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
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include "opt_compat.h"
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/callout.h>
45 #include <sys/kernel.h>
46 #include <sys/sysctl.h>
47 #include <sys/malloc.h>
50 #include <sys/domain.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/protosw.h>
57 #include <sys/random.h>
61 #include <net/route.h>
64 #include <netinet/in.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/ip.h>
68 #include <netinet/ip6.h>
70 #include <netinet/in_pcb.h>
72 #include <netinet6/in6_pcb.h>
74 #include <netinet/in_var.h>
75 #include <netinet/ip_var.h>
77 #include <netinet6/ip6_var.h>
78 #include <netinet6/scope6_var.h>
79 #include <netinet6/nd6.h>
81 #include <netinet/ip_icmp.h>
82 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_seq.h>
85 #include <netinet/tcp_timer.h>
86 #include <netinet/tcp_var.h>
87 #include <netinet/tcp_syncache.h>
88 #include <netinet/tcp_offload.h>
90 #include <netinet6/tcp6_var.h>
92 #include <netinet/tcpip.h>
94 #include <netinet/tcp_debug.h>
96 #include <netinet6/ip6protosw.h>
99 #include <netipsec/ipsec.h>
100 #include <netipsec/xform.h>
102 #include <netipsec/ipsec6.h>
104 #include <netipsec/key.h>
107 #include <machine/in_cksum.h>
110 #include <security/mac/mac_framework.h>
112 int tcp_mssdflt = TCP_MSS;
113 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
114 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
117 int tcp_v6mssdflt = TCP6_MSS;
118 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt,
119 CTLFLAG_RW, &tcp_v6mssdflt , 0,
120 "Default TCP Maximum Segment Size for IPv6");
124 * Minimum MSS we accept and use. This prevents DoS attacks where
125 * we are forced to a ridiculous low MSS like 20 and send hundreds
126 * of packets instead of one. The effect scales with the available
127 * bandwidth and quickly saturates the CPU and network interface
128 * with packet generation and sending. Set to zero to disable MINMSS
129 * checking. This setting prevents us from sending too small packets.
131 int tcp_minmss = TCP_MINMSS;
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
133 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
135 int tcp_do_rfc1323 = 1;
136 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
137 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
139 static int tcp_log_debug = 0;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_debug, CTLFLAG_RW,
141 &tcp_log_debug, 0, "Log errors caused by incoming TCP segments");
143 static int tcp_tcbhashsize = 0;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN,
145 &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable");
147 static int do_tcpdrain = 1;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW,
150 "Enable tcp_drain routine for extra help when low on mbufs");
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
153 &tcbinfo.ipi_count, 0, "Number of active PCBs");
155 static int icmp_may_rst = 1;
156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW,
158 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
160 static int tcp_isn_reseed_interval = 0;
161 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
162 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
165 * TCP bandwidth limiting sysctls. Note that the default lower bound of
166 * 1024 exists only for debugging. A good production default would be
167 * something like 6100.
169 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, inflight, CTLFLAG_RW, 0,
170 "TCP inflight data limiting");
172 static int tcp_inflight_enable = 1;
173 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, enable, CTLFLAG_RW,
174 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
176 static int tcp_inflight_debug = 0;
177 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, debug, CTLFLAG_RW,
178 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
180 static int tcp_inflight_rttthresh;
181 SYSCTL_PROC(_net_inet_tcp_inflight, OID_AUTO, rttthresh, CTLTYPE_INT|CTLFLAG_RW,
182 &tcp_inflight_rttthresh, 0, sysctl_msec_to_ticks, "I",
183 "RTT threshold below which inflight will deactivate itself");
185 static int tcp_inflight_min = 6144;
186 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, min, CTLFLAG_RW,
187 &tcp_inflight_min, 0, "Lower-bound for TCP inflight window");
189 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
190 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, max, CTLFLAG_RW,
191 &tcp_inflight_max, 0, "Upper-bound for TCP inflight window");
193 static int tcp_inflight_stab = 20;
194 SYSCTL_INT(_net_inet_tcp_inflight, OID_AUTO, stab, CTLFLAG_RW,
195 &tcp_inflight_stab, 0, "Inflight Algorithm Stabilization 20 = 2 packets");
197 uma_zone_t sack_hole_zone;
199 static struct inpcb *tcp_notify(struct inpcb *, int);
200 static void tcp_isn_tick(void *);
203 * Target size of TCP PCB hash tables. Must be a power of two.
205 * Note that this can be overridden by the kernel environment
206 * variable net.inet.tcp.tcbhashsize
209 #define TCBHASHSIZE 512
214 * Callouts should be moved into struct tcp directly. They are currently
215 * separate because the tcpcb structure is exported to userland for sysctl
216 * parsing purposes, which do not know about callouts.
223 static uma_zone_t tcpcb_zone;
224 MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers");
225 struct callout isn_callout;
226 static struct mtx isn_mtx;
228 #define ISN_LOCK_INIT() mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF)
229 #define ISN_LOCK() mtx_lock(&isn_mtx)
230 #define ISN_UNLOCK() mtx_unlock(&isn_mtx)
233 * TCP initialization.
236 tcp_zone_change(void *tag)
239 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
240 uma_zone_set_max(tcpcb_zone, maxsockets);
241 tcp_tw_zone_change();
245 tcp_inpcb_init(void *mem, int size, int flags)
247 struct inpcb *inp = mem;
249 INP_LOCK_INIT(inp, "inp", "tcpinp");
257 int hashsize = TCBHASHSIZE;
258 tcp_delacktime = TCPTV_DELACK;
259 tcp_keepinit = TCPTV_KEEP_INIT;
260 tcp_keepidle = TCPTV_KEEP_IDLE;
261 tcp_keepintvl = TCPTV_KEEPINTVL;
262 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
264 tcp_rexmit_min = TCPTV_MIN;
265 if (tcp_rexmit_min < 1)
267 tcp_rexmit_slop = TCPTV_CPU_VAR;
268 tcp_inflight_rttthresh = TCPTV_INFLIGHT_RTTTHRESH;
269 tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT;
271 INP_INFO_LOCK_INIT(&tcbinfo, "tcp");
273 tcbinfo.ipi_listhead = &tcb;
274 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
275 if (!powerof2(hashsize)) {
276 printf("WARNING: TCB hash size not a power of 2\n");
277 hashsize = 512; /* safe default */
279 tcp_tcbhashsize = hashsize;
280 tcbinfo.ipi_hashbase = hashinit(hashsize, M_PCB,
281 &tcbinfo.ipi_hashmask);
282 tcbinfo.ipi_porthashbase = hashinit(hashsize, M_PCB,
283 &tcbinfo.ipi_porthashmask);
284 tcbinfo.ipi_zone = uma_zcreate("inpcb", sizeof(struct inpcb),
285 NULL, NULL, tcp_inpcb_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
286 uma_zone_set_max(tcbinfo.ipi_zone, maxsockets);
288 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
290 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
292 if (max_protohdr < TCP_MINPROTOHDR)
293 max_protohdr = TCP_MINPROTOHDR;
294 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
296 #undef TCP_MINPROTOHDR
298 * These have to be type stable for the benefit of the timers.
