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 1.17 2004/03/14 08:26:31 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>
64 #include <net/netisr.h>
67 #include <netinet/in.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
73 #include <netinet/in_pcb.h>
75 #include <netinet6/in6_pcb.h>
77 #include <netinet/in_var.h>
78 #include <netinet/ip_var.h>
80 #include <netinet6/ip6_var.h>
82 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_seq.h>
85 #include <netinet/tcp_timer.h>
86 #include <netinet/tcp_var.h>
88 #include <netinet6/tcp6_var.h>
90 #include <netinet/tcpip.h>
92 #include <netinet/tcp_debug.h>
94 #include <netinet6/ip6protosw.h>
97 #include <netinet6/ipsec.h>
99 #include <netinet6/ipsec6.h>
104 #include <netipsec/ipsec.h>
106 #include <netipsec/ipsec6.h>
109 #endif /*FAST_IPSEC*/
113 #include <sys/msgport2.h>
115 int tcp_mssdflt = TCP_MSS;
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
117 &tcp_mssdflt , 0, "Default TCP Maximum Segment Size");
120 int tcp_v6mssdflt = TCP6_MSS;
121 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
122 CTLFLAG_RW, &tcp_v6mssdflt , 0,
123 "Default TCP Maximum Segment Size for IPv6");
127 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
128 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
129 &tcp_rttdflt , 0, "Default maximum TCP Round Trip Time");
132 int tcp_do_rfc1323 = 1;
133 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
134 &tcp_do_rfc1323 , 0, "Enable rfc1323 (high performance TCP) extensions");
136 int tcp_do_rfc1644 = 0;
137 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
138 &tcp_do_rfc1644 , 0, "Enable rfc1644 (TTCP) extensions");
140 static int tcp_tcbhashsize = 0;
141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
142 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
144 static int do_tcpdrain = 1;
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
146 "Enable tcp_drain routine for extra help when low on mbufs");
149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
150 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
152 static int icmp_may_rst = 1;
153 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
154 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
156 static int tcp_isn_reseed_interval = 0;
157 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
158 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
161 * TCP bandwidth limiting sysctls. Note that the default lower bound of
162 * 1024 exists only for debugging. A good production default would be
163 * something like 6100.
165 static int tcp_inflight_enable = 0;
166 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
167 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
169 static int tcp_inflight_debug = 0;
170 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
171 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
173 static int tcp_inflight_min = 6144;
174 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
175 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
177 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
178 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
179 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
181 static int tcp_inflight_stab = 20;
182 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
183 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
185 static void tcp_cleartaocache (void);
186 static void tcp_notify (struct inpcb *, int);
189 * Target size of TCP PCB hash tables. Must be a power of two.
191 * Note that this can be overridden by the kernel environment
192 * variable net.inet.tcp.tcbhashsize
195 #define TCBHASHSIZE 512
199 * This is the actual shape of what we allocate using the zone
200 * allocator. Doing it this way allows us to protect both structures
201 * using the same generation count, and also eliminates the overhead
202 * of allocating tcpcbs separately. By hiding the structure here,
203 * we avoid changing most of the rest of the code (although it needs
204 * to be changed, eventually, for greater efficiency).
207 #define ALIGNM1 (ALIGNMENT - 1)
211 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
214 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
215 struct callout inp_tp_delack;
226 struct inpcbporthead *porthashbase;
228 struct inpcbhead *bindhashbase;
230 struct vm_zone *ipi_zone;
231 int hashsize = TCBHASHSIZE;
237 tcp_delacktime = TCPTV_DELACK;
238 tcp_keepinit = TCPTV_KEEP_INIT;
239 tcp_keepidle = TCPTV_KEEP_IDLE;
240 tcp_keepintvl = TCPTV_KEEPINTVL;
241 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
243 tcp_rexmit_min = TCPTV_MIN;
244 tcp_rexmit_slop = TCPTV_CPU_VAR;
246 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
247 if (!powerof2(hashsize)) {
248 printf("WARNING: TCB hash size not a power of 2\n");
249 hashsize = 512; /* safe default */
251 tcp_tcbhashsize = hashsize;
252 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
253 bindhashbase = hashinit(hashsize, M_PCB, &bindhashmask);
254 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
257 for (cpu = 0; cpu < ncpus2; cpu++) {
258 LIST_INIT(&tcbinfo[cpu].listhead);
259 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
260 &tcbinfo[cpu].hashmask);
261 tcbinfo[cpu].porthashbase = porthashbase;
262 tcbinfo[cpu].porthashmask = porthashmask;
263 tcbinfo[cpu].bindhashbase = bindhashbase;
264 tcbinfo[cpu].bindhashmask = bindhashmask;
265 tcbinfo[cpu].ipi_zone = ipi_zone;
268 tcp_reass_maxseg = nmbclusters / 16;
269 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
273 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
275 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
277 if (max_protohdr < TCP_MINPROTOHDR)
278 max_protohdr = TCP_MINPROTOHDR;
279 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
281 #undef TCP_MINPROTOHDR
288 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
289 * tcp_template used to store this data in mbufs, but we now recopy it out
290 * of the tcpcb each time to conserve mbufs.
