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_rttthresh;
197 SYSCTL_PROC(_net_inet_tcp_inflight, OID_AUTO, rttthresh, CTLTYPE_INT|CTLFLAG_RW,
198 &tcp_inflight_rttthresh, 0, sysctl_msec_to_ticks, "I",
199 "RTT threshold below which inflight will deactivate itself");
201 static int tcp_inflight_min = 6144;
202 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW,
203 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
205 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
206 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW,
207 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
209 static int tcp_inflight_stab = 20;
210 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW,
211 &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets");
213 uma_zone_t sack_hole_zone;
215 static struct inpcb *tcp_notify(struct inpcb *, int);
216 static void tcp_discardcb(struct tcpcb *);
217 static void tcp_isn_tick(void *);
220 * Target size of TCP PCB hash tables. Must be a power of two.
222 * Note that this can be overridden by the kernel environment
223 * variable net.inet.tcp.tcbhashsize
226 #define TCBHASHSIZE 512
231 * Callouts should be moved into struct tcp directly. They are currently
232 * separate because the tcpcb structure is exported to userland for sysctl
233 * parsing purposes, which do not know about callouts.
237 struct callout tcpcb_mem_rexmt, tcpcb_mem_persist, tcpcb_mem_keep;
238 struct callout tcpcb_mem_2msl, tcpcb_mem_delack;
241 static uma_zone_t tcpcb_zone;
242 static uma_zone_t tcptw_zone;
243 struct callout isn_callout;
251 int hashsize = TCBHASHSIZE;
253 tcp_delacktime = TCPTV_DELACK;
254 tcp_keepinit = TCPTV_KEEP_INIT;
255 tcp_keepidle = TCPTV_KEEP_IDLE;
256 tcp_keepintvl = TCPTV_KEEPINTVL;
257 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
259 tcp_rexmit_min = TCPTV_MIN;
260 tcp_rexmit_slop = TCPTV_CPU_VAR;
261 tcp_inflight_rttthresh = TCPTV_INFLIGHT_RTTTHRESH;
263 INP_INFO_LOCK_INIT(&tcbinfo, "tcp");
265 tcbinfo.listhead = &tcb;
266 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
267 if (!powerof2(hashsize)) {
268 printf("WARNING: TCB hash size not a power of 2\n");
269 hashsize = 512; /* safe default */
271 tcp_tcbhashsize = hashsize;
272 tcbinfo.hashbase = hashinit(hashsize, M_PCB, &tcbinfo.hashmask);
273 tcbinfo.porthashbase = hashinit(hashsize, M_PCB,
274 &tcbinfo.porthashmask);
275 tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb),
276 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
277 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
279 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
281 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
283 if (max_protohdr < TCP_MINPROTOHDR)
284 max_protohdr = TCP_MINPROTOHDR;
285 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
287 #undef TCP_MINPROTOHDR
289 * These have to be type stable for the benefit of the timers.
291 tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem),
292 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
293 uma_zone_set_max(tcpcb_zone, maxsockets);
294 tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
295 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
296 uma_zone_set_max(tcptw_zone, maxsockets / 5);
301 callout_init(&isn_callout, CALLOUT_MPSAFE);
303 EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL,
304 SHUTDOWN_PRI_DEFAULT);
305 sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole),
306 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 callout_stop(&isn_callout);
318 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
319 * tcp_template used to store this data in mbufs, but we now recopy it out
320 * of the tcpcb each time to conserve mbufs.
323 tcpip_fillheaders(inp, ip_ptr, tcp_ptr)
328 struct tcphdr *th = (struct tcphdr *)tcp_ptr;
330 INP_LOCK_ASSERT(inp);
333 if ((inp->inp_vflag & INP_IPV6) != 0) {
336 ip6 = (struct ip6_hdr *)ip_ptr;
337 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
338 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
339 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
340 (IPV6_VERSION & IPV6_VERSION_MASK);
341 ip6->ip6_nxt = IPPROTO_TCP;
342 ip6->ip6_plen = sizeof(struct tcphdr);
343 ip6->ip6_src = inp->in6p_laddr;
344 ip6->ip6_dst = inp->in6p_faddr;
350 ip = (struct ip *)ip_ptr;
351 ip->ip_v = IPVERSION;
353 ip->ip_tos = inp->inp_ip_tos;
357 ip->ip_ttl = inp->inp_ip_ttl;
359 ip->ip_p = IPPROTO_TCP;
360 ip->ip_src = inp->inp_laddr;
361 ip->ip_dst = inp->inp_faddr;
363 th->th_sport = inp->inp_lport;
364 th->th_dport = inp->inp_fport;
372 th->th_sum = 0; /* in_pseudo() is called later for ipv4 */
376 * Create template to be used to send tcp packets on a connection.
377 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
378 * use for this function is in keepalives, which use tcp_respond.
381 tcpip_maketemplate(inp)
387 m = m_get(M_DONTWAIT, MT_DATA);
390 m->m_len = sizeof(struct tcptemp);
391 n = mtod(m, struct tcptemp *);
393 tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
398 * Send a single message to the TCP at address specified by
399 * the given TCP/IP header. If m == NULL, then we make a copy
400 * of the tcpiphdr at ti and send directly to the addressed host.
401 * This is used to force keep alive messages out using the TCP
402 * template for a connection. If flags are given then we send
403 * a message back to the TCP which originated the * segment ti,
404 * and discard the mbuf containing it and any other attached mbufs.
406 * In any case the ack and sequence number of the transmitted
407 * segment are as specified by the parameters.
