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) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
37 * License terms: all terms for the DragonFly license above plus the following:
39 * 4. All advertising materials mentioning features or use of this software
40 * must display the following acknowledgement:
42 * This product includes software developed by Jeffrey M. Hsu
43 * for the DragonFly Project.
45 * This requirement may be waived with permission from Jeffrey Hsu.
46 * This requirement will sunset and may be removed on July 8 2005,
47 * after which the standard DragonFly license (as shown above) will
52 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
53 * The Regents of the University of California. All rights reserved.
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
63 * 3. All advertising materials mentioning features or use of this software
64 * must display the following acknowledgement:
65 * This product includes software developed by the University of
66 * California, Berkeley and its contributors.
67 * 4. Neither the name of the University nor the names of its contributors
68 * may be used to endorse or promote products derived from this software
69 * without specific prior written permission.
71 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
72 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
73 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
74 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
75 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
76 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
77 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
78 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
79 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
80 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
83 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
84 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
85 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.49 2005/06/02 23:52:42 dillon Exp $
88 #include "opt_compat.h"
89 #include "opt_inet6.h"
90 #include "opt_ipsec.h"
91 #include "opt_tcpdebug.h"
93 #include <sys/param.h>
94 #include <sys/systm.h>
95 #include <sys/callout.h>
96 #include <sys/kernel.h>
97 #include <sys/sysctl.h>
98 #include <sys/malloc.h>
99 #include <sys/mpipe.h>
100 #include <sys/mbuf.h>
102 #include <sys/domain.h>
104 #include <sys/proc.h>
105 #include <sys/socket.h>
106 #include <sys/socketvar.h>
107 #include <sys/protosw.h>
108 #include <sys/random.h>
109 #include <sys/in_cksum.h>
111 #include <vm/vm_zone.h>
113 #include <net/route.h>
115 #include <net/netisr.h>
118 #include <netinet/in.h>
119 #include <netinet/in_systm.h>
120 #include <netinet/ip.h>
121 #include <netinet/ip6.h>
122 #include <netinet/in_pcb.h>
123 #include <netinet6/in6_pcb.h>
124 #include <netinet/in_var.h>
125 #include <netinet/ip_var.h>
126 #include <netinet6/ip6_var.h>
127 #include <netinet/ip_icmp.h>
129 #include <netinet/icmp6.h>
131 #include <netinet/tcp.h>
132 #include <netinet/tcp_fsm.h>
133 #include <netinet/tcp_seq.h>
134 #include <netinet/tcp_timer.h>
135 #include <netinet/tcp_var.h>
136 #include <netinet6/tcp6_var.h>
137 #include <netinet/tcpip.h>
139 #include <netinet/tcp_debug.h>
141 #include <netinet6/ip6protosw.h>
144 #include <netinet6/ipsec.h>
146 #include <netinet6/ipsec6.h>
151 #include <netproto/ipsec/ipsec.h>
153 #include <netproto/ipsec/ipsec6.h>
160 #include <sys/msgport2.h>
162 #include <machine/smp.h>
164 struct inpcbinfo tcbinfo[MAXCPU];
165 struct tcpcbackqhead tcpcbackq[MAXCPU];
167 int tcp_mssdflt = TCP_MSS;
168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
169 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
172 int tcp_v6mssdflt = TCP6_MSS;
173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
174 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
178 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
179 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
180 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
183 int tcp_do_rfc1323 = 1;
184 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
185 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
187 int tcp_do_rfc1644 = 0;
188 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
189 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
191 static int tcp_tcbhashsize = 0;
192 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
193 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
195 static int do_tcpdrain = 1;
196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
197 "Enable tcp_drain routine for extra help when low on mbufs");
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
201 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
203 static int icmp_may_rst = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
205 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
207 static int tcp_isn_reseed_interval = 0;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
209 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
212 * TCP bandwidth limiting sysctls. Note that the default lower bound of
213 * 1024 exists only for debugging. A good production default would be
214 * something like 6100.
216 static int tcp_inflight_enable = 0;
217 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
218 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
220 static int tcp_inflight_debug = 0;
221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
222 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
224 static int tcp_inflight_min = 6144;
225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
226 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
228 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
229 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
230 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
232 static int tcp_inflight_stab = 20;
233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
234 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
236 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
237 static struct malloc_pipe tcptemp_mpipe;
239 static void tcp_willblock(void);
240 static void tcp_cleartaocache (void);
241 static void tcp_notify (struct inpcb *, int);
243 struct tcp_stats tcpstats_percpu[MAXCPU];
246 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
250 for (cpu = 0; cpu < ncpus; ++cpu) {
251 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
252 sizeof(struct tcp_stats))))
254 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
255 sizeof(struct tcp_stats))))
261 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
262 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
264 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
265 &tcpstat, tcp_stats, "TCP statistics");
269 * Target size of TCP PCB hash tables. Must be a power of two.
271 * Note that this can be overridden by the kernel environment
272 * variable net.inet.tcp.tcbhashsize
275 #define TCBHASHSIZE 512
279 * This is the actual shape of what we allocate using the zone
280 * allocator. Doing it this way allows us to protect both structures
281 * using the same generation count, and also eliminates the overhead
282 * of allocating tcpcbs separately. By hiding the structure here,
283 * we avoid changing most of the rest of the code (although it needs
284 * to be changed, eventually, for greater efficiency).
