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 1.45 2005/03/06 05:09:25 hsu 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_ary[MAXCPU];
246 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
250 for (cpu = 0; cpu < ncpus; ++cpu) {
251 if ((error = SYSCTL_OUT(req, &tcpstats_ary[cpu],
252 sizeof(struct tcp_stats))))
254 if ((error = SYSCTL_IN(req, &tcpstats_ary[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.
371 * It is layed out as an array which is has one element for UP,
372 * and SMP_MAXCPU elements for SMP. This allows us to retain
373 * the access mechanism from userland for both UP and SMP.
376 for (cpu = 0; cpu < ncpus; ++cpu) {
377 bzero(&tcpstats_ary[cpu], sizeof(struct tcp_stats));
380 bzero(&tcpstat, sizeof(struct tcp_stats));
389 tcpmsg_service_loop(void *dummy)
393 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
395 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg);
396 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
405 int cpu = mycpu->gd_cpuid;
407 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
408 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
409 tp->t_flags &= ~TF_ONOUTPUTQ;
410 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
417 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
418 * tcp_template used to store this data in mbufs, but we now recopy it out
419 * of the tcpcb each time to conserve mbufs.
422 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
424 struct inpcb *inp = tp->t_inpcb;
425 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
428 if (inp->inp_vflag & INP_IPV6) {
431 ip6 = (struct ip6_hdr *)ip_ptr;
432 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
433 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
434 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
435 (IPV6_VERSION & IPV6_VERSION_MASK);
436 ip6->ip6_nxt = IPPROTO_TCP;
437 ip6->ip6_plen = sizeof(struct tcphdr);
438 ip6->ip6_src = inp->in6p_laddr;
439 ip6->ip6_dst = inp->in6p_faddr;
444 struct ip *ip = (struct ip *) ip_ptr;
446 ip->ip_vhl = IP_VHL_BORING;
453 ip->ip_p = IPPROTO_TCP;
454 ip->ip_src = inp->inp_laddr;
455 ip->ip_dst = inp->inp_faddr;
456 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
458 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
461 tcp_hdr->th_sport = inp->inp_lport;
462 tcp_hdr->th_dport = inp->inp_fport;
467 tcp_hdr->th_flags = 0;
473 * Create template to be used to send tcp packets on a connection.
474 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
475 * use for this function is in keepalives, which use tcp_respond.
478 tcp_maketemplate(struct tcpcb *tp)
482 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
484 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
489 tcp_freetemplate(struct tcptemp *tmp)
491 mpipe_free(&tcptemp_mpipe, tmp);
495 * Send a single message to the TCP at address specified by
496 * the given TCP/IP header. If m == NULL, then we make a copy
497 * of the tcpiphdr at ti and send directly to the addressed host.
498 * This is used to force keep alive messages out using the TCP
499 * template for a connection. If flags are given then we send
500 * a message back to the TCP which originated the * segment ti,
501 * and discard the mbuf containing it and any other attached mbufs.
503 * In any case the ack and sequence number of the transmitted
504 * segment are as specified by the parameters.
506 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
509 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
510 tcp_seq ack, tcp_seq seq, int flags)
514 struct route *ro = NULL;
516 struct ip *ip = ipgen;
519 struct route_in6 *ro6 = NULL;
520 struct route_in6 sro6;
521 struct ip6_hdr *ip6 = ipgen;
523 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
525 const boolean_t isipv6 = FALSE;
529 if (!(flags & TH_RST)) {
530 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
531 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
532 win = (long)TCP_MAXWIN << tp->rcv_scale;
535 ro6 = &tp->t_inpcb->in6p_route;
537 ro = &tp->t_inpcb->inp_route;
541 bzero(ro6, sizeof *ro6);
544 bzero(ro, sizeof *ro);
548 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
552 m->m_data += max_linkhdr;
554 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
555 ip6 = mtod(m, struct ip6_hdr *);
556 nth = (struct tcphdr *)(ip6 + 1);
558 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
559 ip = mtod(m, struct ip *);
560 nth = (struct tcphdr *)(ip + 1);
562 bcopy(th, nth, sizeof(struct tcphdr));
567 m->m_data = (caddr_t)ipgen;
568 /* m_len is set later */
570 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
572 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
573 nth = (struct tcphdr *)(ip6 + 1);
575 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
576 nth = (struct tcphdr *)(ip + 1);
580 * this is usually a case when an extension header
581 * exists between the IPv6 header and the
584 nth->th_sport = th->th_sport;
585 nth->th_dport = th->th_dport;
587 xchg(nth->th_dport, nth->th_sport, n_short);
592 ip6->ip6_vfc = IPV6_VERSION;
593 ip6->ip6_nxt = IPPROTO_TCP;
594 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
595 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
597 tlen += sizeof(struct tcpiphdr);
599 ip->ip_ttl = ip_defttl;
602 m->m_pkthdr.len = tlen;
603 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
604 nth->th_seq = htonl(seq);
605 nth->th_ack = htonl(ack);
607 nth->th_off = sizeof(struct tcphdr) >> 2;
608 nth->th_flags = flags;
610 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
612 nth->th_win = htons((u_short)win);
616 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
617 sizeof(struct ip6_hdr),
618 tlen - sizeof(struct ip6_hdr));
619 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
620 (ro6 && ro6->ro_rt) ?