293 tcp_fillheaders(tp, ip_ptr, tcp_ptr)
298 struct inpcb *inp = tp->t_inpcb;
299 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
302 if ((inp->inp_vflag & INP_IPV6) != 0) {
305 ip6 = (struct ip6_hdr *)ip_ptr;
306 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
307 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
308 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
309 (IPV6_VERSION & IPV6_VERSION_MASK);
310 ip6->ip6_nxt = IPPROTO_TCP;
311 ip6->ip6_plen = sizeof(struct tcphdr);
312 ip6->ip6_src = inp->in6p_laddr;
313 ip6->ip6_dst = inp->in6p_faddr;
318 struct ip *ip = (struct ip *) ip_ptr;
320 ip->ip_vhl = IP_VHL_BORING;
327 ip->ip_p = IPPROTO_TCP;
328 ip->ip_src = inp->inp_laddr;
329 ip->ip_dst = inp->inp_faddr;
330 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
331 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
334 tcp_hdr->th_sport = inp->inp_lport;
335 tcp_hdr->th_dport = inp->inp_fport;
340 tcp_hdr->th_flags = 0;
346 * Create template to be used to send tcp packets on a connection.
347 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
348 * use for this function is in keepalives, which use tcp_respond.
357 m = m_get(M_DONTWAIT, MT_HEADER);
360 m->m_len = sizeof(struct tcptemp);
361 n = mtod(m, struct tcptemp *);
363 tcp_fillheaders(tp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
368 * Send a single message to the TCP at address specified by
369 * the given TCP/IP header. If m == 0, then we make a copy
370 * of the tcpiphdr at ti and send directly to the addressed host.
371 * This is used to force keep alive messages out using the TCP
372 * template for a connection. If flags are given then we send
373 * a message back to the TCP which originated the * segment ti,
374 * and discard the mbuf containing it and any other attached mbufs.
376 * In any case the ack and sequence number of the transmitted
377 * segment are as specified by the parameters.
379 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
382 tcp_respond(tp, ipgen, th, m, ack, seq, flags)
392 struct route *ro = 0;
397 struct route_in6 *ro6 = 0;
398 struct route_in6 sro6;
405 isipv6 = IP_VHL_V(((struct ip *)ipgen)->ip_vhl) == 6;
411 if (!(flags & TH_RST)) {
412 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
413 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
414 win = (long)TCP_MAXWIN << tp->rcv_scale;
418 ro6 = &tp->t_inpcb->in6p_route;
421 ro = &tp->t_inpcb->inp_route;
426 bzero(ro6, sizeof *ro6);
431 bzero(ro, sizeof *ro);
435 m = m_gethdr(M_DONTWAIT, MT_HEADER);
439 m->m_data += max_linkhdr;
442 bcopy((caddr_t)ip6, mtod(m, caddr_t),
443 sizeof(struct ip6_hdr));
444 ip6 = mtod(m, struct ip6_hdr *);
445 nth = (struct tcphdr *)(ip6 + 1);
449 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
450 ip = mtod(m, struct ip *);
451 nth = (struct tcphdr *)(ip + 1);
453 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
458 m->m_data = (caddr_t)ipgen;
459 /* m_len is set later */
461 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
464 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
465 nth = (struct tcphdr *)(ip6 + 1);
469 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
470 nth = (struct tcphdr *)(ip + 1);
474 * this is usually a case when an extension header
475 * exists between the IPv6 header and the
478 nth->th_sport = th->th_sport;
479 nth->th_dport = th->th_dport;
481 xchg(nth->th_dport, nth->th_sport, n_short);
487 ip6->ip6_vfc = IPV6_VERSION;
488 ip6->ip6_nxt = IPPROTO_TCP;
489 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
491 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
495 tlen += sizeof (struct tcpiphdr);
497 ip->ip_ttl = ip_defttl;
500 m->m_pkthdr.len = tlen;
501 m->m_pkthdr.rcvif = (struct ifnet *) 0;
502 nth->th_seq = htonl(seq);
503 nth->th_ack = htonl(ack);
505 nth->th_off = sizeof (struct tcphdr) >> 2;
506 nth->th_flags = flags;
508 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
510 nth->th_win = htons((u_short)win);
515 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
516 sizeof(struct ip6_hdr),
517 tlen - sizeof(struct ip6_hdr));
518 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
525 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
526 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
527 m->m_pkthdr.csum_flags = CSUM_TCP;
528 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
531 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
532 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
536 (void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
537 tp ? tp->t_inpcb : NULL);
538 if (ro6 == &sro6 && ro6->ro_rt) {
545 (void) ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
546 if (ro == &sro && ro->ro_rt) {
554 * Create a new TCP control block, making an
555 * empty reassembly queue and hooking it to the argument
556 * protocol control block. The `inp' parameter must have
557 * come from the zone allocator set up in tcp_init().
