2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of The DragonFly Project nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific, prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the University of
49 * California, Berkeley and its contributors.
50 * 4. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.56 2007/03/04 18:51:59 swildner Exp $
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
85 #include <sys/domain.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/protosw.h>
91 #include <sys/random.h>
92 #include <sys/in_cksum.h>
95 #include <vm/vm_zone.h>
97 #include <net/route.h>
99 #include <net/netisr.h>
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
113 #include <netinet/icmp6.h>
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
123 #include <netinet/tcp_debug.h>
125 #include <netinet6/ip6protosw.h>
128 #include <netinet6/ipsec.h>
130 #include <netinet6/ipsec6.h>
135 #include <netproto/ipsec/ipsec.h>
137 #include <netproto/ipsec/ipsec6.h>
143 #include <sys/msgport2.h>
144 #include <machine/smp.h>
146 #if !defined(KTR_TCP)
147 #define KTR_TCP KTR_ALL
149 KTR_INFO_MASTER(tcp);
150 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
151 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
152 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
153 #define logtcp(name) KTR_LOG(tcp_ ## name)
155 struct inpcbinfo tcbinfo[MAXCPU];
156 struct tcpcbackqhead tcpcbackq[MAXCPU];
158 int tcp_mssdflt = TCP_MSS;
159 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
160 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
163 int tcp_v6mssdflt = TCP6_MSS;
164 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
165 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
169 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
170 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
171 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
174 int tcp_do_rfc1323 = 1;
175 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
176 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
178 int tcp_do_rfc1644 = 0;
179 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
180 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
182 static int tcp_tcbhashsize = 0;
183 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
184 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
186 static int do_tcpdrain = 1;
187 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
188 "Enable tcp_drain routine for extra help when low on mbufs");
191 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
192 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
194 static int icmp_may_rst = 1;
195 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
196 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
198 static int tcp_isn_reseed_interval = 0;
199 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
200 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
203 * TCP bandwidth limiting sysctls. Note that the default lower bound of
204 * 1024 exists only for debugging. A good production default would be
205 * something like 6100.
207 static int tcp_inflight_enable = 0;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
209 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
211 static int tcp_inflight_debug = 0;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
213 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
215 static int tcp_inflight_min = 6144;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
217 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
219 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
221 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
223 static int tcp_inflight_stab = 20;
224 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
225 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
227 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
228 static struct malloc_pipe tcptemp_mpipe;
230 static void tcp_willblock(void);
231 static void tcp_cleartaocache (void);
232 static void tcp_notify (struct inpcb *, int);
234 struct tcp_stats tcpstats_percpu[MAXCPU];
237 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
241 for (cpu = 0; cpu < ncpus; ++cpu) {
242 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
243 sizeof(struct tcp_stats))))
245 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
246 sizeof(struct tcp_stats))))
252 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
253 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
255 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
256 &tcpstat, tcp_stats, "TCP statistics");
260 * Target size of TCP PCB hash tables. Must be a power of two.
262 * Note that this can be overridden by the kernel environment
263 * variable net.inet.tcp.tcbhashsize
266 #define TCBHASHSIZE 512
270 * This is the actual shape of what we allocate using the zone
271 * allocator. Doing it this way allows us to protect both structures
272 * using the same generation count, and also eliminates the overhead
273 * of allocating tcpcbs separately. By hiding the structure here,
274 * we avoid changing most of the rest of the code (although it needs
275 * to be changed, eventually, for greater efficiency).
278 #define ALIGNM1 (ALIGNMENT - 1)
282 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
285 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
286 struct callout inp_tp_delack;
297 struct inpcbporthead *porthashbase;
299 struct vm_zone *ipi_zone;
300 int hashsize = TCBHASHSIZE;
304 * note: tcptemp is used for keepalives, and it is ok for an
305 * allocation to fail so do not specify MPF_INT.
307 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
313 tcp_delacktime = TCPTV_DELACK;
314 tcp_keepinit = TCPTV_KEEP_INIT;
315 tcp_keepidle = TCPTV_KEEP_IDLE;
316 tcp_keepintvl = TCPTV_KEEPINTVL;
317 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
319 tcp_rexmit_min = TCPTV_MIN;
320 tcp_rexmit_slop = TCPTV_CPU_VAR;
322 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
323 if (!powerof2(hashsize)) {
324 kprintf("WARNING: TCB hash size not a power of 2\n");
325 hashsize = 512; /* safe default */
327 tcp_tcbhashsize = hashsize;
328 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
329 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
332 for (cpu = 0; cpu < ncpus2; cpu++) {
333 in_pcbinfo_init(&tcbinfo[cpu]);
334 tcbinfo[cpu].cpu = cpu;
335 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
336 &tcbinfo[cpu].hashmask);
337 tcbinfo[cpu].porthashbase = porthashbase;
338 tcbinfo[cpu].porthashmask = porthashmask;
339 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
340 &tcbinfo[cpu].wildcardhashmask);
341 tcbinfo[cpu].ipi_zone = ipi_zone;
342 TAILQ_INIT(&tcpcbackq[cpu]);
345 tcp_reass_maxseg = nmbclusters / 16;
346 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
349 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
351 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
353 if (max_protohdr < TCP_MINPROTOHDR)
354 max_protohdr = TCP_MINPROTOHDR;
355 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
357 #undef TCP_MINPROTOHDR
360 * Initialize TCP statistics counters for each CPU.
