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 1.12 2004/03/04 01:02:05 hsu 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>
66 #include <netinet/in.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/ip.h>
70 #include <netinet/ip6.h>
72 #include <netinet/in_pcb.h>
74 #include <netinet6/in6_pcb.h>
76 #include <netinet/in_var.h>
77 #include <netinet/ip_var.h>
79 #include <netinet6/ip6_var.h>
81 #include <netinet/tcp.h>
82 #include <netinet/tcp_fsm.h>
83 #include <netinet/tcp_seq.h>
84 #include <netinet/tcp_timer.h>
85 #include <netinet/tcp_var.h>
87 #include <netinet6/tcp6_var.h>
89 #include <netinet/tcpip.h>
91 #include <netinet/tcp_debug.h>
93 #include <netinet6/ip6protosw.h>
96 #include <netinet6/ipsec.h>
98 #include <netinet6/ipsec6.h>
103 #include <netipsec/ipsec.h>
105 #include <netipsec/ipsec6.h>
108 #endif /*FAST_IPSEC*/
112 int tcp_mssdflt = TCP_MSS;
113 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
114 &tcp_mssdflt , 0, "Default TCP Maximum Segment Size");
117 int tcp_v6mssdflt = TCP6_MSS;
118 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
119 CTLFLAG_RW, &tcp_v6mssdflt , 0,
120 "Default TCP Maximum Segment Size for IPv6");
124 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
125 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
126 &tcp_rttdflt , 0, "Default maximum TCP Round Trip Time");
129 int tcp_do_rfc1323 = 1;
130 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
131 &tcp_do_rfc1323 , 0, "Enable rfc1323 (high performance TCP) extensions");
133 int tcp_do_rfc1644 = 0;
134 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
135 &tcp_do_rfc1644 , 0, "Enable rfc1644 (TTCP) extensions");
137 static int tcp_tcbhashsize = 0;
138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
139 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
141 static int do_tcpdrain = 1;
142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
143 "Enable tcp_drain routine for extra help when low on mbufs");
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
146 &tcbinfo.ipi_count, 0, "Number of active PCBs");
148 static int icmp_may_rst = 1;
149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
150 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
152 static int tcp_isn_reseed_interval = 0;
153 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
154 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
157 * TCP bandwidth limiting sysctls. Note that the default lower bound of
158 * 1024 exists only for debugging. A good production default would be
159 * something like 6100.
161 static int tcp_inflight_enable = 0;
162 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
163 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
165 static int tcp_inflight_debug = 0;
166 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
167 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
169 static int tcp_inflight_min = 6144;
170 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
171 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
173 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
174 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
175 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
177 static int tcp_inflight_stab = 20;
178 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
179 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
181 static void tcp_cleartaocache (void);
182 static void tcp_notify (struct inpcb *, int);
185 * Target size of TCP PCB hash tables. Must be a power of two.
187 * Note that this can be overridden by the kernel environment
188 * variable net.inet.tcp.tcbhashsize
191 #define TCBHASHSIZE 512
195 * This is the actual shape of what we allocate using the zone
196 * allocator. Doing it this way allows us to protect both structures
197 * using the same generation count, and also eliminates the overhead
198 * of allocating tcpcbs separately. By hiding the structure here,
199 * we avoid changing most of the rest of the code (although it needs
200 * to be changed, eventually, for greater efficiency).
203 #define ALIGNM1 (ALIGNMENT - 1)
207 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
210 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
211 struct callout inp_tp_delack;
222 int hashsize = TCBHASHSIZE;
227 tcp_delacktime = TCPTV_DELACK;
228 tcp_keepinit = TCPTV_KEEP_INIT;
229 tcp_keepidle = TCPTV_KEEP_IDLE;
230 tcp_keepintvl = TCPTV_KEEPINTVL;
231 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
233 tcp_rexmit_min = TCPTV_MIN;
234 tcp_rexmit_slop = TCPTV_CPU_VAR;
237 tcbinfo.listhead = &tcb;
238 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
239 if (!powerof2(hashsize)) {
240 printf("WARNING: TCB hash size not a power of 2\n");
241 hashsize = 512; /* safe default */
243 tcp_tcbhashsize = hashsize;
244 tcbinfo.hashbase = hashinit(hashsize, M_PCB, &tcbinfo.hashmask);
245 tcbinfo.porthashbase = hashinit(hashsize, M_PCB, &tcbinfo.porthashmask);
246 tcbinfo.bindhashbase = hashinit(hashsize, M_PCB, &tcbinfo.bindhashmask);
247 tcbinfo.ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
250 tcp_reass_maxseg = nmbclusters / 16;
251 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
255 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
257 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
259 if (max_protohdr < TCP_MINPROTOHDR)
260 max_protohdr = TCP_MINPROTOHDR;
261 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
263 #undef TCP_MINPROTOHDR
270 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
271 * tcp_template used to store this data in mbufs, but we now recopy it out
272 * of the tcpcb each time to conserve mbufs.