566 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
569 it = (struct inp_tp *)inp;
571 bzero((char *) tp, sizeof(struct tcpcb));
572 LIST_INIT(&tp->t_segq);
573 tp->t_maxseg = tp->t_maxopd =
575 isipv6 ? tcp_v6mssdflt :
579 /* Set up our timeouts. */
580 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
581 callout_init(tp->tt_persist = &it->inp_tp_persist);
582 callout_init(tp->tt_keep = &it->inp_tp_keep);
583 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
584 callout_init(tp->tt_delack = &it->inp_tp_delack);
587 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
589 tp->t_flags |= TF_REQ_CC;
590 tp->t_inpcb = inp; /* XXX */
592 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
593 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
594 * reasonable initial retransmit time.
596 tp->t_srtt = TCPTV_SRTTBASE;
597 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
598 tp->t_rttmin = tcp_rexmit_min;
599 tp->t_rxtcur = TCPTV_RTOBASE;
600 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
601 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
602 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
603 tp->t_rcvtime = ticks;
604 tp->t_bw_rtttime = ticks;
606 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
607 * because the socket may be bound to an IPv6 wildcard address,
608 * which may match an IPv4-mapped IPv6 address.
610 inp->inp_ip_ttl = ip_defttl;
611 inp->inp_ppcb = (caddr_t)tp;
612 return (tp); /* XXX */
616 * Drop a TCP connection, reporting
617 * the specified error. If connection is synchronized,
618 * then send a RST to peer.
625 struct socket *so = tp->t_inpcb->inp_socket;
627 if (TCPS_HAVERCVDSYN(tp->t_state)) {
628 tp->t_state = TCPS_CLOSED;
629 (void) tcp_output(tp);
630 tcpstat.tcps_drops++;
632 tcpstat.tcps_conndrops++;
633 if (errno == ETIMEDOUT && tp->t_softerror)
634 errno = tp->t_softerror;
635 so->so_error = errno;
636 return (tcp_close(tp));
640 * Close a TCP control block:
641 * discard all space held by the tcp
642 * discard internet protocol block
643 * wake up any sleepers
650 struct inpcb *inp = tp->t_inpcb;
651 struct socket *so = inp->inp_socket;
653 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
659 * Make sure that all of our timers are stopped before we
662 callout_stop(tp->tt_rexmt);
663 callout_stop(tp->tt_persist);
664 callout_stop(tp->tt_keep);
665 callout_stop(tp->tt_2msl);
666 callout_stop(tp->tt_delack);
669 * If we got enough samples through the srtt filter,
670 * save the rtt and rttvar in the routing entry.
671 * 'Enough' is arbitrarily defined as the 16 samples.
672 * 16 samples is enough for the srtt filter to converge
673 * to within 5% of the correct value; fewer samples and
674 * we could save a very bogus rtt.
676 * Don't update the default route's characteristics and don't
677 * update anything that the user "locked".
679 if (tp->t_rttupdated >= 16) {
683 struct sockaddr_in6 *sin6;
685 if ((rt = inp->in6p_route.ro_rt) == NULL)
687 sin6 = (struct sockaddr_in6 *)rt_key(rt);
688 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
693 if ((rt = inp->inp_route.ro_rt) == NULL ||
694 ((struct sockaddr_in *)rt_key(rt))->sin_addr.s_addr
698 if ((rt->rt_rmx.rmx_locks & RTV_RTT) == 0) {
700 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
701 if (rt->rt_rmx.rmx_rtt && i)
703 * filter this update to half the old & half
704 * the new values, converting scale.
705 * See route.h and tcp_var.h for a
706 * description of the scaling constants.
