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) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
37 * License terms: all terms for the DragonFly license above plus the following:
39 * 4. All advertising materials mentioning features or use of this software
40 * must display the following acknowledgement:
42 * This product includes software developed by Jeffrey M. Hsu
43 * for the DragonFly Project.
45 * This requirement may be waived with permission from Jeffrey Hsu.
46 * This requirement will sunset and may be removed on July 8 2005,
47 * after which the standard DragonFly license (as shown above) will
52 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
53 * The Regents of the University of California. All rights reserved.
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
63 * 3. All advertising materials mentioning features or use of this software
64 * must display the following acknowledgement:
65 * This product includes software developed by the University of
66 * California, Berkeley and its contributors.
67 * 4. Neither the name of the University nor the names of its contributors
68 * may be used to endorse or promote products derived from this software
69 * without specific prior written permission.
71 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
72 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
73 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
74 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
75 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
76 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
77 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
78 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
79 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
80 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
83 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
84 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
85 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.55 2006/12/22 23:57:52 swildner Exp $
88 #include "opt_compat.h"
89 #include "opt_inet6.h"
90 #include "opt_ipsec.h"
91 #include "opt_tcpdebug.h"
93 #include <sys/param.h>
94 #include <sys/systm.h>
95 #include <sys/callout.h>
96 #include <sys/kernel.h>
97 #include <sys/sysctl.h>
98 #include <sys/malloc.h>
99 #include <sys/mpipe.h>
100 #include <sys/mbuf.h>
102 #include <sys/domain.h>
104 #include <sys/proc.h>
105 #include <sys/socket.h>
106 #include <sys/socketvar.h>
107 #include <sys/protosw.h>
108 #include <sys/random.h>
109 #include <sys/in_cksum.h>
112 #include <vm/vm_zone.h>
114 #include <net/route.h>
116 #include <net/netisr.h>
119 #include <netinet/in.h>
120 #include <netinet/in_systm.h>
121 #include <netinet/ip.h>
122 #include <netinet/ip6.h>
123 #include <netinet/in_pcb.h>
124 #include <netinet6/in6_pcb.h>
125 #include <netinet/in_var.h>
126 #include <netinet/ip_var.h>
127 #include <netinet6/ip6_var.h>
128 #include <netinet/ip_icmp.h>
130 #include <netinet/icmp6.h>
132 #include <netinet/tcp.h>
133 #include <netinet/tcp_fsm.h>
134 #include <netinet/tcp_seq.h>
135 #include <netinet/tcp_timer.h>
136 #include <netinet/tcp_var.h>
137 #include <netinet6/tcp6_var.h>
138 #include <netinet/tcpip.h>
140 #include <netinet/tcp_debug.h>
142 #include <netinet6/ip6protosw.h>
145 #include <netinet6/ipsec.h>
147 #include <netinet6/ipsec6.h>
152 #include <netproto/ipsec/ipsec.h>
154 #include <netproto/ipsec/ipsec6.h>
160 #include <sys/msgport2.h>
161 #include <machine/smp.h>
163 #if !defined(KTR_TCP)
164 #define KTR_TCP KTR_ALL
166 KTR_INFO_MASTER(tcp);
167 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
168 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
169 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
170 #define logtcp(name) KTR_LOG(tcp_ ## name)
172 struct inpcbinfo tcbinfo[MAXCPU];
173 struct tcpcbackqhead tcpcbackq[MAXCPU];
175 int tcp_mssdflt = TCP_MSS;
176 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
177 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
180 int tcp_v6mssdflt = TCP6_MSS;
181 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
182 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
186 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
187 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
188 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
191 int tcp_do_rfc1323 = 1;
192 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
193 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
195 int tcp_do_rfc1644 = 0;
196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
197 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
199 static int tcp_tcbhashsize = 0;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
201 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
203 static int do_tcpdrain = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
205 "Enable tcp_drain routine for extra help when low on mbufs");
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
209 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
211 static int icmp_may_rst = 1;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
213 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
215 static int tcp_isn_reseed_interval = 0;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
217 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
220 * TCP bandwidth limiting sysctls. Note that the default lower bound of
221 * 1024 exists only for debugging. A good production default would be
222 * something like 6100.
224 static int tcp_inflight_enable = 0;
225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
226 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
228 static int tcp_inflight_debug = 0;
229 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
230 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
232 static int tcp_inflight_min = 6144;
233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
234 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
236 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
238 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
240 static int tcp_inflight_stab = 20;
241 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
242 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
244 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
245 static struct malloc_pipe tcptemp_mpipe;
247 static void tcp_willblock(void);
248 static void tcp_cleartaocache (void);
249 static void tcp_notify (struct inpcb *, int);
251 struct tcp_stats tcpstats_percpu[MAXCPU];
254 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
258 for (cpu = 0; cpu < ncpus; ++cpu) {
259 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
260 sizeof(struct tcp_stats))))
262 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
263 sizeof(struct tcp_stats))))
269 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
270 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
272 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
273 &tcpstat, tcp_stats, "TCP statistics");
277 * Target size of TCP PCB hash tables. Must be a power of two.
279 * Note that this can be overridden by the kernel environment
280 * variable net.inet.tcp.tcbhashsize
283 #define TCBHASHSIZE 512
287 * This is the actual shape of what we allocate using the zone
288 * allocator. Doing it this way allows us to protect both structures
289 * using the same generation count, and also eliminates the overhead
290 * of allocating tcpcbs separately. By hiding the structure here,
291 * we avoid changing most of the rest of the code (although it needs
292 * to be changed, eventually, for greater efficiency).