300 tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem),
301 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
302 uma_zone_set_max(tcpcb_zone, maxsockets);
308 callout_init(&isn_callout, CALLOUT_MPSAFE);
310 EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL,
311 SHUTDOWN_PRI_DEFAULT);
312 sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole),
313 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
314 EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL,
315 EVENTHANDLER_PRI_ANY);
322 callout_stop(&isn_callout);
326 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
327 * tcp_template used to store this data in mbufs, but we now recopy it out
328 * of the tcpcb each time to conserve mbufs.
331 tcpip_fillheaders(struct inpcb *inp, void *ip_ptr, void *tcp_ptr)
333 struct tcphdr *th = (struct tcphdr *)tcp_ptr;
335 INP_LOCK_ASSERT(inp);
338 if ((inp->inp_vflag & INP_IPV6) != 0) {
341 ip6 = (struct ip6_hdr *)ip_ptr;
342 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
343 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
344 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
345 (IPV6_VERSION & IPV6_VERSION_MASK);
346 ip6->ip6_nxt = IPPROTO_TCP;
347 ip6->ip6_plen = sizeof(struct tcphdr);
348 ip6->ip6_src = inp->in6p_laddr;
349 ip6->ip6_dst = inp->in6p_faddr;
355 ip = (struct ip *)ip_ptr;
356 ip->ip_v = IPVERSION;
358 ip->ip_tos = inp->inp_ip_tos;
362 ip->ip_ttl = inp->inp_ip_ttl;
364 ip->ip_p = IPPROTO_TCP;
365 ip->ip_src = inp->inp_laddr;
366 ip->ip_dst = inp->inp_faddr;
368 th->th_sport = inp->inp_lport;
369 th->th_dport = inp->inp_fport;
377 th->th_sum = 0; /* in_pseudo() is called later for ipv4 */
381 * Create template to be used to send tcp packets on a connection.
382 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
383 * use for this function is in keepalives, which use tcp_respond.
386 tcpip_maketemplate(struct inpcb *inp)
391 m = m_get(M_DONTWAIT, MT_DATA);
394 m->m_len = sizeof(struct tcptemp);
395 n = mtod(m, struct tcptemp *);
397 tcpip_fillheaders(inp, (void *)&n->tt_ipgen, (void *)&n->tt_t);
402 * Send a single message to the TCP at address specified by
403 * the given TCP/IP header. If m == NULL, then we make a copy
404 * of the tcpiphdr at ti and send directly to the addressed host.
405 * This is used to force keep alive messages out using the TCP
406 * template for a connection. If flags are given then we send
407 * a message back to the TCP which originated the * segment ti,
408 * and discard the mbuf containing it and any other attached mbufs.
410 * In any case the ack and sequence number of the transmitted
411 * segment are as specified by the parameters.
413 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
416 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
417 tcp_seq ack, tcp_seq seq, int flags)
430 KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL"));
433 isipv6 = ((struct ip *)ipgen)->ip_v == 6;
440 KASSERT(inp != NULL, ("tcp control block w/o inpcb"));
441 INP_LOCK_ASSERT(inp);
446 if (!(flags & TH_RST)) {
447 win = sbspace(&inp->inp_socket->so_rcv);
448 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
449 win = (long)TCP_MAXWIN << tp->rcv_scale;
453 m = m_gethdr(M_DONTWAIT, MT_DATA);
457 m->m_data += max_linkhdr;
460 bcopy((caddr_t)ip6, mtod(m, caddr_t),
461 sizeof(struct ip6_hdr));
462 ip6 = mtod(m, struct ip6_hdr *);
463 nth = (struct tcphdr *)(ip6 + 1);
467 bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
468 ip = mtod(m, struct ip *);
469 nth = (struct tcphdr *)(ip + 1);
471 bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
476 m->m_data = (caddr_t)ipgen;
477 /* m_len is set later */
479 #define xchg(a,b,type) { type t; t=a; a=b; b=t; }
482 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
483 nth = (struct tcphdr *)(ip6 + 1);
487 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
488 nth = (struct tcphdr *)(ip + 1);
492 * this is usually a case when an extension header
493 * exists between the IPv6 header and the
496 nth->th_sport = th->th_sport;
497 nth->th_dport = th->th_dport;
499 xchg(nth->th_dport, nth->th_sport, n_short);
505 ip6->ip6_vfc = IPV6_VERSION;
506 ip6->ip6_nxt = IPPROTO_TCP;
507 ip6->ip6_plen = htons((u_short)(sizeof (struct tcphdr) +
509 tlen += sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
513 tlen += sizeof (struct tcpiphdr);
515 ip->ip_ttl = ip_defttl;
516 if (path_mtu_discovery)
520 m->m_pkthdr.len = tlen;
521 m->m_pkthdr.rcvif = NULL;
525 * Packet is associated with a socket, so allow the
526 * label of the response to reflect the socket label.
528 INP_LOCK_ASSERT(inp);
529 mac_inpcb_create_mbuf(inp, m);
532 * Packet is not associated with a socket, so possibly
533 * update the label in place.
535 mac_netinet_tcp_reply(m);
538 nth->th_seq = htonl(seq);
539 nth->th_ack = htonl(ack);
541 nth->th_off = sizeof (struct tcphdr) >> 2;
542 nth->th_flags = flags;
544 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
546 nth->th_win = htons((u_short)win);
551 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
552 sizeof(struct ip6_hdr),
553 tlen - sizeof(struct ip6_hdr));
554 ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb :
559 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
560 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
561 m->m_pkthdr.csum_flags = CSUM_TCP;
562 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
565 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
566 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
570 (void) ip6_output(m, NULL, NULL, ipflags, NULL, NULL, inp);
573 (void) ip_output(m, NULL, NULL, ipflags, NULL, inp);
577 * Create a new TCP control block, making an
578 * empty reassembly queue and hooking it to the argument
579 * protocol control block. The `inp' parameter must have
580 * come from the zone allocator set up in tcp_init().