275 tcp_fillheaders(tp, ip_ptr, tcp_ptr)
280 struct inpcb *inp = tp->t_inpcb;
281 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
284 if ((inp->inp_vflag & INP_IPV6) != 0) {
287 ip6 = (struct ip6_hdr *)ip_ptr;
288 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
289 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
290 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
291 (IPV6_VERSION & IPV6_VERSION_MASK);
292 ip6->ip6_nxt = IPPROTO_TCP;
293 ip6->ip6_plen = sizeof(struct tcphdr);
294 ip6->ip6_src = inp->in6p_laddr;
295 ip6->ip6_dst = inp->in6p_faddr;
300 struct ip *ip = (struct ip *) ip_ptr;
302 ip->ip_vhl = IP_VHL_BORING;
309 ip->ip_p = IPPROTO_TCP;
310 ip->ip_src = inp->inp_laddr;
311 ip->ip_dst = inp->inp_faddr;
312 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
313 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
316 tcp_hdr->th_sport = inp->inp_lport;
317 tcp_hdr->th_dport = inp->inp_fport;
322 tcp_hdr->th_flags = 0;
328 * Create template to be used to send tcp packets on a connection.
329 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
330 * use for this function is in keepalives, which use tcp_respond.
339 m = m_get(M_DONTWAIT, MT_HEADER);
342 m->m_len = sizeof(struct tcptemp);
343 n = mtod(m, struct tcptemp *);
345 tcp_fillheaders(tp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
350 * Send a single message to the TCP at address specified by
351 * the given TCP/IP header. If m == 0, then we make a copy
352 * of the tcpiphdr at ti and send directly to the addressed host.
353 * This is used to force keep alive messages out using the TCP
354 * template for a connection. If flags are given then we send
355 * a message back to the TCP which originated the * segment ti,
356 * and discard the mbuf containing it and any other attached mbufs.
358 * In any case the ack and sequence number of the transmitted
359 * segment are as specified by the parameters.
361 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
364 tcp_respond(tp, ipgen, th, m, ack, seq, flags)
374 struct route *ro = 0;
379 struct route_in6 *ro6 = 0;
380 struct route_in6 sro6;
387 isipv6 = IP_VHL_V(((struct ip *)ipgen)->ip_vhl) == 6;
393 if (!(flags & TH_RST)) {
394 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
395 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
396 win = (long)TCP_MAXWIN << tp->rcv_scale;
400 ro6 = &tp->t_inpcb->in6p_route;
403 ro = &tp->t_inpcb->inp_route;
408 bzero(ro6, sizeof *ro6);
413 bzero(ro, sizeof *ro);
417 m = m_gethdr(M_DONTWAIT, MT_HEADER);
421 m->m_data += max_linkhdr;
424 bcopy((caddr_t)ip6, mtod(m, caddr_t),
425 sizeof(struct ip6_hdr));
426 ip6 = mtod(m, struct ip6_hdr *);
427 nth = (struct tcphdr *)(ip6 + 1);
431 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
432 ip = mtod(m, struct ip *);
433 nth = (struct tcphdr *)(ip + 1);
435 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
440 m->m_data = (caddr_t)ipgen;
441 /* m_len is set later */
443 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
446 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
447 nth = (struct tcphdr *)(ip6 + 1);
451 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
452 nth = (struct tcphdr *)(ip + 1);
456 * this is usually a case when an extension header
457 * exists between the IPv6 header and the
460 nth->th_sport = th->th_sport;
461 nth->th_dport = th->th_dport;
463 xchg(nth->th_dport, nth->th_sport, n_short);
469 ip6->ip6_vfc = IPV6_VERSION;
470 ip6->ip6_nxt = IPPROTO_TCP;
471 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
473 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
477 tlen += sizeof (struct tcpiphdr);
479 ip->ip_ttl = ip_defttl;
482 m->m_pkthdr.len = tlen;
483 m->m_pkthdr.rcvif = (struct ifnet *) 0;
484 nth->th_seq = htonl(seq);
485 nth->th_ack = htonl(ack);
487 nth->th_off = sizeof (struct tcphdr) >> 2;
488 nth->th_flags = flags;
490 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
492 nth->th_win = htons((u_short)win);
497 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
498 sizeof(struct ip6_hdr),
499 tlen - sizeof(struct ip6_hdr));
500 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
507 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
508 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
509 m->m_pkthdr.csum_flags = CSUM_TCP;
510 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
513 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
514 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
518 (void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
519 tp ? tp->t_inpcb : NULL);
520 if (ro6 == &sro6 && ro6->ro_rt) {
527 (void) ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
528 if (ro == &sro && ro->ro_rt) {
536 * Create a new TCP control block, making an
537 * empty reassembly queue and hooking it to the argument
538 * protocol control block. The `inp' parameter must have
539 * come from the zone allocator set up in tcp_init().