295 #define ALIGNM1 (ALIGNMENT - 1)
299 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
302 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
303 struct callout inp_tp_delack;
314 struct inpcbporthead *porthashbase;
316 struct vm_zone *ipi_zone;
317 int hashsize = TCBHASHSIZE;
321 * note: tcptemp is used for keepalives, and it is ok for an
322 * allocation to fail so do not specify MPF_INT.
324 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
330 tcp_delacktime = TCPTV_DELACK;
331 tcp_keepinit = TCPTV_KEEP_INIT;
332 tcp_keepidle = TCPTV_KEEP_IDLE;
333 tcp_keepintvl = TCPTV_KEEPINTVL;
334 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
336 tcp_rexmit_min = TCPTV_MIN;
337 tcp_rexmit_slop = TCPTV_CPU_VAR;
339 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
340 if (!powerof2(hashsize)) {
341 kprintf("WARNING: TCB hash size not a power of 2\n");
342 hashsize = 512; /* safe default */
344 tcp_tcbhashsize = hashsize;
345 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
346 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
349 for (cpu = 0; cpu < ncpus2; cpu++) {
350 in_pcbinfo_init(&tcbinfo[cpu]);
351 tcbinfo[cpu].cpu = cpu;
352 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
353 &tcbinfo[cpu].hashmask);
354 tcbinfo[cpu].porthashbase = porthashbase;
355 tcbinfo[cpu].porthashmask = porthashmask;
356 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
357 &tcbinfo[cpu].wildcardhashmask);
358 tcbinfo[cpu].ipi_zone = ipi_zone;
359 TAILQ_INIT(&tcpcbackq[cpu]);
362 tcp_reass_maxseg = nmbclusters / 16;
363 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
366 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
368 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
370 if (max_protohdr < TCP_MINPROTOHDR)
371 max_protohdr = TCP_MINPROTOHDR;
372 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
374 #undef TCP_MINPROTOHDR
377 * Initialize TCP statistics counters for each CPU.
380 for (cpu = 0; cpu < ncpus; ++cpu) {
381 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
384 bzero(&tcpstat, sizeof(struct tcp_stats));
393 tcpmsg_service_loop(void *dummy)
397 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
400 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg);
401 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
412 int cpu = mycpu->gd_cpuid;
414 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
415 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
416 tp->t_flags &= ~TF_ONOUTPUTQ;
417 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
424 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
425 * tcp_template used to store this data in mbufs, but we now recopy it out
426 * of the tcpcb each time to conserve mbufs.
429 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
431 struct inpcb *inp = tp->t_inpcb;
432 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
435 if (inp->inp_vflag & INP_IPV6) {
438 ip6 = (struct ip6_hdr *)ip_ptr;
439 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
440 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
441 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
442 (IPV6_VERSION & IPV6_VERSION_MASK);
443 ip6->ip6_nxt = IPPROTO_TCP;
444 ip6->ip6_plen = sizeof(struct tcphdr);
445 ip6->ip6_src = inp->in6p_laddr;
446 ip6->ip6_dst = inp->in6p_faddr;
451 struct ip *ip = (struct ip *) ip_ptr;
453 ip->ip_vhl = IP_VHL_BORING;
460 ip->ip_p = IPPROTO_TCP;
461 ip->ip_src = inp->inp_laddr;
462 ip->ip_dst = inp->inp_faddr;
463 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
465 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
468 tcp_hdr->th_sport = inp->inp_lport;
469 tcp_hdr->th_dport = inp->inp_fport;
474 tcp_hdr->th_flags = 0;
480 * Create template to be used to send tcp packets on a connection.
481 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
482 * use for this function is in keepalives, which use tcp_respond.
485 tcp_maketemplate(struct tcpcb *tp)
489 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
491 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
496 tcp_freetemplate(struct tcptemp *tmp)
498 mpipe_free(&tcptemp_mpipe, tmp);
502 * Send a single message to the TCP at address specified by
503 * the given TCP/IP header. If m == NULL, then we make a copy
504 * of the tcpiphdr at ti and send directly to the addressed host.
505 * This is used to force keep alive messages out using the TCP
506 * template for a connection. If flags are given then we send
507 * a message back to the TCP which originated the * segment ti,
508 * and discard the mbuf containing it and any other attached mbufs.
510 * In any case the ack and sequence number of the transmitted
511 * segment are as specified by the parameters.