709 (rt->rt_rmx.rmx_rtt + i) / 2;
711 rt->rt_rmx.rmx_rtt = i;
712 tcpstat.tcps_cachedrtt++;
714 if ((rt->rt_rmx.rmx_locks & RTV_RTTVAR) == 0) {
716 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
717 if (rt->rt_rmx.rmx_rttvar && i)
718 rt->rt_rmx.rmx_rttvar =
719 (rt->rt_rmx.rmx_rttvar + i) / 2;
721 rt->rt_rmx.rmx_rttvar = i;
722 tcpstat.tcps_cachedrttvar++;
725 * The old comment here said:
726 * update the pipelimit (ssthresh) if it has been updated
727 * already or if a pipesize was specified & the threshhold
728 * got below half the pipesize. I.e., wait for bad news
729 * before we start updating, then update on both good
732 * But we want to save the ssthresh even if no pipesize is
733 * specified explicitly in the route, because such
734 * connections still have an implicit pipesize specified
735 * by the global tcp_sendspace. In the absence of a reliable
736 * way to calculate the pipesize, it will have to do.
738 i = tp->snd_ssthresh;
739 if (rt->rt_rmx.rmx_sendpipe != 0)
740 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe / 2);
742 dosavessthresh = (i < so->so_snd.sb_hiwat / 2);
743 if (((rt->rt_rmx.rmx_locks & RTV_SSTHRESH) == 0 &&
744 i != 0 && rt->rt_rmx.rmx_ssthresh != 0)
747 * convert the limit from user data bytes to
748 * packets then to packet data bytes.
750 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
753 i *= (u_long)(tp->t_maxseg +
755 (isipv6 ? sizeof (struct ip6_hdr) +
756 sizeof (struct tcphdr) :
758 sizeof (struct tcpiphdr)
763 if (rt->rt_rmx.rmx_ssthresh)
764 rt->rt_rmx.rmx_ssthresh =
765 (rt->rt_rmx.rmx_ssthresh + i) / 2;
767 rt->rt_rmx.rmx_ssthresh = i;
768 tcpstat.tcps_cachedssthresh++;
772 /* free the reassembly queue, if any */
773 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
774 LIST_REMOVE(q, tqe_q);
779 inp->inp_ppcb = NULL;
780 soisdisconnected(so);
782 if (INP_CHECK_SOCKAF(so, AF_INET6))
787 tcpstat.tcps_closed++;
788 return ((struct tcpcb *)0);
792 tcp_drain_oncpu(struct inpcbhead *head)
796 struct tseg_qent *te;
798 LIST_FOREACH(inpb, head, inp_list) {
799 if ((tcpb = intotcpcb(inpb))) {
800 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
801 LIST_REMOVE(te, tqe_q);
811 struct netmsg_tcp_drain {
812 struct lwkt_msg nm_lmsg;
813 netisr_fn_t nm_handler;
814 struct inpcbhead *nm_head;
817 static int /* really should be void XXX JH */
818 tcp_drain_handler(struct netmsg *msg0)
820 struct netmsg_tcp_drain *nm = (struct netmsg_tcp_drain *)msg0;
822 tcp_drain_oncpu(nm->nm_head);
824 return (0); /* dummy return value */
839 * Walk the tcpbs, if existing, and flush the reassembly queue,
841 * XXX: The "Net/3" implementation doesn't imply that the TCP
842 * reassembly queue should be flushed, but in a situation
843 * where we're really low on mbufs, this is potentially
847 for (cpu = 0; cpu < ncpus2; cpu++) {
848 struct netmsg_tcp_drain *msg;
850 if (cpu == mycpu->gd_cpuid) {
851 tcp_drain_oncpu(&tcbinfo[cpu].listhead);
853 msg = malloc(sizeof(struct netmsg_tcp_drain),
854 M_LWKTMSG, M_NOWAIT);
857 lwkt_initmsg_rp(&msg->nm_lmsg, &netisr_afree_rport,
859 msg->nm_handler = tcp_drain_handler;
860 msg->nm_head = &tcbinfo[cpu].listhead;
861 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
865 tcp_drain_oncpu(&tcbinfo[0].listhead);
870 * Notify a tcp user of an asynchronous error;
871 * store error as soft error, but wake up user
872 * (for now, won't do anything until can select for soft error).
874 * Do not wake up user since there currently is no mechanism for
875 * reporting soft errors (yet - a kqueue filter may be added).
878 tcp_notify(inp, error)
882 struct tcpcb *tp = (struct tcpcb *)inp->inp_ppcb;
885 * Ignore some errors if we are hooked up.
886 * If connection hasn't completed, has retransmitted several times,
887 * and receives a second error, give up now. This is better
888 * than waiting a long time to establish a connection that
889 * can never complete.
891 if (tp->t_state == TCPS_ESTABLISHED &&
892 (error == EHOSTUNREACH || error == ENETUNREACH ||
893 error == EHOSTDOWN)) {
895 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
899 tp->t_softerror = error;
901 wakeup((caddr_t) &so->so_timeo);
908 tcp_pcblist(SYSCTL_HANDLER_ARGS)
911 struct inpcb *inp, **inp_list;
916 * The process of preparing the TCB list is too time-consuming and
917 * resource-intensive to repeat twice on every request.