621 ro6->ro_rt->rt_ifp : NULL);
623 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
624 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
625 m->m_pkthdr.csum_flags = CSUM_TCP;
626 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
629 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
630 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
633 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
634 tp ? tp->t_inpcb : NULL);
635 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
640 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
641 if ((ro == &sro) && (ro->ro_rt != NULL)) {
649 * Create a new TCP control block, making an
650 * empty reassembly queue and hooking it to the argument
651 * protocol control block. The `inp' parameter must have
652 * come from the zone allocator set up in tcp_init().
655 tcp_newtcpcb(struct inpcb *inp)
660 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
662 const boolean_t isipv6 = FALSE;
665 it = (struct inp_tp *)inp;
667 bzero(tp, sizeof(struct tcpcb));
668 LIST_INIT(&tp->t_segq);
669 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
671 /* Set up our timeouts. */
672 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
673 callout_init(tp->tt_persist = &it->inp_tp_persist);
674 callout_init(tp->tt_keep = &it->inp_tp_keep);
675 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
676 callout_init(tp->tt_delack = &it->inp_tp_delack);
679 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
681 tp->t_flags |= TF_REQ_CC;
682 tp->t_inpcb = inp; /* XXX */
683 tp->t_state = TCPS_CLOSED;
685 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
686 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
687 * reasonable initial retransmit time.
689 tp->t_srtt = TCPTV_SRTTBASE;
691 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
692 tp->t_rttmin = tcp_rexmit_min;
693 tp->t_rxtcur = TCPTV_RTOBASE;
694 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
695 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
696 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
697 tp->t_rcvtime = ticks;
699 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
700 * because the socket may be bound to an IPv6 wildcard address,
701 * which may match an IPv4-mapped IPv6 address.
703 inp->inp_ip_ttl = ip_defttl;
705 tcp_sack_tcpcb_init(tp);
706 return (tp); /* XXX */
710 * Drop a TCP connection, reporting the specified error.
711 * If connection is synchronized, then send a RST to peer.
714 tcp_drop(struct tcpcb *tp, int errno)
716 struct socket *so = tp->t_inpcb->inp_socket;
718 if (TCPS_HAVERCVDSYN(tp->t_state)) {
719 tp->t_state = TCPS_CLOSED;
721 tcpstat.tcps_drops++;
723 tcpstat.tcps_conndrops++;
724 if (errno == ETIMEDOUT && tp->t_softerror)
725 errno = tp->t_softerror;
726 so->so_error = errno;
727 return (tcp_close(tp));
732 struct netmsg_remwildcard {
733 struct lwkt_msg nm_lmsg;
734 struct inpcb *nm_inp;
735 struct inpcbinfo *nm_pcbinfo;
744 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
745 * inp can be detached. We do this by cycling through the cpus, ending up
746 * on the cpu controlling the inp last and then doing the disconnect.
749 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
751 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
754 cpu = msg->nm_pcbinfo->cpu;
756 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
757 /* note: detach removes any wildcard hash entry */
760 in6_pcbdetach(msg->nm_inp);
763 in_pcbdetach(msg->nm_inp);
764 lwkt_replymsg(&msg->nm_lmsg, 0);
766 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
767 cpu = (cpu + 1) % ncpus2;
768 msg->nm_pcbinfo = &tcbinfo[cpu];
769 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_lmsg);
777 * Close a TCP control block:
778 * discard all space held by the tcp
779 * discard internet protocol block
780 * wake up any sleepers
783 tcp_close(struct tcpcb *tp)
786 struct inpcb *inp = tp->t_inpcb;
787 struct socket *so = inp->inp_socket;
789 boolean_t dosavessthresh;
794 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
795 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
797 const boolean_t isipv6 = FALSE;
801 * The tp is not instantly destroyed in the wildcard case. Setting
802 * the state to TCPS_TERMINATING will prevent the TCP stack from
803 * messing with it, though it should be noted that this change may
804 * not take effect on other cpus until we have chained the wildcard
807 * XXX we currently depend on the BGL to synchronize the tp->t_state
808 * update and prevent other tcp protocol threads from accepting new
809 * connections on the listen socket we might be trying to close down.