287 #define ALIGNM1 (ALIGNMENT - 1)
291 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
294 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
295 struct callout inp_tp_delack;
306 struct inpcbporthead *porthashbase;
308 struct vm_zone *ipi_zone;
309 int hashsize = TCBHASHSIZE;
313 * note: tcptemp is used for keepalives, and it is ok for an
314 * allocation to fail so do not specify MPF_INT.
316 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
322 tcp_delacktime = TCPTV_DELACK;
323 tcp_keepinit = TCPTV_KEEP_INIT;
324 tcp_keepidle = TCPTV_KEEP_IDLE;
325 tcp_keepintvl = TCPTV_KEEPINTVL;
326 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
328 tcp_rexmit_min = TCPTV_MIN;
329 tcp_rexmit_slop = TCPTV_CPU_VAR;
331 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
332 if (!powerof2(hashsize)) {
333 printf("WARNING: TCB hash size not a power of 2\n");
334 hashsize = 512; /* safe default */
336 tcp_tcbhashsize = hashsize;
337 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
338 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
341 for (cpu = 0; cpu < ncpus2; cpu++) {
342 in_pcbinfo_init(&tcbinfo[cpu]);
343 tcbinfo[cpu].cpu = cpu;
344 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
345 &tcbinfo[cpu].hashmask);
346 tcbinfo[cpu].porthashbase = porthashbase;
347 tcbinfo[cpu].porthashmask = porthashmask;
348 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
349 &tcbinfo[cpu].wildcardhashmask);
350 tcbinfo[cpu].ipi_zone = ipi_zone;
351 TAILQ_INIT(&tcpcbackq[cpu]);
354 tcp_reass_maxseg = nmbclusters / 16;
355 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
358 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
360 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
362 if (max_protohdr < TCP_MINPROTOHDR)
363 max_protohdr = TCP_MINPROTOHDR;
364 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
366 #undef TCP_MINPROTOHDR
369 * Initialize TCP statistics counters for each CPU.
372 for (cpu = 0; cpu < ncpus; ++cpu) {
373 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
376 bzero(&tcpstat, sizeof(struct tcp_stats));
385 tcpmsg_service_loop(void *dummy)
389 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
391 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg);
392 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
401 int cpu = mycpu->gd_cpuid;
403 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
404 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
405 tp->t_flags &= ~TF_ONOUTPUTQ;
406 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
413 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
414 * tcp_template used to store this data in mbufs, but we now recopy it out
415 * of the tcpcb each time to conserve mbufs.
418 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
420 struct inpcb *inp = tp->t_inpcb;
421 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
424 if (inp->inp_vflag & INP_IPV6) {
427 ip6 = (struct ip6_hdr *)ip_ptr;
428 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
429 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
430 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
431 (IPV6_VERSION & IPV6_VERSION_MASK);
432 ip6->ip6_nxt = IPPROTO_TCP;
433 ip6->ip6_plen = sizeof(struct tcphdr);
434 ip6->ip6_src = inp->in6p_laddr;
435 ip6->ip6_dst = inp->in6p_faddr;
440 struct ip *ip = (struct ip *) ip_ptr;
442 ip->ip_vhl = IP_VHL_BORING;
449 ip->ip_p = IPPROTO_TCP;
450 ip->ip_src = inp->inp_laddr;
451 ip->ip_dst = inp->inp_faddr;
452 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
454 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
457 tcp_hdr->th_sport = inp->inp_lport;
458 tcp_hdr->th_dport = inp->inp_fport;
463 tcp_hdr->th_flags = 0;
469 * Create template to be used to send tcp packets on a connection.
470 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
471 * use for this function is in keepalives, which use tcp_respond.
474 tcp_maketemplate(struct tcpcb *tp)
478 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
480 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
485 tcp_freetemplate(struct tcptemp *tmp)
487 mpipe_free(&tcptemp_mpipe, tmp);
491 * Send a single message to the TCP at address specified by
492 * the given TCP/IP header. If m == NULL, then we make a copy
493 * of the tcpiphdr at ti and send directly to the addressed host.
494 * This is used to force keep alive messages out using the TCP
495 * template for a connection. If flags are given then we send
496 * a message back to the TCP which originated the * segment ti,
497 * and discard the mbuf containing it and any other attached mbufs.
499 * In any case the ack and sequence number of the transmitted
500 * segment are as specified by the parameters.
