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"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_tcpdebug.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/callout.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mpipe.h>
86 #include <sys/domain.h>
90 #include <sys/socket.h>
91 #include <sys/socketvar.h>
92 #include <sys/protosw.h>
93 #include <sys/random.h>
94 #include <sys/in_cksum.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_timer2.h>
120 #include <netinet/tcp_var.h>
121 #include <netinet6/tcp6_var.h>
122 #include <netinet/tcpip.h>
124 #include <netinet/tcp_debug.h>
126 #include <netinet6/ip6protosw.h>
129 #include <netinet6/ipsec.h>
130 #include <netproto/key/key.h>
132 #include <netinet6/ipsec6.h>
137 #include <netproto/ipsec/ipsec.h>
139 #include <netproto/ipsec/ipsec6.h>
145 #include <machine/smp.h>
147 #include <sys/msgport2.h>
148 #include <sys/mplock2.h>
149 #include <net/netmsg2.h>
151 #if !defined(KTR_TCP)
152 #define KTR_TCP KTR_ALL
154 KTR_INFO_MASTER(tcp);
156 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
158 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
160 #define logtcp(name) KTR_LOG(tcp_ ## name)
162 struct inpcbinfo tcbinfo[MAXCPU];
163 struct tcpcbackqhead tcpcbackq[MAXCPU];
165 static struct lwkt_token tcp_port_token =
166 LWKT_TOKEN_INITIALIZER(tcp_port_token);
168 int tcp_mssdflt = TCP_MSS;
169 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
170 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
173 int tcp_v6mssdflt = TCP6_MSS;
174 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
175 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
179 * Minimum MSS we accept and use. This prevents DoS attacks where
180 * we are forced to a ridiculous low MSS like 20 and send hundreds
181 * of packets instead of one. The effect scales with the available
182 * bandwidth and quickly saturates the CPU and network interface
183 * with packet generation and sending. Set to zero to disable MINMSS
184 * checking. This setting prevents us from sending too small packets.
186 int tcp_minmss = TCP_MINMSS;
187 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
188 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
191 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
192 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
193 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
196 int tcp_do_rfc1323 = 1;
197 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
198 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
200 static int tcp_tcbhashsize = 0;
201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
202 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
204 static int do_tcpdrain = 1;
205 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
206 "Enable tcp_drain routine for extra help when low on mbufs");
208 static int icmp_may_rst = 1;
209 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
210 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
212 static int tcp_isn_reseed_interval = 0;
213 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
214 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
217 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
218 * by default, but with generous values which should allow maximal
219 * bandwidth. In particular, the slop defaults to 50 (5 packets).
221 * The reason for doing this is that the limiter is the only mechanism we
222 * have which seems to do a really good job preventing receiver RX rings
223 * on network interfaces from getting blown out. Even though GigE/10GigE
224 * is supposed to flow control it looks like either it doesn't actually
225 * do it or Open Source drivers do not properly enable it.
227 * People using the limiter to reduce bottlenecks on slower WAN connections
228 * should set the slop to 20 (2 packets).
230 static int tcp_inflight_enable = 1;
231 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
232 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
234 static int tcp_inflight_debug = 0;
235 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
236 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
238 static int tcp_inflight_min = 6144;
239 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
240 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
242 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
243 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
244 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
246 static int tcp_inflight_stab = 50;
247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
248 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
250 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
251 static struct malloc_pipe tcptemp_mpipe;
253 static void tcp_willblock(void);
254 static void tcp_notify (struct inpcb *, int);
256 struct tcp_stats tcpstats_percpu[MAXCPU];
259 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
263 for (cpu = 0; cpu < ncpus; ++cpu) {
264 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
265 sizeof(struct tcp_stats))))
267 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
268 sizeof(struct tcp_stats))))
274 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
275 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
277 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
278 &tcpstat, tcp_stats, "TCP statistics");
282 * Target size of TCP PCB hash tables. Must be a power of two.
284 * Note that this can be overridden by the kernel environment
285 * variable net.inet.tcp.tcbhashsize
288 #define TCBHASHSIZE 512
292 * This is the actual shape of what we allocate using the zone
293 * allocator. Doing it this way allows us to protect both structures
294 * using the same generation count, and also eliminates the overhead
295 * of allocating tcpcbs separately. By hiding the structure here,
296 * we avoid changing most of the rest of the code (although it needs
297 * to be changed, eventually, for greater efficiency).
300 #define ALIGNM1 (ALIGNMENT - 1)
304 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
307 struct tcp_callout inp_tp_rexmt;
308 struct tcp_callout inp_tp_persist;
309 struct tcp_callout inp_tp_keep;
310 struct tcp_callout inp_tp_2msl;
311 struct tcp_callout inp_tp_delack;
312 struct netmsg_tcp_timer inp_tp_timermsg;
323 struct inpcbporthead *porthashbase;
324 struct inpcbinfo *ticb;
326 int hashsize = TCBHASHSIZE;
330 * note: tcptemp is used for keepalives, and it is ok for an
331 * allocation to fail so do not specify MPF_INT.