811 KKASSERT(tp->t_state != TCPS_TERMINATING);
812 tp->t_state = TCPS_TERMINATING;
815 * Make sure that all of our timers are stopped before we
818 callout_stop(tp->tt_rexmt);
819 callout_stop(tp->tt_persist);
820 callout_stop(tp->tt_keep);
821 callout_stop(tp->tt_2msl);
822 callout_stop(tp->tt_delack);
824 if (tp->t_flags & TF_ONOUTPUTQ) {
825 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
826 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
827 tp->t_flags &= ~TF_ONOUTPUTQ;
831 * If we got enough samples through the srtt filter,
832 * save the rtt and rttvar in the routing entry.
833 * 'Enough' is arbitrarily defined as the 16 samples.
834 * 16 samples is enough for the srtt filter to converge
835 * to within 5% of the correct value; fewer samples and
836 * we could save a very bogus rtt.
838 * Don't update the default route's characteristics and don't
839 * update anything that the user "locked".
841 if (tp->t_rttupdated >= 16) {
845 struct sockaddr_in6 *sin6;
847 if ((rt = inp->in6p_route.ro_rt) == NULL)
849 sin6 = (struct sockaddr_in6 *)rt_key(rt);
850 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
853 if ((rt = inp->inp_route.ro_rt) == NULL ||
854 ((struct sockaddr_in *)rt_key(rt))->
855 sin_addr.s_addr == INADDR_ANY)
858 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
859 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
860 if (rt->rt_rmx.rmx_rtt && i)
862 * filter this update to half the old & half
863 * the new values, converting scale.
864 * See route.h and tcp_var.h for a
865 * description of the scaling constants.
868 (rt->rt_rmx.rmx_rtt + i) / 2;
870 rt->rt_rmx.rmx_rtt = i;
871 tcpstat.tcps_cachedrtt++;
873 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
875 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
876 if (rt->rt_rmx.rmx_rttvar && i)
877 rt->rt_rmx.rmx_rttvar =
878 (rt->rt_rmx.rmx_rttvar + i) / 2;
880 rt->rt_rmx.rmx_rttvar = i;
881 tcpstat.tcps_cachedrttvar++;
884 * The old comment here said:
885 * update the pipelimit (ssthresh) if it has been updated
886 * already or if a pipesize was specified & the threshhold
887 * got below half the pipesize. I.e., wait for bad news
888 * before we start updating, then update on both good
891 * But we want to save the ssthresh even if no pipesize is
892 * specified explicitly in the route, because such
893 * connections still have an implicit pipesize specified
894 * by the global tcp_sendspace. In the absence of a reliable
895 * way to calculate the pipesize, it will have to do.
897 i = tp->snd_ssthresh;
898 if (rt->rt_rmx.rmx_sendpipe != 0)
899 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
901 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
902 if (dosavessthresh ||
903 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
904 (rt->rt_rmx.rmx_ssthresh != 0))) {
906 * convert the limit from user data bytes to
907 * packets then to packet data bytes.
909 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
914 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
915 sizeof(struct tcpiphdr));
916 if (rt->rt_rmx.rmx_ssthresh)
917 rt->rt_rmx.rmx_ssthresh =
918 (rt->rt_rmx.rmx_ssthresh + i) / 2;
920 rt->rt_rmx.rmx_ssthresh = i;
921 tcpstat.tcps_cachedssthresh++;
926 /* free the reassembly queue, if any */
927 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
928 LIST_REMOVE(q, tqe_q);
933 /* throw away SACK blocks in scoreboard*/
935 tcp_sack_cleanup(&tp->scb);
937 inp->inp_ppcb = NULL;
938 soisdisconnected(so);
940 * Discard the inp. In the SMP case a wildcard inp's hash (created
941 * by a listen socket or an INADDR_ANY udp socket) is replicated
942 * for each protocol thread and must be removed in the context of
943 * that thread. This is accomplished by chaining the message
946 * If the inp is not wildcarded we simply detach, which will remove
947 * the any hashes still present for this inp.