502 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
505 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
506 tcp_seq ack, tcp_seq seq, int flags)
510 struct route *ro = NULL;
512 struct ip *ip = ipgen;
515 struct route_in6 *ro6 = NULL;
516 struct route_in6 sro6;
517 struct ip6_hdr *ip6 = ipgen;
519 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
521 const boolean_t isipv6 = FALSE;
525 if (!(flags & TH_RST)) {
526 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
527 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
528 win = (long)TCP_MAXWIN << tp->rcv_scale;
531 ro6 = &tp->t_inpcb->in6p_route;
533 ro = &tp->t_inpcb->inp_route;
537 bzero(ro6, sizeof *ro6);
540 bzero(ro, sizeof *ro);
544 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
548 m->m_data += max_linkhdr;
550 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
551 ip6 = mtod(m, struct ip6_hdr *);
552 nth = (struct tcphdr *)(ip6 + 1);
554 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
555 ip = mtod(m, struct ip *);
556 nth = (struct tcphdr *)(ip + 1);
558 bcopy(th, nth, sizeof(struct tcphdr));
563 m->m_data = (caddr_t)ipgen;
564 /* m_len is set later */
566 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
568 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
569 nth = (struct tcphdr *)(ip6 + 1);
571 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
572 nth = (struct tcphdr *)(ip + 1);
576 * this is usually a case when an extension header
577 * exists between the IPv6 header and the
580 nth->th_sport = th->th_sport;
581 nth->th_dport = th->th_dport;
583 xchg(nth->th_dport, nth->th_sport, n_short);
588 ip6->ip6_vfc = IPV6_VERSION;
589 ip6->ip6_nxt = IPPROTO_TCP;
590 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
591 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
593 tlen += sizeof(struct tcpiphdr);
595 ip->ip_ttl = ip_defttl;
598 m->m_pkthdr.len = tlen;
599 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
600 nth->th_seq = htonl(seq);
601 nth->th_ack = htonl(ack);
603 nth->th_off = sizeof(struct tcphdr) >> 2;
604 nth->th_flags = flags;
606 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
608 nth->th_win = htons((u_short)win);
612 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
613 sizeof(struct ip6_hdr),
614 tlen - sizeof(struct ip6_hdr));
615 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
616 (ro6 && ro6->ro_rt) ?
617 ro6->ro_rt->rt_ifp : NULL);
619 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
620 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
621 m->m_pkthdr.csum_flags = CSUM_TCP;
622 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
625 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
626 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
629 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
630 tp ? tp->t_inpcb : NULL);
631 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
636 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
637 if ((ro == &sro) && (ro->ro_rt != NULL)) {
645 * Create a new TCP control block, making an
646 * empty reassembly queue and hooking it to the argument
647 * protocol control block. The `inp' parameter must have
648 * come from the zone allocator set up in tcp_init().
651 tcp_newtcpcb(struct inpcb *inp)
656 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
658 const boolean_t isipv6 = FALSE;
661 it = (struct inp_tp *)inp;
663 bzero(tp, sizeof(struct tcpcb));
664 LIST_INIT(&tp->t_segq);
665 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
667 /* Set up our timeouts. */
668 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
669 callout_init(tp->tt_persist = &it->inp_tp_persist);
670 callout_init(tp->tt_keep = &it->inp_tp_keep);
671 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
672 callout_init(tp->tt_delack = &it->inp_tp_delack);
675 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
677 tp->t_flags |= TF_REQ_CC;
678 tp->t_inpcb = inp; /* XXX */
679 tp->t_state = TCPS_CLOSED;
681 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
682 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
683 * reasonable initial retransmit time.
685 tp->t_srtt = TCPTV_SRTTBASE;
687 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
688 tp->t_rttmin = tcp_rexmit_min;
689 tp->t_rxtcur = TCPTV_RTOBASE;
690 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
691 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
692 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
693 tp->t_rcvtime = ticks;
695 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
696 * because the socket may be bound to an IPv6 wildcard address,
697 * which may match an IPv4-mapped IPv6 address.
699 inp->inp_ip_ttl = ip_defttl;
701 tcp_sack_tcpcb_init(tp);
702 return (tp); /* XXX */
706 * Drop a TCP connection, reporting the specified error.
707 * If connection is synchronized, then send a RST to peer.
710 tcp_drop(struct tcpcb *tp, int errno)
712 struct socket *so = tp->t_inpcb->inp_socket;
714 if (TCPS_HAVERCVDSYN(tp->t_state)) {
715 tp->t_state = TCPS_CLOSED;
717 tcpstat.tcps_drops++;
719 tcpstat.tcps_conndrops++;
720 if (errno == ETIMEDOUT && tp->t_softerror)
721 errno = tp->t_softerror;
722 so->so_error = errno;
723 return (tcp_close(tp));
728 struct netmsg_remwildcard {
729 struct lwkt_msg nm_lmsg;
730 struct inpcb *nm_inp;
731 struct inpcbinfo *nm_pcbinfo;
740 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
741 * inp can be detached. We do this by cycling through the cpus, ending up
742 * on the cpu controlling the inp last and then doing the disconnect.
745 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
747 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
750 cpu = msg->nm_pcbinfo->cpu;
752 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
753 /* note: detach removes any wildcard hash entry */
756 in6_pcbdetach(msg->nm_inp);
759 in_pcbdetach(msg->nm_inp);
760 lwkt_replymsg(&msg->nm_lmsg, 0);
762 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
763 cpu = (cpu + 1) % ncpus2;
764 msg->nm_pcbinfo = &tcbinfo[cpu];
765 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg);
773 * Close a TCP control block:
774 * discard all space held by the tcp
775 * discard internet protocol block
776 * wake up any sleepers
779 tcp_close(struct tcpcb *tp)
782 struct inpcb *inp = tp->t_inpcb;
783 struct socket *so = inp->inp_socket;
785 boolean_t dosavessthresh;
790 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
791 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
793 const boolean_t isipv6 = FALSE;
797 * The tp is not instantly destroyed in the wildcard case. Setting
798 * the state to TCPS_TERMINATING will prevent the TCP stack from
799 * messing with it, though it should be noted that this change may
800 * not take effect on other cpus until we have chained the wildcard
803 * XXX we currently depend on the BGL to synchronize the tp->t_state
804 * update and prevent other tcp protocol threads from accepting new
805 * connections on the listen socket we might be trying to close down.