583 tcp_newtcpcb(struct inpcb *inp)
585 struct tcpcb_mem *tm;
588 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
591 tm = uma_zalloc(tcpcb_zone, M_NOWAIT | M_ZERO);
595 tp->t_timers = &tm->tt;
596 /* LIST_INIT(&tp->t_segq); */ /* XXX covered by M_ZERO */
597 tp->t_maxseg = tp->t_maxopd =
599 isipv6 ? tcp_v6mssdflt :
603 /* Set up our timeouts. */
604 callout_init(&tp->t_timers->tt_rexmt, CALLOUT_MPSAFE);
605 callout_init(&tp->t_timers->tt_persist, CALLOUT_MPSAFE);
606 callout_init(&tp->t_timers->tt_keep, CALLOUT_MPSAFE);
607 callout_init(&tp->t_timers->tt_2msl, CALLOUT_MPSAFE);
608 callout_init(&tp->t_timers->tt_delack, CALLOUT_MPSAFE);
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_reset(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 callout_stop(&tp->t_timers->tt_rexmt);
682 callout_stop(&tp->t_timers->tt_persist);
683 callout_stop(&tp->t_timers->tt_keep);
684 callout_stop(&tp->t_timers->tt_2msl);
685 callout_stop(&tp->t_timers->tt_delack);
688 * If we got enough samples through the srtt filter,
689 * save the rtt and rttvar in the routing entry.
690 * 'Enough' is arbitrarily defined as 4 rtt samples.
691 * 4 samples is enough for the srtt filter to converge
692 * to within enough % of the correct value; fewer samples
693 * and we could save a bogus rtt. The danger is not high
694 * as tcp quickly recovers from everything.
695 * XXX: Works very well but needs some more statistics!
697 if (tp->t_rttupdated >= 4) {
698 struct hc_metrics_lite metrics;
701 bzero(&metrics, sizeof(metrics));
703 * Update the ssthresh always when the conditions below
704 * are satisfied. This gives us better new start value
705 * for the congestion avoidance for new connections.
706 * ssthresh is only set if packet loss occured on a session.
708 * XXXRW: 'so' may be NULL here, and/or socket buffer may be
709 * being torn down. Ideally this code would not use 'so'.
711 ssthresh = tp->snd_ssthresh;
712 if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) {
714 * convert the limit from user data bytes to
715 * packets then to packet data bytes.
717 ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg;
720 ssthresh *= (u_long)(tp->t_maxseg +
722 (isipv6 ? sizeof (struct ip6_hdr) +
723 sizeof (struct tcphdr) :
725 sizeof (struct tcpiphdr)
732 metrics.rmx_ssthresh = ssthresh;
734 metrics.rmx_rtt = tp->t_srtt;
735 metrics.rmx_rttvar = tp->t_rttvar;
736 /* XXX: This wraps if the pipe is more than 4 Gbit per second */
737 metrics.rmx_bandwidth = tp->snd_bandwidth;
738 metrics.rmx_cwnd = tp->snd_cwnd;
739 metrics.rmx_sendpipe = 0;
740 metrics.rmx_recvpipe = 0;
742 tcp_hc_update(&inp->inp_inc, &metrics);
745 /* free the reassembly queue, if any */
746 while ((q = LIST_FIRST(&tp->t_segq)) != NULL) {
747 LIST_REMOVE(q, tqe_q);
749 uma_zfree(tcp_reass_zone, q);
753 /* Disconnect offload device, if any. */
754 tcp_offload_detach(tp);
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);
775 /* Notify any offload devices of listener close */
776 if (tp->t_state == TCPS_LISTEN)
777 tcp_offload_listen_close(tp);
779 tcpstat.tcps_closed++;
780 KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL"));
781 so = inp->inp_socket;
782 soisdisconnected(so);
783 if (inp->inp_vflag & INP_SOCKREF) {
784 KASSERT(so->so_state & SS_PROTOREF,
785 ("tcp_close: !SS_PROTOREF"));
786 inp->inp_vflag &= ~INP_SOCKREF;
790 so->so_state &= ~SS_PROTOREF;
804 struct tseg_qent *te;
807 * Walk the tcpbs, if existing, and flush the reassembly queue,
809 * XXX: The "Net/3" implementation doesn't imply that the TCP
810 * reassembly queue should be flushed, but in a situation
811 * where we're really low on mbufs, this is potentially
814 INP_INFO_RLOCK(&tcbinfo);
815 LIST_FOREACH(inpb, tcbinfo.ipi_listhead, inp_list) {
816 if (inpb->inp_vflag & INP_TIMEWAIT)
819 if ((tcpb = intotcpcb(inpb)) != NULL) {
820 while ((te = LIST_FIRST(&tcpb->t_segq))
822 LIST_REMOVE(te, tqe_q);
824 uma_zfree(tcp_reass_zone, te);
828 tcp_clean_sackreport(tcpb);
832 INP_INFO_RUNLOCK(&tcbinfo);
837 * Notify a tcp user of an asynchronous error;
838 * store error as soft error, but wake up user
839 * (for now, won't do anything until can select for soft error).
841 * Do not wake up user since there currently is no mechanism for
842 * reporting soft errors (yet - a kqueue filter may be added).
844 static struct inpcb *
845 tcp_notify(struct inpcb *inp, int error)
849 INP_INFO_WLOCK_ASSERT(&tcbinfo);
850 INP_LOCK_ASSERT(inp);
852 if ((inp->inp_vflag & INP_TIMEWAIT) ||
853 (inp->inp_vflag & INP_DROPPED))
857 KASSERT(tp != NULL, ("tcp_notify: tp == NULL"));
860 * Ignore some errors if we are hooked up.
861 * If connection hasn't completed, has retransmitted several times,
862 * and receives a second error, give up now. This is better
863 * than waiting a long time to establish a connection that
864 * can never complete.
866 if (tp->t_state == TCPS_ESTABLISHED &&
867 (error == EHOSTUNREACH || error == ENETUNREACH ||
868 error == EHOSTDOWN)) {
870 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
872 tp = tcp_drop(tp, error);
878 tp->t_softerror = error;
882 wakeup( &so->so_timeo);
889 tcp_pcblist(SYSCTL_HANDLER_ARGS)
891 int error, i, m, n, pcb_count;
892 struct inpcb *inp, **inp_list;
897 * The process of preparing the TCB list is too time-consuming and
898 * resource-intensive to repeat twice on every request.
900 if (req->oldptr == NULL) {
901 m = syncache_pcbcount();
902 n = tcbinfo.ipi_count;
903 req->oldidx = 2 * (sizeof xig)
904 + ((m + n) + n/8) * sizeof(struct xtcpcb);
908 if (req->newptr != NULL)
912 * OK, now we're committed to doing something.
914 INP_INFO_RLOCK(&tcbinfo);
915 gencnt = tcbinfo.ipi_gencnt;
916 n = tcbinfo.ipi_count;
917 INP_INFO_RUNLOCK(&tcbinfo);
919 m = syncache_pcbcount();
921 error = sysctl_wire_old_buffer(req, 2 * (sizeof xig)
922 + (n + m) * sizeof(struct xtcpcb));
926 xig.xig_len = sizeof xig;
927 xig.xig_count = n + m;
928 xig.xig_gen = gencnt;
929 xig.xig_sogen = so_gencnt;
930 error = SYSCTL_OUT(req, &xig, sizeof xig);
934 error = syncache_pcblist(req, m, &pcb_count);
938 inp_list = malloc(n * sizeof *inp_list, M_TEMP, M_WAITOK);
939 if (inp_list == NULL)
942 INP_INFO_RLOCK(&tcbinfo);
943 for (inp = LIST_FIRST(tcbinfo.ipi_listhead), i = 0; inp != NULL && i
944 < n; inp = LIST_NEXT(inp, inp_list)) {
946 if (inp->inp_gencnt <= gencnt) {
948 * XXX: This use of cr_cansee(), introduced with
949 * TCP state changes, is not quite right, but for
950 * now, better than nothing.