548 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
551 it = (struct inp_tp *)inp;
553 bzero((char *) tp, sizeof(struct tcpcb));
554 LIST_INIT(&tp->t_segq);
555 tp->t_maxseg = tp->t_maxopd =
557 isipv6 ? tcp_v6mssdflt :
561 /* Set up our timeouts. */
562 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
563 callout_init(tp->tt_persist = &it->inp_tp_persist);
564 callout_init(tp->tt_keep = &it->inp_tp_keep);
565 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
566 callout_init(tp->tt_delack = &it->inp_tp_delack);
569 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
571 tp->t_flags |= TF_REQ_CC;
572 tp->t_inpcb = inp; /* XXX */
574 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
575 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
576 * reasonable initial retransmit time.
578 tp->t_srtt = TCPTV_SRTTBASE;
579 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
580 tp->t_rttmin = tcp_rexmit_min;
581 tp->t_rxtcur = TCPTV_RTOBASE;
582 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
583 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
584 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
585 tp->t_rcvtime = ticks;
586 tp->t_bw_rtttime = ticks;
588 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
589 * because the socket may be bound to an IPv6 wildcard address,
590 * which may match an IPv4-mapped IPv6 address.
592 inp->inp_ip_ttl = ip_defttl;
593 inp->inp_ppcb = (caddr_t)tp;
594 return (tp); /* XXX */
598 * Drop a TCP connection, reporting
599 * the specified error. If connection is synchronized,
600 * then send a RST to peer.
607 struct socket *so = tp->t_inpcb->inp_socket;
609 if (TCPS_HAVERCVDSYN(tp->t_state)) {
610 tp->t_state = TCPS_CLOSED;
611 (void) tcp_output(tp);
612 tcpstat.tcps_drops++;
614 tcpstat.tcps_conndrops++;
615 if (errno == ETIMEDOUT && tp->t_softerror)
616 errno = tp->t_softerror;
617 so->so_error = errno;
618 return (tcp_close(tp));
622 * Close a TCP control block:
623 * discard all space held by the tcp
624 * discard internet protocol block
625 * wake up any sleepers
632 struct inpcb *inp = tp->t_inpcb;
633 struct socket *so = inp->inp_socket;
635 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
641 * Make sure that all of our timers are stopped before we
644 callout_stop(tp->tt_rexmt);
645 callout_stop(tp->tt_persist);
646 callout_stop(tp->tt_keep);
647 callout_stop(tp->tt_2msl);
648 callout_stop(tp->tt_delack);
651 * If we got enough samples through the srtt filter,
652 * save the rtt and rttvar in the routing entry.
653 * 'Enough' is arbitrarily defined as the 16 samples.
654 * 16 samples is enough for the srtt filter to converge
655 * to within 5% of the correct value; fewer samples and
656 * we could save a very bogus rtt.
658 * Don't update the default route's characteristics and don't
659 * update anything that the user "locked".
661 if (tp->t_rttupdated >= 16) {
665 struct sockaddr_in6 *sin6;
667 if ((rt = inp->in6p_route.ro_rt) == NULL)
669 sin6 = (struct sockaddr_in6 *)rt_key(rt);
670 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
675 if ((rt = inp->inp_route.ro_rt) == NULL ||
676 ((struct sockaddr_in *)rt_key(rt))->sin_addr.s_addr
680 if ((rt->rt_rmx.rmx_locks & RTV_RTT) == 0) {
682 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
683 if (rt->rt_rmx.rmx_rtt && i)
685 * filter this update to half the old & half
686 * the new values, converting scale.
687 * See route.h and tcp_var.h for a
688 * description of the scaling constants.
