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"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/callout.h>
43 #include <sys/kernel.h>
44 #include <sys/sysctl.h>
45 #include <sys/malloc.h>
48 #include <sys/domain.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/protosw.h>
55 #include <sys/random.h>
59 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_systm.h>
64 #include <netinet/ip.h>
66 #include <netinet/ip6.h>
68 #include <netinet/in_pcb.h>
70 #include <netinet6/in6_pcb.h>
72 #include <netinet/in_var.h>
73 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 #include <netinet6/scope6_var.h>
77 #include <netinet6/nd6.h>
79 #include <netinet/ip_icmp.h>
80 #include <netinet/tcp.h>
81 #include <netinet/tcp_fsm.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
86 #include <netinet6/tcp6_var.h>
88 #include <netinet/tcpip.h>
90 #include <netinet/tcp_debug.h>
92 #include <netinet6/ip6protosw.h>
95 #include <netinet6/ipsec.h>
97 #include <netinet6/ipsec6.h>
99 #include <netkey/key.h>
103 #include <netipsec/ipsec.h>
104 #include <netipsec/xform.h>
106 #include <netipsec/ipsec6.h>
108 #include <netipsec/key.h>
110 #endif /*FAST_IPSEC*/
112 #include <machine/in_cksum.h>
115 #include <security/mac/mac_framework.h>
117 int tcp_mssdflt = TCP_MSS;
118 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
119 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
122 int tcp_v6mssdflt = TCP6_MSS;
123 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
124 CTLFLAG_RW, &tcp_v6mssdflt , 0,
125 "Default TCP Maximum Segment Size for IPv6");
129 * Minimum MSS we accept and use. This prevents DoS attacks where
130 * we are forced to a ridiculous low MSS like 20 and send hundreds
131 * of packets instead of one. The effect scales with the available
132 * bandwidth and quickly saturates the CPU and network interface
133 * with packet generation and sending. Set to zero to disable MINMSS
134 * checking. This setting prevents us from sending too small packets.
136 int tcp_minmss = TCP_MINMSS;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
138 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
140 int tcp_do_rfc1323 = 1;
141 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
142 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
144 static int tcp_tcbhashsize = 0;
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN,
146 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
148 static int do_tcpdrain = 1;
149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW,
151 "Enable tcp_drain routine for extra help when low on mbufs");
153 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
154 &tcbinfo.ipi_count, 0, "Number of active PCBs");
156 static int icmp_may_rst = 1;
157 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW,
159 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
161 static int tcp_isn_reseed_interval = 0;
162 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
163 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
166 * TCP bandwidth limiting sysctls. Note that the default lower bound of
167 * 1024 exists only for debugging. A good production default would be
168 * something like 6100.
170 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, inflight, CTLFLAG_RW, 0,
171 "TCP inflight data limiting");
173 static int tcp_inflight_enable = 1;
174 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, enable, CTLFLAG_RW,
175 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
177 static int tcp_inflight_debug = 0;
178 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, debug, CTLFLAG_RW,
179 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
181 static int tcp_inflight_rttthresh;
182 SYSCTL_PROC(_net_inet_tcp_inflight, OID_AUTO, rttthresh, CTLTYPE_INT|CTLFLAG_RW,
183 &tcp_inflight_rttthresh, 0, sysctl_msec_to_ticks, "I",
184 "RTT threshold below which inflight will deactivate itself");
186 static int tcp_inflight_min = 6144;
187 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW,
188 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
190 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
191 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW,
192 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
194 static int tcp_inflight_stab = 20;
195 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW,
196 &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets");
198 uma_zone_t sack_hole_zone;
200 static struct inpcb *tcp_notify(struct inpcb *, int);
201 static void tcp_isn_tick(void *);
204 * Target size of TCP PCB hash tables. Must be a power of two.
206 * Note that this can be overridden by the kernel environment
207 * variable net.inet.tcp.tcbhashsize
210 #define TCBHASHSIZE 512
215 * Callouts should be moved into struct tcp directly. They are currently
216 * separate because the tcpcb structure is exported to userland for sysctl
217 * parsing purposes, which do not know about callouts.
224 static uma_zone_t tcpcb_zone;
225 MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers");
226 struct callout isn_callout;
227 static struct mtx isn_mtx;
229 #define ISN_LOCK_INIT() mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF)
230 #define ISN_LOCK() mtx_lock(&isn_mtx)
231 #define ISN_UNLOCK() mtx_unlock(&isn_mtx)
234 * TCP initialization.
237 tcp_zone_change(void *tag)
240 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
241 uma_zone_set_max(tcpcb_zone, maxsockets);
242 tcp_tw_zone_change();
246 tcp_inpcb_init(void *mem, int size, int flags)
248 struct inpcb *inp = mem;
250 INP_LOCK_INIT(inp, "inp", "tcpinp");
258 int hashsize = TCBHASHSIZE;
259 tcp_delacktime = TCPTV_DELACK;
260 tcp_keepinit = TCPTV_KEEP_INIT;
261 tcp_keepidle = TCPTV_KEEP_IDLE;
262 tcp_keepintvl = TCPTV_KEEPINTVL;
263 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
265 tcp_rexmit_min = TCPTV_MIN;
266 tcp_rexmit_slop = TCPTV_CPU_VAR;
267 tcp_inflight_rttthresh = TCPTV_INFLIGHT_RTTTHRESH;
268 tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT;
270 INP_INFO_LOCK_INIT(&tcbinfo, "tcp");
272 tcbinfo.ipi_listhead = &tcb;
273 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
274 if (!powerof2(hashsize)) {
275 printf("WARNING: TCB hash size not a power of 2\n");
276 hashsize = 512; /* safe default */
278 tcp_tcbhashsize = hashsize;
279 tcbinfo.ipi_hashbase = hashinit(hashsize, M_PCB,
280 &tcbinfo.ipi_hashmask);
281 tcbinfo.ipi_porthashbase = hashinit(hashsize, M_PCB,
282 &tcbinfo.ipi_porthashmask);
283 tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb),
284 NULL, NULL, tcp_inpcb_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
285 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
287 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
289 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
291 if (max_protohdr < TCP_MINPROTOHDR)
292 max_protohdr = TCP_MINPROTOHDR;
293 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
295 #undef TCP_MINPROTOHDR
297 * These have to be type stable for the benefit of the timers.