513 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
516 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
517 tcp_seq ack, tcp_seq seq, int flags)
521 struct route *ro = NULL;
523 struct ip *ip = ipgen;
526 struct route_in6 *ro6 = NULL;
527 struct route_in6 sro6;
528 struct ip6_hdr *ip6 = ipgen;
530 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
532 const boolean_t isipv6 = FALSE;
536 if (!(flags & TH_RST)) {
537 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
538 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
539 win = (long)TCP_MAXWIN << tp->rcv_scale;
542 ro6 = &tp->t_inpcb->in6p_route;
544 ro = &tp->t_inpcb->inp_route;
548 bzero(ro6, sizeof *ro6);
551 bzero(ro, sizeof *ro);
555 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
559 m->m_data += max_linkhdr;
561 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
562 ip6 = mtod(m, struct ip6_hdr *);
563 nth = (struct tcphdr *)(ip6 + 1);
565 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
566 ip = mtod(m, struct ip *);
567 nth = (struct tcphdr *)(ip + 1);
569 bcopy(th, nth, sizeof(struct tcphdr));
574 m->m_data = (caddr_t)ipgen;
575 /* m_len is set later */
577 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
579 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
580 nth = (struct tcphdr *)(ip6 + 1);
582 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
583 nth = (struct tcphdr *)(ip + 1);
587 * this is usually a case when an extension header
588 * exists between the IPv6 header and the
591 nth->th_sport = th->th_sport;
592 nth->th_dport = th->th_dport;
594 xchg(nth->th_dport, nth->th_sport, n_short);
599 ip6->ip6_vfc = IPV6_VERSION;
600 ip6->ip6_nxt = IPPROTO_TCP;
601 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
602 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
604 tlen += sizeof(struct tcpiphdr);
606 ip->ip_ttl = ip_defttl;
609 m->m_pkthdr.len = tlen;
610 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
611 nth->th_seq = htonl(seq);
612 nth->th_ack = htonl(ack);
614 nth->th_off = sizeof(struct tcphdr) >> 2;
615 nth->th_flags = flags;
617 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
619 nth->th_win = htons((u_short)win);
623 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
624 sizeof(struct ip6_hdr),
625 tlen - sizeof(struct ip6_hdr));
626 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
627 (ro6 && ro6->ro_rt) ?
628 ro6->ro_rt->rt_ifp : NULL);
630 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
631 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
632 m->m_pkthdr.csum_flags = CSUM_TCP;
633 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
636 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
637 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
640 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
641 tp ? tp->t_inpcb : NULL);
642 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
647 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
648 if ((ro == &sro) && (ro->ro_rt != NULL)) {
656 * Create a new TCP control block, making an
657 * empty reassembly queue and hooking it to the argument
658 * protocol control block. The `inp' parameter must have
659 * come from the zone allocator set up in tcp_init().
662 tcp_newtcpcb(struct inpcb *inp)
667 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
669 const boolean_t isipv6 = FALSE;
672 it = (struct inp_tp *)inp;
674 bzero(tp, sizeof(struct tcpcb));
675 LIST_INIT(&tp->t_segq);
676 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
678 /* Set up our timeouts. */
679 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
680 callout_init(tp->tt_persist = &it->inp_tp_persist);
681 callout_init(tp->tt_keep = &it->inp_tp_keep);
682 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
683 callout_init(tp->tt_delack = &it->inp_tp_delack);
686 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
688 tp->t_flags |= TF_REQ_CC;
689 tp->t_inpcb = inp; /* XXX */
690 tp->t_state = TCPS_CLOSED;
692 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
693 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
694 * reasonable initial retransmit time.
696 tp->t_srtt = TCPTV_SRTTBASE;
698 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
699 tp->t_rttmin = tcp_rexmit_min;
700 tp->t_rxtcur = TCPTV_RTOBASE;
701 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
702 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
703 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
704 tp->t_rcvtime = ticks;
706 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
707 * because the socket may be bound to an IPv6 wildcard address,
708 * which may match an IPv4-mapped IPv6 address.
710 inp->inp_ip_ttl = ip_defttl;
712 tcp_sack_tcpcb_init(tp);
713 return (tp); /* XXX */
717 * Drop a TCP connection, reporting the specified error.
718 * If connection is synchronized, then send a RST to peer.
721 tcp_drop(struct tcpcb *tp, int error)
723 struct socket *so = tp->t_inpcb->inp_socket;
725 if (TCPS_HAVERCVDSYN(tp->t_state)) {
726 tp->t_state = TCPS_CLOSED;
728 tcpstat.tcps_drops++;
730 tcpstat.tcps_conndrops++;
731 if (error == ETIMEDOUT && tp->t_softerror)
732 error = tp->t_softerror;
733 so->so_error = error;
734 return (tcp_close(tp));
739 struct netmsg_remwildcard {
740 struct lwkt_msg nm_lmsg;
741 struct inpcb *nm_inp;
742 struct inpcbinfo *nm_pcbinfo;
751 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
752 * inp can be detached. We do this by cycling through the cpus, ending up
753 * on the cpu controlling the inp last and then doing the disconnect.
756 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
758 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
761 cpu = msg->nm_pcbinfo->cpu;
763 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
764 /* note: detach removes any wildcard hash entry */
767 in6_pcbdetach(msg->nm_inp);
770 in_pcbdetach(msg->nm_inp);
771 lwkt_replymsg(&msg->nm_lmsg, 0);
773 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
774 cpu = (cpu + 1) % ncpus2;
775 msg->nm_pcbinfo = &tcbinfo[cpu];
776 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg);
784 * Close a TCP control block:
785 * discard all space held by the tcp
786 * discard internet protocol block
787 * wake up any sleepers
790 tcp_close(struct tcpcb *tp)
793 struct inpcb *inp = tp->t_inpcb;
794 struct socket *so = inp->inp_socket;
796 boolean_t dosavessthresh;
801 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
802 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
804 const boolean_t isipv6 = FALSE;
808 * The tp is not instantly destroyed in the wildcard case. Setting
809 * the state to TCPS_TERMINATING will prevent the TCP stack from
810 * messing with it, though it should be noted that this change may
811 * not take effect on other cpus until we have chained the wildcard
814 * XXX we currently depend on the BGL to synchronize the tp->t_state
815 * update and prevent other tcp protocol threads from accepting new
816 * connections on the listen socket we might be trying to close down.
