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 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
33 #include "opt_compat.h"
35 #include "opt_inet6.h"
36 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
39 #include "opt_tcp_sack.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/callout.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
47 #include <sys/malloc.h>
50 #include <sys/domain.h>
53 #include <sys/socket.h>
54 #include <sys/socketvar.h>
55 #include <sys/protosw.h>
56 #include <sys/random.h>
60 #include <net/route.h>
63 #include <netinet/in.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
67 #include <netinet/ip6.h>
69 #include <netinet/in_pcb.h>
71 #include <netinet6/in6_pcb.h>
73 #include <netinet/in_var.h>
74 #include <netinet/ip_var.h>
76 #include <netinet6/ip6_var.h>
77 #include <netinet6/scope6_var.h>
78 #include <netinet6/nd6.h>
80 #include <netinet/ip_icmp.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>
100 #include <netkey/key.h>
104 #include <netipsec/ipsec.h>
105 #include <netipsec/xform.h>
107 #include <netipsec/ipsec6.h>
109 #include <netipsec/key.h>
111 #endif /*FAST_IPSEC*/
113 #include <machine/in_cksum.h>
116 int tcp_mssdflt = TCP_MSS;
117 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
118 &tcp_mssdflt , 0, "Default TCP Maximum Segment Size");
121 int tcp_v6mssdflt = TCP6_MSS;
122 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
123 CTLFLAG_RW, &tcp_v6mssdflt , 0,
124 "Default TCP Maximum Segment Size for IPv6");
128 * Minimum MSS we accept and use. This prevents DoS attacks where
129 * we are forced to a ridiculous low MSS like 20 and send hundreds
130 * of packets instead of one. The effect scales with the available
131 * bandwidth and quickly saturates the CPU and network interface
132 * with packet generation and sending. Set to zero to disable MINMSS
133 * checking. This setting prevents us from sending too small packets.
135 int tcp_minmss = TCP_MINMSS;
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
137 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
139 * Number of TCP segments per second we accept from remote host
140 * before we start to calculate average segment size. If average
141 * segment size drops below the minimum TCP MSS we assume a DoS
142 * attack and reset+drop the connection. Care has to be taken not to
143 * set this value too small to not kill interactive type connections
144 * (telnet, SSH) which send many small packets.
146 int tcp_minmssoverload = TCP_MINMSSOVERLOAD;
147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmssoverload, CTLFLAG_RW,
148 &tcp_minmssoverload , 0, "Number of TCP Segments per Second allowed to"
149 "be under the MINMSS Size");
152 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
153 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
154 &tcp_rttdflt , 0, "Default maximum TCP Round Trip Time");
157 int tcp_do_rfc1323 = 1;
158 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
159 &tcp_do_rfc1323 , 0, "Enable rfc1323 (high performance TCP) extensions");
161 static int tcp_tcbhashsize = 0;
162 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN,
163 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
165 static int do_tcpdrain = 1;
166 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
167 "Enable tcp_drain routine for extra help when low on mbufs");
169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
170 &tcbinfo.ipi_count, 0, "Number of active PCBs");
172 static int icmp_may_rst = 1;
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
174 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
176 static int tcp_isn_reseed_interval = 0;
177 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
178 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
181 * TCP bandwidth limiting sysctls. Note that the default lower bound of
182 * 1024 exists only for debugging. A good production default would be
183 * something like 6100.
185 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, inflight, CTLFLAG_RW, 0,
186 "TCP inflight data limiting");
188 static int tcp_inflight_enable = 1;
189 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, enable, CTLFLAG_RW,
190 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
192 static int tcp_inflight_debug = 0;
193 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, debug, CTLFLAG_RW,
194 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
196 static int tcp_inflight_min = 6144;
197 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW,
198 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
200 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
201 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW,
202 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
204 static int tcp_inflight_stab = 20;
205 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW,
206 &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets");
208 uma_zone_t sack_hole_zone;
210 static struct inpcb *tcp_notify(struct inpcb *, int);
211 static void tcp_discardcb(struct tcpcb *);
212 static void tcp_isn_tick(void *);
215 * Target size of TCP PCB hash tables. Must be a power of two.
217 * Note that this can be overridden by the kernel environment
218 * variable net.inet.tcp.tcbhashsize
221 #define TCBHASHSIZE 512
226 * Callouts should be moved into struct tcp directly. They are currently
227 * separate because the tcpcb structure is exported to userland for sysctl
228 * parsing purposes, which do not know about callouts.
232 struct callout tcpcb_mem_rexmt, tcpcb_mem_persist, tcpcb_mem_keep;
233 struct callout tcpcb_mem_2msl, tcpcb_mem_delack;
236 static uma_zone_t tcpcb_zone;
237 static uma_zone_t tcptw_zone;
238 struct callout isn_callout;
246 int hashsize = TCBHASHSIZE;
248 tcp_delacktime = TCPTV_DELACK;
249 tcp_keepinit = TCPTV_KEEP_INIT;
250 tcp_keepidle = TCPTV_KEEP_IDLE;
251 tcp_keepintvl = TCPTV_KEEPINTVL;
252 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
254 tcp_rexmit_min = TCPTV_MIN;
255 tcp_rexmit_slop = TCPTV_CPU_VAR;
257 INP_INFO_LOCK_INIT(&tcbinfo, "tcp");
259 tcbinfo.listhead = &tcb;
260 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
261 if (!powerof2(hashsize)) {
262 printf("WARNING: TCB hash size not a power of 2\n");
263 hashsize = 512; /* safe default */
265 tcp_tcbhashsize = hashsize;
266 tcbinfo.hashbase = hashinit(hashsize, M_PCB, &tcbinfo.hashmask);
267 tcbinfo.porthashbase = hashinit(hashsize, M_PCB,
268 &tcbinfo.porthashmask);
269 tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb),
270 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
271 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
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
283 * These have to be type stable for the benefit of the timers.
285 tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem),
286 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
287 uma_zone_set_max(tcpcb_zone, maxsockets);
288 tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
289 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
290 uma_zone_set_max(tcptw_zone, maxsockets / 5);
295 callout_init(&isn_callout, CALLOUT_MPSAFE);
297 EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL,
298 SHUTDOWN_PRI_DEFAULT);
299 sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole),
300 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 callout_stop(&isn_callout);
312 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
313 * tcp_template used to store this data in mbufs, but we now recopy it out
314 * of the tcpcb each time to conserve mbufs.