299 tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem),
300 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
301 uma_zone_set_max(tcpcb_zone, maxsockets);
307 callout_init(&isn_callout, CALLOUT_MPSAFE);
309 EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL,
310 SHUTDOWN_PRI_DEFAULT);
311 sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole),
312 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL,
314 EVENTHANDLER_PRI_ANY);
321 callout_stop(&isn_callout);
325 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
326 * tcp_template used to store this data in mbufs, but we now recopy it out
327 * of the tcpcb each time to conserve mbufs.
330 tcpip_fillheaders(struct inpcb *inp, void *ip_ptr, void *tcp_ptr)
332 struct tcphdr *th = (struct tcphdr *)tcp_ptr;
334 INP_LOCK_ASSERT(inp);
337 if ((inp->inp_vflag & INP_IPV6) != 0) {
340 ip6 = (struct ip6_hdr *)ip_ptr;
341 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
342 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
343 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
344 (IPV6_VERSION & IPV6_VERSION_MASK);
345 ip6->ip6_nxt = IPPROTO_TCP;
346 ip6->ip6_plen = sizeof(struct tcphdr);
347 ip6->ip6_src = inp->in6p_laddr;
348 ip6->ip6_dst = inp->in6p_faddr;
354 ip = (struct ip *)ip_ptr;
355 ip->ip_v = IPVERSION;
357 ip->ip_tos = inp->inp_ip_tos;
361 ip->ip_ttl = inp->inp_ip_ttl;
363 ip->ip_p = IPPROTO_TCP;
364 ip->ip_src = inp->inp_laddr;
365 ip->ip_dst = inp->inp_faddr;
367 th->th_sport = inp->inp_lport;
368 th->th_dport = inp->inp_fport;
376 th->th_sum = 0; /* in_pseudo() is called later for ipv4 */
380 * Create template to be used to send tcp packets on a connection.
381 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
382 * use for this function is in keepalives, which use tcp_respond.
385 tcpip_maketemplate(struct inpcb *inp)
390 m = m_get(M_DONTWAIT, MT_DATA);
393 m->m_len = sizeof(struct tcptemp);
394 n = mtod(m, struct tcptemp *);
396 tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
401 * Send a single message to the TCP at address specified by
402 * the given TCP/IP header. If m == NULL, then we make a copy
403 * of the tcpiphdr at ti and send directly to the addressed host.
404 * This is used to force keep alive messages out using the TCP
405 * template for a connection. If flags are given then we send
406 * a message back to the TCP which originated the * segment ti,
407 * and discard the mbuf containing it and any other attached mbufs.
409 * In any case the ack and sequence number of the transmitted
410 * segment are as specified by the parameters.
412 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
415 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
416 tcp_seq ack, tcp_seq seq, int flags)
429 KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL"));
432 isipv6 = ((struct ip *)ipgen)->ip_v == 6;
439 KASSERT(inp != NULL, ("tcp control block w/o inpcb"));
440 INP_LOCK_ASSERT(inp);
445 if (!(flags & TH_RST)) {
446 win = sbspace(&inp->inp_socket->so_rcv);
447 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
448 win = (long)TCP_MAXWIN << tp->rcv_scale;
452 m = m_gethdr(M_DONTWAIT, MT_DATA);
456 m->m_data += max_linkhdr;
459 bcopy((caddr_t)ip6, mtod(m, caddr_t),
460 sizeof(struct ip6_hdr));
461 ip6 = mtod(m, struct ip6_hdr *);
462 nth = (struct tcphdr *)(ip6 + 1);
466 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
467 ip = mtod(m, struct ip *);
468 nth = (struct tcphdr *)(ip + 1);
470 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
475 m->m_data = (caddr_t)ipgen;
476 /* m_len is set later */
478 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
481 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
482 nth = (struct tcphdr *)(ip6 + 1);
486 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
487 nth = (struct tcphdr *)(ip + 1);
491 * this is usually a case when an extension header
492 * exists between the IPv6 header and the
495 nth->th_sport = th->th_sport;
496 nth->th_dport = th->th_dport;
498 xchg(nth->th_dport, nth->th_sport, n_short);
504 ip6->ip6_vfc = IPV6_VERSION;
505 ip6->ip6_nxt = IPPROTO_TCP;
506 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
508 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
512 tlen += sizeof (struct tcpiphdr);
514 ip->ip_ttl = ip_defttl;
515 if (path_mtu_discovery)
519 m->m_pkthdr.len = tlen;
520 m->m_pkthdr.rcvif = NULL;
524 * Packet is associated with a socket, so allow the
525 * label of the response to reflect the socket label.
527 INP_LOCK_ASSERT(inp);
528 mac_create_mbuf_from_inpcb(inp, m);
531 * Packet is not associated with a socket, so possibly
532 * update the label in place.
534 mac_reflect_mbuf_tcp(m);
537 nth->th_seq = htonl(seq);
538 nth->th_ack = htonl(ack);
540 nth->th_off = sizeof (struct tcphdr) >> 2;
541 nth->th_flags = flags;
543 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
545 nth->th_win = htons((u_short)win);
550 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
551 sizeof(struct ip6_hdr),
552 tlen - sizeof(struct ip6_hdr));
553 ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb :
558 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
559 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
560 m->m_pkthdr.csum_flags = CSUM_TCP;
561 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
564 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
565 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
569 (void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp);
572 (void) ip_output(m, NULL, NULL, ipflags, NULL, inp);
576 * Create a new TCP control block, making an
577 * empty reassembly queue and hooking it to the argument
578 * protocol control block. The `inp' parameter must have
579 * come from the zone allocator set up in tcp_init().
582 tcp_newtcpcb(struct inpcb *inp)
584 struct tcpcb_mem *tm;
587 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
590 tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO);
594 tp->t_timers = &tm->tt;
595 /* LIST_INIT(&tp->t_segq); */ /* XXX covered by M_ZERO */
596 tp->t_maxseg = tp->t_maxopd =
598 isipv6 ? tcp_v6mssdflt :
602 /* Set up our timeouts. */
603 if (NET_CALLOUT_MPSAFE)
604 callout_init_mtx(&tp->t_timers->tt_timer, &inp->inp_mtx,
605 CALLOUT_RETURNUNLOCKED);
607 callout_init_mtx(&tp->t_timers->tt_timer, &inp->inp_mtx,
608 (CALLOUT_RETURNUNLOCKED|CALLOUT_NETGIANT));
611 tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
613 tp->t_flags |= TF_SACK_PERMIT;
614 TAILQ_INIT(&tp->snd_holes);
615 tp->t_inpcb = inp; /* XXX */
617 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
618 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
619 * reasonable initial retransmit time.