952 if (inp->inp_vflag & INP_TIMEWAIT) {
953 if (intotw(inp) != NULL)
954 error = cr_cansee(req->td->td_ucred,
955 intotw(inp)->tw_cred);
957 error = EINVAL; /* Skip this inp. */
959 error = cr_canseesocket(req->td->td_ucred,
966 INP_INFO_RUNLOCK(&tcbinfo);
970 for (i = 0; i < n; i++) {
973 if (inp->inp_gencnt <= gencnt) {
977 bzero(&xt, sizeof(xt));
978 xt.xt_len = sizeof xt;
979 /* XXX should avoid extra copy */
980 bcopy(inp, &xt.xt_inp, sizeof *inp);
981 inp_ppcb = inp->inp_ppcb;
982 if (inp_ppcb == NULL)
983 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
984 else if (inp->inp_vflag & INP_TIMEWAIT) {
985 bzero((char *) &xt.xt_tp, sizeof xt.xt_tp);
986 xt.xt_tp.t_state = TCPS_TIME_WAIT;
988 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
989 if (inp->inp_socket != NULL)
990 sotoxsocket(inp->inp_socket, &xt.xt_socket);
992 bzero(&xt.xt_socket, sizeof xt.xt_socket);
993 xt.xt_socket.xso_protocol = IPPROTO_TCP;
995 xt.xt_inp.inp_gencnt = inp->inp_gencnt;
997 error = SYSCTL_OUT(req, &xt, sizeof xt);
1004 * Give the user an updated idea of our state.
1005 * If the generation differs from what we told
1006 * her before, she knows that something happened
1007 * while we were processing this request, and it
1008 * might be necessary to retry.
1010 INP_INFO_RLOCK(&tcbinfo);
1011 xig.xig_gen = tcbinfo.ipi_gencnt;
1012 xig.xig_sogen = so_gencnt;
1013 xig.xig_count = tcbinfo.ipi_count + pcb_count;
1014 INP_INFO_RUNLOCK(&tcbinfo);
1015 error = SYSCTL_OUT(req, &xig, sizeof xig);
1017 free(inp_list, M_TEMP);
1021 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1022 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1025 tcp_getcred(SYSCTL_HANDLER_ARGS)
1028 struct sockaddr_in addrs[2];
1032 error = priv_check(req->td, PRIV_NETINET_GETCRED);
1035 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1038 INP_INFO_RLOCK(&tcbinfo);
1039 inp = in_pcblookup_hash(&tcbinfo, addrs[1].sin_addr, addrs[1].sin_port,
1040 addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1046 if (inp->inp_socket == NULL) {
1050 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1053 cru2x(inp->inp_socket->so_cred, &xuc);
1057 INP_INFO_RUNLOCK(&tcbinfo);
1059 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1063 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred,
1064 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1065 tcp_getcred, "S,xucred", "Get the xucred of a TCP connection");
1069 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1072 struct sockaddr_in6 addrs[2];
1074 int error, mapped = 0;
1076 error = priv_check(req->td, PRIV_NETINET_GETCRED);
1079 error = SYSCTL_IN(req, addrs, sizeof(addrs));
1082 if ((error = sa6_embedscope(&addrs[0], ip6_use_defzone)) != 0 ||
1083 (error = sa6_embedscope(&addrs[1], ip6_use_defzone)) != 0) {
1086 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1087 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1093 INP_INFO_RLOCK(&tcbinfo);
1095 inp = in_pcblookup_hash(&tcbinfo,
1096 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1098 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1102 inp = in6_pcblookup_hash(&tcbinfo,
1103 &addrs[1].sin6_addr, addrs[1].sin6_port,
1104 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1110 if (inp->inp_socket == NULL) {
1114 error = cr_canseesocket(req->td->td_ucred, inp->inp_socket);
1117 cru2x(inp->inp_socket->so_cred, &xuc);
1121 INP_INFO_RUNLOCK(&tcbinfo);
1123 error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred));
1127 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred,
1128 CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0,
1129 tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection");
1134 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1136 struct ip *ip = vip;
1138 struct in_addr faddr;
1141 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1143 struct in_conninfo inc;
1144 tcp_seq icmp_tcp_seq;
1147 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1148 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1151 if (cmd == PRC_MSGSIZE)
1152 notify = tcp_mtudisc;
1153 else if (icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1154 cmd == PRC_UNREACH_PORT || cmd == PRC_TIMXCEED_INTRANS) && ip)
1155 notify = tcp_drop_syn_sent;
1157 * Redirects don't need to be handled up here.
1159 else if (PRC_IS_REDIRECT(cmd))
1162 * Source quench is depreciated.
1164 else if (cmd == PRC_QUENCH)
1167 * Hostdead is ugly because it goes linearly through all PCBs.
1168 * XXX: We never get this from ICMP, otherwise it makes an
1169 * excellent DoS attack on machines with many connections.
1171 else if (cmd == PRC_HOSTDEAD)
1173 else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0)
1176 icp = (struct icmp *)((caddr_t)ip
1177 - offsetof(struct icmp, icmp_ip));
1178 th = (struct tcphdr *)((caddr_t)ip
1179 + (ip->ip_hl << 2));
1180 INP_INFO_WLOCK(&tcbinfo);
1181 inp = in_pcblookup_hash(&tcbinfo, faddr, th->th_dport,
1182 ip->ip_src, th->th_sport, 0, NULL);
1185 if (!(inp->inp_vflag & INP_TIMEWAIT) &&
1186 !(inp->inp_vflag & INP_DROPPED) &&
1187 !(inp->inp_socket == NULL)) {
1188 icmp_tcp_seq = htonl(th->th_seq);
1189 tp = intotcpcb(inp);
1190 if (SEQ_GEQ(icmp_tcp_seq, tp->snd_una) &&
1191 SEQ_LT(icmp_tcp_seq, tp->snd_max)) {
1192 if (cmd == PRC_MSGSIZE) {
1195 * If we got a needfrag set the MTU
1196 * in the route to the suggested new
1197 * value (if given) and then notify.
1199 bzero(&inc, sizeof(inc));
1200 inc.inc_flags = 0; /* IPv4 */
1201 inc.inc_faddr = faddr;
1203 mtu = ntohs(icp->icmp_nextmtu);
1205 * If no alternative MTU was
1206 * proposed, try the next smaller
1207 * one. ip->ip_len has already
1208 * been swapped in icmp_input().
