2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of The DragonFly Project nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific, prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the University of
49 * California, Berkeley and its contributors.
50 * 4. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.63 2008/11/11 10:46:58 sephe Exp $
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
85 #include <sys/domain.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/protosw.h>
91 #include <sys/random.h>
92 #include <sys/in_cksum.h>
95 #include <vm/vm_zone.h>
97 #include <net/route.h>
99 #include <net/netisr.h>
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/ip6.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet6/in6_pcb.h>
108 #include <netinet/in_var.h>
109 #include <netinet/ip_var.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet/ip_icmp.h>
113 #include <netinet/icmp6.h>
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
123 #include <netinet/tcp_debug.h>
125 #include <netinet6/ip6protosw.h>
128 #include <netinet6/ipsec.h>
130 #include <netinet6/ipsec6.h>
135 #include <netproto/ipsec/ipsec.h>
137 #include <netproto/ipsec/ipsec6.h>
143 #include <sys/msgport2.h>
144 #include <machine/smp.h>
146 #include <net/netmsg2.h>
148 #if !defined(KTR_TCP)
149 #define KTR_TCP KTR_ALL
151 KTR_INFO_MASTER(tcp);
152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
155 #define logtcp(name) KTR_LOG(tcp_ ## name)
157 struct inpcbinfo tcbinfo[MAXCPU];
158 struct tcpcbackqhead tcpcbackq[MAXCPU];
160 int tcp_mpsafe_proto = 0;
161 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
163 static int tcp_mpsafe_thread = 0;
164 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
166 &tcp_mpsafe_thread, 0,
167 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
169 int tcp_mssdflt = TCP_MSS;
170 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
171 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
174 int tcp_v6mssdflt = TCP6_MSS;
175 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
176 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
180 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
181 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
182 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
185 int tcp_do_rfc1323 = 1;
186 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
187 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
189 int tcp_do_rfc1644 = 0;
190 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
191 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
193 static int tcp_tcbhashsize = 0;
194 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
195 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
197 static int do_tcpdrain = 1;
198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
199 "Enable tcp_drain routine for extra help when low on mbufs");
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
203 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
205 static int icmp_may_rst = 1;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
207 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
209 static int tcp_isn_reseed_interval = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
211 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
214 * TCP bandwidth limiting sysctls. Note that the default lower bound of
215 * 1024 exists only for debugging. A good production default would be
216 * something like 6100.
218 static int tcp_inflight_enable = 0;
219 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
220 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
222 static int tcp_inflight_debug = 0;
223 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
224 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
226 static int tcp_inflight_min = 6144;
227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
228 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
230 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
232 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
234 static int tcp_inflight_stab = 20;
235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
236 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
238 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
239 static struct malloc_pipe tcptemp_mpipe;
241 static void tcp_willblock(int);
242 static void tcp_cleartaocache (void);
243 static void tcp_notify (struct inpcb *, int);
245 struct tcp_stats tcpstats_percpu[MAXCPU];
248 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
252 for (cpu = 0; cpu < ncpus; ++cpu) {
253 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
254 sizeof(struct tcp_stats))))
256 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
257 sizeof(struct tcp_stats))))
263 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
264 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
266 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
267 &tcpstat, tcp_stats, "TCP statistics");
271 * Target size of TCP PCB hash tables. Must be a power of two.
273 * Note that this can be overridden by the kernel environment
274 * variable net.inet.tcp.tcbhashsize
277 #define TCBHASHSIZE 512
281 * This is the actual shape of what we allocate using the zone
282 * allocator. Doing it this way allows us to protect both structures
283 * using the same generation count, and also eliminates the overhead
284 * of allocating tcpcbs separately. By hiding the structure here,
285 * we avoid changing most of the rest of the code (although it needs
286 * to be changed, eventually, for greater efficiency).
289 #define ALIGNM1 (ALIGNMENT - 1)
293 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
296 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
297 struct callout inp_tp_delack;
308 struct inpcbporthead *porthashbase;
310 struct vm_zone *ipi_zone;
311 int hashsize = TCBHASHSIZE;
315 * note: tcptemp is used for keepalives, and it is ok for an
316 * allocation to fail so do not specify MPF_INT.
318 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
324 tcp_delacktime = TCPTV_DELACK;
325 tcp_keepinit = TCPTV_KEEP_INIT;
326 tcp_keepidle = TCPTV_KEEP_IDLE;
327 tcp_keepintvl = TCPTV_KEEPINTVL;
328 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
330 tcp_rexmit_min = TCPTV_MIN;
331 tcp_rexmit_slop = TCPTV_CPU_VAR;
333 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
334 if (!powerof2(hashsize)) {
335 kprintf("WARNING: TCB hash size not a power of 2\n");
336 hashsize = 512; /* safe default */
338 tcp_tcbhashsize = hashsize;
339 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
340 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
343 for (cpu = 0; cpu < ncpus2; cpu++) {
344 in_pcbinfo_init(&tcbinfo[cpu]);
345 tcbinfo[cpu].cpu = cpu;
346 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
347 &tcbinfo[cpu].hashmask);
348 tcbinfo[cpu].porthashbase = porthashbase;
349 tcbinfo[cpu].porthashmask = porthashmask;
350 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
351 &tcbinfo[cpu].wildcardhashmask);
352 tcbinfo[cpu].ipi_zone = ipi_zone;
353 TAILQ_INIT(&tcpcbackq[cpu]);
356 tcp_reass_maxseg = nmbclusters / 16;
357 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
360 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
362 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
364 if (max_protohdr < TCP_MINPROTOHDR)
365 max_protohdr = TCP_MINPROTOHDR;
366 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
368 #undef TCP_MINPROTOHDR
371 * Initialize TCP statistics counters for each CPU.
