {
1331 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1337 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1339 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1342 struct netmsg_tcp_notify {
1343 struct netmsg_base base;
1344 void (*nm_notify)(struct inpcb *, int);
1345 struct in_addr nm_faddr;
1350 tcp_notifyall_oncpu(netmsg_t msg)
1352 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1355 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1356 nm->nm_arg, nm->nm_notify);
1358 nextcpu = mycpuid + 1;
1359 if (nextcpu < ncpus2)
1360 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg);
1362 lwkt_replymsg(&nm->base.lmsg, 0);
1366 tcp_ctlinput(netmsg_t msg)
1368 int cmd = msg->ctlinput.nm_cmd;
1369 struct sockaddr *sa = msg->ctlinput.nm_arg;
1370 struct ip *ip = msg->ctlinput.nm_extra;
1372 struct in_addr faddr;
1375 void (*notify)(struct inpcb *, int) = tcp_notify;
1379 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1383 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1384 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1387 arg = inetctlerrmap[cmd];
1388 if (cmd == PRC_QUENCH) {
1389 notify = tcp_quench;
1390 } else if (icmp_may_rst &&
1391 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1392 cmd == PRC_UNREACH_PORT ||
1393 cmd == PRC_TIMXCEED_INTRANS) &&
1395 notify = tcp_drop_syn_sent;
1396 } else if (cmd == PRC_MSGSIZE) {
1397 struct icmp *icmp = (struct icmp *)
1398 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1400 arg = ntohs(icmp->icmp_nextmtu);
1401 notify = tcp_mtudisc;
1402 } else if (PRC_IS_REDIRECT(cmd)) {
1404 notify = in_rtchange;
1405 } else if (cmd == PRC_HOSTDEAD) {
1411 th = (struct tcphdr *)((caddr_t)ip +
1412 (IP_VHL_HL(ip->ip_vhl) << 2));
1413 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1414 ip->ip_src.s_addr, th->th_sport);
1415 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1416 ip->ip_src, th->th_sport, 0, NULL);
1417 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1418 icmpseq = htonl(th->th_seq);
1419 tp = intotcpcb(inp);
1420 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1421 SEQ_LT(icmpseq, tp->snd_max))
1422 (*notify)(inp, arg);
1424 struct in_conninfo inc;
1426 inc.inc_fport = th->th_dport;
1427 inc.inc_lport = th->th_sport;
1428 inc.inc_faddr = faddr;
1429 inc.inc_laddr = ip->ip_src;
1433 syncache_unreach(&inc, th);
1437 struct netmsg_tcp_notify *nm;
1439 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1440 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1441 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1442 0, tcp_notifyall_oncpu);
1443 nm->nm_faddr = faddr;
1445 nm->nm_notify = notify;
1447 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg);
1450 lwkt_replymsg(&msg->lmsg, 0);
1456 tcp6_ctlinput(netmsg_t msg)
1458 int cmd = msg->ctlinput.nm_cmd;
1459 struct sockaddr *sa = msg->ctlinput.nm_arg;
1460 void *d = msg->ctlinput.nm_extra;
1462 void (*notify) (struct inpcb *, int) = tcp_notify;
1463 struct ip6_hdr *ip6;
1465 struct ip6ctlparam *ip6cp = NULL;
1466 const struct sockaddr_in6 *sa6_src = NULL;
1468 struct tcp_portonly {
1474 if (sa->sa_family != AF_INET6 ||
1475 sa->sa_len != sizeof(struct sockaddr_in6)) {
1480 if (cmd == PRC_QUENCH)
1481 notify = tcp_quench;
1482 else if (cmd == PRC_MSGSIZE) {
1483 struct ip6ctlparam *ip6cp = d;
1484 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1486 arg = ntohl(icmp6->icmp6_mtu);
1487 notify = tcp_mtudisc;
1488 } else if (!PRC_IS_REDIRECT(cmd) &&
1489 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1493 /* if the parameter is from icmp6, decode it. */
1495 ip6cp = (struct ip6ctlparam *)d;
1497 ip6 = ip6cp->ip6c_ip6;
1498 off = ip6cp->ip6c_off;
1499 sa6_src = ip6cp->ip6c_src;
1503 off = 0; /* fool gcc */
1508 struct in_conninfo inc;
1510 * XXX: We assume that when IPV6 is non NULL,
1511 * M and OFF are valid.