691 (rt->rt_rmx.rmx_rtt + i) / 2;
693 rt->rt_rmx.rmx_rtt = i;
694 tcpstat.tcps_cachedrtt++;
696 if ((rt->rt_rmx.rmx_locks & RTV_RTTVAR) == 0) {
698 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
699 if (rt->rt_rmx.rmx_rttvar && i)
700 rt->rt_rmx.rmx_rttvar =
701 (rt->rt_rmx.rmx_rttvar + i) / 2;
703 rt->rt_rmx.rmx_rttvar = i;
704 tcpstat.tcps_cachedrttvar++;
707 * The old comment here said:
708 * update the pipelimit (ssthresh) if it has been updated
709 * already or if a pipesize was specified & the threshhold
710 * got below half the pipesize. I.e., wait for bad news
711 * before we start updating, then update on both good
714 * But we want to save the ssthresh even if no pipesize is
715 * specified explicitly in the route, because such
716 * connections still have an implicit pipesize specified
717 * by the global tcp_sendspace. In the absence of a reliable
718 * way to calculate the pipesize, it will have to do.
720 i = tp->snd_ssthresh;
721 if (rt->rt_rmx.rmx_sendpipe != 0)
722 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe / 2);
724 dosavessthresh = (i < so->so_snd.sb_hiwat / 2);
725 if (((rt->rt_rmx.rmx_locks & RTV_SSTHRESH) == 0 &&
726 i != 0 && rt->rt_rmx.rmx_ssthresh != 0)
729 * convert the limit from user data bytes to
730 * packets then to packet data bytes.
732 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
735 i *= (u_long)(tp->t_maxseg +
737 (isipv6 ? sizeof (struct ip6_hdr) +
738 sizeof (struct tcphdr) :
740 sizeof (struct tcpiphdr)
745 if (rt->rt_rmx.rmx_ssthresh)
746 rt->rt_rmx.rmx_ssthresh =
747 (rt->rt_rmx.rmx_ssthresh + i) / 2;
749 rt->rt_rmx.rmx_ssthresh = i;
750 tcpstat.tcps_cachedssthresh++;
754 /* free the reassembly queue, if any */
755 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
756 LIST_REMOVE(q, tqe_q);
761 inp->inp_ppcb = NULL;
762 soisdisconnected(so);
764 if (INP_CHECK_SOCKAF(so, AF_INET6))
769 tcpstat.tcps_closed++;
770 return ((struct tcpcb *)0);
780 struct tseg_qent *te;
783 * Walk the tcpbs, if existing, and flush the reassembly queue,
785 * XXX: The "Net/3" implementation doesn't imply that the TCP
786 * reassembly queue should be flushed, but in a situation
787 * where we're really low on mbufs, this is potentially
790 LIST_FOREACH(inpb, tcbinfo.listhead, inp_list) {
791 if ((tcpb = intotcpcb(inpb))) {
792 while ((te = LIST_FIRST(&tcpb->t_segq))
794 LIST_REMOVE(te, tqe_q);
806 * Notify a tcp user of an asynchronous error;
807 * store error as soft error, but wake up user
808 * (for now, won't do anything until can select for soft error).
810 * Do not wake up user since there currently is no mechanism for
811 * reporting soft errors (yet - a kqueue filter may be added).
814 tcp_notify(inp, error)
818 struct tcpcb *tp = (struct tcpcb *)inp->inp_ppcb;
821 * Ignore some errors if we are hooked up.
822 * If connection hasn't completed, has retransmitted several times,
823 * and receives a second error, give up now. This is better
824 * than waiting a long time to establish a connection that
825 * can never complete.
827 if (tp->t_state == TCPS_ESTABLISHED &&
828 (error == EHOSTUNREACH || error == ENETUNREACH ||
829 error == EHOSTDOWN)) {
831 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
835 tp->t_softerror = error;
837 wakeup((caddr_t) &so->so_timeo);
844 tcp_pcblist(SYSCTL_HANDLER_ARGS)
847 struct inpcb *inp, **inp_list;
852 * The process of preparing the TCB list is too time-consuming and
853 * resource-intensive to repeat twice on every request.
855 if (req->oldptr == 0) {
856 n = tcbinfo.ipi_count;
857 req->oldidx = 2 * (sizeof xig)
858 + (n + n/8) * sizeof(struct xtcpcb);
862 if (req->newptr != 0)
866 * OK, now we're committed to doing something.