333 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
334 25, -1, 0, NULL, NULL, NULL);
336 tcp_delacktime = TCPTV_DELACK;
337 tcp_keepinit = TCPTV_KEEP_INIT;
338 tcp_keepidle = TCPTV_KEEP_IDLE;
339 tcp_keepintvl = TCPTV_KEEPINTVL;
340 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
342 tcp_rexmit_min = TCPTV_MIN;
343 tcp_rexmit_slop = TCPTV_CPU_VAR;
345 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
346 if (!powerof2(hashsize)) {
347 kprintf("WARNING: TCB hash size not a power of 2\n");
348 hashsize = 512; /* safe default */
350 tcp_tcbhashsize = hashsize;
351 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
353 for (cpu = 0; cpu < ncpus2; cpu++) {
354 ticb = &tcbinfo[cpu];
355 in_pcbinfo_init(ticb);
357 ticb->hashbase = hashinit(hashsize, M_PCB,
359 ticb->porthashbase = porthashbase;
360 ticb->porthashmask = porthashmask;
361 ticb->porttoken = &tcp_port_token;
363 ticb->porthashbase = hashinit(hashsize, M_PCB,
364 &ticb->porthashmask);
366 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
367 &ticb->wildcardhashmask);
368 ticb->ipi_size = sizeof(struct inp_tp);
369 TAILQ_INIT(&tcpcbackq[cpu]);
372 tcp_reass_maxseg = nmbclusters / 16;
373 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
376 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
378 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
380 if (max_protohdr < TCP_MINPROTOHDR)
381 max_protohdr = TCP_MINPROTOHDR;
382 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
384 #undef TCP_MINPROTOHDR
387 * Initialize TCP statistics counters for each CPU.
390 for (cpu = 0; cpu < ncpus; ++cpu) {
391 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
394 bzero(&tcpstat, sizeof(struct tcp_stats));
398 netisr_register_rollup(tcp_willblock);
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);
416 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
417 * tcp_template used to store this data in mbufs, but we now recopy it out
418 * of the tcpcb each time to conserve mbufs.
421 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
423 struct inpcb *inp = tp->t_inpcb;
424 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
427 if (inp->inp_vflag & INP_IPV6) {
430 ip6 = (struct ip6_hdr *)ip_ptr;
431 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
432 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
433 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
434 (IPV6_VERSION & IPV6_VERSION_MASK);
435 ip6->ip6_nxt = IPPROTO_TCP;
436 ip6->ip6_plen = sizeof(struct tcphdr);
437 ip6->ip6_src = inp->in6p_laddr;
438 ip6->ip6_dst = inp->in6p_faddr;
443 struct ip *ip = (struct ip *) ip_ptr;
445 ip->ip_vhl = IP_VHL_BORING;
452 ip->ip_p = IPPROTO_TCP;
453 ip->ip_src = inp->inp_laddr;
454 ip->ip_dst = inp->inp_faddr;
455 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
457 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
460 tcp_hdr->th_sport = inp->inp_lport;
461 tcp_hdr->th_dport = inp->inp_fport;
466 tcp_hdr->th_flags = 0;
472 * Create template to be used to send tcp packets on a connection.
473 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
474 * use for this function is in keepalives, which use tcp_respond.
477 tcp_maketemplate(struct tcpcb *tp)
481 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
483 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
488 tcp_freetemplate(struct tcptemp *tmp)
490 mpipe_free(&tcptemp_mpipe, tmp);
494 * Send a single message to the TCP at address specified by
495 * the given TCP/IP header. If m == NULL, then we make a copy
496 * of the tcpiphdr at ti and send directly to the addressed host.
497 * This is used to force keep alive messages out using the TCP
498 * template for a connection. If flags are given then we send
499 * a message back to the TCP which originated the * segment ti,
500 * and discard the mbuf containing it and any other attached mbufs.
502 * In any case the ack and sequence number of the transmitted
503 * segment are as specified by the parameters.
505 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
508 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
509 tcp_seq ack, tcp_seq seq, int flags)
513 struct route *ro = NULL;
515 struct ip *ip = ipgen;
518 struct route_in6 *ro6 = NULL;
519 struct route_in6 sro6;
520 struct ip6_hdr *ip6 = ipgen;
521 boolean_t use_tmpro = TRUE;
523 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
525 const boolean_t isipv6 = FALSE;
529 if (!(flags & TH_RST)) {
530 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
533 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
534 win = (long)TCP_MAXWIN << tp->rcv_scale;
537 * Don't use the route cache of a listen socket,
538 * it is not MPSAFE; use temporary route cache.
540 if (tp->t_state != TCPS_LISTEN) {
542 ro6 = &tp->t_inpcb->in6p_route;
544 ro = &tp->t_inpcb->inp_route;
551 bzero(ro6, sizeof *ro6);
554 bzero(ro, sizeof *ro);
558 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
562 m->m_data += max_linkhdr;
564 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
565 ip6 = mtod(m, struct ip6_hdr *);
566 nth = (struct tcphdr *)(ip6 + 1);
568 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
569 ip = mtod(m, struct ip *);
570 nth = (struct tcphdr *)(ip + 1);
572 bcopy(th, nth, sizeof(struct tcphdr));
577 m->m_data = (caddr_t)ipgen;
578 /* m_len is set later */
580 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
582 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
583 nth = (struct tcphdr *)(ip6 + 1);
585 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
586 nth = (struct tcphdr *)(ip + 1);
590 * this is usually a case when an extension header
591 * exists between the IPv6 header and the
594 nth->th_sport = th->th_sport;
595 nth->th_dport = th->th_dport;
597 xchg(nth->th_dport, nth->th_sport, n_short);
602 ip6->ip6_vfc = IPV6_VERSION;
603 ip6->ip6_nxt = IPPROTO_TCP;
604 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
605 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
607 tlen += sizeof(struct tcpiphdr);
609 ip->ip_ttl = ip_defttl;
612 m->m_pkthdr.len = tlen;
613 m->m_pkthdr.rcvif = NULL;
614 nth->th_seq = htonl(seq);
615 nth->th_ack = htonl(ack);
617 nth->th_off = sizeof(struct tcphdr) >> 2;
618 nth->th_flags = flags;
620 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
622 nth->th_win = htons((u_short)win);
626 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
627 sizeof(struct ip6_hdr),
628 tlen - sizeof(struct ip6_hdr));
629 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
630 (ro6 && ro6->ro_rt) ?