363 for (cpu = 0; cpu < ncpus; ++cpu) {
364 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
367 bzero(&tcpstat, sizeof(struct tcp_stats));
376 tcpmsg_service_loop(void *dummy)
380 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
383 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg);
384 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
395 int cpu = mycpu->gd_cpuid;
397 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
398 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
399 tp->t_flags &= ~TF_ONOUTPUTQ;
400 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
407 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
408 * tcp_template used to store this data in mbufs, but we now recopy it out
409 * of the tcpcb each time to conserve mbufs.
412 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
414 struct inpcb *inp = tp->t_inpcb;
415 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
418 if (inp->inp_vflag & INP_IPV6) {
421 ip6 = (struct ip6_hdr *)ip_ptr;
422 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
423 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
424 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
425 (IPV6_VERSION & IPV6_VERSION_MASK);
426 ip6->ip6_nxt = IPPROTO_TCP;
427 ip6->ip6_plen = sizeof(struct tcphdr);
428 ip6->ip6_src = inp->in6p_laddr;
429 ip6->ip6_dst = inp->in6p_faddr;
434 struct ip *ip = (struct ip *) ip_ptr;
436 ip->ip_vhl = IP_VHL_BORING;
443 ip->ip_p = IPPROTO_TCP;
444 ip->ip_src = inp->inp_laddr;
445 ip->ip_dst = inp->inp_faddr;
446 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
448 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
451 tcp_hdr->th_sport = inp->inp_lport;
452 tcp_hdr->th_dport = inp->inp_fport;
457 tcp_hdr->th_flags = 0;
463 * Create template to be used to send tcp packets on a connection.
464 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
465 * use for this function is in keepalives, which use tcp_respond.
468 tcp_maketemplate(struct tcpcb *tp)
472 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
474 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
479 tcp_freetemplate(struct tcptemp *tmp)
481 mpipe_free(&tcptemp_mpipe, tmp);
485 * Send a single message to the TCP at address specified by
486 * the given TCP/IP header. If m == NULL, then we make a copy
487 * of the tcpiphdr at ti and send directly to the addressed host.
488 * This is used to force keep alive messages out using the TCP
489 * template for a connection. If flags are given then we send
490 * a message back to the TCP which originated the * segment ti,
491 * and discard the mbuf containing it and any other attached mbufs.
493 * In any case the ack and sequence number of the transmitted
494 * segment are as specified by the parameters.
496 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
499 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
500 tcp_seq ack, tcp_seq seq, int flags)
504 struct route *ro = NULL;
506 struct ip *ip = ipgen;
509 struct route_in6 *ro6 = NULL;
510 struct route_in6 sro6;
511 struct ip6_hdr *ip6 = ipgen;
513 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
515 const boolean_t isipv6 = FALSE;
519 if (!(flags & TH_RST)) {
520 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
521 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
522 win = (long)TCP_MAXWIN << tp->rcv_scale;
525 ro6 = &tp->t_inpcb->in6p_route;
527 ro = &tp->t_inpcb->inp_route;
531 bzero(ro6, sizeof *ro6);
534 bzero(ro, sizeof *ro);
538 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
542 m->m_data += max_linkhdr;
544 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
545 ip6 = mtod(m, struct ip6_hdr *);
546 nth = (struct tcphdr *)(ip6 + 1);
548 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
549 ip = mtod(m, struct ip *);
550 nth = (struct tcphdr *)(ip + 1);
552 bcopy(th, nth, sizeof(struct tcphdr));
557 m->m_data = (caddr_t)ipgen;
558 /* m_len is set later */
560 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
562 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
563 nth = (struct tcphdr *)(ip6 + 1);
565 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
566 nth = (struct tcphdr *)(ip + 1);
570 * this is usually a case when an extension header
571 * exists between the IPv6 header and the
574 nth->th_sport = th->th_sport;
575 nth->th_dport = th->th_dport;
577 xchg(nth->th_dport, nth->th_sport, n_short);
582 ip6->ip6_vfc = IPV6_VERSION;
583 ip6->ip6_nxt = IPPROTO_TCP;
584 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
585 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
587 tlen += sizeof(struct tcpiphdr);
589 ip->ip_ttl = ip_defttl;
592 m->m_pkthdr.len = tlen;
593 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
594 nth->th_seq = htonl(seq);
595 nth->th_ack = htonl(ack);
597 nth->th_off = sizeof(struct tcphdr) >> 2;
598 nth->th_flags = flags;
600 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
602 nth->th_win = htons((u_short)win);
606 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
607 sizeof(struct ip6_hdr),
608 tlen - sizeof(struct ip6_hdr));
609 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
610 (ro6 && ro6->ro_rt) ?