1211 mtu = ip_next_mtu(ip->ip_len,
1213 if (mtu < max(296, (tcp_minmss)
1214 + sizeof(struct tcpiphdr)))
1218 + sizeof(struct tcpiphdr);
1220 * Only cache the the MTU if it
1221 * is smaller than the interface
1222 * or route MTU. tcp_mtudisc()
1223 * will do right thing by itself.
1225 if (mtu <= tcp_maxmtu(&inc, NULL))
1226 tcp_hc_updatemtu(&inc, mtu);
1229 inp = (*notify)(inp, inetctlerrmap[cmd]);
1235 inc.inc_fport = th->th_dport;
1236 inc.inc_lport = th->th_sport;
1237 inc.inc_faddr = faddr;
1238 inc.inc_laddr = ip->ip_src;
1242 syncache_unreach(&inc, th);
1244 INP_INFO_WUNLOCK(&tcbinfo);
1246 in_pcbnotifyall(&tcbinfo, faddr, inetctlerrmap[cmd], notify);
1251 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1254 struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify;
1255 struct ip6_hdr *ip6;
1257 struct ip6ctlparam *ip6cp = NULL;
1258 const struct sockaddr_in6 *sa6_src = NULL;
1260 struct tcp_portonly {
1265 if (sa->sa_family != AF_INET6 ||
1266 sa->sa_len != sizeof(struct sockaddr_in6))
1269 if (cmd == PRC_MSGSIZE)
1270 notify = tcp_mtudisc;
1271 else if (!PRC_IS_REDIRECT(cmd) &&
1272 ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1274 /* Source quench is depreciated. */
1275 else if (cmd == PRC_QUENCH)
1278 /* if the parameter is from icmp6, decode it. */
1280 ip6cp = (struct ip6ctlparam *)d;
1282 ip6 = ip6cp->ip6c_ip6;
1283 off = ip6cp->ip6c_off;
1284 sa6_src = ip6cp->ip6c_src;
1288 off = 0; /* fool gcc */
1293 struct in_conninfo inc;
1295 * XXX: We assume that when IPV6 is non NULL,
1296 * M and OFF are valid.
1299 /* check if we can safely examine src and dst ports */
1300 if (m->m_pkthdr.len < off + sizeof(*thp))
1303 bzero(&th, sizeof(th));
1304 m_copydata(m, off, sizeof(*thp), (caddr_t)&th);
1306 in6_pcbnotify(&tcbinfo, sa, th.th_dport,
1307 (struct sockaddr *)ip6cp->ip6c_src,
1308 th.th_sport, cmd, NULL, notify);
1310 inc.inc_fport = th.th_dport;
1311 inc.inc_lport = th.th_sport;
1312 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1313 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1315 INP_INFO_WLOCK(&tcbinfo);
1316 syncache_unreach(&inc, &th);
1317 INP_INFO_WUNLOCK(&tcbinfo);
1319 in6_pcbnotify(&tcbinfo, sa, 0, (const struct sockaddr *)sa6_src,
1320 0, cmd, NULL, notify);
1326 * Following is where TCP initial sequence number generation occurs.
1328 * There are two places where we must use initial sequence numbers:
1329 * 1. In SYN-ACK packets.
1330 * 2. In SYN packets.
1332 * All ISNs for SYN-ACK packets are generated by the syncache. See
1333 * tcp_syncache.c for details.
1335 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1336 * depends on this property. In addition, these ISNs should be
1337 * unguessable so as to prevent connection hijacking. To satisfy
1338 * the requirements of this situation, the algorithm outlined in
1339 * RFC 1948 is used, with only small modifications.
1341 * Implementation details:
1343 * Time is based off the system timer, and is corrected so that it
1344 * increases by one megabyte per second. This allows for proper
1345 * recycling on high speed LANs while still leaving over an hour
1348 * As reading the *exact* system time is too expensive to be done
1349 * whenever setting up a TCP connection, we increment the time
1350 * offset in two ways. First, a small random positive increment
1351 * is added to isn_offset for each connection that is set up.
1352 * Second, the function tcp_isn_tick fires once per clock tick
1353 * and increments isn_offset as necessary so that sequence numbers
1354 * are incremented at approximately ISN_BYTES_PER_SECOND. The
1355 * random positive increments serve only to ensure that the same
1356 * exact sequence number is never sent out twice (as could otherwise
1357 * happen when a port is recycled in less than the system tick
1360 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1361 * between seeding of isn_secret. This is normally set to zero,
1362 * as reseeding should not be necessary.
1364 * Locking of the global variables isn_secret, isn_last_reseed, isn_offset,
1365 * isn_offset_old, and isn_ctx is performed using the TCP pcbinfo lock. In
1366 * general, this means holding an exclusive (write) lock.
1369 #define ISN_BYTES_PER_SECOND 1048576
1370 #define ISN_STATIC_INCREMENT 4096
1371 #define ISN_RANDOM_INCREMENT (4096 - 1)
1373 static u_char isn_secret[32];
1374 static int isn_last_reseed;
1375 static u_int32_t isn_offset, isn_offset_old;
1376 static MD5_CTX isn_ctx;
1379 tcp_new_isn(struct tcpcb *tp)
1381 u_int32_t md5_buffer[4];
1384 INP_LOCK_ASSERT(tp->t_inpcb);
1387 /* Seed if this is the first use, reseed if requested. */
1388 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1389 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1391 read_random(&isn_secret, sizeof(isn_secret));
1392 isn_last_reseed = ticks;
1395 /* Compute the md5 hash and return the ISN. */
1397 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_fport, sizeof(u_short));
1398 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_lport, sizeof(u_short));
1400 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) {
1401 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1402 sizeof(struct in6_addr));
1403 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1404 sizeof(struct in6_addr));
1408 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1409 sizeof(struct in_addr));
1410 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1411 sizeof(struct in_addr));
1413 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1414 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1415 new_isn = (tcp_seq) md5_buffer[0];
1416 isn_offset += ISN_STATIC_INCREMENT +
1417 (arc4random() & ISN_RANDOM_INCREMENT);
1418 new_isn += isn_offset;
1424 * Increment the offset to the next ISN_BYTES_PER_SECOND / 100 boundary
1425 * to keep time flowing at a relatively constant rate. If the random
1426 * increments have already pushed us past the projected offset, do nothing.
1429 tcp_isn_tick(void *xtp)
1431 u_int32_t projected_offset;
1434 projected_offset = isn_offset_old + ISN_BYTES_PER_SECOND / 100;
1436 if (SEQ_GT(projected_offset, isn_offset))
1437 isn_offset = projected_offset;
1439 isn_offset_old = isn_offset;
1440 callout_reset(&isn_callout, hz/100, tcp_isn_tick, NULL);
1445 * When a specific ICMP unreachable message is received and the
1446 * connection state is SYN-SENT, drop the connection. This behavior
1447 * is controlled by the icmp_may_rst sysctl.