409 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
412 tcp_respond(tp, ipgen, th, m, ack, seq, flags)
415 register struct tcphdr *th;
416 register struct mbuf *m;
431 KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL"));
434 isipv6 = ((struct ip *)ipgen)->ip_v == 6;
441 KASSERT(inp != NULL, ("tcp control block w/o inpcb"));
442 INP_INFO_WLOCK_ASSERT(&tcbinfo);
443 INP_LOCK_ASSERT(inp);
448 if (!(flags & TH_RST)) {
449 win = sbspace(&inp->inp_socket->so_rcv);
450 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
451 win = (long)TCP_MAXWIN << tp->rcv_scale;
455 m = m_gethdr(M_DONTWAIT, MT_DATA);
459 m->m_data += max_linkhdr;
462 bcopy((caddr_t)ip6, mtod(m, caddr_t),
463 sizeof(struct ip6_hdr));
464 ip6 = mtod(m, struct ip6_hdr *);
465 nth = (struct tcphdr *)(ip6 + 1);
469 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
470 ip = mtod(m, struct ip *);
471 nth = (struct tcphdr *)(ip + 1);
473 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
478 m->m_data = (caddr_t)ipgen;
479 /* m_len is set later */
481 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
484 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
485 nth = (struct tcphdr *)(ip6 + 1);
489 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
490 nth = (struct tcphdr *)(ip + 1);
494 * this is usually a case when an extension header
495 * exists between the IPv6 header and the
498 nth->th_sport = th->th_sport;
499 nth->th_dport = th->th_dport;
501 xchg(nth->th_dport, nth->th_sport, n_short);
507 ip6->ip6_vfc = IPV6_VERSION;
508 ip6->ip6_nxt = IPPROTO_TCP;
509 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
511 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
515 tlen += sizeof (struct tcpiphdr);
517 ip->ip_ttl = ip_defttl;
518 if (path_mtu_discovery)
522 m->m_pkthdr.len = tlen;
523 m->m_pkthdr.rcvif = NULL;
527 * Packet is associated with a socket, so allow the
528 * label of the response to reflect the socket label.
530 INP_LOCK_ASSERT(inp);
531 mac_create_mbuf_from_inpcb(inp, m);
534 * Packet is not associated with a socket, so possibly
535 * update the label in place.
537 mac_reflect_mbuf_tcp(m);
540 nth->th_seq = htonl(seq);
541 nth->th_ack = htonl(ack);
543 nth->th_off = sizeof (struct tcphdr) >> 2;
544 nth->th_flags = flags;
546 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
548 nth->th_win = htons((u_short)win);
553 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
554 sizeof(struct ip6_hdr),
555 tlen - sizeof(struct ip6_hdr));
556 ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb :
561 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
562 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
563 m->m_pkthdr.csum_flags = CSUM_TCP;
564 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
567 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
568 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
572 (void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp);
575 (void) ip_output(m, NULL, NULL, ipflags, NULL, inp);
579 * Create a new TCP control block, making an
580 * empty reassembly queue and hooking it to the argument
581 * protocol control block. The `inp' parameter must have
582 * come from the zone allocator set up in tcp_init().
588 struct tcpcb_mem *tm;
591 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
594 tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO);
598 /* LIST_INIT(&tp->t_segq); */ /* XXX covered by M_ZERO */
599 tp->t_maxseg = tp->t_maxopd =
601 isipv6 ? tcp_v6mssdflt :
605 /* Set up our timeouts. */
606 callout_init(tp->tt_rexmt = &tm->tcpcb_mem_rexmt, NET_CALLOUT_MPSAFE);
607 callout_init(tp->tt_persist = &tm->tcpcb_mem_persist, NET_CALLOUT_MPSAFE);
608 callout_init(tp->tt_keep = &tm->tcpcb_mem_keep, NET_CALLOUT_MPSAFE);
609 callout_init(tp->tt_2msl = &tm->tcpcb_mem_2msl, NET_CALLOUT_MPSAFE);
610 callout_init(tp->tt_delack = &tm->tcpcb_mem_delack, NET_CALLOUT_MPSAFE);
613 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
614 tp->sack_enable = tcp_do_sack;
615 TAILQ_INIT(&tp->snd_holes);
616 tp->t_inpcb = inp; /* XXX */
618 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
619 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
620 * reasonable initial retransmit time.
622 tp->t_srtt = TCPTV_SRTTBASE;
623 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
624 tp->t_rttmin = tcp_rexmit_min;
625 tp->t_rxtcur = TCPTV_RTOBASE;
626 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
627 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
628 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
629 tp->t_rcvtime = ticks;
630 tp->t_bw_rtttime = ticks;
632 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
633 * because the socket may be bound to an IPv6 wildcard address,
634 * which may match an IPv4-mapped IPv6 address.
636 inp->inp_ip_ttl = ip_defttl;
637 inp->inp_ppcb = (caddr_t)tp;
638 return (tp); /* XXX */
642 * Drop a TCP connection, reporting
643 * the specified error. If connection is synchronized,
644 * then send a RST to peer.
648 register struct tcpcb *tp;
651 struct socket *so = tp->t_inpcb->inp_socket;
653 INP_INFO_WLOCK_ASSERT(&tcbinfo);
654 INP_LOCK_ASSERT(tp->t_inpcb);
656 if (TCPS_HAVERCVDSYN(tp->t_state)) {
657 tp->t_state = TCPS_CLOSED;
658 (void) tcp_output(tp);
659 tcpstat.tcps_drops++;
661 tcpstat.tcps_conndrops++;
662 if (errno == ETIMEDOUT && tp->t_softerror)
663 errno = tp->t_softerror;
664 so->so_error = errno;
665 return (tcp_close(tp));
673 struct inpcb *inp = tp->t_inpcb;
674 struct socket *so = inp->inp_socket;
676 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
679 INP_LOCK_ASSERT(inp);
682 * Make sure that all of our timers are stopped before we
685 callout_stop(tp->tt_rexmt);
686 callout_stop(tp->tt_persist);
687 callout_stop(tp->tt_keep);
688 callout_stop(tp->tt_2msl);
689 callout_stop(tp->tt_delack);
692 * If we got enough samples through the srtt filter,
693 * save the rtt and rttvar in the routing entry.
694 * 'Enough' is arbitrarily defined as 4 rtt samples.
695 * 4 samples is enough for the srtt filter to converge
696 * to within enough % of the correct value; fewer samples
697 * and we could save a bogus rtt. The danger is not high
698 * as tcp quickly recovers from everything.
699 * XXX: Works very well but needs some more statistics!
701 if (tp->t_rttupdated >= 4) {
702 struct hc_metrics_lite metrics;
705 bzero(&metrics, sizeof(metrics));
707 * Update the ssthresh always when the conditions below
708 * are satisfied. This gives us better new start value
709 * for the congestion avoidance for new connections.
710 * ssthresh is only set if packet loss occured on a session.
712 ssthresh = tp->snd_ssthresh;
713 if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) {
715 * convert the limit from user data bytes to
716 * packets then to packet data bytes.