621 tp->t_srtt = TCPTV_SRTTBASE;
622 tp->t_rttvar = ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
623 tp->t_rttmin = tcp_rexmit_min;
624 tp->t_rxtcur = TCPTV_RTOBASE;
625 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
626 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
627 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
628 tp->t_rcvtime = ticks;
629 tp->t_bw_rtttime = ticks;
631 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
632 * because the socket may be bound to an IPv6 wildcard address,
633 * which may match an IPv4-mapped IPv6 address.
635 inp->inp_ip_ttl = ip_defttl;
637 return (tp); /* XXX */
641 * Drop a TCP connection, reporting
642 * the specified error. If connection is synchronized,
643 * then send a RST to peer.
646 tcp_drop(struct tcpcb *tp, int errno)
648 struct socket *so = tp->t_inpcb->inp_socket;
650 INP_INFO_WLOCK_ASSERT(&tcbinfo);
651 INP_LOCK_ASSERT(tp->t_inpcb);
653 if (TCPS_HAVERCVDSYN(tp->t_state)) {
654 tp->t_state = TCPS_CLOSED;
655 (void) tcp_output(tp);
656 tcpstat.tcps_drops++;
658 tcpstat.tcps_conndrops++;
659 if (errno == ETIMEDOUT && tp->t_softerror)
660 errno = tp->t_softerror;
661 so->so_error = errno;
662 return (tcp_close(tp));
666 tcp_discardcb(struct tcpcb *tp)
669 struct inpcb *inp = tp->t_inpcb;
670 struct socket *so = inp->inp_socket;
672 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
675 INP_LOCK_ASSERT(inp);
678 * Make sure that all of our timers are stopped before we
681 * XXX: callout_stop() may race and a callout may already
682 * try to obtain the INP_LOCK. Only callout_drain() would
683 * stop this but it would cause a LOR thus we can't use it.
684 * The tcp_timer() function contains a lot of checks to
685 * handle this case rather gracefully.
687 tp->t_timers->tt_active = 0;
688 callout_stop(&tp->t_timers->tt_timer);
691 * If we got enough samples through the srtt filter,
692 * save the rtt and rttvar in the routing entry.
693 * 'Enough' is arbitrarily defined as 4 rtt samples.
694 * 4 samples is enough for the srtt filter to converge
695 * to within enough % of the correct value; fewer samples
696 * and we could save a bogus rtt. The danger is not high
697 * as tcp quickly recovers from everything.
698 * XXX: Works very well but needs some more statistics!
700 if (tp->t_rttupdated >= 4) {
701 struct hc_metrics_lite metrics;
704 bzero(&metrics, sizeof(metrics));
706 * Update the ssthresh always when the conditions below
707 * are satisfied. This gives us better new start value
708 * for the congestion avoidance for new connections.
709 * ssthresh is only set if packet loss occured on a session.
711 * XXXRW: 'so' may be NULL here, and/or socket buffer may be
712 * being torn down. Ideally this code would not use 'so'.
714 ssthresh = tp->snd_ssthresh;
715 if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) {
717 * convert the limit from user data bytes to
718 * packets then to packet data bytes.
720 ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg;
723 ssthresh *= (u_long)(tp->t_maxseg +
725 (isipv6 ? sizeof (struct ip6_hdr) +
726 sizeof (struct tcphdr) :
728 sizeof (struct tcpiphdr)
735 metrics.rmx_ssthresh = ssthresh;
737 metrics.rmx_rtt = tp->t_srtt;
738 metrics.rmx_rttvar = tp->t_rttvar;
739 /* XXX: This wraps if the pipe is more than 4 Gbit per second */
740 metrics.rmx_bandwidth = tp->snd_bandwidth;
741 metrics.rmx_cwnd = tp->snd_cwnd;
742 metrics.rmx_sendpipe = 0;
743 metrics.rmx_recvpipe = 0;
745 tcp_hc_update(&inp->inp_inc, &metrics);
748 /* free the reassembly queue, if any */
749 while ((q = LIST_FIRST(&tp->t_segq)) != NULL) {
750 LIST_REMOVE(q, tqe_q);
752 uma_zfree(tcp_reass_zone, q);
756 tcp_free_sackholes(tp);
757 inp->inp_ppcb = NULL;
759 uma_zfree(tcpcb_zone, tp);
763 * Attempt to close a TCP control block, marking it as dropped, and freeing
764 * the socket if we hold the only reference.
767 tcp_close(struct tcpcb *tp)
769 struct inpcb *inp = tp->t_inpcb;
772 INP_INFO_WLOCK_ASSERT(&tcbinfo);
773 INP_LOCK_ASSERT(inp);
776 tcpstat.tcps_closed++;
777 KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL"));
778 so = inp->inp_socket;
779 soisdisconnected(so);
780 if (inp->inp_vflag & INP_SOCKREF) {
781 KASSERT(so->so_state & SS_PROTOREF,
782 ("tcp_close: !SS_PROTOREF"));
783 inp->inp_vflag &= ~INP_SOCKREF;
787 so->so_state &= ~SS_PROTOREF;
801 struct tseg_qent *te;
804 * Walk the tcpbs, if existing, and flush the reassembly queue,
806 * XXX: The "Net/3" implementation doesn't imply that the TCP
807 * reassembly queue should be flushed, but in a situation
808 * where we're really low on mbufs, this is potentially
811 INP_INFO_RLOCK(&tcbinfo);
812 LIST_FOREACH(inpb, tcbinfo.ipi_listhead, inp_list) {
813 if (inpb->inp_vflag & INP_TIMEWAIT)
816 if ((tcpb = intotcpcb(inpb)) != NULL) {
817 while ((te = LIST_FIRST(&tcpb->t_segq))
819 LIST_REMOVE(te, tqe_q);
821 uma_zfree(tcp_reass_zone, te);
825 tcp_clean_sackreport(tcpb);
829 INP_INFO_RUNLOCK(&tcbinfo);
834 * Notify a tcp user of an asynchronous error;
835 * store error as soft error, but wake up user
836 * (for now, won't do anything until can select for soft error).
838 * Do not wake up user since there currently is no mechanism for
839 * reporting soft errors (yet - a kqueue filter may be added).
841 static struct inpcb *
842 tcp_notify(struct inpcb *inp, int error)
846 INP_INFO_WLOCK_ASSERT(&tcbinfo);
847 INP_LOCK_ASSERT(inp);
849 if ((inp->inp_vflag & INP_TIMEWAIT) ||
850 (inp->inp_vflag & INP_DROPPED))
854 KASSERT(tp != NULL, ("tcp_notify: tp == NULL"));
857 * Ignore some errors if we are hooked up.