317 tcpip_fillheaders(inp, ip_ptr, tcp_ptr)
322 struct tcphdr *th = (struct tcphdr *)tcp_ptr;
324 INP_LOCK_ASSERT(inp);
327 if ((inp->inp_vflag & INP_IPV6) != 0) {
330 ip6 = (struct ip6_hdr *)ip_ptr;
331 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
332 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
333 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
334 (IPV6_VERSION & IPV6_VERSION_MASK);
335 ip6->ip6_nxt = IPPROTO_TCP;
336 ip6->ip6_plen = sizeof(struct tcphdr);
337 ip6->ip6_src = inp->in6p_laddr;
338 ip6->ip6_dst = inp->in6p_faddr;
344 ip = (struct ip *)ip_ptr;
345 ip->ip_v = IPVERSION;
347 ip->ip_tos = inp->inp_ip_tos;
351 ip->ip_ttl = inp->inp_ip_ttl;
353 ip->ip_p = IPPROTO_TCP;
354 ip->ip_src = inp->inp_laddr;
355 ip->ip_dst = inp->inp_faddr;
357 th->th_sport = inp->inp_lport;
358 th->th_dport = inp->inp_fport;
366 th->th_sum = 0; /* in_pseudo() is called later for ipv4 */
370 * Create template to be used to send tcp packets on a connection.
371 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
372 * use for this function is in keepalives, which use tcp_respond.
375 tcpip_maketemplate(inp)
381 m = m_get(M_DONTWAIT, MT_DATA);
384 m->m_len = sizeof(struct tcptemp);
385 n = mtod(m, struct tcptemp *);
387 tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
392 * Send a single message to the TCP at address specified by
393 * the given TCP/IP header. If m == NULL, then we make a copy
394 * of the tcpiphdr at ti and send directly to the addressed host.
395 * This is used to force keep alive messages out using the TCP
396 * template for a connection. If flags are given then we send
397 * a message back to the TCP which originated the * segment ti,
398 * and discard the mbuf containing it and any other attached mbufs.
400 * In any case the ack and sequence number of the transmitted
401 * segment are as specified by the parameters.
403 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
406 tcp_respond(tp, ipgen, th, m, ack, seq, flags)
409 register struct tcphdr *th;
410 register struct mbuf *m;
425 KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL"));
428 isipv6 = ((struct ip *)ipgen)->ip_v == 6;
435 KASSERT(inp != NULL, ("tcp control block w/o inpcb"));
436 INP_INFO_WLOCK_ASSERT(&tcbinfo);
437 INP_LOCK_ASSERT(inp);
442 if (!(flags & TH_RST)) {
443 win = sbspace(&inp->inp_socket->so_rcv);
444 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
445 win = (long)TCP_MAXWIN << tp->rcv_scale;
449 m = m_gethdr(M_DONTWAIT, MT_DATA);
453 m->m_data += max_linkhdr;
456 bcopy((caddr_t)ip6, mtod(m, caddr_t),
457 sizeof(struct ip6_hdr));
458 ip6 = mtod(m, struct ip6_hdr *);
459 nth = (struct tcphdr *)(ip6 + 1);
463 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
464 ip = mtod(m, struct ip *);
465 nth = (struct tcphdr *)(ip + 1);
467 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
472 m->m_data = (caddr_t)ipgen;
473 /* m_len is set later */
475 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
478 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
479 nth = (struct tcphdr *)(ip6 + 1);
483 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
484 nth = (struct tcphdr *)(ip + 1);
488 * this is usually a case when an extension header
489 * exists between the IPv6 header and the
492 nth->th_sport = th->th_sport;
493 nth->th_dport = th->th_dport;
495 xchg(nth->th_dport, nth->th_sport, n_short);
501 ip6->ip6_vfc = IPV6_VERSION;
502 ip6->ip6_nxt = IPPROTO_TCP;
503 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
505 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
509 tlen += sizeof (struct tcpiphdr);
511 ip->ip_ttl = ip_defttl;
512 if (path_mtu_discovery)
516 m->m_pkthdr.len = tlen;
517 m->m_pkthdr.rcvif = NULL;
521 * Packet is associated with a socket, so allow the
522 * label of the response to reflect the socket label.
524 INP_LOCK_ASSERT(inp);
525 mac_create_mbuf_from_inpcb(inp, m);
528 * Packet is not associated with a socket, so possibly
529 * update the label in place.
531 mac_reflect_mbuf_tcp(m);
534 nth->th_seq = htonl(seq);
535 nth->th_ack = htonl(ack);
537 nth->th_off = sizeof (struct tcphdr) >> 2;
538 nth->th_flags = flags;
540 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
542 nth->th_win = htons((u_short)win);
547 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
548 sizeof(struct ip6_hdr),
549 tlen - sizeof(struct ip6_hdr));
550 ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb :
555 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
556 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
557 m->m_pkthdr.csum_flags = CSUM_TCP;
558 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
561 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
562 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
566 (void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp);
569 (void) ip_output(m, NULL, NULL, ipflags, NULL, inp);
573 * Create a new TCP control block, making an
574 * empty reassembly queue and hooking it to the argument
575 * protocol control block. The `inp' parameter must have
576 * come from the zone allocator set up in tcp_init().
582 struct tcpcb_mem *tm;
585 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
588 tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO);
592 /* LIST_INIT(&tp->t_segq); */ /* XXX covered by M_ZERO */
593 tp->t_maxseg = tp->t_maxopd =
595 isipv6 ? tcp_v6mssdflt :
599 /* Set up our timeouts. */
600 callout_init(tp->tt_rexmt = &tm->tcpcb_mem_rexmt, NET_CALLOUT_MPSAFE);
601 callout_init(tp->tt_persist = &tm->tcpcb_mem_persist, NET_CALLOUT_MPSAFE);
602 callout_init(tp->tt_keep = &tm->tcpcb_mem_keep, NET_CALLOUT_MPSAFE);
603 callout_init(tp->tt_2msl = &tm->tcpcb_mem_2msl, NET_CALLOUT_MPSAFE);
604 callout_init(tp->tt_delack = &tm->tcpcb_mem_delack, NET_CALLOUT_MPSAFE);
607 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
608 tp->sack_enable = tcp_do_sack;
609 TAILQ_INIT(&tp->snd_holes);
610 tp->t_inpcb = inp; /* XXX */
612 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
613 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
614 * reasonable initial retransmit time.
616 tp->t_srtt = TCPTV_SRTTBASE;
617 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
618 tp->t_rttmin = tcp_rexmit_min;
619 tp->t_rxtcur = TCPTV_RTOBASE;
620 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
621 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
622 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
623 tp->t_rcvtime = ticks;
624 tp->t_bw_rtttime = ticks;
626 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
627 * because the socket may be bound to an IPv6 wildcard address,
628 * which may match an IPv4-mapped IPv6 address.
630 inp->inp_ip_ttl = ip_defttl;
631 inp->inp_ppcb = (caddr_t)tp;
632 return (tp); /* XXX */
636 * Drop a TCP connection, reporting
637 * the specified error. If connection is synchronized,
638 * then send a RST to peer.