718 ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg;
721 ssthresh *= (u_long)(tp->t_maxseg +
723 (isipv6 ? sizeof (struct ip6_hdr) +
724 sizeof (struct tcphdr) :
726 sizeof (struct tcpiphdr)
733 metrics.rmx_ssthresh = ssthresh;
735 metrics.rmx_rtt = tp->t_srtt;
736 metrics.rmx_rttvar = tp->t_rttvar;
737 /* XXX: This wraps if the pipe is more than 4 Gbit per second */
738 metrics.rmx_bandwidth = tp->snd_bandwidth;
739 metrics.rmx_cwnd = tp->snd_cwnd;
740 metrics.rmx_sendpipe = 0;
741 metrics.rmx_recvpipe = 0;
743 tcp_hc_update(&inp->inp_inc, &metrics);
746 /* free the reassembly queue, if any */
747 while ((q = LIST_FIRST(&tp->t_segq)) != NULL) {
748 LIST_REMOVE(q, tqe_q);
750 uma_zfree(tcp_reass_zone, q);
754 tcp_free_sackholes(tp);
755 inp->inp_ppcb = NULL;
757 uma_zfree(tcpcb_zone, tp);
758 soisdisconnected(so);
762 * Close a TCP control block:
763 * discard all space held by the tcp
764 * discard internet protocol block
765 * wake up any sleepers
771 struct inpcb *inp = tp->t_inpcb;
773 struct socket *so = inp->inp_socket;
776 INP_INFO_WLOCK_ASSERT(&tcbinfo);
777 INP_LOCK_ASSERT(inp);
781 if (INP_CHECK_SOCKAF(so, AF_INET6))
786 tcpstat.tcps_closed++;
797 struct tseg_qent *te;
800 * Walk the tcpbs, if existing, and flush the reassembly queue,
802 * XXX: The "Net/3" implementation doesn't imply that the TCP
803 * reassembly queue should be flushed, but in a situation
804 * where we're really low on mbufs, this is potentially
807 INP_INFO_RLOCK(&tcbinfo);
808 LIST_FOREACH(inpb, tcbinfo.listhead, inp_list) {
809 if (inpb->inp_vflag & INP_TIMEWAIT)
812 if ((tcpb = intotcpcb(inpb)) != NULL) {
813 while ((te = LIST_FIRST(&tcpb->t_segq))
815 LIST_REMOVE(te, tqe_q);
817 uma_zfree(tcp_reass_zone, te);
821 tcp_clean_sackreport(tcpb);
825 INP_INFO_RUNLOCK(&tcbinfo);
830 * Notify a tcp user of an asynchronous error;
831 * store error as soft error, but wake up user
832 * (for now, won't do anything until can select for soft error).
834 * Do not wake up user since there currently is no mechanism for
835 * reporting soft errors (yet - a kqueue filter may be added).
837 static struct inpcb *
838 tcp_notify(inp, error)
842 struct tcpcb *tp = (struct tcpcb *)inp->inp_ppcb;
844 INP_INFO_WLOCK_ASSERT(&tcbinfo);
845 INP_LOCK_ASSERT(inp);
848 * Ignore some errors if we are hooked up.
849 * If connection hasn't completed, has retransmitted several times,
850 * and receives a second error, give up now. This is better
851 * than waiting a long time to establish a connection that
852 * can never complete.
854 if (tp->t_state == TCPS_ESTABLISHED &&
855 (error == EHOSTUNREACH || error == ENETUNREACH ||
856 error == EHOSTDOWN)) {
858 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
861 return (struct inpcb *)0;
863 tp->t_softerror = error;
867 wakeup( &so->so_timeo);
874 tcp_pcblist(SYSCTL_HANDLER_ARGS)
877 struct inpcb *inp, **inp_list;
882 * The process of preparing the TCB list is too time-consuming and
883 * resource-intensive to repeat twice on every request.
885 if (req->oldptr == NULL) {
886 n = tcbinfo.ipi_count;
887 req->oldidx = 2 * (sizeof xig)
888 + (n + n/8) * sizeof(struct xtcpcb);
892 if (req->newptr != NULL)
896 * OK, now we're committed to doing something.
898 INP_INFO_RLOCK(&tcbinfo);
899 gencnt = tcbinfo.ipi_gencnt;
900 n = tcbinfo.ipi_count;
901 INP_INFO_RUNLOCK(&tcbinfo);
903 error = sysctl_wire_old_buffer(req, 2 * (sizeof xig)
904 + n * sizeof(struct xtcpcb));
908 xig.xig_len = sizeof xig;
910 xig.xig_gen = gencnt;
911 xig.xig_sogen = so_gencnt;
912 error = SYSCTL_OUT(req, &xig, sizeof xig);
916 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
917 if (inp_list == NULL)
920 INP_INFO_RLOCK(&tcbinfo);
921 for (inp = LIST_FIRST(tcbinfo.listhead), i = 0; inp != NULL && i < n;
922 inp = LIST_NEXT(inp, inp_list)) {
924 if (inp->inp_gencnt <= gencnt) {
926 * XXX: This use of cr_cansee(), introduced with
927 * TCP state changes, is not quite right, but for
928 * now, better than nothing.
930 if (inp->inp_vflag & INP_TIMEWAIT)
931 error = cr_cansee(req->td->td_ucred,
932 intotw(inp)->tw_cred);
934 error = cr_canseesocket(req->td->td_ucred,
941 INP_INFO_RUNLOCK(&tcbinfo);
945 for (i = 0; i < n; i++) {
947 if (inp->inp_gencnt <= gencnt) {
951 bzero(&xt, sizeof(xt));
952 xt.xt_len = sizeof xt;
953 /* XXX should avoid extra copy */
954 bcopy(inp, &xt.xt_inp, sizeof *inp);
955 inp_ppcb = inp->inp_ppcb;
956 if (inp_ppcb == NULL)
957 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
958 else if (inp->inp_vflag & INP_TIMEWAIT) {
959 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
960 xt.xt_tp.t_state = TCPS_TIME_WAIT;
962 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
963 if (inp->inp_socket != NULL)
964 sotoxsocket(inp->inp_socket, &xt.xt_socket);
966 bzero(&xt.xt_socket, sizeof xt.xt_socket);
967 xt.xt_socket.xso_protocol = IPPROTO_TCP;
969 xt.xt_inp.inp_gencnt = inp->inp_gencnt;
970 error = SYSCTL_OUT(req, &xt, sizeof xt);
975 * Give the user an updated idea of our state.
976 * If the generation differs from what we told
977 * her before, she knows that something happened
978 * while we were processing this request, and it
979 * might be necessary to retry.