818 KKASSERT(tp->t_state != TCPS_TERMINATING);
819 tp->t_state = TCPS_TERMINATING;
822 * Make sure that all of our timers are stopped before we
825 callout_stop(tp->tt_rexmt);
826 callout_stop(tp->tt_persist);
827 callout_stop(tp->tt_keep);
828 callout_stop(tp->tt_2msl);
829 callout_stop(tp->tt_delack);
831 if (tp->t_flags & TF_ONOUTPUTQ) {
832 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
833 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
834 tp->t_flags &= ~TF_ONOUTPUTQ;
838 * If we got enough samples through the srtt filter,
839 * save the rtt and rttvar in the routing entry.
840 * 'Enough' is arbitrarily defined as the 16 samples.
841 * 16 samples is enough for the srtt filter to converge
842 * to within 5% of the correct value; fewer samples and
843 * we could save a very bogus rtt.
845 * Don't update the default route's characteristics and don't
846 * update anything that the user "locked".
848 if (tp->t_rttupdated >= 16) {
852 struct sockaddr_in6 *sin6;
854 if ((rt = inp->in6p_route.ro_rt) == NULL)
856 sin6 = (struct sockaddr_in6 *)rt_key(rt);
857 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
860 if ((rt = inp->inp_route.ro_rt) == NULL ||
861 ((struct sockaddr_in *)rt_key(rt))->
862 sin_addr.s_addr == INADDR_ANY)
865 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
866 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
867 if (rt->rt_rmx.rmx_rtt && i)
869 * filter this update to half the old & half
870 * the new values, converting scale.
871 * See route.h and tcp_var.h for a
872 * description of the scaling constants.
875 (rt->rt_rmx.rmx_rtt + i) / 2;
877 rt->rt_rmx.rmx_rtt = i;
878 tcpstat.tcps_cachedrtt++;
880 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
882 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
883 if (rt->rt_rmx.rmx_rttvar && i)
884 rt->rt_rmx.rmx_rttvar =
885 (rt->rt_rmx.rmx_rttvar + i) / 2;
887 rt->rt_rmx.rmx_rttvar = i;
888 tcpstat.tcps_cachedrttvar++;
891 * The old comment here said:
892 * update the pipelimit (ssthresh) if it has been updated
893 * already or if a pipesize was specified & the threshhold
894 * got below half the pipesize. I.e., wait for bad news
895 * before we start updating, then update on both good
898 * But we want to save the ssthresh even if no pipesize is
899 * specified explicitly in the route, because such
900 * connections still have an implicit pipesize specified
901 * by the global tcp_sendspace. In the absence of a reliable
902 * way to calculate the pipesize, it will have to do.
904 i = tp->snd_ssthresh;
905 if (rt->rt_rmx.rmx_sendpipe != 0)
906 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
908 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
909 if (dosavessthresh ||
910 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
911 (rt->rt_rmx.rmx_ssthresh != 0))) {
913 * convert the limit from user data bytes to
914 * packets then to packet data bytes.
916 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
921 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
922 sizeof(struct tcpiphdr));
923 if (rt->rt_rmx.rmx_ssthresh)
924 rt->rt_rmx.rmx_ssthresh =
925 (rt->rt_rmx.rmx_ssthresh + i) / 2;
927 rt->rt_rmx.rmx_ssthresh = i;
928 tcpstat.tcps_cachedssthresh++;
933 /* free the reassembly queue, if any */
934 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
935 LIST_REMOVE(q, tqe_q);
940 /* throw away SACK blocks in scoreboard*/
942 tcp_sack_cleanup(&tp->scb);
944 inp->inp_ppcb = NULL;
945 soisdisconnected(so);
947 * Discard the inp. In the SMP case a wildcard inp's hash (created
948 * by a listen socket or an INADDR_ANY udp socket) is replicated
949 * for each protocol thread and must be removed in the context of
950 * that thread. This is accomplished by chaining the message
953 * If the inp is not wildcarded we simply detach, which will remove
954 * the any hashes still present for this inp.
957 if (inp->inp_flags & INP_WILDCARD_MP) {
958 struct netmsg_remwildcard *msg;
960 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
961 msg = kmalloc(sizeof(struct netmsg_remwildcard),
962 M_LWKTMSG, M_INTWAIT);
963 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
964 lwkt_cmd_func(in_pcbremwildcardhash_handler),
967 msg->nm_isinet6 = isafinet6;
970 msg->nm_pcbinfo = &tcbinfo[cpu];
971 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
975 /* note: detach removes any wildcard hash entry */
983 tcpstat.tcps_closed++;
988 tcp_drain_oncpu(struct inpcbhead *head)
992 struct tseg_qent *te;
994 LIST_FOREACH(inpb, head, inp_list) {
995 if (inpb->inp_flags & INP_PLACEMARKER)
997 if ((tcpb = intotcpcb(inpb))) {
998 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
999 LIST_REMOVE(te, tqe_q);
1009 struct netmsg_tcp_drain {
1010 struct lwkt_msg nm_lmsg;
1011 struct inpcbhead *nm_head;
1015 tcp_drain_handler(lwkt_msg_t lmsg)
1017 struct netmsg_tcp_drain *nm = (void *)lmsg;
1019 tcp_drain_oncpu(nm->nm_head);
1020 lwkt_replymsg(lmsg, 0);
1036 * Walk the tcpbs, if existing, and flush the reassembly queue,
1037 * if there is one...