642 register struct tcpcb *tp;
645 struct socket *so = tp->t_inpcb->inp_socket;
647 INP_INFO_WLOCK_ASSERT(&tcbinfo);
648 INP_LOCK_ASSERT(tp->t_inpcb);
650 if (TCPS_HAVERCVDSYN(tp->t_state)) {
651 tp->t_state = TCPS_CLOSED;
652 (void) tcp_output(tp);
653 tcpstat.tcps_drops++;
655 tcpstat.tcps_conndrops++;
656 if (errno == ETIMEDOUT && tp->t_softerror)
657 errno = tp->t_softerror;
658 so->so_error = errno;
659 return (tcp_close(tp));
667 struct inpcb *inp = tp->t_inpcb;
668 struct socket *so = inp->inp_socket;
670 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
673 INP_LOCK_ASSERT(inp);
676 * Make sure that all of our timers are stopped before we
679 callout_stop(tp->tt_rexmt);
680 callout_stop(tp->tt_persist);
681 callout_stop(tp->tt_keep);
682 callout_stop(tp->tt_2msl);
683 callout_stop(tp->tt_delack);
686 * If we got enough samples through the srtt filter,
687 * save the rtt and rttvar in the routing entry.
688 * 'Enough' is arbitrarily defined as 4 rtt samples.
689 * 4 samples is enough for the srtt filter to converge
690 * to within enough % of the correct value; fewer samples
691 * and we could save a bogus rtt. The danger is not high
692 * as tcp quickly recovers from everything.
693 * XXX: Works very well but needs some more statistics!
695 if (tp->t_rttupdated >= 4) {
696 struct hc_metrics_lite metrics;
699 bzero(&metrics, sizeof(metrics));
701 * Update the ssthresh always when the conditions below
702 * are satisfied. This gives us better new start value
703 * for the congestion avoidance for new connections.
704 * ssthresh is only set if packet loss occured on a session.
706 ssthresh = tp->snd_ssthresh;
707 if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) {
709 * convert the limit from user data bytes to
710 * packets then to packet data bytes.
712 ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg;
715 ssthresh *= (u_long)(tp->t_maxseg +
717 (isipv6 ? sizeof (struct ip6_hdr) +
718 sizeof (struct tcphdr) :
720 sizeof (struct tcpiphdr)
727 metrics.rmx_ssthresh = ssthresh;
729 metrics.rmx_rtt = tp->t_srtt;
730 metrics.rmx_rttvar = tp->t_rttvar;
731 /* XXX: This wraps if the pipe is more than 4 Gbit per second */
732 metrics.rmx_bandwidth = tp->snd_bandwidth;
733 metrics.rmx_cwnd = tp->snd_cwnd;
734 metrics.rmx_sendpipe = 0;
735 metrics.rmx_recvpipe = 0;
737 tcp_hc_update(&inp->inp_inc, &metrics);
740 /* free the reassembly queue, if any */
741 while ((q = LIST_FIRST(&tp->t_segq)) != NULL) {
742 LIST_REMOVE(q, tqe_q);
744 uma_zfree(tcp_reass_zone, q);
748 tcp_free_sackholes(tp);
749 inp->inp_ppcb = NULL;
751 uma_zfree(tcpcb_zone, tp);
752 soisdisconnected(so);
756 * Close a TCP control block:
757 * discard all space held by the tcp
758 * discard internet protocol block
759 * wake up any sleepers
765 struct inpcb *inp = tp->t_inpcb;
767 struct socket *so = inp->inp_socket;
770 INP_INFO_WLOCK_ASSERT(&tcbinfo);
771 INP_LOCK_ASSERT(inp);
775 if (INP_CHECK_SOCKAF(so, AF_INET6))
780 tcpstat.tcps_closed++;
791 struct tseg_qent *te;
794 * Walk the tcpbs, if existing, and flush the reassembly queue,
796 * XXX: The "Net/3" implementation doesn't imply that the TCP
797 * reassembly queue should be flushed, but in a situation
798 * where we're really low on mbufs, this is potentially
801 INP_INFO_RLOCK(&tcbinfo);
802 LIST_FOREACH(inpb, tcbinfo.listhead, inp_list) {
803 if (inpb->inp_vflag & INP_TIMEWAIT)
806 if ((tcpb = intotcpcb(inpb)) != NULL) {
807 while ((te = LIST_FIRST(&tcpb->t_segq))
809 LIST_REMOVE(te, tqe_q);
811 uma_zfree(tcp_reass_zone, te);
815 tcp_clean_sackreport(tcpb);
819 INP_INFO_RUNLOCK(&tcbinfo);
824 * Notify a tcp user of an asynchronous error;
825 * store error as soft error, but wake up user
826 * (for now, won't do anything until can select for soft error).
828 * Do not wake up user since there currently is no mechanism for
829 * reporting soft errors (yet - a kqueue filter may be added).
831 static struct inpcb *
832 tcp_notify(inp, error)
836 struct tcpcb *tp = (struct tcpcb *)inp->inp_ppcb;
838 INP_INFO_WLOCK_ASSERT(&tcbinfo);
839 INP_LOCK_ASSERT(inp);
842 * Ignore some errors if we are hooked up.
843 * If connection hasn't completed, has retransmitted several times,
844 * and receives a second error, give up now. This is better
845 * than waiting a long time to establish a connection that
846 * can never complete.
848 if (tp->t_state == TCPS_ESTABLISHED &&
849 (error == EHOSTUNREACH || error == ENETUNREACH ||
850 error == EHOSTDOWN)) {
852 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
855 return (struct inpcb *)0;
857 tp->t_softerror = error;
861 wakeup( &so->so_timeo);
868 tcp_pcblist(SYSCTL_HANDLER_ARGS)
871 struct inpcb *inp, **inp_list;
876 * The process of preparing the TCB list is too time-consuming and
877 * resource-intensive to repeat twice on every request.
879 if (req->oldptr == NULL) {
880 n = tcbinfo.ipi_count;
881 req->oldidx = 2 * (sizeof xig)
882 + (n + n/8) * sizeof(struct xtcpcb);
886 if (req->newptr != NULL)
890 * OK, now we're committed to doing something.
892 INP_INFO_RLOCK(&tcbinfo);
893 gencnt = tcbinfo.ipi_gencnt;
894 n = tcbinfo.ipi_count;
895 INP_INFO_RUNLOCK(&tcbinfo);
897 error = sysctl_wire_old_buffer(req, 2 * (sizeof xig)
898 + n * sizeof(struct xtcpcb));
902 xig.xig_len = sizeof xig;
904 xig.xig_gen = gencnt;
905 xig.xig_sogen = so_gencnt;
906 error = SYSCTL_OUT(req, &xig, sizeof xig);
910 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
911 if (inp_list == NULL)
914 INP_INFO_RLOCK(&tcbinfo);
915 for (inp = LIST_FIRST(tcbinfo.listhead), i = 0; inp != NULL && i < n;
916 inp = LIST_NEXT(inp, inp_list)) {
918 if (inp->inp_gencnt <= gencnt) {
920 * XXX: This use of cr_cansee(), introduced with
921 * TCP state changes, is not quite right, but for
922 * now, better than nothing.