631 ro6->ro_rt->rt_ifp : NULL);
633 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
634 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
635 m->m_pkthdr.csum_flags = CSUM_TCP;
636 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
639 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
640 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
643 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
644 tp ? tp->t_inpcb : NULL);
645 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
650 ipflags |= IP_DEBUGROUTE;
651 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
652 if ((ro == &sro) && (ro->ro_rt != NULL)) {
660 * Create a new TCP control block, making an
661 * empty reassembly queue and hooking it to the argument
662 * protocol control block. The `inp' parameter must have
663 * come from the zone allocator set up in tcp_init().
666 tcp_newtcpcb(struct inpcb *inp)
671 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
673 const boolean_t isipv6 = FALSE;
676 it = (struct inp_tp *)inp;
678 bzero(tp, sizeof(struct tcpcb));
679 LIST_INIT(&tp->t_segq);
680 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
682 /* Set up our timeouts. */
683 tp->tt_rexmt = &it->inp_tp_rexmt;
684 tp->tt_persist = &it->inp_tp_persist;
685 tp->tt_keep = &it->inp_tp_keep;
686 tp->tt_2msl = &it->inp_tp_2msl;
687 tp->tt_delack = &it->inp_tp_delack;
691 * Zero out timer message. We don't create it here,
692 * since the current CPU may not be the owner of this
695 tp->tt_msg = &it->inp_tp_timermsg;
696 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
699 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
700 tp->t_inpcb = inp; /* XXX */
701 tp->t_state = TCPS_CLOSED;
703 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
704 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
705 * reasonable initial retransmit time.
707 tp->t_srtt = TCPTV_SRTTBASE;
709 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
710 tp->t_rttmin = tcp_rexmit_min;
711 tp->t_rxtcur = TCPTV_RTOBASE;
712 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
713 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
714 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
715 tp->t_rcvtime = ticks;
717 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
718 * because the socket may be bound to an IPv6 wildcard address,
719 * which may match an IPv4-mapped IPv6 address.
721 inp->inp_ip_ttl = ip_defttl;
723 tcp_sack_tcpcb_init(tp);
724 return (tp); /* XXX */
728 * Drop a TCP connection, reporting the specified error.
729 * If connection is synchronized, then send a RST to peer.
732 tcp_drop(struct tcpcb *tp, int error)
734 struct socket *so = tp->t_inpcb->inp_socket;
736 if (TCPS_HAVERCVDSYN(tp->t_state)) {
737 tp->t_state = TCPS_CLOSED;
739 tcpstat.tcps_drops++;
741 tcpstat.tcps_conndrops++;
742 if (error == ETIMEDOUT && tp->t_softerror)
743 error = tp->t_softerror;
744 so->so_error = error;
745 return (tcp_close(tp));
750 struct netmsg_listen_detach {
751 struct netmsg_base base;
756 tcp_listen_detach_handler(netmsg_t msg)
758 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
759 struct tcpcb *tp = nmsg->nm_tp;
760 int cpu = mycpuid, nextcpu;
762 if (tp->t_flags & TF_LISTEN)
763 syncache_destroy(tp);
765 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
768 if (nextcpu < ncpus2)
769 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg);
771 lwkt_replymsg(&nmsg->base.lmsg, 0);
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;
791 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
792 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
794 const boolean_t isipv6 = FALSE;
799 * INP_WILDCARD_MP indicates that listen(2) has been called on
800 * this socket. This implies:
801 * - A wildcard inp's hash is replicated for each protocol thread.
802 * - Syncache for this inp grows independently in each protocol
804 * - There is more than one cpu
806 * We have to chain a message to the rest of the protocol threads
807 * to cleanup the wildcard hash and the syncache. The cleanup
808 * in the current protocol thread is defered till the end of this
812 * After cleanup the inp's hash and syncache entries, this inp will
813 * no longer be available to the rest of the protocol threads, so we
814 * are safe to whack the inp in the following code.
816 if (inp->inp_flags & INP_WILDCARD_MP) {
817 struct netmsg_listen_detach nmsg;
819 KKASSERT(so->so_port == cpu_portfn(0));
820 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
821 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
823 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
824 MSGF_PRIORITY, tcp_listen_detach_handler);
826 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0);
828 inp->inp_flags &= ~INP_WILDCARD_MP;
832 KKASSERT(tp->t_state != TCPS_TERMINATING);
833 tp->t_state = TCPS_TERMINATING;
836 * Make sure that all of our timers are stopped before we
837 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
838 * timers are never used. If timer message is never created
839 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
841 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
842 tcp_callout_stop(tp, tp->tt_rexmt);
843 tcp_callout_stop(tp, tp->tt_persist);
844 tcp_callout_stop(tp, tp->tt_keep);
845 tcp_callout_stop(tp, tp->tt_2msl);
846 tcp_callout_stop(tp, tp->tt_delack);
849 if (tp->t_flags & TF_ONOUTPUTQ) {
850 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
851 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
852 tp->t_flags &= ~TF_ONOUTPUTQ;
856 * If we got enough samples through the srtt filter,
857 * save the rtt and rttvar in the routing entry.