858 * If connection hasn't completed, has retransmitted several times,
859 * and receives a second error, give up now. This is better
860 * than waiting a long time to establish a connection that
861 * can never complete.
863 if (tp->t_state == TCPS_ESTABLISHED &&
864 (error == EHOSTUNREACH || error == ENETUNREACH ||
865 error == EHOSTDOWN)) {
867 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
869 tp = tcp_drop(tp, error);
875 tp->t_softerror = error;
879 wakeup( &so->so_timeo);
886 tcp_pcblist(SYSCTL_HANDLER_ARGS)
889 struct inpcb *inp, **inp_list;
894 * The process of preparing the TCB list is too time-consuming and
895 * resource-intensive to repeat twice on every request.
897 if (req->oldptr == NULL) {
898 n = tcbinfo.ipi_count;
899 req->oldidx = 2 * (sizeof xig)
900 + (n + n/8) * sizeof(struct xtcpcb);
904 if (req->newptr != NULL)
908 * OK, now we're committed to doing something.
910 INP_INFO_RLOCK(&tcbinfo);
911 gencnt = tcbinfo.ipi_gencnt;
912 n = tcbinfo.ipi_count;
913 INP_INFO_RUNLOCK(&tcbinfo);
915 error = sysctl_wire_old_buffer(req, 2 * (sizeof xig)
916 + n * sizeof(struct xtcpcb));
920 xig.xig_len = sizeof xig;
922 xig.xig_gen = gencnt;
923 xig.xig_sogen = so_gencnt;
924 error = SYSCTL_OUT(req, &xig, sizeof xig);
928 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
929 if (inp_list == NULL)
932 INP_INFO_RLOCK(&tcbinfo);
933 for (inp = LIST_FIRST(tcbinfo.ipi_listhead), i = 0; inp != NULL && i
934 < n; inp = LIST_NEXT(inp, inp_list)) {
936 if (inp->inp_gencnt <= gencnt) {
938 * XXX: This use of cr_cansee(), introduced with
939 * TCP state changes, is not quite right, but for
940 * now, better than nothing.
942 if (inp->inp_vflag & INP_TIMEWAIT) {
943 if (intotw(inp) != NULL)
944 error = cr_cansee(req->td->td_ucred,
945 intotw(inp)->tw_cred);
947 error = EINVAL; /* Skip this inp. */
949 error = cr_canseesocket(req->td->td_ucred,
956 INP_INFO_RUNLOCK(&tcbinfo);
960 for (i = 0; i < n; i++) {
963 if (inp->inp_gencnt <= gencnt) {
967 bzero(&xt, sizeof(xt));
968 xt.xt_len = sizeof xt;
969 /* XXX should avoid extra copy */
970 bcopy(inp, &xt.xt_inp, sizeof *inp);
971 inp_ppcb = inp->inp_ppcb;
972 if (inp_ppcb == NULL)
973 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
974 else if (inp->inp_vflag & INP_TIMEWAIT) {
975 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
976 xt.xt_tp.t_state = TCPS_TIME_WAIT;
978 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
979 if (inp->inp_socket != NULL)
980 sotoxsocket(inp->inp_socket, &xt.xt_socket);
982 bzero(&xt.xt_socket, sizeof xt.xt_socket);
983 xt.xt_socket.xso_protocol = IPPROTO_TCP;
985 xt.xt_inp.inp_gencnt = inp->inp_gencnt;
987 error = SYSCTL_OUT(req, &xt, sizeof xt);
994 * Give the user an updated idea of our state.
995 * If the generation differs from what we told
996 * her before, she knows that something happened
997 * while we were processing this request, and it
998 * might be necessary to retry.
1000 INP_INFO_RLOCK(&tcbinfo);
1001 xig.xig_gen = tcbinfo.ipi_gencnt;
1002 xig.xig_sogen = so_gencnt;
1003 xig.xig_count = tcbinfo.ipi_count;
1004 INP_INFO_RUNLOCK(&tcbinfo);
1005 error = SYSCTL_OUT(req, &xig, sizeof xig);
1007 free(inp_list, M_TEMP);
1011 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1012 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1015 tcp_getcred(SYSCTL_HANDLER_ARGS)
1018 struct sockaddr_in addrs[2];
1022 error = priv_check_cred(req->td->td_ucred, PRIV_NETINET_GETCRED,
1026 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1029 INP_INFO_RLOCK(&tcbinfo);
1030 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
1031 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1037 if (inp->inp_socket == NULL) {
1041 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1044 cru2x(inp->inp_socket->so_cred, &xuc);
1048 INP_INFO_RUNLOCK(&tcbinfo);
1050 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1054 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred,
1055 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1056 tcp_getcred, "S,xucred", "Get the xucred of a TCP connection");
1060 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1063 struct sockaddr_in6 addrs[2];
1065 int error, mapped = 0;
1067 error = priv_check_cred(req->td->td_ucred, PRIV_NETINET_GETCRED,
1071 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1074 if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 ||
1075 (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) {
1078 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1079 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1085 INP_INFO_RLOCK(&tcbinfo);
1087 inp = in_pcblookup_hash(&tcbinfo,
1088 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1090 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1094 inp = in6_pcblookup_hash(&tcbinfo,
1095 &addrs[1].sin6_addr, addrs[1].sin6_port,
1096 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1102 if (inp->inp_socket == NULL) {
1106 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1109 cru2x(inp->inp_socket->so_cred, &xuc);
1113 INP_INFO_RUNLOCK(&tcbinfo);
1115 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1119 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred,
1120 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1121 tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection");
1126 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1128 struct ip *ip = vip;
1130 struct in_addr faddr;
1133 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1135 struct in_conninfo inc;
1136 tcp_seq icmp_tcp_seq;
1139 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1140 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1143 if (cmd == PRC_MSGSIZE)
1144 notify = tcp_mtudisc;
1145 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1146 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1147 notify = tcp_drop_syn_sent;
1149 * Redirects don't need to be handled up here.
1151 else if (PRC_IS_REDIRECT(cmd))
1154 * Source quench is depreciated.
1156 else if (cmd == PRC_QUENCH)
1159 * Hostdead is ugly because it goes linearly through all PCBs.
1160 * XXX: We never get this from ICMP, otherwise it makes an
1161 * excellent DoS attack on machines with many connections.