374 for (cpu = 0; cpu < ncpus; ++cpu) {
375 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
378 bzero(&tcpstat, sizeof(struct tcp_stats));
386 tcpmsg_service_loop(void *dummy)
392 * Thread was started with TDF_MPSAFE
396 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
399 mplocked = netmsg_service(msg, tcp_mpsafe_thread,
401 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
404 tcp_willblock(mplocked);
410 tcp_willblock(int mplocked)
413 int cpu = mycpu->gd_cpuid;
416 if (!mplocked && !tcp_mpsafe_proto) {
417 if (TAILQ_EMPTY(&tcpcbackq[cpu]))
425 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
426 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
427 tp->t_flags &= ~TF_ONOUTPUTQ;
428 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
438 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
439 * tcp_template used to store this data in mbufs, but we now recopy it out
440 * of the tcpcb each time to conserve mbufs.
443 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
445 struct inpcb *inp = tp->t_inpcb;
446 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
449 if (inp->inp_vflag & INP_IPV6) {
452 ip6 = (struct ip6_hdr *)ip_ptr;
453 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
454 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
455 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
456 (IPV6_VERSION & IPV6_VERSION_MASK);
457 ip6->ip6_nxt = IPPROTO_TCP;
458 ip6->ip6_plen = sizeof(struct tcphdr);
459 ip6->ip6_src = inp->in6p_laddr;
460 ip6->ip6_dst = inp->in6p_faddr;
465 struct ip *ip = (struct ip *) ip_ptr;
467 ip->ip_vhl = IP_VHL_BORING;
474 ip->ip_p = IPPROTO_TCP;
475 ip->ip_src = inp->inp_laddr;
476 ip->ip_dst = inp->inp_faddr;
477 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
479 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
482 tcp_hdr->th_sport = inp->inp_lport;
483 tcp_hdr->th_dport = inp->inp_fport;
488 tcp_hdr->th_flags = 0;
494 * Create template to be used to send tcp packets on a connection.
495 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
496 * use for this function is in keepalives, which use tcp_respond.
499 tcp_maketemplate(struct tcpcb *tp)
503 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
505 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
510 tcp_freetemplate(struct tcptemp *tmp)
512 mpipe_free(&tcptemp_mpipe, tmp);
516 * Send a single message to the TCP at address specified by
517 * the given TCP/IP header. If m == NULL, then we make a copy
518 * of the tcpiphdr at ti and send directly to the addressed host.
519 * This is used to force keep alive messages out using the TCP
520 * template for a connection. If flags are given then we send
521 * a message back to the TCP which originated the * segment ti,
522 * and discard the mbuf containing it and any other attached mbufs.
524 * In any case the ack and sequence number of the transmitted
525 * segment are as specified by the parameters.
527 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
530 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
531 tcp_seq ack, tcp_seq seq, int flags)
535 struct route *ro = NULL;
537 struct ip *ip = ipgen;
540 struct route_in6 *ro6 = NULL;
541 struct route_in6 sro6;
542 struct ip6_hdr *ip6 = ipgen;
543 boolean_t use_tmpro = TRUE;
545 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
547 const boolean_t isipv6 = FALSE;
551 if (!(flags & TH_RST)) {
552 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
553 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
554 win = (long)TCP_MAXWIN << tp->rcv_scale;
557 * Don't use the route cache of a listen socket,
558 * it is not MPSAFE; use temporary route cache.
560 if (tp->t_state != TCPS_LISTEN) {
562 ro6 = &tp->t_inpcb->in6p_route;
564 ro = &tp->t_inpcb->inp_route;
571 bzero(ro6, sizeof *ro6);
574 bzero(ro, sizeof *ro);
578 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
582 m->m_data += max_linkhdr;
584 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
585 ip6 = mtod(m, struct ip6_hdr *);
586 nth = (struct tcphdr *)(ip6 + 1);
588 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
589 ip = mtod(m, struct ip *);
590 nth = (struct tcphdr *)(ip + 1);
592 bcopy(th, nth, sizeof(struct tcphdr));
597 m->m_data = (caddr_t)ipgen;
598 /* m_len is set later */
600 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
602 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
603 nth = (struct tcphdr *)(ip6 + 1);
605 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
606 nth = (struct tcphdr *)(ip + 1);
610 * this is usually a case when an extension header
611 * exists between the IPv6 header and the
614 nth->th_sport = th->th_sport;
615 nth->th_dport = th->th_dport;
617 xchg(nth->th_dport, nth->th_sport, n_short);
622 ip6->ip6_vfc = IPV6_VERSION;
623 ip6->ip6_nxt = IPPROTO_TCP;
624 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
625 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
627 tlen += sizeof(struct tcpiphdr);
629 ip->ip_ttl = ip_defttl;
632 m->m_pkthdr.len = tlen;
633 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
634 nth->th_seq = htonl(seq);
635 nth->th_ack = htonl(ack);
637 nth->th_off = sizeof(struct tcphdr) >> 2;
638 nth->th_flags = flags;
640 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
642 nth->th_win = htons((u_short)win);
646 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
647 sizeof(struct ip6_hdr),
648 tlen - sizeof(struct ip6_hdr));
649 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
650 (ro6 && ro6->ro_rt) ?