807 KKASSERT(tp->t_state != TCPS_TERMINATING);
808 tp->t_state = TCPS_TERMINATING;
811 * Make sure that all of our timers are stopped before we
814 callout_stop(tp->tt_rexmt);
815 callout_stop(tp->tt_persist);
816 callout_stop(tp->tt_keep);
817 callout_stop(tp->tt_2msl);
818 callout_stop(tp->tt_delack);
820 if (tp->t_flags & TF_ONOUTPUTQ) {
821 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
822 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
823 tp->t_flags &= ~TF_ONOUTPUTQ;
827 * If we got enough samples through the srtt filter,
828 * save the rtt and rttvar in the routing entry.
829 * 'Enough' is arbitrarily defined as the 16 samples.
830 * 16 samples is enough for the srtt filter to converge
831 * to within 5% of the correct value; fewer samples and
832 * we could save a very bogus rtt.
834 * Don't update the default route's characteristics and don't
835 * update anything that the user "locked".
837 if (tp->t_rttupdated >= 16) {
841 struct sockaddr_in6 *sin6;
843 if ((rt = inp->in6p_route.ro_rt) == NULL)
845 sin6 = (struct sockaddr_in6 *)rt_key(rt);
846 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
849 if ((rt = inp->inp_route.ro_rt) == NULL ||
850 ((struct sockaddr_in *)rt_key(rt))->
851 sin_addr.s_addr == INADDR_ANY)
854 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
855 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
856 if (rt->rt_rmx.rmx_rtt && i)
858 * filter this update to half the old & half
859 * the new values, converting scale.
860 * See route.h and tcp_var.h for a
861 * description of the scaling constants.
864 (rt->rt_rmx.rmx_rtt + i) / 2;
866 rt->rt_rmx.rmx_rtt = i;
867 tcpstat.tcps_cachedrtt++;
869 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
871 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
872 if (rt->rt_rmx.rmx_rttvar && i)
873 rt->rt_rmx.rmx_rttvar =
874 (rt->rt_rmx.rmx_rttvar + i) / 2;
876 rt->rt_rmx.rmx_rttvar = i;
877 tcpstat.tcps_cachedrttvar++;
880 * The old comment here said:
881 * update the pipelimit (ssthresh) if it has been updated
882 * already or if a pipesize was specified & the threshhold
883 * got below half the pipesize. I.e., wait for bad news
884 * before we start updating, then update on both good
887 * But we want to save the ssthresh even if no pipesize is
888 * specified explicitly in the route, because such
889 * connections still have an implicit pipesize specified
890 * by the global tcp_sendspace. In the absence of a reliable
891 * way to calculate the pipesize, it will have to do.
893 i = tp->snd_ssthresh;
894 if (rt->rt_rmx.rmx_sendpipe != 0)
895 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
897 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
898 if (dosavessthresh ||
899 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
900 (rt->rt_rmx.rmx_ssthresh != 0))) {
902 * convert the limit from user data bytes to
903 * packets then to packet data bytes.
905 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
910 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
911 sizeof(struct tcpiphdr));
912 if (rt->rt_rmx.rmx_ssthresh)
913 rt->rt_rmx.rmx_ssthresh =
914 (rt->rt_rmx.rmx_ssthresh + i) / 2;
916 rt->rt_rmx.rmx_ssthresh = i;
917 tcpstat.tcps_cachedssthresh++;
922 /* free the reassembly queue, if any */
923 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
924 LIST_REMOVE(q, tqe_q);
929 /* throw away SACK blocks in scoreboard*/
931 tcp_sack_cleanup(&tp->scb);
933 inp->inp_ppcb = NULL;
934 soisdisconnected(so);
936 * Discard the inp. In the SMP case a wildcard inp's hash (created
937 * by a listen socket or an INADDR_ANY udp socket) is replicated
938 * for each protocol thread and must be removed in the context of
939 * that thread. This is accomplished by chaining the message
942 * If the inp is not wildcarded we simply detach, which will remove
943 * the any hashes still present for this inp.
946 if (inp->inp_flags & INP_WILDCARD_MP) {
947 struct netmsg_remwildcard *msg;
949 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
950 msg = malloc(sizeof(struct netmsg_remwildcard),
951 M_LWKTMSG, M_INTWAIT);
952 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
953 lwkt_cmd_func(in_pcbremwildcardhash_handler),
956 msg->nm_isinet6 = isafinet6;
959 msg->nm_pcbinfo = &tcbinfo[cpu];
960 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
964 /* note: detach removes any wildcard hash entry */
972 tcpstat.tcps_closed++;
977 tcp_drain_oncpu(struct inpcbhead *head)
981 struct tseg_qent *te;
983 LIST_FOREACH(inpb, head, inp_list) {
984 if (inpb->inp_flags & INP_PLACEMARKER)
986 if ((tcpb = intotcpcb(inpb))) {
987 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
988 LIST_REMOVE(te, tqe_q);
998 struct netmsg_tcp_drain {
999 struct lwkt_msg nm_lmsg;
1000 struct inpcbhead *nm_head;
1004 tcp_drain_handler(lwkt_msg_t lmsg)
1006 struct netmsg_tcp_drain *nm = (void *)lmsg;
1008 tcp_drain_oncpu(nm->nm_head);
1009 lwkt_replymsg(lmsg, 0);
1025 * Walk the tcpbs, if existing, and flush the reassembly queue,
1026 * if there is one...