858 * 'Enough' is arbitrarily defined as the 16 samples.
859 * 16 samples is enough for the srtt filter to converge
860 * to within 5% of the correct value; fewer samples and
861 * we could save a very bogus rtt.
863 * Don't update the default route's characteristics and don't
864 * update anything that the user "locked".
866 if (tp->t_rttupdated >= 16) {
870 struct sockaddr_in6 *sin6;
872 if ((rt = inp->in6p_route.ro_rt) == NULL)
874 sin6 = (struct sockaddr_in6 *)rt_key(rt);
875 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
878 if ((rt = inp->inp_route.ro_rt) == NULL ||
879 ((struct sockaddr_in *)rt_key(rt))->
880 sin_addr.s_addr == INADDR_ANY)
883 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
884 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
885 if (rt->rt_rmx.rmx_rtt && i)
887 * filter this update to half the old & half
888 * the new values, converting scale.
889 * See route.h and tcp_var.h for a
890 * description of the scaling constants.
893 (rt->rt_rmx.rmx_rtt + i) / 2;
895 rt->rt_rmx.rmx_rtt = i;
896 tcpstat.tcps_cachedrtt++;
898 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
900 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
901 if (rt->rt_rmx.rmx_rttvar && i)
902 rt->rt_rmx.rmx_rttvar =
903 (rt->rt_rmx.rmx_rttvar + i) / 2;
905 rt->rt_rmx.rmx_rttvar = i;
906 tcpstat.tcps_cachedrttvar++;
909 * The old comment here said:
910 * update the pipelimit (ssthresh) if it has been updated
911 * already or if a pipesize was specified & the threshhold
912 * got below half the pipesize. I.e., wait for bad news
913 * before we start updating, then update on both good
916 * But we want to save the ssthresh even if no pipesize is
917 * specified explicitly in the route, because such
918 * connections still have an implicit pipesize specified
919 * by the global tcp_sendspace. In the absence of a reliable
920 * way to calculate the pipesize, it will have to do.
922 i = tp->snd_ssthresh;
923 if (rt->rt_rmx.rmx_sendpipe != 0)
924 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
926 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
927 if (dosavessthresh ||
928 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
929 (rt->rt_rmx.rmx_ssthresh != 0))) {
931 * convert the limit from user data bytes to
932 * packets then to packet data bytes.
934 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
939 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
940 sizeof(struct tcpiphdr));
941 if (rt->rt_rmx.rmx_ssthresh)
942 rt->rt_rmx.rmx_ssthresh =
943 (rt->rt_rmx.rmx_ssthresh + i) / 2;
945 rt->rt_rmx.rmx_ssthresh = i;
946 tcpstat.tcps_cachedssthresh++;
951 /* free the reassembly queue, if any */
952 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
953 LIST_REMOVE(q, tqe_q);
956 atomic_add_int(&tcp_reass_qsize, -1);
958 /* throw away SACK blocks in scoreboard*/
960 tcp_sack_cleanup(&tp->scb);
962 inp->inp_ppcb = NULL;
963 soisdisconnected(so);
964 /* note: pcb detached later on */
966 tcp_destroy_timermsg(tp);
968 if (tp->t_flags & TF_LISTEN)
969 syncache_destroy(tp);
973 * pcbdetach removes any wildcard hash entry on the current CPU.
982 tcpstat.tcps_closed++;
987 tcp_drain_oncpu(struct inpcbhead *head)
989 struct inpcb *marker;
992 struct tseg_qent *te;
995 * Allows us to block while running the list
997 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
998 marker->inp_flags |= INP_PLACEMARKER;
999 LIST_INSERT_HEAD(head, marker, inp_list);
1001 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1002 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1003 (tcpb = intotcpcb(inpb)) != NULL &&
1004 (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1005 LIST_REMOVE(te, tqe_q);
1008 atomic_add_int(&tcp_reass_qsize, -1);
1011 LIST_REMOVE(marker, inp_list);
1012 LIST_INSERT_AFTER(inpb, marker, inp_list);
1015 LIST_REMOVE(marker, inp_list);
1016 kfree(marker, M_TEMP);
1020 struct netmsg_tcp_drain {
1021 struct netmsg_base base;
1022 struct inpcbhead *nm_head;
1026 tcp_drain_handler(netmsg_t msg)
1028 struct netmsg_tcp_drain *nm = (void *)msg;
1030 tcp_drain_oncpu(nm->nm_head);
1031 lwkt_replymsg(&nm->base.lmsg, 0);
1046 * Walk the tcpbs, if existing, and flush the reassembly queue,
1047 * if there is one...