1163 else if (cmd == PRC_HOSTDEAD)
1165 else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0)
1168 icp = (struct icmp *)((caddr_t)ip
1169 - offsetof(struct icmp, icmp_ip));
1170 th = (struct tcphdr *)((caddr_t)ip
1171 + (ip->ip_hl << 2));
1172 INP_INFO_WLOCK(&tcbinfo);
1173 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1174 ip->ip_src, th->th_sport, 0, NULL);
1177 if (!(inp->inp_vflag & INP_TIMEWAIT) &&
1178 !(inp->inp_vflag & INP_DROPPED) &&
1179 !(inp->inp_socket == NULL)) {
1180 icmp_tcp_seq = htonl(th->th_seq);
1181 tp = intotcpcb(inp);
1182 if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) &&
1183 SEQ_LT(icmp_tcp_seq, tp->snd_max)) {
1184 if (cmd == PRC_MSGSIZE) {
1187 * If we got a needfrag set the MTU
1188 * in the route to the suggested new
1189 * value (if given) and then notify.
1191 bzero(&inc, sizeof(inc));
1192 inc.inc_flags = 0; /* IPv4 */
1193 inc.inc_faddr = faddr;
1195 mtu = ntohs(icp->icmp_nextmtu);
1197 * If no alternative MTU was
1198 * proposed, try the next smaller
1199 * one. ip->ip_len has already
1200 * been swapped in icmp_input().
1203 mtu = ip_next_mtu(ip->ip_len,
1205 if (mtu < max(296, (tcp_minmss)
1206 + sizeof(struct tcpiphdr)))
1210 + sizeof(struct tcpiphdr);
1212 * Only cache the the MTU if it
1213 * is smaller than the interface
1214 * or route MTU. tcp_mtudisc()
1215 * will do right thing by itself.
1217 if (mtu <= tcp_maxmtu(&inc, NULL))
1218 tcp_hc_updatemtu(&inc, mtu);
1221 inp = (*notify)(inp, inetctlerrmap[cmd]);
1227 inc.inc_fport = th->th_dport;
1228 inc.inc_lport = th->th_sport;
1229 inc.inc_faddr = faddr;
1230 inc.inc_laddr = ip->ip_src;
1234 syncache_unreach(&inc, th);
1236 INP_INFO_WUNLOCK(&tcbinfo);
1238 in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify);
1243 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1246 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1247 struct ip6_hdr *ip6;
1249 struct ip6ctlparam *ip6cp = NULL;
1250 const struct sockaddr_in6 *sa6_src = NULL;
1252 struct tcp_portonly {
1257 if (sa->sa_family != AF_INET6 ||
1258 sa->sa_len != sizeof(struct sockaddr_in6))
1261 if (cmd == PRC_MSGSIZE)
1262 notify = tcp_mtudisc;
1263 else if (!PRC_IS_REDIRECT(cmd) &&
1264 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1266 /* Source quench is depreciated. */
1267 else if (cmd == PRC_QUENCH)
1270 /* if the parameter is from icmp6, decode it. */
1272 ip6cp = (struct ip6ctlparam *)d;
1274 ip6 = ip6cp->ip6c_ip6;
1275 off = ip6cp->ip6c_off;
1276 sa6_src = ip6cp->ip6c_src;
1280 off = 0; /* fool gcc */
1285 struct in_conninfo inc;
1287 * XXX: We assume that when IPV6 is non NULL,
1288 * M and OFF are valid.
1291 /* check if we can safely examine src and dst ports */
1292 if (m->m_pkthdr.len < off + sizeof(*thp))
1295 bzero(&th, sizeof(th));
1296 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1298 in6_pcbnotify(&tcbinfo, sa, th.th_dport,
1299 (struct sockaddr *)ip6cp->ip6c_src,
1300 th.th_sport, cmd, NULL, notify);
1302 inc.inc_fport = th.th_dport;
1303 inc.inc_lport = th.th_sport;
1304 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1305 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1307 INP_INFO_WLOCK(&tcbinfo);
1308 syncache_unreach(&inc, &th);
1309 INP_INFO_WUNLOCK(&tcbinfo);
1311 in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src,
1312 0, cmd, NULL, notify);
1318 * Following is where TCP initial sequence number generation occurs.
1320 * There are two places where we must use initial sequence numbers:
1321 * 1. In SYN-ACK packets.
1322 * 2. In SYN packets.
1324 * All ISNs for SYN-ACK packets are generated by the syncache. See
1325 * tcp_syncache.c for details.
1327 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1328 * depends on this property. In addition, these ISNs should be
1329 * unguessable so as to prevent connection hijacking. To satisfy
1330 * the requirements of this situation, the algorithm outlined in
1331 * RFC 1948 is used, with only small modifications.
1333 * Implementation details:
1335 * Time is based off the system timer, and is corrected so that it
1336 * increases by one megabyte per second. This allows for proper
1337 * recycling on high speed LANs while still leaving over an hour
1340 * As reading the *exact* system time is too expensive to be done
1341 * whenever setting up a TCP connection, we increment the time
1342 * offset in two ways. First, a small random positive increment
1343 * is added to isn_offset for each connection that is set up.
1344 * Second, the function tcp_isn_tick fires once per clock tick
1345 * and increments isn_offset as necessary so that sequence numbers
1346 * are incremented at approximately ISN_BYTES_PER_SECOND. The
1347 * random positive increments serve only to ensure that the same
1348 * exact sequence number is never sent out twice (as could otherwise
1349 * happen when a port is recycled in less than the system tick
1352 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1353 * between seeding of isn_secret. This is normally set to zero,
1354 * as reseeding should not be necessary.
1356 * Locking of the global variables isn_secret, isn_last_reseed, isn_offset,
1357 * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock. In
1358 * general, this means holding an exclusive (write) lock.
1361 #define ISN_BYTES_PER_SECOND 1048576
1362 #define ISN_STATIC_INCREMENT 4096
1363 #define ISN_RANDOM_INCREMENT (4096 - 1)
1365 static u_char isn_secret[32];
1366 static int isn_last_reseed;
1367 static u_int32_t isn_offset, isn_offset_old;
1368 static MD5_CTX isn_ctx;
1371 tcp_new_isn(struct tcpcb *tp)
1373 u_int32_t md5_buffer[4];
1376 INP_LOCK_ASSERT(tp->t_inpcb);
1379 /* Seed if this is the first use, reseed if requested. */
1380 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1381 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1383 read_random(&isn_secret, sizeof(isn_secret));
1384 isn_last_reseed = ticks;
1387 /* Compute the md5 hash and return the ISN. */
1389 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1390 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1392 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1393 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1394 sizeof(struct in6_addr));
1395 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1396 sizeof(struct in6_addr));
1400 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1401 sizeof(struct in_addr));
1402 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1403 sizeof(struct in_addr));
1405 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1406 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1407 new_isn = (tcp_seq) md5_buffer[0];
1408 isn_offset += ISN_STATIC_INCREMENT +
1409 (arc4random() & ISN_RANDOM_INCREMENT);
1410 new_isn += isn_offset;
1416 * Increment the offset to the next ISN_BYTES_PER_SECOND / hz boundary
1417 * to keep time flowing at a relatively constant rate. If the random
1418 * increments have already pushed us past the projected offset, do nothing.