950 if (inp->inp_flags & INP_WILDCARD_MP) {
951 struct netmsg_remwildcard *msg;
953 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
954 msg = malloc(sizeof(struct netmsg_remwildcard),
955 M_LWKTMSG, M_INTWAIT);
956 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
957 lwkt_cmd_func(in_pcbremwildcardhash_handler),
960 msg->nm_isinet6 = isafinet6;
963 msg->nm_pcbinfo = &tcbinfo[cpu];
964 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
968 /* note: detach removes any wildcard hash entry */
976 tcpstat.tcps_closed++;
981 tcp_drain_oncpu(struct inpcbhead *head)
985 struct tseg_qent *te;
987 LIST_FOREACH(inpb, head, inp_list) {
988 if (inpb->inp_flags & INP_PLACEMARKER)
990 if ((tcpb = intotcpcb(inpb))) {
991 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
992 LIST_REMOVE(te, tqe_q);
1002 struct netmsg_tcp_drain {
1003 struct lwkt_msg nm_lmsg;
1004 struct inpcbhead *nm_head;
1008 tcp_drain_handler(lwkt_msg_t lmsg)
1010 struct netmsg_tcp_drain *nm = (void *)lmsg;
1012 tcp_drain_oncpu(nm->nm_head);
1013 lwkt_replymsg(lmsg, 0);
1029 * Walk the tcpbs, if existing, and flush the reassembly queue,
1030 * if there is one...
1031 * XXX: The "Net/3" implementation doesn't imply that the TCP
1032 * reassembly queue should be flushed, but in a situation
1033 * where we're really low on mbufs, this is potentially
1037 for (cpu = 0; cpu < ncpus2; cpu++) {
1038 struct netmsg_tcp_drain *msg;
1040 if (cpu == mycpu->gd_cpuid) {
1041 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1043 msg = malloc(sizeof(struct netmsg_tcp_drain),
1044 M_LWKTMSG, M_NOWAIT);
1047 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
1048 lwkt_cmd_func(tcp_drain_handler),
1050 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1051 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
1055 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1060 * Notify a tcp user of an asynchronous error;
1061 * store error as soft error, but wake up user
1062 * (for now, won't do anything until can select for soft error).
1064 * Do not wake up user since there currently is no mechanism for
1065 * reporting soft errors (yet - a kqueue filter may be added).
1068 tcp_notify(struct inpcb *inp, int error)
1070 struct tcpcb *tp = intotcpcb(inp);
1073 * Ignore some errors if we are hooked up.
1074 * If connection hasn't completed, has retransmitted several times,
1075 * and receives a second error, give up now. This is better
1076 * than waiting a long time to establish a connection that
1077 * can never complete.
1079 if (tp->t_state == TCPS_ESTABLISHED &&
1080 (error == EHOSTUNREACH || error == ENETUNREACH ||
1081 error == EHOSTDOWN)) {
1083 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1085 tcp_drop(tp, error);
1087 tp->t_softerror = error;
1089 wakeup(&so->so_timeo);
1096 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1099 struct inpcb *marker;
1109 * The process of preparing the TCB list is too time-consuming and
1110 * resource-intensive to repeat twice on every request.
1112 if (req->oldptr == NULL) {
1113 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1114 gd = globaldata_find(ccpu);
1115 n += tcbinfo[gd->gd_cpuid].ipi_count;
1117 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1121 if (req->newptr != NULL)
1124 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1125 marker->inp_flags |= INP_PLACEMARKER;
1128 * OK, now we're committed to doing something. Run the inpcb list
1129 * for each cpu in the system and construct the output. Use a
1130 * list placemarker to deal with list changes occuring during
1131 * copyout blockages (but otherwise depend on being on the correct
1132 * cpu to avoid races).
1134 origcpu = mycpu->gd_cpuid;
1135 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1141 cpu_id = (origcpu + ccpu) % ncpus;
1142 if ((smp_active_mask & (1 << cpu_id)) == 0)
1144 rgd = globaldata_find(cpu_id);
1145 lwkt_setcpu_self(rgd);
1147 /* indicate change of CPU */
1150 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1151 n = tcbinfo[cpu_id].ipi_count;
1153 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1155 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1157 * process a snapshot of pcbs, ignoring placemarkers
1158 * and using our own to allow SYSCTL_OUT to block.