1027 * XXX: The "Net/3" implementation doesn't imply that the TCP
1028 * reassembly queue should be flushed, but in a situation
1029 * where we're really low on mbufs, this is potentially
1033 for (cpu = 0; cpu < ncpus2; cpu++) {
1034 struct netmsg_tcp_drain *msg;
1036 if (cpu == mycpu->gd_cpuid) {
1037 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1039 msg = malloc(sizeof(struct netmsg_tcp_drain),
1040 M_LWKTMSG, M_NOWAIT);
1043 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
1044 lwkt_cmd_func(tcp_drain_handler),
1046 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1047 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
1051 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1056 * Notify a tcp user of an asynchronous error;
1057 * store error as soft error, but wake up user
1058 * (for now, won't do anything until can select for soft error).
1060 * Do not wake up user since there currently is no mechanism for
1061 * reporting soft errors (yet - a kqueue filter may be added).
1064 tcp_notify(struct inpcb *inp, int error)
1066 struct tcpcb *tp = intotcpcb(inp);
1069 * Ignore some errors if we are hooked up.
1070 * If connection hasn't completed, has retransmitted several times,
1071 * and receives a second error, give up now. This is better
1072 * than waiting a long time to establish a connection that
1073 * can never complete.
1075 if (tp->t_state == TCPS_ESTABLISHED &&
1076 (error == EHOSTUNREACH || error == ENETUNREACH ||
1077 error == EHOSTDOWN)) {
1079 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1081 tcp_drop(tp, error);
1083 tp->t_softerror = error;
1085 wakeup(&so->so_timeo);
1092 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1095 struct inpcb *marker;
1105 * The process of preparing the TCB list is too time-consuming and
1106 * resource-intensive to repeat twice on every request.
1108 if (req->oldptr == NULL) {
1109 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1110 gd = globaldata_find(ccpu);
1111 n += tcbinfo[gd->gd_cpuid].ipi_count;
1113 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1117 if (req->newptr != NULL)
1120 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1121 marker->inp_flags |= INP_PLACEMARKER;
1124 * OK, now we're committed to doing something. Run the inpcb list
1125 * for each cpu in the system and construct the output. Use a
1126 * list placemarker to deal with list changes occuring during
1127 * copyout blockages (but otherwise depend on being on the correct
1128 * cpu to avoid races).
1130 origcpu = mycpu->gd_cpuid;
1131 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1137 cpu_id = (origcpu + ccpu) % ncpus;
1138 if ((smp_active_mask & (1 << cpu_id)) == 0)
1140 rgd = globaldata_find(cpu_id);
1141 lwkt_setcpu_self(rgd);
1143 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1144 n = tcbinfo[cpu_id].ipi_count;
1146 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1148 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1150 * process a snapshot of pcbs, ignoring placemarkers
1151 * and using our own to allow SYSCTL_OUT to block.
1153 LIST_REMOVE(marker, inp_list);
1154 LIST_INSERT_AFTER(inp, marker, inp_list);
1156 if (inp->inp_flags & INP_PLACEMARKER)
1158 if (inp->inp_gencnt > gencnt)
1160 if (prison_xinpcb(req->td, inp))
1163 xt.xt_len = sizeof xt;
1164 bcopy(inp, &xt.xt_inp, sizeof *inp);
1165 inp_ppcb = inp->inp_ppcb;
1166 if (inp_ppcb != NULL)
1167 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1169 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1170 if (inp->inp_socket)
1171 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1172 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1176 LIST_REMOVE(marker, inp_list);
1177 if (error == 0 && i < n) {
1178 bzero(&xt, sizeof xt);
1179 xt.xt_len = sizeof xt;
1181 error = SYSCTL_OUT(req, &xt, sizeof xt);
1190 * Make sure we are on the same cpu we were on originally, since
1191 * higher level callers expect this. Also don't pollute caches with
1192 * migrated userland data by (eventually) returning to userland
1193 * on a different cpu.