1514 /* check if we can safely examine src and dst ports */
1515 if (m->m_pkthdr.len < off + sizeof *thp)
1518 bzero(&th, sizeof th);
1519 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1521 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1522 (struct sockaddr *)ip6cp->ip6c_src,
1523 th.th_sport, cmd, arg, notify);
1525 inc.inc_fport = th.th_dport;
1526 inc.inc_lport = th.th_sport;
1527 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1528 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1530 syncache_unreach(&inc, &th);
1532 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1533 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1536 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1542 * Following is where TCP initial sequence number generation occurs.
1544 * There are two places where we must use initial sequence numbers:
1545 * 1. In SYN-ACK packets.
1546 * 2. In SYN packets.
1548 * All ISNs for SYN-ACK packets are generated by the syncache. See
1549 * tcp_syncache.c for details.
1551 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1552 * depends on this property. In addition, these ISNs should be
1553 * unguessable so as to prevent connection hijacking. To satisfy
1554 * the requirements of this situation, the algorithm outlined in
1555 * RFC 1948 is used to generate sequence numbers.
1557 * Implementation details:
1559 * Time is based off the system timer, and is corrected so that it
1560 * increases by one megabyte per second. This allows for proper
1561 * recycling on high speed LANs while still leaving over an hour
1564 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1565 * between seeding of isn_secret. This is normally set to zero,
1566 * as reseeding should not be necessary.
1570 #define ISN_BYTES_PER_SECOND 1048576
1572 u_char isn_secret[32];
1573 int isn_last_reseed;
1577 tcp_new_isn(struct tcpcb *tp)
1579 u_int32_t md5_buffer[4];
1582 /* Seed if this is the first use, reseed if requested. */
1583 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1584 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1586 read_random_unlimited(&isn_secret, sizeof isn_secret);
1587 isn_last_reseed = ticks;
1590 /* Compute the md5 hash and return the ISN. */
1592 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1593 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1595 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1596 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1597 sizeof(struct in6_addr));
1598 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1599 sizeof(struct in6_addr));
1603 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1604 sizeof(struct in_addr));
1605 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1606 sizeof(struct in_addr));
1608 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1609 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1610 new_isn = (tcp_seq) md5_buffer[0];
1611 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1616 * When a source quench is received, close congestion window
1617 * to one segment. We will gradually open it again as we proceed.
1620 tcp_quench(struct inpcb *inp, int error)
1622 struct tcpcb *tp = intotcpcb(inp);
1625 tp->snd_cwnd = tp->t_maxseg;
1631 * When a specific ICMP unreachable message is received and the
1632 * connection state is SYN-SENT, drop the connection. This behavior
1633 * is controlled by the icmp_may_rst sysctl.
1636 tcp_drop_syn_sent(struct inpcb *inp, int error)
1638 struct tcpcb *tp = intotcpcb(inp);
1640 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1641 tcp_drop(tp, error);
1645 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1646 * based on the new value in the route. Also nudge TCP to send something,
1647 * since we know the packet we just sent was dropped.
1648 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1651 tcp_mtudisc(struct inpcb *inp, int mtu)
1653 struct tcpcb *tp = intotcpcb(inp);
1655 struct socket *so = inp->inp_socket;
1658 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1660 const boolean_t isipv6 = FALSE;
1667 * If no MTU is provided in the ICMP message, use the
1668 * next lower likely value, as specified in RFC 1191.