869 gencnt = tcbinfo.ipi_gencnt;
870 n = tcbinfo.ipi_count;
873 xig.xig_len = sizeof xig;
875 xig.xig_gen = gencnt;
876 xig.xig_sogen = so_gencnt;
877 error = SYSCTL_OUT(req, &xig, sizeof xig);
881 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
886 for (inp = LIST_FIRST(tcbinfo.listhead), i = 0; inp && i < n;
887 inp = LIST_NEXT(inp, inp_list)) {
888 if (inp->inp_gencnt <= gencnt && !prison_xinpcb(req->td, inp))
895 for (i = 0; i < n; i++) {
897 if (inp->inp_gencnt <= gencnt) {
900 xt.xt_len = sizeof xt;
901 /* XXX should avoid extra copy */
902 bcopy(inp, &xt.xt_inp, sizeof *inp);
903 inp_ppcb = inp->inp_ppcb;
904 if (inp_ppcb != NULL)
905 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
907 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
909 sotoxsocket(inp->inp_socket, &xt.xt_socket);
910 error = SYSCTL_OUT(req, &xt, sizeof xt);
915 * Give the user an updated idea of our state.
916 * If the generation differs from what we told
917 * her before, she knows that something happened
918 * while we were processing this request, and it
919 * might be necessary to retry.
922 xig.xig_gen = tcbinfo.ipi_gencnt;
923 xig.xig_sogen = so_gencnt;
924 xig.xig_count = tcbinfo.ipi_count;
926 error = SYSCTL_OUT(req, &xig, sizeof xig);
928 free(inp_list, M_TEMP);
932 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
933 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
936 tcp_getcred(SYSCTL_HANDLER_ARGS)
938 struct sockaddr_in addrs[2];
942 error = suser(req->td);
945 error = SYSCTL_IN(req, addrs, sizeof(addrs));
949 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
950 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
951 if (inp == NULL || inp->inp_socket == NULL) {
955 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
961 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW,
962 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
966 tcp6_getcred(SYSCTL_HANDLER_ARGS)
968 struct sockaddr_in6 addrs[2];
970 int error, s, mapped = 0;
972 error = suser(req->td);
975 error = SYSCTL_IN(req, addrs, sizeof(addrs));
978 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
979 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
986 inp = in_pcblookup_hash(&tcbinfo,
987 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
989 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
993 inp = in6_pcblookup_hash(&tcbinfo, &addrs[1].sin6_addr,
995 &addrs[0].sin6_addr, addrs[0].sin6_port,
997 if (inp == NULL || inp->inp_socket == NULL) {
1001 error = SYSCTL_OUT(req, inp->inp_socket->so_cred,
1002 sizeof(struct ucred));
1008 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW,
1010 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1015 tcp_ctlinput(cmd, sa, vip)
1017 struct sockaddr *sa;
1020 struct ip *ip = vip;
1022 struct in_addr faddr;
1025 void (*notify) (struct inpcb *, int) = tcp_notify;
1029 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1030 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1033 if (cmd == PRC_QUENCH)
1034 notify = tcp_quench;
1035 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1036 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1037 notify = tcp_drop_syn_sent;
1038 else if (cmd == PRC_MSGSIZE)
1039 notify = tcp_mtudisc;
1040 else if (PRC_IS_REDIRECT(cmd)) {
1042 notify = in_rtchange;
1043 } else if (cmd == PRC_HOSTDEAD)
1045 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0)
1049 th = (struct tcphdr *)((caddr_t)ip
1050 + (IP_VHL_HL(ip->ip_vhl) << 2));
1051 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1052 ip->ip_src, th->th_sport, 0, NULL);
1053 if (inp != NULL && inp->inp_socket != NULL) {
1054 icmp_seq = htonl(th->th_seq);
1055 tp = intotcpcb(inp);
1056 if (SEQ_GEQ(icmp_seq, tp->snd_una) &&
1057 SEQ_LT(icmp_seq, tp->snd_max))
1058 (*notify)(inp, inetctlerrmap[cmd]);
1060 struct in_conninfo inc;
1062 inc.inc_fport = th->th_dport;
1063 inc.inc_lport = th->th_sport;
1064 inc.inc_faddr = faddr;
1065 inc.inc_laddr = ip->ip_src;
1069 syncache_unreach(&inc, th);
1073 in_pcbnotifyall(&tcb, faddr, inetctlerrmap[cmd], notify);
1078 tcp6_ctlinput(cmd, sa, d)
1080 struct sockaddr *sa;
1084 void (*notify) (struct inpcb *, int) = tcp_notify;
1085 struct ip6_hdr *ip6;
1087 struct ip6ctlparam *ip6cp = NULL;
1088 const struct sockaddr_in6 *sa6_src = NULL;
1090 struct tcp_portonly {
1095 if (sa->sa_family != AF_INET6 ||
1096 sa->sa_len != sizeof(struct sockaddr_in6))
1099 if (cmd == PRC_QUENCH)
1100 notify = tcp_quench;
1101 else if (cmd == PRC_MSGSIZE)
1102 notify = tcp_mtudisc;
1103 else if (!PRC_IS_REDIRECT(cmd) &&
1104 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1107 /* if the parameter is from icmp6, decode it. */
1109 ip6cp = (struct ip6ctlparam *)d;
1111 ip6 = ip6cp->ip6c_ip6;
1112 off = ip6cp->ip6c_off;
1113 sa6_src = ip6cp->ip6c_src;
1117 off = 0; /* fool gcc */
1122 struct in_conninfo inc;
1124 * XXX: We assume that when IPV6 is non NULL,
1125 * M and OFF are valid.
