2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. All advertising materials mentioning features or use of this software
14 * must display the following acknowledgement:
15 * This product includes software developed by the University of
16 * California, Berkeley and its contributors.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
34 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
35 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.28 2004/04/21 18:13:56 dillon Exp $
38 #include "opt_compat.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_tcpdebug.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/callout.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/malloc.h>
51 #include <sys/domain.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/protosw.h>
57 #include <sys/random.h>
58 #include <sys/in_cksum.h>
60 #include <vm/vm_zone.h>
62 #include <net/route.h>
64 #include <net/netisr.h>
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
70 #include <netinet/ip6.h>
71 #include <netinet/in_pcb.h>
72 #include <netinet6/in6_pcb.h>
73 #include <netinet/in_var.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 #include <netinet/tcp.h>
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet6/tcp6_var.h>
82 #include <netinet/tcpip.h>
84 #include <netinet/tcp_debug.h>
86 #include <netinet6/ip6protosw.h>
89 #include <netinet6/ipsec.h>
91 #include <netinet6/ipsec6.h>
96 #include <netipsec/ipsec.h>
98 #include <netipsec/ipsec6.h>
105 #include <sys/msgport2.h>
107 int tcp_mssdflt = TCP_MSS;
108 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
109 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
112 int tcp_v6mssdflt = TCP6_MSS;
113 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
114 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
118 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
119 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
120 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
123 int tcp_do_rfc1323 = 1;
124 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
125 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
127 int tcp_do_rfc1644 = 0;
128 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
129 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
131 static int tcp_tcbhashsize = 0;
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
133 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
135 static int do_tcpdrain = 1;
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
137 "Enable tcp_drain routine for extra help when low on mbufs");
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
141 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
143 static int icmp_may_rst = 1;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
145 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
147 static int tcp_isn_reseed_interval = 0;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
149 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
152 * TCP bandwidth limiting sysctls. Note that the default lower bound of
153 * 1024 exists only for debugging. A good production default would be
154 * something like 6100.
156 static int tcp_inflight_enable = 0;
157 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
158 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
160 static int tcp_inflight_debug = 0;
161 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
162 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
164 static int tcp_inflight_min = 6144;
165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
166 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
168 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
170 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
172 static int tcp_inflight_stab = 20;
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
174 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
176 static void tcp_cleartaocache (void);
177 static void tcp_notify (struct inpcb *, int);
179 struct tcp_stats tcpstats_ary[MAXCPU];
182 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
186 for (cpu = 0; cpu < ncpus; ++cpu) {
187 if ((error = SYSCTL_OUT(req, (void *)&tcpstats_ary[cpu],
188 sizeof(struct tcp_stats))))
190 if ((error = SYSCTL_IN(req, (void *)&tcpstats_ary[cpu],
191 sizeof(struct tcp_stats))))
197 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
198 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
200 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
201 &tcpstat, tcp_stats, "TCP statistics");
205 * Target size of TCP PCB hash tables. Must be a power of two.
207 * Note that this can be overridden by the kernel environment
208 * variable net.inet.tcp.tcbhashsize
211 #define TCBHASHSIZE 512
215 * This is the actual shape of what we allocate using the zone
216 * allocator. Doing it this way allows us to protect both structures
217 * using the same generation count, and also eliminates the overhead
218 * of allocating tcpcbs separately. By hiding the structure here,
219 * we avoid changing most of the rest of the code (although it needs
220 * to be changed, eventually, for greater efficiency).
223 #define ALIGNM1 (ALIGNMENT - 1)
227 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
230 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
231 struct callout inp_tp_delack;
242 struct inpcbporthead *porthashbase;
244 struct inpcontainerhead *wildcardhashbase;
245 u_long wildcardhashmask;
246 struct vm_zone *ipi_zone;
247 int hashsize = TCBHASHSIZE;
253 tcp_delacktime = TCPTV_DELACK;
254 tcp_keepinit = TCPTV_KEEP_INIT;
255 tcp_keepidle = TCPTV_KEEP_IDLE;
256 tcp_keepintvl = TCPTV_KEEPINTVL;
257 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
259 tcp_rexmit_min = TCPTV_MIN;
260 tcp_rexmit_slop = TCPTV_CPU_VAR;
262 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
263 if (!powerof2(hashsize)) {
264 printf("WARNING: TCB hash size not a power of 2\n");
265 hashsize = 512; /* safe default */
267 tcp_tcbhashsize = hashsize;
268 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
269 wildcardhashbase = hashinit(hashsize, M_PCB, &wildcardhashmask);
270 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
273 for (cpu = 0; cpu < ncpus2; cpu++) {
274 LIST_INIT(&tcbinfo[cpu].listhead);
275 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
276 &tcbinfo[cpu].hashmask);
277 tcbinfo[cpu].porthashbase = porthashbase;
278 tcbinfo[cpu].porthashmask = porthashmask;
279 tcbinfo[cpu].wildcardhashbase = wildcardhashbase;
280 tcbinfo[cpu].wildcardhashmask = wildcardhashmask;
281 tcbinfo[cpu].ipi_zone = ipi_zone;
284 tcp_reass_maxseg = nmbclusters / 16;
285 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
288 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
290 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
292 if (max_protohdr < TCP_MINPROTOHDR)
293 max_protohdr = TCP_MINPROTOHDR;
294 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
296 #undef TCP_MINPROTOHDR
299 * Initialize TCP statistics.