651 ro6->ro_rt->rt_ifp : NULL);
653 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
654 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
655 m->m_pkthdr.csum_flags = CSUM_TCP;
656 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
659 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
660 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
663 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
664 tp ? tp->t_inpcb : NULL);
665 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
670 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
671 if ((ro == &sro) && (ro->ro_rt != NULL)) {
679 * Create a new TCP control block, making an
680 * empty reassembly queue and hooking it to the argument
681 * protocol control block. The `inp' parameter must have
682 * come from the zone allocator set up in tcp_init().
685 tcp_newtcpcb(struct inpcb *inp)
690 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
692 const boolean_t isipv6 = FALSE;
695 it = (struct inp_tp *)inp;
697 bzero(tp, sizeof(struct tcpcb));
698 LIST_INIT(&tp->t_segq);
699 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
701 /* Set up our timeouts. */
702 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
703 callout_init(tp->tt_persist = &it->inp_tp_persist);
704 callout_init(tp->tt_keep = &it->inp_tp_keep);
705 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
706 callout_init(tp->tt_delack = &it->inp_tp_delack);
709 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
711 tp->t_flags |= TF_REQ_CC;
712 tp->t_inpcb = inp; /* XXX */
713 tp->t_state = TCPS_CLOSED;
715 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
716 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
717 * reasonable initial retransmit time.
719 tp->t_srtt = TCPTV_SRTTBASE;
721 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
722 tp->t_rttmin = tcp_rexmit_min;
723 tp->t_rxtcur = TCPTV_RTOBASE;
724 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
725 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
726 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
727 tp->t_rcvtime = ticks;
729 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
730 * because the socket may be bound to an IPv6 wildcard address,
731 * which may match an IPv4-mapped IPv6 address.
733 inp->inp_ip_ttl = ip_defttl;
735 tcp_sack_tcpcb_init(tp);
736 return (tp); /* XXX */
740 * Drop a TCP connection, reporting the specified error.
741 * If connection is synchronized, then send a RST to peer.
744 tcp_drop(struct tcpcb *tp, int error)
746 struct socket *so = tp->t_inpcb->inp_socket;
748 if (TCPS_HAVERCVDSYN(tp->t_state)) {
749 tp->t_state = TCPS_CLOSED;
751 tcpstat.tcps_drops++;
753 tcpstat.tcps_conndrops++;
754 if (error == ETIMEDOUT && tp->t_softerror)
755 error = tp->t_softerror;
756 so->so_error = error;
757 return (tcp_close(tp));
762 struct netmsg_remwildcard {
763 struct netmsg nm_netmsg;
764 struct inpcb *nm_inp;
765 struct inpcbinfo *nm_pcbinfo;
774 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
775 * inp can be detached. We do this by cycling through the cpus, ending up
776 * on the cpu controlling the inp last and then doing the disconnect.
779 in_pcbremwildcardhash_handler(struct netmsg *msg0)
781 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
784 cpu = msg->nm_pcbinfo->cpu;
786 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
787 /* note: detach removes any wildcard hash entry */
790 in6_pcbdetach(msg->nm_inp);
793 in_pcbdetach(msg->nm_inp);
794 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
796 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
797 cpu = (cpu + 1) % ncpus2;
798 msg->nm_pcbinfo = &tcbinfo[cpu];
799 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
806 * Close a TCP control block:
807 * discard all space held by the tcp
808 * discard internet protocol block
809 * wake up any sleepers
812 tcp_close(struct tcpcb *tp)
815 struct inpcb *inp = tp->t_inpcb;
816 struct socket *so = inp->inp_socket;
818 boolean_t dosavessthresh;
823 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
824 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
826 const boolean_t isipv6 = FALSE;
830 * The tp is not instantly destroyed in the wildcard case. Setting
831 * the state to TCPS_TERMINATING will prevent the TCP stack from
832 * messing with it, though it should be noted that this change may
833 * not take effect on other cpus until we have chained the wildcard
836 * XXX we currently depend on the BGL to synchronize the tp->t_state
837 * update and prevent other tcp protocol threads from accepting new
838 * connections on the listen socket we might be trying to close down.
840 KKASSERT(tp->t_state != TCPS_TERMINATING);
841 tp->t_state = TCPS_TERMINATING;
844 * Make sure that all of our timers are stopped before we
847 callout_stop(tp->tt_rexmt);
848 callout_stop(tp->tt_persist);
849 callout_stop(tp->tt_keep);
850 callout_stop(tp->tt_2msl);
851 callout_stop(tp->tt_delack);
853 if (tp->t_flags & TF_ONOUTPUTQ) {
854 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
855 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
856 tp->t_flags &= ~TF_ONOUTPUTQ;
860 * If we got enough samples through the srtt filter,
861 * save the rtt and rttvar in the routing entry.
862 * 'Enough' is arbitrarily defined as the 16 samples.