1673 oldmtu = tp->t_maxopd +
1675 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1676 sizeof(struct tcpiphdr));
1677 mtu = ip_next_mtu(oldmtu, 0);
1681 rt = tcp_rtlookup6(&inp->inp_inc);
1683 rt = tcp_rtlookup(&inp->inp_inc);
1685 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1686 mtu = rt->rt_rmx.rmx_mtu;
1690 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1691 sizeof(struct tcpiphdr));
1694 * XXX - The following conditional probably violates the TCP
1695 * spec. The problem is that, since we don't know the
1696 * other end's MSS, we are supposed to use a conservative
1697 * default. But, if we do that, then MTU discovery will
1698 * never actually take place, because the conservative
1699 * default is much less than the MTUs typically seen
1700 * on the Internet today. For the moment, we'll sweep
1701 * this under the carpet.
1703 * The conservative default might not actually be a problem
1704 * if the only case this occurs is when sending an initial
1705 * SYN with options and data to a host we've never talked
1706 * to before. Then, they will reply with an MSS value which
1707 * will get recorded and the new parameters should get
1708 * recomputed. For Further Study.
1710 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1711 maxopd = rt->rt_rmx.rmx_mssopt;
1715 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1716 sizeof(struct tcpiphdr));
1718 if (tp->t_maxopd <= maxopd)
1720 tp->t_maxopd = maxopd;
1723 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1724 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1725 mss -= TCPOLEN_TSTAMP_APPA;
1727 /* round down to multiple of MCLBYTES */
1728 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1730 mss &= ~(MCLBYTES - 1);
1733 mss = (mss / MCLBYTES) * MCLBYTES;
1736 if (so->so_snd.ssb_hiwat < mss)
1737 mss = so->so_snd.ssb_hiwat;
1741 tp->snd_nxt = tp->snd_una;
1743 tcpstat.tcps_mturesent++;
1747 * Look-up the routing entry to the peer of this inpcb. If no route
1748 * is found and it cannot be allocated the return NULL. This routine
1749 * is called by TCP routines that access the rmx structure and by tcp_mss
1750 * to get the interface MTU.
1753 tcp_rtlookup(struct in_conninfo *inc)
1755 struct route *ro = &inc->inc_route;
1757 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1758 /* No route yet, so try to acquire one */
1759 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1761 * unused portions of the structure MUST be zero'd
1762 * out because rtalloc() treats it as opaque data
1764 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1765 ro->ro_dst.sa_family = AF_INET;
1766 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1767 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1777 tcp_rtlookup6(struct in_conninfo *inc)
1779 struct route_in6 *ro6 = &inc->inc6_route;
1781 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1782 /* No route yet, so try to acquire one */
1783 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1785 * unused portions of the structure MUST be zero'd
1786 * out because rtalloc() treats it as opaque data
1788 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1789 ro6->ro_dst.sin6_family = AF_INET6;
1790 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1791 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1792 rtalloc((struct route *)ro6);
1795 return (ro6->ro_rt);
1800 /* compute ESP/AH header size for TCP, including outer IP header. */
1802 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1810 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1812 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1817 if (inp->inp_vflag & INP_IPV6) {
1818 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1820 th = (struct tcphdr *)(ip6 + 1);
1821 m->m_pkthdr.len = m->m_len =
1822 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1823 tcp_fillheaders(tp, ip6, th, FALSE);
1824 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1828 ip = mtod(m, struct ip *);
1829 th = (struct tcphdr *)(ip + 1);
1830 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1831 tcp_fillheaders(tp, ip, th, FALSE);
1832 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1841 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1843 * This code attempts to calculate the bandwidth-delay product as a
1844 * means of determining the optimal window size to maximize bandwidth,
1845 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1846 * routers. This code also does a fairly good job keeping RTTs in check
1847 * across slow links like modems. We implement an algorithm which is very
1848 * similar (but not meant to be) TCP/Vegas. The code operates on the
1849 * transmitter side of a TCP connection and so only effects the transmit
1850 * side of the connection.