981 INP_INFO_RLOCK(&tcbinfo);
982 xig.xig_gen = tcbinfo.ipi_gencnt;
983 xig.xig_sogen = so_gencnt;
984 xig.xig_count = tcbinfo.ipi_count;
985 INP_INFO_RUNLOCK(&tcbinfo);
986 error = SYSCTL_OUT(req, &xig, sizeof xig);
988 free(inp_list, M_TEMP);
992 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
993 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
996 tcp_getcred(SYSCTL_HANDLER_ARGS)
999 struct sockaddr_in addrs[2];
1003 error = suser_cred(req->td->td_ucred, SUSER_ALLOWJAIL);
1006 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1009 INP_INFO_RLOCK(&tcbinfo);
1010 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
1011 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1017 if (inp->inp_socket == NULL) {
1021 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1024 cru2x(inp->inp_socket->so_cred, &xuc);
1028 INP_INFO_RUNLOCK(&tcbinfo);
1030 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1034 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred,
1035 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1036 tcp_getcred, "S,xucred", "Get the xucred of a TCP connection");
1040 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1043 struct sockaddr_in6 addrs[2];
1045 int error, mapped = 0;
1047 error = suser_cred(req->td->td_ucred, SUSER_ALLOWJAIL);
1050 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1053 if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 ||
1054 (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) {
1057 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1058 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1064 INP_INFO_RLOCK(&tcbinfo);
1066 inp = in_pcblookup_hash(&tcbinfo,
1067 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1069 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1073 inp = in6_pcblookup_hash(&tcbinfo,
1074 &addrs[1].sin6_addr, addrs[1].sin6_port,
1075 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1081 if (inp->inp_socket == NULL) {
1085 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1088 cru2x(inp->inp_socket->so_cred, &xuc);
1092 INP_INFO_RUNLOCK(&tcbinfo);
1094 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1098 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred,
1099 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1100 tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection");
1105 tcp_ctlinput(cmd, sa, vip)
1107 struct sockaddr *sa;
1110 struct ip *ip = vip;
1112 struct in_addr faddr;
1115 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1117 struct in_conninfo inc;
1118 tcp_seq icmp_tcp_seq;
1121 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1122 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1125 if (cmd == PRC_MSGSIZE)
1126 notify = tcp_mtudisc;
1127 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1128 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1129 notify = tcp_drop_syn_sent;
1131 * Redirects don't need to be handled up here.
1133 else if (PRC_IS_REDIRECT(cmd))
1136 * Source quench is depreciated.
1138 else if (cmd == PRC_QUENCH)
1141 * Hostdead is ugly because it goes linearly through all PCBs.
1142 * XXX: We never get this from ICMP, otherwise it makes an
1143 * excellent DoS attack on machines with many connections.
1145 else if (cmd == PRC_HOSTDEAD)
1147 else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0)
1150 icp = (struct icmp *)((caddr_t)ip
1151 - offsetof(struct icmp, icmp_ip));
1152 th = (struct tcphdr *)((caddr_t)ip
1153 + (ip->ip_hl << 2));
1154 INP_INFO_WLOCK(&tcbinfo);
1155 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1156 ip->ip_src, th->th_sport, 0, NULL);
1159 if (inp->inp_socket != NULL) {
1160 icmp_tcp_seq = htonl(th->th_seq);
1161 tp = intotcpcb(inp);
1162 if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) &&
1163 SEQ_LT(icmp_tcp_seq, tp->snd_max)) {
1164 if (cmd == PRC_MSGSIZE) {
1167 * If we got a needfrag set the MTU
1168 * in the route to the suggested new
1169 * value (if given) and then notify.
1171 bzero(&inc, sizeof(inc));
1172 inc.inc_flags = 0; /* IPv4 */
1173 inc.inc_faddr = faddr;
1175 mtu = ntohs(icp->icmp_nextmtu);
1177 * If no alternative MTU was
1178 * proposed, try the next smaller
1179 * one. ip->ip_len has already
1180 * been swapped in icmp_input().
1183 mtu = ip_next_mtu(ip->ip_len,
1185 if (mtu < max(296, (tcp_minmss)
1186 + sizeof(struct tcpiphdr)))
1190 + sizeof(struct tcpiphdr);
1192 * Only cache the the MTU if it
1193 * is smaller than the interface
1194 * or route MTU. tcp_mtudisc()
1195 * will do right thing by itself.
1197 if (mtu <= tcp_maxmtu(&inc))
1198 tcp_hc_updatemtu(&inc, mtu);
1201 inp = (*notify)(inp, inetctlerrmap[cmd]);
1207 inc.inc_fport = th->th_dport;
1208 inc.inc_lport = th->th_sport;
1209 inc.inc_faddr = faddr;
1210 inc.inc_laddr = ip->ip_src;
1214 syncache_unreach(&inc, th);
1216 INP_INFO_WUNLOCK(&tcbinfo);
1218 in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify);
1223 tcp6_ctlinput(cmd, sa, d)
1225 struct sockaddr *sa;
1229 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1230 struct ip6_hdr *ip6;
1232 struct ip6ctlparam *ip6cp = NULL;
1233 const struct sockaddr_in6 *sa6_src = NULL;
1235 struct tcp_portonly {
1240 if (sa->sa_family != AF_INET6 ||
1241 sa->sa_len != sizeof(struct sockaddr_in6))
1244 if (cmd == PRC_MSGSIZE)
1245 notify = tcp_mtudisc;
1246 else if (!PRC_IS_REDIRECT(cmd) &&
1247 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1249 /* Source quench is depreciated. */
1250 else if (cmd == PRC_QUENCH)
1253 /* if the parameter is from icmp6, decode it. */
1255 ip6cp = (struct ip6ctlparam *)d;
1257 ip6 = ip6cp->ip6c_ip6;
1258 off = ip6cp->ip6c_off;
1259 sa6_src = ip6cp->ip6c_src;
1263 off = 0; /* fool gcc */
1268 struct in_conninfo inc;
1270 * XXX: We assume that when IPV6 is non NULL,
1271 * M and OFF are valid.