924 if (inp->inp_vflag & INP_TIMEWAIT)
925 error = cr_cansee(req->td->td_ucred,
926 intotw(inp)->tw_cred);
928 error = cr_canseesocket(req->td->td_ucred,
935 INP_INFO_RUNLOCK(&tcbinfo);
939 for (i = 0; i < n; i++) {
941 if (inp->inp_gencnt <= gencnt) {
945 bzero(&xt, sizeof(xt));
946 xt.xt_len = sizeof xt;
947 /* XXX should avoid extra copy */
948 bcopy(inp, &xt.xt_inp, sizeof *inp);
949 inp_ppcb = inp->inp_ppcb;
950 if (inp_ppcb == NULL)
951 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
952 else if (inp->inp_vflag & INP_TIMEWAIT) {
953 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
954 xt.xt_tp.t_state = TCPS_TIME_WAIT;
956 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
957 if (inp->inp_socket != NULL)
958 sotoxsocket(inp->inp_socket, &xt.xt_socket);
960 bzero(&xt.xt_socket, sizeof xt.xt_socket);
961 xt.xt_socket.xso_protocol = IPPROTO_TCP;
963 xt.xt_inp.inp_gencnt = inp->inp_gencnt;
964 error = SYSCTL_OUT(req, &xt, sizeof xt);
969 * Give the user an updated idea of our state.
970 * If the generation differs from what we told
971 * her before, she knows that something happened
972 * while we were processing this request, and it
973 * might be necessary to retry.
975 INP_INFO_RLOCK(&tcbinfo);
976 xig.xig_gen = tcbinfo.ipi_gencnt;
977 xig.xig_sogen = so_gencnt;
978 xig.xig_count = tcbinfo.ipi_count;
979 INP_INFO_RUNLOCK(&tcbinfo);
980 error = SYSCTL_OUT(req, &xig, sizeof xig);
982 free(inp_list, M_TEMP);
986 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
987 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
990 tcp_getcred(SYSCTL_HANDLER_ARGS)
993 struct sockaddr_in addrs[2];
997 error = suser_cred(req->td->td_ucred, SUSER_ALLOWJAIL);
1000 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1003 INP_INFO_RLOCK(&tcbinfo);
1004 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
1005 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1011 if (inp->inp_socket == NULL) {
1015 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1018 cru2x(inp->inp_socket->so_cred, &xuc);
1022 INP_INFO_RUNLOCK(&tcbinfo);
1024 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1028 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred,
1029 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1030 tcp_getcred, "S,xucred", "Get the xucred of a TCP connection");
1034 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1037 struct sockaddr_in6 addrs[2];
1039 int error, mapped = 0;
1041 error = suser_cred(req->td->td_ucred, SUSER_ALLOWJAIL);
1044 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1047 if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 ||
1048 (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) {
1051 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1052 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1058 INP_INFO_RLOCK(&tcbinfo);
1060 inp = in_pcblookup_hash(&tcbinfo,
1061 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1063 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1067 inp = in6_pcblookup_hash(&tcbinfo,
1068 &addrs[1].sin6_addr, addrs[1].sin6_port,
1069 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1075 if (inp->inp_socket == NULL) {
1079 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1082 cru2x(inp->inp_socket->so_cred, &xuc);
1086 INP_INFO_RUNLOCK(&tcbinfo);
1088 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1092 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred,
1093 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1094 tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection");
1099 tcp_ctlinput(cmd, sa, vip)
1101 struct sockaddr *sa;
1104 struct ip *ip = vip;
1106 struct in_addr faddr;
1109 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1111 struct in_conninfo inc;
1112 tcp_seq icmp_tcp_seq;
1115 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1116 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1119 if (cmd == PRC_MSGSIZE)
1120 notify = tcp_mtudisc;
1121 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1122 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1123 notify = tcp_drop_syn_sent;
1125 * Redirects don't need to be handled up here.
1127 else if (PRC_IS_REDIRECT(cmd))
1130 * Source quench is depreciated.
1132 else if (cmd == PRC_QUENCH)
1135 * Hostdead is ugly because it goes linearly through all PCBs.
1136 * XXX: We never get this from ICMP, otherwise it makes an
1137 * excellent DoS attack on machines with many connections.
1139 else if (cmd == PRC_HOSTDEAD)
1141 else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0)
1144 icp = (struct icmp *)((caddr_t)ip
1145 - offsetof(struct icmp, icmp_ip));
1146 th = (struct tcphdr *)((caddr_t)ip
1147 + (ip->ip_hl << 2));
1148 INP_INFO_WLOCK(&tcbinfo);
1149 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1150 ip->ip_src, th->th_sport, 0, NULL);
1153 if (inp->inp_socket != NULL) {
1154 icmp_tcp_seq = htonl(th->th_seq);
1155 tp = intotcpcb(inp);
1156 if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) &&
1157 SEQ_LT(icmp_tcp_seq, tp->snd_max)) {
1158 if (cmd == PRC_MSGSIZE) {
1161 * If we got a needfrag set the MTU
1162 * in the route to the suggested new
1163 * value (if given) and then notify.
1165 bzero(&inc, sizeof(inc));
1166 inc.inc_flags = 0; /* IPv4 */
1167 inc.inc_faddr = faddr;
1169 mtu = ntohs(icp->icmp_nextmtu);
1171 * If no alternative MTU was
1172 * proposed, try the next smaller
1173 * one. ip->ip_len has already
1174 * been swapped in icmp_input().
1177 mtu = ip_next_mtu(ip->ip_len,
1179 if (mtu < max(296, (tcp_minmss)
1180 + sizeof(struct tcpiphdr)))
1184 + sizeof(struct tcpiphdr);
1186 * Only cache the the MTU if it
1187 * is smaller than the interface
1188 * or route MTU. tcp_mtudisc()
1189 * will do right thing by itself.