863 * 16 samples is enough for the srtt filter to converge
864 * to within 5% of the correct value; fewer samples and
865 * we could save a very bogus rtt.
867 * Don't update the default route's characteristics and don't
868 * update anything that the user "locked".
870 if (tp->t_rttupdated >= 16) {
874 struct sockaddr_in6 *sin6;
876 if ((rt = inp->in6p_route.ro_rt) == NULL)
878 sin6 = (struct sockaddr_in6 *)rt_key(rt);
879 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
882 if ((rt = inp->inp_route.ro_rt) == NULL ||
883 ((struct sockaddr_in *)rt_key(rt))->
884 sin_addr.s_addr == INADDR_ANY)
887 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
888 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
889 if (rt->rt_rmx.rmx_rtt && i)
891 * filter this update to half the old & half
892 * the new values, converting scale.
893 * See route.h and tcp_var.h for a
894 * description of the scaling constants.
897 (rt->rt_rmx.rmx_rtt + i) / 2;
899 rt->rt_rmx.rmx_rtt = i;
900 tcpstat.tcps_cachedrtt++;
902 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
904 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
905 if (rt->rt_rmx.rmx_rttvar && i)
906 rt->rt_rmx.rmx_rttvar =
907 (rt->rt_rmx.rmx_rttvar + i) / 2;
909 rt->rt_rmx.rmx_rttvar = i;
910 tcpstat.tcps_cachedrttvar++;
913 * The old comment here said:
914 * update the pipelimit (ssthresh) if it has been updated
915 * already or if a pipesize was specified & the threshhold
916 * got below half the pipesize. I.e., wait for bad news
917 * before we start updating, then update on both good
920 * But we want to save the ssthresh even if no pipesize is
921 * specified explicitly in the route, because such
922 * connections still have an implicit pipesize specified
923 * by the global tcp_sendspace. In the absence of a reliable
924 * way to calculate the pipesize, it will have to do.
926 i = tp->snd_ssthresh;
927 if (rt->rt_rmx.rmx_sendpipe != 0)
928 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
930 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
931 if (dosavessthresh ||
932 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
933 (rt->rt_rmx.rmx_ssthresh != 0))) {
935 * convert the limit from user data bytes to
936 * packets then to packet data bytes.
938 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
943 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
944 sizeof(struct tcpiphdr));
945 if (rt->rt_rmx.rmx_ssthresh)
946 rt->rt_rmx.rmx_ssthresh =
947 (rt->rt_rmx.rmx_ssthresh + i) / 2;
949 rt->rt_rmx.rmx_ssthresh = i;
950 tcpstat.tcps_cachedssthresh++;
955 /* free the reassembly queue, if any */
956 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
957 LIST_REMOVE(q, tqe_q);
962 /* throw away SACK blocks in scoreboard*/
964 tcp_sack_cleanup(&tp->scb);
966 inp->inp_ppcb = NULL;
967 soisdisconnected(so);
969 * Discard the inp. In the SMP case a wildcard inp's hash (created
970 * by a listen socket or an INADDR_ANY udp socket) is replicated
971 * for each protocol thread and must be removed in the context of
972 * that thread. This is accomplished by chaining the message
975 * If the inp is not wildcarded we simply detach, which will remove
976 * the any hashes still present for this inp.
979 if (inp->inp_flags & INP_WILDCARD_MP) {
980 struct netmsg_remwildcard *msg;
982 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
983 msg = kmalloc(sizeof(struct netmsg_remwildcard),
984 M_LWKTMSG, M_INTWAIT);
985 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
986 in_pcbremwildcardhash_handler);
988 msg->nm_isinet6 = isafinet6;
991 msg->nm_pcbinfo = &tcbinfo[cpu];
992 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
996 /* note: detach removes any wildcard hash entry */
1004 tcpstat.tcps_closed++;
1008 static __inline void
1009 tcp_drain_oncpu(struct inpcbhead *head)
1013 struct tseg_qent *te;
1015 LIST_FOREACH(inpb, head, inp_list) {
1016 if (inpb->inp_flags & INP_PLACEMARKER)
1018 if ((tcpb = intotcpcb(inpb))) {
1019 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1020 LIST_REMOVE(te, tqe_q);
1030 struct netmsg_tcp_drain {
1031 struct netmsg nm_netmsg;
1032 struct inpcbhead *nm_head;
1036 tcp_drain_handler(netmsg_t netmsg)
1038 struct netmsg_tcp_drain *nm = (void *)netmsg;
1040 tcp_drain_oncpu(nm->nm_head);
1041 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1056 * Walk the tcpbs, if existing, and flush the reassembly queue,
1057 * if there is one...
1058 * XXX: The "Net/3" implementation doesn't imply that the TCP
1059 * reassembly queue should be flushed, but in a situation
1060 * where we're really low on mbufs, this is potentially
1064 for (cpu = 0; cpu < ncpus2; cpu++) {
1065 struct netmsg_tcp_drain *msg;
1067 if (cpu == mycpu->gd_cpuid) {
1068 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1070 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1071 M_LWKTMSG, M_NOWAIT);
1074 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1076 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1077 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1081 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1086 * Notify a tcp user of an asynchronous error;
1087 * store error as soft error, but wake up user
1088 * (for now, won't do anything until can select for soft error).