1274 /* check if we can safely examine src and dst ports */
1275 if (m->m_pkthdr.len < off + sizeof(*thp))
1278 bzero(&th, sizeof(th));
1279 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1281 in6_pcbnotify(&tcbinfo, sa, th.th_dport,
1282 (struct sockaddr *)ip6cp->ip6c_src,
1283 th.th_sport, cmd, NULL, notify);
1285 inc.inc_fport = th.th_dport;
1286 inc.inc_lport = th.th_sport;
1287 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1288 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1290 INP_INFO_WLOCK(&tcbinfo);
1291 syncache_unreach(&inc, &th);
1292 INP_INFO_WUNLOCK(&tcbinfo);
1294 in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src,
1295 0, cmd, NULL, notify);
1301 * Following is where TCP initial sequence number generation occurs.
1303 * There are two places where we must use initial sequence numbers:
1304 * 1. In SYN-ACK packets.
1305 * 2. In SYN packets.
1307 * All ISNs for SYN-ACK packets are generated by the syncache. See
1308 * tcp_syncache.c for details.
1310 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1311 * depends on this property. In addition, these ISNs should be
1312 * unguessable so as to prevent connection hijacking. To satisfy
1313 * the requirements of this situation, the algorithm outlined in
1314 * RFC 1948 is used, with only small modifications.
1316 * Implementation details:
1318 * Time is based off the system timer, and is corrected so that it
1319 * increases by one megabyte per second. This allows for proper
1320 * recycling on high speed LANs while still leaving over an hour
1323 * As reading the *exact* system time is too expensive to be done
1324 * whenever setting up a TCP connection, we increment the time
1325 * offset in two ways. First, a small random positive increment
1326 * is added to isn_offset for each connection that is set up.
1327 * Second, the function tcp_isn_tick fires once per clock tick
1328 * and increments isn_offset as necessary so that sequence numbers
1329 * are incremented at approximately ISN_BYTES_PER_SECOND. The
1330 * random positive increments serve only to ensure that the same
1331 * exact sequence number is never sent out twice (as could otherwise
1332 * happen when a port is recycled in less than the system tick
1335 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1336 * between seeding of isn_secret. This is normally set to zero,
1337 * as reseeding should not be necessary.
1339 * Locking of the global variables isn_secret, isn_last_reseed, isn_offset,
1340 * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock. In
1341 * general, this means holding an exclusive (write) lock.
1344 #define ISN_BYTES_PER_SECOND 1048576
1345 #define ISN_STATIC_INCREMENT 4096
1346 #define ISN_RANDOM_INCREMENT (4096 - 1)
1348 static u_char isn_secret[32];
1349 static int isn_last_reseed;
1350 static u_int32_t isn_offset, isn_offset_old;
1351 static MD5_CTX isn_ctx;
1357 u_int32_t md5_buffer[4];
1360 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1361 INP_LOCK_ASSERT(tp->t_inpcb);
1363 /* Seed if this is the first use, reseed if requested. */
1364 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1365 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1367 read_random(&isn_secret, sizeof(isn_secret));
1368 isn_last_reseed = ticks;
1371 /* Compute the md5 hash and return the ISN. */
1373 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1374 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1376 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1377 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1378 sizeof(struct in6_addr));
1379 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1380 sizeof(struct in6_addr));
1384 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1385 sizeof(struct in_addr));
1386 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1387 sizeof(struct in_addr));
1389 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1390 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1391 new_isn = (tcp_seq) md5_buffer[0];
1392 isn_offset += ISN_STATIC_INCREMENT +
1393 (arc4random() & ISN_RANDOM_INCREMENT);
1394 new_isn += isn_offset;
1399 * Increment the offset to the next ISN_BYTES_PER_SECOND / hz boundary
1400 * to keep time flowing at a relatively constant rate. If the random
1401 * increments have already pushed us past the projected offset, do nothing.
1407 u_int32_t projected_offset;
1409 INP_INFO_WLOCK(&tcbinfo);
1410 projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100;
1412 if (projected_offset > isn_offset)
1413 isn_offset = projected_offset;
1415 isn_offset_old = isn_offset;
1416 callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL);
1417 INP_INFO_WUNLOCK(&tcbinfo);
1421 * When a specific ICMP unreachable message is received and the
1422 * connection state is SYN-SENT, drop the connection. This behavior
1423 * is controlled by the icmp_may_rst sysctl.
1426 tcp_drop_syn_sent(inp, errno)
1430 struct tcpcb *tp = intotcpcb(inp);
1432 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1433 INP_LOCK_ASSERT(inp);
1435 if (tp != NULL && tp->t_state == TCPS_SYN_SENT) {
1436 tcp_drop(tp, errno);
1443 * When `need fragmentation' ICMP is received, update our idea of the MSS
1444 * based on the new value in the route. Also nudge TCP to send something,
1445 * since we know the packet we just sent was dropped.
1446 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1449 tcp_mtudisc(inp, errno)
1453 struct tcpcb *tp = intotcpcb(inp);
1454 struct socket *so = inp->inp_socket;
1462 INP_LOCK_ASSERT(inp);
1465 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1467 maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */
1470 isipv6 ? tcp_maxmtu6(&inp->inp_inc) :
1472 tcp_maxmtu(&inp->inp_inc);
1476 maxmtu = min(maxmtu, romtu);
1478 tp->t_maxopd = tp->t_maxseg =
1480 isipv6 ? tcp_v6mssdflt :
1488 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1490 sizeof(struct tcpiphdr)
1497 * XXX - The above conditional probably violates the TCP
1498 * spec. The problem is that, since we don't know the
1499 * other end's MSS, we are supposed to use a conservative
1500 * default. But, if we do that, then MTU discovery will
1501 * never actually take place, because the conservative
1502 * default is much less than the MTUs typically seen
1503 * on the Internet today. For the moment, we'll sweep
1504 * this under the carpet.
1506 * The conservative default might not actually be a problem
1507 * if the only case this occurs is when sending an initial
1508 * SYN with options and data to a host we've never talked
1509 * to before. Then, they will reply with an MSS value which
1510 * will get recorded and the new parameters should get
1511 * recomputed. For Further Study.
1513 if (tp->t_maxopd <= mss)
1517 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1518 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1519 mss -= TCPOLEN_TSTAMP_APPA;
1520 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1522 mss &= ~(MCLBYTES-1);
1525 mss = mss / MCLBYTES * MCLBYTES;
1527 if (so->so_snd.sb_hiwat < mss)
1528 mss = so->so_snd.sb_hiwat;
1532 tcpstat.tcps_mturesent++;
1534 tp->snd_nxt = tp->snd_una;
1541 * Look-up the routing entry to the peer of this inpcb. If no route
1542 * is found and it cannot be allocated, then return NULL. This routine
1543 * is called by TCP routines that access the rmx structure and by tcp_mss
1544 * to get the interface MTU.