1852 * BACKGROUND: TCP makes no provision for the management of buffer space
1853 * at the end points or at the intermediate routers and switches. A TCP
1854 * stream, whether using NewReno or not, will eventually buffer as
1855 * many packets as it is able and the only reason this typically works is
1856 * due to the fairly small default buffers made available for a connection
1857 * (typicaly 16K or 32K). As machines use larger windows and/or window
1858 * scaling it is now fairly easy for even a single TCP connection to blow-out
1859 * all available buffer space not only on the local interface, but on
1860 * intermediate routers and switches as well. NewReno makes a misguided
1861 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1862 * then backing off, then steadily increasing the window again until another
1863 * failure occurs, ad-infinitum. This results in terrible oscillation that
1864 * is only made worse as network loads increase and the idea of intentionally
1865 * blowing out network buffers is, frankly, a terrible way to manage network
1868 * It is far better to limit the transmit window prior to the failure
1869 * condition being achieved. There are two general ways to do this: First
1870 * you can 'scan' through different transmit window sizes and locate the
1871 * point where the RTT stops increasing, indicating that you have filled the
1872 * pipe, then scan backwards until you note that RTT stops decreasing, then
1873 * repeat ad-infinitum. This method works in principle but has severe
1874 * implementation issues due to RTT variances, timer granularity, and
1875 * instability in the algorithm which can lead to many false positives and
1876 * create oscillations as well as interact badly with other TCP streams
1877 * implementing the same algorithm.
1879 * The second method is to limit the window to the bandwidth delay product
1880 * of the link. This is the method we implement. RTT variances and our
1881 * own manipulation of the congestion window, bwnd, can potentially
1882 * destabilize the algorithm. For this reason we have to stabilize the
1883 * elements used to calculate the window. We do this by using the minimum
1884 * observed RTT, the long term average of the observed bandwidth, and
1885 * by adding two segments worth of slop. It isn't perfect but it is able
1886 * to react to changing conditions and gives us a very stable basis on
1887 * which to extend the algorithm.
1890 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1898 * If inflight_enable is disabled in the middle of a tcp connection,
1899 * make sure snd_bwnd is effectively disabled.
1901 if (!tcp_inflight_enable) {
1902 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1903 tp->snd_bandwidth = 0;
1908 * Validate the delta time. If a connection is new or has been idle
1909 * a long time we have to reset the bandwidth calculator.
1912 delta_ticks = save_ticks - tp->t_bw_rtttime;
1913 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1914 tp->t_bw_rtttime = ticks;
1915 tp->t_bw_rtseq = ack_seq;
1916 if (tp->snd_bandwidth == 0)
1917 tp->snd_bandwidth = tcp_inflight_min;
1920 if (delta_ticks == 0)
1924 * Sanity check, plus ignore pure window update acks.
1926 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1930 * Figure out the bandwidth. Due to the tick granularity this
1931 * is a very rough number and it MUST be averaged over a fairly
1932 * long period of time. XXX we need to take into account a link
1933 * that is not using all available bandwidth, but for now our
1934 * slop will ramp us up if this case occurs and the bandwidth later
1937 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1938 tp->t_bw_rtttime = save_ticks;
1939 tp->t_bw_rtseq = ack_seq;
1940 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1942 tp->snd_bandwidth = bw;
1945 * Calculate the semi-static bandwidth delay product, plus two maximal
1946 * segments. The additional slop puts us squarely in the sweet
1947 * spot and also handles the bandwidth run-up case. Without the
1948 * slop we could be locking ourselves into a lower bandwidth.
1950 * Situations Handled:
1951 * (1) Prevents over-queueing of packets on LANs, especially on
1952 * high speed LANs, allowing larger TCP buffers to be
1953 * specified, and also does a good job preventing
1954 * over-queueing of packets over choke points like modems
1955 * (at least for the transmit side).