919 if (req->oldptr == 0) {
920 n = tcbinfo[mycpu->gd_cpuid].ipi_count;
921 req->oldidx = 2 * (sizeof xig)
922 + (n + n/8) * sizeof(struct xtcpcb);
926 if (req->newptr != 0)
930 * OK, now we're committed to doing something.
933 gencnt = tcbinfo[mycpu->gd_cpuid].ipi_gencnt;
934 n = tcbinfo[mycpu->gd_cpuid].ipi_count;
937 xig.xig_len = sizeof xig;
939 xig.xig_gen = gencnt;
940 xig.xig_sogen = so_gencnt;
941 error = SYSCTL_OUT(req, &xig, sizeof xig);
945 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
950 for (inp = LIST_FIRST(&tcbinfo[mycpu->gd_cpuid].listhead), i = 0;
951 inp && i < n; inp = LIST_NEXT(inp, inp_list)) {
952 if (inp->inp_gencnt <= gencnt && !prison_xinpcb(req->td, inp))
959 for (i = 0; i < n; i++) {
961 if (inp->inp_gencnt <= gencnt) {
964 xt.xt_len = sizeof xt;
965 /* XXX should avoid extra copy */
966 bcopy(inp, &xt.xt_inp, sizeof *inp);
967 inp_ppcb = inp->inp_ppcb;
968 if (inp_ppcb != NULL)
969 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
971 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
973 sotoxsocket(inp->inp_socket, &xt.xt_socket);
974 error = SYSCTL_OUT(req, &xt, sizeof xt);
979 * Give the user an updated idea of our state.
980 * If the generation differs from what we told
981 * her before, she knows that something happened
982 * while we were processing this request, and it
983 * might be necessary to retry.
986 xig.xig_gen = tcbinfo[mycpu->gd_cpuid].ipi_gencnt;
987 xig.xig_sogen = so_gencnt;
988 xig.xig_count = tcbinfo[mycpu->gd_cpuid].ipi_count;
990 error = SYSCTL_OUT(req, &xig, sizeof xig);
992 free(inp_list, M_TEMP);
996 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
997 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1000 tcp_getcred(SYSCTL_HANDLER_ARGS)
1002 struct sockaddr_in addrs[2];
1007 error = suser(req->td);
1010 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1014 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1015 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1016 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1017 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1018 if (inp == NULL || inp->inp_socket == NULL) {
1022 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1028 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW,
1029 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1033 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1035 struct sockaddr_in6 addrs[2];
1037 int error, s, mapped = 0;
1039 error = suser(req->td);
1042 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1045 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1046 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1053 inp = in_pcblookup_hash(&tcbinfo[0],
1054 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1056 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1060 inp = in6_pcblookup_hash(&tcbinfo[0],
1061 &addrs[1].sin6_addr, addrs[1].sin6_port,
1062 &addrs[0].sin6_addr, addrs[0].sin6_port,
1065 if (inp == NULL || inp->inp_socket == NULL) {
1069 error = SYSCTL_OUT(req, inp->inp_socket->so_cred,
1070 sizeof(struct ucred));
1076 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW,
1078 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1083 tcp_ctlinput(cmd, sa, vip)
1085 struct sockaddr *sa;
1088 struct ip *ip = vip;
1090 struct in_addr faddr;
1093 void (*notify) (struct inpcb *, int) = tcp_notify;
1098 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1099 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1102 if (cmd == PRC_QUENCH)
1103 notify = tcp_quench;
1104 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1105 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1106 notify = tcp_drop_syn_sent;
1107 else if (cmd == PRC_MSGSIZE)
1108 notify = tcp_mtudisc;
1109 else if (PRC_IS_REDIRECT(cmd)) {
1111 notify = in_rtchange;
1112 } else if (cmd == PRC_HOSTDEAD)
1114 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0)
1118 th = (struct tcphdr *)((caddr_t)ip
1119 + (IP_VHL_HL(ip->ip_vhl) << 2));
1120 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1121 ip->ip_src.s_addr, th->th_sport);
1122 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1123 ip->ip_src, th->th_sport, 0, NULL);
1124 if (inp != NULL && inp->inp_socket != NULL) {
1125 icmp_seq = htonl(th->th_seq);
1126 tp = intotcpcb(inp);
1127 if (SEQ_GEQ(icmp_seq, tp->snd_una) &&
1128 SEQ_LT(icmp_seq, tp->snd_max))
1129 (*notify)(inp, inetctlerrmap[cmd]);
1131 struct in_conninfo inc;
1133 inc.inc_fport = th->th_dport;
1134 inc.inc_lport = th->th_sport;
1135 inc.inc_faddr = faddr;
1136 inc.inc_laddr = ip->ip_src;
1140 syncache_unreach(&inc, th);