1128 /* check if we can safely examine src and dst ports */
1129 if (m->m_pkthdr.len < off + sizeof(*thp))
1132 bzero(&th, sizeof(th));
1133 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1135 in6_pcbnotify(&tcb, sa, th.th_dport,
1136 (struct sockaddr *)ip6cp->ip6c_src,
1137 th.th_sport, cmd, notify);
1139 inc.inc_fport = th.th_dport;
1140 inc.inc_lport = th.th_sport;
1141 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1142 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1144 syncache_unreach(&inc, &th);
1146 in6_pcbnotify(&tcb, sa, 0, (const struct sockaddr *)sa6_src,
1153 * Following is where TCP initial sequence number generation occurs.
1155 * There are two places where we must use initial sequence numbers:
1156 * 1. In SYN-ACK packets.
1157 * 2. In SYN packets.
1159 * All ISNs for SYN-ACK packets are generated by the syncache. See
1160 * tcp_syncache.c for details.
1162 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1163 * depends on this property. In addition, these ISNs should be
1164 * unguessable so as to prevent connection hijacking. To satisfy
1165 * the requirements of this situation, the algorithm outlined in
1166 * RFC 1948 is used to generate sequence numbers.
1168 * Implementation details:
1170 * Time is based off the system timer, and is corrected so that it
1171 * increases by one megabyte per second. This allows for proper
1172 * recycling on high speed LANs while still leaving over an hour
1175 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1176 * between seeding of isn_secret. This is normally set to zero,
1177 * as reseeding should not be necessary.
1181 #define ISN_BYTES_PER_SECOND 1048576
1183 u_char isn_secret[32];
1184 int isn_last_reseed;
1191 u_int32_t md5_buffer[4];
1194 /* Seed if this is the first use, reseed if requested. */
1195 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1196 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1198 read_random_unlimited(&isn_secret, sizeof(isn_secret));
1199 isn_last_reseed = ticks;
1202 /* Compute the md5 hash and return the ISN. */
1204 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1205 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1207 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1208 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1209 sizeof(struct in6_addr));
1210 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1211 sizeof(struct in6_addr));
1215 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1216 sizeof(struct in_addr));
1217 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1218 sizeof(struct in_addr));
1220 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1221 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1222 new_isn = (tcp_seq) md5_buffer[0];
1223 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1228 * When a source quench is received, close congestion window
1229 * to one segment. We will gradually open it again as we proceed.
1232 tcp_quench(inp, errno)
1236 struct tcpcb *tp = intotcpcb(inp);
1239 tp->snd_cwnd = tp->t_maxseg;
1243 * When a specific ICMP unreachable message is received and the
1244 * connection state is SYN-SENT, drop the connection. This behavior
1245 * is controlled by the icmp_may_rst sysctl.
1248 tcp_drop_syn_sent(inp, errno)
1252 struct tcpcb *tp = intotcpcb(inp);
1254 if (tp && tp->t_state == TCPS_SYN_SENT)
1255 tcp_drop(tp, errno);
1259 * When `need fragmentation' ICMP is received, update our idea of the MSS
1260 * based on the new value in the route. Also nudge TCP to send something,
1261 * since we know the packet we just sent was dropped.