611 ro6->ro_rt->rt_ifp : NULL);
613 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
614 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
615 m->m_pkthdr.csum_flags = CSUM_TCP;
616 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
619 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
620 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
623 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
624 tp ? tp->t_inpcb : NULL);
625 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
630 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
631 if ((ro == &sro) && (ro->ro_rt != NULL)) {
639 * Create a new TCP control block, making an
640 * empty reassembly queue and hooking it to the argument
641 * protocol control block. The `inp' parameter must have
642 * come from the zone allocator set up in tcp_init().
645 tcp_newtcpcb(struct inpcb *inp)
650 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
652 const boolean_t isipv6 = FALSE;
655 it = (struct inp_tp *)inp;
657 bzero(tp, sizeof(struct tcpcb));
658 LIST_INIT(&tp->t_segq);
659 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
661 /* Set up our timeouts. */
662 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
663 callout_init(tp->tt_persist = &it->inp_tp_persist);
664 callout_init(tp->tt_keep = &it->inp_tp_keep);
665 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
666 callout_init(tp->tt_delack = &it->inp_tp_delack);
669 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
671 tp->t_flags |= TF_REQ_CC;
672 tp->t_inpcb = inp; /* XXX */
673 tp->t_state = TCPS_CLOSED;
675 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
676 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
677 * reasonable initial retransmit time.
679 tp->t_srtt = TCPTV_SRTTBASE;
681 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
682 tp->t_rttmin = tcp_rexmit_min;
683 tp->t_rxtcur = TCPTV_RTOBASE;
684 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
685 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
686 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
687 tp->t_rcvtime = ticks;
689 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
690 * because the socket may be bound to an IPv6 wildcard address,
691 * which may match an IPv4-mapped IPv6 address.
693 inp->inp_ip_ttl = ip_defttl;
695 tcp_sack_tcpcb_init(tp);
696 return (tp); /* XXX */
700 * Drop a TCP connection, reporting the specified error.
701 * If connection is synchronized, then send a RST to peer.
704 tcp_drop(struct tcpcb *tp, int error)
706 struct socket *so = tp->t_inpcb->inp_socket;
708 if (TCPS_HAVERCVDSYN(tp->t_state)) {
709 tp->t_state = TCPS_CLOSED;
711 tcpstat.tcps_drops++;
713 tcpstat.tcps_conndrops++;
714 if (error == ETIMEDOUT && tp->t_softerror)
715 error = tp->t_softerror;
716 so->so_error = error;
717 return (tcp_close(tp));
722 struct netmsg_remwildcard {
723 struct lwkt_msg nm_lmsg;
724 struct inpcb *nm_inp;
725 struct inpcbinfo *nm_pcbinfo;
734 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
735 * inp can be detached. We do this by cycling through the cpus, ending up
736 * on the cpu controlling the inp last and then doing the disconnect.
739 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
741 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
744 cpu = msg->nm_pcbinfo->cpu;
746 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
747 /* note: detach removes any wildcard hash entry */
750 in6_pcbdetach(msg->nm_inp);
753 in_pcbdetach(msg->nm_inp);
754 lwkt_replymsg(&msg->nm_lmsg, 0);
756 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
757 cpu = (cpu + 1) % ncpus2;
758 msg->nm_pcbinfo = &tcbinfo[cpu];
759 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg);
767 * Close a TCP control block:
768 * discard all space held by the tcp
769 * discard internet protocol block
770 * wake up any sleepers
773 tcp_close(struct tcpcb *tp)
776 struct inpcb *inp = tp->t_inpcb;
777 struct socket *so = inp->inp_socket;
779 boolean_t dosavessthresh;
784 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
785 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
787 const boolean_t isipv6 = FALSE;
791 * The tp is not instantly destroyed in the wildcard case. Setting
792 * the state to TCPS_TERMINATING will prevent the TCP stack from
793 * messing with it, though it should be noted that this change may
794 * not take effect on other cpus until we have chained the wildcard
797 * XXX we currently depend on the BGL to synchronize the tp->t_state
798 * update and prevent other tcp protocol threads from accepting new
799 * connections on the listen socket we might be trying to close down.
801 KKASSERT(tp->t_state != TCPS_TERMINATING);
802 tp->t_state = TCPS_TERMINATING;
805 * Make sure that all of our timers are stopped before we
808 callout_stop(tp->tt_rexmt);
809 callout_stop(tp->tt_persist);
810 callout_stop(tp->tt_keep);
811 callout_stop(tp->tt_2msl);
812 callout_stop(tp->tt_delack);
814 if (tp->t_flags & TF_ONOUTPUTQ) {
815 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
816 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
817 tp->t_flags &= ~TF_ONOUTPUTQ;
821 * If we got enough samples through the srtt filter,
822 * save the rtt and rttvar in the routing entry.
823 * 'Enough' is arbitrarily defined as the 16 samples.
824 * 16 samples is enough for the srtt filter to converge
825 * to within 5% of the correct value; fewer samples and
826 * we could save a very bogus rtt.
828 * Don't update the default route's characteristics and don't
829 * update anything that the user "locked".