1191 if (mtu <= tcp_maxmtu(&inc))
1192 tcp_hc_updatemtu(&inc, mtu);
1195 inp = (*notify)(inp, inetctlerrmap[cmd]);
1201 inc.inc_fport = th->th_dport;
1202 inc.inc_lport = th->th_sport;
1203 inc.inc_faddr = faddr;
1204 inc.inc_laddr = ip->ip_src;
1208 syncache_unreach(&inc, th);
1210 INP_INFO_WUNLOCK(&tcbinfo);
1212 in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify);
1217 tcp6_ctlinput(cmd, sa, d)
1219 struct sockaddr *sa;
1223 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1224 struct ip6_hdr *ip6;
1226 struct ip6ctlparam *ip6cp = NULL;
1227 const struct sockaddr_in6 *sa6_src = NULL;
1229 struct tcp_portonly {
1234 if (sa->sa_family != AF_INET6 ||
1235 sa->sa_len != sizeof(struct sockaddr_in6))
1238 if (cmd == PRC_MSGSIZE)
1239 notify = tcp_mtudisc;
1240 else if (!PRC_IS_REDIRECT(cmd) &&
1241 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1243 /* Source quench is depreciated. */
1244 else if (cmd == PRC_QUENCH)
1247 /* if the parameter is from icmp6, decode it. */
1249 ip6cp = (struct ip6ctlparam *)d;
1251 ip6 = ip6cp->ip6c_ip6;
1252 off = ip6cp->ip6c_off;
1253 sa6_src = ip6cp->ip6c_src;
1257 off = 0; /* fool gcc */
1262 struct in_conninfo inc;
1264 * XXX: We assume that when IPV6 is non NULL,
1265 * M and OFF are valid.
1268 /* check if we can safely examine src and dst ports */
1269 if (m->m_pkthdr.len < off + sizeof(*thp))
1272 bzero(&th, sizeof(th));
1273 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1275 in6_pcbnotify(&tcbinfo, sa, th.th_dport,
1276 (struct sockaddr *)ip6cp->ip6c_src,
1277 th.th_sport, cmd, NULL, notify);
1279 inc.inc_fport = th.th_dport;
1280 inc.inc_lport = th.th_sport;
1281 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1282 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1284 INP_INFO_WLOCK(&tcbinfo);
1285 syncache_unreach(&inc, &th);
1286 INP_INFO_WUNLOCK(&tcbinfo);
1288 in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src,
1289 0, cmd, NULL, notify);
1295 * Following is where TCP initial sequence number generation occurs.
1297 * There are two places where we must use initial sequence numbers:
1298 * 1. In SYN-ACK packets.
1299 * 2. In SYN packets.
1301 * All ISNs for SYN-ACK packets are generated by the syncache. See
1302 * tcp_syncache.c for details.
1304 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1305 * depends on this property. In addition, these ISNs should be
1306 * unguessable so as to prevent connection hijacking. To satisfy
1307 * the requirements of this situation, the algorithm outlined in
1308 * RFC 1948 is used, with only small modifications.
1310 * Implementation details:
1312 * Time is based off the system timer, and is corrected so that it
1313 * increases by one megabyte per second. This allows for proper
1314 * recycling on high speed LANs while still leaving over an hour
1317 * As reading the *exact* system time is too expensive to be done
1318 * whenever setting up a TCP connection, we increment the time
1319 * offset in two ways. First, a small random positive increment
1320 * is added to isn_offset for each connection that is set up.
1321 * Second, the function tcp_isn_tick fires once per clock tick
1322 * and increments isn_offset as necessary so that sequence numbers
1323 * are incremented at approximately ISN_BYTES_PER_SECOND. The
1324 * random positive increments serve only to ensure that the same
1325 * exact sequence number is never sent out twice (as could otherwise
1326 * happen when a port is recycled in less than the system tick
1329 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1330 * between seeding of isn_secret. This is normally set to zero,
1331 * as reseeding should not be necessary.
1333 * Locking of the global variables isn_secret, isn_last_reseed, isn_offset,
1334 * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock. In
1335 * general, this means holding an exclusive (write) lock.
1338 #define ISN_BYTES_PER_SECOND 1048576
1339 #define ISN_STATIC_INCREMENT 4096
1340 #define ISN_RANDOM_INCREMENT (4096 - 1)
1342 static u_char isn_secret[32];
1343 static int isn_last_reseed;
1344 static u_int32_t isn_offset, isn_offset_old;
1345 static MD5_CTX isn_ctx;
1351 u_int32_t md5_buffer[4];
1354 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1355 INP_LOCK_ASSERT(tp->t_inpcb);
1357 /* Seed if this is the first use, reseed if requested. */
1358 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1359 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1361 read_random(&isn_secret, sizeof(isn_secret));
1362 isn_last_reseed = ticks;
1365 /* Compute the md5 hash and return the ISN. */
1367 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1368 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1370 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1371 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1372 sizeof(struct in6_addr));
1373 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1374 sizeof(struct in6_addr));
1378 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1379 sizeof(struct in_addr));
1380 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1381 sizeof(struct in_addr));
1383 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1384 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1385 new_isn = (tcp_seq) md5_buffer[0];
1386 isn_offset += ISN_STATIC_INCREMENT +
1387 (arc4random() & ISN_RANDOM_INCREMENT);
1388 new_isn += isn_offset;
1393 * Increment the offset to the next ISN_BYTES_PER_SECOND / hz boundary
1394 * to keep time flowing at a relatively constant rate. If the random
1395 * increments have already pushed us past the projected offset, do nothing.
1401 u_int32_t projected_offset;
1403 INP_INFO_WLOCK(&tcbinfo);
1404 projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100;
1406 if (projected_offset > isn_offset)
1407 isn_offset = projected_offset;
1409 isn_offset_old = isn_offset;
1410 callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL);
1411 INP_INFO_WUNLOCK(&tcbinfo);
1415 * When a specific ICMP unreachable message is received and the
1416 * connection state is SYN-SENT, drop the connection. This behavior
1417 * is controlled by the icmp_may_rst sysctl.
1420 tcp_drop_syn_sent(inp, errno)
1424 struct tcpcb *tp = intotcpcb(inp);
1426 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1427 INP_LOCK_ASSERT(inp);
1429 if (tp != NULL && tp->t_state == TCPS_SYN_SENT) {
1430 tcp_drop(tp, errno);
1437 * When `need fragmentation' ICMP is received, update our idea of the MSS
1438 * based on the new value in the route. Also nudge TCP to send something,
1439 * since we know the packet we just sent was dropped.
1440 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1443 tcp_mtudisc(inp, errno)
1447 struct tcpcb *tp = intotcpcb(inp);
1448 struct socket *so = inp->inp_socket;
1456 INP_LOCK_ASSERT(inp);
1459 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1461 maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */
1464 isipv6 ? tcp_maxmtu6(&inp->inp_inc) :
1466 tcp_maxmtu(&inp->inp_inc);
1470 maxmtu = min(maxmtu, romtu);
1472 tp->t_maxopd = tp->t_maxseg =
1474 isipv6 ? tcp_v6mssdflt :
1482 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1484 sizeof(struct tcpiphdr)
1491 * XXX - The above conditional probably violates the TCP
1492 * spec. The problem is that, since we don't know the
1493 * other end's MSS, we are supposed to use a conservative
1494 * default. But, if we do that, then MTU discovery will
1495 * never actually take place, because the conservative
1496 * default is much less than the MTUs typically seen
1497 * on the Internet today. For the moment, we'll sweep
1498 * this under the carpet.