301 * It is layed out as an array which is has one element for UP,
302 * and SMP_MAXCPU elements for SMP. This allows us to retain
303 * the access mechanism from userland for both UP and SMP.
306 for (cpu = 0; cpu < ncpus; ++cpu) {
307 bzero(&tcpstats_ary[cpu], sizeof(struct tcp_stats));
310 bzero(&tcpstat, sizeof(struct tcp_stats));
318 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
319 * tcp_template used to store this data in mbufs, but we now recopy it out
320 * of the tcpcb each time to conserve mbufs.
323 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
325 struct inpcb *inp = tp->t_inpcb;
326 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
329 if (inp->inp_vflag & INP_IPV6) {
332 ip6 = (struct ip6_hdr *)ip_ptr;
333 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
334 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
335 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
336 (IPV6_VERSION & IPV6_VERSION_MASK);
337 ip6->ip6_nxt = IPPROTO_TCP;
338 ip6->ip6_plen = sizeof(struct tcphdr);
339 ip6->ip6_src = inp->in6p_laddr;
340 ip6->ip6_dst = inp->in6p_faddr;
345 struct ip *ip = (struct ip *) ip_ptr;
347 ip->ip_vhl = IP_VHL_BORING;
354 ip->ip_p = IPPROTO_TCP;
355 ip->ip_src = inp->inp_laddr;
356 ip->ip_dst = inp->inp_faddr;
357 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
359 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
362 tcp_hdr->th_sport = inp->inp_lport;
363 tcp_hdr->th_dport = inp->inp_fport;
368 tcp_hdr->th_flags = 0;
374 * Create template to be used to send tcp packets on a connection.
375 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
376 * use for this function is in keepalives, which use tcp_respond.
379 tcp_maketemplate(struct tcpcb *tp)
384 m = m_get(M_DONTWAIT, MT_HEADER);
387 m->m_len = sizeof(struct tcptemp);
388 n = mtod(m, struct tcptemp *);
390 tcp_fillheaders(tp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
395 * Send a single message to the TCP at address specified by
396 * the given TCP/IP header. If m == NULL, then we make a copy
397 * of the tcpiphdr at ti and send directly to the addressed host.
398 * This is used to force keep alive messages out using the TCP
399 * template for a connection. If flags are given then we send
400 * a message back to the TCP which originated the * segment ti,
401 * and discard the mbuf containing it and any other attached mbufs.
403 * In any case the ack and sequence number of the transmitted
404 * segment are as specified by the parameters.
406 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
409 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
410 tcp_seq ack, tcp_seq seq, int flags)
414 struct route *ro = NULL;
416 struct ip *ip = ipgen;
419 struct route_in6 *ro6 = NULL;
420 struct route_in6 sro6;
421 struct ip6_hdr *ip6 = ipgen;
423 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
425 const boolean_t isipv6 = FALSE;
429 if (!(flags & TH_RST)) {
430 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
431 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
432 win = (long)TCP_MAXWIN << tp->rcv_scale;
435 ro6 = &tp->t_inpcb->in6p_route;
437 ro = &tp->t_inpcb->inp_route;
441 bzero(ro6, sizeof *ro6);
444 bzero(ro, sizeof *ro);
448 m = m_gethdr(M_DONTWAIT, MT_HEADER);
452 m->m_data += max_linkhdr;
454 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
455 ip6 = mtod(m, struct ip6_hdr *);
456 nth = (struct tcphdr *)(ip6 + 1);
458 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
459 ip = mtod(m, struct ip *);
460 nth = (struct tcphdr *)(ip + 1);
462 bcopy(th, nth, sizeof(struct tcphdr));
467 m->m_data = (caddr_t)ipgen;
468 /* m_len is set later */
470 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
472 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
473 nth = (struct tcphdr *)(ip6 + 1);
475 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
476 nth = (struct tcphdr *)(ip + 1);
480 * this is usually a case when an extension header
481 * exists between the IPv6 header and the
484 nth->th_sport = th->th_sport;
485 nth->th_dport = th->th_dport;
487 xchg(nth->th_dport, nth->th_sport, n_short);
492 ip6->ip6_vfc = IPV6_VERSION;
493 ip6->ip6_nxt = IPPROTO_TCP;
494 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
495 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
497 tlen += sizeof(struct tcpiphdr);
499 ip->ip_ttl = ip_defttl;
502 m->m_pkthdr.len = tlen;
503 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
504 nth->th_seq = htonl(seq);
505 nth->th_ack = htonl(ack);
507 nth->th_off = sizeof(struct tcphdr) >> 2;
508 nth->th_flags = flags;
510 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
512 nth->th_win = htons((u_short)win);
516 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
517 sizeof(struct ip6_hdr),
518 tlen - sizeof(struct ip6_hdr));
519 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
520 (ro6 && ro6->ro_rt) ?