1450 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1454 INP_INFO_WLOCK_ASSERT(&tcbinfo);
1455 INP_LOCK_ASSERT(inp);
1457 if ((inp->inp_vflag & INP_TIMEWAIT) ||
1458 (inp->inp_vflag & INP_DROPPED))
1461 tp = intotcpcb(inp);
1462 if (tp->t_state != TCPS_SYN_SENT)
1465 tp = tcp_drop(tp, errno);
1473 * When `need fragmentation' ICMP is received, update our idea of the MSS
1474 * based on the new value in the route. Also nudge TCP to send something,
1475 * since we know the packet we just sent was dropped.
1476 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1479 tcp_mtudisc(struct inpcb *inp, int errno)
1482 struct socket *so = inp->inp_socket;
1490 INP_LOCK_ASSERT(inp);
1491 if ((inp->inp_vflag & INP_TIMEWAIT) ||
1492 (inp->inp_vflag & INP_DROPPED))
1495 tp = intotcpcb(inp);
1496 KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL"));
1499 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
1501 maxmtu = tcp_hc_getmtu(&inp->inp_inc); /* IPv4 and IPv6 */
1504 isipv6 ? tcp_maxmtu6(&inp->inp_inc, NULL) :
1506 tcp_maxmtu(&inp->inp_inc, NULL);
1510 maxmtu = min(maxmtu, romtu);
1512 tp->t_maxopd = tp->t_maxseg =
1514 isipv6 ? tcp_v6mssdflt :
1521 (isipv6 ? sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1523 sizeof(struct tcpiphdr)
1530 * XXX - The above conditional probably violates the TCP
1531 * spec. The problem is that, since we don't know the
1532 * other end's MSS, we are supposed to use a conservative
1533 * default. But, if we do that, then MTU discovery will
1534 * never actually take place, because the conservative
1535 * default is much less than the MTUs typically seen
1536 * on the Internet today. For the moment, we'll sweep
1537 * this under the carpet.
1539 * The conservative default might not actually be a problem
1540 * if the only case this occurs is when sending an initial
1541 * SYN with options and data to a host we've never talked
1542 * to before. Then, they will reply with an MSS value which
1543 * will get recorded and the new parameters should get
1544 * recomputed. For Further Study.
1546 if (tp->t_maxopd <= mss)
1550 if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1551 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1552 mss -= TCPOLEN_TSTAMP_APPA;
1553 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1555 mss &= ~(MCLBYTES-1);
1558 mss = mss / MCLBYTES * MCLBYTES;
1560 if (so->so_snd.sb_hiwat < mss)
1561 mss = so->so_snd.sb_hiwat;
1565 tcpstat.tcps_mturesent++;
1567 tp->snd_nxt = tp->snd_una;
1568 tcp_free_sackholes(tp);
1569 tp->snd_recover = tp->snd_max;
1570 if (tp->t_flags & TF_SACK_PERMIT)
1571 EXIT_FASTRECOVERY(tp);
1572 tcp_output_send(tp);
1577 * Look-up the routing entry to the peer of this inpcb. If no route
1578 * is found and it cannot be allocated, then return NULL. This routine
1579 * is called by TCP routines that access the rmx structure and by tcp_mss
1580 * to get the interface MTU.
1583 tcp_maxmtu(struct in_conninfo *inc, int *flags)
1586 struct sockaddr_in *dst;
1590 KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer"));
1592 bzero(&sro, sizeof(sro));
1593 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1594 dst = (struct sockaddr_in *)&sro.ro_dst;
1595 dst->sin_family = AF_INET;
1596 dst->sin_len = sizeof(*dst);
1597 dst->sin_addr = inc->inc_faddr;
1598 rtalloc_ign(&sro, RTF_CLONING);
1600 if (sro.ro_rt != NULL) {
1601 ifp = sro.ro_rt->rt_ifp;
1602 if (sro.ro_rt->rt_rmx.rmx_mtu == 0)
1603 maxmtu = ifp->if_mtu;
1605 maxmtu = min(sro.ro_rt->rt_rmx.rmx_mtu, ifp->if_mtu);
1607 /* Report additional interface capabilities. */
1608 if (flags != NULL) {
1609 if (ifp->if_capenable & IFCAP_TSO4 &&
1610 ifp->if_hwassist & CSUM_TSO)
1620 tcp_maxmtu6(struct in_conninfo *inc, int *flags)
1622 struct route_in6 sro6;
1626 KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer"));
1628 bzero(&sro6, sizeof(sro6));
1629 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1630 sro6.ro_dst.sin6_family = AF_INET6;
1631 sro6.ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1632 sro6.ro_dst.sin6_addr = inc->inc6_faddr;
1633 rtalloc_ign((struct route *)&sro6, RTF_CLONING);
1635 if (sro6.ro_rt != NULL) {
1636 ifp = sro6.ro_rt->rt_ifp;
1637 if (sro6.ro_rt->rt_rmx.rmx_mtu == 0)
1638 maxmtu = IN6_LINKMTU(sro6.ro_rt->rt_ifp);
1640 maxmtu = min(sro6.ro_rt->rt_rmx.rmx_mtu,
1641 IN6_LINKMTU(sro6.ro_rt->rt_ifp));
1643 /* Report additional interface capabilities. */
1644 if (flags != NULL) {
1645 if (ifp->if_capenable & IFCAP_TSO6 &&
1646 ifp->if_hwassist & CSUM_TSO)
1657 /* compute ESP/AH header size for TCP, including outer IP header. */
1659 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1666 struct ip6_hdr *ip6;
1670 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1672 MGETHDR(m, M_DONTWAIT, MT_DATA);
1677 if ((inp->inp_vflag & INP_IPV6) != 0) {
1678 ip6 = mtod(m, struct ip6_hdr *);
1679 th = (struct tcphdr *)(ip6 + 1);
1680 m->m_pkthdr.len = m->m_len =
1681 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1682 tcpip_fillheaders(inp, ip6, th);
1683 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1687 ip = mtod(m, struct ip *);
1688 th = (struct tcphdr *)(ip + 1);
1689 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1690 tcpip_fillheaders(inp, ip, th);
1691 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1700 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1702 * This code attempts to calculate the bandwidth-delay product as a
1703 * means of determining the optimal window size to maximize bandwidth,
1704 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1705 * routers. This code also does a fairly good job keeping RTTs in check
1706 * across slow links like modems. We implement an algorithm which is very
1707 * similar (but not meant to be) TCP/Vegas. The code operates on the
1708 * transmitter side of a TCP connection and so only effects the transmit
1709 * side of the connection.