1421 tcp_isn_tick(void *xtp)
1423 u_int32_t projected_offset;
1426 projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100;
1428 if (projected_offset > isn_offset)
1429 isn_offset = projected_offset;
1431 isn_offset_old = isn_offset;
1432 callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL);
1437 * When a specific ICMP unreachable message is received and the
1438 * connection state is SYN-SENT, drop the connection. This behavior
1439 * is controlled by the icmp_may_rst sysctl.
1442 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1446 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1447 INP_LOCK_ASSERT(inp);
1449 if ((inp->inp_vflag & INP_TIMEWAIT) ||
1450 (inp->inp_vflag & INP_DROPPED))
1453 tp = intotcpcb(inp);
1454 if (tp->t_state != TCPS_SYN_SENT)
1457 tp = tcp_drop(tp, errno);
1465 * When `need fragmentation' ICMP is received, update our idea of the MSS
1466 * based on the new value in the route. Also nudge TCP to send something,
1467 * since we know the packet we just sent was dropped.
1468 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1471 tcp_mtudisc(struct inpcb *inp, int errno)
1474 struct socket *so = inp->inp_socket;
1482 INP_LOCK_ASSERT(inp);
1483 if ((inp->inp_vflag & INP_TIMEWAIT) ||
1484 (inp->inp_vflag & INP_DROPPED))
1487 tp = intotcpcb(inp);
1488 KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL"));
1491 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1493 maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */
1496 isipv6 ? tcp_maxmtu6(&inp->inp_inc, NULL) :
1498 tcp_maxmtu(&inp->inp_inc, NULL);
1502 maxmtu = min(maxmtu, romtu);
1504 tp->t_maxopd = tp->t_maxseg =
1506 isipv6 ? tcp_v6mssdflt :
1513 (isipv6 ? sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1515 sizeof(struct tcpiphdr)
1522 * XXX - The above conditional probably violates the TCP
1523 * spec. The problem is that, since we don't know the
1524 * other end's MSS, we are supposed to use a conservative
1525 * default. But, if we do that, then MTU discovery will
1526 * never actually take place, because the conservative
1527 * default is much less than the MTUs typically seen
1528 * on the Internet today. For the moment, we'll sweep
1529 * this under the carpet.
1531 * The conservative default might not actually be a problem
1532 * if the only case this occurs is when sending an initial
1533 * SYN with options and data to a host we've never talked
1534 * to before. Then, they will reply with an MSS value which
1535 * will get recorded and the new parameters should get
1536 * recomputed. For Further Study.
1538 if (tp->t_maxopd <= mss)
1542 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1543 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1544 mss -= TCPOLEN_TSTAMP_APPA;
1545 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1547 mss &= ~(MCLBYTES-1);
1550 mss = mss / MCLBYTES * MCLBYTES;
1552 if (so->so_snd.sb_hiwat < mss)
1553 mss = so->so_snd.sb_hiwat;
1557 tcpstat.tcps_mturesent++;
1559 tp->snd_nxt = tp->snd_una;
1560 tcp_free_sackholes(tp);
1561 tp->snd_recover = tp->snd_max;
1562 if (tp->t_flags & TF_SACK_PERMIT)
1563 EXIT_FASTRECOVERY(tp);
1569 * Look-up the routing entry to the peer of this inpcb. If no route
1570 * is found and it cannot be allocated, then return NULL. This routine
1571 * is called by TCP routines that access the rmx structure and by tcp_mss
1572 * to get the interface MTU.
1575 tcp_maxmtu(struct in_conninfo *inc, int *flags)
1578 struct sockaddr_in *dst;
1582 KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer"));
1584 bzero(&sro, sizeof(sro));
1585 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1586 dst = (struct sockaddr_in *)&sro.ro_dst;
1587 dst->sin_family = AF_INET;
1588 dst->sin_len = sizeof(*dst);
1589 dst->sin_addr = inc->inc_faddr;
1590 rtalloc_ign(&sro, RTF_CLONING);
1592 if (sro.ro_rt != NULL) {
1593 ifp = sro.ro_rt->rt_ifp;
1594 if (sro.ro_rt->rt_rmx.rmx_mtu == 0)
1595 maxmtu = ifp->if_mtu;
1597 maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu);
1599 /* Report additional interface capabilities. */
1600 if (flags != NULL) {
1601 if (ifp->if_capenable & IFCAP_TSO4 &&
1602 ifp->if_hwassist & CSUM_TSO)
1612 tcp_maxmtu6(struct in_conninfo *inc, int *flags)
1614 struct route_in6 sro6;
1618 KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer"));
1620 bzero(&sro6, sizeof(sro6));
1621 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1622 sro6.ro_dst.sin6_family = AF_INET6;
1623 sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1624 sro6.ro_dst.sin6_addr = inc->inc6_faddr;
1625 rtalloc_ign((struct route *)&sro6, RTF_CLONING);
1627 if (sro6.ro_rt != NULL) {
1628 ifp = sro6.ro_rt->rt_ifp;
1629 if (sro6.ro_rt->rt_rmx.rmx_mtu == 0)
1630 maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp);
1632 maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu,
1633 IN6_LINKMTU(sro6.ro_rt->rt_ifp));
1635 /* Report additional interface capabilities. */
1636 if (flags != NULL) {
1637 if (ifp->if_capenable & IFCAP_TSO6 &&
1638 ifp->if_hwassist & CSUM_TSO)
1649 /* compute ESP/AH header size for TCP, including outer IP header. */
1651 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1658 struct ip6_hdr *ip6;
1662 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1664 MGETHDR(m, M_DONTWAIT, MT_DATA);
1669 if ((inp->inp_vflag & INP_IPV6) != 0) {
1670 ip6 = mtod(m, struct ip6_hdr *);
1671 th = (struct tcphdr *)(ip6 + 1);
1672 m->m_pkthdr.len = m->m_len =
1673 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1674 tcpip_fillheaders(inp, ip6, th);
1675 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1679 ip = mtod(m, struct ip *);
1680 th = (struct tcphdr *)(ip + 1);
1681 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1682 tcpip_fillheaders(inp, ip, th);
1683 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1692 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1694 * This code attempts to calculate the bandwidth-delay product as a
1695 * means of determining the optimal window size to maximize bandwidth,
1696 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1697 * routers. This code also does a fairly good job keeping RTTs in check
1698 * across slow links like modems. We implement an algorithm which is very
1699 * similar (but not meant to be) TCP/Vegas. The code operates on the
1700 * transmitter side of a TCP connection and so only effects the transmit
1701 * side of the connection.