521 ro6->ro_rt->rt_ifp : NULL);
523 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
524 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
525 m->m_pkthdr.csum_flags = CSUM_TCP;
526 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
529 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
530 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
533 (void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
534 tp ? tp->t_inpcb : NULL);
535 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
540 (void)ip_output(m, NULL, ro, ipflags, NULL,
541 tp ? tp->t_inpcb : NULL);
542 if ((ro == &sro) && (ro->ro_rt != NULL)) {
550 * Create a new TCP control block, making an
551 * empty reassembly queue and hooking it to the argument
552 * protocol control block. The `inp' parameter must have
553 * come from the zone allocator set up in tcp_init().
556 tcp_newtcpcb(struct inpcb *inp)
561 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
563 const boolean_t isipv6 = FALSE;
566 it = (struct inp_tp *)inp;
568 bzero(tp, sizeof(struct tcpcb));
569 LIST_INIT(&tp->t_segq);
570 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
572 /* Set up our timeouts. */
573 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
574 callout_init(tp->tt_persist = &it->inp_tp_persist);
575 callout_init(tp->tt_keep = &it->inp_tp_keep);
576 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
577 callout_init(tp->tt_delack = &it->inp_tp_delack);
580 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
582 tp->t_flags |= TF_REQ_CC;
583 tp->t_inpcb = inp; /* XXX */
585 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
586 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
587 * reasonable initial retransmit time.
589 tp->t_srtt = TCPTV_SRTTBASE;
591 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
592 tp->t_rttmin = tcp_rexmit_min;
593 tp->t_rxtcur = TCPTV_RTOBASE;
594 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
595 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
596 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
597 tp->t_rcvtime = ticks;
598 tp->t_bw_rtttime = ticks;
600 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
601 * because the socket may be bound to an IPv6 wildcard address,
602 * which may match an IPv4-mapped IPv6 address.
604 inp->inp_ip_ttl = ip_defttl;
605 inp->inp_ppcb = (caddr_t)tp;
606 return (tp); /* XXX */
610 * Drop a TCP connection, reporting the specified error.
611 * If connection is synchronized, then send a RST to peer.
614 tcp_drop(struct tcpcb *tp, int errno)
616 struct socket *so = tp->t_inpcb->inp_socket;
618 if (TCPS_HAVERCVDSYN(tp->t_state)) {
619 tp->t_state = TCPS_CLOSED;
620 (void) tcp_output(tp);
621 tcpstat.tcps_drops++;
623 tcpstat.tcps_conndrops++;
624 if (errno == ETIMEDOUT && tp->t_softerror)
625 errno = tp->t_softerror;
626 so->so_error = errno;
627 return (tcp_close(tp));
631 * Close a TCP control block:
632 * discard all space held by the tcp
633 * discard internet protocol block
634 * wake up any sleepers
637 tcp_close(struct tcpcb *tp)
640 struct inpcb *inp = tp->t_inpcb;
641 struct socket *so = inp->inp_socket;
643 boolean_t dosavessthresh;
645 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
647 const boolean_t isipv6 = FALSE;
651 * Make sure that all of our timers are stopped before we
654 callout_stop(tp->tt_rexmt);
655 callout_stop(tp->tt_persist);
656 callout_stop(tp->tt_keep);
657 callout_stop(tp->tt_2msl);
658 callout_stop(tp->tt_delack);
661 * If we got enough samples through the srtt filter,
662 * save the rtt and rttvar in the routing entry.
663 * 'Enough' is arbitrarily defined as the 16 samples.
664 * 16 samples is enough for the srtt filter to converge
665 * to within 5% of the correct value; fewer samples and
666 * we could save a very bogus rtt.
668 * Don't update the default route's characteristics and don't
669 * update anything that the user "locked".