1711 * BACKGROUND: TCP makes no provision for the management of buffer space
1712 * at the end points or at the intermediate routers and switches. A TCP
1713 * stream, whether using NewReno or not, will eventually buffer as
1714 * many packets as it is able and the only reason this typically works is
1715 * due to the fairly small default buffers made available for a connection
1716 * (typicaly 16K or 32K). As machines use larger windows and/or window
1717 * scaling it is now fairly easy for even a single TCP connection to blow-out
1718 * all available buffer space not only on the local interface, but on
1719 * intermediate routers and switches as well. NewReno makes a misguided
1720 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1721 * then backing off, then steadily increasing the window again until another
1722 * failure occurs, ad-infinitum. This results in terrible oscillation that
1723 * is only made worse as network loads increase and the idea of intentionally
1724 * blowing out network buffers is, frankly, a terrible way to manage network
1727 * It is far better to limit the transmit window prior to the failure
1728 * condition being achieved. There are two general ways to do this: First
1729 * you can 'scan' through different transmit window sizes and locate the
1730 * point where the RTT stops increasing, indicating that you have filled the
1731 * pipe, then scan backwards until you note that RTT stops decreasing, then
1732 * repeat ad-infinitum. This method works in principle but has severe
1733 * implementation issues due to RTT variances, timer granularity, and
1734 * instability in the algorithm which can lead to many false positives and
1735 * create oscillations as well as interact badly with other TCP streams
1736 * implementing the same algorithm.
1738 * The second method is to limit the window to the bandwidth delay product
1739 * of the link. This is the method we implement. RTT variances and our
1740 * own manipulation of the congestion window, bwnd, can potentially
1741 * destabilize the algorithm. For this reason we have to stabilize the
1742 * elements used to calculate the window. We do this by using the minimum
1743 * observed RTT, the long term average of the observed bandwidth, and
1744 * by adding two segments worth of slop. It isn't perfect but it is able
1745 * to react to changing conditions and gives us a very stable basis on
1746 * which to extend the algorithm.
1749 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1755 INP_LOCK_ASSERT(tp->t_inpcb);
1758 * If inflight_enable is disabled in the middle of a tcp connection,
1759 * make sure snd_bwnd is effectively disabled.
1761 if (tcp_inflight_enable == 0 || tp->t_rttlow < tcp_inflight_rttthresh) {
1762 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1763 tp->snd_bandwidth = 0;
1768 * Figure out the bandwidth. Due to the tick granularity this
1769 * is a very rough number and it MUST be averaged over a fairly
1770 * long period of time. XXX we need to take into account a link
1771 * that is not using all available bandwidth, but for now our
1772 * slop will ramp us up if this case occurs and the bandwidth later
1775 * Note: if ticks rollover 'bw' may wind up negative. We must
1776 * effectively reset t_bw_rtttime for this case.
1779 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1782 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1783 (save_ticks - tp->t_bw_rtttime);
1784 tp->t_bw_rtttime = save_ticks;
1785 tp->t_bw_rtseq = ack_seq;
1786 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1788 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1790 tp->snd_bandwidth = bw;
1793 * Calculate the semi-static bandwidth delay product, plus two maximal
1794 * segments. The additional slop puts us squarely in the sweet
1795 * spot and also handles the bandwidth run-up case and stabilization.
1796 * Without the slop we could be locking ourselves into a lower
1799 * Situations Handled:
1800 * (1) Prevents over-queueing of packets on LANs, especially on
1801 * high speed LANs, allowing larger TCP buffers to be
1802 * specified, and also does a good job preventing
1803 * over-queueing of packets over choke points like modems
1804 * (at least for the transmit side).
1806 * (2) Is able to handle changing network loads (bandwidth
1807 * drops so bwnd drops, bandwidth increases so bwnd
1810 * (3) Theoretically should stabilize in the face of multiple
1811 * connections implementing the same algorithm (this may need
1814 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1815 * be adjusted with a sysctl but typically only needs to be
1816 * on very slow connections. A value no smaller then 5
1817 * should be used, but only reduce this default if you have
1820 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1821 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) + tcp_inflight_stab * tp->t_maxseg / 10;
1824 if (tcp_inflight_debug > 0) {
1826 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1828 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1837 if ((long)bwnd < tcp_inflight_min)
1838 bwnd = tcp_inflight_min;
1839 if (bwnd > tcp_inflight_max)
1840 bwnd = tcp_inflight_max;
1841 if ((long)bwnd < tp->t_maxseg * 2)
1842 bwnd = tp->t_maxseg * 2;
1843 tp->snd_bwnd = bwnd;
1846 #ifdef TCP_SIGNATURE
1848 * Callback function invoked by m_apply() to digest TCP segment data
1849 * contained within an mbuf chain.
1852 tcp_signature_apply(void *fstate, void *data, u_int len)
1855 MD5Update(fstate, (u_char *)data, len);
1860 * Compute TCP-MD5 hash of a TCPv4 segment. (RFC2385)
1863 * m pointer to head of mbuf chain
1864 * off0 offset to TCP header within the mbuf chain
1865 * len length of TCP segment data, excluding options
1866 * optlen length of TCP segment options
1867 * buf pointer to storage for computed MD5 digest
1868 * direction direction of flow (IPSEC_DIR_INBOUND or OUTBOUND)
1870 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1871 * When called from tcp_input(), we can be sure that th_sum has been
1872 * zeroed out and verified already.
1874 * This function is for IPv4 use only. Calling this function with an
1875 * IPv6 packet in the mbuf chain will yield undefined results.
1877 * Return 0 if successful, otherwise return -1.
1879 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1880 * search with the destination IP address, and a 'magic SPI' to be
1881 * determined by the application. This is hardcoded elsewhere to 1179
1882 * right now. Another branch of this code exists which uses the SPD to
1883 * specify per-application flows but it is unstable.
1886 tcp_signature_compute(struct mbuf *m, int off0, int len, int optlen,
1887 u_char *buf, u_int direction)
1889 union sockaddr_union dst;
1890 struct ippseudo ippseudo;
1894 struct ipovly *ipovly;
1895 struct secasvar *sav;
1899 KASSERT(m != NULL, ("NULL mbuf chain"));
1900 KASSERT(buf != NULL, ("NULL signature pointer"));
1902 /* Extract the destination from the IP header in the mbuf. */
1903 ip = mtod(m, struct ip *);
1904 bzero(&dst, sizeof(union sockaddr_union));
1905 dst.sa.sa_len = sizeof(struct sockaddr_in);
1906 dst.sa.sa_family = AF_INET;
1907 dst.sin.sin_addr = (direction == IPSEC_DIR_INBOUND) ?