831 if (tp->t_rttupdated >= 16) {
835 struct sockaddr_in6 *sin6;
837 if ((rt = inp->in6p_route.ro_rt) == NULL)
839 sin6 = (struct sockaddr_in6 *)rt_key(rt);
840 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
843 if ((rt = inp->inp_route.ro_rt) == NULL ||
844 ((struct sockaddr_in *)rt_key(rt))->
845 sin_addr.s_addr == INADDR_ANY)
848 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
849 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
850 if (rt->rt_rmx.rmx_rtt && i)
852 * filter this update to half the old & half
853 * the new values, converting scale.
854 * See route.h and tcp_var.h for a
855 * description of the scaling constants.
858 (rt->rt_rmx.rmx_rtt + i) / 2;
860 rt->rt_rmx.rmx_rtt = i;
861 tcpstat.tcps_cachedrtt++;
863 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
865 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
866 if (rt->rt_rmx.rmx_rttvar && i)
867 rt->rt_rmx.rmx_rttvar =
868 (rt->rt_rmx.rmx_rttvar + i) / 2;
870 rt->rt_rmx.rmx_rttvar = i;
871 tcpstat.tcps_cachedrttvar++;
874 * The old comment here said:
875 * update the pipelimit (ssthresh) if it has been updated
876 * already or if a pipesize was specified & the threshhold
877 * got below half the pipesize. I.e., wait for bad news
878 * before we start updating, then update on both good
881 * But we want to save the ssthresh even if no pipesize is
882 * specified explicitly in the route, because such
883 * connections still have an implicit pipesize specified
884 * by the global tcp_sendspace. In the absence of a reliable
885 * way to calculate the pipesize, it will have to do.
887 i = tp->snd_ssthresh;
888 if (rt->rt_rmx.rmx_sendpipe != 0)
889 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
891 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
892 if (dosavessthresh ||
893 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
894 (rt->rt_rmx.rmx_ssthresh != 0))) {
896 * convert the limit from user data bytes to
897 * packets then to packet data bytes.
899 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
904 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
905 sizeof(struct tcpiphdr));
906 if (rt->rt_rmx.rmx_ssthresh)
907 rt->rt_rmx.rmx_ssthresh =
908 (rt->rt_rmx.rmx_ssthresh + i) / 2;
910 rt->rt_rmx.rmx_ssthresh = i;
911 tcpstat.tcps_cachedssthresh++;
916 /* free the reassembly queue, if any */
917 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
918 LIST_REMOVE(q, tqe_q);
923 /* throw away SACK blocks in scoreboard*/
925 tcp_sack_cleanup(&tp->scb);
927 inp->inp_ppcb = NULL;
928 soisdisconnected(so);
930 * Discard the inp. In the SMP case a wildcard inp's hash (created
931 * by a listen socket or an INADDR_ANY udp socket) is replicated
932 * for each protocol thread and must be removed in the context of
933 * that thread. This is accomplished by chaining the message
936 * If the inp is not wildcarded we simply detach, which will remove
937 * the any hashes still present for this inp.
940 if (inp->inp_flags & INP_WILDCARD_MP) {
941 struct netmsg_remwildcard *msg;
943 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
944 msg = kmalloc(sizeof(struct netmsg_remwildcard),
945 M_LWKTMSG, M_INTWAIT);
946 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
947 lwkt_cmd_func(in_pcbremwildcardhash_handler),
950 msg->nm_isinet6 = isafinet6;
953 msg->nm_pcbinfo = &tcbinfo[cpu];
954 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
958 /* note: detach removes any wildcard hash entry */
966 tcpstat.tcps_closed++;
971 tcp_drain_oncpu(struct inpcbhead *head)
975 struct tseg_qent *te;
977 LIST_FOREACH(inpb, head, inp_list) {
978 if (inpb->inp_flags & INP_PLACEMARKER)
980 if ((tcpb = intotcpcb(inpb))) {
981 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
982 LIST_REMOVE(te, tqe_q);
992 struct netmsg_tcp_drain {
993 struct lwkt_msg nm_lmsg;
994 struct inpcbhead *nm_head;
998 tcp_drain_handler(lwkt_msg_t lmsg)
1000 struct netmsg_tcp_drain *nm = (void *)lmsg;
1002 tcp_drain_oncpu(nm->nm_head);
1003 lwkt_replymsg(lmsg, 0);
1019 * Walk the tcpbs, if existing, and flush the reassembly queue,
1020 * if there is one...
1021 * XXX: The "Net/3" implementation doesn't imply that the TCP
1022 * reassembly queue should be flushed, but in a situation
1023 * where we're really low on mbufs, this is potentially
1027 for (cpu = 0; cpu < ncpus2; cpu++) {
1028 struct netmsg_tcp_drain *msg;
1030 if (cpu == mycpu->gd_cpuid) {
1031 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1033 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1034 M_LWKTMSG, M_NOWAIT);
1037 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
1038 lwkt_cmd_func(tcp_drain_handler),
1040 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1041 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
1045 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1050 * Notify a tcp user of an asynchronous error;
1051 * store error as soft error, but wake up user
1052 * (for now, won't do anything until can select for soft error).