1144 for (cpu = 0; cpu < ncpus2; cpu++)
1145 in_pcbnotifyall(&tcbinfo[cpu].listhead, faddr,
1146 inetctlerrmap[cmd], notify);
1152 tcp6_ctlinput(cmd, sa, d)
1154 struct sockaddr *sa;
1158 void (*notify) (struct inpcb *, int) = tcp_notify;
1159 struct ip6_hdr *ip6;
1161 struct ip6ctlparam *ip6cp = NULL;
1162 const struct sockaddr_in6 *sa6_src = NULL;
1164 struct tcp_portonly {
1169 if (sa->sa_family != AF_INET6 ||
1170 sa->sa_len != sizeof(struct sockaddr_in6))
1173 if (cmd == PRC_QUENCH)
1174 notify = tcp_quench;
1175 else if (cmd == PRC_MSGSIZE)
1176 notify = tcp_mtudisc;
1177 else if (!PRC_IS_REDIRECT(cmd) &&
1178 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1181 /* if the parameter is from icmp6, decode it. */
1183 ip6cp = (struct ip6ctlparam *)d;
1185 ip6 = ip6cp->ip6c_ip6;
1186 off = ip6cp->ip6c_off;
1187 sa6_src = ip6cp->ip6c_src;
1191 off = 0; /* fool gcc */
1196 struct in_conninfo inc;
1198 * XXX: We assume that when IPV6 is non NULL,
1199 * M and OFF are valid.
1202 /* check if we can safely examine src and dst ports */
1203 if (m->m_pkthdr.len < off + sizeof(*thp))
1206 bzero(&th, sizeof(th));
1207 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1209 in6_pcbnotify(&tcbinfo[0].listhead, sa, th.th_dport,
1210 (struct sockaddr *)ip6cp->ip6c_src,
1211 th.th_sport, cmd, notify);
1213 inc.inc_fport = th.th_dport;
1214 inc.inc_lport = th.th_sport;
1215 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1216 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1218 syncache_unreach(&inc, &th);
1220 in6_pcbnotify(&tcbinfo[0].listhead, sa, 0,
1221 (const struct sockaddr *)sa6_src, 0, cmd, notify);
1227 * Following is where TCP initial sequence number generation occurs.
1229 * There are two places where we must use initial sequence numbers:
1230 * 1. In SYN-ACK packets.
1231 * 2. In SYN packets.
1233 * All ISNs for SYN-ACK packets are generated by the syncache. See
1234 * tcp_syncache.c for details.
1236 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1237 * depends on this property. In addition, these ISNs should be
1238 * unguessable so as to prevent connection hijacking. To satisfy
1239 * the requirements of this situation, the algorithm outlined in
1240 * RFC 1948 is used to generate sequence numbers.
1242 * Implementation details:
1244 * Time is based off the system timer, and is corrected so that it
1245 * increases by one megabyte per second. This allows for proper
1246 * recycling on high speed LANs while still leaving over an hour
1249 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1250 * between seeding of isn_secret. This is normally set to zero,
1251 * as reseeding should not be necessary.
1255 #define ISN_BYTES_PER_SECOND 1048576
1257 u_char isn_secret[32];
1258 int isn_last_reseed;
1265 u_int32_t md5_buffer[4];
1268 /* Seed if this is the first use, reseed if requested. */
1269 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1270 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1272 read_random_unlimited(&isn_secret, sizeof(isn_secret));
1273 isn_last_reseed = ticks;
1276 /* Compute the md5 hash and return the ISN. */
1278 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1279 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1281 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1282 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1283 sizeof(struct in6_addr));
1284 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1285 sizeof(struct in6_addr));
1289 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1290 sizeof(struct in_addr));
1291 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1292 sizeof(struct in_addr));
1294 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1295 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1296 new_isn = (tcp_seq) md5_buffer[0];
1297 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1302 * When a source quench is received, close congestion window
1303 * to one segment. We will gradually open it again as we proceed.
1306 tcp_quench(inp, errno)
1310 struct tcpcb *tp = intotcpcb(inp);
1313 tp->snd_cwnd = tp->t_maxseg;
1317 * When a specific ICMP unreachable message is received and the
1318 * connection state is SYN-SENT, drop the connection. This behavior
1319 * is controlled by the icmp_may_rst sysctl.
1322 tcp_drop_syn_sent(inp, errno)
1326 struct tcpcb *tp = intotcpcb(inp);
1328 if (tp && tp->t_state == TCPS_SYN_SENT)
1329 tcp_drop(tp, errno);
1333 * When `need fragmentation' ICMP is received, update our idea of the MSS
1334 * based on the new value in the route. Also nudge TCP to send something,
1335 * since we know the packet we just sent was dropped.