1048 * XXX: The "Net/3" implementation doesn't imply that the TCP
1049 * reassembly queue should be flushed, but in a situation
1050 * where we're really low on mbufs, this is potentially
1054 for (cpu = 0; cpu < ncpus2; cpu++) {
1055 struct netmsg_tcp_drain *nm;
1057 if (cpu == mycpu->gd_cpuid) {
1058 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1060 nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1061 M_LWKTMSG, M_NOWAIT);
1064 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1065 0, tcp_drain_handler);
1066 nm->nm_head = &tcbinfo[cpu].pcblisthead;
1067 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg);
1071 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1076 * Notify a tcp user of an asynchronous error;
1077 * store error as soft error, but wake up user
1078 * (for now, won't do anything until can select for soft error).
1080 * Do not wake up user since there currently is no mechanism for
1081 * reporting soft errors (yet - a kqueue filter may be added).
1084 tcp_notify(struct inpcb *inp, int error)
1086 struct tcpcb *tp = intotcpcb(inp);
1089 * Ignore some errors if we are hooked up.
1090 * If connection hasn't completed, has retransmitted several times,
1091 * and receives a second error, give up now. This is better
1092 * than waiting a long time to establish a connection that
1093 * can never complete.
1095 if (tp->t_state == TCPS_ESTABLISHED &&
1096 (error == EHOSTUNREACH || error == ENETUNREACH ||
1097 error == EHOSTDOWN)) {
1099 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1101 tcp_drop(tp, error);
1103 tp->t_softerror = error;
1105 wakeup(&so->so_timeo);
1112 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1115 struct inpcb *marker;
1124 * The process of preparing the TCB list is too time-consuming and
1125 * resource-intensive to repeat twice on every request.
1127 if (req->oldptr == NULL) {
1128 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1129 gd = globaldata_find(ccpu);
1130 n += tcbinfo[gd->gd_cpuid].ipi_count;
1132 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1136 if (req->newptr != NULL)
1139 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1140 marker->inp_flags |= INP_PLACEMARKER;
1143 * OK, now we're committed to doing something. Run the inpcb list
1144 * for each cpu in the system and construct the output. Use a
1145 * list placemarker to deal with list changes occuring during
1146 * copyout blockages (but otherwise depend on being on the correct
1147 * cpu to avoid races).
1149 origcpu = mycpu->gd_cpuid;
1150 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1156 cpu_id = (origcpu + ccpu) % ncpus;
1157 if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1159 rgd = globaldata_find(cpu_id);
1160 lwkt_setcpu_self(rgd);
1162 n = tcbinfo[cpu_id].ipi_count;
1164 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1166 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1168 * process a snapshot of pcbs, ignoring placemarkers
1169 * and using our own to allow SYSCTL_OUT to block.
1171 LIST_REMOVE(marker, inp_list);
1172 LIST_INSERT_AFTER(inp, marker, inp_list);
1174 if (inp->inp_flags & INP_PLACEMARKER)
1176 if (prison_xinpcb(req->td, inp))
1179 xt.xt_len = sizeof xt;
1180 bcopy(inp, &xt.xt_inp, sizeof *inp);
1181 inp_ppcb = inp->inp_ppcb;
1182 if (inp_ppcb != NULL)
1183 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1185 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1186 if (inp->inp_socket)
1187 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1188 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1192 LIST_REMOVE(marker, inp_list);
1193 if (error == 0 && i < n) {
1194 bzero(&xt, sizeof xt);
1195 xt.xt_len = sizeof xt;
1197 error = SYSCTL_OUT(req, &xt, sizeof xt);
1206 * Make sure we are on the same cpu we were on originally, since
1207 * higher level callers expect this. Also don't pollute caches with
1208 * migrated userland data by (eventually) returning to userland
1209 * on a different cpu.
1211 lwkt_setcpu_self(globaldata_find(origcpu));
1212 kfree(marker, M_TEMP);
1216 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1217 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1220 tcp_getcred(SYSCTL_HANDLER_ARGS)
1222 struct sockaddr_in addrs[2];
1227 error = priv_check(req->td, PRIV_ROOT);
1230 error = SYSCTL_IN(req, addrs, sizeof addrs);
1234 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1235 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1236 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1237 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1238 if (inp == NULL || inp->inp_socket == NULL) {
1242 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1248 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1249 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1253 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1255 struct sockaddr_in6 addrs[2];
1258 boolean_t mapped = FALSE;
1260 error = priv_check(req->td, PRIV_ROOT);
1263 error = SYSCTL_IN(req, addrs, sizeof addrs);
1266 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1267 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1274 inp = in_pcblookup_hash(&tcbinfo[0],
1275 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1277 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1281 inp = in6_pcblookup_hash(&tcbinfo[0],
1282 &addrs[1].sin6_addr, addrs[1].sin6_port,
1283 &addrs[0].sin6_addr, addrs[0].sin6_port,
1286 if (inp == NULL || inp->inp_socket == NULL) {