1090 * Do not wake up user since there currently is no mechanism for
1091 * reporting soft errors (yet - a kqueue filter may be added).
1094 tcp_notify(struct inpcb *inp, int error)
1096 struct tcpcb *tp = intotcpcb(inp);
1099 * Ignore some errors if we are hooked up.
1100 * If connection hasn't completed, has retransmitted several times,
1101 * and receives a second error, give up now. This is better
1102 * than waiting a long time to establish a connection that
1103 * can never complete.
1105 if (tp->t_state == TCPS_ESTABLISHED &&
1106 (error == EHOSTUNREACH || error == ENETUNREACH ||
1107 error == EHOSTDOWN)) {
1109 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1111 tcp_drop(tp, error);
1113 tp->t_softerror = error;
1115 wakeup(&so->so_timeo);
1122 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1125 struct inpcb *marker;
1135 * The process of preparing the TCB list is too time-consuming and
1136 * resource-intensive to repeat twice on every request.
1138 if (req->oldptr == NULL) {
1139 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1140 gd = globaldata_find(ccpu);
1141 n += tcbinfo[gd->gd_cpuid].ipi_count;
1143 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1147 if (req->newptr != NULL)
1150 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1151 marker->inp_flags |= INP_PLACEMARKER;
1154 * OK, now we're committed to doing something. Run the inpcb list
1155 * for each cpu in the system and construct the output. Use a
1156 * list placemarker to deal with list changes occuring during
1157 * copyout blockages (but otherwise depend on being on the correct
1158 * cpu to avoid races).
1160 origcpu = mycpu->gd_cpuid;
1161 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1167 cpu_id = (origcpu + ccpu) % ncpus;
1168 if ((smp_active_mask & (1 << cpu_id)) == 0)
1170 rgd = globaldata_find(cpu_id);
1171 lwkt_setcpu_self(rgd);
1173 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1174 n = tcbinfo[cpu_id].ipi_count;
1176 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1178 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1180 * process a snapshot of pcbs, ignoring placemarkers
1181 * and using our own to allow SYSCTL_OUT to block.
1183 LIST_REMOVE(marker, inp_list);
1184 LIST_INSERT_AFTER(inp, marker, inp_list);
1186 if (inp->inp_flags & INP_PLACEMARKER)
1188 if (inp->inp_gencnt > gencnt)
1190 if (prison_xinpcb(req->td, inp))
1193 xt.xt_len = sizeof xt;
1194 bcopy(inp, &xt.xt_inp, sizeof *inp);
1195 inp_ppcb = inp->inp_ppcb;
1196 if (inp_ppcb != NULL)
1197 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1199 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1200 if (inp->inp_socket)
1201 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1202 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1206 LIST_REMOVE(marker, inp_list);
1207 if (error == 0 && i < n) {
1208 bzero(&xt, sizeof xt);
1209 xt.xt_len = sizeof xt;
1211 error = SYSCTL_OUT(req, &xt, sizeof xt);
1220 * Make sure we are on the same cpu we were on originally, since
1221 * higher level callers expect this. Also don't pollute caches with
1222 * migrated userland data by (eventually) returning to userland
1223 * on a different cpu.
1225 lwkt_setcpu_self(globaldata_find(origcpu));
1226 kfree(marker, M_TEMP);
1230 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1231 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1234 tcp_getcred(SYSCTL_HANDLER_ARGS)
1236 struct sockaddr_in addrs[2];
1241 error = suser(req->td);
1244 error = SYSCTL_IN(req, addrs, sizeof addrs);
1248 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1249 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1250 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1251 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1252 if (inp == NULL || inp->inp_socket == NULL) {
1256 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1262 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1263 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1267 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1269 struct sockaddr_in6 addrs[2];
1272 boolean_t mapped = FALSE;
1274 error = suser(req->td);
1277 error = SYSCTL_IN(req, addrs, sizeof addrs);
1280 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1281 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1288 inp = in_pcblookup_hash(&tcbinfo[0],
1289 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1291 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1295 inp = in6_pcblookup_hash(&tcbinfo[0],
1296 &addrs[1].sin6_addr, addrs[1].sin6_port,
1297 &addrs[0].sin6_addr, addrs[0].sin6_port,
1300 if (inp == NULL || inp->inp_socket == NULL) {
1304 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1310 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1312 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1315 struct netmsg_tcp_notify {
1316 struct netmsg nm_nmsg;
1317 void (*nm_notify)(struct inpcb *, int);
1318 struct in_addr nm_faddr;
1323 tcp_notifyall_oncpu(struct netmsg *netmsg)
1325 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1328 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1329 nmsg->nm_arg, nmsg->nm_notify);
1331 nextcpu = mycpuid + 1;
1332 if (nextcpu < ncpus2)
1333 lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1335 lwkt_replymsg(&netmsg->nm_lmsg, 0);
1339 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1341 struct ip *ip = vip;
1343 struct in_addr faddr;
1346 void (*notify)(struct inpcb *, int) = tcp_notify;
1350 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1354 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1355 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1358 arg = inetctlerrmap[cmd];