1160 LIST_REMOVE(marker, inp_list);
1161 LIST_INSERT_AFTER(inp, marker, inp_list);
1163 if (inp->inp_flags & INP_PLACEMARKER)
1165 if (inp->inp_gencnt > gencnt)
1167 if (prison_xinpcb(req->td, inp))
1170 xt.xt_len = sizeof xt;
1171 bcopy(inp, &xt.xt_inp, sizeof *inp);
1172 inp_ppcb = inp->inp_ppcb;
1173 if (inp_ppcb != NULL)
1174 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1176 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1177 if (inp->inp_socket)
1178 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1179 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1183 LIST_REMOVE(marker, inp_list);
1184 if (error == 0 && i < n) {
1185 bzero(&xt, sizeof xt);
1186 xt.xt_len = sizeof xt;
1188 error = SYSCTL_OUT(req, &xt, sizeof xt);
1197 * Make sure we are on the same cpu we were on originally, since
1198 * higher level callers expect this. Also don't pollute caches with
1199 * migrated userland data by (eventually) returning to userland
1200 * on a different cpu.
1202 lwkt_setcpu_self(globaldata_find(origcpu));
1203 free(marker, M_TEMP);
1207 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1208 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1211 tcp_getcred(SYSCTL_HANDLER_ARGS)
1213 struct sockaddr_in addrs[2];
1218 error = suser(req->td);
1221 error = SYSCTL_IN(req, addrs, sizeof addrs);
1226 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1227 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1228 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1229 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1230 if (inp == NULL || inp->inp_socket == NULL) {
1234 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1240 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1241 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1245 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1247 struct sockaddr_in6 addrs[2];
1250 boolean_t mapped = FALSE;
1252 error = suser(req->td);
1255 error = SYSCTL_IN(req, addrs, sizeof addrs);
1258 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1259 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1266 inp = in_pcblookup_hash(&tcbinfo[0],
1267 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1269 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1273 inp = in6_pcblookup_hash(&tcbinfo[0],
1274 &addrs[1].sin6_addr, addrs[1].sin6_port,
1275 &addrs[0].sin6_addr, addrs[0].sin6_port,
1278 if (inp == NULL || inp->inp_socket == NULL) {
1282 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1288 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1290 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1294 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1296 struct ip *ip = vip;
1298 struct in_addr faddr;
1301 void (*notify)(struct inpcb *, int) = tcp_notify;
1305 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1309 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1310 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1313 arg = inetctlerrmap[cmd];
1314 if (cmd == PRC_QUENCH) {
1315 notify = tcp_quench;
1316 } else if (icmp_may_rst &&
1317 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1318 cmd == PRC_UNREACH_PORT ||
1319 cmd == PRC_TIMXCEED_INTRANS) &&
1321 notify = tcp_drop_syn_sent;
1322 } else if (cmd == PRC_MSGSIZE) {
1323 struct icmp *icmp = (struct icmp *)
1324 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1326 arg = ntohs(icmp->icmp_nextmtu);
1327 notify = tcp_mtudisc;
1328 } else if (PRC_IS_REDIRECT(cmd)) {
1330 notify = in_rtchange;
1331 } else if (cmd == PRC_HOSTDEAD) {
1337 th = (struct tcphdr *)((caddr_t)ip +
1338 (IP_VHL_HL(ip->ip_vhl) << 2));
1339 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1340 ip->ip_src.s_addr, th->th_sport);
1341 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1342 ip->ip_src, th->th_sport, 0, NULL);
1343 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1344 icmpseq = htonl(th->th_seq);
1345 tp = intotcpcb(inp);
1346 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1347 SEQ_LT(icmpseq, tp->snd_max))
1348 (*notify)(inp, arg);
1350 struct in_conninfo inc;
1352 inc.inc_fport = th->th_dport;
1353 inc.inc_lport = th->th_sport;
1354 inc.inc_faddr = faddr;
1355 inc.inc_laddr = ip->ip_src;
1359 syncache_unreach(&inc, th);
1363 for (cpu = 0; cpu < ncpus2; cpu++) {
1364 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr, arg,
1372 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1375 void (*notify) (struct inpcb *, int) = tcp_notify;
1376 struct ip6_hdr *ip6;
1378 struct ip6ctlparam *ip6cp = NULL;
1379 const struct sockaddr_in6 *sa6_src = NULL;
1381 struct tcp_portonly {
1387 if (sa->sa_family != AF_INET6 ||
1388 sa->sa_len != sizeof(struct sockaddr_in6))
1392 if (cmd == PRC_QUENCH)
1393 notify = tcp_quench;
1394 else if (cmd == PRC_MSGSIZE) {
1395 struct ip6ctlparam *ip6cp = d;
1396 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1398 arg = ntohl(icmp6->icmp6_mtu);
1399 notify = tcp_mtudisc;
1400 } else if (!PRC_IS_REDIRECT(cmd) &&
1401 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1405 /* if the parameter is from icmp6, decode it. */
1407 ip6cp = (struct ip6ctlparam *)d;
1409 ip6 = ip6cp->ip6c_ip6;
1410 off = ip6cp->ip6c_off;
1411 sa6_src = ip6cp->ip6c_src;
1415 off = 0; /* fool gcc */
1420 struct in_conninfo inc;
1422 * XXX: We assume that when IPV6 is non NULL,
1423 * M and OFF are valid.