1290 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1296 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1298 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1301 struct netmsg_tcp_notify {
1302 struct netmsg_base base;
1303 void (*nm_notify)(struct inpcb *, int);
1304 struct in_addr nm_faddr;
1309 tcp_notifyall_oncpu(netmsg_t msg)
1311 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1314 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1315 nm->nm_arg, nm->nm_notify);
1317 nextcpu = mycpuid + 1;
1318 if (nextcpu < ncpus2)
1319 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg);
1321 lwkt_replymsg(&nm->base.lmsg, 0);
1325 tcp_ctlinput(netmsg_t msg)
1327 int cmd = msg->ctlinput.nm_cmd;
1328 struct sockaddr *sa = msg->ctlinput.nm_arg;
1329 struct ip *ip = msg->ctlinput.nm_extra;
1331 struct in_addr faddr;
1334 void (*notify)(struct inpcb *, int) = tcp_notify;
1338 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1342 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1343 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1346 arg = inetctlerrmap[cmd];
1347 if (cmd == PRC_QUENCH) {
1348 notify = tcp_quench;
1349 } else if (icmp_may_rst &&
1350 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1351 cmd == PRC_UNREACH_PORT ||
1352 cmd == PRC_TIMXCEED_INTRANS) &&
1354 notify = tcp_drop_syn_sent;
1355 } else if (cmd == PRC_MSGSIZE) {
1356 struct icmp *icmp = (struct icmp *)
1357 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1359 arg = ntohs(icmp->icmp_nextmtu);
1360 notify = tcp_mtudisc;
1361 } else if (PRC_IS_REDIRECT(cmd)) {
1363 notify = in_rtchange;
1364 } else if (cmd == PRC_HOSTDEAD) {
1370 th = (struct tcphdr *)((caddr_t)ip +
1371 (IP_VHL_HL(ip->ip_vhl) << 2));
1372 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1373 ip->ip_src.s_addr, th->th_sport);
1374 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1375 ip->ip_src, th->th_sport, 0, NULL);
1376 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1377 icmpseq = htonl(th->th_seq);
1378 tp = intotcpcb(inp);
1379 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1380 SEQ_LT(icmpseq, tp->snd_max))
1381 (*notify)(inp, arg);
1383 struct in_conninfo inc;
1385 inc.inc_fport = th->th_dport;
1386 inc.inc_lport = th->th_sport;
1387 inc.inc_faddr = faddr;
1388 inc.inc_laddr = ip->ip_src;
1392 syncache_unreach(&inc, th);
1396 struct netmsg_tcp_notify *nm;
1398 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1399 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1400 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1401 0, tcp_notifyall_oncpu);
1402 nm->nm_faddr = faddr;
1404 nm->nm_notify = notify;
1406 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg);
1409 lwkt_replymsg(&msg->lmsg, 0);
1415 tcp6_ctlinput(netmsg_t msg)
1417 int cmd = msg->ctlinput.nm_cmd;
1418 struct sockaddr *sa = msg->ctlinput.nm_arg;
1419 void *d = msg->ctlinput.nm_extra;
1421 void (*notify) (struct inpcb *, int) = tcp_notify;
1422 struct ip6_hdr *ip6;
1424 struct ip6ctlparam *ip6cp = NULL;
1425 const struct sockaddr_in6 *sa6_src = NULL;
1427 struct tcp_portonly {
1433 if (sa->sa_family != AF_INET6 ||
1434 sa->sa_len != sizeof(struct sockaddr_in6)) {
1439 if (cmd == PRC_QUENCH)
1440 notify = tcp_quench;
1441 else if (cmd == PRC_MSGSIZE) {
1442 struct ip6ctlparam *ip6cp = d;
1443 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1445 arg = ntohl(icmp6->icmp6_mtu);
1446 notify = tcp_mtudisc;
1447 } else if (!PRC_IS_REDIRECT(cmd) &&
1448 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1452 /* if the parameter is from icmp6, decode it. */
1454 ip6cp = (struct ip6ctlparam *)d;
1456 ip6 = ip6cp->ip6c_ip6;
1457 off = ip6cp->ip6c_off;
1458 sa6_src = ip6cp->ip6c_src;
1462 off = 0; /* fool gcc */
1467 struct in_conninfo inc;
1469 * XXX: We assume that when IPV6 is non NULL,
1470 * M and OFF are valid.
1473 /* check if we can safely examine src and dst ports */
1474 if (m->m_pkthdr.len < off + sizeof *thp)
1477 bzero(&th, sizeof th);
1478 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1480 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1481 (struct sockaddr *)ip6cp->ip6c_src,
1482 th.th_sport, cmd, arg, notify);
1484 inc.inc_fport = th.th_dport;
1485 inc.inc_lport = th.th_sport;
1486 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1487 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1489 syncache_unreach(&inc, &th);
1491 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1492 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1495 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
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 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1645 mtu = rt->rt_rmx.rmx_mtu;
1649 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1650 sizeof(struct tcpiphdr));
1653 * XXX - The following conditional probably violates the TCP
1654 * spec. The problem is that, since we don't know the
1655 * other end's MSS, we are supposed to use a conservative
1656 * default. But, if we do that, then MTU discovery will
1657 * never actually take place, because the conservative
1658 * default is much less than the MTUs typically seen
1659 * on the Internet today. For the moment, we'll sweep
1660 * this under the carpet.
1662 * The conservative default might not actually be a problem
1663 * if the only case this occurs is when sending an initial
1664 * SYN with options and data to a host we've never talked
1665 * to before. Then, they will reply with an MSS value which
1666 * will get recorded and the new parameters should get
1667 * recomputed. For Further Study.