1359 if (cmd == PRC_QUENCH) {
1360 notify = tcp_quench;
1361 } else if (icmp_may_rst &&
1362 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1363 cmd == PRC_UNREACH_PORT ||
1364 cmd == PRC_TIMXCEED_INTRANS) &&
1366 notify = tcp_drop_syn_sent;
1367 } else if (cmd == PRC_MSGSIZE) {
1368 struct icmp *icmp = (struct icmp *)
1369 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1371 arg = ntohs(icmp->icmp_nextmtu);
1372 notify = tcp_mtudisc;
1373 } else if (PRC_IS_REDIRECT(cmd)) {
1375 notify = in_rtchange;
1376 } else if (cmd == PRC_HOSTDEAD) {
1382 th = (struct tcphdr *)((caddr_t)ip +
1383 (IP_VHL_HL(ip->ip_vhl) << 2));
1384 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1385 ip->ip_src.s_addr, th->th_sport);
1386 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1387 ip->ip_src, th->th_sport, 0, NULL);
1388 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1389 icmpseq = htonl(th->th_seq);
1390 tp = intotcpcb(inp);
1391 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1392 SEQ_LT(icmpseq, tp->snd_max))
1393 (*notify)(inp, arg);
1395 struct in_conninfo inc;
1397 inc.inc_fport = th->th_dport;
1398 inc.inc_lport = th->th_sport;
1399 inc.inc_faddr = faddr;
1400 inc.inc_laddr = ip->ip_src;
1404 syncache_unreach(&inc, th);
1408 struct netmsg_tcp_notify nmsg;
1410 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1411 netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0,
1412 tcp_notifyall_oncpu);
1413 nmsg.nm_faddr = faddr;
1415 nmsg.nm_notify = notify;
1417 lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1423 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1426 void (*notify) (struct inpcb *, int) = tcp_notify;
1427 struct ip6_hdr *ip6;
1429 struct ip6ctlparam *ip6cp = NULL;
1430 const struct sockaddr_in6 *sa6_src = NULL;
1432 struct tcp_portonly {
1438 if (sa->sa_family != AF_INET6 ||
1439 sa->sa_len != sizeof(struct sockaddr_in6))
1443 if (cmd == PRC_QUENCH)
1444 notify = tcp_quench;
1445 else if (cmd == PRC_MSGSIZE) {
1446 struct ip6ctlparam *ip6cp = d;
1447 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1449 arg = ntohl(icmp6->icmp6_mtu);
1450 notify = tcp_mtudisc;
1451 } else if (!PRC_IS_REDIRECT(cmd) &&
1452 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1456 /* if the parameter is from icmp6, decode it. */
1458 ip6cp = (struct ip6ctlparam *)d;
1460 ip6 = ip6cp->ip6c_ip6;
1461 off = ip6cp->ip6c_off;
1462 sa6_src = ip6cp->ip6c_src;
1466 off = 0; /* fool gcc */
1471 struct in_conninfo inc;
1473 * XXX: We assume that when IPV6 is non NULL,
1474 * M and OFF are valid.
1477 /* check if we can safely examine src and dst ports */
1478 if (m->m_pkthdr.len < off + sizeof *thp)
1481 bzero(&th, sizeof th);
1482 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1484 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1485 (struct sockaddr *)ip6cp->ip6c_src,
1486 th.th_sport, cmd, arg, notify);
1488 inc.inc_fport = th.th_dport;
1489 inc.inc_lport = th.th_sport;
1490 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1491 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1493 syncache_unreach(&inc, &th);
1495 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1496 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1501 * Following is where TCP initial sequence number generation occurs.
1503 * There are two places where we must use initial sequence numbers:
1504 * 1. In SYN-ACK packets.
1505 * 2. In SYN packets.
1507 * All ISNs for SYN-ACK packets are generated by the syncache. See
1508 * tcp_syncache.c for details.
1510 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1511 * depends on this property. In addition, these ISNs should be
1512 * unguessable so as to prevent connection hijacking. To satisfy
1513 * the requirements of this situation, the algorithm outlined in
1514 * RFC 1948 is used to generate sequence numbers.
1516 * Implementation details:
1518 * Time is based off the system timer, and is corrected so that it
1519 * increases by one megabyte per second. This allows for proper
1520 * recycling on high speed LANs while still leaving over an hour
1523 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1524 * between seeding of isn_secret. This is normally set to zero,
1525 * as reseeding should not be necessary.
1529 #define ISN_BYTES_PER_SECOND 1048576
1531 u_char isn_secret[32];
1532 int isn_last_reseed;
1536 tcp_new_isn(struct tcpcb *tp)
1538 u_int32_t md5_buffer[4];
1541 /* Seed if this is the first use, reseed if requested. */
1542 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1543 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1545 read_random_unlimited(&isn_secret, sizeof isn_secret);
1546 isn_last_reseed = ticks;
1549 /* Compute the md5 hash and return the ISN. */
1551 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1552 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1554 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1555 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1556 sizeof(struct in6_addr));
1557 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1558 sizeof(struct in6_addr));
1562 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1563 sizeof(struct in_addr));
1564 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1565 sizeof(struct in_addr));
1567 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1568 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1569 new_isn = (tcp_seq) md5_buffer[0];
1570 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1575 * When a source quench is received, close congestion window
1576 * to one segment. We will gradually open it again as we proceed.