1336 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1339 tcp_mtudisc(inp, errno)
1343 struct tcpcb *tp = intotcpcb(inp);
1345 struct rmxp_tao *taop;
1346 struct socket *so = inp->inp_socket;
1350 int isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1356 rt = tcp_rtlookup6(&inp->inp_inc);
1359 rt = tcp_rtlookup(&inp->inp_inc);
1360 if (!rt || !rt->rt_rmx.rmx_mtu) {
1361 tp->t_maxopd = tp->t_maxseg =
1363 isipv6 ? tcp_v6mssdflt :
1368 taop = rmx_taop(rt->rt_rmx);
1369 offered = taop->tao_mssopt;
1370 mss = rt->rt_rmx.rmx_mtu -
1373 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1375 sizeof(struct tcpiphdr)
1382 mss = min(mss, offered);
1384 * XXX - The above conditional probably violates the TCP
1385 * spec. The problem is that, since we don't know the
1386 * other end's MSS, we are supposed to use a conservative
1387 * default. But, if we do that, then MTU discovery will
1388 * never actually take place, because the conservative
1389 * default is much less than the MTUs typically seen
1390 * on the Internet today. For the moment, we'll sweep
1391 * this under the carpet.
1393 * The conservative default might not actually be a problem
1394 * if the only case this occurs is when sending an initial
1395 * SYN with options and data to a host we've never talked
1396 * to before. Then, they will reply with an MSS value which
1397 * will get recorded and the new parameters should get
1398 * recomputed. For Further Study.
1400 if (tp->t_maxopd <= mss)
1404 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1405 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1406 mss -= TCPOLEN_TSTAMP_APPA;
1407 if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
1408 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC)
1409 mss -= TCPOLEN_CC_APPA;
1410 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1412 mss &= ~(MCLBYTES-1);
1415 mss = mss / MCLBYTES * MCLBYTES;
1417 if (so->so_snd.sb_hiwat < mss)
1418 mss = so->so_snd.sb_hiwat;
1422 tcpstat.tcps_mturesent++;
1424 tp->snd_nxt = tp->snd_una;
1430 * Look-up the routing entry to the peer of this inpcb. If no route
1431 * is found and it cannot be allocated the return NULL. This routine
1432 * is called by TCP routines that access the rmx structure and by tcp_mss
1433 * to get the interface MTU.
1437 struct in_conninfo *inc;
1442 ro = &inc->inc_route;
1444 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1445 /* No route yet, so try to acquire one */
1446 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1447 ro->ro_dst.sa_family = AF_INET;
1448 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1449 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1461 struct in_conninfo *inc;
1463 struct route_in6 *ro6;
1466 ro6 = &inc->inc6_route;
1468 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1469 /* No route yet, so try to acquire one */
1470 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1471 ro6->ro_dst.sin6_family = AF_INET6;
1472 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1473 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1474 rtalloc((struct route *)ro6);
1483 /* compute ESP/AH header size for TCP, including outer IP header. */
1485 ipsec_hdrsiz_tcp(tp)
1493 struct ip6_hdr *ip6;
1497 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1499 MGETHDR(m, M_DONTWAIT, MT_DATA);
1504 if ((inp->inp_vflag & INP_IPV6) != 0) {
1505 ip6 = mtod(m, struct ip6_hdr *);
1506 th = (struct tcphdr *)(ip6 + 1);
1507 m->m_pkthdr.len = m->m_len =
1508 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1509 tcp_fillheaders(tp, ip6, th);
1510 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1514 ip = mtod(m, struct ip *);
1515 th = (struct tcphdr *)(ip + 1);
1516 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1517 tcp_fillheaders(tp, ip, th);
1518 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1527 * Return a pointer to the cached information about the remote host.
1528 * The cached information is stored in the protocol specific part of
1529 * the route metrics.
1532 tcp_gettaocache(inc)
1533 struct in_conninfo *inc;
1538 if (inc->inc_isipv6)
1539 rt = tcp_rtlookup6(inc);
1542 rt = tcp_rtlookup(inc);
1544 /* Make sure this is a host route and is up. */
1546 (rt->rt_flags & (RTF_UP|RTF_HOST)) != (RTF_UP|RTF_HOST))
1549 return rmx_taop(rt->rt_rmx);
1553 * Clear all the TAO cache entries, called from tcp_init.
1556 * This routine is just an empty one, because we assume that the routing
1557 * routing tables are initialized at the same time when TCP, so there is
1558 * nothing in the cache left over.