1500 * The conservative default might not actually be a problem
1501 * if the only case this occurs is when sending an initial
1502 * SYN with options and data to a host we've never talked
1503 * to before. Then, they will reply with an MSS value which
1504 * will get recorded and the new parameters should get
1505 * recomputed. For Further Study.
1507 if (tp->t_maxopd <= mss)
1511 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1512 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1513 mss -= TCPOLEN_TSTAMP_APPA;
1514 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1516 mss &= ~(MCLBYTES-1);
1519 mss = mss / MCLBYTES * MCLBYTES;
1521 if (so->so_snd.sb_hiwat < mss)
1522 mss = so->so_snd.sb_hiwat;
1526 tcpstat.tcps_mturesent++;
1528 tp->snd_nxt = tp->snd_una;
1535 * Look-up the routing entry to the peer of this inpcb. If no route
1536 * is found and it cannot be allocated, then return NULL. This routine
1537 * is called by TCP routines that access the rmx structure and by tcp_mss
1538 * to get the interface MTU.
1542 struct in_conninfo *inc;
1545 struct sockaddr_in *dst;
1549 KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer"));
1551 bzero(&sro, sizeof(sro));
1552 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1553 dst = (struct sockaddr_in *)&sro.ro_dst;
1554 dst->sin_family = AF_INET;
1555 dst->sin_len = sizeof(*dst);
1556 dst->sin_addr = inc->inc_faddr;
1557 rtalloc_ign(&sro, RTF_CLONING);
1559 if (sro.ro_rt != NULL) {
1560 ifp = sro.ro_rt->rt_ifp;
1561 if (sro.ro_rt->rt_rmx.rmx_mtu == 0)
1562 maxmtu = ifp->if_mtu;
1564 maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu);
1573 struct in_conninfo *inc;
1575 struct route_in6 sro6;
1579 KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer"));
1581 bzero(&sro6, sizeof(sro6));
1582 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1583 sro6.ro_dst.sin6_family = AF_INET6;
1584 sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1585 sro6.ro_dst.sin6_addr = inc->inc6_faddr;
1586 rtalloc_ign((struct route *)&sro6, RTF_CLONING);
1588 if (sro6.ro_rt != NULL) {
1589 ifp = sro6.ro_rt->rt_ifp;
1590 if (sro6.ro_rt->rt_rmx.rmx_mtu == 0)
1591 maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp);
1593 maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu,
1594 IN6_LINKMTU(sro6.ro_rt->rt_ifp));
1603 /* compute ESP/AH header size for TCP, including outer IP header. */
1605 ipsec_hdrsiz_tcp(tp)
1613 struct ip6_hdr *ip6;
1617 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1619 MGETHDR(m, M_DONTWAIT, MT_DATA);
1624 if ((inp->inp_vflag & INP_IPV6) != 0) {
1625 ip6 = mtod(m, struct ip6_hdr *);
1626 th = (struct tcphdr *)(ip6 + 1);
1627 m->m_pkthdr.len = m->m_len =
1628 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1629 tcpip_fillheaders(inp, ip6, th);
1630 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1634 ip = mtod(m, struct ip *);
1635 th = (struct tcphdr *)(ip + 1);
1636 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1637 tcpip_fillheaders(inp, ip, th);
1638 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1647 * Move a TCP connection into TIME_WAIT state.
1648 * tcbinfo is locked.
1649 * inp is locked, and is unlocked before returning.
1657 int tw_time, acknow;
1660 INP_INFO_WLOCK_ASSERT(&tcbinfo); /* tcp_timer_2msl_reset(). */
1661 INP_LOCK_ASSERT(tp->t_inpcb);
1663 tw = uma_zalloc(tcptw_zone, M_NOWAIT);
1665 tw = tcp_timer_2msl_tw(1);
1675 * Recover last window size sent.
1677 tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
1680 * Set t_recent if timestamps are used on the connection.
1682 if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
1683 (TF_REQ_TSTMP|TF_RCVD_TSTMP))
1684 tw->t_recent = tp->ts_recent;
1688 tw->snd_nxt = tp->snd_nxt;
1689 tw->rcv_nxt = tp->rcv_nxt;
1692 tw->t_starttime = tp->t_starttime;
1697 * be used for fin-wait-2 state also, then we may need
1698 * a ts_recent from the last segment.
1700 tw_time = 2 * tcp_msl;
1701 acknow = tp->t_flags & TF_ACKNOW;
1703 so = inp->inp_socket;
1707 tw->tw_cred = crhold(so->so_cred);
1708 tw->tw_so_options = so->so_options;
1710 inp->inp_socket = NULL;
1712 tcp_twrespond(tw, TH_ACK);
1713 inp->inp_ppcb = (caddr_t)tw;
1714 inp->inp_vflag |= INP_TIMEWAIT;
1715 tcp_timer_2msl_reset(tw, tw_time);
1720 * The appromixate rate of ISN increase of Microsoft TCP stacks;
1721 * the actual rate is slightly higher due to the addition of
1722 * random positive increments.
1724 * Most other new OSes use semi-randomized ISN values, so we
1725 * do not need to worry about them.
1727 #define MS_ISN_BYTES_PER_SECOND 250000
1730 * Determine if the ISN we will generate has advanced beyond the last
1731 * sequence number used by the previous connection. If so, indicate
1732 * that it is safe to recycle this tw socket by returning 1.
1734 * XXXRW: This function should assert the inpcb lock as it does multiple
1735 * non-atomic reads from the tcptw, but is currently called without it from
1736 * in_pcb.c:in_pcblookup_local().