1579 tcp_quench(struct inpcb *inp, int error)
1581 struct tcpcb *tp = intotcpcb(inp);
1584 tp->snd_cwnd = tp->t_maxseg;
1590 * When a specific ICMP unreachable message is received and the
1591 * connection state is SYN-SENT, drop the connection. This behavior
1592 * is controlled by the icmp_may_rst sysctl.
1595 tcp_drop_syn_sent(struct inpcb *inp, int error)
1597 struct tcpcb *tp = intotcpcb(inp);
1599 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1600 tcp_drop(tp, error);
1604 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1605 * based on the new value in the route. Also nudge TCP to send something,
1606 * since we know the packet we just sent was dropped.
1607 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1610 tcp_mtudisc(struct inpcb *inp, int mtu)
1612 struct tcpcb *tp = intotcpcb(inp);
1614 struct socket *so = inp->inp_socket;
1617 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1619 const boolean_t isipv6 = FALSE;
1626 * If no MTU is provided in the ICMP message, use the
1627 * next lower likely value, as specified in RFC 1191.
1632 oldmtu = tp->t_maxopd +
1634 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1635 sizeof(struct tcpiphdr));
1636 mtu = ip_next_mtu(oldmtu, 0);
1640 rt = tcp_rtlookup6(&inp->inp_inc);
1642 rt = tcp_rtlookup(&inp->inp_inc);
1644 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1646 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1647 mtu = rt->rt_rmx.rmx_mtu;
1651 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1652 sizeof(struct tcpiphdr));
1655 * XXX - The following conditional probably violates the TCP
1656 * spec. The problem is that, since we don't know the
1657 * other end's MSS, we are supposed to use a conservative
1658 * default. But, if we do that, then MTU discovery will
1659 * never actually take place, because the conservative
1660 * default is much less than the MTUs typically seen
1661 * on the Internet today. For the moment, we'll sweep
1662 * this under the carpet.
1664 * The conservative default might not actually be a problem
1665 * if the only case this occurs is when sending an initial
1666 * SYN with options and data to a host we've never talked
1667 * to before. Then, they will reply with an MSS value which
1668 * will get recorded and the new parameters should get
1669 * recomputed. For Further Study.
1671 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1672 maxopd = taop->tao_mssopt;
1676 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1677 sizeof(struct tcpiphdr));
1679 if (tp->t_maxopd <= maxopd)
1681 tp->t_maxopd = maxopd;
1684 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1685 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1686 mss -= TCPOLEN_TSTAMP_APPA;
1688 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1689 (TF_REQ_CC | TF_RCVD_CC))
1690 mss -= TCPOLEN_CC_APPA;
1692 /* round down to multiple of MCLBYTES */
1693 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1695 mss &= ~(MCLBYTES - 1);
1698 mss = (mss / MCLBYTES) * MCLBYTES;
1701 if (so->so_snd.ssb_hiwat < mss)
1702 mss = so->so_snd.ssb_hiwat;
1706 tp->snd_nxt = tp->snd_una;
1708 tcpstat.tcps_mturesent++;
1712 * Look-up the routing entry to the peer of this inpcb. If no route
1713 * is found and it cannot be allocated the return NULL. This routine
1714 * is called by TCP routines that access the rmx structure and by tcp_mss
1715 * to get the interface MTU.
1718 tcp_rtlookup(struct in_conninfo *inc)
1720 struct route *ro = &inc->inc_route;
1722 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1723 /* No route yet, so try to acquire one */
1724 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1726 * unused portions of the structure MUST be zero'd
1727 * out because rtalloc() treats it as opaque data
1729 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1730 ro->ro_dst.sa_family = AF_INET;
1731 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1732 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1742 tcp_rtlookup6(struct in_conninfo *inc)
1744 struct route_in6 *ro6 = &inc->inc6_route;
1746 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1747 /* No route yet, so try to acquire one */
1748 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1750 * unused portions of the structure MUST be zero'd
1751 * out because rtalloc() treats it as opaque data
1753 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1754 ro6->ro_dst.sin6_family = AF_INET6;
1755 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1756 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1757 rtalloc((struct route *)ro6);
1760 return (ro6->ro_rt);
1765 /* compute ESP/AH header size for TCP, including outer IP header. */
1767 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1775 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1777 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1782 if (inp->inp_vflag & INP_IPV6) {
1783 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1785 th = (struct tcphdr *)(ip6 + 1);
1786 m->m_pkthdr.len = m->m_len =
1787 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1788 tcp_fillheaders(tp, ip6, th);
1789 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1793 ip = mtod(m, struct ip *);
1794 th = (struct tcphdr *)(ip + 1);
1795 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1796 tcp_fillheaders(tp, ip, th);
1797 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1806 * Return a pointer to the cached information about the remote host.
1807 * The cached information is stored in the protocol specific part of
1808 * the route metrics.
1811 tcp_gettaocache(struct in_conninfo *inc)
1816 if (inc->inc_isipv6)
1817 rt = tcp_rtlookup6(inc);
1820 rt = tcp_rtlookup(inc);
1822 /* Make sure this is a host route and is up. */
1824 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1827 return (rmx_taop(rt->rt_rmx));
1831 * Clear all the TAO cache entries, called from tcp_init.
1834 * This routine is just an empty one, because we assume that the routing
1835 * routing tables are initialized at the same time when TCP, so there is
1836 * nothing in the cache left over.