1548 struct in_conninfo *inc;
1551 struct sockaddr_in *dst;
1555 KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer"));
1557 bzero(&sro, sizeof(sro));
1558 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1559 dst = (struct sockaddr_in *)&sro.ro_dst;
1560 dst->sin_family = AF_INET;
1561 dst->sin_len = sizeof(*dst);
1562 dst->sin_addr = inc->inc_faddr;
1563 rtalloc_ign(&sro, RTF_CLONING);
1565 if (sro.ro_rt != NULL) {
1566 ifp = sro.ro_rt->rt_ifp;
1567 if (sro.ro_rt->rt_rmx.rmx_mtu == 0)
1568 maxmtu = ifp->if_mtu;
1570 maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu);
1579 struct in_conninfo *inc;
1581 struct route_in6 sro6;
1585 KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer"));
1587 bzero(&sro6, sizeof(sro6));
1588 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1589 sro6.ro_dst.sin6_family = AF_INET6;
1590 sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1591 sro6.ro_dst.sin6_addr = inc->inc6_faddr;
1592 rtalloc_ign((struct route *)&sro6, RTF_CLONING);
1594 if (sro6.ro_rt != NULL) {
1595 ifp = sro6.ro_rt->rt_ifp;
1596 if (sro6.ro_rt->rt_rmx.rmx_mtu == 0)
1597 maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp);
1599 maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu,
1600 IN6_LINKMTU(sro6.ro_rt->rt_ifp));
1609 /* compute ESP/AH header size for TCP, including outer IP header. */
1611 ipsec_hdrsiz_tcp(tp)
1619 struct ip6_hdr *ip6;
1623 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1625 MGETHDR(m, M_DONTWAIT, MT_DATA);
1630 if ((inp->inp_vflag & INP_IPV6) != 0) {
1631 ip6 = mtod(m, struct ip6_hdr *);
1632 th = (struct tcphdr *)(ip6 + 1);
1633 m->m_pkthdr.len = m->m_len =
1634 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1635 tcpip_fillheaders(inp, ip6, th);
1636 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1640 ip = mtod(m, struct ip *);
1641 th = (struct tcphdr *)(ip + 1);
1642 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1643 tcpip_fillheaders(inp, ip, th);
1644 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1653 * Move a TCP connection into TIME_WAIT state.
1654 * tcbinfo is locked.
1655 * inp is locked, and is unlocked before returning.
1663 int tw_time, acknow;
1666 INP_INFO_WLOCK_ASSERT(&tcbinfo); /* tcp_timer_2msl_reset(). */
1667 INP_LOCK_ASSERT(tp->t_inpcb);
1669 tw = uma_zalloc(tcptw_zone, M_NOWAIT);
1671 tw = tcp_timer_2msl_tw(1);
1681 * Recover last window size sent.
1683 tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
1686 * Set t_recent if timestamps are used on the connection.
1688 if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
1689 (TF_REQ_TSTMP|TF_RCVD_TSTMP))
1690 tw->t_recent = tp->ts_recent;
1694 tw->snd_nxt = tp->snd_nxt;
1695 tw->rcv_nxt = tp->rcv_nxt;
1698 tw->t_starttime = tp->t_starttime;
1703 * be used for fin-wait-2 state also, then we may need
1704 * a ts_recent from the last segment.
1706 tw_time = 2 * tcp_msl;
1707 acknow = tp->t_flags & TF_ACKNOW;
1709 so = inp->inp_socket;
1713 tw->tw_cred = crhold(so->so_cred);
1714 tw->tw_so_options = so->so_options;
1716 inp->inp_socket = NULL;
1718 tcp_twrespond(tw, TH_ACK);
1719 inp->inp_ppcb = (caddr_t)tw;
1720 inp->inp_vflag |= INP_TIMEWAIT;
1721 tcp_timer_2msl_reset(tw, tw_time);
1726 * The appromixate rate of ISN increase of Microsoft TCP stacks;
1727 * the actual rate is slightly higher due to the addition of
1728 * random positive increments.
1730 * Most other new OSes use semi-randomized ISN values, so we
1731 * do not need to worry about them.
1733 #define MS_ISN_BYTES_PER_SECOND 250000
1736 * Determine if the ISN we will generate has advanced beyond the last
1737 * sequence number used by the previous connection. If so, indicate
1738 * that it is safe to recycle this tw socket by returning 1.
1740 * XXXRW: This function should assert the inpcb lock as it does multiple
1741 * non-atomic reads from the tcptw, but is currently called without it from
1742 * in_pcb.c:in_pcblookup_local().
1745 tcp_twrecycleable(struct tcptw *tw)
1747 tcp_seq new_iss = tw->iss;
1748 tcp_seq new_irs = tw->irs;
1750 new_iss += (ticks - tw->t_starttime) * (ISN_BYTES_PER_SECOND / hz);
1751 new_irs += (ticks - tw->t_starttime) * (MS_ISN_BYTES_PER_SECOND / hz);
1753 if (SEQ_GT(new_iss, tw->snd_nxt) && SEQ_GT(new_irs, tw->rcv_nxt))
1760 tcp_twclose(struct tcptw *tw, int reuse)
1765 INP_INFO_WLOCK_ASSERT(&tcbinfo); /* tcp_timer_2msl_stop(). */
1766 INP_LOCK_ASSERT(inp);
1768 tw->tw_inpcb = NULL;
1769 tcp_timer_2msl_stop(tw);
1770 inp->inp_ppcb = NULL;
1772 if (inp->inp_vflag & INP_IPV6PROTO)
1777 tcpstat.tcps_closed++;
1778 crfree(tw->tw_cred);
1782 uma_zfree(tcptw_zone, tw);
1787 tcp_twrespond(struct tcptw *tw, int flags)
1789 struct inpcb *inp = tw->tw_inpcb;
1792 struct ip *ip = NULL;
1794 u_int hdrlen, optlen;
1797 struct ip6_hdr *ip6 = NULL;
1798 int isipv6 = inp->inp_inc.inc_isipv6;
1801 INP_LOCK_ASSERT(inp);
1803 m = m_gethdr(M_DONTWAIT, MT_DATA);
1806 m->m_data += max_linkhdr;
1809 mac_create_mbuf_from_inpcb(inp, m);
1814 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1815 ip6 = mtod(m, struct ip6_hdr *);
1816 th = (struct tcphdr *)(ip6 + 1);
1817 tcpip_fillheaders(inp, ip6, th);
1821 hdrlen = sizeof(struct tcpiphdr);
1822 ip = mtod(m, struct ip *);
1823 th = (struct tcphdr *)(ip + 1);
1824 tcpip_fillheaders(inp, ip, th);
1826 optp = (u_int8_t *)(th + 1);
1829 * Send a timestamp and echo-reply if both our side and our peer
1830 * have sent timestamps in our SYN's and this is not a RST.