1669 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1670 maxopd = rt->rt_rmx.rmx_mssopt;
1674 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1675 sizeof(struct tcpiphdr));
1677 if (tp->t_maxopd <= maxopd)
1679 tp->t_maxopd = maxopd;
1682 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1683 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1684 mss -= TCPOLEN_TSTAMP_APPA;
1686 /* round down to multiple of MCLBYTES */
1687 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1689 mss &= ~(MCLBYTES - 1);
1692 mss = (mss / MCLBYTES) * MCLBYTES;
1695 if (so->so_snd.ssb_hiwat < mss)
1696 mss = so->so_snd.ssb_hiwat;
1700 tp->snd_nxt = tp->snd_una;
1702 tcpstat.tcps_mturesent++;
1706 * Look-up the routing entry to the peer of this inpcb. If no route
1707 * is found and it cannot be allocated the return NULL. This routine
1708 * is called by TCP routines that access the rmx structure and by tcp_mss
1709 * to get the interface MTU.
1712 tcp_rtlookup(struct in_conninfo *inc)
1714 struct route *ro = &inc->inc_route;
1716 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1717 /* No route yet, so try to acquire one */
1718 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1720 * unused portions of the structure MUST be zero'd
1721 * out because rtalloc() treats it as opaque data
1723 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1724 ro->ro_dst.sa_family = AF_INET;
1725 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1726 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1736 tcp_rtlookup6(struct in_conninfo *inc)
1738 struct route_in6 *ro6 = &inc->inc6_route;
1740 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1741 /* No route yet, so try to acquire one */
1742 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1744 * unused portions of the structure MUST be zero'd
1745 * out because rtalloc() treats it as opaque data
1747 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1748 ro6->ro_dst.sin6_family = AF_INET6;
1749 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1750 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1751 rtalloc((struct route *)ro6);
1754 return (ro6->ro_rt);
1759 /* compute ESP/AH header size for TCP, including outer IP header. */
1761 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1769 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1771 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1776 if (inp->inp_vflag & INP_IPV6) {
1777 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1779 th = (struct tcphdr *)(ip6 + 1);
1780 m->m_pkthdr.len = m->m_len =
1781 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1782 tcp_fillheaders(tp, ip6, th);
1783 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1787 ip = mtod(m, struct ip *);
1788 th = (struct tcphdr *)(ip + 1);
1789 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1790 tcp_fillheaders(tp, ip, th);
1791 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1800 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1802 * This code attempts to calculate the bandwidth-delay product as a
1803 * means of determining the optimal window size to maximize bandwidth,
1804 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1805 * routers. This code also does a fairly good job keeping RTTs in check
1806 * across slow links like modems. We implement an algorithm which is very
1807 * similar (but not meant to be) TCP/Vegas. The code operates on the
1808 * transmitter side of a TCP connection and so only effects the transmit
1809 * side of the connection.
1811 * BACKGROUND: TCP makes no provision for the management of buffer space
1812 * at the end points or at the intermediate routers and switches. A TCP
1813 * stream, whether using NewReno or not, will eventually buffer as
1814 * many packets as it is able and the only reason this typically works is
1815 * due to the fairly small default buffers made available for a connection
1816 * (typicaly 16K or 32K). As machines use larger windows and/or window
1817 * scaling it is now fairly easy for even a single TCP connection to blow-out
1818 * all available buffer space not only on the local interface, but on
1819 * intermediate routers and switches as well. NewReno makes a misguided
1820 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1821 * then backing off, then steadily increasing the window again until another
1822 * failure occurs, ad-infinitum. This results in terrible oscillation that
1823 * is only made worse as network loads increase and the idea of intentionally
1824 * blowing out network buffers is, frankly, a terrible way to manage network
1827 * It is far better to limit the transmit window prior to the failure
1828 * condition being achieved. There are two general ways to do this: First
1829 * you can 'scan' through different transmit window sizes and locate the
1830 * point where the RTT stops increasing, indicating that you have filled the
1831 * pipe, then scan backwards until you note that RTT stops decreasing, then
1832 * repeat ad-infinitum. This method works in principle but has severe
1833 * implementation issues due to RTT variances, timer granularity, and
1834 * instability in the algorithm which can lead to many false positives and
1835 * create oscillations as well as interact badly with other TCP streams
1836 * implementing the same algorithm.
1838 * The second method is to limit the window to the bandwidth delay product
1839 * of the link. This is the method we implement. RTT variances and our
1840 * own manipulation of the congestion window, bwnd, can potentially
1841 * destabilize the algorithm. For this reason we have to stabilize the
1842 * elements used to calculate the window. We do this by using the minimum
1843 * observed RTT, the long term average of the observed bandwidth, and
1844 * by adding two segments worth of slop. It isn't perfect but it is able
1845 * to react to changing conditions and gives us a very stable basis on
1846 * which to extend the algorithm.
1849 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1857 * If inflight_enable is disabled in the middle of a tcp connection,
1858 * make sure snd_bwnd is effectively disabled.
1860 if (!tcp_inflight_enable) {
1861 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1862 tp->snd_bandwidth = 0;
1867 * Validate the delta time. If a connection is new or has been idle
1868 * a long time we have to reset the bandwidth calculator.
1871 delta_ticks = save_ticks - tp->t_bw_rtttime;
1872 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1873 tp->t_bw_rtttime = ticks;
1874 tp->t_bw_rtseq = ack_seq;
1875 if (tp->snd_bandwidth == 0)
1876 tp->snd_bandwidth = tcp_inflight_min;
1879 if (delta_ticks == 0)
1883 * Sanity check, plus ignore pure window update acks.