671 if (tp->t_rttupdated >= 16) {
675 struct sockaddr_in6 *sin6;
677 if ((rt = inp->in6p_route.ro_rt) == NULL)
679 sin6 = (struct sockaddr_in6 *)rt_key(rt);
680 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
683 if ((rt = inp->inp_route.ro_rt) == NULL ||
684 ((struct sockaddr_in *)rt_key(rt))->
685 sin_addr.s_addr == INADDR_ANY)
688 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
689 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
690 if (rt->rt_rmx.rmx_rtt && i)
692 * filter this update to half the old & half
693 * the new values, converting scale.
694 * See route.h and tcp_var.h for a
695 * description of the scaling constants.
698 (rt->rt_rmx.rmx_rtt + i) / 2;
700 rt->rt_rmx.rmx_rtt = i;
701 tcpstat.tcps_cachedrtt++;
703 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
705 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
706 if (rt->rt_rmx.rmx_rttvar && i)
707 rt->rt_rmx.rmx_rttvar =
708 (rt->rt_rmx.rmx_rttvar + i) / 2;
710 rt->rt_rmx.rmx_rttvar = i;
711 tcpstat.tcps_cachedrttvar++;
714 * The old comment here said:
715 * update the pipelimit (ssthresh) if it has been updated
716 * already or if a pipesize was specified & the threshhold
717 * got below half the pipesize. I.e., wait for bad news
718 * before we start updating, then update on both good
721 * But we want to save the ssthresh even if no pipesize is
722 * specified explicitly in the route, because such
723 * connections still have an implicit pipesize specified
724 * by the global tcp_sendspace. In the absence of a reliable
725 * way to calculate the pipesize, it will have to do.
727 i = tp->snd_ssthresh;
728 if (rt->rt_rmx.rmx_sendpipe != 0)
729 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
731 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
732 if (dosavessthresh ||
733 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
734 (rt->rt_rmx.rmx_ssthresh != 0))) {
736 * convert the limit from user data bytes to
737 * packets then to packet data bytes.
739 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
744 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
745 sizeof(struct tcpiphdr));
746 if (rt->rt_rmx.rmx_ssthresh)
747 rt->rt_rmx.rmx_ssthresh =
748 (rt->rt_rmx.rmx_ssthresh + i) / 2;
750 rt->rt_rmx.rmx_ssthresh = i;
751 tcpstat.tcps_cachedssthresh++;
756 /* free the reassembly queue, if any */
757 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
758 LIST_REMOVE(q, tqe_q);
763 inp->inp_ppcb = NULL;
764 soisdisconnected(so);
766 if (INP_CHECK_SOCKAF(so, AF_INET6))
771 tcpstat.tcps_closed++;
776 tcp_drain_oncpu(struct inpcbhead *head)
780 struct tseg_qent *te;
782 LIST_FOREACH(inpb, head, inp_list) {
783 if ((tcpb = intotcpcb(inpb))) {
784 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
785 LIST_REMOVE(te, tqe_q);
795 struct netmsg_tcp_drain {
796 struct lwkt_msg nm_lmsg;
797 struct inpcbhead *nm_head;
801 tcp_drain_handler(lwkt_msg_t lmsg)
803 struct netmsg_tcp_drain *nm = (void *)lmsg;
805 tcp_drain_oncpu(nm->nm_head);
806 lwkt_replymsg(lmsg, 0);
822 * Walk the tcpbs, if existing, and flush the reassembly queue,
824 * XXX: The "Net/3" implementation doesn't imply that the TCP
825 * reassembly queue should be flushed, but in a situation
826 * where we're really low on mbufs, this is potentially
830 for (cpu = 0; cpu < ncpus2; cpu++) {
831 struct netmsg_tcp_drain *msg;
833 if (cpu == mycpu->gd_cpuid) {
834 tcp_drain_oncpu(&tcbinfo[cpu].listhead);
836 msg = malloc(sizeof(struct netmsg_tcp_drain),
837 M_LWKTMSG, M_NOWAIT);
840 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
841 lwkt_cmd_func(tcp_drain_handler),
843 msg->nm_head = &tcbinfo[cpu].listhead;
844 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
848 tcp_drain_oncpu(&tcbinfo[0].listhead);
853 * Notify a tcp user of an asynchronous error;
854 * store error as soft error, but wake up user
855 * (for now, won't do anything until can select for soft error).
857 * Do not wake up user since there currently is no mechanism for
858 * reporting soft errors (yet - a kqueue filter may be added).
861 tcp_notify(struct inpcb *inp, int error)
863 struct tcpcb *tp = intotcpcb(inp);
866 * Ignore some errors if we are hooked up.
867 * If connection hasn't completed, has retransmitted several times,
868 * and receives a second error, give up now. This is better
869 * than waiting a long time to establish a connection that
870 * can never complete.