1195 lwkt_setcpu_self(globaldata_find(origcpu));
1196 free(marker, M_TEMP);
1200 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1201 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1204 tcp_getcred(SYSCTL_HANDLER_ARGS)
1206 struct sockaddr_in addrs[2];
1211 error = suser(req->td);
1214 error = SYSCTL_IN(req, addrs, sizeof addrs);
1218 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1219 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1220 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1221 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1222 if (inp == NULL || inp->inp_socket == NULL) {
1226 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1232 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1233 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1237 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1239 struct sockaddr_in6 addrs[2];
1242 boolean_t mapped = FALSE;
1244 error = suser(req->td);
1247 error = SYSCTL_IN(req, addrs, sizeof addrs);
1250 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1251 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1258 inp = in_pcblookup_hash(&tcbinfo[0],
1259 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1261 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1265 inp = in6_pcblookup_hash(&tcbinfo[0],
1266 &addrs[1].sin6_addr, addrs[1].sin6_port,
1267 &addrs[0].sin6_addr, addrs[0].sin6_port,
1270 if (inp == NULL || inp->inp_socket == NULL) {
1274 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1280 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1282 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1286 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1288 struct ip *ip = vip;
1290 struct in_addr faddr;
1293 void (*notify)(struct inpcb *, int) = tcp_notify;
1297 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1301 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1302 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1305 arg = inetctlerrmap[cmd];
1306 if (cmd == PRC_QUENCH) {
1307 notify = tcp_quench;
1308 } else if (icmp_may_rst &&
1309 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1310 cmd == PRC_UNREACH_PORT ||
1311 cmd == PRC_TIMXCEED_INTRANS) &&
1313 notify = tcp_drop_syn_sent;
1314 } else if (cmd == PRC_MSGSIZE) {
1315 struct icmp *icmp = (struct icmp *)
1316 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1318 arg = ntohs(icmp->icmp_nextmtu);
1319 notify = tcp_mtudisc;
1320 } else if (PRC_IS_REDIRECT(cmd)) {
1322 notify = in_rtchange;
1323 } else if (cmd == PRC_HOSTDEAD) {
1329 th = (struct tcphdr *)((caddr_t)ip +
1330 (IP_VHL_HL(ip->ip_vhl) << 2));
1331 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1332 ip->ip_src.s_addr, th->th_sport);
1333 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1334 ip->ip_src, th->th_sport, 0, NULL);
1335 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1336 icmpseq = htonl(th->th_seq);
1337 tp = intotcpcb(inp);
1338 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1339 SEQ_LT(icmpseq, tp->snd_max))
1340 (*notify)(inp, arg);
1342 struct in_conninfo inc;
1344 inc.inc_fport = th->th_dport;
1345 inc.inc_lport = th->th_sport;
1346 inc.inc_faddr = faddr;
1347 inc.inc_laddr = ip->ip_src;
1351 syncache_unreach(&inc, th);
1355 for (cpu = 0; cpu < ncpus2; cpu++) {
1356 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1364 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1367 void (*notify) (struct inpcb *, int) = tcp_notify;
1368 struct ip6_hdr *ip6;
1370 struct ip6ctlparam *ip6cp = NULL;
1371 const struct sockaddr_in6 *sa6_src = NULL;
1373 struct tcp_portonly {
1379 if (sa->sa_family != AF_INET6 ||
1380 sa->sa_len != sizeof(struct sockaddr_in6))
1384 if (cmd == PRC_QUENCH)
1385 notify = tcp_quench;
1386 else if (cmd == PRC_MSGSIZE) {
1387 struct ip6ctlparam *ip6cp = d;
1388 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1390 arg = ntohl(icmp6->icmp6_mtu);
1391 notify = tcp_mtudisc;
1392 } else if (!PRC_IS_REDIRECT(cmd) &&
1393 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1397 /* if the parameter is from icmp6, decode it. */
1399 ip6cp = (struct ip6ctlparam *)d;
1401 ip6 = ip6cp->ip6c_ip6;
1402 off = ip6cp->ip6c_off;
1403 sa6_src = ip6cp->ip6c_src;
1407 off = 0; /* fool gcc */
1412 struct in_conninfo inc;
1414 * XXX: We assume that when IPV6 is non NULL,
1415 * M and OFF are valid.
1418 /* check if we can safely examine src and dst ports */
1419 if (m->m_pkthdr.len < off + sizeof *thp)
1422 bzero(&th, sizeof th);
1423 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1425 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1426 (struct sockaddr *)ip6cp->ip6c_src,
1427 th.th_sport, cmd, arg, notify);
1429 inc.inc_fport = th.th_dport;
1430 inc.inc_lport = th.th_sport;
1431 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1432 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1434 syncache_unreach(&inc, &th);
1436 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1437 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1442 * Following is where TCP initial sequence number generation occurs.
1444 * There are two places where we must use initial sequence numbers:
1445 * 1. In SYN-ACK packets.
1446 * 2. In SYN packets.
1448 * All ISNs for SYN-ACK packets are generated by the syncache. See
1449 * tcp_syncache.c for details.
1451 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1452 * depends on this property. In addition, these ISNs should be
1453 * unguessable so as to prevent connection hijacking. To satisfy
1454 * the requirements of this situation, the algorithm outlined in
1455 * RFC 1948 is used to generate sequence numbers.
1457 * Implementation details:
1459 * Time is based off the system timer, and is corrected so that it
1460 * increases by one megabyte per second. This allows for proper
1461 * recycling on high speed LANs while still leaving over an hour
1464 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1465 * between seeding of isn_secret. This is normally set to zero,
1466 * as reseeding should not be necessary.
1470 #define ISN_BYTES_PER_SECOND 1048576
1472 u_char isn_secret[32];
1473 int isn_last_reseed;
1477 tcp_new_isn(struct tcpcb *tp)
1479 u_int32_t md5_buffer[4];
1482 /* Seed if this is the first use, reseed if requested. */
1483 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1484 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1486 read_random_unlimited(&isn_secret, sizeof isn_secret);
1487 isn_last_reseed = ticks;
1490 /* Compute the md5 hash and return the ISN. */
1492 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1493 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1495 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1496 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1497 sizeof(struct in6_addr));
1498 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1499 sizeof(struct in6_addr));
1503 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1504 sizeof(struct in_addr));
1505 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1506 sizeof(struct in_addr));
1508 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1509 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1510 new_isn = (tcp_seq) md5_buffer[0];
1511 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1516 * When a source quench is received, close congestion window
1517 * to one segment. We will gradually open it again as we proceed.