1832 if (tw->t_recent && flags == TH_ACK) {
1833 u_int32_t *lp = (u_int32_t *)optp;
1835 /* Form timestamp option as shown in appendix A of RFC 1323. */
1836 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
1837 *lp++ = htonl(ticks);
1838 *lp = htonl(tw->t_recent);
1839 optp += TCPOLEN_TSTAMP_APPA;
1842 optlen = optp - (u_int8_t *)(th + 1);
1844 m->m_len = hdrlen + optlen;
1845 m->m_pkthdr.len = m->m_len;
1847 KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
1849 th->th_seq = htonl(tw->snd_nxt);
1850 th->th_ack = htonl(tw->rcv_nxt);
1851 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1852 th->th_flags = flags;
1853 th->th_win = htons(tw->last_win);
1857 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1858 sizeof(struct tcphdr) + optlen);
1859 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1860 error = ip6_output(m, inp->in6p_outputopts, NULL,
1861 (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
1865 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1866 htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
1867 m->m_pkthdr.csum_flags = CSUM_TCP;
1868 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1869 ip->ip_len = m->m_pkthdr.len;
1870 if (path_mtu_discovery)
1871 ip->ip_off |= IP_DF;
1872 error = ip_output(m, inp->inp_options, NULL,
1873 ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1877 tcpstat.tcps_sndacks++;
1879 tcpstat.tcps_sndctrl++;
1880 tcpstat.tcps_sndtotal++;
1885 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1887 * This code attempts to calculate the bandwidth-delay product as a
1888 * means of determining the optimal window size to maximize bandwidth,
1889 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1890 * routers. This code also does a fairly good job keeping RTTs in check
1891 * across slow links like modems. We implement an algorithm which is very
1892 * similar (but not meant to be) TCP/Vegas. The code operates on the
1893 * transmitter side of a TCP connection and so only effects the transmit
1894 * side of the connection.
1896 * BACKGROUND: TCP makes no provision for the management of buffer space
1897 * at the end points or at the intermediate routers and switches. A TCP
1898 * stream, whether using NewReno or not, will eventually buffer as
1899 * many packets as it is able and the only reason this typically works is
1900 * due to the fairly small default buffers made available for a connection
1901 * (typicaly 16K or 32K). As machines use larger windows and/or window
1902 * scaling it is now fairly easy for even a single TCP connection to blow-out
1903 * all available buffer space not only on the local interface, but on
1904 * intermediate routers and switches as well. NewReno makes a misguided
1905 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1906 * then backing off, then steadily increasing the window again until another
1907 * failure occurs, ad-infinitum. This results in terrible oscillation that
1908 * is only made worse as network loads increase and the idea of intentionally
1909 * blowing out network buffers is, frankly, a terrible way to manage network
1912 * It is far better to limit the transmit window prior to the failure
1913 * condition being achieved. There are two general ways to do this: First
1914 * you can 'scan' through different transmit window sizes and locate the
1915 * point where the RTT stops increasing, indicating that you have filled the
1916 * pipe, then scan backwards until you note that RTT stops decreasing, then
1917 * repeat ad-infinitum. This method works in principle but has severe
1918 * implementation issues due to RTT variances, timer granularity, and
1919 * instability in the algorithm which can lead to many false positives and
1920 * create oscillations as well as interact badly with other TCP streams
1921 * implementing the same algorithm.
1923 * The second method is to limit the window to the bandwidth delay product
1924 * of the link. This is the method we implement. RTT variances and our
1925 * own manipulation of the congestion window, bwnd, can potentially
1926 * destabilize the algorithm. For this reason we have to stabilize the
1927 * elements used to calculate the window. We do this by using the minimum
1928 * observed RTT, the long term average of the observed bandwidth, and
1929 * by adding two segments worth of slop. It isn't perfect but it is able
1930 * to react to changing conditions and gives us a very stable basis on
1931 * which to extend the algorithm.
1934 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1940 INP_LOCK_ASSERT(tp->t_inpcb);
1943 * If inflight_enable is disabled in the middle of a tcp connection,
1944 * make sure snd_bwnd is effectively disabled.
1946 if (tcp_inflight_enable == 0 || tp->t_rttlow < tcp_inflight_rttthresh) {
1947 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1948 tp->snd_bandwidth = 0;
1953 * Figure out the bandwidth. Due to the tick granularity this
1954 * is a very rough number and it MUST be averaged over a fairly
1955 * long period of time. XXX we need to take into account a link
1956 * that is not using all available bandwidth, but for now our
1957 * slop will ramp us up if this case occurs and the bandwidth later
1960 * Note: if ticks rollover 'bw' may wind up negative. We must
1961 * effectively reset t_bw_rtttime for this case.
1964 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1967 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1968 (save_ticks - tp->t_bw_rtttime);
1969 tp->t_bw_rtttime = save_ticks;
1970 tp->t_bw_rtseq = ack_seq;
1971 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1973 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1975 tp->snd_bandwidth = bw;
1978 * Calculate the semi-static bandwidth delay product, plus two maximal
1979 * segments. The additional slop puts us squarely in the sweet
1980 * spot and also handles the bandwidth run-up case and stabilization.
1981 * Without the slop we could be locking ourselves into a lower
1984 * Situations Handled:
1985 * (1) Prevents over-queueing of packets on LANs, especially on
1986 * high speed LANs, allowing larger TCP buffers to be
1987 * specified, and also does a good job preventing
1988 * over-queueing of packets over choke points like modems
1989 * (at least for the transmit side).