872 if (tp->t_state == TCPS_ESTABLISHED &&
873 (error == EHOSTUNREACH || error == ENETUNREACH ||
874 error == EHOSTDOWN)) {
876 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
880 tp->t_softerror = error;
882 wakeup((caddr_t) &so->so_timeo);
889 tcp_pcblist(SYSCTL_HANDLER_ARGS)
892 struct inpcb *inp, **inp_list;
897 * The process of preparing the TCB list is too time-consuming and
898 * resource-intensive to repeat twice on every request.
900 if (req->oldptr == NULL) {
901 n = tcbinfo[mycpu->gd_cpuid].ipi_count;
902 req->oldidx = 2 * (sizeof xig) +
903 (n + n/8) * sizeof(struct xtcpcb);
907 if (req->newptr != NULL)
911 * OK, now we're committed to doing something.
914 gencnt = tcbinfo[mycpu->gd_cpuid].ipi_gencnt;
915 n = tcbinfo[mycpu->gd_cpuid].ipi_count;
918 xig.xig_len = sizeof xig;
920 xig.xig_gen = gencnt;
921 xig.xig_sogen = so_gencnt;
922 error = SYSCTL_OUT(req, &xig, sizeof xig);
926 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
927 if (inp_list == NULL)
931 for (inp = LIST_FIRST(&tcbinfo[mycpu->gd_cpuid].listhead), i = 0;
932 inp && i < n; inp = LIST_NEXT(inp, inp_list)) {
933 if (inp->inp_gencnt <= gencnt && !prison_xinpcb(req->td, inp))
940 for (i = 0; i < n; i++) {
942 if (inp->inp_gencnt <= gencnt) {
945 xt.xt_len = sizeof xt;
946 /* XXX should avoid extra copy */
947 bcopy(inp, &xt.xt_inp, sizeof *inp);
948 inp_ppcb = inp->inp_ppcb;
949 if (inp_ppcb != NULL)
950 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
952 bzero(&xt.xt_tp, sizeof xt.xt_tp);
954 sotoxsocket(inp->inp_socket, &xt.xt_socket);
955 error = SYSCTL_OUT(req, &xt, sizeof xt);
960 * Give the user an updated idea of our state.
961 * If the generation differs from what we told
962 * her before, she knows that something happened
963 * while we were processing this request, and it
964 * might be necessary to retry.
967 xig.xig_gen = tcbinfo[mycpu->gd_cpuid].ipi_gencnt;
968 xig.xig_sogen = so_gencnt;
969 xig.xig_count = tcbinfo[mycpu->gd_cpuid].ipi_count;
971 error = SYSCTL_OUT(req, &xig, sizeof xig);
973 free(inp_list, M_TEMP);
977 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
978 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
981 tcp_getcred(SYSCTL_HANDLER_ARGS)
983 struct sockaddr_in addrs[2];
988 error = suser(req->td);
991 error = SYSCTL_IN(req, addrs, sizeof addrs);
996 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
997 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
998 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
999 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1000 if (inp == NULL || inp->inp_socket == NULL) {
1004 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1010 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1011 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1015 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1017 struct sockaddr_in6 addrs[2];
1020 boolean_t mapped = FALSE;
1022 error = suser(req->td);
1025 error = SYSCTL_IN(req, addrs, sizeof addrs);
1028 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1029 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1036 inp = in_pcblookup_hash(&tcbinfo[0],
1037 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1039 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1043 inp = in6_pcblookup_hash(&tcbinfo[0],
1044 &addrs[1].sin6_addr, addrs[1].sin6_port,
1045 &addrs[0].sin6_addr, addrs[0].sin6_port,
1048 if (inp == NULL || inp->inp_socket == NULL) {
1052 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1058 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1060 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1064 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1066 struct ip *ip = vip;
1068 struct in_addr faddr;
1071 void (*notify)(struct inpcb *, int) = tcp_notify;
1076 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1077 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1080 if (cmd == PRC_QUENCH)
1081 notify = tcp_quench;
1082 else if (icmp_may_rst &&
1083 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT ||
1084 cmd == PRC_TIMXCEED_INTRANS) &&
1086 notify = tcp_drop_syn_sent;
1087 else if (cmd == PRC_MSGSIZE)
1088 notify = tcp_mtudisc;
1089 else if (PRC_IS_REDIRECT(cmd)) {
1091 notify = in_rtchange;
1092 } else if (cmd == PRC_HOSTDEAD)
1094 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0)
1098 th = (struct tcphdr *)((caddr_t)ip +
1099 (IP_VHL_HL(ip->ip_vhl) << 2));
1100 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1101 ip->ip_src.s_addr, th->th_sport);
1102 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1103 ip->ip_src, th->th_sport, 0, NULL);
1104 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1105 icmp_seq = htonl(th->th_seq);