1566 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1568 * This code attempts to calculate the bandwidth-delay product as a
1569 * means of determining the optimal window size to maximize bandwidth,
1570 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1571 * routers. This code also does a fairly good job keeping RTTs in check
1572 * across slow links like modems. We implement an algorithm which is very
1573 * similar (but not meant to be) TCP/Vegas. The code operates on the
1574 * transmitter side of a TCP connection and so only effects the transmit
1575 * side of the connection.
1577 * BACKGROUND: TCP makes no provision for the management of buffer space
1578 * at the end points or at the intermediate routers and switches. A TCP
1579 * stream, whether using NewReno or not, will eventually buffer as
1580 * many packets as it is able and the only reason this typically works is
1581 * due to the fairly small default buffers made available for a connection
1582 * (typicaly 16K or 32K). As machines use larger windows and/or window
1583 * scaling it is now fairly easy for even a single TCP connection to blow-out
1584 * all available buffer space not only on the local interface, but on
1585 * intermediate routers and switches as well. NewReno makes a misguided
1586 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1587 * then backing off, then steadily increasing the window again until another
1588 * failure occurs, ad-infinitum. This results in terrible oscillation that
1589 * is only made worse as network loads increase and the idea of intentionally
1590 * blowing out network buffers is, frankly, a terrible way to manage network
1593 * It is far better to limit the transmit window prior to the failure
1594 * condition being achieved. There are two general ways to do this: First
1595 * you can 'scan' through different transmit window sizes and locate the
1596 * point where the RTT stops increasing, indicating that you have filled the
1597 * pipe, then scan backwards until you note that RTT stops decreasing, then
1598 * repeat ad-infinitum. This method works in principle but has severe
1599 * implementation issues due to RTT variances, timer granularity, and
1600 * instability in the algorithm which can lead to many false positives and
1601 * create oscillations as well as interact badly with other TCP streams
1602 * implementing the same algorithm.
1604 * The second method is to limit the window to the bandwidth delay product
1605 * of the link. This is the method we implement. RTT variances and our
1606 * own manipulation of the congestion window, bwnd, can potentially
1607 * destabilize the algorithm. For this reason we have to stabilize the
1608 * elements used to calculate the window. We do this by using the minimum
1609 * observed RTT, the long term average of the observed bandwidth, and
1610 * by adding two segments worth of slop. It isn't perfect but it is able
1611 * to react to changing conditions and gives us a very stable basis on
1612 * which to extend the algorithm.
1615 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1622 * If inflight_enable is disabled in the middle of a tcp connection,
1623 * make sure snd_bwnd is effectively disabled.
1625 if (tcp_inflight_enable == 0) {
1626 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1627 tp->snd_bandwidth = 0;
1632 * Figure out the bandwidth. Due to the tick granularity this
1633 * is a very rough number and it MUST be averaged over a fairly
1634 * long period of time. XXX we need to take into account a link
1635 * that is not using all available bandwidth, but for now our
1636 * slop will ramp us up if this case occurs and the bandwidth later
1639 * Note: if ticks rollover 'bw' may wind up negative. We must
1640 * effectively reset t_bw_rtttime for this case.
1643 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1646 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1647 (save_ticks - tp->t_bw_rtttime);
1648 tp->t_bw_rtttime = save_ticks;
1649 tp->t_bw_rtseq = ack_seq;
1650 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1652 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1654 tp->snd_bandwidth = bw;
1657 * Calculate the semi-static bandwidth delay product, plus two maximal
1658 * segments. The additional slop puts us squarely in the sweet
1659 * spot and also handles the bandwidth run-up case. Without the
1660 * slop we could be locking ourselves into a lower bandwidth.
1662 * Situations Handled:
1663 * (1) Prevents over-queueing of packets on LANs, especially on
1664 * high speed LANs, allowing larger TCP buffers to be
1665 * specified, and also does a good job preventing
1666 * over-queueing of packets over choke points like modems
1667 * (at least for the transmit side).
1669 * (2) Is able to handle changing network loads (bandwidth
1670 * drops so bwnd drops, bandwidth increases so bwnd
1673 * (3) Theoretically should stabilize in the face of multiple
1674 * connections implementing the same algorithm (this may need
1677 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1678 * be adjusted with a sysctl but typically only needs to be on
1679 * very slow connections. A value no smaller then 5 should
1680 * be used, but only reduce this default if you have no other
1683 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1684 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * (int)tp->t_maxseg / 10;
1687 if (tcp_inflight_debug > 0) {
1689 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1691 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1700 if ((long)bwnd < tcp_inflight_min)
1701 bwnd = tcp_inflight_min;
1702 if (bwnd > tcp_inflight_max)
1703 bwnd = tcp_inflight_max;
1704 if ((long)bwnd < tp->t_maxseg * 2)
1705 bwnd = tp->t_maxseg * 2;
1706 tp->snd_bwnd = bwnd;