1426 /* check if we can safely examine src and dst ports */
1427 if (m->m_pkthdr.len < off + sizeof *thp)
1430 bzero(&th, sizeof th);
1431 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1433 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1434 (struct sockaddr *)ip6cp->ip6c_src,
1435 th.th_sport, cmd, arg, notify);
1437 inc.inc_fport = th.th_dport;
1438 inc.inc_lport = th.th_sport;
1439 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1440 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1442 syncache_unreach(&inc, &th);
1444 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1445 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1450 * Following is where TCP initial sequence number generation occurs.
1452 * There are two places where we must use initial sequence numbers:
1453 * 1. In SYN-ACK packets.
1454 * 2. In SYN packets.
1456 * All ISNs for SYN-ACK packets are generated by the syncache. See
1457 * tcp_syncache.c for details.
1459 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1460 * depends on this property. In addition, these ISNs should be
1461 * unguessable so as to prevent connection hijacking. To satisfy
1462 * the requirements of this situation, the algorithm outlined in
1463 * RFC 1948 is used to generate sequence numbers.
1465 * Implementation details:
1467 * Time is based off the system timer, and is corrected so that it
1468 * increases by one megabyte per second. This allows for proper
1469 * recycling on high speed LANs while still leaving over an hour
1472 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1473 * between seeding of isn_secret. This is normally set to zero,
1474 * as reseeding should not be necessary.
1478 #define ISN_BYTES_PER_SECOND 1048576
1480 u_char isn_secret[32];
1481 int isn_last_reseed;
1485 tcp_new_isn(struct tcpcb *tp)
1487 u_int32_t md5_buffer[4];
1490 /* Seed if this is the first use, reseed if requested. */
1491 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1492 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1494 read_random_unlimited(&isn_secret, sizeof isn_secret);
1495 isn_last_reseed = ticks;
1498 /* Compute the md5 hash and return the ISN. */
1500 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1501 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1503 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1504 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1505 sizeof(struct in6_addr));
1506 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1507 sizeof(struct in6_addr));
1511 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1512 sizeof(struct in_addr));
1513 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1514 sizeof(struct in_addr));
1516 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1517 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1518 new_isn = (tcp_seq) md5_buffer[0];
1519 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1524 * When a source quench is received, close congestion window
1525 * to one segment. We will gradually open it again as we proceed.
1528 tcp_quench(struct inpcb *inp, int errno)
1530 struct tcpcb *tp = intotcpcb(inp);
1533 tp->snd_cwnd = tp->t_maxseg;
1537 * When a specific ICMP unreachable message is received and the
1538 * connection state is SYN-SENT, drop the connection. This behavior
1539 * is controlled by the icmp_may_rst sysctl.
1542 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1544 struct tcpcb *tp = intotcpcb(inp);
1546 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1547 tcp_drop(tp, errno);
1551 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1552 * based on the new value in the route. Also nudge TCP to send something,
1553 * since we know the packet we just sent was dropped.
1554 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1557 tcp_mtudisc(struct inpcb *inp, int mtu)
1559 struct tcpcb *tp = intotcpcb(inp);
1561 struct socket *so = inp->inp_socket;
1564 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1566 const boolean_t isipv6 = FALSE;
1573 * If no MTU is provided in the ICMP message, use the
1574 * next lower likely value, as specified in RFC 1191.
1579 oldmtu = tp->t_maxopd +
1581 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1582 sizeof(struct tcpiphdr));
1583 mtu = ip_next_mtu(oldmtu, 0);
1587 rt = tcp_rtlookup6(&inp->inp_inc);
1589 rt = tcp_rtlookup(&inp->inp_inc);
1591 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1593 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1594 mtu = rt->rt_rmx.rmx_mtu;
1598 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1599 sizeof(struct tcpiphdr));
1602 * XXX - The following conditional probably violates the TCP
1603 * spec. The problem is that, since we don't know the
1604 * other end's MSS, we are supposed to use a conservative
1605 * default. But, if we do that, then MTU discovery will
1606 * never actually take place, because the conservative
1607 * default is much less than the MTUs typically seen
1608 * on the Internet today. For the moment, we'll sweep
1609 * this under the carpet.