1106 tp = intotcpcb(inp);
1107 if (SEQ_GEQ(icmp_seq, tp->snd_una) &&
1108 SEQ_LT(icmp_seq, tp->snd_max))
1109 (*notify)(inp, inetctlerrmap[cmd]);
1111 struct in_conninfo inc;
1113 inc.inc_fport = th->th_dport;
1114 inc.inc_lport = th->th_sport;
1115 inc.inc_faddr = faddr;
1116 inc.inc_laddr = ip->ip_src;
1120 syncache_unreach(&inc, th);
1124 for (cpu = 0; cpu < ncpus2; cpu++) {
1125 in_pcbnotifyall(&tcbinfo[cpu].listhead, faddr,
1126 inetctlerrmap[cmd], notify);
1133 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1136 void (*notify) (struct inpcb *, int) = tcp_notify;
1137 struct ip6_hdr *ip6;
1139 struct ip6ctlparam *ip6cp = NULL;
1140 const struct sockaddr_in6 *sa6_src = NULL;
1142 struct tcp_portonly {
1147 if (sa->sa_family != AF_INET6 ||
1148 sa->sa_len != sizeof(struct sockaddr_in6))
1151 if (cmd == PRC_QUENCH)
1152 notify = tcp_quench;
1153 else if (cmd == PRC_MSGSIZE)
1154 notify = tcp_mtudisc;
1155 else if (!PRC_IS_REDIRECT(cmd) &&
1156 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1159 /* if the parameter is from icmp6, decode it. */
1161 ip6cp = (struct ip6ctlparam *)d;
1163 ip6 = ip6cp->ip6c_ip6;
1164 off = ip6cp->ip6c_off;
1165 sa6_src = ip6cp->ip6c_src;
1169 off = 0; /* fool gcc */
1174 struct in_conninfo inc;
1176 * XXX: We assume that when IPV6 is non NULL,
1177 * M and OFF are valid.
1180 /* check if we can safely examine src and dst ports */
1181 if (m->m_pkthdr.len < off + sizeof *thp)
1184 bzero(&th, sizeof th);
1185 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1187 in6_pcbnotify(&tcbinfo[0].listhead, sa, th.th_dport,
1188 (struct sockaddr *)ip6cp->ip6c_src,
1189 th.th_sport, cmd, notify);
1191 inc.inc_fport = th.th_dport;
1192 inc.inc_lport = th.th_sport;
1193 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1194 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1196 syncache_unreach(&inc, &th);
1198 in6_pcbnotify(&tcbinfo[0].listhead, sa, 0,
1199 (const struct sockaddr *)sa6_src, 0, cmd, notify);
1204 * Following is where TCP initial sequence number generation occurs.
1206 * There are two places where we must use initial sequence numbers:
1207 * 1. In SYN-ACK packets.
1208 * 2. In SYN packets.
1210 * All ISNs for SYN-ACK packets are generated by the syncache. See
1211 * tcp_syncache.c for details.
1213 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1214 * depends on this property. In addition, these ISNs should be
1215 * unguessable so as to prevent connection hijacking. To satisfy
1216 * the requirements of this situation, the algorithm outlined in
1217 * RFC 1948 is used to generate sequence numbers.
1219 * Implementation details:
1221 * Time is based off the system timer, and is corrected so that it
1222 * increases by one megabyte per second. This allows for proper
1223 * recycling on high speed LANs while still leaving over an hour
1226 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1227 * between seeding of isn_secret. This is normally set to zero,
1228 * as reseeding should not be necessary.
1232 #define ISN_BYTES_PER_SECOND 1048576
1234 u_char isn_secret[32];
1235 int isn_last_reseed;
1239 tcp_new_isn(struct tcpcb *tp)
1241 u_int32_t md5_buffer[4];
1244 /* Seed if this is the first use, reseed if requested. */
1245 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1246 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1248 read_random_unlimited(&isn_secret, sizeof isn_secret);
1249 isn_last_reseed = ticks;
1252 /* Compute the md5 hash and return the ISN. */
1254 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1255 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1257 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1258 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1259 sizeof(struct in6_addr));
1260 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1261 sizeof(struct in6_addr));
1265 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1266 sizeof(struct in_addr));
1267 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1268 sizeof(struct in_addr));
1270 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1271 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1272 new_isn = (tcp_seq) md5_buffer[0];
1273 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1278 * When a source quench is received, close congestion window
1279 * to one segment. We will gradually open it again as we proceed.
1282 tcp_quench(struct inpcb *inp, int errno)
1284 struct tcpcb *tp = intotcpcb(inp);
1287 tp->snd_cwnd = tp->t_maxseg;
1291 * When a specific ICMP unreachable message is received and the
1292 * connection state is SYN-SENT, drop the connection. This behavior
1293 * is controlled by the icmp_may_rst sysctl.