1262 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1265 tcp_mtudisc(inp, errno)
1269 struct tcpcb *tp = intotcpcb(inp);
1271 struct rmxp_tao *taop;
1272 struct socket *so = inp->inp_socket;
1276 int isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1282 rt = tcp_rtlookup6(&inp->inp_inc);
1285 rt = tcp_rtlookup(&inp->inp_inc);
1286 if (!rt || !rt->rt_rmx.rmx_mtu) {
1287 tp->t_maxopd = tp->t_maxseg =
1289 isipv6 ? tcp_v6mssdflt :
1294 taop = rmx_taop(rt->rt_rmx);
1295 offered = taop->tao_mssopt;
1296 mss = rt->rt_rmx.rmx_mtu -
1299 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1301 sizeof(struct tcpiphdr)
1308 mss = min(mss, offered);
1310 * XXX - The above conditional probably violates the TCP
1311 * spec. The problem is that, since we don't know the
1312 * other end's MSS, we are supposed to use a conservative
1313 * default. But, if we do that, then MTU discovery will
1314 * never actually take place, because the conservative
1315 * default is much less than the MTUs typically seen
1316 * on the Internet today. For the moment, we'll sweep
1317 * this under the carpet.
1319 * The conservative default might not actually be a problem
1320 * if the only case this occurs is when sending an initial
1321 * SYN with options and data to a host we've never talked
1322 * to before. Then, they will reply with an MSS value which
1323 * will get recorded and the new parameters should get
1324 * recomputed. For Further Study.
1326 if (tp->t_maxopd <= mss)
1330 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1331 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1332 mss -= TCPOLEN_TSTAMP_APPA;
1333 if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
1334 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC)
1335 mss -= TCPOLEN_CC_APPA;
1336 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1338 mss &= ~(MCLBYTES-1);
1341 mss = mss / MCLBYTES * MCLBYTES;
1343 if (so->so_snd.sb_hiwat < mss)
1344 mss = so->so_snd.sb_hiwat;
1348 tcpstat.tcps_mturesent++;
1350 tp->snd_nxt = tp->snd_una;
1356 * Look-up the routing entry to the peer of this inpcb. If no route
1357 * is found and it cannot be allocated the return NULL. This routine
1358 * is called by TCP routines that access the rmx structure and by tcp_mss
1359 * to get the interface MTU.
1363 struct in_conninfo *inc;
1368 ro = &inc->inc_route;
1370 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1371 /* No route yet, so try to acquire one */
1372 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1373 ro->ro_dst.sa_family = AF_INET;
1374 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1375 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1387 struct in_conninfo *inc;
1389 struct route_in6 *ro6;
1392 ro6 = &inc->inc6_route;
1394 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1395 /* No route yet, so try to acquire one */
1396 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1397 ro6->ro_dst.sin6_family = AF_INET6;
1398 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1399 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1400 rtalloc((struct route *)ro6);
1409 /* compute ESP/AH header size for TCP, including outer IP header. */
1411 ipsec_hdrsiz_tcp(tp)
1419 struct ip6_hdr *ip6;
1423 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1425 MGETHDR(m, M_DONTWAIT, MT_DATA);
1430 if ((inp->inp_vflag & INP_IPV6) != 0) {
1431 ip6 = mtod(m, struct ip6_hdr *);
1432 th = (struct tcphdr *)(ip6 + 1);
1433 m->m_pkthdr.len = m->m_len =
1434 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1435 tcp_fillheaders(tp, ip6, th);
1436 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1440 ip = mtod(m, struct ip *);
1441 th = (struct tcphdr *)(ip + 1);
1442 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1443 tcp_fillheaders(tp, ip, th);
1444 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1453 * Return a pointer to the cached information about the remote host.
1454 * The cached information is stored in the protocol specific part of
1455 * the route metrics.
1458 tcp_gettaocache(inc)
1459 struct in_conninfo *inc;
1464 if (inc->inc_isipv6)
1465 rt = tcp_rtlookup6(inc);
1468 rt = tcp_rtlookup(inc);
1470 /* Make sure this is a host route and is up. */
1472 (rt->rt_flags & (RTF_UP|RTF_HOST)) != (RTF_UP|RTF_HOST))
1475 return rmx_taop(rt->rt_rmx);
1479 * Clear all the TAO cache entries, called from tcp_init.
1482 * This routine is just an empty one, because we assume that the routing
1483 * routing tables are initialized at the same time when TCP, so there is
1484 * nothing in the cache left over.
1492 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1494 * This code attempts to calculate the bandwidth-delay product as a
1495 * means of determining the optimal window size to maximize bandwidth,
1496 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1497 * routers. This code also does a fairly good job keeping RTTs in check
1498 * across slow links like modems. We implement an algorithm which is very
1499 * similar (but not meant to be) TCP/Vegas. The code operates on the
1500 * transmitter side of a TCP connection and so only effects the transmit
1501 * side of the connection.