1038 * XXX: The "Net/3" implementation doesn't imply that the TCP
1039 * reassembly queue should be flushed, but in a situation
1040 * where we're really low on mbufs, this is potentially
1044 for (cpu = 0; cpu < ncpus2; cpu++) {
1045 struct netmsg_tcp_drain *msg;
1047 if (cpu == mycpu->gd_cpuid) {
1048 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1050 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1051 M_LWKTMSG, M_NOWAIT);
1054 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
1055 lwkt_cmd_func(tcp_drain_handler),
1057 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1058 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
1062 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1067 * Notify a tcp user of an asynchronous error;
1068 * store error as soft error, but wake up user
1069 * (for now, won't do anything until can select for soft error).
1071 * Do not wake up user since there currently is no mechanism for
1072 * reporting soft errors (yet - a kqueue filter may be added).
1075 tcp_notify(struct inpcb *inp, int error)
1077 struct tcpcb *tp = intotcpcb(inp);
1080 * Ignore some errors if we are hooked up.
1081 * If connection hasn't completed, has retransmitted several times,
1082 * and receives a second error, give up now. This is better
1083 * than waiting a long time to establish a connection that
1084 * can never complete.
1086 if (tp->t_state == TCPS_ESTABLISHED &&
1087 (error == EHOSTUNREACH || error == ENETUNREACH ||
1088 error == EHOSTDOWN)) {
1090 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1092 tcp_drop(tp, error);
1094 tp->t_softerror = error;
1096 wakeup(&so->so_timeo);
1103 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1106 struct inpcb *marker;
1116 * The process of preparing the TCB list is too time-consuming and
1117 * resource-intensive to repeat twice on every request.
1119 if (req->oldptr == NULL) {
1120 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1121 gd = globaldata_find(ccpu);
1122 n += tcbinfo[gd->gd_cpuid].ipi_count;
1124 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1128 if (req->newptr != NULL)
1131 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1132 marker->inp_flags |= INP_PLACEMARKER;
1135 * OK, now we're committed to doing something. Run the inpcb list
1136 * for each cpu in the system and construct the output. Use a
1137 * list placemarker to deal with list changes occuring during
1138 * copyout blockages (but otherwise depend on being on the correct
1139 * cpu to avoid races).
1141 origcpu = mycpu->gd_cpuid;
1142 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1148 cpu_id = (origcpu + ccpu) % ncpus;
1149 if ((smp_active_mask & (1 << cpu_id)) == 0)
1151 rgd = globaldata_find(cpu_id);
1152 lwkt_setcpu_self(rgd);
1154 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1155 n = tcbinfo[cpu_id].ipi_count;
1157 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1159 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1161 * process a snapshot of pcbs, ignoring placemarkers
1162 * and using our own to allow SYSCTL_OUT to block.
1164 LIST_REMOVE(marker, inp_list);
1165 LIST_INSERT_AFTER(inp, marker, inp_list);
1167 if (inp->inp_flags & INP_PLACEMARKER)
1169 if (inp->inp_gencnt > gencnt)
1171 if (prison_xinpcb(req->td, inp))
1174 xt.xt_len = sizeof xt;
1175 bcopy(inp, &xt.xt_inp, sizeof *inp);
1176 inp_ppcb = inp->inp_ppcb;
1177 if (inp_ppcb != NULL)
1178 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1180 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1181 if (inp->inp_socket)
1182 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1183 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1187 LIST_REMOVE(marker, inp_list);
1188 if (error == 0 && i < n) {
1189 bzero(&xt, sizeof xt);
1190 xt.xt_len = sizeof xt;
1192 error = SYSCTL_OUT(req, &xt, sizeof xt);
1201 * Make sure we are on the same cpu we were on originally, since
1202 * higher level callers expect this. Also don't pollute caches with
1203 * migrated userland data by (eventually) returning to userland
1204 * on a different cpu.