1703 * BACKGROUND: TCP makes no provision for the management of buffer space
1704 * at the end points or at the intermediate routers and switches. A TCP
1705 * stream, whether using NewReno or not, will eventually buffer as
1706 * many packets as it is able and the only reason this typically works is
1707 * due to the fairly small default buffers made available for a connection
1708 * (typicaly 16K or 32K). As machines use larger windows and/or window
1709 * scaling it is now fairly easy for even a single TCP connection to blow-out
1710 * all available buffer space not only on the local interface, but on
1711 * intermediate routers and switches as well. NewReno makes a misguided
1712 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1713 * then backing off, then steadily increasing the window again until another
1714 * failure occurs, ad-infinitum. This results in terrible oscillation that
1715 * is only made worse as network loads increase and the idea of intentionally
1716 * blowing out network buffers is, frankly, a terrible way to manage network
1719 * It is far better to limit the transmit window prior to the failure
1720 * condition being achieved. There are two general ways to do this: First
1721 * you can 'scan' through different transmit window sizes and locate the
1722 * point where the RTT stops increasing, indicating that you have filled the
1723 * pipe, then scan backwards until you note that RTT stops decreasing, then
1724 * repeat ad-infinitum. This method works in principle but has severe
1725 * implementation issues due to RTT variances, timer granularity, and
1726 * instability in the algorithm which can lead to many false positives and
1727 * create oscillations as well as interact badly with other TCP streams
1728 * implementing the same algorithm.
1730 * The second method is to limit the window to the bandwidth delay product
1731 * of the link. This is the method we implement. RTT variances and our
1732 * own manipulation of the congestion window, bwnd, can potentially
1733 * destabilize the algorithm. For this reason we have to stabilize the
1734 * elements used to calculate the window. We do this by using the minimum
1735 * observed RTT, the long term average of the observed bandwidth, and
1736 * by adding two segments worth of slop. It isn't perfect but it is able
1737 * to react to changing conditions and gives us a very stable basis on
1738 * which to extend the algorithm.
1741 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1747 INP_LOCK_ASSERT(tp->t_inpcb);
1750 * If inflight_enable is disabled in the middle of a tcp connection,
1751 * make sure snd_bwnd is effectively disabled.
1753 if (tcp_inflight_enable == 0 || tp->t_rttlow < tcp_inflight_rttthresh) {
1754 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1755 tp->snd_bandwidth = 0;
1760 * Figure out the bandwidth. Due to the tick granularity this
1761 * is a very rough number and it MUST be averaged over a fairly
1762 * long period of time. XXX we need to take into account a link
1763 * that is not using all available bandwidth, but for now our
1764 * slop will ramp us up if this case occurs and the bandwidth later
1767 * Note: if ticks rollover 'bw' may wind up negative. We must
1768 * effectively reset t_bw_rtttime for this case.
1771 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1774 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1775 (save_ticks - tp->t_bw_rtttime);
1776 tp->t_bw_rtttime = save_ticks;
1777 tp->t_bw_rtseq = ack_seq;
1778 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1780 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1782 tp->snd_bandwidth = bw;
1785 * Calculate the semi-static bandwidth delay product, plus two maximal
1786 * segments. The additional slop puts us squarely in the sweet
1787 * spot and also handles the bandwidth run-up case and stabilization.
1788 * Without the slop we could be locking ourselves into a lower
1791 * Situations Handled:
1792 * (1) Prevents over-queueing of packets on LANs, especially on
1793 * high speed LANs, allowing larger TCP buffers to be
1794 * specified, and also does a good job preventing
1795 * over-queueing of packets over choke points like modems
1796 * (at least for the transmit side).
1798 * (2) Is able to handle changing network loads (bandwidth
1799 * drops so bwnd drops, bandwidth increases so bwnd
1802 * (3) Theoretically should stabilize in the face of multiple
1803 * connections implementing the same algorithm (this may need
1806 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1807 * be adjusted with a sysctl but typically only needs to be
1808 * on very slow connections. A value no smaller then 5
1809 * should be used, but only reduce this default if you have
1812 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1813 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10;
1816 if (tcp_inflight_debug > 0) {
1818 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1820 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1829 if ((long)bwnd < tcp_inflight_min)
1830 bwnd = tcp_inflight_min;
1831 if (bwnd > tcp_inflight_max)
1832 bwnd = tcp_inflight_max;
1833 if ((long)bwnd < tp->t_maxseg * 2)
1834 bwnd = tp->t_maxseg * 2;
1835 tp->snd_bwnd = bwnd;
1838 #ifdef TCP_SIGNATURE
1840 * Callback function invoked by m_apply() to digest TCP segment data
1841 * contained within an mbuf chain.
1844 tcp_signature_apply(void *fstate, void *data, u_int len)
1847 MD5Update(fstate, (u_char *)data, len);
1852 * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385)
1855 * m pointer to head of mbuf chain
1856 * off0 offset to TCP header within the mbuf chain
1857 * len length of TCP segment data, excluding options
1858 * optlen length of TCP segment options
1859 * buf pointer to storage for computed MD5 digest
1860 * direction direction of flow (IPSEC_DIR_INBOUND or OUTBOUND)
1862 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1863 * When called from tcp_input(), we can be sure that th_sum has been
1864 * zeroed out and verified already.
1866 * This function is for IPv4 use only. Calling this function with an
1867 * IPv6 packet in the mbuf chain will yield undefined results.
1869 * Return 0 if successful, otherwise return -1.
1871 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1872 * search with the destination IP address, and a 'magic SPI' to be
1873 * determined by the application. This is hardcoded elsewhere to 1179
1874 * right now. Another branch of this code exists which uses the SPD to
1875 * specify per-application flows but it is unstable.
1878 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen,
1879 u_char *buf, u_int direction)
1881 union sockaddr_union dst;
1882 struct ippseudo ippseudo;
1886 struct ipovly *ipovly;
1887 struct secasvar *sav;
1891 KASSERT(m != NULL, ("NULL mbuf chain"));
1892 KASSERT(buf != NULL, ("NULL signature pointer"));
1894 /* Extract the destination from the IP header in the mbuf. */
1895 ip = mtod(m, struct ip *);
1896 bzero(&dst, sizeof(union sockaddr_union));
1897 dst.sa.sa_len = sizeof(struct sockaddr_in);
1898 dst.sa.sa_family = AF_INET;
1899 dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ?