1054 * Do not wake up user since there currently is no mechanism for
1055 * reporting soft errors (yet - a kqueue filter may be added).
1058 tcp_notify(struct inpcb *inp, int error)
1060 struct tcpcb *tp = intotcpcb(inp);
1063 * Ignore some errors if we are hooked up.
1064 * If connection hasn't completed, has retransmitted several times,
1065 * and receives a second error, give up now. This is better
1066 * than waiting a long time to establish a connection that
1067 * can never complete.
1069 if (tp->t_state == TCPS_ESTABLISHED &&
1070 (error == EHOSTUNREACH || error == ENETUNREACH ||
1071 error == EHOSTDOWN)) {
1073 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1075 tcp_drop(tp, error);
1077 tp->t_softerror = error;
1079 wakeup(&so->so_timeo);
1086 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1089 struct inpcb *marker;
1099 * The process of preparing the TCB list is too time-consuming and
1100 * resource-intensive to repeat twice on every request.
1102 if (req->oldptr == NULL) {
1103 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1104 gd = globaldata_find(ccpu);
1105 n += tcbinfo[gd->gd_cpuid].ipi_count;
1107 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1111 if (req->newptr != NULL)
1114 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1115 marker->inp_flags |= INP_PLACEMARKER;
1118 * OK, now we're committed to doing something. Run the inpcb list
1119 * for each cpu in the system and construct the output. Use a
1120 * list placemarker to deal with list changes occuring during
1121 * copyout blockages (but otherwise depend on being on the correct
1122 * cpu to avoid races).
1124 origcpu = mycpu->gd_cpuid;
1125 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1131 cpu_id = (origcpu + ccpu) % ncpus;
1132 if ((smp_active_mask & (1 << cpu_id)) == 0)
1134 rgd = globaldata_find(cpu_id);
1135 lwkt_setcpu_self(rgd);
1137 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1138 n = tcbinfo[cpu_id].ipi_count;
1140 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1142 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1144 * process a snapshot of pcbs, ignoring placemarkers
1145 * and using our own to allow SYSCTL_OUT to block.
1147 LIST_REMOVE(marker, inp_list);
1148 LIST_INSERT_AFTER(inp, marker, inp_list);
1150 if (inp->inp_flags & INP_PLACEMARKER)
1152 if (inp->inp_gencnt > gencnt)
1154 if (prison_xinpcb(req->td, inp))
1157 xt.xt_len = sizeof xt;
1158 bcopy(inp, &xt.xt_inp, sizeof *inp);
1159 inp_ppcb = inp->inp_ppcb;
1160 if (inp_ppcb != NULL)
1161 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1163 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1164 if (inp->inp_socket)
1165 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1166 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1170 LIST_REMOVE(marker, inp_list);
1171 if (error == 0 && i < n) {
1172 bzero(&xt, sizeof xt);
1173 xt.xt_len = sizeof xt;
1175 error = SYSCTL_OUT(req, &xt, sizeof xt);
1184 * Make sure we are on the same cpu we were on originally, since
1185 * higher level callers expect this. Also don't pollute caches with
1186 * migrated userland data by (eventually) returning to userland
1187 * on a different cpu.
1189 lwkt_setcpu_self(globaldata_find(origcpu));
1190 kfree(marker, M_TEMP);
1194 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1195 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1198 tcp_getcred(SYSCTL_HANDLER_ARGS)
1200 struct sockaddr_in addrs[2];
1205 error = suser(req->td);
1208 error = SYSCTL_IN(req, addrs, sizeof addrs);
1212 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1213 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1214 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1215 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1216 if (inp == NULL || inp->inp_socket == NULL) {
1220 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1226 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1227 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1231 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1233 struct sockaddr_in6 addrs[2];
1236 boolean_t mapped = FALSE;
1238 error = suser(req->td);
1241 error = SYSCTL_IN(req, addrs, sizeof addrs);
1244 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1245 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1252 inp = in_pcblookup_hash(&tcbinfo[0],
1253 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1255 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1259 inp = in6_pcblookup_hash(&tcbinfo[0],
1260 &addrs[1].sin6_addr, addrs[1].sin6_port,
1261 &addrs[0].sin6_addr, addrs[0].sin6_port,
1264 if (inp == NULL || inp->inp_socket == NULL) {
1268 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1274 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1276 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1280 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1282 struct ip *ip = vip;
1284 struct in_addr faddr;
1287 void (*notify)(struct inpcb *, int) = tcp_notify;
1291 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1295 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1296 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1299 arg = inetctlerrmap[cmd];
1300 if (cmd == PRC_QUENCH) {
1301 notify = tcp_quench;
1302 } else if (icmp_may_rst &&
1303 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1304 cmd == PRC_UNREACH_PORT ||
1305 cmd == PRC_TIMXCEED_INTRANS) &&
1307 notify = tcp_drop_syn_sent;