1206 lwkt_setcpu_self(globaldata_find(origcpu));
1207 kfree(marker, M_TEMP);
1211 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1212 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1215 tcp_getcred(SYSCTL_HANDLER_ARGS)
1217 struct sockaddr_in addrs[2];
1222 error = suser(req->td);
1225 error = SYSCTL_IN(req, addrs, sizeof addrs);
1229 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1230 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1231 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1232 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1233 if (inp == NULL || inp->inp_socket == NULL) {
1237 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1243 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1244 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1248 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1250 struct sockaddr_in6 addrs[2];
1253 boolean_t mapped = FALSE;
1255 error = suser(req->td);
1258 error = SYSCTL_IN(req, addrs, sizeof addrs);
1261 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1262 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1269 inp = in_pcblookup_hash(&tcbinfo[0],
1270 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1272 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1276 inp = in6_pcblookup_hash(&tcbinfo[0],
1277 &addrs[1].sin6_addr, addrs[1].sin6_port,
1278 &addrs[0].sin6_addr, addrs[0].sin6_port,
1281 if (inp == NULL || inp->inp_socket == NULL) {
1285 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1291 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1293 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1297 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1299 struct ip *ip = vip;
1301 struct in_addr faddr;
1304 void (*notify)(struct inpcb *, int) = tcp_notify;
1308 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1312 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1313 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1316 arg = inetctlerrmap[cmd];
1317 if (cmd == PRC_QUENCH) {
1318 notify = tcp_quench;
1319 } else if (icmp_may_rst &&
1320 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1321 cmd == PRC_UNREACH_PORT ||
1322 cmd == PRC_TIMXCEED_INTRANS) &&
1324 notify = tcp_drop_syn_sent;
1325 } else if (cmd == PRC_MSGSIZE) {
1326 struct icmp *icmp = (struct icmp *)
1327 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1329 arg = ntohs(icmp->icmp_nextmtu);
1330 notify = tcp_mtudisc;
1331 } else if (PRC_IS_REDIRECT(cmd)) {
1333 notify = in_rtchange;
1334 } else if (cmd == PRC_HOSTDEAD) {
1340 th = (struct tcphdr *)((caddr_t)ip +
1341 (IP_VHL_HL(ip->ip_vhl) << 2));
1342 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1343 ip->ip_src.s_addr, th->th_sport);
1344 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1345 ip->ip_src, th->th_sport, 0, NULL);
1346 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1347 icmpseq = htonl(th->th_seq);
1348 tp = intotcpcb(inp);
1349 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1350 SEQ_LT(icmpseq, tp->snd_max))
1351 (*notify)(inp, arg);
1353 struct in_conninfo inc;
1355 inc.inc_fport = th->th_dport;
1356 inc.inc_lport = th->th_sport;
1357 inc.inc_faddr = faddr;
1358 inc.inc_laddr = ip->ip_src;
1362 syncache_unreach(&inc, th);
1366 for (cpu = 0; cpu < ncpus2; cpu++) {
1367 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1375 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1378 void (*notify) (struct inpcb *, int) = tcp_notify;
1379 struct ip6_hdr *ip6;
1381 struct ip6ctlparam *ip6cp = NULL;
1382 const struct sockaddr_in6 *sa6_src = NULL;
1384 struct tcp_portonly {
1390 if (sa->sa_family != AF_INET6 ||
1391 sa->sa_len != sizeof(struct sockaddr_in6))
1395 if (cmd == PRC_QUENCH)
1396 notify = tcp_quench;
1397 else if (cmd == PRC_MSGSIZE) {
1398 struct ip6ctlparam *ip6cp = d;
1399 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1401 arg = ntohl(icmp6->icmp6_mtu);
1402 notify = tcp_mtudisc;
1403 } else if (!PRC_IS_REDIRECT(cmd) &&
1404 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1408 /* if the parameter is from icmp6, decode it. */
1410 ip6cp = (struct ip6ctlparam *)d;
1412 ip6 = ip6cp->ip6c_ip6;
1413 off = ip6cp->ip6c_off;
1414 sa6_src = ip6cp->ip6c_src;
1418 off = 0; /* fool gcc */
1423 struct in_conninfo inc;
1425 * XXX: We assume that when IPV6 is non NULL,
1426 * M and OFF are valid.
1429 /* check if we can safely examine src and dst ports */
1430 if (m->m_pkthdr.len < off + sizeof *thp)
1433 bzero(&th, sizeof th);
1434 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1436 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1437 (struct sockaddr *)ip6cp->ip6c_src,
1438 th.th_sport, cmd, arg, notify);
1440 inc.inc_fport = th.th_dport;
1441 inc.inc_lport = th.th_sport;
1442 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1443 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1445 syncache_unreach(&inc, &th);
1447 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1448 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1453 * Following is where TCP initial sequence number generation occurs.
1455 * There are two places where we must use initial sequence numbers:
1456 * 1. In SYN-ACK packets.
1457 * 2. In SYN packets.
1459 * All ISNs for SYN-ACK packets are generated by the syncache. See
1460 * tcp_syncache.c for details.
1462 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1463 * depends on this property. In addition, these ISNs should be
1464 * unguessable so as to prevent connection hijacking. To satisfy
1465 * the requirements of this situation, the algorithm outlined in
1466 * RFC 1948 is used to generate sequence numbers.
1468 * Implementation details:
1470 * Time is based off the system timer, and is corrected so that it
1471 * increases by one megabyte per second. This allows for proper
1472 * recycling on high speed LANs while still leaving over an hour
1475 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1476 * between seeding of isn_secret. This is normally set to zero,
1477 * as reseeding should not be necessary.
1481 #define ISN_BYTES_PER_SECOND 1048576
1483 u_char isn_secret[32];
1484 int isn_last_reseed;
1488 tcp_new_isn(struct tcpcb *tp)
1490 u_int32_t md5_buffer[4];
1493 /* Seed if this is the first use, reseed if requested. */
1494 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1495 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1497 read_random_unlimited(&isn_secret, sizeof isn_secret);
1498 isn_last_reseed = ticks;
1501 /* Compute the md5 hash and return the ISN. */
1503 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1504 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1506 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1507 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1508 sizeof(struct in6_addr));
1509 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1510 sizeof(struct in6_addr));
1514 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1515 sizeof(struct in_addr));
1516 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1517 sizeof(struct in_addr));
1519 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1520 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1521 new_isn = (tcp_seq) md5_buffer[0];
1522 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1527 * When a source quench is received, close congestion window
1528 * to one segment. We will gradually open it again as we proceed.