1991 * (2) Is able to handle changing network loads (bandwidth
1992 * drops so bwnd drops, bandwidth increases so bwnd
1995 * (3) Theoretically should stabilize in the face of multiple
1996 * connections implementing the same algorithm (this may need
1999 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2000 * be adjusted with a sysctl but typically only needs to be
2001 * on very slow connections. A value no smaller then 5
2002 * should be used, but only reduce this default if you have
2005 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
2006 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10;
2009 if (tcp_inflight_debug > 0) {
2011 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2013 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2022 if ((long)bwnd < tcp_inflight_min)
2023 bwnd = tcp_inflight_min;
2024 if (bwnd > tcp_inflight_max)
2025 bwnd = tcp_inflight_max;
2026 if ((long)bwnd < tp->t_maxseg * 2)
2027 bwnd = tp->t_maxseg * 2;
2028 tp->snd_bwnd = bwnd;
2031 #ifdef TCP_SIGNATURE
2033 * Callback function invoked by m_apply() to digest TCP segment data
2034 * contained within an mbuf chain.
2037 tcp_signature_apply(void *fstate, void *data, u_int len)
2040 MD5Update(fstate, (u_char *)data, len);
2045 * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385)
2048 * m pointer to head of mbuf chain
2049 * off0 offset to TCP header within the mbuf chain
2050 * len length of TCP segment data, excluding options
2051 * optlen length of TCP segment options
2052 * buf pointer to storage for computed MD5 digest
2053 * direction direction of flow (IPSEC_DIR_INBOUND or OUTBOUND)
2055 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2056 * When called from tcp_input(), we can be sure that th_sum has been
2057 * zeroed out and verified already.
2059 * This function is for IPv4 use only. Calling this function with an
2060 * IPv6 packet in the mbuf chain will yield undefined results.
2062 * Return 0 if successful, otherwise return -1.
2064 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2065 * search with the destination IP address, and a 'magic SPI' to be
2066 * determined by the application. This is hardcoded elsewhere to 1179
2067 * right now. Another branch of this code exists which uses the SPD to
2068 * specify per-application flows but it is unstable.
2071 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen,
2072 u_char *buf, u_int direction)
2074 union sockaddr_union dst;
2075 struct ippseudo ippseudo;
2079 struct ipovly *ipovly;
2080 struct secasvar *sav;
2084 KASSERT(m != NULL, ("NULL mbuf chain"));
2085 KASSERT(buf != NULL, ("NULL signature pointer"));
2087 /* Extract the destination from the IP header in the mbuf. */
2088 ip = mtod(m, struct ip *);
2089 bzero(&dst, sizeof(union sockaddr_union));
2090 dst.sa.sa_len = sizeof(struct sockaddr_in);
2091 dst.sa.sa_family = AF_INET;
2092 dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ?
2093 ip->ip_src : ip->ip_dst;
2095 /* Look up an SADB entry which matches the address of the peer. */
2096 sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI));
2098 printf("%s: SADB lookup failed for %s\n", __func__,
2099 inet_ntoa(dst.sin.sin_addr));
2104 ipovly = (struct ipovly *)ip;
2105 th = (struct tcphdr *)((u_char *)ip + off0);
2106 doff = off0 + sizeof(struct tcphdr) + optlen;
2109 * Step 1: Update MD5 hash with IP pseudo-header.
2111 * XXX The ippseudo header MUST be digested in network byte order,
2112 * or else we'll fail the regression test. Assume all fields we've
2113 * been doing arithmetic on have been in host byte order.
2114 * XXX One cannot depend on ipovly->ih_len here. When called from
2115 * tcp_output(), the underlying ip_len member has not yet been set.
2117 ippseudo.ippseudo_src = ipovly->ih_src;
2118 ippseudo.ippseudo_dst = ipovly->ih_dst;
2119 ippseudo.ippseudo_pad = 0;
2120 ippseudo.ippseudo_p = IPPROTO_TCP;
2121 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2122 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2125 * Step 2: Update MD5 hash with TCP header, excluding options.
2126 * The TCP checksum must be set to zero.
2128 savecsum = th->th_sum;
2130 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2131 th->th_sum = savecsum;
2134 * Step 3: Update MD5 hash with TCP segment data.
2135 * Use m_apply() to avoid an early m_pullup().
2138 m_apply(m, doff, len, tcp_signature_apply, &ctx);
2141 * Step 4: Update MD5 hash with shared secret.
2143 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2144 MD5Final(buf, &ctx);
2146 key_sa_recordxfer(sav, m);
2150 #endif /* TCP_SIGNATURE */
2153 sysctl_drop(SYSCTL_HANDLER_ARGS)
2155 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2156 struct sockaddr_storage addrs[2];
2160 struct sockaddr_in *fin, *lin;
2162 struct sockaddr_in6 *fin6, *lin6;
2163 struct in6_addr f6, l6;
2174 if (req->oldptr != NULL || req->oldlen != 0)
2176 if (req->newptr == NULL)
2178 if (req->newlen < sizeof(addrs))
2180 error = SYSCTL_IN(req, &addrs, sizeof(addrs));
2184 switch (addrs[0].ss_family) {
2187 fin6 = (struct sockaddr_in6 *)&addrs[0];
2188 lin6 = (struct sockaddr_in6 *)&addrs[1];
2189 if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
2190 lin6->sin6_len != sizeof(struct sockaddr_in6))
2192 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) {
2193 if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2195 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]);
2196 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]);
2197 fin = (struct sockaddr_in *)&addrs[0];
2198 lin = (struct sockaddr_in *)&addrs[1];
2201 error = sa6_embedscope(fin6, ip6_use_defzone);
2204 error = sa6_embedscope(lin6, ip6_use_defzone);
2210 fin = (struct sockaddr_in *)&addrs[0];
2211 lin = (struct sockaddr_in *)&addrs[1];
2212 if (fin->sin_len != sizeof(struct sockaddr_in) ||
2213 lin->sin_len != sizeof(struct sockaddr_in))
2219 INP_INFO_WLOCK(&tcbinfo);
2220 switch (addrs[0].ss_family) {
2223 inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port,
2224 &l6, lin6->sin6_port, 0, NULL);
2228 inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port,
2229 lin->sin_addr, lin->sin_port, 0, NULL);
2234 if ((tw = intotw(inp)) &&
2235 (inp->inp_vflag & INP_TIMEWAIT) != 0) {
2236 (void) tcp_twclose(tw, 0);
2237 } else if ((tp = intotcpcb(inp)) &&
2238 ((inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) {
2239 tp = tcp_drop(tp, ECONNABORTED);
2246 INP_INFO_WUNLOCK(&tcbinfo);
2250 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop,
2251 CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL,
2252 0, sysctl_drop, "", "Drop TCP connection");