1611 * The conservative default might not actually be a problem
1612 * if the only case this occurs is when sending an initial
1613 * SYN with options and data to a host we've never talked
1614 * to before. Then, they will reply with an MSS value which
1615 * will get recorded and the new parameters should get
1616 * recomputed. For Further Study.
1618 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1619 maxopd = taop->tao_mssopt;
1623 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1624 sizeof(struct tcpiphdr));
1626 if (tp->t_maxopd <= maxopd)
1628 tp->t_maxopd = maxopd;
1631 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1632 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1633 mss -= TCPOLEN_TSTAMP_APPA;
1635 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1636 (TF_REQ_CC | TF_RCVD_CC))
1637 mss -= TCPOLEN_CC_APPA;
1639 /* round down to multiple of MCLBYTES */
1640 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1642 mss &= ~(MCLBYTES - 1);
1645 mss = (mss / MCLBYTES) * MCLBYTES;
1648 if (so->so_snd.sb_hiwat < mss)
1649 mss = so->so_snd.sb_hiwat;
1653 tp->snd_nxt = tp->snd_una;
1655 tcpstat.tcps_mturesent++;
1659 * Look-up the routing entry to the peer of this inpcb. If no route
1660 * is found and it cannot be allocated the return NULL. This routine
1661 * is called by TCP routines that access the rmx structure and by tcp_mss
1662 * to get the interface MTU.
1665 tcp_rtlookup(struct in_conninfo *inc)
1667 struct route *ro = &inc->inc_route;
1669 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1670 /* No route yet, so try to acquire one */
1671 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1673 * unused portions of the structure MUST be zero'd
1674 * out because rtalloc() treats it as opaque data
1676 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1677 ro->ro_dst.sa_family = AF_INET;
1678 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1679 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1689 tcp_rtlookup6(struct in_conninfo *inc)
1691 struct route_in6 *ro6 = &inc->inc6_route;
1693 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1694 /* No route yet, so try to acquire one */
1695 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1697 * unused portions of the structure MUST be zero'd
1698 * out because rtalloc() treats it as opaque data
1700 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1701 ro6->ro_dst.sin6_family = AF_INET6;
1702 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1703 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1704 rtalloc((struct route *)ro6);
1707 return (ro6->ro_rt);
1712 /* compute ESP/AH header size for TCP, including outer IP header. */
1714 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1722 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1724 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1729 if (inp->inp_vflag & INP_IPV6) {
1730 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1732 th = (struct tcphdr *)(ip6 + 1);
1733 m->m_pkthdr.len = m->m_len =
1734 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1735 tcp_fillheaders(tp, ip6, th);
1736 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1740 ip = mtod(m, struct ip *);
1741 th = (struct tcphdr *)(ip + 1);
1742 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1743 tcp_fillheaders(tp, ip, th);
1744 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1753 * Return a pointer to the cached information about the remote host.
1754 * The cached information is stored in the protocol specific part of
1755 * the route metrics.
1758 tcp_gettaocache(struct in_conninfo *inc)
1763 if (inc->inc_isipv6)
1764 rt = tcp_rtlookup6(inc);
1767 rt = tcp_rtlookup(inc);
1769 /* Make sure this is a host route and is up. */
1771 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1774 return (rmx_taop(rt->rt_rmx));
1778 * Clear all the TAO cache entries, called from tcp_init.
1781 * This routine is just an empty one, because we assume that the routing
1782 * routing tables are initialized at the same time when TCP, so there is
1783 * nothing in the cache left over.
1791 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1793 * This code attempts to calculate the bandwidth-delay product as a
1794 * means of determining the optimal window size to maximize bandwidth,
1795 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1796 * routers. This code also does a fairly good job keeping RTTs in check
1797 * across slow links like modems. We implement an algorithm which is very
1798 * similar (but not meant to be) TCP/Vegas. The code operates on the
1799 * transmitter side of a TCP connection and so only effects the transmit
1800 * side of the connection.