1739 tcp_twrecycleable(struct tcptw *tw)
1741 tcp_seq new_iss = tw->iss;
1742 tcp_seq new_irs = tw->irs;
1744 new_iss += (ticks - tw->t_starttime) * (ISN_BYTES_PER_SECOND / hz);
1745 new_irs += (ticks - tw->t_starttime) * (MS_ISN_BYTES_PER_SECOND / hz);
1747 if (SEQ_GT(new_iss, tw->snd_nxt) && SEQ_GT(new_irs, tw->rcv_nxt))
1754 tcp_twclose(struct tcptw *tw, int reuse)
1759 INP_INFO_WLOCK_ASSERT(&tcbinfo); /* tcp_timer_2msl_stop(). */
1760 INP_LOCK_ASSERT(inp);
1762 tw->tw_inpcb = NULL;
1763 tcp_timer_2msl_stop(tw);
1764 inp->inp_ppcb = NULL;
1766 if (inp->inp_vflag & INP_IPV6PROTO)
1771 tcpstat.tcps_closed++;
1772 crfree(tw->tw_cred);
1776 uma_zfree(tcptw_zone, tw);
1781 tcp_twrespond(struct tcptw *tw, int flags)
1783 struct inpcb *inp = tw->tw_inpcb;
1786 struct ip *ip = NULL;
1788 u_int hdrlen, optlen;
1791 struct ip6_hdr *ip6 = NULL;
1792 int isipv6 = inp->inp_inc.inc_isipv6;
1795 INP_LOCK_ASSERT(inp);
1797 m = m_gethdr(M_DONTWAIT, MT_DATA);
1800 m->m_data += max_linkhdr;
1803 mac_create_mbuf_from_inpcb(inp, m);
1808 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1809 ip6 = mtod(m, struct ip6_hdr *);
1810 th = (struct tcphdr *)(ip6 + 1);
1811 tcpip_fillheaders(inp, ip6, th);
1815 hdrlen = sizeof(struct tcpiphdr);
1816 ip = mtod(m, struct ip *);
1817 th = (struct tcphdr *)(ip + 1);
1818 tcpip_fillheaders(inp, ip, th);
1820 optp = (u_int8_t *)(th + 1);
1823 * Send a timestamp and echo-reply if both our side and our peer
1824 * have sent timestamps in our SYN's and this is not a RST.
1826 if (tw->t_recent && flags == TH_ACK) {
1827 u_int32_t *lp = (u_int32_t *)optp;
1829 /* Form timestamp option as shown in appendix A of RFC 1323. */
1830 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1831 *lp++ = htonl(ticks);
1832 *lp = htonl(tw->t_recent);
1833 optp += TCPOLEN_TSTAMP_APPA;
1836 optlen = optp - (u_int8_t *)(th + 1);
1838 m->m_len = hdrlen + optlen;
1839 m->m_pkthdr.len = m->m_len;
1841 KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
1843 th->th_seq = htonl(tw->snd_nxt);
1844 th->th_ack = htonl(tw->rcv_nxt);
1845 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1846 th->th_flags = flags;
1847 th->th_win = htons(tw->last_win);
1851 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1852 sizeof(struct tcphdr) + optlen);
1853 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1854 error = ip6_output(m, inp->in6p_outputopts, NULL,
1855 (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
1859 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1860 htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
1861 m->m_pkthdr.csum_flags = CSUM_TCP;
1862 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1863 ip->ip_len = m->m_pkthdr.len;
1864 if (path_mtu_discovery)
1865 ip->ip_off |= IP_DF;
1866 error = ip_output(m, inp->inp_options, NULL,
1867 ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1871 tcpstat.tcps_sndacks++;
1873 tcpstat.tcps_sndctrl++;
1874 tcpstat.tcps_sndtotal++;
1879 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1881 * This code attempts to calculate the bandwidth-delay product as a
1882 * means of determining the optimal window size to maximize bandwidth,
1883 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1884 * routers. This code also does a fairly good job keeping RTTs in check
1885 * across slow links like modems. We implement an algorithm which is very
1886 * similar (but not meant to be) TCP/Vegas. The code operates on the
1887 * transmitter side of a TCP connection and so only effects the transmit
1888 * side of the connection.
1890 * BACKGROUND: TCP makes no provision for the management of buffer space
1891 * at the end points or at the intermediate routers and switches. A TCP
1892 * stream, whether using NewReno or not, will eventually buffer as
1893 * many packets as it is able and the only reason this typically works is
1894 * due to the fairly small default buffers made available for a connection
1895 * (typicaly 16K or 32K). As machines use larger windows and/or window
1896 * scaling it is now fairly easy for even a single TCP connection to blow-out
1897 * all available buffer space not only on the local interface, but on
1898 * intermediate routers and switches as well. NewReno makes a misguided
1899 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1900 * then backing off, then steadily increasing the window again until another
1901 * failure occurs, ad-infinitum. This results in terrible oscillation that
1902 * is only made worse as network loads increase and the idea of intentionally
1903 * blowing out network buffers is, frankly, a terrible way to manage network
1906 * It is far better to limit the transmit window prior to the failure
1907 * condition being achieved. There are two general ways to do this: First
1908 * you can 'scan' through different transmit window sizes and locate the
1909 * point where the RTT stops increasing, indicating that you have filled the
1910 * pipe, then scan backwards until you note that RTT stops decreasing, then
1911 * repeat ad-infinitum. This method works in principle but has severe
1912 * implementation issues due to RTT variances, timer granularity, and
1913 * instability in the algorithm which can lead to many false positives and
1914 * create oscillations as well as interact badly with other TCP streams
1915 * implementing the same algorithm.
1917 * The second method is to limit the window to the bandwidth delay product
1918 * of the link. This is the method we implement. RTT variances and our
1919 * own manipulation of the congestion window, bwnd, can potentially
1920 * destabilize the algorithm. For this reason we have to stabilize the
1921 * elements used to calculate the window. We do this by using the minimum
1922 * observed RTT, the long term average of the observed bandwidth, and
1923 * by adding two segments worth of slop. It isn't perfect but it is able
1924 * to react to changing conditions and gives us a very stable basis on
1925 * which to extend the algorithm.
1928 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1934 INP_LOCK_ASSERT(tp->t_inpcb);
1937 * If inflight_enable is disabled in the middle of a tcp connection,
1938 * make sure snd_bwnd is effectively disabled.
1940 if (tcp_inflight_enable == 0) {
1941 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1942 tp->snd_bandwidth = 0;
1947 * Figure out the bandwidth. Due to the tick granularity this
1948 * is a very rough number and it MUST be averaged over a fairly
1949 * long period of time. XXX we need to take into account a link
1950 * that is not using all available bandwidth, but for now our
1951 * slop will ramp us up if this case occurs and the bandwidth later
1954 * Note: if ticks rollover 'bw' may wind up negative. We must
1955 * effectively reset t_bw_rtttime for this case.
1958 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1961 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1962 (save_ticks - tp->t_bw_rtttime);
1963 tp->t_bw_rtttime = save_ticks;
1964 tp->t_bw_rtseq = ack_seq;
1965 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1967 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1969 tp->snd_bandwidth = bw;
1972 * Calculate the semi-static bandwidth delay product, plus two maximal
1973 * segments. The additional slop puts us squarely in the sweet
1974 * spot and also handles the bandwidth run-up case and stabilization.
1975 * Without the slop we could be locking ourselves into a lower
1978 * Situations Handled:
1979 * (1) Prevents over-queueing of packets on LANs, especially on
1980 * high speed LANs, allowing larger TCP buffers to be
1981 * specified, and also does a good job preventing
1982 * over-queueing of packets over choke points like modems
1983 * (at least for the transmit side).