1308 } else if (cmd == PRC_MSGSIZE) {
1309 struct icmp *icmp = (struct icmp *)
1310 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1312 arg = ntohs(icmp->icmp_nextmtu);
1313 notify = tcp_mtudisc;
1314 } else if (PRC_IS_REDIRECT(cmd)) {
1316 notify = in_rtchange;
1317 } else if (cmd == PRC_HOSTDEAD) {
1323 th = (struct tcphdr *)((caddr_t)ip +
1324 (IP_VHL_HL(ip->ip_vhl) << 2));
1325 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1326 ip->ip_src.s_addr, th->th_sport);
1327 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1328 ip->ip_src, th->th_sport, 0, NULL);
1329 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1330 icmpseq = htonl(th->th_seq);
1331 tp = intotcpcb(inp);
1332 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1333 SEQ_LT(icmpseq, tp->snd_max))
1334 (*notify)(inp, arg);
1336 struct in_conninfo inc;
1338 inc.inc_fport = th->th_dport;
1339 inc.inc_lport = th->th_sport;
1340 inc.inc_faddr = faddr;
1341 inc.inc_laddr = ip->ip_src;
1345 syncache_unreach(&inc, th);
1349 for (cpu = 0; cpu < ncpus2; cpu++) {
1350 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1358 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1361 void (*notify) (struct inpcb *, int) = tcp_notify;
1362 struct ip6_hdr *ip6;
1364 struct ip6ctlparam *ip6cp = NULL;
1365 const struct sockaddr_in6 *sa6_src = NULL;
1367 struct tcp_portonly {
1373 if (sa->sa_family != AF_INET6 ||
1374 sa->sa_len != sizeof(struct sockaddr_in6))
1378 if (cmd == PRC_QUENCH)
1379 notify = tcp_quench;
1380 else if (cmd == PRC_MSGSIZE) {
1381 struct ip6ctlparam *ip6cp = d;
1382 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1384 arg = ntohl(icmp6->icmp6_mtu);
1385 notify = tcp_mtudisc;
1386 } else if (!PRC_IS_REDIRECT(cmd) &&
1387 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1391 /* if the parameter is from icmp6, decode it. */
1393 ip6cp = (struct ip6ctlparam *)d;
1395 ip6 = ip6cp->ip6c_ip6;
1396 off = ip6cp->ip6c_off;
1397 sa6_src = ip6cp->ip6c_src;
1401 off = 0; /* fool gcc */
1406 struct in_conninfo inc;
1408 * XXX: We assume that when IPV6 is non NULL,
1409 * M and OFF are valid.
1412 /* check if we can safely examine src and dst ports */
1413 if (m->m_pkthdr.len < off + sizeof *thp)
1416 bzero(&th, sizeof th);
1417 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1419 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1420 (struct sockaddr *)ip6cp->ip6c_src,
1421 th.th_sport, cmd, arg, notify);
1423 inc.inc_fport = th.th_dport;
1424 inc.inc_lport = th.th_sport;
1425 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1426 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1428 syncache_unreach(&inc, &th);
1430 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1431 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1436 * Following is where TCP initial sequence number generation occurs.
1438 * There are two places where we must use initial sequence numbers:
1439 * 1. In SYN-ACK packets.
1440 * 2. In SYN packets.
1442 * All ISNs for SYN-ACK packets are generated by the syncache. See
1443 * tcp_syncache.c for details.
1445 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1446 * depends on this property. In addition, these ISNs should be
1447 * unguessable so as to prevent connection hijacking. To satisfy
1448 * the requirements of this situation, the algorithm outlined in
1449 * RFC 1948 is used to generate sequence numbers.
1451 * Implementation details:
1453 * Time is based off the system timer, and is corrected so that it
1454 * increases by one megabyte per second. This allows for proper
1455 * recycling on high speed LANs while still leaving over an hour
1458 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1459 * between seeding of isn_secret. This is normally set to zero,
1460 * as reseeding should not be necessary.
1464 #define ISN_BYTES_PER_SECOND 1048576
1466 u_char isn_secret[32];
1467 int isn_last_reseed;
1471 tcp_new_isn(struct tcpcb *tp)
1473 u_int32_t md5_buffer[4];
1476 /* Seed if this is the first use, reseed if requested. */
1477 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1478 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1480 read_random_unlimited(&isn_secret, sizeof isn_secret);
1481 isn_last_reseed = ticks;
1484 /* Compute the md5 hash and return the ISN. */
1486 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1487 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1489 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1490 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1491 sizeof(struct in6_addr));
1492 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1493 sizeof(struct in6_addr));
1497 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1498 sizeof(struct in_addr));
1499 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1500 sizeof(struct in_addr));
1502 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1503 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1504 new_isn = (tcp_seq) md5_buffer[0];
1505 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1510 * When a source quench is received, close congestion window
1511 * to one segment. We will gradually open it again as we proceed.
1514 tcp_quench(struct inpcb *inp, int error)
1516 struct tcpcb *tp = intotcpcb(inp);
1519 tp->snd_cwnd = tp->t_maxseg;
1525 * When a specific ICMP unreachable message is received and the
1526 * connection state is SYN-SENT, drop the connection. This behavior
1527 * is controlled by the icmp_may_rst sysctl.