1957 * (2) Is able to handle changing network loads (bandwidth
1958 * drops so bwnd drops, bandwidth increases so bwnd
1961 * (3) Theoretically should stabilize in the face of multiple
1962 * connections implementing the same algorithm (this may need
1965 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1966 * be adjusted with a sysctl but typically only needs to be on
1967 * very slow connections. A value no smaller then 5 should
1968 * be used, but only reduce this default if you have no other
1972 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1973 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1974 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1977 if (tcp_inflight_debug > 0) {
1979 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1981 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1982 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1985 if ((long)bwnd < tcp_inflight_min)
1986 bwnd = tcp_inflight_min;
1987 if (bwnd > tcp_inflight_max)
1988 bwnd = tcp_inflight_max;
1989 if ((long)bwnd < tp->t_maxseg * 2)
1990 bwnd = tp->t_maxseg * 2;
1991 tp->snd_bwnd = bwnd;
1995 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs)
1998 struct inpcb *inp = tp->t_inpcb;
2000 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2002 const boolean_t isipv6 = FALSE;
2006 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT)
2007 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
2008 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT)
2009 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
2012 rt = tcp_rtlookup6(&inp->inp_inc);
2014 rt = tcp_rtlookup(&inp->inp_inc);
2016 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT ||
2017 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) {
2018 *maxsegs = tcp_iw_maxsegs;
2019 *capsegs = tcp_iw_capsegs;
2022 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs;
2023 *capsegs = rt->rt_rmx.rmx_iwcapsegs;
2027 tcp_initial_window(struct tcpcb *tp)
2029 if (tcp_do_rfc3390) {
2032 * "If the SYN or SYN/ACK is lost, the initial window
2033 * used by a sender after a correctly transmitted SYN
2034 * MUST be one segment consisting of MSS bytes."
2036 * However, we do something a little bit more aggressive
2037 * then RFC3390 here:
2038 * - Only if time spent in the SYN or SYN|ACK retransmition
2039 * >= 3 seconds, the IW is reduced. We do this mainly
2040 * because when RFC3390 is published, the initial RTO is
2041 * still 3 seconds (the threshold we test here), while
2042 * after RFC6298, the initial RTO is 1 second. This
2043 * behaviour probably still falls within the spirit of
2045 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2046 * Mainly to avoid sender and receiver deadlock until
2047 * delayed ACK timer expires. And even RFC2581 does not
2048 * try to reduce IW upon SYN or SYN|ACK retransmition
2052 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2054 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) {
2055 return (2 * tp->t_maxseg);
2057 u_long maxsegs, capsegs;
2059 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs);
2060 return min(maxsegs * tp->t_maxseg,
2061 max(2 * tp->t_maxseg, capsegs * 1460));
2065 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2067 * Mainly to avoid sender and receiver deadlock
2068 * until delayed ACK timer expires.
2070 return (2 * tp->t_maxseg);
2074 #ifdef TCP_SIGNATURE
2076 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2078 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2079 * When called from tcp_input(), we can be sure that th_sum has been
2080 * zeroed out and verified already.
2082 * Return 0 if successful, otherwise return -1.
2084 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2085 * search with the destination IP address, and a 'magic SPI' to be
2086 * determined by the application. This is hardcoded elsewhere to 1179
2087 * right now. Another branch of this code exists which uses the SPD to
2088 * specify per-application flows but it is unstable.
2091 tcpsignature_compute(
2092 struct mbuf *m, /* mbuf chain */
2093 int len, /* length of TCP data */
2094 int optlen, /* length of TCP options */
2095 u_char *buf, /* storage for MD5 digest */
2096 u_int direction) /* direction of flow */
2098 struct ippseudo ippseudo;
2102 struct ipovly *ipovly;
2103 struct secasvar *sav;
2106 struct ip6_hdr *ip6;
2107 struct in6_addr in6;
2113 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2114 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2116 * Extract the destination from the IP header in the mbuf.