1839 tcp_cleartaocache(void)
1844 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1846 * This code attempts to calculate the bandwidth-delay product as a
1847 * means of determining the optimal window size to maximize bandwidth,
1848 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1849 * routers. This code also does a fairly good job keeping RTTs in check
1850 * across slow links like modems. We implement an algorithm which is very
1851 * similar (but not meant to be) TCP/Vegas. The code operates on the
1852 * transmitter side of a TCP connection and so only effects the transmit
1853 * side of the connection.
1855 * BACKGROUND: TCP makes no provision for the management of buffer space
1856 * at the end points or at the intermediate routers and switches. A TCP
1857 * stream, whether using NewReno or not, will eventually buffer as
1858 * many packets as it is able and the only reason this typically works is
1859 * due to the fairly small default buffers made available for a connection
1860 * (typicaly 16K or 32K). As machines use larger windows and/or window
1861 * scaling it is now fairly easy for even a single TCP connection to blow-out
1862 * all available buffer space not only on the local interface, but on
1863 * intermediate routers and switches as well. NewReno makes a misguided
1864 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1865 * then backing off, then steadily increasing the window again until another
1866 * failure occurs, ad-infinitum. This results in terrible oscillation that
1867 * is only made worse as network loads increase and the idea of intentionally
1868 * blowing out network buffers is, frankly, a terrible way to manage network
1871 * It is far better to limit the transmit window prior to the failure
1872 * condition being achieved. There are two general ways to do this: First
1873 * you can 'scan' through different transmit window sizes and locate the
1874 * point where the RTT stops increasing, indicating that you have filled the
1875 * pipe, then scan backwards until you note that RTT stops decreasing, then
1876 * repeat ad-infinitum. This method works in principle but has severe
1877 * implementation issues due to RTT variances, timer granularity, and
1878 * instability in the algorithm which can lead to many false positives and
1879 * create oscillations as well as interact badly with other TCP streams
1880 * implementing the same algorithm.
1882 * The second method is to limit the window to the bandwidth delay product
1883 * of the link. This is the method we implement. RTT variances and our
1884 * own manipulation of the congestion window, bwnd, can potentially
1885 * destabilize the algorithm. For this reason we have to stabilize the
1886 * elements used to calculate the window. We do this by using the minimum
1887 * observed RTT, the long term average of the observed bandwidth, and
1888 * by adding two segments worth of slop. It isn't perfect but it is able
1889 * to react to changing conditions and gives us a very stable basis on
1890 * which to extend the algorithm.
1893 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1901 * If inflight_enable is disabled in the middle of a tcp connection,
1902 * make sure snd_bwnd is effectively disabled.
1904 if (!tcp_inflight_enable) {
1905 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1906 tp->snd_bandwidth = 0;
1911 * Validate the delta time. If a connection is new or has been idle
1912 * a long time we have to reset the bandwidth calculator.
1915 delta_ticks = save_ticks - tp->t_bw_rtttime;
1916 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1917 tp->t_bw_rtttime = ticks;
1918 tp->t_bw_rtseq = ack_seq;
1919 if (tp->snd_bandwidth == 0)
1920 tp->snd_bandwidth = tcp_inflight_min;
1923 if (delta_ticks == 0)
1927 * Sanity check, plus ignore pure window update acks.
1929 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1933 * Figure out the bandwidth. Due to the tick granularity this
1934 * is a very rough number and it MUST be averaged over a fairly
1935 * long period of time. XXX we need to take into account a link
1936 * that is not using all available bandwidth, but for now our
1937 * slop will ramp us up if this case occurs and the bandwidth later
1940 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1941 tp->t_bw_rtttime = save_ticks;
1942 tp->t_bw_rtseq = ack_seq;
1943 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1945 tp->snd_bandwidth = bw;
1948 * Calculate the semi-static bandwidth delay product, plus two maximal
1949 * segments. The additional slop puts us squarely in the sweet
1950 * spot and also handles the bandwidth run-up case. Without the
1951 * slop we could be locking ourselves into a lower bandwidth.
1953 * Situations Handled:
1954 * (1) Prevents over-queueing of packets on LANs, especially on
1955 * high speed LANs, allowing larger TCP buffers to be
1956 * specified, and also does a good job preventing
1957 * over-queueing of packets over choke points like modems
1958 * (at least for the transmit side).
1960 * (2) Is able to handle changing network loads (bandwidth
1961 * drops so bwnd drops, bandwidth increases so bwnd
1964 * (3) Theoretically should stabilize in the face of multiple
1965 * connections implementing the same algorithm (this may need
1968 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1969 * be adjusted with a sysctl but typically only needs to be on
1970 * very slow connections. A value no smaller then 5 should
1971 * be used, but only reduce this default if you have no other
1975 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1976 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1977 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1980 if (tcp_inflight_debug > 0) {
1982 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1984 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1985 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1988 if ((long)bwnd < tcp_inflight_min)
1989 bwnd = tcp_inflight_min;
1990 if (bwnd > tcp_inflight_max)
1991 bwnd = tcp_inflight_max;
1992 if ((long)bwnd < tp->t_maxseg * 2)
1993 bwnd = tp->t_maxseg * 2;
1994 tp->snd_bwnd = bwnd;