1520 tcp_quench(struct inpcb *inp, int errno)
1522 struct tcpcb *tp = intotcpcb(inp);
1525 tp->snd_cwnd = tp->t_maxseg;
1531 * When a specific ICMP unreachable message is received and the
1532 * connection state is SYN-SENT, drop the connection. This behavior
1533 * is controlled by the icmp_may_rst sysctl.
1536 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1538 struct tcpcb *tp = intotcpcb(inp);
1540 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1541 tcp_drop(tp, errno);
1545 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1546 * based on the new value in the route. Also nudge TCP to send something,
1547 * since we know the packet we just sent was dropped.
1548 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1551 tcp_mtudisc(struct inpcb *inp, int mtu)
1553 struct tcpcb *tp = intotcpcb(inp);
1555 struct socket *so = inp->inp_socket;
1558 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1560 const boolean_t isipv6 = FALSE;
1567 * If no MTU is provided in the ICMP message, use the
1568 * next lower likely value, as specified in RFC 1191.
1573 oldmtu = tp->t_maxopd +
1575 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1576 sizeof(struct tcpiphdr));
1577 mtu = ip_next_mtu(oldmtu, 0);
1581 rt = tcp_rtlookup6(&inp->inp_inc);
1583 rt = tcp_rtlookup(&inp->inp_inc);
1585 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1587 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1588 mtu = rt->rt_rmx.rmx_mtu;
1592 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1593 sizeof(struct tcpiphdr));
1596 * XXX - The following conditional probably violates the TCP
1597 * spec. The problem is that, since we don't know the
1598 * other end's MSS, we are supposed to use a conservative
1599 * default. But, if we do that, then MTU discovery will
1600 * never actually take place, because the conservative
1601 * default is much less than the MTUs typically seen
1602 * on the Internet today. For the moment, we'll sweep
1603 * this under the carpet.
1605 * The conservative default might not actually be a problem
1606 * if the only case this occurs is when sending an initial
1607 * SYN with options and data to a host we've never talked
1608 * to before. Then, they will reply with an MSS value which
1609 * will get recorded and the new parameters should get
1610 * recomputed. For Further Study.
1612 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1613 maxopd = taop->tao_mssopt;
1617 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1618 sizeof(struct tcpiphdr));
1620 if (tp->t_maxopd <= maxopd)
1622 tp->t_maxopd = maxopd;
1625 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1626 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1627 mss -= TCPOLEN_TSTAMP_APPA;
1629 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1630 (TF_REQ_CC | TF_RCVD_CC))
1631 mss -= TCPOLEN_CC_APPA;
1633 /* round down to multiple of MCLBYTES */
1634 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1636 mss &= ~(MCLBYTES - 1);
1639 mss = (mss / MCLBYTES) * MCLBYTES;
1642 if (so->so_snd.sb_hiwat < mss)
1643 mss = so->so_snd.sb_hiwat;
1647 tp->snd_nxt = tp->snd_una;
1649 tcpstat.tcps_mturesent++;
1653 * Look-up the routing entry to the peer of this inpcb. If no route
1654 * is found and it cannot be allocated the return NULL. This routine
1655 * is called by TCP routines that access the rmx structure and by tcp_mss
1656 * to get the interface MTU.
1659 tcp_rtlookup(struct in_conninfo *inc)
1661 struct route *ro = &inc->inc_route;
1663 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1664 /* No route yet, so try to acquire one */
1665 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1667 * unused portions of the structure MUST be zero'd
1668 * out because rtalloc() treats it as opaque data
1670 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1671 ro->ro_dst.sa_family = AF_INET;
1672 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1673 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1683 tcp_rtlookup6(struct in_conninfo *inc)
1685 struct route_in6 *ro6 = &inc->inc6_route;
1687 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1688 /* No route yet, so try to acquire one */
1689 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1691 * unused portions of the structure MUST be zero'd
1692 * out because rtalloc() treats it as opaque data
1694 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1695 ro6->ro_dst.sin6_family = AF_INET6;
1696 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1697 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1698 rtalloc((struct route *)ro6);
1701 return (ro6->ro_rt);
1706 /* compute ESP/AH header size for TCP, including outer IP header. */
1708 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1716 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1718 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1723 if (inp->inp_vflag & INP_IPV6) {
1724 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1726 th = (struct tcphdr *)(ip6 + 1);
1727 m->m_pkthdr.len = m->m_len =
1728 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1729 tcp_fillheaders(tp, ip6, th);
1730 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1734 ip = mtod(m, struct ip *);
1735 th = (struct tcphdr *)(ip + 1);
1736 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1737 tcp_fillheaders(tp, ip, th);
1738 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1747 * Return a pointer to the cached information about the remote host.
1748 * The cached information is stored in the protocol specific part of
1749 * the route metrics.
1752 tcp_gettaocache(struct in_conninfo *inc)
1757 if (inc->inc_isipv6)
1758 rt = tcp_rtlookup6(inc);
1761 rt = tcp_rtlookup(inc);
1763 /* Make sure this is a host route and is up. */
1765 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1768 return (rmx_taop(rt->rt_rmx));
1772 * Clear all the TAO cache entries, called from tcp_init.