1908 ip->ip_src : ip->ip_dst;
1910 /* Look up an SADB entry which matches the address of the peer. */
1911 sav = KEY_ALLOCSA(&dst, IPPROTO_TCP, htonl(TCP_SIG_SPI));
1913 printf("%s: SADB lookup failed for %s\n", __func__,
1914 inet_ntoa(dst.sin.sin_addr));
1919 ipovly = (struct ipovly *)ip;
1920 th = (struct tcphdr *)((u_char *)ip + off0);
1921 doff = off0 + sizeof(struct tcphdr) + optlen;
1924 * Step 1: Update MD5 hash with IP pseudo-header.
1926 * XXX The ippseudo header MUST be digested in network byte order,
1927 * or else we'll fail the regression test. Assume all fields we've
1928 * been doing arithmetic on have been in host byte order.
1929 * XXX One cannot depend on ipovly->ih_len here. When called from
1930 * tcp_output(), the underlying ip_len member has not yet been set.
1932 ippseudo.ippseudo_src = ipovly->ih_src;
1933 ippseudo.ippseudo_dst = ipovly->ih_dst;
1934 ippseudo.ippseudo_pad = 0;
1935 ippseudo.ippseudo_p = IPPROTO_TCP;
1936 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
1937 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
1940 * Step 2: Update MD5 hash with TCP header, excluding options.
1941 * The TCP checksum must be set to zero.
1943 savecsum = th->th_sum;
1945 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
1946 th->th_sum = savecsum;
1949 * Step 3: Update MD5 hash with TCP segment data.
1950 * Use m_apply() to avoid an early m_pullup().
1953 m_apply(m, doff, len, tcp_signature_apply, &ctx);
1956 * Step 4: Update MD5 hash with shared secret.
1958 MD5Update(&ctx, sav->key_auth->key_data, _KEYLEN(sav->key_auth));
1959 MD5Final(buf, &ctx);
1961 key_sa_recordxfer(sav, m);
1965 #endif /* TCP_SIGNATURE */
1968 sysctl_drop(SYSCTL_HANDLER_ARGS)
1970 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
1971 struct sockaddr_storage addrs[2];
1975 struct sockaddr_in *fin, *lin;
1977 struct sockaddr_in6 *fin6, *lin6;
1978 struct in6_addr f6, l6;
1989 if (req->oldptr != NULL || req->oldlen != 0)
1991 if (req->newptr == NULL)
1993 if (req->newlen < sizeof(addrs))
1995 error = SYSCTL_IN(req, &addrs, sizeof(addrs));
1999 switch (addrs[0].ss_family) {
2002 fin6 = (struct sockaddr_in6 *)&addrs[0];
2003 lin6 = (struct sockaddr_in6 *)&addrs[1];
2004 if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
2005 lin6->sin6_len != sizeof(struct sockaddr_in6))
2007 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) {
2008 if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2010 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]);
2011 in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]);
2012 fin = (struct sockaddr_in *)&addrs[0];
2013 lin = (struct sockaddr_in *)&addrs[1];
2016 error = sa6_embedscope(fin6, ip6_use_defzone);
2019 error = sa6_embedscope(lin6, ip6_use_defzone);
2025 fin = (struct sockaddr_in *)&addrs[0];
2026 lin = (struct sockaddr_in *)&addrs[1];
2027 if (fin->sin_len != sizeof(struct sockaddr_in) ||
2028 lin->sin_len != sizeof(struct sockaddr_in))
2034 INP_INFO_WLOCK(&tcbinfo);
2035 switch (addrs[0].ss_family) {
2038 inp = in6_pcblookup_hash(&tcbinfo, &f6, fin6->sin6_port,
2039 &l6, lin6->sin6_port, 0, NULL);
2043 inp = in_pcblookup_hash(&tcbinfo, fin->sin_addr, fin->sin_port,
2044 lin->sin_addr, lin->sin_port, 0, NULL);
2049 if (inp->inp_vflag & INP_TIMEWAIT) {
2051 * XXXRW: There currently exists a state where an
2052 * inpcb is present, but its timewait state has been
2053 * discarded. For now, don't allow dropping of this
2061 } else if (!(inp->inp_vflag & INP_DROPPED) &&
2062 !(inp->inp_socket->so_options & SO_ACCEPTCONN)) {
2063 tp = intotcpcb(inp);
2064 tp = tcp_drop(tp, ECONNABORTED);
2071 INP_INFO_WUNLOCK(&tcbinfo);
2075 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop,
2076 CTLTYPE_STRUCT|CTLFLAG_WR|CTLFLAG_SKIP, NULL,
2077 0, sysctl_drop, "", "Drop TCP connection");
2080 * Generate a standardized TCP log line for use throughout the
2081 * tcp subsystem. Memory allocation is done with M_NOWAIT to
2082 * allow use in the interrupt context.
2084 * NB: The caller MUST free(s, M_TCPLOG) the returned string.
2085 * NB: The function may return NULL if memory allocation failed.
2087 * Due to header inclusion and ordering limitations the struct ip
2088 * and ip6_hdr pointers have to be passed as void pointers.
2091 tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr,
2098 const struct ip6_hdr *ip6;
2100 ip6 = (const struct ip6_hdr *)ip6hdr;
2102 ip = (struct ip *)ip4hdr;
2105 * The log line looks like this:
2106 * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2<SYN>"
2108 size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") +
2109 sizeof(PRINT_TH_FLAGS) + 1 +
2111 2 * INET6_ADDRSTRLEN;
2113 2 * INET_ADDRSTRLEN;
2116 /* Is logging enabled? */
2117 if (tcp_log_debug == 0 && tcp_log_in_vain == 0)
2120 s = malloc(size, M_TCPLOG, M_ZERO|M_NOWAIT);
2124 strcat(s, "TCP: [");
2127 if (inc && inc->inc_isipv6 == 0) {
2128 inet_ntoa_r(inc->inc_faddr, sp);
2130 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2132 inet_ntoa_r(inc->inc_laddr, sp);
2134 sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2137 ip6_sprintf(sp, &inc->inc6_faddr);
2139 sprintf(sp, "]:%i to [", ntohs(inc->inc_fport));
2141 ip6_sprintf(sp, &inc->inc6_laddr);
2143 sprintf(sp, "]:%i", ntohs(inc->inc_lport));
2144 } else if (ip6 && th) {
2145 ip6_sprintf(sp, &ip6->ip6_src);
2147 sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2149 ip6_sprintf(sp, &ip6->ip6_dst);
2151 sprintf(sp, "]:%i", ntohs(th->th_dport));
2153 } else if (ip && th) {
2154 inet_ntoa_r(ip->ip_src, sp);
2156 sprintf(sp, "]:%i to [", ntohs(th->th_sport));
2158 inet_ntoa_r(ip->ip_dst, sp);
2160 sprintf(sp, "]:%i", ntohs(th->th_dport));
2167 sprintf(sp, " tcpflags 0x%b", th->th_flags, PRINT_TH_FLAGS);
2168 if (*(s + size - 1) != '\0')
2169 panic("%s: string too long", __func__);