1900 ip->ip_src : ip->ip_dst;
1902 /* Look up an SADB entry which matches the address of the peer. */
1903 sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI));
1905 printf("%s: SADB lookup failed for %s\n", __func__,
1906 inet_ntoa(dst.sin.sin_addr));
1911 ipovly = (struct ipovly *)ip;
1912 th = (struct tcphdr *)((u_char *)ip + off0);
1913 doff = off0 + sizeof(struct tcphdr) + optlen;
1916 * Step 1: Update MD5 hash with IP pseudo-header.
1918 * XXX The ippseudo header MUST be digested in network byte order,
1919 * or else we'll fail the regression test. Assume all fields we've
1920 * been doing arithmetic on have been in host byte order.
1921 * XXX One cannot depend on ipovly->ih_len here. When called from
1922 * tcp_output(), the underlying ip_len member has not yet been set.
1924 ippseudo.ippseudo_src = ipovly->ih_src;
1925 ippseudo.ippseudo_dst = ipovly->ih_dst;
1926 ippseudo.ippseudo_pad = 0;
1927 ippseudo.ippseudo_p = IPPROTO_TCP;
1928 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
1929 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
1932 * Step 2: Update MD5 hash with TCP header, excluding options.
1933 * The TCP checksum must be set to zero.
1935 savecsum = th->th_sum;
1937 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
1938 th->th_sum = savecsum;
1941 * Step 3: Update MD5 hash with TCP segment data.
1942 * Use m_apply() to avoid an early m_pullup().
1945 m_apply(m, doff, len, tcp_signature_apply, &ctx);
1948 * Step 4: Update MD5 hash with shared secret.
1950 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
1951 MD5Final(buf, &ctx);
1953 key_sa_recordxfer(sav, m);
1957 #endif /* TCP_SIGNATURE */
1960 sysctl_drop(SYSCTL_HANDLER_ARGS)
1962 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
1963 struct sockaddr_storage addrs[2];
1967 struct sockaddr_in *fin, *lin;
1969 struct sockaddr_in6 *fin6, *lin6;
1970 struct in6_addr f6, l6;
1981 if (req->oldptr != NULL || req->oldlen != 0)
1983 if (req->newptr == NULL)
1985 if (req->newlen < sizeof(addrs))
1987 error = SYSCTL_IN(req, &addrs, sizeof(addrs));
1991 switch (addrs[0].ss_family) {
1994 fin6 = (struct sockaddr_in6 *)&addrs[0];
1995 lin6 = (struct sockaddr_in6 *)&addrs[1];
1996 if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
1997 lin6->sin6_len != sizeof(struct sockaddr_in6))
1999 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) {
2000 if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2002 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]);
2003 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]);
2004 fin = (struct sockaddr_in *)&addrs[0];
2005 lin = (struct sockaddr_in *)&addrs[1];
2008 error = sa6_embedscope(fin6, ip6_use_defzone);
2011 error = sa6_embedscope(lin6, ip6_use_defzone);
2017 fin = (struct sockaddr_in *)&addrs[0];
2018 lin = (struct sockaddr_in *)&addrs[1];
2019 if (fin->sin_len != sizeof(struct sockaddr_in) ||
2020 lin->sin_len != sizeof(struct sockaddr_in))
2026 INP_INFO_WLOCK(&tcbinfo);
2027 switch (addrs[0].ss_family) {
2030 inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port,
2031 &l6, lin6->sin6_port, 0, NULL);
2035 inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port,
2036 lin->sin_addr, lin->sin_port, 0, NULL);
2041 if (inp->inp_vflag & INP_TIMEWAIT) {
2043 * XXXRW: There currently exists a state where an
2044 * inpcb is present, but its timewait state has been
2045 * discarded. For now, don't allow dropping of this
2051 } else if (!(inp->inp_vflag & INP_DROPPED) &&
2052 !(inp->inp_socket->so_options & SO_ACCEPTCONN)) {
2053 tp = intotcpcb(inp);
2054 tcp_drop(tp, ECONNABORTED);
2059 INP_INFO_WUNLOCK(&tcbinfo);
2063 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop,
2064 CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL,
2065 0, sysctl_drop, "", "Drop TCP connection");
2068 * Generate a standardized TCP log line for use throughout the
2069 * tcp subsystem. Memory allocation is done with M_NOWAIT to
2070 * allow use in the interrupt context.
2072 * NB: The caller MUST free(s, M_TCPLOG) the returned string.
2073 * NB: The function may return NULL if memory allocation failed.
2075 * Due to header inclusion and ordering limitations the struct ip
2076 * and ip6_hdr pointers have to be passed as void pointers.
2079 tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr,
2086 struct ip6_hdr *ip6;
2088 ip6 = (struct ip6_hdr *)ip6hdr;
2090 ip = (struct ip *)ip4hdr;
2093 * The log line looks like this:
2094 * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2<SYN>"
2096 size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") +
2097 sizeof(PRINT_TH_FLAGS) + 1 +
2099 2 * INET6_ADDRSTRLEN;
2101 2 * INET_ADDRSTRLEN;
2104 s = sp = malloc(size, M_TCPLOG, (M_ZERO|M_NOWAIT));
2108 strcat(s, "TCP: [");
2111 if (inc && inc->inc_isipv6 == 0) {
2112 inet_ntoa_r(inc->inc_faddr, sp);
2114 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2116 inet_ntoa_r(inc->inc_laddr, sp);
2118 sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2121 ip6_sprintf(sp, &inc->inc6_faddr);
2123 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2125 ip6_sprintf(sp, &inc->inc6_laddr);
2127 sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2128 } else if (ip6 && th) {
2129 ip6_sprintf(sp, &ip6->ip6_src);
2131 sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2133 ip6_sprintf(sp, &ip6->ip6_dst);
2135 sprintf(sp, "]:%i", ntohs(th->th_dport));
2137 } else if (ip && th) {
2138 inet_ntoa_r(ip->ip_src, sp);
2140 sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2142 inet_ntoa_r(ip->ip_dst, sp);
2144 sprintf(sp, "]:%i", ntohs(th->th_dport));
2151 sprintf(sp, " tcpflags 0x%b", th->th_flags, PRINT_TH_FLAGS);
2152 if (*(s + size - 1) != '\0')
2153 panic("%s: string too long", __func__);