1296 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1298 struct tcpcb *tp = intotcpcb(inp);
1300 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1301 tcp_drop(tp, errno);
1305 * When `need fragmentation' ICMP is received, update our idea of the MSS
1306 * based on the new value in the route. Also nudge TCP to send something,
1307 * since we know the packet we just sent was dropped.
1308 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1311 tcp_mtudisc(struct inpcb *inp, int errno)
1313 struct tcpcb *tp = intotcpcb(inp);
1315 struct rmxp_tao *taop;
1316 struct socket *so = inp->inp_socket;
1320 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1322 const boolean_t isipv6 = FALSE;
1327 rt = tcp_rtlookup6(&inp->inp_inc);
1329 rt = tcp_rtlookup(&inp->inp_inc);
1330 if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) {
1331 tp->t_maxopd = tp->t_maxseg =
1332 isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
1335 taop = rmx_taop(rt->rt_rmx);
1336 offered = taop->tao_mssopt;
1337 mss = rt->rt_rmx.rmx_mtu -
1339 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1340 sizeof(struct tcpiphdr));
1343 mss = min(mss, offered);
1345 * XXX - The above conditional probably violates the TCP
1346 * spec. The problem is that, since we don't know the
1347 * other end's MSS, we are supposed to use a conservative
1348 * default. But, if we do that, then MTU discovery will
1349 * never actually take place, because the conservative
1350 * default is much less than the MTUs typically seen
1351 * on the Internet today. For the moment, we'll sweep
1352 * this under the carpet.
1354 * The conservative default might not actually be a problem
1355 * if the only case this occurs is when sending an initial
1356 * SYN with options and data to a host we've never talked
1357 * to before. Then, they will reply with an MSS value which
1358 * will get recorded and the new parameters should get
1359 * recomputed. For Further Study.
1361 if (tp->t_maxopd <= mss)
1365 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1366 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1367 mss -= TCPOLEN_TSTAMP_APPA;
1368 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
1369 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC)
1370 mss -= TCPOLEN_CC_APPA;
1371 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1373 mss &= ~(MCLBYTES - 1);
1376 mss = mss / MCLBYTES * MCLBYTES;
1378 if (so->so_snd.sb_hiwat < mss)
1379 mss = so->so_snd.sb_hiwat;
1383 tcpstat.tcps_mturesent++;
1385 tp->snd_nxt = tp->snd_una;
1391 * Look-up the routing entry to the peer of this inpcb. If no route
1392 * is found and it cannot be allocated the return NULL. This routine
1393 * is called by TCP routines that access the rmx structure and by tcp_mss
1394 * to get the interface MTU.
1397 tcp_rtlookup(struct in_conninfo *inc)
1402 ro = &inc->inc_route;
1404 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1405 /* No route yet, so try to acquire one */
1406 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1407 ro->ro_dst.sa_family = AF_INET;
1408 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1409 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1420 tcp_rtlookup6(struct in_conninfo *inc)
1422 struct route_in6 *ro6;
1425 ro6 = &inc->inc6_route;
1427 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1428 /* No route yet, so try to acquire one */
1429 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1430 ro6->ro_dst.sin6_family = AF_INET6;
1431 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1432 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1433 rtalloc((struct route *)ro6);
1442 /* compute ESP/AH header size for TCP, including outer IP header. */
1444 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1452 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1454 MGETHDR(m, M_DONTWAIT, MT_DATA);
1459 if (inp->inp_vflag & INP_IPV6) {
1460 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1462 th = (struct tcphdr *)(ip6 + 1);
1463 m->m_pkthdr.len = m->m_len =
1464 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1465 tcp_fillheaders(tp, ip6, th);
1466 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1470 ip = mtod(m, struct ip *);
1471 th = (struct tcphdr *)(ip + 1);
1472 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1473 tcp_fillheaders(tp, ip, th);
1474 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1483 * Return a pointer to the cached information about the remote host.
1484 * The cached information is stored in the protocol specific part of
1485 * the route metrics.
1488 tcp_gettaocache(struct in_conninfo *inc)
1493 if (inc->inc_isipv6)
1494 rt = tcp_rtlookup6(inc);
1497 rt = tcp_rtlookup(inc);
1499 /* Make sure this is a host route and is up. */
1501 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1504 return (rmx_taop(rt->rt_rmx));
1508 * Clear all the TAO cache entries, called from tcp_init.
1511 * This routine is just an empty one, because we assume that the routing
1512 * routing tables are initialized at the same time when TCP, so there is
1513 * nothing in the cache left over.