1531 tcp_quench(struct inpcb *inp, int error)
1533 struct tcpcb *tp = intotcpcb(inp);
1536 tp->snd_cwnd = tp->t_maxseg;
1542 * When a specific ICMP unreachable message is received and the
1543 * connection state is SYN-SENT, drop the connection. This behavior
1544 * is controlled by the icmp_may_rst sysctl.
1547 tcp_drop_syn_sent(struct inpcb *inp, int error)
1549 struct tcpcb *tp = intotcpcb(inp);
1551 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1552 tcp_drop(tp, error);
1556 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1557 * based on the new value in the route. Also nudge TCP to send something,
1558 * since we know the packet we just sent was dropped.
1559 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1562 tcp_mtudisc(struct inpcb *inp, int mtu)
1564 struct tcpcb *tp = intotcpcb(inp);
1566 struct socket *so = inp->inp_socket;
1569 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1571 const boolean_t isipv6 = FALSE;
1578 * If no MTU is provided in the ICMP message, use the
1579 * next lower likely value, as specified in RFC 1191.
1584 oldmtu = tp->t_maxopd +
1586 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1587 sizeof(struct tcpiphdr));
1588 mtu = ip_next_mtu(oldmtu, 0);
1592 rt = tcp_rtlookup6(&inp->inp_inc);
1594 rt = tcp_rtlookup(&inp->inp_inc);
1596 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1598 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1599 mtu = rt->rt_rmx.rmx_mtu;
1603 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1604 sizeof(struct tcpiphdr));
1607 * XXX - The following conditional probably violates the TCP
1608 * spec. The problem is that, since we don't know the
1609 * other end's MSS, we are supposed to use a conservative
1610 * default. But, if we do that, then MTU discovery will
1611 * never actually take place, because the conservative
1612 * default is much less than the MTUs typically seen
1613 * on the Internet today. For the moment, we'll sweep
1614 * this under the carpet.
1616 * The conservative default might not actually be a problem
1617 * if the only case this occurs is when sending an initial
1618 * SYN with options and data to a host we've never talked
1619 * to before. Then, they will reply with an MSS value which
1620 * will get recorded and the new parameters should get
1621 * recomputed. For Further Study.
1623 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1624 maxopd = taop->tao_mssopt;
1628 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1629 sizeof(struct tcpiphdr));
1631 if (tp->t_maxopd <= maxopd)
1633 tp->t_maxopd = maxopd;
1636 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1637 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1638 mss -= TCPOLEN_TSTAMP_APPA;
1640 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1641 (TF_REQ_CC | TF_RCVD_CC))
1642 mss -= TCPOLEN_CC_APPA;
1644 /* round down to multiple of MCLBYTES */
1645 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1647 mss &= ~(MCLBYTES - 1);
1650 mss = (mss / MCLBYTES) * MCLBYTES;
1653 if (so->so_snd.sb_hiwat < mss)
1654 mss = so->so_snd.sb_hiwat;
1658 tp->snd_nxt = tp->snd_una;
1660 tcpstat.tcps_mturesent++;
1664 * Look-up the routing entry to the peer of this inpcb. If no route
1665 * is found and it cannot be allocated the return NULL. This routine
1666 * is called by TCP routines that access the rmx structure and by tcp_mss
1667 * to get the interface MTU.
1670 tcp_rtlookup(struct in_conninfo *inc)
1672 struct route *ro = &inc->inc_route;
1674 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1675 /* No route yet, so try to acquire one */
1676 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1678 * unused portions of the structure MUST be zero'd
1679 * out because rtalloc() treats it as opaque data
1681 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1682 ro->ro_dst.sa_family = AF_INET;
1683 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1684 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1694 tcp_rtlookup6(struct in_conninfo *inc)
1696 struct route_in6 *ro6 = &inc->inc6_route;
1698 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1699 /* No route yet, so try to acquire one */
1700 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1702 * unused portions of the structure MUST be zero'd
1703 * out because rtalloc() treats it as opaque data
1705 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1706 ro6->ro_dst.sin6_family = AF_INET6;
1707 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1708 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1709 rtalloc((struct route *)ro6);
1712 return (ro6->ro_rt);
1717 /* compute ESP/AH header size for TCP, including outer IP header. */
1719 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1727 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1729 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1734 if (inp->inp_vflag & INP_IPV6) {
1735 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1737 th = (struct tcphdr *)(ip6 + 1);
1738 m->m_pkthdr.len = m->m_len =
1739 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1740 tcp_fillheaders(tp, ip6, th);
1741 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1745 ip = mtod(m, struct ip *);
1746 th = (struct tcphdr *)(ip + 1);
1747 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1748 tcp_fillheaders(tp, ip, th);
1749 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1758 * Return a pointer to the cached information about the remote host.
1759 * The cached information is stored in the protocol specific part of
1760 * the route metrics.
1763 tcp_gettaocache(struct in_conninfo *inc)
1768 if (inc->inc_isipv6)
1769 rt = tcp_rtlookup6(inc);
1772 rt = tcp_rtlookup(inc);
1774 /* Make sure this is a host route and is up. */
1776 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1779 return (rmx_taop(rt->rt_rmx));
1783 * Clear all the TAO cache entries, called from tcp_init.