1530 tcp_drop_syn_sent(struct inpcb *inp, int error)
1532 struct tcpcb *tp = intotcpcb(inp);
1534 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1535 tcp_drop(tp, error);
1539 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1540 * based on the new value in the route. Also nudge TCP to send something,
1541 * since we know the packet we just sent was dropped.
1542 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1545 tcp_mtudisc(struct inpcb *inp, int mtu)
1547 struct tcpcb *tp = intotcpcb(inp);
1549 struct socket *so = inp->inp_socket;
1552 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1554 const boolean_t isipv6 = FALSE;
1561 * If no MTU is provided in the ICMP message, use the
1562 * next lower likely value, as specified in RFC 1191.
1567 oldmtu = tp->t_maxopd +
1569 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1570 sizeof(struct tcpiphdr));
1571 mtu = ip_next_mtu(oldmtu, 0);
1575 rt = tcp_rtlookup6(&inp->inp_inc);
1577 rt = tcp_rtlookup(&inp->inp_inc);
1579 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1581 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1582 mtu = rt->rt_rmx.rmx_mtu;
1586 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1587 sizeof(struct tcpiphdr));
1590 * XXX - The following conditional probably violates the TCP
1591 * spec. The problem is that, since we don't know the
1592 * other end's MSS, we are supposed to use a conservative
1593 * default. But, if we do that, then MTU discovery will
1594 * never actually take place, because the conservative
1595 * default is much less than the MTUs typically seen
1596 * on the Internet today. For the moment, we'll sweep
1597 * this under the carpet.
1599 * The conservative default might not actually be a problem
1600 * if the only case this occurs is when sending an initial
1601 * SYN with options and data to a host we've never talked
1602 * to before. Then, they will reply with an MSS value which
1603 * will get recorded and the new parameters should get
1604 * recomputed. For Further Study.
1606 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1607 maxopd = taop->tao_mssopt;
1611 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1612 sizeof(struct tcpiphdr));
1614 if (tp->t_maxopd <= maxopd)
1616 tp->t_maxopd = maxopd;
1619 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1620 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1621 mss -= TCPOLEN_TSTAMP_APPA;
1623 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1624 (TF_REQ_CC | TF_RCVD_CC))
1625 mss -= TCPOLEN_CC_APPA;
1627 /* round down to multiple of MCLBYTES */
1628 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1630 mss &= ~(MCLBYTES - 1);
1633 mss = (mss / MCLBYTES) * MCLBYTES;
1636 if (so->so_snd.sb_hiwat < mss)
1637 mss = so->so_snd.sb_hiwat;
1641 tp->snd_nxt = tp->snd_una;
1643 tcpstat.tcps_mturesent++;
1647 * Look-up the routing entry to the peer of this inpcb. If no route
1648 * is found and it cannot be allocated the return NULL. This routine
1649 * is called by TCP routines that access the rmx structure and by tcp_mss
1650 * to get the interface MTU.
1653 tcp_rtlookup(struct in_conninfo *inc)
1655 struct route *ro = &inc->inc_route;
1657 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1658 /* No route yet, so try to acquire one */
1659 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1661 * unused portions of the structure MUST be zero'd
1662 * out because rtalloc() treats it as opaque data
1664 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1665 ro->ro_dst.sa_family = AF_INET;
1666 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1667 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1677 tcp_rtlookup6(struct in_conninfo *inc)
1679 struct route_in6 *ro6 = &inc->inc6_route;
1681 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1682 /* No route yet, so try to acquire one */
1683 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1685 * unused portions of the structure MUST be zero'd
1686 * out because rtalloc() treats it as opaque data
1688 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1689 ro6->ro_dst.sin6_family = AF_INET6;
1690 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1691 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1692 rtalloc((struct route *)ro6);
1695 return (ro6->ro_rt);
1700 /* compute ESP/AH header size for TCP, including outer IP header. */
1702 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1710 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1712 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1717 if (inp->inp_vflag & INP_IPV6) {
1718 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1720 th = (struct tcphdr *)(ip6 + 1);
1721 m->m_pkthdr.len = m->m_len =
1722 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1723 tcp_fillheaders(tp, ip6, th);
1724 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1728 ip = mtod(m, struct ip *);
1729 th = (struct tcphdr *)(ip + 1);
1730 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1731 tcp_fillheaders(tp, ip, th);
1732 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1741 * Return a pointer to the cached information about the remote host.
1742 * The cached information is stored in the protocol specific part of
1743 * the route metrics.
1746 tcp_gettaocache(struct in_conninfo *inc)
1751 if (inc->inc_isipv6)
1752 rt = tcp_rtlookup6(inc);
1755 rt = tcp_rtlookup(inc);
1757 /* Make sure this is a host route and is up. */
1759 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1762 return (rmx_taop(rt->rt_rmx));
1766 * Clear all the TAO cache entries, called from tcp_init.
1769 * This routine is just an empty one, because we assume that the routing
1770 * routing tables are initialized at the same time when TCP, so there is
1771 * nothing in the cache left over.