1775 * This routine is just an empty one, because we assume that the routing
1776 * routing tables are initialized at the same time when TCP, so there is
1777 * nothing in the cache left over.
1785 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1787 * This code attempts to calculate the bandwidth-delay product as a
1788 * means of determining the optimal window size to maximize bandwidth,
1789 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1790 * routers. This code also does a fairly good job keeping RTTs in check
1791 * across slow links like modems. We implement an algorithm which is very
1792 * similar (but not meant to be) TCP/Vegas. The code operates on the
1793 * transmitter side of a TCP connection and so only effects the transmit
1794 * side of the connection.
1796 * BACKGROUND: TCP makes no provision for the management of buffer space
1797 * at the end points or at the intermediate routers and switches. A TCP
1798 * stream, whether using NewReno or not, will eventually buffer as
1799 * many packets as it is able and the only reason this typically works is
1800 * due to the fairly small default buffers made available for a connection
1801 * (typicaly 16K or 32K). As machines use larger windows and/or window
1802 * scaling it is now fairly easy for even a single TCP connection to blow-out
1803 * all available buffer space not only on the local interface, but on
1804 * intermediate routers and switches as well. NewReno makes a misguided
1805 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1806 * then backing off, then steadily increasing the window again until another
1807 * failure occurs, ad-infinitum. This results in terrible oscillation that
1808 * is only made worse as network loads increase and the idea of intentionally
1809 * blowing out network buffers is, frankly, a terrible way to manage network
1812 * It is far better to limit the transmit window prior to the failure
1813 * condition being achieved. There are two general ways to do this: First
1814 * you can 'scan' through different transmit window sizes and locate the
1815 * point where the RTT stops increasing, indicating that you have filled the
1816 * pipe, then scan backwards until you note that RTT stops decreasing, then
1817 * repeat ad-infinitum. This method works in principle but has severe
1818 * implementation issues due to RTT variances, timer granularity, and
1819 * instability in the algorithm which can lead to many false positives and
1820 * create oscillations as well as interact badly with other TCP streams
1821 * implementing the same algorithm.
1823 * The second method is to limit the window to the bandwidth delay product
1824 * of the link. This is the method we implement. RTT variances and our
1825 * own manipulation of the congestion window, bwnd, can potentially
1826 * destabilize the algorithm. For this reason we have to stabilize the
1827 * elements used to calculate the window. We do this by using the minimum
1828 * observed RTT, the long term average of the observed bandwidth, and
1829 * by adding two segments worth of slop. It isn't perfect but it is able
1830 * to react to changing conditions and gives us a very stable basis on
1831 * which to extend the algorithm.
1834 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1842 * If inflight_enable is disabled in the middle of a tcp connection,
1843 * make sure snd_bwnd is effectively disabled.
1845 if (!tcp_inflight_enable) {
1846 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1847 tp->snd_bandwidth = 0;
1852 * Validate the delta time. If a connection is new or has been idle
1853 * a long time we have to reset the bandwidth calculator.
1856 delta_ticks = save_ticks - tp->t_bw_rtttime;
1857 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1858 tp->t_bw_rtttime = ticks;
1859 tp->t_bw_rtseq = ack_seq;
1860 if (tp->snd_bandwidth == 0)
1861 tp->snd_bandwidth = tcp_inflight_min;
1864 if (delta_ticks == 0)
1868 * Sanity check, plus ignore pure window update acks.
1870 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1874 * Figure out the bandwidth. Due to the tick granularity this
1875 * is a very rough number and it MUST be averaged over a fairly
1876 * long period of time. XXX we need to take into account a link
1877 * that is not using all available bandwidth, but for now our
1878 * slop will ramp us up if this case occurs and the bandwidth later
1881 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1882 tp->t_bw_rtttime = save_ticks;
1883 tp->t_bw_rtseq = ack_seq;
1884 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1886 tp->snd_bandwidth = bw;
1889 * Calculate the semi-static bandwidth delay product, plus two maximal
1890 * segments. The additional slop puts us squarely in the sweet
1891 * spot and also handles the bandwidth run-up case. Without the
1892 * slop we could be locking ourselves into a lower bandwidth.
1894 * Situations Handled:
1895 * (1) Prevents over-queueing of packets on LANs, especially on
1896 * high speed LANs, allowing larger TCP buffers to be
1897 * specified, and also does a good job preventing
1898 * over-queueing of packets over choke points like modems
1899 * (at least for the transmit side).
1901 * (2) Is able to handle changing network loads (bandwidth
1902 * drops so bwnd drops, bandwidth increases so bwnd
1905 * (3) Theoretically should stabilize in the face of multiple
1906 * connections implementing the same algorithm (this may need
1909 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1910 * be adjusted with a sysctl but typically only needs to be on
1911 * very slow connections. A value no smaller then 5 should
1912 * be used, but only reduce this default if you have no other
1916 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1917 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1918 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1921 if (tcp_inflight_debug > 0) {
1923 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1925 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1926 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1929 if ((long)bwnd < tcp_inflight_min)
1930 bwnd = tcp_inflight_min;
1931 if (bwnd > tcp_inflight_max)
1932 bwnd = tcp_inflight_max;
1933 if ((long)bwnd < tp->t_maxseg * 2)
1934 bwnd = tp->t_maxseg * 2;
1935 tp->snd_bwnd = bwnd;