1802 * BACKGROUND: TCP makes no provision for the management of buffer space
1803 * at the end points or at the intermediate routers and switches. A TCP
1804 * stream, whether using NewReno or not, will eventually buffer as
1805 * many packets as it is able and the only reason this typically works is
1806 * due to the fairly small default buffers made available for a connection
1807 * (typicaly 16K or 32K). As machines use larger windows and/or window
1808 * scaling it is now fairly easy for even a single TCP connection to blow-out
1809 * all available buffer space not only on the local interface, but on
1810 * intermediate routers and switches as well. NewReno makes a misguided
1811 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1812 * then backing off, then steadily increasing the window again until another
1813 * failure occurs, ad-infinitum. This results in terrible oscillation that
1814 * is only made worse as network loads increase and the idea of intentionally
1815 * blowing out network buffers is, frankly, a terrible way to manage network
1818 * It is far better to limit the transmit window prior to the failure
1819 * condition being achieved. There are two general ways to do this: First
1820 * you can 'scan' through different transmit window sizes and locate the
1821 * point where the RTT stops increasing, indicating that you have filled the
1822 * pipe, then scan backwards until you note that RTT stops decreasing, then
1823 * repeat ad-infinitum. This method works in principle but has severe
1824 * implementation issues due to RTT variances, timer granularity, and
1825 * instability in the algorithm which can lead to many false positives and
1826 * create oscillations as well as interact badly with other TCP streams
1827 * implementing the same algorithm.
1829 * The second method is to limit the window to the bandwidth delay product
1830 * of the link. This is the method we implement. RTT variances and our
1831 * own manipulation of the congestion window, bwnd, can potentially
1832 * destabilize the algorithm. For this reason we have to stabilize the
1833 * elements used to calculate the window. We do this by using the minimum
1834 * observed RTT, the long term average of the observed bandwidth, and
1835 * by adding two segments worth of slop. It isn't perfect but it is able
1836 * to react to changing conditions and gives us a very stable basis on
1837 * which to extend the algorithm.
1840 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1848 * If inflight_enable is disabled in the middle of a tcp connection,
1849 * make sure snd_bwnd is effectively disabled.
1851 if (!tcp_inflight_enable) {
1852 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1853 tp->snd_bandwidth = 0;
1858 * Validate the delta time. If a connection is new or has been idle
1859 * a long time we have to reset the bandwidth calculator.
1862 delta_ticks = save_ticks - tp->t_bw_rtttime;
1863 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1864 tp->t_bw_rtttime = ticks;
1865 tp->t_bw_rtseq = ack_seq;
1866 if (tp->snd_bandwidth == 0)
1867 tp->snd_bandwidth = tcp_inflight_min;
1870 if (delta_ticks == 0)
1874 * Sanity check, plus ignore pure window update acks.
1876 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1880 * Figure out the bandwidth. Due to the tick granularity this
1881 * is a very rough number and it MUST be averaged over a fairly
1882 * long period of time. XXX we need to take into account a link
1883 * that is not using all available bandwidth, but for now our
1884 * slop will ramp us up if this case occurs and the bandwidth later
1887 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1888 tp->t_bw_rtttime = save_ticks;
1889 tp->t_bw_rtseq = ack_seq;
1890 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1892 tp->snd_bandwidth = bw;
1895 * Calculate the semi-static bandwidth delay product, plus two maximal
1896 * segments. The additional slop puts us squarely in the sweet
1897 * spot and also handles the bandwidth run-up case. Without the
1898 * slop we could be locking ourselves into a lower bandwidth.
1900 * Situations Handled:
1901 * (1) Prevents over-queueing of packets on LANs, especially on
1902 * high speed LANs, allowing larger TCP buffers to be
1903 * specified, and also does a good job preventing
1904 * over-queueing of packets over choke points like modems
1905 * (at least for the transmit side).
1907 * (2) Is able to handle changing network loads (bandwidth
1908 * drops so bwnd drops, bandwidth increases so bwnd
1911 * (3) Theoretically should stabilize in the face of multiple
1912 * connections implementing the same algorithm (this may need
1915 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1916 * be adjusted with a sysctl but typically only needs to be on
1917 * very slow connections. A value no smaller then 5 should
1918 * be used, but only reduce this default if you have no other
1922 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1923 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1924 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1927 if (tcp_inflight_debug > 0) {
1929 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1931 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1932 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1935 if ((long)bwnd < tcp_inflight_min)
1936 bwnd = tcp_inflight_min;
1937 if (bwnd > tcp_inflight_max)
1938 bwnd = tcp_inflight_max;
1939 if ((long)bwnd < tp->t_maxseg * 2)
1940 bwnd = tp->t_maxseg * 2;
1941 tp->snd_bwnd = bwnd;