1503 * BACKGROUND: TCP makes no provision for the management of buffer space
1504 * at the end points or at the intermediate routers and switches. A TCP
1505 * stream, whether using NewReno or not, will eventually buffer as
1506 * many packets as it is able and the only reason this typically works is
1507 * due to the fairly small default buffers made available for a connection
1508 * (typicaly 16K or 32K). As machines use larger windows and/or window
1509 * scaling it is now fairly easy for even a single TCP connection to blow-out
1510 * all available buffer space not only on the local interface, but on
1511 * intermediate routers and switches as well. NewReno makes a misguided
1512 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1513 * then backing off, then steadily increasing the window again until another
1514 * failure occurs, ad-infinitum. This results in terrible oscillation that
1515 * is only made worse as network loads increase and the idea of intentionally
1516 * blowing out network buffers is, frankly, a terrible way to manage network
1519 * It is far better to limit the transmit window prior to the failure
1520 * condition being achieved. There are two general ways to do this: First
1521 * you can 'scan' through different transmit window sizes and locate the
1522 * point where the RTT stops increasing, indicating that you have filled the
1523 * pipe, then scan backwards until you note that RTT stops decreasing, then
1524 * repeat ad-infinitum. This method works in principle but has severe
1525 * implementation issues due to RTT variances, timer granularity, and
1526 * instability in the algorithm which can lead to many false positives and
1527 * create oscillations as well as interact badly with other TCP streams
1528 * implementing the same algorithm.
1530 * The second method is to limit the window to the bandwidth delay product
1531 * of the link. This is the method we implement. RTT variances and our
1532 * own manipulation of the congestion window, bwnd, can potentially
1533 * destabilize the algorithm. For this reason we have to stabilize the
1534 * elements used to calculate the window. We do this by using the minimum
1535 * observed RTT, the long term average of the observed bandwidth, and
1536 * by adding two segments worth of slop. It isn't perfect but it is able
1537 * to react to changing conditions and gives us a very stable basis on
1538 * which to extend the algorithm.
1541 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1548 * If inflight_enable is disabled in the middle of a tcp connection,
1549 * make sure snd_bwnd is effectively disabled.
1551 if (tcp_inflight_enable == 0) {
1552 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1553 tp->snd_bandwidth = 0;
1558 * Figure out the bandwidth. Due to the tick granularity this
1559 * is a very rough number and it MUST be averaged over a fairly
1560 * long period of time. XXX we need to take into account a link
1561 * that is not using all available bandwidth, but for now our
1562 * slop will ramp us up if this case occurs and the bandwidth later
1565 * Note: if ticks rollover 'bw' may wind up negative. We must
1566 * effectively reset t_bw_rtttime for this case.
1569 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1572 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1573 (save_ticks - tp->t_bw_rtttime);
1574 tp->t_bw_rtttime = save_ticks;
1575 tp->t_bw_rtseq = ack_seq;
1576 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1578 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1580 tp->snd_bandwidth = bw;
1583 * Calculate the semi-static bandwidth delay product, plus two maximal
1584 * segments. The additional slop puts us squarely in the sweet
1585 * spot and also handles the bandwidth run-up case. Without the
1586 * slop we could be locking ourselves into a lower bandwidth.
1588 * Situations Handled:
1589 * (1) Prevents over-queueing of packets on LANs, especially on
1590 * high speed LANs, allowing larger TCP buffers to be
1591 * specified, and also does a good job preventing
1592 * over-queueing of packets over choke points like modems
1593 * (at least for the transmit side).
1595 * (2) Is able to handle changing network loads (bandwidth
1596 * drops so bwnd drops, bandwidth increases so bwnd
1599 * (3) Theoretically should stabilize in the face of multiple
1600 * connections implementing the same algorithm (this may need
1603 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1604 * be adjusted with a sysctl but typically only needs to be on
1605 * very slow connections. A value no smaller then 5 should
1606 * be used, but only reduce this default if you have no other
1609 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1610 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * (int)tp->t_maxseg / 10;
1613 if (tcp_inflight_debug > 0) {
1615 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1617 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1626 if ((long)bwnd < tcp_inflight_min)
1627 bwnd = tcp_inflight_min;
1628 if (bwnd > tcp_inflight_max)
1629 bwnd = tcp_inflight_max;
1630 if ((long)bwnd < tp->t_maxseg * 2)
1631 bwnd = tp->t_maxseg * 2;
1632 tp->snd_bwnd = bwnd;