1786 * This routine is just an empty one, because we assume that the routing
1787 * routing tables are initialized at the same time when TCP, so there is
1788 * nothing in the cache left over.
1791 tcp_cleartaocache(void)
1796 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1798 * This code attempts to calculate the bandwidth-delay product as a
1799 * means of determining the optimal window size to maximize bandwidth,
1800 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1801 * routers. This code also does a fairly good job keeping RTTs in check
1802 * across slow links like modems. We implement an algorithm which is very
1803 * similar (but not meant to be) TCP/Vegas. The code operates on the
1804 * transmitter side of a TCP connection and so only effects the transmit
1805 * side of the connection.
1807 * BACKGROUND: TCP makes no provision for the management of buffer space
1808 * at the end points or at the intermediate routers and switches. A TCP
1809 * stream, whether using NewReno or not, will eventually buffer as
1810 * many packets as it is able and the only reason this typically works is
1811 * due to the fairly small default buffers made available for a connection
1812 * (typicaly 16K or 32K). As machines use larger windows and/or window
1813 * scaling it is now fairly easy for even a single TCP connection to blow-out
1814 * all available buffer space not only on the local interface, but on
1815 * intermediate routers and switches as well. NewReno makes a misguided
1816 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1817 * then backing off, then steadily increasing the window again until another
1818 * failure occurs, ad-infinitum. This results in terrible oscillation that
1819 * is only made worse as network loads increase and the idea of intentionally
1820 * blowing out network buffers is, frankly, a terrible way to manage network
1823 * It is far better to limit the transmit window prior to the failure
1824 * condition being achieved. There are two general ways to do this: First
1825 * you can 'scan' through different transmit window sizes and locate the
1826 * point where the RTT stops increasing, indicating that you have filled the
1827 * pipe, then scan backwards until you note that RTT stops decreasing, then
1828 * repeat ad-infinitum. This method works in principle but has severe
1829 * implementation issues due to RTT variances, timer granularity, and
1830 * instability in the algorithm which can lead to many false positives and
1831 * create oscillations as well as interact badly with other TCP streams
1832 * implementing the same algorithm.
1834 * The second method is to limit the window to the bandwidth delay product
1835 * of the link. This is the method we implement. RTT variances and our
1836 * own manipulation of the congestion window, bwnd, can potentially
1837 * destabilize the algorithm. For this reason we have to stabilize the
1838 * elements used to calculate the window. We do this by using the minimum
1839 * observed RTT, the long term average of the observed bandwidth, and
1840 * by adding two segments worth of slop. It isn't perfect but it is able
1841 * to react to changing conditions and gives us a very stable basis on
1842 * which to extend the algorithm.
1845 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1853 * If inflight_enable is disabled in the middle of a tcp connection,
1854 * make sure snd_bwnd is effectively disabled.
1856 if (!tcp_inflight_enable) {
1857 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1858 tp->snd_bandwidth = 0;
1863 * Validate the delta time. If a connection is new or has been idle
1864 * a long time we have to reset the bandwidth calculator.
1867 delta_ticks = save_ticks - tp->t_bw_rtttime;
1868 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1869 tp->t_bw_rtttime = ticks;
1870 tp->t_bw_rtseq = ack_seq;
1871 if (tp->snd_bandwidth == 0)
1872 tp->snd_bandwidth = tcp_inflight_min;
1875 if (delta_ticks == 0)
1879 * Sanity check, plus ignore pure window update acks.
1881 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1885 * Figure out the bandwidth. Due to the tick granularity this
1886 * is a very rough number and it MUST be averaged over a fairly
1887 * long period of time. XXX we need to take into account a link
1888 * that is not using all available bandwidth, but for now our
1889 * slop will ramp us up if this case occurs and the bandwidth later
1892 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1893 tp->t_bw_rtttime = save_ticks;
1894 tp->t_bw_rtseq = ack_seq;
1895 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1897 tp->snd_bandwidth = bw;
1900 * Calculate the semi-static bandwidth delay product, plus two maximal
1901 * segments. The additional slop puts us squarely in the sweet
1902 * spot and also handles the bandwidth run-up case. Without the
1903 * slop we could be locking ourselves into a lower bandwidth.
1905 * Situations Handled:
1906 * (1) Prevents over-queueing of packets on LANs, especially on
1907 * high speed LANs, allowing larger TCP buffers to be
1908 * specified, and also does a good job preventing
1909 * over-queueing of packets over choke points like modems
1910 * (at least for the transmit side).
1912 * (2) Is able to handle changing network loads (bandwidth
1913 * drops so bwnd drops, bandwidth increases so bwnd
1916 * (3) Theoretically should stabilize in the face of multiple
1917 * connections implementing the same algorithm (this may need
1920 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1921 * be adjusted with a sysctl but typically only needs to be on
1922 * very slow connections. A value no smaller then 5 should
1923 * be used, but only reduce this default if you have no other
1927 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1928 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1929 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1932 if (tcp_inflight_debug > 0) {
1934 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1936 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1937 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1940 if ((long)bwnd < tcp_inflight_min)
1941 bwnd = tcp_inflight_min;
1942 if (bwnd > tcp_inflight_max)
1943 bwnd = tcp_inflight_max;
1944 if ((long)bwnd < tp->t_maxseg * 2)
1945 bwnd = tp->t_maxseg * 2;
1946 tp->snd_bwnd = bwnd;