1885 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1889 * Figure out the bandwidth. Due to the tick granularity this
1890 * is a very rough number and it MUST be averaged over a fairly
1891 * long period of time. XXX we need to take into account a link
1892 * that is not using all available bandwidth, but for now our
1893 * slop will ramp us up if this case occurs and the bandwidth later
1896 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1897 tp->t_bw_rtttime = save_ticks;
1898 tp->t_bw_rtseq = ack_seq;
1899 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1901 tp->snd_bandwidth = bw;
1904 * Calculate the semi-static bandwidth delay product, plus two maximal
1905 * segments. The additional slop puts us squarely in the sweet
1906 * spot and also handles the bandwidth run-up case. Without the
1907 * slop we could be locking ourselves into a lower bandwidth.
1909 * Situations Handled:
1910 * (1) Prevents over-queueing of packets on LANs, especially on
1911 * high speed LANs, allowing larger TCP buffers to be
1912 * specified, and also does a good job preventing
1913 * over-queueing of packets over choke points like modems
1914 * (at least for the transmit side).
1916 * (2) Is able to handle changing network loads (bandwidth
1917 * drops so bwnd drops, bandwidth increases so bwnd
1920 * (3) Theoretically should stabilize in the face of multiple
1921 * connections implementing the same algorithm (this may need
1924 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1925 * be adjusted with a sysctl but typically only needs to be on
1926 * very slow connections. A value no smaller then 5 should
1927 * be used, but only reduce this default if you have no other
1931 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1932 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1933 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1936 if (tcp_inflight_debug > 0) {
1938 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1940 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1941 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1944 if ((long)bwnd < tcp_inflight_min)
1945 bwnd = tcp_inflight_min;
1946 if (bwnd > tcp_inflight_max)
1947 bwnd = tcp_inflight_max;
1948 if ((long)bwnd < tp->t_maxseg * 2)
1949 bwnd = tp->t_maxseg * 2;
1950 tp->snd_bwnd = bwnd;
1953 #ifdef TCP_SIGNATURE
1955 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1957 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1958 * When called from tcp_input(), we can be sure that th_sum has been
1959 * zeroed out and verified already.
1961 * Return 0 if successful, otherwise return -1.
1963 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1964 * search with the destination IP address, and a 'magic SPI' to be
1965 * determined by the application. This is hardcoded elsewhere to 1179
1966 * right now. Another branch of this code exists which uses the SPD to
1967 * specify per-application flows but it is unstable.
1970 tcpsignature_compute(
1971 struct mbuf *m, /* mbuf chain */
1972 int len, /* length of TCP data */
1973 int optlen, /* length of TCP options */
1974 u_char *buf, /* storage for MD5 digest */
1975 u_int direction) /* direction of flow */
1977 struct ippseudo ippseudo;
1981 struct ipovly *ipovly;
1982 struct secasvar *sav;
1985 struct ip6_hdr *ip6;
1986 struct in6_addr in6;
1992 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
1993 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
1995 * Extract the destination from the IP header in the mbuf.
1997 ip = mtod(m, struct ip *);
1999 ip6 = NULL; /* Make the compiler happy. */
2002 * Look up an SADB entry which matches the address found in
2005 switch (IP_VHL_V(ip->ip_vhl)) {
2007 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2008 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2011 case (IPV6_VERSION >> 4):
2012 ip6 = mtod(m, struct ip6_hdr *);
2013 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2014 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2023 kprintf("%s: SADB lookup failed\n", __func__);
2029 * Step 1: Update MD5 hash with IP pseudo-header.
2031 * XXX The ippseudo header MUST be digested in network byte order,
2032 * or else we'll fail the regression test. Assume all fields we've
2033 * been doing arithmetic on have been in host byte order.
2034 * XXX One cannot depend on ipovly->ih_len here. When called from
2035 * tcp_output(), the underlying ip_len member has not yet been set.
2037 switch (IP_VHL_V(ip->ip_vhl)) {
2039 ipovly = (struct ipovly *)ip;
2040 ippseudo.ippseudo_src = ipovly->ih_src;
2041 ippseudo.ippseudo_dst = ipovly->ih_dst;
2042 ippseudo.ippseudo_pad = 0;
2043 ippseudo.ippseudo_p = IPPROTO_TCP;
2044 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2045 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2046 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2047 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2051 * RFC 2385, 2.0 Proposal
2052 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2053 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2054 * extended next header value (to form 32 bits), and 32-bit segment
2056 * Note: Upper-Layer Packet Length comes before Next Header.
2058 case (IPV6_VERSION >> 4):
2060 in6_clearscope(&in6);
2061 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2063 in6_clearscope(&in6);
2064 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2065 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2066 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2068 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2069 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2070 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2072 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2073 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2074 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2083 * Step 2: Update MD5 hash with TCP header, excluding options.
2084 * The TCP checksum must be set to zero.
2086 savecsum = th->th_sum;
2088 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2089 th->th_sum = savecsum;
2091 * Step 3: Update MD5 hash with TCP segment data.
2092 * Use m_apply() to avoid an early m_pullup().
2095 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2097 * Step 4: Update MD5 hash with shared secret.
2099 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2100 MD5Final(buf, &ctx);
2101 key_sa_recordxfer(sav, m);
2107 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2110 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2113 #endif /* TCP_SIGNATURE */