1521 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1523 * This code attempts to calculate the bandwidth-delay product as a
1524 * means of determining the optimal window size to maximize bandwidth,
1525 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1526 * routers. This code also does a fairly good job keeping RTTs in check
1527 * across slow links like modems. We implement an algorithm which is very
1528 * similar (but not meant to be) TCP/Vegas. The code operates on the
1529 * transmitter side of a TCP connection and so only effects the transmit
1530 * side of the connection.
1532 * BACKGROUND: TCP makes no provision for the management of buffer space
1533 * at the end points or at the intermediate routers and switches. A TCP
1534 * stream, whether using NewReno or not, will eventually buffer as
1535 * many packets as it is able and the only reason this typically works is
1536 * due to the fairly small default buffers made available for a connection
1537 * (typicaly 16K or 32K). As machines use larger windows and/or window
1538 * scaling it is now fairly easy for even a single TCP connection to blow-out
1539 * all available buffer space not only on the local interface, but on
1540 * intermediate routers and switches as well. NewReno makes a misguided
1541 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1542 * then backing off, then steadily increasing the window again until another
1543 * failure occurs, ad-infinitum. This results in terrible oscillation that
1544 * is only made worse as network loads increase and the idea of intentionally
1545 * blowing out network buffers is, frankly, a terrible way to manage network
1548 * It is far better to limit the transmit window prior to the failure
1549 * condition being achieved. There are two general ways to do this: First
1550 * you can 'scan' through different transmit window sizes and locate the
1551 * point where the RTT stops increasing, indicating that you have filled the
1552 * pipe, then scan backwards until you note that RTT stops decreasing, then
1553 * repeat ad-infinitum. This method works in principle but has severe
1554 * implementation issues due to RTT variances, timer granularity, and
1555 * instability in the algorithm which can lead to many false positives and
1556 * create oscillations as well as interact badly with other TCP streams
1557 * implementing the same algorithm.
1559 * The second method is to limit the window to the bandwidth delay product
1560 * of the link. This is the method we implement. RTT variances and our
1561 * own manipulation of the congestion window, bwnd, can potentially
1562 * destabilize the algorithm. For this reason we have to stabilize the
1563 * elements used to calculate the window. We do this by using the minimum
1564 * observed RTT, the long term average of the observed bandwidth, and
1565 * by adding two segments worth of slop. It isn't perfect but it is able
1566 * to react to changing conditions and gives us a very stable basis on
1567 * which to extend the algorithm.
1570 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1577 * If inflight_enable is disabled in the middle of a tcp connection,
1578 * make sure snd_bwnd is effectively disabled.
1580 if (!tcp_inflight_enable) {
1581 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1582 tp->snd_bandwidth = 0;
1587 * Figure out the bandwidth. Due to the tick granularity this
1588 * is a very rough number and it MUST be averaged over a fairly
1589 * long period of time. XXX we need to take into account a link
1590 * that is not using all available bandwidth, but for now our
1591 * slop will ramp us up if this case occurs and the bandwidth later
1594 * Note: if ticks rollover 'bw' may wind up negative. We must
1595 * effectively reset t_bw_rtttime for this case.
1598 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1601 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1602 (save_ticks - tp->t_bw_rtttime);
1603 tp->t_bw_rtttime = save_ticks;
1604 tp->t_bw_rtseq = ack_seq;
1605 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1607 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1609 tp->snd_bandwidth = bw;
1612 * Calculate the semi-static bandwidth delay product, plus two maximal
1613 * segments. The additional slop puts us squarely in the sweet
1614 * spot and also handles the bandwidth run-up case. Without the
1615 * slop we could be locking ourselves into a lower bandwidth.
1617 * Situations Handled:
1618 * (1) Prevents over-queueing of packets on LANs, especially on
1619 * high speed LANs, allowing larger TCP buffers to be
1620 * specified, and also does a good job preventing
1621 * over-queueing of packets over choke points like modems
1622 * (at least for the transmit side).
1624 * (2) Is able to handle changing network loads (bandwidth
1625 * drops so bwnd drops, bandwidth increases so bwnd
1628 * (3) Theoretically should stabilize in the face of multiple
1629 * connections implementing the same algorithm (this may need
1632 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1633 * be adjusted with a sysctl but typically only needs to be on
1634 * very slow connections. A value no smaller then 5 should
1635 * be used, but only reduce this default if you have no other
1639 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1640 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1641 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1644 if (tcp_inflight_debug > 0) {
1646 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1648 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1649 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1652 if ((long)bwnd < tcp_inflight_min)
1653 bwnd = tcp_inflight_min;
1654 if (bwnd > tcp_inflight_max)
1655 bwnd = tcp_inflight_max;
1656 if ((long)bwnd < tp->t_maxseg * 2)
1657 bwnd = tp->t_maxseg * 2;
1658 tp->snd_bwnd = bwnd;