1985 * (2) Is able to handle changing network loads (bandwidth
1986 * drops so bwnd drops, bandwidth increases so bwnd
1989 * (3) Theoretically should stabilize in the face of multiple
1990 * connections implementing the same algorithm (this may need
1993 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1994 * be adjusted with a sysctl but typically only needs to be
1995 * on very slow connections. A value no smaller then 5
1996 * should be used, but only reduce this default if you have
1999 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
2000 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10;
2003 if (tcp_inflight_debug > 0) {
2005 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2007 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2016 if ((long)bwnd < tcp_inflight_min)
2017 bwnd = tcp_inflight_min;
2018 if (bwnd > tcp_inflight_max)
2019 bwnd = tcp_inflight_max;
2020 if ((long)bwnd < tp->t_maxseg * 2)
2021 bwnd = tp->t_maxseg * 2;
2022 tp->snd_bwnd = bwnd;
2025 #ifdef TCP_SIGNATURE
2027 * Callback function invoked by m_apply() to digest TCP segment data
2028 * contained within an mbuf chain.
2031 tcp_signature_apply(void *fstate, void *data, u_int len)
2034 MD5Update(fstate, (u_char *)data, len);
2039 * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385)
2042 * m pointer to head of mbuf chain
2043 * off0 offset to TCP header within the mbuf chain
2044 * len length of TCP segment data, excluding options
2045 * optlen length of TCP segment options
2046 * buf pointer to storage for computed MD5 digest
2047 * direction direction of flow (IPSEC_DIR_INBOUND or OUTBOUND)
2049 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2050 * When called from tcp_input(), we can be sure that th_sum has been
2051 * zeroed out and verified already.
2053 * This function is for IPv4 use only. Calling this function with an
2054 * IPv6 packet in the mbuf chain will yield undefined results.
2056 * Return 0 if successful, otherwise return -1.
2058 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2059 * search with the destination IP address, and a 'magic SPI' to be
2060 * determined by the application. This is hardcoded elsewhere to 1179
2061 * right now. Another branch of this code exists which uses the SPD to
2062 * specify per-application flows but it is unstable.
2065 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen,
2066 u_char *buf, u_int direction)
2068 union sockaddr_union dst;
2069 struct ippseudo ippseudo;
2073 struct ipovly *ipovly;
2074 struct secasvar *sav;
2078 KASSERT(m != NULL, ("NULL mbuf chain"));
2079 KASSERT(buf != NULL, ("NULL signature pointer"));
2081 /* Extract the destination from the IP header in the mbuf. */
2082 ip = mtod(m, struct ip *);
2083 bzero(&dst, sizeof(union sockaddr_union));
2084 dst.sa.sa_len = sizeof(struct sockaddr_in);
2085 dst.sa.sa_family = AF_INET;
2086 dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ?
2087 ip->ip_src : ip->ip_dst;
2089 /* Look up an SADB entry which matches the address of the peer. */
2090 sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI));
2092 printf("%s: SADB lookup failed for %s\n", __func__,
2093 inet_ntoa(dst.sin.sin_addr));
2098 ipovly = (struct ipovly *)ip;
2099 th = (struct tcphdr *)((u_char *)ip + off0);
2100 doff = off0 + sizeof(struct tcphdr) + optlen;
2103 * Step 1: Update MD5 hash with IP pseudo-header.
2105 * XXX The ippseudo header MUST be digested in network byte order,
2106 * or else we'll fail the regression test. Assume all fields we've
2107 * been doing arithmetic on have been in host byte order.
2108 * XXX One cannot depend on ipovly->ih_len here. When called from
2109 * tcp_output(), the underlying ip_len member has not yet been set.
2111 ippseudo.ippseudo_src = ipovly->ih_src;
2112 ippseudo.ippseudo_dst = ipovly->ih_dst;
2113 ippseudo.ippseudo_pad = 0;
2114 ippseudo.ippseudo_p = IPPROTO_TCP;
2115 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2116 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2119 * Step 2: Update MD5 hash with TCP header, excluding options.
2120 * The TCP checksum must be set to zero.
2122 savecsum = th->th_sum;
2124 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2125 th->th_sum = savecsum;
2128 * Step 3: Update MD5 hash with TCP segment data.
2129 * Use m_apply() to avoid an early m_pullup().
2132 m_apply(m, doff, len, tcp_signature_apply, &ctx);
2135 * Step 4: Update MD5 hash with shared secret.
2137 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2138 MD5Final(buf, &ctx);
2140 key_sa_recordxfer(sav, m);
2144 #endif /* TCP_SIGNATURE */
2147 sysctl_drop(SYSCTL_HANDLER_ARGS)
2149 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2150 struct sockaddr_storage addrs[2];
2154 struct sockaddr_in *fin, *lin;
2156 struct sockaddr_in6 *fin6, *lin6;
2157 struct in6_addr f6, l6;
2168 if (req->oldptr != NULL || req->oldlen != 0)
2170 if (req->newptr == NULL)
2172 if (req->newlen < sizeof(addrs))
2174 error = SYSCTL_IN(req, &addrs, sizeof(addrs));
2178 switch (addrs[0].ss_family) {
2181 fin6 = (struct sockaddr_in6 *)&addrs[0];
2182 lin6 = (struct sockaddr_in6 *)&addrs[1];
2183 if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
2184 lin6->sin6_len != sizeof(struct sockaddr_in6))
2186 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) {
2187 if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2189 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]);
2190 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]);
2191 fin = (struct sockaddr_in *)&addrs[0];
2192 lin = (struct sockaddr_in *)&addrs[1];
2195 error = sa6_embedscope(fin6, ip6_use_defzone);
2198 error = sa6_embedscope(lin6, ip6_use_defzone);
2204 fin = (struct sockaddr_in *)&addrs[0];
2205 lin = (struct sockaddr_in *)&addrs[1];
2206 if (fin->sin_len != sizeof(struct sockaddr_in) ||
2207 lin->sin_len != sizeof(struct sockaddr_in))
2213 INP_INFO_WLOCK(&tcbinfo);
2214 switch (addrs[0].ss_family) {
2217 inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port,
2218 &l6, lin6->sin6_port, 0, NULL);
2222 inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port,
2223 lin->sin_addr, lin->sin_port, 0, NULL);
2228 if ((tw = intotw(inp)) &&
2229 (inp->inp_vflag & INP_TIMEWAIT) != 0) {
2230 (void) tcp_twclose(tw, 0);
2231 } else if ((tp = intotcpcb(inp)) &&
2232 ((inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) {
2233 tp = tcp_drop(tp, ECONNABORTED);
2240 INP_INFO_WUNLOCK(&tcbinfo);
2244 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop,
2245 CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL,
2246 0, sysctl_drop, "", "Drop TCP connection");