2118 ip = mtod(m, struct ip *);
2120 ip6 = NULL; /* Make the compiler happy. */
2123 * Look up an SADB entry which matches the address found in
2126 switch (IP_VHL_V(ip->ip_vhl)) {
2128 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2129 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2132 case (IPV6_VERSION >> 4):
2133 ip6 = mtod(m, struct ip6_hdr *);
2134 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2135 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2144 kprintf("%s: SADB lookup failed\n", __func__);
2150 * Step 1: Update MD5 hash with IP pseudo-header.
2152 * XXX The ippseudo header MUST be digested in network byte order,
2153 * or else we'll fail the regression test. Assume all fields we've
2154 * been doing arithmetic on have been in host byte order.
2155 * XXX One cannot depend on ipovly->ih_len here. When called from
2156 * tcp_output(), the underlying ip_len member has not yet been set.
2158 switch (IP_VHL_V(ip->ip_vhl)) {
2160 ipovly = (struct ipovly *)ip;
2161 ippseudo.ippseudo_src = ipovly->ih_src;
2162 ippseudo.ippseudo_dst = ipovly->ih_dst;
2163 ippseudo.ippseudo_pad = 0;
2164 ippseudo.ippseudo_p = IPPROTO_TCP;
2165 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2166 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2167 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2168 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2172 * RFC 2385, 2.0 Proposal
2173 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2174 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2175 * extended next header value (to form 32 bits), and 32-bit segment
2177 * Note: Upper-Layer Packet Length comes before Next Header.
2179 case (IPV6_VERSION >> 4):
2181 in6_clearscope(&in6);
2182 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2184 in6_clearscope(&in6);
2185 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2186 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2187 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2189 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2190 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2191 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2193 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2194 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2195 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2204 * Step 2: Update MD5 hash with TCP header, excluding options.
2205 * The TCP checksum must be set to zero.
2207 savecsum = th->th_sum;
2209 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2210 th->th_sum = savecsum;
2212 * Step 3: Update MD5 hash with TCP segment data.
2213 * Use m_apply() to avoid an early m_pullup().
2216 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2218 * Step 4: Update MD5 hash with shared secret.
2220 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2221 MD5Final(buf, &ctx);
2222 key_sa_recordxfer(sav, m);
2228 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2231 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2234 #endif /* TCP_SIGNATURE */
2237 tcp_tso_pullup(struct mbuf **mp, int hoff, struct ip **ip0, int *iphlen0,
2238 struct tcphdr **th0, int *thoff0)
2240 struct mbuf *m = *mp;
2246 len = hoff + sizeof(struct ip);
2248 /* The fixed IP header must reside completely in the first mbuf. */
2249 if (m->m_len < len) {
2250 m = m_pullup(m, len);
2255 ip = mtodoff(m, struct ip *, hoff);
2256 iphlen = IP_VHL_HL(ip->ip_vhl) << 2;
2258 /* The full IP header must reside completely in the one mbuf. */
2259 if (m->m_len < hoff + iphlen) {
2260 m = m_pullup(m, hoff + iphlen);
2263 ip = mtodoff(m, struct ip *, hoff);
2266 KASSERT(ip->ip_p == IPPROTO_TCP, ("not tcp %d", ip->ip_p));
2268 if (m->m_len < hoff + iphlen + sizeof(struct tcphdr)) {
2269 m = m_pullup(m, hoff + iphlen + sizeof(struct tcphdr));
2272 ip = mtodoff(m, struct ip *, hoff);
2275 th = (struct tcphdr *)((caddr_t)ip + iphlen);
2276 thoff = th->th_off << 2;
2278 if (m->m_len < hoff + iphlen + thoff) {
2279 m = m_pullup(m, hoff + iphlen + thoff);
2282 ip = mtodoff(m, struct ip *, hoff);
2283 th = (struct tcphdr *)((caddr_t)ip + iphlen);