1774 tcp_cleartaocache(void)
1779 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1781 * This code attempts to calculate the bandwidth-delay product as a
1782 * means of determining the optimal window size to maximize bandwidth,
1783 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1784 * routers. This code also does a fairly good job keeping RTTs in check
1785 * across slow links like modems. We implement an algorithm which is very
1786 * similar (but not meant to be) TCP/Vegas. The code operates on the
1787 * transmitter side of a TCP connection and so only effects the transmit
1788 * side of the connection.
1790 * BACKGROUND: TCP makes no provision for the management of buffer space
1791 * at the end points or at the intermediate routers and switches. A TCP
1792 * stream, whether using NewReno or not, will eventually buffer as
1793 * many packets as it is able and the only reason this typically works is
1794 * due to the fairly small default buffers made available for a connection
1795 * (typicaly 16K or 32K). As machines use larger windows and/or window
1796 * scaling it is now fairly easy for even a single TCP connection to blow-out
1797 * all available buffer space not only on the local interface, but on
1798 * intermediate routers and switches as well. NewReno makes a misguided
1799 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1800 * then backing off, then steadily increasing the window again until another
1801 * failure occurs, ad-infinitum. This results in terrible oscillation that
1802 * is only made worse as network loads increase and the idea of intentionally
1803 * blowing out network buffers is, frankly, a terrible way to manage network
1806 * It is far better to limit the transmit window prior to the failure
1807 * condition being achieved. There are two general ways to do this: First
1808 * you can 'scan' through different transmit window sizes and locate the
1809 * point where the RTT stops increasing, indicating that you have filled the
1810 * pipe, then scan backwards until you note that RTT stops decreasing, then
1811 * repeat ad-infinitum. This method works in principle but has severe
1812 * implementation issues due to RTT variances, timer granularity, and
1813 * instability in the algorithm which can lead to many false positives and
1814 * create oscillations as well as interact badly with other TCP streams
1815 * implementing the same algorithm.
1817 * The second method is to limit the window to the bandwidth delay product
1818 * of the link. This is the method we implement. RTT variances and our
1819 * own manipulation of the congestion window, bwnd, can potentially
1820 * destabilize the algorithm. For this reason we have to stabilize the
1821 * elements used to calculate the window. We do this by using the minimum
1822 * observed RTT, the long term average of the observed bandwidth, and
1823 * by adding two segments worth of slop. It isn't perfect but it is able
1824 * to react to changing conditions and gives us a very stable basis on
1825 * which to extend the algorithm.
1828 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1836 * If inflight_enable is disabled in the middle of a tcp connection,
1837 * make sure snd_bwnd is effectively disabled.
1839 if (!tcp_inflight_enable) {
1840 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1841 tp->snd_bandwidth = 0;
1846 * Validate the delta time. If a connection is new or has been idle
1847 * a long time we have to reset the bandwidth calculator.
1850 delta_ticks = save_ticks - tp->t_bw_rtttime;
1851 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1852 tp->t_bw_rtttime = ticks;
1853 tp->t_bw_rtseq = ack_seq;
1854 if (tp->snd_bandwidth == 0)
1855 tp->snd_bandwidth = tcp_inflight_min;
1858 if (delta_ticks == 0)
1862 * Sanity check, plus ignore pure window update acks.
1864 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1868 * Figure out the bandwidth. Due to the tick granularity this
1869 * is a very rough number and it MUST be averaged over a fairly
1870 * long period of time. XXX we need to take into account a link
1871 * that is not using all available bandwidth, but for now our
1872 * slop will ramp us up if this case occurs and the bandwidth later
1875 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1876 tp->t_bw_rtttime = save_ticks;
1877 tp->t_bw_rtseq = ack_seq;
1878 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1880 tp->snd_bandwidth = bw;
1883 * Calculate the semi-static bandwidth delay product, plus two maximal
1884 * segments. The additional slop puts us squarely in the sweet
1885 * spot and also handles the bandwidth run-up case. Without the
1886 * slop we could be locking ourselves into a lower bandwidth.
1888 * Situations Handled:
1889 * (1) Prevents over-queueing of packets on LANs, especially on
1890 * high speed LANs, allowing larger TCP buffers to be
1891 * specified, and also does a good job preventing
1892 * over-queueing of packets over choke points like modems
1893 * (at least for the transmit side).
1895 * (2) Is able to handle changing network loads (bandwidth
1896 * drops so bwnd drops, bandwidth increases so bwnd
1899 * (3) Theoretically should stabilize in the face of multiple
1900 * connections implementing the same algorithm (this may need
1903 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1904 * be adjusted with a sysctl but typically only needs to be on
1905 * very slow connections. A value no smaller then 5 should
1906 * be used, but only reduce this default if you have no other
1910 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1911 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1912 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1915 if (tcp_inflight_debug > 0) {
1917 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1919 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1920 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1923 if ((long)bwnd < tcp_inflight_min)
1924 bwnd = tcp_inflight_min;
1925 if (bwnd > tcp_inflight_max)
1926 bwnd = tcp_inflight_max;
1927 if ((long)bwnd < tp->t_maxseg * 2)
1928 bwnd = tp->t_maxseg * 2;
1929 tp->snd_bwnd = bwnd;